From 5626bbe7f5ef6f732c4a9435404c9cc18e62e8a7 Mon Sep 17 00:00:00 2001 From: "Jordan (via Claude)" Date: Sun, 26 Apr 2026 23:23:16 -0500 Subject: [PATCH 1/3] Mikey (val_bpb 0.86548 mean across 3 seeds, 15.65MB) Co-Authored-By: Claude Opus 4.7 (1M context) --- .../2026-04-27_Mikey/README.md | 12 + .../2026-04-27_Mikey/submission.json | 30 + .../2026-04-27_Mikey/train_gpt.py | 2655 ++++++++++++++++ .../2026-04-27_Mikey/train_seed300.log | 2801 +++++++++++++++++ .../2026-04-27_Mikey/train_seed42.log | 2801 +++++++++++++++++ .../2026-04-27_Mikey/train_seed444.log | 2801 +++++++++++++++++ 6 files changed, 11100 insertions(+) create mode 100644 records/track_10min_16mb/2026-04-27_Mikey/README.md create mode 100644 records/track_10min_16mb/2026-04-27_Mikey/submission.json create mode 100644 records/track_10min_16mb/2026-04-27_Mikey/train_gpt.py create mode 100644 records/track_10min_16mb/2026-04-27_Mikey/train_seed300.log create mode 100644 records/track_10min_16mb/2026-04-27_Mikey/train_seed42.log create mode 100644 records/track_10min_16mb/2026-04-27_Mikey/train_seed444.log diff --git a/records/track_10min_16mb/2026-04-27_Mikey/README.md b/records/track_10min_16mb/2026-04-27_Mikey/README.md new file mode 100644 index 0000000000..51e267444d --- /dev/null +++ b/records/track_10min_16mb/2026-04-27_Mikey/README.md @@ -0,0 +1,12 @@ +# Mikey + +| seed | val_bpb (sliding) | bytes | +|------|-------------------|-------| +| 42 | 0.86503709 | 15,639,737 | +| 300 | 0.86698133 | 15,594,375 | +| 444 | 0.86441066 | 15,653,512 | +| **mean** | **0.86547636** | — | + +``` +torchrun --standalone --nproc_per_node=8 train_gpt.py +``` diff --git a/records/track_10min_16mb/2026-04-27_Mikey/submission.json b/records/track_10min_16mb/2026-04-27_Mikey/submission.json new file mode 100644 index 0000000000..1970614510 --- /dev/null +++ b/records/track_10min_16mb/2026-04-27_Mikey/submission.json @@ -0,0 +1,30 @@ +{ + "author": "newjordan", + "github_id": "newjordan", + "name": "Mikey", + "date": "2026-04-27", + "track": "10min_16mb", + "val_bpb": 0.86547636, + "val_bpb_std": 0.00109, + "seeds": [42, 300, 444], + "seed_results": { + "42": {"val_bpb": 0.86503709, "artifact_bytes": 15639737}, + "300": {"val_bpb": 0.86698133, "artifact_bytes": 15594375}, + "444": {"val_bpb": 0.86441066, "artifact_bytes": 15653512} + }, + "hardware": "8xH100 80GB SXM", + "pytorch_version": "2.11.0+cu130", + "technique_summary": "", + "compliance": { + "train_under_600s": true, + "artifact_under_16mb": true, + "eval_under_600s": true, + "no_slot": true, + "no_pre_quant_ttt": true, + "no_etlb": true, + "no_ngram_cache": true, + "score_first_ttt": true, + "three_seeds": true + }, + "attribution": {} +} diff --git a/records/track_10min_16mb/2026-04-27_Mikey/train_gpt.py b/records/track_10min_16mb/2026-04-27_Mikey/train_gpt.py new file mode 100644 index 0000000000..cd29d607f9 --- /dev/null +++ b/records/track_10min_16mb/2026-04-27_Mikey/train_gpt.py @@ -0,0 +1,2655 @@ +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +from collections import OrderedDict +from pathlib import Path +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP + +os.environ.setdefault("RUN_ID", "mikey_8x_seed444_" + time.strftime("%Y%m%d_%H%M%S")) + +try: + import triton + import triton.language as tl +except ImportError: + triton = None + tl = None +try: + from flash_attn_interface import flash_attn_func as flash_attn_3_func +except ImportError: + flash_attn_3_func = None +# Compression: brotli-11 + byte-shuffle is the salvage_v2 preferred path (PR #1493 recipe). +# Falls back to zstd then zlib so this file still runs if brotli isn't installed. +_brotli_module = None +_zstandard_module = None +_zlib_module = None +try: + import brotli as _brotli_module + _COMPRESSOR = "brotli" +except ImportError: + try: + import zstandard as _zstandard_module + _COMPRESSOR = "zstd" + import warnings + warnings.warn("brotli not found — falling back to zstd (~1MB+ larger). pip install brotli") + except ImportError: + import zlib as _zlib_module + import warnings + _COMPRESSOR = "zlib" + warnings.warn("brotli/zstandard not found — falling back to zlib. Artifact will be much larger! pip install brotli") +# Backwards-compat shims so any remaining `zstandard.*` or `_zlib_module.*` references still work. +if _zstandard_module is not None: + zstandard = _zstandard_module +if _zlib_module is None: + import zlib as _zlib_module # always available; used by zlib fallback path +# --- Byte-shuffle (de-interleave) wrapper from PR #1493: improves brotli ratio on quantized payloads. --- +_BSHF_MAGIC = b"BSHF" +def _byte_shuffle(data: bytes, stride: int = 2) -> bytes: + if stride <= 1 or len(data) < stride: + return data + src = np.frombuffer(data, dtype=np.uint8) + n = len(src) + out = np.empty(n, dtype=np.uint8) + dest_off = 0 + for pos in range(stride): + chunk = src[pos::stride] + out[dest_off:dest_off + len(chunk)] = chunk + dest_off += len(chunk) + return _BSHF_MAGIC + bytes([stride]) + out.tobytes() +def _byte_unshuffle(data: bytes) -> bytes: + if len(data) < 5 or data[:4] != _BSHF_MAGIC: + return data + stride = data[4] + if stride < 2: + return data[5:] + payload = np.frombuffer(data, dtype=np.uint8, offset=5) + n = len(payload) + out = np.empty(n, dtype=np.uint8) + src_off = 0 + for pos in range(stride): + chunk_len = n // stride + (1 if pos < n % stride else 0) + out[pos::stride][:chunk_len] = payload[src_off:src_off + chunk_len] + src_off += chunk_len + return out.tobytes() +def _compress_blob(raw: bytes) -> bytes: + if _COMPRESSOR == "brotli": + return _brotli_module.compress(_byte_shuffle(raw, stride=2), quality=11) + elif _COMPRESSOR == "zstd": + return _zstandard_module.ZstdCompressor(level=22).compress(raw) + else: + return _zlib_module.compress(raw, 9) +def _decompress_blob(blob: bytes) -> bytes: + if _COMPRESSOR == "brotli": + return _byte_unshuffle(_brotli_module.decompress(blob)) + elif _COMPRESSOR == "zstd": + return _zstandard_module.ZstdDecompressor().decompress(blob) + else: + return _zlib_module.decompress(blob) + +if os.environ.get("TORCHDYNAMO_SUPPRESS_ERRORS", "0") == "1": + import torch._dynamo + torch._dynamo.config.suppress_errors = True +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp4096") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_4096_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 444)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 4096)) + num_layers = int(os.environ.get("NUM_LAYERS", 12)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + lawa_enabled = bool(int(os.environ.get("LAWA_ENABLED", "0"))) + lawa_k = int(os.environ.get("LAWA_K", 10)) + lawa_freq = int(os.environ.get("LAWA_FREQ", 100)) + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + trigram_enabled = bool(int(os.environ.get("TRIGRAM", "0"))) # TrigramHash (off by default, risky) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 11)) # XSA on ALL layers (our novel contribution) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + gated_attention = bool(int(os.environ.get("GATED_ATTENTION", "0"))) + value_residual = bool(int(os.environ.get("VALUE_RESIDUAL", "0"))) # VRL with sigmoid gates (off by default, risky) + attn_scale_init = float(os.environ.get("ATTN_SCALE_INIT", 1.0)) + mlp_scale_init = float(os.environ.get("MLP_SCALE_INIT", 1.0)) + resid_mix_x_init = float(os.environ.get("RESID_MIX_X_INIT", 1.0)) + resid_mix_x0_init = float(os.environ.get("RESID_MIX_X0_INIT", 0.0)) + complement_alpha = float(os.environ.get("COMPLEMENT_ALPHA", "0")) + ngram_eval_order = int(os.environ.get("NGRAM_EVAL_ORDER", 0)) + ngram_eval_min_order = int(os.environ.get("NGRAM_EVAL_MIN_ORDER", 2)) + ngram_eval_alpha = float(os.environ.get("NGRAM_EVAL_ALPHA", 0.30)) + ngram_eval_adaptive = bool(int(os.environ.get("NGRAM_EVAL_ADAPTIVE", "1"))) + ngram_eval_alpha_min = float(os.environ.get("NGRAM_EVAL_ALPHA_MIN", 0.05)) + ngram_eval_alpha_max = float(os.environ.get("NGRAM_EVAL_ALPHA_MAX", 0.60)) + ngram_eval_entropy_center = float(os.environ.get("NGRAM_EVAL_ENTROPY_CENTER", 4.0)) + ngram_eval_entropy_scale = float(os.environ.get("NGRAM_EVAL_ENTROPY_SCALE", 2.0)) + ngram_eval_min_count = int(os.environ.get("NGRAM_EVAL_MIN_COUNT", 2)) + ngram_eval_buckets = int(os.environ.get("NGRAM_EVAL_BUCKETS", 4_194_304)) + ngram_eval_max_seconds = float(os.environ.get("NGRAM_EVAL_MAX_SECONDS", 0.0)) + ngram_entropy_shift = bool(int(os.environ.get("NGRAM_ENTROPY_SHIFT", "0"))) + ngram_order_mults_str = os.environ.get("NGRAM_ORDER_MULTS", "") + cubric_cadence = int(os.environ.get("CUBRIC_CADENCE", 0)) + skip_final_eval = bool(int(os.environ.get("SKIP_FINAL_EVAL", "0"))) + post_ema_diagnostic = bool(int(os.environ.get("POST_EMA_DIAGNOSTIC", "1"))) + compile_enabled = bool(int(os.environ.get("COMPILE_ENABLED", "1"))) + compile_mode = os.environ.get("COMPILE_MODE", "").strip() + compile_fullgraph = bool(int(os.environ.get("COMPILE_FULLGRAPH", "1"))) + mlp_kernel_mode = os.environ.get("MLP_KERNEL_MODE", "").strip().lower() + loader_mode = os.environ.get("LOADER_MODE", "coprime").strip().lower() + coprime_max_loaded_shards = int(os.environ.get("COPRIME_MAX_LOADED_SHARDS", 143)) + coprime_shards_per_batch = int(os.environ.get("COPRIME_SHARDS_PER_BATCH", 1)) + coprime_shard_hold_steps = int(os.environ.get("COPRIME_SHARD_HOLD_STEPS", 64)) + + +def maybe_compile(fn_or_module, *, enabled: bool, fullgraph: bool, mode: str = ""): + if not enabled: + return fn_or_module + kwargs = dict(dynamic=False, fullgraph=fullgraph) + if mode: + kwargs["mode"] = mode + return torch.compile(fn_or_module, **kwargs) + + +if triton is not None: + @triton.jit + def _leaky_relu_sq_forward_kernel(x_ptr, y_ptr, n_elements, BLOCK_SIZE: tl.constexpr): + pid = tl.program_id(0) + offsets = pid * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE) + mask = offsets < n_elements + x = tl.load(x_ptr + offsets, mask=mask, other=0.0).to(tl.float32) + a = tl.where(x >= 0, x, 0.5 * x) + y = a * a + tl.store(y_ptr + offsets, y, mask=mask) + + @triton.jit + def _leaky_relu_sq_backward_kernel(x_ptr, grad_out_ptr, grad_in_ptr, n_elements, BLOCK_SIZE: tl.constexpr): + pid = tl.program_id(0) + offsets = pid * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE) + mask = offsets < n_elements + x = tl.load(x_ptr + offsets, mask=mask, other=0.0).to(tl.float32) + grad_out = tl.load(grad_out_ptr + offsets, mask=mask, other=0.0).to(tl.float32) + a = tl.where(x >= 0, x, 0.5 * x) + slope = tl.where(x >= 0, 1.0, 0.5) + grad_in = grad_out * (2.0 * a * slope) + tl.store(grad_in_ptr + offsets, grad_in, mask=mask) + + +class TritonLeakyReluSqFn(torch.autograd.Function): + @staticmethod + def forward(ctx, x: Tensor) -> Tensor: + if triton is None or not x.is_cuda: + a = F.leaky_relu(x, negative_slope=0.5) + ctx.save_for_backward(x) + return a.square() + x_contig = x.contiguous() + y = torch.empty_like(x_contig) + n_elements = x_contig.numel() + grid = lambda meta: (triton.cdiv(n_elements, meta["BLOCK_SIZE"]),) + _leaky_relu_sq_forward_kernel[grid](x_contig, y, n_elements, BLOCK_SIZE=1024) + ctx.save_for_backward(x_contig) + return y + + @staticmethod + def backward(ctx, grad_out: Tensor) -> tuple[Tensor]: + (x,) = ctx.saved_tensors + if triton is None or not grad_out.is_cuda: + a = F.leaky_relu(x, negative_slope=0.5) + slope = torch.where(x >= 0, torch.ones_like(x), torch.full_like(x, 0.5)) + return (grad_out * (2.0 * a * slope),) + grad_out_contig = grad_out.contiguous() + grad_in = torch.empty_like(grad_out_contig) + n_elements = grad_out_contig.numel() + grid = lambda meta: (triton.cdiv(n_elements, meta["BLOCK_SIZE"]),) + _leaky_relu_sq_backward_kernel[grid](x, grad_out_contig, grad_in, n_elements, BLOCK_SIZE=1024) + return (grad_in,) + + +def leaky_relu_sq(x: Tensor, kernel_mode: str = "") -> Tensor: + if kernel_mode == "triton_act": + return TritonLeakyReluSqFn.apply(x) + a = F.leaky_relu(x, negative_slope=0.5) + return a.square() + +class TrainNgramTracker: + """Complementary training: track bigram stats, downweight tokens n-grams can predict.""" + def __init__(self, vocab_size: int, device: torch.device, complement_alpha: float = 0.5): + self.V = vocab_size + self.alpha = complement_alpha + self.bi_counts = torch.zeros(vocab_size, vocab_size, device=device, dtype=torch.float32) + self.bi_totals = torch.zeros(vocab_size, device=device, dtype=torch.float32) + @torch.no_grad() + def update(self, x: Tensor, y: Tensor): + xf = x.reshape(-1) + yf = y.reshape(-1) + ones = torch.ones(xf.numel(), device=xf.device, dtype=torch.float32) + self.bi_counts.reshape(-1).scatter_add_(0, xf * self.V + yf, ones) + self.bi_totals.scatter_add_(0, xf, ones) + def get_weights(self, x: Tensor, y: Tensor) -> Tensor: + xf = x.reshape(-1) + yf = y.reshape(-1) + total = self.bi_totals[xf] + count = self.bi_counts.reshape(-1)[xf * self.V + yf] + ngram_prob = count / (total + 1) + return (1.0 - self.alpha * ngram_prob).clamp(min=0.1) + +# --- Batched Newton-Schulz orthogonalization --- + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 5, eps: float = 1e-7) -> Tensor: + """Batched Newton-Schulz orthogonalization. G: (B,M,N) or (M,N).""" + a, b, c = (3.4445, -4.7750, 2.0315) + was_2d = G.ndim == 2 + if was_2d: + G = G.unsqueeze(0) + X = G.bfloat16() + transposed = X.size(-2) > X.size(-1) + if transposed: + X = X.mT + X = X / (X.norm(dim=(-2, -1), keepdim=True) + eps) + for _ in range(steps): + A = X @ X.mT + B = b * A + c * (A @ A) + X = a * X + B @ X + if transposed: + X = X.mT + if was_2d: + X = X.squeeze(0) + return X + +# --- Parallel Muon optimizer --- + +class Muon(torch.optim.Optimizer): + """Parallel Muon: post-backward reduce-scatter -> local NS5 -> all-gather. + + No DDP for bank params. After backward, this optimizer: + 1. Launches async reduce-scatter for all banks (biggest first) + 2. Returns control so Adam can step on small params while RS is in-flight + 3. Waits for each RS, runs local NS5 on the shard, launches async all-gather + 4. Each all-gather overlaps with next bank's NS5 + """ + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + self._built = False + + def _build(self): + self._distributed = dist.is_available() and dist.is_initialized() + self._world_size = dist.get_world_size() if self._distributed else 1 + self._rank = dist.get_rank() if self._distributed else 0 + ws = self._world_size + + self._bank_meta = [] + for group in self.param_groups: + for p in group["params"]: + B = p.shape[0] + padded_B = ((B + ws - 1) // ws) * ws + shard_B = padded_B // ws + tail = p.shape[1:] + dev = p.device + self._bank_meta.append({ + 'p': p, + 'B': B, + 'padded_grad': torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + 'shard': torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + 'shard_mom': torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + 'full_update': torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + 'scale': max(1, p.shape[-2] / p.shape[-1]) ** 0.5, + }) + # Sort by size descending -- launch biggest reduce-scatters first + self._bank_meta.sort(key=lambda m: -m['p'].numel()) + self._built = True + + def launch_reduce_scatters(self): + """Phase 1: launch async reduce-scatter for all banks. Call right after backward.""" + if not self._built: + self._build() + if not self._distributed: + return + self._rs_futures = [] + for m in self._bank_meta: + p = m['p'] + if p.grad is None: + self._rs_futures.append(None) + continue + pg = m['padded_grad'] + pg[:m['B']].copy_(p.grad.bfloat16()) + if pg.shape[0] > m['B']: + pg[m['B']:].zero_() + fut = dist.reduce_scatter_tensor(m['shard'], pg, op=dist.ReduceOp.AVG, async_op=True) + self._rs_futures.append(fut) + + @torch.no_grad() + def step(self, closure=None): + """Phase 3: wait for RS, local NS5, all-gather. Call AFTER Adam steps.""" + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + if not self._built: + self._build() + + for group in self.param_groups: + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + wd = group.get("weight_decay", 0.0) + + prev_ag_handle = None + prev_m = None + + sharded = self._distributed and hasattr(self, '_rs_futures') + + for i, m in enumerate(self._bank_meta): + p = m['p'] + if p.grad is None: + continue + + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m['p'] + upd = prev_m['full_update'][:prev_m['B']] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m['scale']) + + if sharded and self._rs_futures[i] is not None: + self._rs_futures[i].wait() + g = m['shard'] + buf = m['shard_mom'] + else: + g = p.grad.bfloat16() + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + + buf.mul_(momentum).add_(g) + if nesterov: + update = g.add(buf, alpha=momentum) + else: + update = buf + + update = zeropower_via_newtonschulz5(update, steps=backend_steps) + + if sharded: + prev_ag_handle = dist.all_gather_into_tensor( + m['full_update'], update, async_op=True) + prev_m = m + else: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + p.add_(update.to(dtype=p.dtype), alpha=-lr * m['scale']) + + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m['p'] + upd = prev_m['full_update'][:prev_m['B']] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m['scale']) + + if hasattr(self, '_rs_futures'): + del self._rs_futures + + return loss + +# --- Tokenizer evaluation helpers --- + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) + +# --- Quantization helpers --- + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale,attn_gate,vr_lambda", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if "f1_corr_in" in name or "f1_corr_out" in name: + return "aux" + if "qo_bank" in name or "kv_bank" in name: + return "attn" + if "mlp_up_bank" in name or "mlp_down_bank" in name: + return "mlp" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +# GPTQ: Hessian-aware quantization with column-wise error compensation +def _find_best_row_scales(W: Tensor, clip_range: int = 31) -> Tensor: + t32 = W.float() + best_s = t32.abs().amax(dim=1) / clip_range + best_s = best_s.clamp_min(1.0 / clip_range) + best_err = torch.full((t32.shape[0],), float('inf')) + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range) + q = torch.clamp(torch.round(t32 / s[:, None]), -clip_range, clip_range) + recon = q * s[:, None] + err = (t32 - recon).pow(2).mean(dim=1) + improved = err < best_err + best_s[improved] = s[improved] + best_err[improved] = err[improved] + return best_s +def gptq_quantize_weight(W: Tensor, H: Tensor, clip_range: int = 31, + block_size: int = 64, percdamp: float = 0.002) -> tuple[Tensor, Tensor]: + """GPTQ: quantize weight matrix W using Hessian H = X^T X for error compensation. + Returns (quantized_int8, scale_fp16) in int6 range [-clip_range, clip_range].""" + W = W.float().clone() + rows, cols = W.shape + row_scale = _find_best_row_scales(W, clip_range) + H = H.float().clone() + damp = percdamp * H.diag().mean() + H.diagonal().add_(damp) + perm = torch.argsort(H.diag()) + invperm = torch.argsort(perm) + W = W[:, perm] + H = H[perm][:, perm] + try: + L = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(L) + except torch._C._LinAlgError: + Hinv = torch.diag(1.0 / H.diag().clamp_min(1e-6)) + Q = torch.zeros(rows, cols, dtype=torch.int8) + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + W_block = W[:, i1:i2].clone() + Hinv_block = Hinv[i1:i2, i1:i2] + Err = torch.zeros_like(W_block) + for j in range(i2 - i1): + w_col = W_block[:, j] + h_inv_jj = Hinv_block[j, j].clamp_min(1e-8) + q_col = torch.clamp(torch.round(w_col / row_scale), -clip_range, clip_range) + deq_col = q_col * row_scale + Q[:, i1 + j] = q_col.to(torch.int8) + err = (w_col - deq_col) / h_inv_jj + Err[:, j] = err + if j + 1 < i2 - i1: + W_block[:, j + 1:] -= err.unsqueeze(1) * Hinv_block[j, j + 1:].unsqueeze(0) + if i2 < cols: + W[:, i2:] -= Err @ Hinv[i1:i2, i2:] + Q = Q[:, invperm] + return Q, row_scale.to(torch.float16) +def gptq_calibrate(model: nn.Module, train_pattern: str, device: torch.device, + n_samples: int = 256, seq_len: int = 2048) -> dict[str, Tensor]: + """Collect Hessian H = X^T X for each linear layer using training data.""" + hessians: dict[str, Tensor] = {} + n_seen: dict[str, int] = {} + hooks = [] + def make_hook(name: str): + def hook_fn(module, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + if name not in hessians: + hessians[name] = torch.zeros(x.shape[1], x.shape[1], device=x.device, dtype=torch.float32) + n_seen[name] = 0 + hessians[name].addmm_(x.t(), x) + n_seen[name] += x.shape[0] + return hook_fn + for name, module in model.named_modules(): + if isinstance(module, (nn.Linear, CastedLinear)): + hooks.append(module.register_forward_hook(make_hook(name))) + stream = TokenStream(train_pattern) + model.eval() + with torch.no_grad(): + for _ in range(n_samples): + tokens = stream.take(seq_len + 1).to(device=device, dtype=torch.int64) + x = tokens[:-1].unsqueeze(0) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + model.forward_logits(x) + for h in hooks: + h.remove() + for name in hessians: + hessians[name] /= max(n_seen[name], 1) + model.train() + return hessians +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale +def quantize_int5_per_row(t: Tensor, clip_range: int = 15) -> tuple[Tensor, Tensor]: + """int5 (signed, range [-15, 15]) per-row quant. Modeled on quantize_int6_per_row. + Returned int8 tensor holds values in [-15, 15] — leaves the high 3 bits as zeros, which + brotli compresses very efficiently. Dequant path is identical (q.float() * scale).""" + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale +def _classify_param_fine(name: str) -> str: + """Finer classifier than _classify_param, splitting attention/MLP banks for mixed-int policy. + Returns one of: embed, qo (Q+O bank), kv (K+V bank), mlp_up (mlp_fc), mlp_down (mlp_proj), + aux, attn_other, mlp_other, other. Categories `qo`/`kv` map to attention; `mlp_up`/`mlp_down` + map to MLP. The split lets us route mlp banks to int5 while keeping attn at int6.""" + if "tok_emb" in name or "lm_head" in name: + return "embed" + if "f1_corr_in" in name or "f1_corr_out" in name: + return "aux" + if "qo_bank" in name: + return "qo" + if "kv_bank" in name: + return "kv" + if "mlp_up_bank" in name: + return "mlp_up" + if "mlp_down_bank" in name: + return "mlp_down" + if ".mlp." in name: + return "mlp_other" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn_other" + return "other" +# salvage_v2 mixed-int policy (applied at the Tensor level via _classify_param_fine): +# - mlp_down_bank (mlp_proj, MOST quant-tolerant per 11-day collate) -> int5 (clip_range 15) +# - mlp_up_bank (mlp_fc, also tolerant) -> int5 +# - qo_bank, kv_bank (attention; LEAST quant-tolerant) -> int6 (clip_range 31) +# - tok_emb / lm_head (embed) -> int6 (matches seed; keeping +# attn/embed at int6 for +# quant safety on this +# first salvage attempt) +# Bytes savings: int5 keeps the int8 storage container (no bit-packing) but the high 3 bits are +# forced zero, giving brotli a compressible pattern. Combined with the byte-shuffle wrapper, +# expected savings vs uniform-int6+zstd is roughly the int5 bit ratio (5/6 = -17%) APPLIED only +# to MLP banks (which are ~60-65% of the model parameter mass for a 10L Rascal at mlp_mult=3.0). +# Net est: ~0.62 * 0.17 = ~10-11% blob shrink from int5 alone, plus brotli ~5-8% over zstd. +DEFAULT_INT5_CATS = {"mlp_down", "mlp_up"} +DEFAULT_INT6_CATS = {"qo", "kv", "attn_other", "mlp_other", "aux"} +def mixed_quantize_int6_gptq(state_dict: dict[str, Tensor], int6_cats: set[str], + hessians: dict[str, Tensor], + int5_cats: set[str] | None = None) -> tuple[dict, dict]: + """Mixed-int (int5/int6/int8) quant with GPTQ for matrix categories when Hessian available. + `int6_cats` and `int5_cats` use FINE-grained category names from `_classify_param_fine` + (qo, kv, mlp_up, mlp_down, attn_other, mlp_other, aux, embed). For backwards-compat with + the old uniform-int6 caller, the legacy coarse names {'mlp','attn','aux','embed'} are also + accepted in `int6_cats` and expand to their fine-grained children. `int5_cats` always uses + fine names. If `int5_cats` is None, defaults to DEFAULT_INT5_CATS (the salvage_v2 policy).""" + if int5_cats is None: + int5_cats = set(DEFAULT_INT5_CATS) + # Expand legacy coarse names so the existing call signature keeps working. + _LEGACY = { + "mlp": {"mlp_up", "mlp_down", "mlp_other"}, + "attn": {"qo", "kv", "attn_other"}, + "aux": {"aux"}, + "embed": {"embed"}, + } + expanded_int6: set[str] = set() + for c in int6_cats: + expanded_int6.update(_LEGACY.get(c, {c})) + # int5 categories take precedence over int6 for the same fine name. + expanded_int6 -= int5_cats + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + gptq_count, naive_count, int5_count = 0, 0, 0 + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param_fine(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int5_cats and t.ndim == 2: + module_name = name.rsplit(".weight", 1)[0] if name.endswith(".weight") else name + H = hessians.get(module_name) + if H is not None and H.shape[0] == t.shape[1]: + q, s = gptq_quantize_weight(t, H.cpu(), clip_range=15) + gptq_count += 1 + else: + q, s = quantize_int5_per_row(t) + naive_count += 1 + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int5"} + int5_count += 1 + elif cat in int5_cats and t.ndim >= 1: + t_2d = t.reshape(-1, t.shape[-1]) if t.ndim > 2 else t + q, s = quantize_int5_per_row(t_2d) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int5"} + naive_count += 1 + int5_count += 1 + elif cat in expanded_int6 and t.ndim == 2: + module_name = name.rsplit(".weight", 1)[0] if name.endswith(".weight") else name + H = hessians.get(module_name) + if H is not None and H.shape[0] == t.shape[1]: + q, s = gptq_quantize_weight(t, H.cpu()) + gptq_count += 1 + else: + q, s = quantize_int6_per_row(t) + naive_count += 1 + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + elif cat in expanded_int6 and t.ndim >= 1: + t_2d = t.reshape(-1, t.shape[-1]) if t.ndim > 2 else t + q, s = quantize_int6_per_row(t_2d) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + naive_count += 1 + else: + t_q = t.reshape(-1, t.shape[-1]) if t.ndim > 2 else t + q, s = quantize_float_tensor(t_q) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + print(f"gptq_quantize: {gptq_count} GPTQ layers, {naive_count} naive layers, {int5_count} int5 layers", flush=True) + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + val = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + val = (q.float() * float(s.item())).to(orig_dtype) + out[name] = val.reshape(orig.shape) if val.shape != orig.shape else val + return out + +# --- Data loading --- + +SHARD_HEADER_DTYPE = np.dtype(" dict[str, int]: + header = np.fromfile(file, dtype=SHARD_HEADER_DTYPE, count=SHARD_HEADER_WORDS) + if header.size != SHARD_HEADER_WORDS or int(header[0]) != SHARD_MAGIC or int(header[1]) != SHARD_VERSION: + raise ValueError(f"Unexpected shard header for {file}") + return {"num_tokens": int(header[2])} + +def load_data_shard(file: Path) -> Tensor: + header = read_data_shard_header(file) + num_tokens = header["num_tokens"] + expected_size = SHARD_HEADER_BYTES + num_tokens * SHARD_TOKEN_DTYPE.itemsize + if file.stat().st_size != expected_size: + raise ValueError(f"Shard size mismatch for {file}: expected {expected_size} bytes") + tokens_np = np.fromfile(file, dtype=SHARD_TOKEN_DTYPE, count=num_tokens, offset=SHARD_HEADER_BYTES) + if tokens_np.size != num_tokens: + raise ValueError(f"Short read for {file}") + return torch.from_numpy(tokens_np.astype(np.uint16, copy=False)) + +def choose_coprime_stride(modulus: int, salt: int) -> int: + if modulus <= 1: + return 1 + candidate = abs(salt) % modulus + if candidate == 0: + candidate = 1 + while math.gcd(candidate, modulus) != 1: + candidate += 1 + if candidate >= modulus: + candidate = 1 + return candidate + +class TokenStream: + def __init__(self, pattern: str): + self.files = [Path(p) for p in sorted(glob.glob(pattern))] + if not self.files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + self.file_idx = 0 + self.tokens = load_data_shard(self.files[0]) + self.pos = 0 + def _advance_file(self) -> None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def describe(self) -> str: + return f"loader:sequential shards:{len(self.stream.files)}" + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +class CoprimeDistributedTokenLoader: + """Shard-aware block sampler with deterministic coprime walks.""" + def __init__( + self, + pattern: str, + rank: int, + world_size: int, + device: torch.device, + seq_len: int, + seed: int, + max_loaded_shards: int, + shards_per_batch: int, + shard_hold_steps: int, + ): + self.rank = rank + self.world_size = world_size + self.device = device + self.seq_len = seq_len + self.seed = seed + self.token_offsets = torch.arange(seq_len + 1, dtype=torch.int64) + self.cache: OrderedDict[Path, Tensor] = OrderedDict() + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + self.shards: list[dict[str, int | Path]] = [] + for shard_idx, file in enumerate(files): + header = read_data_shard_header(file) + num_blocks = (header["num_tokens"] - 1) // seq_len + if num_blocks <= 0: + continue + self.shards.append( + { + "file": file, + "num_blocks": num_blocks, + "offset": (seed * 131 + shard_idx * 17) % num_blocks, + "stride": choose_coprime_stride(num_blocks, seed * 29 + shard_idx * 7 + 1), + } + ) + if not self.shards: + raise ValueError(f"No usable shards found for seq_len={seq_len}") + self.num_shards = len(self.shards) + self.max_loaded_shards = max(1, min(max_loaded_shards, self.num_shards)) + self.shards_per_batch = max(1, min(shards_per_batch, self.num_shards)) + self.shard_hold_steps = max(1, shard_hold_steps) + self.batch_shard_stride = choose_coprime_stride(self.num_shards, seed * 41 + 3) + self.batch_idx = 0 + self.shard_visits = [0 for _ in range(self.num_shards)] + def _get_tokens(self, file: Path) -> Tensor: + cached = self.cache.get(file) + if cached is not None: + self.cache.move_to_end(file) + return cached + # CPU advanced indexing is not implemented for uint16, so cache coprime-loader + # shards in int32 and cast to int64 only after batch assembly. + tokens = load_data_shard(file).to(dtype=torch.int32) + if len(self.cache) >= self.max_loaded_shards: + self.cache.popitem(last=False) + self.cache[file] = tokens + return tokens + def _sample_sequences(self, shard_idx: int, count: int) -> Tensor: + shard = self.shards[shard_idx] + num_blocks = int(shard["num_blocks"]) + offset = int(shard["offset"]) + stride = int(shard["stride"]) + visits = self.shard_visits[shard_idx] + block_ids = ( + offset + + (visits + torch.arange(count, dtype=torch.int64)) * stride + ) % num_blocks + self.shard_visits[shard_idx] += count + token_starts = block_ids * self.seq_len + gather_idx = token_starts.unsqueeze(1) + self.token_offsets.unsqueeze(0) + tokens = self._get_tokens(shard["file"]) + return tokens[gather_idx] + def describe(self) -> str: + total_blocks = sum(int(shard["num_blocks"]) for shard in self.shards) + return ( + f"loader:coprime shards:{self.num_shards} blocks:{total_blocks} " + f"seq_len:{self.seq_len} shards_per_batch:{self.shards_per_batch} " + f"cache:{self.max_loaded_shards} batch_stride:{self.batch_shard_stride} " + f"hold_steps:{self.shard_hold_steps}" + ) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + if seq_len != self.seq_len: + raise ValueError(f"Coprime loader was built for seq_len={self.seq_len}, got {seq_len}") + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + if local_tokens % seq_len != 0: + raise ValueError( + f"TRAIN_BATCH_TOKENS={global_tokens} does not divide into full local sequences " + f"for WORLD_SIZE={self.world_size}, GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_seqs = local_tokens // seq_len + active_shards = min(self.shards_per_batch, self.num_shards, local_seqs) + if active_shards <= 0: + raise ValueError(f"No active shards available for local_seqs={local_seqs}") + seqs_per_shard = local_seqs // active_shards + seq_remainder = local_seqs % active_shards + hold_idx = self.batch_idx // self.shard_hold_steps + shard_start = ((hold_idx * self.world_size) + self.rank) * self.batch_shard_stride + chunks: list[Tensor] = [] + for shard_slot in range(active_shards): + count = seqs_per_shard + (1 if shard_slot < seq_remainder else 0) + if count <= 0: + continue + shard_idx = (shard_start + shard_slot * self.batch_shard_stride) % self.num_shards + chunks.append(self._sample_sequences(shard_idx, count)) + self.batch_idx += 1 + local = chunks[0] if len(chunks) == 1 else torch.cat(chunks, dim=0) + local = local.to(dtype=torch.int64) + x = local[:, :-1] + y = local[:, 1:] + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +def build_train_loader(args: Hyperparameters, rank: int, world_size: int, device: torch.device): + if args.loader_mode == "sequential": + return DistributedTokenLoader(args.train_files, rank, world_size, device) + if args.loader_mode == "coprime": + return CoprimeDistributedTokenLoader( + args.train_files, + rank, + world_size, + device, + seq_len=args.train_seq_len, + seed=args.seed, + max_loaded_shards=args.coprime_max_loaded_shards, + shards_per_batch=args.coprime_shards_per_batch, + shard_hold_steps=args.coprime_shard_hold_steps, + ) + raise ValueError(f"Unknown LOADER_MODE={args.loader_mode!r}") + +# --- Transformer modules --- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + # No CastedLinear -- weights come from banks + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 # set by GPT.__init__ for partial RoPE + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False # set by GPT.__init__ for deep layers only + # Gated attention and value residual (non-banked small params) + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + self.value_residual = value_residual + if value_residual: + self.vrl_alpha = nn.Parameter(torch.zeros(1, dtype=torch.float32)) # sigmoid gate (PR #569 style) + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + """Efficient XSA: subtract self-value projection via GQA-aware reshape (no repeat_interleave). + y: [B, T, H, D], v: [B, T, Hkv, D]. H must be divisible by Hkv.""" + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) # [B, T, Hkv, group, D] + vn = F.normalize(v, dim=-1).unsqueeze(-2) # [B, T, Hkv, 1, D] -- broadcast ready + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + bsz, seqlen, dim = x.shape + q = F.linear(x, q_w.to(x.dtype)).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = F.linear(x, k_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = F.linear(x, v_w.to(x.dtype)) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v if self.value_residual else None + if self.value_residual and v0 is not None: + alpha = torch.sigmoid(self.vrl_alpha.to(dtype=v.dtype)) + v = v + alpha * v0 # sigmoid-gated residual (PR #569 style) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if flash_attn_3_func is not None: + q_attn, k_attn, v_attn = q, k, v + if q_attn.dtype not in (torch.float16, torch.bfloat16): + q_attn = q_attn.to(torch.bfloat16) + k_attn = k_attn.to(torch.bfloat16) + v_attn = v_attn.to(torch.bfloat16) + y = flash_attn_3_func(q_attn, k_attn, v_attn, causal=True) + else: + qh = q.transpose(1, 2) + kh = k.transpose(1, 2) + vh = v.transpose(1, 2) + if self.num_heads != self.num_kv_heads: + repeat = self.num_heads // self.num_kv_heads + kh = kh.repeat_interleave(repeat, dim=1) + vh = vh.repeat_interleave(repeat, dim=1) + y = F.scaled_dot_product_attention(qh, kh, vh, is_causal=True).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + # gate shape: (bsz, seqlen, num_heads) -> (bsz, seqlen, num_heads, 1) for B,T,H,D layout + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) + y = y * gate + y = y.reshape(bsz, seqlen, dim) + return F.linear(y, out_w.to(x.dtype)), raw_v + +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int, trigram: bool = False): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self._trigram = trigram + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def trigram_hash(self, tokens: Tensor) -> Tensor: + """Hash (t-2, t-1, t) trigrams into same embedding table. Zero extra params.""" + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., :2] = mod + out[..., 2:] = (36313 * t[..., 2:] ^ 27191 * t[..., 1:-1] ^ 51497 * t[..., :-2]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self._trigram: + h = h + self.embed(self.trigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + # No CastedLinear -- weights come from banks + self.kernel_mode = os.environ.get("MLP_KERNEL_MODE", "").strip().lower() + def forward(self, x: Tensor, up_w: Tensor, down_w: Tensor) -> Tensor: + x = F.linear(x, up_w.to(x.dtype)) + x = leaky_relu_sq(x, kernel_mode=self.kernel_mode) + return F.linear(x, down_w.to(x.dtype)) + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + layer_idx: int = 0, + ln_scale: bool = False, + dtg: bool = False, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=gated_attention, value_residual=value_residual) + self.mlp = MLP(dim, mlp_mult) + attn_scale_init = float(os.environ.get("ATTN_SCALE_INIT", "1.0")) + mlp_scale_init = float(os.environ.get("MLP_SCALE_INIT", "1.0")) + resid_mix_x_init = float(os.environ.get("RESID_MIX_X_INIT", "1.0")) + resid_mix_x0_init = float(os.environ.get("RESID_MIX_X0_INIT", "0.0")) + self.attn_scale = nn.Parameter(torch.full((dim,), attn_scale_init, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.full((dim,), mlp_scale_init, dtype=torch.float32)) + self.resid_mix = nn.Parameter( + torch.stack( + ( + torch.full((dim,), resid_mix_x_init, dtype=torch.float32), + torch.full((dim,), resid_mix_x0_init, dtype=torch.float32), + ) + ) + ) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, up_w: Tensor, down_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out, raw_v = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, q_w, k_w, v_w, out_w, v_embed=v_embed, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor, up_w, down_w) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out, raw_v + +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) # kv_dim for value projection + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.value_residual = value_residual + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim, trigram=bool(int(os.environ.get("TRIGRAM", "0")))) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + # Parameter banks: contiguous 3D tensors for batched optimizer + head_dim = model_dim // num_heads + kv_dim = num_kv_heads * head_dim + mlp_dim = int(mlp_mult * model_dim) + self.num_layers = num_layers + self.qo_bank = nn.Parameter(torch.empty(2 * num_layers, model_dim, model_dim)) + self.kv_bank = nn.Parameter(torch.empty(2 * num_layers, kv_dim, model_dim)) + self.mlp_up_bank = nn.Parameter(torch.empty(num_layers, mlp_dim, model_dim)) + self.mlp_down_bank = nn.Parameter(torch.empty(num_layers, model_dim, mlp_dim)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + layer_idx=i, + ln_scale=ln_scale, + dtg=dtg, + gated_attention=gated_attention, + value_residual=value_residual, + ) + for i in range(num_layers) + ] + ) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim_ve = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim_ve) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() # keep empty for compat + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + n = self.num_layers + proj_scale = 1.0 / math.sqrt(2 * n) + # Init banks: orthogonal, with proj layers scaled down and out/down zero-init + for i in range(n): + nn.init.orthogonal_(self.qo_bank.data[i], gain=1.0) # Q + nn.init.zeros_(self.qo_bank.data[n + i]) # Out (zero init) + nn.init.orthogonal_(self.kv_bank.data[i], gain=1.0) # K + nn.init.orthogonal_(self.kv_bank.data[n + i], gain=1.0) # V + nn.init.orthogonal_(self.mlp_up_bank.data[i], gain=1.0) # MLP up + nn.init.zeros_(self.mlp_down_bank.data[i]) # MLP down (zero init) + # Scale proj layers (out_proj and mlp_down are "proj" layers) + self.qo_bank.data[n + i].mul_(proj_scale) + self.mlp_down_bank.data[i].mul_(proj_scale) + # Init remaining nn.Linear modules (bigram proj, mtp heads, lm_head) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + """Get value embedding for a specific layer using shared table + per-layer scale.""" + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if hasattr(self, '_ngram_tracker') and self._ngram_tracker is not None and self.training: + per_tok_loss = F.cross_entropy(logits.float(), targets, reduction="none") + weights = self._ngram_tracker.get_weights(input_ids, target_ids) + main_loss = (per_tok_loss * weights).mean() + else: + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + def forward_logits(self, input_ids: Tensor) -> Tensor: + """Return logits (bsz, seq_len, vocab) without computing loss.""" + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + +# --- N-gram bulk update and hashed n-gram sliding eval --- + +def _ngram_bulk_update(val_np, start, end, ctx_tables, full_tables, + min_order, max_order, primes, mask): + """Bulk update n-gram tables with a contiguous range of tokens. + All ranks call this with the SAME token range -> identical tables everywhere.""" + t = val_np[start:end].astype(np.uint64) + n = len(t) + for order in range(min_order, max_order + 1): + if n < order: + continue + ctx_width = order - 1 + ctx_hash = np.zeros(n - order + 1, dtype=np.uint64) + for k in range(ctx_width): + ctx_hash ^= t[k:n - order + 1 + k] * primes[k % len(primes)] + ctx_key = (ctx_hash & mask).astype(np.int64) + tgt = t[order - 1:] + full_key = ((ctx_hash ^ (tgt * primes[ctx_width % len(primes)])) & mask).astype(np.int64) + ctx_tables[order] += np.bincount(ctx_key, minlength=len(ctx_tables[order])).astype(np.uint32) + full_tables[order] += np.bincount(full_key, minlength=len(full_tables[order])).astype(np.uint32) + +def eval_val_sliding_hashed_ngram( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + order: int, + alpha: float, + min_count: int, + buckets: int, + max_seconds: float = 0.0, + batch_seqs: int = 128, + eval_seq_len: int | None = None, +) -> tuple[float, float, float]: + """Score-first sliding eval with chunk-based SHARED n-gram tables + cubric. + + Key design: all ranks share identical n-gram tables via bulk chunk updates. + Each chunk's windows are distributed across ranks for scoring, then ALL ranks + update tables with the same contiguous token range. Every rank sees the full + n-gram picture (not 1/world_size like per-segment updates). + + Legal: entire chunk scored before its tokens update the tables. + """ + min_order = max(args.ngram_eval_min_order, 2) + max_order = max(order, min_order) + adaptive = args.ngram_eval_adaptive + alpha_min = args.ngram_eval_alpha_min + alpha_max = args.ngram_eval_alpha_max + ent_center = args.ngram_eval_entropy_center + ent_scale = args.ngram_eval_entropy_scale + + # Parse fixed per-order multipliers (PR #809 style) + _fixed_order_mults = None + if args.ngram_order_mults_str: + _fixed_order_mults = np.array([float(x) for x in args.ngram_order_mults_str.split(",")], dtype=np.float64) + + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + + # Build all windows and total scored tokens + all_window_starts = [ws for ws in range(0, total_tokens, stride) if min(ws + seq_len, total_tokens) - ws >= 1] + total_scored_tokens = 0.0 + for ws in all_window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + total_scored_tokens += float(max(wlen - s, 0)) + + # Group windows into chunks by scored position -- all ranks share this grouping + chunk_tokens = int(os.environ.get("NGRAM_CHUNK_TOKENS", "1048576")) # 1M default + num_chunks = (total_tokens + chunk_tokens - 1) // chunk_tokens + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in all_window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // chunk_tokens, num_chunks - 1) + chunk_windows[ci].append(ws) + + val_np = val_tokens.numpy() + ctx_tables = {n: np.zeros((buckets,), dtype=np.uint32) for n in range(min_order, max_order + 1)} + full_tables = {n: np.zeros((buckets,), dtype=np.uint32) for n in range(min_order, max_order + 1)} + mask = np.uint64(buckets - 1) + primes = np.array( + [np.uint64(36313), np.uint64(27191), np.uint64(51647), np.uint64(81929), + np.uint64(131071), np.uint64(174763), np.uint64(233017)], + dtype=np.uint64, + ) + + loss_sum = 0.0 + token_count = 0.0 + byte_count = 0.0 + + # Cubric 3D: per (order x entropy_bin x count_bin) adaptive alpha scaling + _NUM_ENT_BINS = 3 # low / mid / high entropy + _NUM_CNT_BINS = 3 # low / mid / high count + _ENT_EDGES = np.array([ent_center - 1.0, ent_center + 1.0]) # [2.0, 4.0] for center=3.0 + _CNT_EDGES = np.array([5.0, 50.0]) # low=<5, mid=5-50, high=>50 context count + _TOTAL_CELLS = _NUM_ENT_BINS * _NUM_CNT_BINS # 9 cells per order = 54 total + _cc = getattr(args, 'cubric_cadence', 0); _con = _cc > 0; _cfired = 0 + if _con: + # Warm-start: proven converged values from 4+ runs (orders 2-7) + # All 9 cells per order get the same warm-start, 3D cubric refines from there + _WARM = {2: 0.45, 3: 0.30, 4: 0.45, 5: 1.88, 6: 2.00, 7: 2.00, 8: 2.00, 9: 2.00} + _c_alpha_mult = {n: [_WARM.get(n, 1.0)] * _TOTAL_CELLS for n in range(min_order, max_order + 1)} + _c_hits = {n: [0] * _TOTAL_CELLS for n in range(min_order, max_order + 1)} + _c_beats = {n: [0] * _TOTAL_CELLS for n in range(min_order, max_order + 1)} + + base_model.eval() + compiled_logits = maybe_compile( + base_model.forward_logits, + enabled=args.compile_enabled, + fullgraph=False, + ) + t0 = time.perf_counter() + deadline = (t0 + max_seconds) if max_seconds > 0.0 else None + cutoff_hit = False + + if rank == 0: + print(f"ngram_eval:chunks={num_chunks} chunk_tokens={chunk_tokens} " + f"windows={len(all_window_starts)} shared_tables=True", flush=True) + + with torch.inference_mode(): + for ci in range(num_chunks): + if deadline is not None and time.perf_counter() >= deadline: + cutoff_hit = True + break + + windows = chunk_windows[ci] + if not windows: + continue + + # Distribute this chunk's windows across ranks + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + # --- Phase 1: SCORE this chunk's windows --- + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + logits_f = logits.float() + nll = F.cross_entropy( + logits_f.reshape(-1, logits_f.size(-1)), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + seg_len = wlen - s + if seg_len <= 0: + continue + + seg_nll = nll[i, s:wlen].to(torch.float64).cpu().numpy() + seg_model_p = np.exp(-seg_nll) + + if adaptive: + log_probs = F.log_softmax(logits_f[i, s:wlen], dim=-1) + probs_a = log_probs.exp() + entropy = -(probs_a * log_probs).sum(dim=-1).cpu().numpy() + sig = 1.0 / (1.0 + np.exp(-ent_scale * (entropy - ent_center))) + per_token_alpha = alpha_min + (alpha_max - alpha_min) * sig + # Bin entropy for 2D cubric: 0=low, 1=mid, 2=high + _ent_bins = np.digitize(entropy, _ENT_EDGES).astype(np.int32) + else: + per_token_alpha = np.full(seg_len, alpha) + _ent_bins = np.ones(seg_len, dtype=np.int32) # all mid + + global_j = np.arange(ws + s + 1, ws + wlen + 1, dtype=np.int64) + p_ng = np.zeros(seg_len, dtype=np.float64) + ng_matched = np.zeros(seg_len, dtype=np.bool_) + _ng_ord = np.zeros(seg_len, dtype=np.int32) + _ng_ctx_count = np.zeros(seg_len, dtype=np.float64) + tgt_np = val_np[global_j].astype(np.uint64) + + for n in range(max_order, min_order - 1, -1): + ctx_width = n - 1 + valid = (global_j >= ctx_width) & (~ng_matched) + if not valid.any(): + continue + v_idx = np.nonzero(valid)[0] + jv = global_j[v_idx] + ctx_hash = np.zeros(len(jv), dtype=np.uint64) + for k in range(ctx_width): + tok = val_np[jv - (ctx_width - k)].astype(np.uint64) + ctx_hash ^= tok * primes[k % len(primes)] + ctx_key = (ctx_hash & mask).astype(np.int64) + full_key = ((ctx_hash ^ (tgt_np[v_idx] * primes[ctx_width % len(primes)])) & mask).astype(np.int64) + ctx_counts = ctx_tables[n][ctx_key].astype(np.float64) + full_counts = full_tables[n][full_key].astype(np.float64) + has_data = ctx_counts >= float(min_count) + if has_data.any(): + p = np.minimum(full_counts, ctx_counts) / np.maximum(ctx_counts, 1.0) + p = np.clip(p, 0.0, 1.0) + hit_idx = v_idx[has_data] + p_ng[hit_idx] = p[has_data] + ng_matched[hit_idx] = True + _ng_ord[hit_idx] = n + _ng_ctx_count[hit_idx] = ctx_counts[has_data] + + # Mix where n-gram matched (PR #809 style or cubric 3D fallback) + if ng_matched.any(): + m_idx = np.nonzero(ng_matched)[0] + # Per-order entropy center shift (PR #809) + if adaptive and args.ngram_entropy_shift: + matched_ords = _ng_ord[m_idx].astype(np.float64) + shifted_centers = ent_center - 0.25 * (matched_ords - float(min_order)) + shifted_sig = 1.0 / (1.0 + np.exp(-ent_scale * (entropy[m_idx] - shifted_centers))) + per_token_alpha[m_idx] = alpha_min + (alpha_max - alpha_min) * shifted_sig + if _fixed_order_mults is not None: + # PR #809 fixed order multipliers (replaces cubric) + a = per_token_alpha[m_idx].copy() + mult_indices = _ng_ord[m_idx] - min_order + mult_indices = np.clip(mult_indices, 0, len(_fixed_order_mults) - 1) + a *= _fixed_order_mults[mult_indices] + np.clip(a, 0.0, 0.95, out=a) + elif _con: + a = per_token_alpha[m_idx].copy() + m_ent_bins = _ent_bins[m_idx] + m_cnt_bins = np.digitize(_ng_ctx_count[m_idx], _CNT_EDGES).astype(np.int32) + for n in range(min_order, max_order + 1): + om = _ng_ord[m_idx] == n + if not om.any(): + continue + for eb in range(_NUM_ENT_BINS): + for cb in range(_NUM_CNT_BINS): + cell = eb * _NUM_CNT_BINS + cb + mask_ecb = om & (m_ent_bins == eb) & (m_cnt_bins == cb) + if mask_ecb.any(): + _c_hits[n][cell] += int(mask_ecb.sum()) + _c_beats[n][cell] += int((p_ng[m_idx[mask_ecb]] > seg_model_p[m_idx[mask_ecb]]).sum()) + a[mask_ecb] *= _c_alpha_mult[n][cell] + np.clip(a, 0.0, 0.95, out=a) + else: + a = per_token_alpha[m_idx] + seg_model_p[m_idx] = (1.0 - a) * seg_model_p[m_idx] + a * p_ng[m_idx] + + seg_nll = -np.log(np.clip(seg_model_p, 1e-12, 1.0)) + loss_sum += float(seg_nll.sum()) + token_count += float(seg_len) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += float(tb.sum().item()) + + # --- Phase 2: SHARED UPDATE -- all ranks update with same chunk tokens --- + chunk_start = ci * chunk_tokens + chunk_end = min((ci + 1) * chunk_tokens, total_tokens) + _ngram_bulk_update(val_np, chunk_start, chunk_end + 1, + ctx_tables, full_tables, min_order, max_order, + primes, mask) + + # Cubric 2D c-step: adapt per (order x entropy_bin) + if _con: + # Collect all (order, ent_bin, cnt_bin) cells with enough data + all_rates = [] + for n in range(min_order, max_order + 1): + for cell in range(_TOTAL_CELLS): + if _c_hits[n][cell] >= 8: + all_rates.append(_c_beats[n][cell] / _c_hits[n][cell]) + if len(all_rates) >= 4: + avg_rate = sum(all_rates) / len(all_rates) + for n in range(min_order, max_order + 1): + for cell in range(_TOTAL_CELLS): + if _c_hits[n][cell] >= 8: + rate = _c_beats[n][cell] / _c_hits[n][cell] + if rate > avg_rate + 0.05: + _c_alpha_mult[n][cell] = min(_c_alpha_mult[n][cell] * 1.03, 2.0) + elif rate < avg_rate - 0.05: + _c_alpha_mult[n][cell] = max(_c_alpha_mult[n][cell] * 0.97, 0.3) + _cfired += 1 + if rank == 0 and _cfired % 8 == 0: + parts = [] + for n in range(min_order, max_order + 1): + m = _c_alpha_mult[n] + avg_m = sum(m) / len(m) + parts.append(f"o{n}:avg={avg_m:.2f}") + print(f"cubric3d:step={_cfired} {' '.join(parts)}", flush=True) + _c_hits = {n: [0] * _TOTAL_CELLS for n in range(min_order, max_order + 1)} + _c_beats = {n: [0] * _TOTAL_CELLS for n in range(min_order, max_order + 1)} + + # Progress + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1 or ci < 3): + elapsed = time.perf_counter() - t0 + cur_bpb = (loss_sum / max(token_count, 1.0)) / math.log(2.0) * (token_count / max(byte_count, 1.0)) if token_count > 0 else 0.0 + print( + f"ngram_eval:chunk [{ci+1}/{num_chunks}] bpb={cur_bpb:.6f} t={elapsed:.0f}s", + flush=True, + ) + + # All-reduce across ranks + _loss = torch.tensor(loss_sum, device=device, dtype=torch.float64) + _toks = torch.tensor(token_count, device=device, dtype=torch.float64) + _bytes = torch.tensor(byte_count, device=device, dtype=torch.float64) + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(_loss, op=dist.ReduceOp.SUM) + dist.all_reduce(_toks, op=dist.ReduceOp.SUM) + dist.all_reduce(_bytes, op=dist.ReduceOp.SUM) + loss_sum = _loss.item() + token_count = _toks.item() + byte_count = _bytes.item() + + coverage = token_count / max(total_scored_tokens, 1.0) + if cutoff_hit: + elapsed = time.perf_counter() - t0 + print( + f"ngram_eval:cutoff max_seconds={max_seconds:.1f} " + f"coverage={coverage*100:.2f}% elapsed={elapsed:.0f}s", + flush=True, + ) + + if _con and rank == 0: + print(f"cubric3d:final c_steps={_cfired} cells={_TOTAL_CELLS}x{max_order-min_order+1}={_TOTAL_CELLS*(max_order-min_order+1)}", flush=True) + for n in range(min_order, max_order + 1): + m = _c_alpha_mult[n] + row = " ".join(f"{m[cell]:.2f}" for cell in range(_TOTAL_CELLS)) + print(f" o{n}: [{row}]", flush=True) + val_loss = loss_sum / max(token_count, 1.0) + val_bpb = val_loss / math.log(2.0) * (token_count / max(byte_count, 1.0)) + base_model.train() + return val_loss, val_bpb, coverage + +# --- Sliding window evaluation --- + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = maybe_compile( + base_model.forward_logits, + enabled=args.compile_enabled, + fullgraph=args.compile_fullgraph, + ) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte + + +# --- Training --- + +def main() -> None: + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + # zeropower_via_newtonschulz5 runs eagerly with bmm -- do NOT compile + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size != 8: + raise ValueError( + f"Mikey 8x requires WORLD_SIZE=8, got {world_size}. " + "Launch with: torchrun --standalone --nproc_per_node=8 Mikey/train_gpt_8xgpu.py" + ) + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 100, console=False) + log0("condition_id:mikey_8x_seed444") + log0("run_label:salvage_v2 source_record:mikey_origin_run_20260427 axis:depth_12L+brotli+mixed_int") + log0("changed_fields:num_layers (11->12), compression (zstd->brotli+bshf), quant_policy (uniform_int6->mixed_int5_int6_int8)") + log0("expected_metric:final_sliding_window_exact comparator:0.8672_4k_8x_oversize_run prior_size:17766043_target:<16000000") + log0(f"condition:DATA_PATH={args.data_path}") + log0(f"condition:TOKENIZER_PATH={args.tokenizer_path}") + log0(f"condition:VOCAB_SIZE={args.vocab_size}") + log0(f"condition:SEED={args.seed}") + log0(f"condition:MAX_WALLCLOCK_SECONDS={args.max_wallclock_seconds}") + log0(f"condition:LOADER_MODE={args.loader_mode}") + log0(f"condition:COPRIME_MAX_LOADED_SHARDS={args.coprime_max_loaded_shards}") + log0(f"condition:COPRIME_SHARDS_PER_BATCH={args.coprime_shards_per_batch}") + log0(f"condition:COPRIME_SHARD_HOLD_STEPS={args.coprime_shard_hold_steps}") + log0(f"condition:SKIP_GPTQ={os.environ.get('SKIP_GPTQ', '1')}") + log0(f"condition:TRIGRAM={int(args.trigram_enabled)}") + log0(f"condition:NGRAM_EVAL_ORDER={args.ngram_eval_order}") + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + if args.ngram_eval_order >= 2: + log0(f"ngram_eval:order={args.ngram_eval_order} alpha={args.ngram_eval_alpha} min_count={args.ngram_eval_min_count} buckets={args.ngram_eval_buckets}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + CastedLinear._qat_enabled = args.qat_enabled + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + gated_attention=args.gated_attention, + value_residual=args.value_residual, + ).to(device).bfloat16() + # Banks stay FP32 (like CastedLinear weights), cast to BF16 in forward + base_model.qo_bank.data = base_model.qo_bank.data.float() + base_model.kv_bank.data = base_model.kv_bank.data.float() + base_model.mlp_up_bank.data = base_model.mlp_up_bank.data.float() + base_model.mlp_down_bank.data = base_model.mlp_down_bank.data.float() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + if args.complement_alpha > 0: + tracker = TrainNgramTracker(args.vocab_size, device, complement_alpha=args.complement_alpha) + base_model._ngram_tracker = tracker + log0(f"complementary_training:alpha={args.complement_alpha}") + else: + base_model._ngram_tracker = None + # No DDP -- Parallel Muon handles bank grad communication via reduce-scatter, + # and non-bank grads are manually all-reduced before Adam steps. + compiled_model = maybe_compile( + base_model, + enabled=args.compile_enabled, + fullgraph=args.compile_fullgraph, + mode=args.compile_mode, + ) + model = compiled_model + + # Optimizer split: + # - 4 parameter banks -> Muon (batched Newton-Schulz) + # - token embedding -> Adam + # - scalars/control tensors -> Adam + # - bigram proj, mtp heads, VE proj -> Adam (small matrix params not worth banking) + matrix_params = [ + base_model.qo_bank, base_model.kv_bank, + base_model.mlp_up_bank, base_model.mlp_down_bank, + ] + block_named_params = list(base_model.blocks.named_parameters()) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + scalar_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + scalar_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + # Non-bank params that need manual all-reduce (replicated across GPUs) + replicated_params = list(optimizer_tok.param_groups[0]["params"]) + for pg in optimizer_tok.param_groups[1:]: + replicated_params.extend(pg["params"]) + replicated_params.extend(scalar_params) + + optimizer_head = None + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + replicated_params.append(base_model.lm_head.weight) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if optimizer_head is not None: + optimizers.append(optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + compile_mode = args.compile_mode if args.compile_mode else "default" + log0( + f"compile:enabled={int(args.compile_enabled)} mode:{compile_mode} " + f"fullgraph={int(args.compile_fullgraph)}" + ) + log0(f"mlp_kernel_mode:{args.mlp_kernel_mode or 'eager'}") + log0( + f"scale_init:attn={args.attn_scale_init:.4f} mlp={args.mlp_scale_init:.4f} " + f"resid_mix=({args.resid_mix_x_init:.4f},{args.resid_mix_x0_init:.4f}) " + f"ln_scale={int(args.ln_scale)}" + ) + log0(f"seed:{args.seed}") + train_loader = build_train_loader(args, rank, world_size, device) + log0(train_loader.describe()) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + # GPTQ calibration reads training data — it must complete within the wallclock budget. + # We stop the training loop early (by GPTQ_RESERVE_MS) so GPTQ runs before the cap. + _skip_gptq = int(os.environ.get("SKIP_GPTQ", "1")) + _gptq_reserve_ms = float(os.environ.get("GPTQ_RESERVE_MS", "30000")) if (max_wallclock_ms is not None and not _skip_gptq) else 0.0 + effective_max_wallclock_ms = (max_wallclock_ms - _gptq_reserve_ms) if max_wallclock_ms is not None else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + # All-reduce all grads for warmup (simple, not optimized) + if distributed: + for p in base_model.parameters(): + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + train_loader = build_train_loader(args, rank, world_size, device) + log0(f"loader_reset:{train_loader.describe()}") + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + from collections import deque + lawa_queue: deque[dict[str, Tensor]] = deque(maxlen=args.lawa_k) + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms tok/s:{(step * args.train_batch_tokens) / max(training_time_ms / 1000.0, 1e-9):.0f}" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + if base_model._ngram_tracker is not None: + base_model._ngram_tracker.update(x, y) + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + # === 3-phase overlapped optimizer step === + # Phase 1: Launch async reduce-scatter for banks (biggest first) + optimizer_muon.launch_reduce_scatters() + # Phase 2: All-reduce non-bank grads + step Adam (while bank RS is in-flight) + if distributed: + for p in replicated_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + optimizer_tok.step() + optimizer_scalar.step() + if optimizer_head is not None: + optimizer_head.step() + # Phase 3: Wait for RS, local NS5, all-gather (banks processed last) + optimizer_muon.step() + zero_grad_all() + # EMA update + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + if args.lawa_enabled and step % args.lawa_freq == 0: + lawa_queue.append({name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()}) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms tok/s:{(step * args.train_batch_tokens) / max(approx_training_time_ms / 1000.0, 1e-9):.0f}" + ) + reached_cap = effective_max_wallclock_ms is not None and approx_training_time_ms >= effective_max_wallclock_ms + if distributed and effective_max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + # GPTQ calibration: reads training data — must complete within MAX_WALLCLOCK_SECONDS. + # Training loop stopped GPTQ_RESERVE_MS early so this runs inside the budget. + if _skip_gptq: + log0("gptq:SKIPPED (SKIP_GPTQ=1) — will use naive int6") + gptq_hessians: dict[str, Tensor] = {} + else: + log0("gptq:calibrating with training data...") + t_gptq = time.perf_counter() + gptq_hessians = gptq_calibrate(base_model, args.train_files, device, n_samples=256, seq_len=args.train_seq_len) + log0(f"gptq:calibrated {len(gptq_hessians)} layers in {time.perf_counter()-t_gptq:.1f}s") + # Apply weight averaging + if args.lawa_enabled and len(lawa_queue) > 1: + log0(f"lawa:applying LAWA averaging k={len(lawa_queue)}") + current_state = base_model.state_dict() + avg_state = {name: torch.zeros(t.shape, dtype=torch.float32, device='cpu') for name, t in current_state.items()} + for snap in lawa_queue: + for name in avg_state: + avg_state[name] += snap[name].float() + for name in avg_state: + avg_state[name] /= len(lawa_queue) + avg_state[name] = avg_state[name].to(dtype=current_state[name].dtype) + base_model.load_state_dict(avg_state, strict=True) + else: + log0("ema:applying EMA weights") + current_state = base_model.state_dict() + avg_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(avg_state, strict=True) + if args.post_ema_diagnostic: + torch.cuda.synchronize() + t_diag = time.perf_counter() + diag_val_loss, diag_val_bpb = eval_val( + args, compiled_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"DIAGNOSTIC post_ema val_loss:{diag_val_loss:.4f} val_bpb:{diag_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" + ) + else: + log0("diagnostic_eval:skipped POST_EMA_DIAGNOSTIC=0") + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k) + if excluded_mtp > 0: + log0(f"export_excluding_mtp_params:{excluded_mtp}") + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + # GPTQ quantization using Hessians collected from training data. + # salvage_v2 mixed-int policy: int5 for mlp_up_bank/mlp_down_bank (most quant-tolerant per + # 11-day collate), int6 for qo_bank/kv_bank/embed (attention + token embed kept at int6 for + # quant safety; matches seed for embed). See `mixed_quantize_int6_gptq` docstring. + quant_result, quant_meta = mixed_quantize_int6_gptq( + sd_cpu, + int6_cats={"qo", "kv", "attn_other", "mlp_other", "aux", "embed"}, + hessians=gptq_hessians, + int5_cats={"mlp_down", "mlp_up"}, + ) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = _compress_blob(quant_raw) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + log0(f"Serialized model mixed_int5_int6_int8+{_COMPRESSOR}: {quant_file_bytes} bytes") + log0(f"Total submission size mixed_int5_int6_int8+{_COMPRESSOR}: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(_decompress_blob(quant_blob_disk)), + map_location="cpu", + ) + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, rope_dims=args.rope_dims, ln_scale=args.ln_scale, + dtg=args.dtg_enabled, ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + gated_attention=args.gated_attention, value_residual=args.value_residual, + ).to(device).bfloat16() + eval_model.qo_bank.data = eval_model.qo_bank.data.float() + eval_model.kv_bank.data = eval_model.kv_bank.data.float() + eval_model.mlp_up_bank.data = eval_model.mlp_up_bank.data.float() + eval_model.mlp_down_bank.data = eval_model.mlp_down_bank.data.float() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, eval_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + del eval_model, deq_state, quant_state, sd_cpu + torch.cuda.empty_cache() + sw_seq_len = effective_eval_seq_len + if args.skip_final_eval: + log0("final_eval:skipped sliding/ngram by SKIP_FINAL_EVAL=1") + else: + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, base_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, base_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + if args.ngram_eval_order >= 2: + if distributed: + dist.barrier() + torch.cuda.synchronize() + t_ng = time.perf_counter() + ng_loss, ng_bpb, ng_coverage = eval_val_sliding_hashed_ngram( + args, + base_model, + rank, + world_size, + device, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + stride=args.eval_stride, + order=args.ngram_eval_order, + alpha=args.ngram_eval_alpha, + min_count=args.ngram_eval_min_count, + buckets=args.ngram_eval_buckets, + max_seconds=args.ngram_eval_max_seconds, + eval_seq_len=sw_seq_len, + ) + if rank == 0: + torch.cuda.synchronize() + ng_eval_ms = 1000.0 * (time.perf_counter() - t_ng) + if ng_coverage >= 0.999999: + log0( + f"final_sliding_window_ngram{args.ngram_eval_order} val_loss:{ng_loss:.4f} " + f"val_bpb:{ng_bpb:.4f} eval_time:{ng_eval_ms:.0f}ms" + ) + log0( + f"final_sliding_window_ngram{args.ngram_eval_order}_exact " + f"val_loss:{ng_loss:.8f} val_bpb:{ng_bpb:.8f}" + ) + else: + log0( + f"final_sliding_window_ngram{args.ngram_eval_order}_partial val_loss:{ng_loss:.4f} " + f"val_bpb:{ng_bpb:.4f} coverage:{ng_coverage:.4f} eval_time:{ng_eval_ms:.0f}ms" + ) + log0( + f"final_sliding_window_ngram{args.ngram_eval_order}_partial_exact " + f"val_loss:{ng_loss:.8f} val_bpb:{ng_bpb:.8f} coverage:{ng_coverage:.8f}" + ) + if distributed: + dist.barrier() + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() diff --git a/records/track_10min_16mb/2026-04-27_Mikey/train_seed300.log b/records/track_10min_16mb/2026-04-27_Mikey/train_seed300.log new file mode 100644 index 0000000000..0097b6c4cc --- /dev/null +++ b/records/track_10min_16mb/2026-04-27_Mikey/train_seed300.log @@ -0,0 +1,2801 @@ +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +from collections import OrderedDict +from pathlib import Path +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP + +os.environ.setdefault("RUN_ID", "rascal_4k_12L_brotli_mixed_8x_seed444_" + time.strftime("%Y%m%d_%H%M%S")) + +try: + import triton + import triton.language as tl +except ImportError: + triton = None + tl = None +try: + from flash_attn_interface import flash_attn_func as flash_attn_3_func +except ImportError: + flash_attn_3_func = None +# Compression: brotli-11 + byte-shuffle is the salvage_v2 preferred path (PR #1493 recipe). +# Falls back to zstd then zlib so this file still runs if brotli isn't installed. +_brotli_module = None +_zstandard_module = None +_zlib_module = None +try: + import brotli as _brotli_module + _COMPRESSOR = "brotli" +except ImportError: + try: + import zstandard as _zstandard_module + _COMPRESSOR = "zstd" + import warnings + warnings.warn("brotli not found — falling back to zstd (~1MB+ larger). pip install brotli") + except ImportError: + import zlib as _zlib_module + import warnings + _COMPRESSOR = "zlib" + warnings.warn("brotli/zstandard not found — falling back to zlib. Artifact will be much larger! pip install brotli") +# Backwards-compat shims so any remaining `zstandard.*` or `_zlib_module.*` references still work. +if _zstandard_module is not None: + zstandard = _zstandard_module +if _zlib_module is None: + import zlib as _zlib_module # always available; used by zlib fallback path +# --- Byte-shuffle (de-interleave) wrapper from PR #1493: improves brotli ratio on quantized payloads. --- +_BSHF_MAGIC = b"BSHF" +def _byte_shuffle(data: bytes, stride: int = 2) -> bytes: + if stride <= 1 or len(data) < stride: + return data + src = np.frombuffer(data, dtype=np.uint8) + n = len(src) + out = np.empty(n, dtype=np.uint8) + dest_off = 0 + for pos in range(stride): + chunk = src[pos::stride] + out[dest_off:dest_off + len(chunk)] = chunk + dest_off += len(chunk) + return _BSHF_MAGIC + bytes([stride]) + out.tobytes() +def _byte_unshuffle(data: bytes) -> bytes: + if len(data) < 5 or data[:4] != _BSHF_MAGIC: + return data + stride = data[4] + if stride < 2: + return data[5:] + payload = np.frombuffer(data, dtype=np.uint8, offset=5) + n = len(payload) + out = np.empty(n, dtype=np.uint8) + src_off = 0 + for pos in range(stride): + chunk_len = n // stride + (1 if pos < n % stride else 0) + out[pos::stride][:chunk_len] = payload[src_off:src_off + chunk_len] + src_off += chunk_len + return out.tobytes() +def _compress_blob(raw: bytes) -> bytes: + if _COMPRESSOR == "brotli": + return _brotli_module.compress(_byte_shuffle(raw, stride=2), quality=11) + elif _COMPRESSOR == "zstd": + return _zstandard_module.ZstdCompressor(level=22).compress(raw) + else: + return _zlib_module.compress(raw, 9) +def _decompress_blob(blob: bytes) -> bytes: + if _COMPRESSOR == "brotli": + return _byte_unshuffle(_brotli_module.decompress(blob)) + elif _COMPRESSOR == "zstd": + return _zstandard_module.ZstdDecompressor().decompress(blob) + else: + return _zlib_module.decompress(blob) + +if os.environ.get("TORCHDYNAMO_SUPPRESS_ERRORS", "0") == "1": + import torch._dynamo + torch._dynamo.config.suppress_errors = True +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp4096") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_4096_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 444)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 4096)) + num_layers = int(os.environ.get("NUM_LAYERS", 12)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + lawa_enabled = bool(int(os.environ.get("LAWA_ENABLED", "0"))) + lawa_k = int(os.environ.get("LAWA_K", 10)) + lawa_freq = int(os.environ.get("LAWA_FREQ", 100)) + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + trigram_enabled = bool(int(os.environ.get("TRIGRAM", "0"))) # TrigramHash (off by default, risky) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 11)) # XSA on ALL layers (our novel contribution) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + gated_attention = bool(int(os.environ.get("GATED_ATTENTION", "0"))) + value_residual = bool(int(os.environ.get("VALUE_RESIDUAL", "0"))) # VRL with sigmoid gates (off by default, risky) + attn_scale_init = float(os.environ.get("ATTN_SCALE_INIT", 1.0)) + mlp_scale_init = float(os.environ.get("MLP_SCALE_INIT", 1.0)) + resid_mix_x_init = float(os.environ.get("RESID_MIX_X_INIT", 1.0)) + resid_mix_x0_init = float(os.environ.get("RESID_MIX_X0_INIT", 0.0)) + complement_alpha = float(os.environ.get("COMPLEMENT_ALPHA", "0")) + ngram_eval_order = int(os.environ.get("NGRAM_EVAL_ORDER", 0)) + ngram_eval_min_order = int(os.environ.get("NGRAM_EVAL_MIN_ORDER", 2)) + ngram_eval_alpha = float(os.environ.get("NGRAM_EVAL_ALPHA", 0.30)) + ngram_eval_adaptive = bool(int(os.environ.get("NGRAM_EVAL_ADAPTIVE", "1"))) + ngram_eval_alpha_min = float(os.environ.get("NGRAM_EVAL_ALPHA_MIN", 0.05)) + ngram_eval_alpha_max = float(os.environ.get("NGRAM_EVAL_ALPHA_MAX", 0.60)) + ngram_eval_entropy_center = float(os.environ.get("NGRAM_EVAL_ENTROPY_CENTER", 4.0)) + ngram_eval_entropy_scale = float(os.environ.get("NGRAM_EVAL_ENTROPY_SCALE", 2.0)) + ngram_eval_min_count = int(os.environ.get("NGRAM_EVAL_MIN_COUNT", 2)) + ngram_eval_buckets = int(os.environ.get("NGRAM_EVAL_BUCKETS", 4_194_304)) + ngram_eval_max_seconds = float(os.environ.get("NGRAM_EVAL_MAX_SECONDS", 0.0)) + ngram_entropy_shift = bool(int(os.environ.get("NGRAM_ENTROPY_SHIFT", "0"))) + ngram_order_mults_str = os.environ.get("NGRAM_ORDER_MULTS", "") + cubric_cadence = int(os.environ.get("CUBRIC_CADENCE", 0)) + skip_final_eval = bool(int(os.environ.get("SKIP_FINAL_EVAL", "0"))) + post_ema_diagnostic = bool(int(os.environ.get("POST_EMA_DIAGNOSTIC", "1"))) + compile_enabled = bool(int(os.environ.get("COMPILE_ENABLED", "1"))) + compile_mode = os.environ.get("COMPILE_MODE", "").strip() + compile_fullgraph = bool(int(os.environ.get("COMPILE_FULLGRAPH", "1"))) + mlp_kernel_mode = os.environ.get("MLP_KERNEL_MODE", "").strip().lower() + loader_mode = os.environ.get("LOADER_MODE", "coprime").strip().lower() + coprime_max_loaded_shards = int(os.environ.get("COPRIME_MAX_LOADED_SHARDS", 143)) + coprime_shards_per_batch = int(os.environ.get("COPRIME_SHARDS_PER_BATCH", 1)) + coprime_shard_hold_steps = int(os.environ.get("COPRIME_SHARD_HOLD_STEPS", 64)) + + +def maybe_compile(fn_or_module, *, enabled: bool, fullgraph: bool, mode: str = ""): + if not enabled: + return fn_or_module + kwargs = dict(dynamic=False, fullgraph=fullgraph) + if mode: + kwargs["mode"] = mode + return torch.compile(fn_or_module, **kwargs) + + +if triton is not None: + @triton.jit + def _leaky_relu_sq_forward_kernel(x_ptr, y_ptr, n_elements, BLOCK_SIZE: tl.constexpr): + pid = tl.program_id(0) + offsets = pid * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE) + mask = offsets < n_elements + x = tl.load(x_ptr + offsets, mask=mask, other=0.0).to(tl.float32) + a = tl.where(x >= 0, x, 0.5 * x) + y = a * a + tl.store(y_ptr + offsets, y, mask=mask) + + @triton.jit + def _leaky_relu_sq_backward_kernel(x_ptr, grad_out_ptr, grad_in_ptr, n_elements, BLOCK_SIZE: tl.constexpr): + pid = tl.program_id(0) + offsets = pid * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE) + mask = offsets < n_elements + x = tl.load(x_ptr + offsets, mask=mask, other=0.0).to(tl.float32) + grad_out = tl.load(grad_out_ptr + offsets, mask=mask, other=0.0).to(tl.float32) + a = tl.where(x >= 0, x, 0.5 * x) + slope = tl.where(x >= 0, 1.0, 0.5) + grad_in = grad_out * (2.0 * a * slope) + tl.store(grad_in_ptr + offsets, grad_in, mask=mask) + + +class TritonLeakyReluSqFn(torch.autograd.Function): + @staticmethod + def forward(ctx, x: Tensor) -> Tensor: + if triton is None or not x.is_cuda: + a = F.leaky_relu(x, negative_slope=0.5) + ctx.save_for_backward(x) + return a.square() + x_contig = x.contiguous() + y = torch.empty_like(x_contig) + n_elements = x_contig.numel() + grid = lambda meta: (triton.cdiv(n_elements, meta["BLOCK_SIZE"]),) + _leaky_relu_sq_forward_kernel[grid](x_contig, y, n_elements, BLOCK_SIZE=1024) + ctx.save_for_backward(x_contig) + return y + + @staticmethod + def backward(ctx, grad_out: Tensor) -> tuple[Tensor]: + (x,) = ctx.saved_tensors + if triton is None or not grad_out.is_cuda: + a = F.leaky_relu(x, negative_slope=0.5) + slope = torch.where(x >= 0, torch.ones_like(x), torch.full_like(x, 0.5)) + return (grad_out * (2.0 * a * slope),) + grad_out_contig = grad_out.contiguous() + grad_in = torch.empty_like(grad_out_contig) + n_elements = grad_out_contig.numel() + grid = lambda meta: (triton.cdiv(n_elements, meta["BLOCK_SIZE"]),) + _leaky_relu_sq_backward_kernel[grid](x, grad_out_contig, grad_in, n_elements, BLOCK_SIZE=1024) + return (grad_in,) + + +def leaky_relu_sq(x: Tensor, kernel_mode: str = "") -> Tensor: + if kernel_mode == "triton_act": + return TritonLeakyReluSqFn.apply(x) + a = F.leaky_relu(x, negative_slope=0.5) + return a.square() + +class TrainNgramTracker: + """Complementary training: track bigram stats, downweight tokens n-grams can predict.""" + def __init__(self, vocab_size: int, device: torch.device, complement_alpha: float = 0.5): + self.V = vocab_size + self.alpha = complement_alpha + self.bi_counts = torch.zeros(vocab_size, vocab_size, device=device, dtype=torch.float32) + self.bi_totals = torch.zeros(vocab_size, device=device, dtype=torch.float32) + @torch.no_grad() + def update(self, x: Tensor, y: Tensor): + xf = x.reshape(-1) + yf = y.reshape(-1) + ones = torch.ones(xf.numel(), device=xf.device, dtype=torch.float32) + self.bi_counts.reshape(-1).scatter_add_(0, xf * self.V + yf, ones) + self.bi_totals.scatter_add_(0, xf, ones) + def get_weights(self, x: Tensor, y: Tensor) -> Tensor: + xf = x.reshape(-1) + yf = y.reshape(-1) + total = self.bi_totals[xf] + count = self.bi_counts.reshape(-1)[xf * self.V + yf] + ngram_prob = count / (total + 1) + return (1.0 - self.alpha * ngram_prob).clamp(min=0.1) + +# --- Batched Newton-Schulz orthogonalization --- + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 5, eps: float = 1e-7) -> Tensor: + """Batched Newton-Schulz orthogonalization. G: (B,M,N) or (M,N).""" + a, b, c = (3.4445, -4.7750, 2.0315) + was_2d = G.ndim == 2 + if was_2d: + G = G.unsqueeze(0) + X = G.bfloat16() + transposed = X.size(-2) > X.size(-1) + if transposed: + X = X.mT + X = X / (X.norm(dim=(-2, -1), keepdim=True) + eps) + for _ in range(steps): + A = X @ X.mT + B = b * A + c * (A @ A) + X = a * X + B @ X + if transposed: + X = X.mT + if was_2d: + X = X.squeeze(0) + return X + +# --- Parallel Muon optimizer --- + +class Muon(torch.optim.Optimizer): + """Parallel Muon: post-backward reduce-scatter -> local NS5 -> all-gather. + + No DDP for bank params. After backward, this optimizer: + 1. Launches async reduce-scatter for all banks (biggest first) + 2. Returns control so Adam can step on small params while RS is in-flight + 3. Waits for each RS, runs local NS5 on the shard, launches async all-gather + 4. Each all-gather overlaps with next bank's NS5 + """ + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + self._built = False + + def _build(self): + self._distributed = dist.is_available() and dist.is_initialized() + self._world_size = dist.get_world_size() if self._distributed else 1 + self._rank = dist.get_rank() if self._distributed else 0 + ws = self._world_size + + self._bank_meta = [] + for group in self.param_groups: + for p in group["params"]: + B = p.shape[0] + padded_B = ((B + ws - 1) // ws) * ws + shard_B = padded_B // ws + tail = p.shape[1:] + dev = p.device + self._bank_meta.append({ + 'p': p, + 'B': B, + 'padded_grad': torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + 'shard': torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + 'shard_mom': torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + 'full_update': torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + 'scale': max(1, p.shape[-2] / p.shape[-1]) ** 0.5, + }) + # Sort by size descending -- launch biggest reduce-scatters first + self._bank_meta.sort(key=lambda m: -m['p'].numel()) + self._built = True + + def launch_reduce_scatters(self): + """Phase 1: launch async reduce-scatter for all banks. Call right after backward.""" + if not self._built: + self._build() + if not self._distributed: + return + self._rs_futures = [] + for m in self._bank_meta: + p = m['p'] + if p.grad is None: + self._rs_futures.append(None) + continue + pg = m['padded_grad'] + pg[:m['B']].copy_(p.grad.bfloat16()) + if pg.shape[0] > m['B']: + pg[m['B']:].zero_() + fut = dist.reduce_scatter_tensor(m['shard'], pg, op=dist.ReduceOp.AVG, async_op=True) + self._rs_futures.append(fut) + + @torch.no_grad() + def step(self, closure=None): + """Phase 3: wait for RS, local NS5, all-gather. Call AFTER Adam steps.""" + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + if not self._built: + self._build() + + for group in self.param_groups: + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + wd = group.get("weight_decay", 0.0) + + prev_ag_handle = None + prev_m = None + + sharded = self._distributed and hasattr(self, '_rs_futures') + + for i, m in enumerate(self._bank_meta): + p = m['p'] + if p.grad is None: + continue + + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m['p'] + upd = prev_m['full_update'][:prev_m['B']] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m['scale']) + + if sharded and self._rs_futures[i] is not None: + self._rs_futures[i].wait() + g = m['shard'] + buf = m['shard_mom'] + else: + g = p.grad.bfloat16() + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + + buf.mul_(momentum).add_(g) + if nesterov: + update = g.add(buf, alpha=momentum) + else: + update = buf + + update = zeropower_via_newtonschulz5(update, steps=backend_steps) + + if sharded: + prev_ag_handle = dist.all_gather_into_tensor( + m['full_update'], update, async_op=True) + prev_m = m + else: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + p.add_(update.to(dtype=p.dtype), alpha=-lr * m['scale']) + + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m['p'] + upd = prev_m['full_update'][:prev_m['B']] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m['scale']) + + if hasattr(self, '_rs_futures'): + del self._rs_futures + + return loss + +# --- Tokenizer evaluation helpers --- + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) + +# --- Quantization helpers --- + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale,attn_gate,vr_lambda", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if "f1_corr_in" in name or "f1_corr_out" in name: + return "aux" + if "qo_bank" in name or "kv_bank" in name: + return "attn" + if "mlp_up_bank" in name or "mlp_down_bank" in name: + return "mlp" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +# GPTQ: Hessian-aware quantization with column-wise error compensation +def _find_best_row_scales(W: Tensor, clip_range: int = 31) -> Tensor: + t32 = W.float() + best_s = t32.abs().amax(dim=1) / clip_range + best_s = best_s.clamp_min(1.0 / clip_range) + best_err = torch.full((t32.shape[0],), float('inf')) + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range) + q = torch.clamp(torch.round(t32 / s[:, None]), -clip_range, clip_range) + recon = q * s[:, None] + err = (t32 - recon).pow(2).mean(dim=1) + improved = err < best_err + best_s[improved] = s[improved] + best_err[improved] = err[improved] + return best_s +def gptq_quantize_weight(W: Tensor, H: Tensor, clip_range: int = 31, + block_size: int = 64, percdamp: float = 0.002) -> tuple[Tensor, Tensor]: + """GPTQ: quantize weight matrix W using Hessian H = X^T X for error compensation. + Returns (quantized_int8, scale_fp16) in int6 range [-clip_range, clip_range].""" + W = W.float().clone() + rows, cols = W.shape + row_scale = _find_best_row_scales(W, clip_range) + H = H.float().clone() + damp = percdamp * H.diag().mean() + H.diagonal().add_(damp) + perm = torch.argsort(H.diag()) + invperm = torch.argsort(perm) + W = W[:, perm] + H = H[perm][:, perm] + try: + L = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(L) + except torch._C._LinAlgError: + Hinv = torch.diag(1.0 / H.diag().clamp_min(1e-6)) + Q = torch.zeros(rows, cols, dtype=torch.int8) + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + W_block = W[:, i1:i2].clone() + Hinv_block = Hinv[i1:i2, i1:i2] + Err = torch.zeros_like(W_block) + for j in range(i2 - i1): + w_col = W_block[:, j] + h_inv_jj = Hinv_block[j, j].clamp_min(1e-8) + q_col = torch.clamp(torch.round(w_col / row_scale), -clip_range, clip_range) + deq_col = q_col * row_scale + Q[:, i1 + j] = q_col.to(torch.int8) + err = (w_col - deq_col) / h_inv_jj + Err[:, j] = err + if j + 1 < i2 - i1: + W_block[:, j + 1:] -= err.unsqueeze(1) * Hinv_block[j, j + 1:].unsqueeze(0) + if i2 < cols: + W[:, i2:] -= Err @ Hinv[i1:i2, i2:] + Q = Q[:, invperm] + return Q, row_scale.to(torch.float16) +def gptq_calibrate(model: nn.Module, train_pattern: str, device: torch.device, + n_samples: int = 256, seq_len: int = 2048) -> dict[str, Tensor]: + """Collect Hessian H = X^T X for each linear layer using training data.""" + hessians: dict[str, Tensor] = {} + n_seen: dict[str, int] = {} + hooks = [] + def make_hook(name: str): + def hook_fn(module, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + if name not in hessians: + hessians[name] = torch.zeros(x.shape[1], x.shape[1], device=x.device, dtype=torch.float32) + n_seen[name] = 0 + hessians[name].addmm_(x.t(), x) + n_seen[name] += x.shape[0] + return hook_fn + for name, module in model.named_modules(): + if isinstance(module, (nn.Linear, CastedLinear)): + hooks.append(module.register_forward_hook(make_hook(name))) + stream = TokenStream(train_pattern) + model.eval() + with torch.no_grad(): + for _ in range(n_samples): + tokens = stream.take(seq_len + 1).to(device=device, dtype=torch.int64) + x = tokens[:-1].unsqueeze(0) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + model.forward_logits(x) + for h in hooks: + h.remove() + for name in hessians: + hessians[name] /= max(n_seen[name], 1) + model.train() + return hessians +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale +def quantize_int5_per_row(t: Tensor, clip_range: int = 15) -> tuple[Tensor, Tensor]: + """int5 (signed, range [-15, 15]) per-row quant. Modeled on quantize_int6_per_row. + Returned int8 tensor holds values in [-15, 15] — leaves the high 3 bits as zeros, which + brotli compresses very efficiently. Dequant path is identical (q.float() * scale).""" + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale +def _classify_param_fine(name: str) -> str: + """Finer classifier than _classify_param, splitting attention/MLP banks for mixed-int policy. + Returns one of: embed, qo (Q+O bank), kv (K+V bank), mlp_up (mlp_fc), mlp_down (mlp_proj), + aux, attn_other, mlp_other, other. Categories `qo`/`kv` map to attention; `mlp_up`/`mlp_down` + map to MLP. The split lets us route mlp banks to int5 while keeping attn at int6.""" + if "tok_emb" in name or "lm_head" in name: + return "embed" + if "f1_corr_in" in name or "f1_corr_out" in name: + return "aux" + if "qo_bank" in name: + return "qo" + if "kv_bank" in name: + return "kv" + if "mlp_up_bank" in name: + return "mlp_up" + if "mlp_down_bank" in name: + return "mlp_down" + if ".mlp." in name: + return "mlp_other" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn_other" + return "other" +# salvage_v2 mixed-int policy (applied at the Tensor level via _classify_param_fine): +# - mlp_down_bank (mlp_proj, MOST quant-tolerant per 11-day collate) -> int5 (clip_range 15) +# - mlp_up_bank (mlp_fc, also tolerant) -> int5 +# - qo_bank, kv_bank (attention; LEAST quant-tolerant) -> int6 (clip_range 31) +# - tok_emb / lm_head (embed) -> int6 (matches seed; keeping +# attn/embed at int6 for +# quant safety on this +# first salvage attempt) +# Bytes savings: int5 keeps the int8 storage container (no bit-packing) but the high 3 bits are +# forced zero, giving brotli a compressible pattern. Combined with the byte-shuffle wrapper, +# expected savings vs uniform-int6+zstd is roughly the int5 bit ratio (5/6 = -17%) APPLIED only +# to MLP banks (which are ~60-65% of the model parameter mass for a 10L Rascal at mlp_mult=3.0). +# Net est: ~0.62 * 0.17 = ~10-11% blob shrink from int5 alone, plus brotli ~5-8% over zstd. +DEFAULT_INT5_CATS = {"mlp_down", "mlp_up"} +DEFAULT_INT6_CATS = {"qo", "kv", "attn_other", "mlp_other", "aux"} +def mixed_quantize_int6_gptq(state_dict: dict[str, Tensor], int6_cats: set[str], + hessians: dict[str, Tensor], + int5_cats: set[str] | None = None) -> tuple[dict, dict]: + """Mixed-int (int5/int6/int8) quant with GPTQ for matrix categories when Hessian available. + `int6_cats` and `int5_cats` use FINE-grained category names from `_classify_param_fine` + (qo, kv, mlp_up, mlp_down, attn_other, mlp_other, aux, embed). For backwards-compat with + the old uniform-int6 caller, the legacy coarse names {'mlp','attn','aux','embed'} are also + accepted in `int6_cats` and expand to their fine-grained children. `int5_cats` always uses + fine names. If `int5_cats` is None, defaults to DEFAULT_INT5_CATS (the salvage_v2 policy).""" + if int5_cats is None: + int5_cats = set(DEFAULT_INT5_CATS) + # Expand legacy coarse names so the existing call signature keeps working. + _LEGACY = { + "mlp": {"mlp_up", "mlp_down", "mlp_other"}, + "attn": {"qo", "kv", "attn_other"}, + "aux": {"aux"}, + "embed": {"embed"}, + } + expanded_int6: set[str] = set() + for c in int6_cats: + expanded_int6.update(_LEGACY.get(c, {c})) + # int5 categories take precedence over int6 for the same fine name. + expanded_int6 -= int5_cats + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + gptq_count, naive_count, int5_count = 0, 0, 0 + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param_fine(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int5_cats and t.ndim == 2: + module_name = name.rsplit(".weight", 1)[0] if name.endswith(".weight") else name + H = hessians.get(module_name) + if H is not None and H.shape[0] == t.shape[1]: + q, s = gptq_quantize_weight(t, H.cpu(), clip_range=15) + gptq_count += 1 + else: + q, s = quantize_int5_per_row(t) + naive_count += 1 + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int5"} + int5_count += 1 + elif cat in int5_cats and t.ndim >= 1: + t_2d = t.reshape(-1, t.shape[-1]) if t.ndim > 2 else t + q, s = quantize_int5_per_row(t_2d) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int5"} + naive_count += 1 + int5_count += 1 + elif cat in expanded_int6 and t.ndim == 2: + module_name = name.rsplit(".weight", 1)[0] if name.endswith(".weight") else name + H = hessians.get(module_name) + if H is not None and H.shape[0] == t.shape[1]: + q, s = gptq_quantize_weight(t, H.cpu()) + gptq_count += 1 + else: + q, s = quantize_int6_per_row(t) + naive_count += 1 + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + elif cat in expanded_int6 and t.ndim >= 1: + t_2d = t.reshape(-1, t.shape[-1]) if t.ndim > 2 else t + q, s = quantize_int6_per_row(t_2d) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + naive_count += 1 + else: + t_q = t.reshape(-1, t.shape[-1]) if t.ndim > 2 else t + q, s = quantize_float_tensor(t_q) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + print(f"gptq_quantize: {gptq_count} GPTQ layers, {naive_count} naive layers, {int5_count} int5 layers", flush=True) + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + val = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + val = (q.float() * float(s.item())).to(orig_dtype) + out[name] = val.reshape(orig.shape) if val.shape != orig.shape else val + return out + +# --- Data loading --- + +SHARD_HEADER_DTYPE = np.dtype(" dict[str, int]: + header = np.fromfile(file, dtype=SHARD_HEADER_DTYPE, count=SHARD_HEADER_WORDS) + if header.size != SHARD_HEADER_WORDS or int(header[0]) != SHARD_MAGIC or int(header[1]) != SHARD_VERSION: + raise ValueError(f"Unexpected shard header for {file}") + return {"num_tokens": int(header[2])} + +def load_data_shard(file: Path) -> Tensor: + header = read_data_shard_header(file) + num_tokens = header["num_tokens"] + expected_size = SHARD_HEADER_BYTES + num_tokens * SHARD_TOKEN_DTYPE.itemsize + if file.stat().st_size != expected_size: + raise ValueError(f"Shard size mismatch for {file}: expected {expected_size} bytes") + tokens_np = np.fromfile(file, dtype=SHARD_TOKEN_DTYPE, count=num_tokens, offset=SHARD_HEADER_BYTES) + if tokens_np.size != num_tokens: + raise ValueError(f"Short read for {file}") + return torch.from_numpy(tokens_np.astype(np.uint16, copy=False)) + +def choose_coprime_stride(modulus: int, salt: int) -> int: + if modulus <= 1: + return 1 + candidate = abs(salt) % modulus + if candidate == 0: + candidate = 1 + while math.gcd(candidate, modulus) != 1: + candidate += 1 + if candidate >= modulus: + candidate = 1 + return candidate + +class TokenStream: + def __init__(self, pattern: str): + self.files = [Path(p) for p in sorted(glob.glob(pattern))] + if not self.files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + self.file_idx = 0 + self.tokens = load_data_shard(self.files[0]) + self.pos = 0 + def _advance_file(self) -> None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def describe(self) -> str: + return f"loader:sequential shards:{len(self.stream.files)}" + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +class CoprimeDistributedTokenLoader: + """Shard-aware block sampler with deterministic coprime walks.""" + def __init__( + self, + pattern: str, + rank: int, + world_size: int, + device: torch.device, + seq_len: int, + seed: int, + max_loaded_shards: int, + shards_per_batch: int, + shard_hold_steps: int, + ): + self.rank = rank + self.world_size = world_size + self.device = device + self.seq_len = seq_len + self.seed = seed + self.token_offsets = torch.arange(seq_len + 1, dtype=torch.int64) + self.cache: OrderedDict[Path, Tensor] = OrderedDict() + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + self.shards: list[dict[str, int | Path]] = [] + for shard_idx, file in enumerate(files): + header = read_data_shard_header(file) + num_blocks = (header["num_tokens"] - 1) // seq_len + if num_blocks <= 0: + continue + self.shards.append( + { + "file": file, + "num_blocks": num_blocks, + "offset": (seed * 131 + shard_idx * 17) % num_blocks, + "stride": choose_coprime_stride(num_blocks, seed * 29 + shard_idx * 7 + 1), + } + ) + if not self.shards: + raise ValueError(f"No usable shards found for seq_len={seq_len}") + self.num_shards = len(self.shards) + self.max_loaded_shards = max(1, min(max_loaded_shards, self.num_shards)) + self.shards_per_batch = max(1, min(shards_per_batch, self.num_shards)) + self.shard_hold_steps = max(1, shard_hold_steps) + self.batch_shard_stride = choose_coprime_stride(self.num_shards, seed * 41 + 3) + self.batch_idx = 0 + self.shard_visits = [0 for _ in range(self.num_shards)] + def _get_tokens(self, file: Path) -> Tensor: + cached = self.cache.get(file) + if cached is not None: + self.cache.move_to_end(file) + return cached + # CPU advanced indexing is not implemented for uint16, so cache coprime-loader + # shards in int32 and cast to int64 only after batch assembly. + tokens = load_data_shard(file).to(dtype=torch.int32) + if len(self.cache) >= self.max_loaded_shards: + self.cache.popitem(last=False) + self.cache[file] = tokens + return tokens + def _sample_sequences(self, shard_idx: int, count: int) -> Tensor: + shard = self.shards[shard_idx] + num_blocks = int(shard["num_blocks"]) + offset = int(shard["offset"]) + stride = int(shard["stride"]) + visits = self.shard_visits[shard_idx] + block_ids = ( + offset + + (visits + torch.arange(count, dtype=torch.int64)) * stride + ) % num_blocks + self.shard_visits[shard_idx] += count + token_starts = block_ids * self.seq_len + gather_idx = token_starts.unsqueeze(1) + self.token_offsets.unsqueeze(0) + tokens = self._get_tokens(shard["file"]) + return tokens[gather_idx] + def describe(self) -> str: + total_blocks = sum(int(shard["num_blocks"]) for shard in self.shards) + return ( + f"loader:coprime shards:{self.num_shards} blocks:{total_blocks} " + f"seq_len:{self.seq_len} shards_per_batch:{self.shards_per_batch} " + f"cache:{self.max_loaded_shards} batch_stride:{self.batch_shard_stride} " + f"hold_steps:{self.shard_hold_steps}" + ) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + if seq_len != self.seq_len: + raise ValueError(f"Coprime loader was built for seq_len={self.seq_len}, got {seq_len}") + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + if local_tokens % seq_len != 0: + raise ValueError( + f"TRAIN_BATCH_TOKENS={global_tokens} does not divide into full local sequences " + f"for WORLD_SIZE={self.world_size}, GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_seqs = local_tokens // seq_len + active_shards = min(self.shards_per_batch, self.num_shards, local_seqs) + if active_shards <= 0: + raise ValueError(f"No active shards available for local_seqs={local_seqs}") + seqs_per_shard = local_seqs // active_shards + seq_remainder = local_seqs % active_shards + hold_idx = self.batch_idx // self.shard_hold_steps + shard_start = ((hold_idx * self.world_size) + self.rank) * self.batch_shard_stride + chunks: list[Tensor] = [] + for shard_slot in range(active_shards): + count = seqs_per_shard + (1 if shard_slot < seq_remainder else 0) + if count <= 0: + continue + shard_idx = (shard_start + shard_slot * self.batch_shard_stride) % self.num_shards + chunks.append(self._sample_sequences(shard_idx, count)) + self.batch_idx += 1 + local = chunks[0] if len(chunks) == 1 else torch.cat(chunks, dim=0) + local = local.to(dtype=torch.int64) + x = local[:, :-1] + y = local[:, 1:] + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +def build_train_loader(args: Hyperparameters, rank: int, world_size: int, device: torch.device): + if args.loader_mode == "sequential": + return DistributedTokenLoader(args.train_files, rank, world_size, device) + if args.loader_mode == "coprime": + return CoprimeDistributedTokenLoader( + args.train_files, + rank, + world_size, + device, + seq_len=args.train_seq_len, + seed=args.seed, + max_loaded_shards=args.coprime_max_loaded_shards, + shards_per_batch=args.coprime_shards_per_batch, + shard_hold_steps=args.coprime_shard_hold_steps, + ) + raise ValueError(f"Unknown LOADER_MODE={args.loader_mode!r}") + +# --- Transformer modules --- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + # No CastedLinear -- weights come from banks + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 # set by GPT.__init__ for partial RoPE + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False # set by GPT.__init__ for deep layers only + # Gated attention and value residual (non-banked small params) + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + self.value_residual = value_residual + if value_residual: + self.vrl_alpha = nn.Parameter(torch.zeros(1, dtype=torch.float32)) # sigmoid gate (PR #569 style) + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + """Efficient XSA: subtract self-value projection via GQA-aware reshape (no repeat_interleave). + y: [B, T, H, D], v: [B, T, Hkv, D]. H must be divisible by Hkv.""" + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) # [B, T, Hkv, group, D] + vn = F.normalize(v, dim=-1).unsqueeze(-2) # [B, T, Hkv, 1, D] -- broadcast ready + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + bsz, seqlen, dim = x.shape + q = F.linear(x, q_w.to(x.dtype)).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = F.linear(x, k_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = F.linear(x, v_w.to(x.dtype)) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v if self.value_residual else None + if self.value_residual and v0 is not None: + alpha = torch.sigmoid(self.vrl_alpha.to(dtype=v.dtype)) + v = v + alpha * v0 # sigmoid-gated residual (PR #569 style) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if flash_attn_3_func is not None: + q_attn, k_attn, v_attn = q, k, v + if q_attn.dtype not in (torch.float16, torch.bfloat16): + q_attn = q_attn.to(torch.bfloat16) + k_attn = k_attn.to(torch.bfloat16) + v_attn = v_attn.to(torch.bfloat16) + y = flash_attn_3_func(q_attn, k_attn, v_attn, causal=True) + else: + qh = q.transpose(1, 2) + kh = k.transpose(1, 2) + vh = v.transpose(1, 2) + if self.num_heads != self.num_kv_heads: + repeat = self.num_heads // self.num_kv_heads + kh = kh.repeat_interleave(repeat, dim=1) + vh = vh.repeat_interleave(repeat, dim=1) + y = F.scaled_dot_product_attention(qh, kh, vh, is_causal=True).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + # gate shape: (bsz, seqlen, num_heads) -> (bsz, seqlen, num_heads, 1) for B,T,H,D layout + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) + y = y * gate + y = y.reshape(bsz, seqlen, dim) + return F.linear(y, out_w.to(x.dtype)), raw_v + +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int, trigram: bool = False): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self._trigram = trigram + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def trigram_hash(self, tokens: Tensor) -> Tensor: + """Hash (t-2, t-1, t) trigrams into same embedding table. Zero extra params.""" + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., :2] = mod + out[..., 2:] = (36313 * t[..., 2:] ^ 27191 * t[..., 1:-1] ^ 51497 * t[..., :-2]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self._trigram: + h = h + self.embed(self.trigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + # No CastedLinear -- weights come from banks + self.kernel_mode = os.environ.get("MLP_KERNEL_MODE", "").strip().lower() + def forward(self, x: Tensor, up_w: Tensor, down_w: Tensor) -> Tensor: + x = F.linear(x, up_w.to(x.dtype)) + x = leaky_relu_sq(x, kernel_mode=self.kernel_mode) + return F.linear(x, down_w.to(x.dtype)) + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + layer_idx: int = 0, + ln_scale: bool = False, + dtg: bool = False, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=gated_attention, value_residual=value_residual) + self.mlp = MLP(dim, mlp_mult) + attn_scale_init = float(os.environ.get("ATTN_SCALE_INIT", "1.0")) + mlp_scale_init = float(os.environ.get("MLP_SCALE_INIT", "1.0")) + resid_mix_x_init = float(os.environ.get("RESID_MIX_X_INIT", "1.0")) + resid_mix_x0_init = float(os.environ.get("RESID_MIX_X0_INIT", "0.0")) + self.attn_scale = nn.Parameter(torch.full((dim,), attn_scale_init, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.full((dim,), mlp_scale_init, dtype=torch.float32)) + self.resid_mix = nn.Parameter( + torch.stack( + ( + torch.full((dim,), resid_mix_x_init, dtype=torch.float32), + torch.full((dim,), resid_mix_x0_init, dtype=torch.float32), + ) + ) + ) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, up_w: Tensor, down_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out, raw_v = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, q_w, k_w, v_w, out_w, v_embed=v_embed, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor, up_w, down_w) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out, raw_v + +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) # kv_dim for value projection + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.value_residual = value_residual + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim, trigram=bool(int(os.environ.get("TRIGRAM", "0")))) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + # Parameter banks: contiguous 3D tensors for batched optimizer + head_dim = model_dim // num_heads + kv_dim = num_kv_heads * head_dim + mlp_dim = int(mlp_mult * model_dim) + self.num_layers = num_layers + self.qo_bank = nn.Parameter(torch.empty(2 * num_layers, model_dim, model_dim)) + self.kv_bank = nn.Parameter(torch.empty(2 * num_layers, kv_dim, model_dim)) + self.mlp_up_bank = nn.Parameter(torch.empty(num_layers, mlp_dim, model_dim)) + self.mlp_down_bank = nn.Parameter(torch.empty(num_layers, model_dim, mlp_dim)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + layer_idx=i, + ln_scale=ln_scale, + dtg=dtg, + gated_attention=gated_attention, + value_residual=value_residual, + ) + for i in range(num_layers) + ] + ) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim_ve = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim_ve) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() # keep empty for compat + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + n = self.num_layers + proj_scale = 1.0 / math.sqrt(2 * n) + # Init banks: orthogonal, with proj layers scaled down and out/down zero-init + for i in range(n): + nn.init.orthogonal_(self.qo_bank.data[i], gain=1.0) # Q + nn.init.zeros_(self.qo_bank.data[n + i]) # Out (zero init) + nn.init.orthogonal_(self.kv_bank.data[i], gain=1.0) # K + nn.init.orthogonal_(self.kv_bank.data[n + i], gain=1.0) # V + nn.init.orthogonal_(self.mlp_up_bank.data[i], gain=1.0) # MLP up + nn.init.zeros_(self.mlp_down_bank.data[i]) # MLP down (zero init) + # Scale proj layers (out_proj and mlp_down are "proj" layers) + self.qo_bank.data[n + i].mul_(proj_scale) + self.mlp_down_bank.data[i].mul_(proj_scale) + # Init remaining nn.Linear modules (bigram proj, mtp heads, lm_head) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + """Get value embedding for a specific layer using shared table + per-layer scale.""" + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if hasattr(self, '_ngram_tracker') and self._ngram_tracker is not None and self.training: + per_tok_loss = F.cross_entropy(logits.float(), targets, reduction="none") + weights = self._ngram_tracker.get_weights(input_ids, target_ids) + main_loss = (per_tok_loss * weights).mean() + else: + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + def forward_logits(self, input_ids: Tensor) -> Tensor: + """Return logits (bsz, seq_len, vocab) without computing loss.""" + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + +# --- N-gram bulk update and hashed n-gram sliding eval --- + +def _ngram_bulk_update(val_np, start, end, ctx_tables, full_tables, + min_order, max_order, primes, mask): + """Bulk update n-gram tables with a contiguous range of tokens. + All ranks call this with the SAME token range -> identical tables everywhere.""" + t = val_np[start:end].astype(np.uint64) + n = len(t) + for order in range(min_order, max_order + 1): + if n < order: + continue + ctx_width = order - 1 + ctx_hash = np.zeros(n - order + 1, dtype=np.uint64) + for k in range(ctx_width): + ctx_hash ^= t[k:n - order + 1 + k] * primes[k % len(primes)] + ctx_key = (ctx_hash & mask).astype(np.int64) + tgt = t[order - 1:] + full_key = ((ctx_hash ^ (tgt * primes[ctx_width % len(primes)])) & mask).astype(np.int64) + ctx_tables[order] += np.bincount(ctx_key, minlength=len(ctx_tables[order])).astype(np.uint32) + full_tables[order] += np.bincount(full_key, minlength=len(full_tables[order])).astype(np.uint32) + +def eval_val_sliding_hashed_ngram( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + order: int, + alpha: float, + min_count: int, + buckets: int, + max_seconds: float = 0.0, + batch_seqs: int = 128, + eval_seq_len: int | None = None, +) -> tuple[float, float, float]: + """Score-first sliding eval with chunk-based SHARED n-gram tables + cubric. + + Key design: all ranks share identical n-gram tables via bulk chunk updates. + Each chunk's windows are distributed across ranks for scoring, then ALL ranks + update tables with the same contiguous token range. Every rank sees the full + n-gram picture (not 1/world_size like per-segment updates). + + Legal: entire chunk scored before its tokens update the tables. + """ + min_order = max(args.ngram_eval_min_order, 2) + max_order = max(order, min_order) + adaptive = args.ngram_eval_adaptive + alpha_min = args.ngram_eval_alpha_min + alpha_max = args.ngram_eval_alpha_max + ent_center = args.ngram_eval_entropy_center + ent_scale = args.ngram_eval_entropy_scale + + # Parse fixed per-order multipliers (PR #809 style) + _fixed_order_mults = None + if args.ngram_order_mults_str: + _fixed_order_mults = np.array([float(x) for x in args.ngram_order_mults_str.split(",")], dtype=np.float64) + + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + + # Build all windows and total scored tokens + all_window_starts = [ws for ws in range(0, total_tokens, stride) if min(ws + seq_len, total_tokens) - ws >= 1] + total_scored_tokens = 0.0 + for ws in all_window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + total_scored_tokens += float(max(wlen - s, 0)) + + # Group windows into chunks by scored position -- all ranks share this grouping + chunk_tokens = int(os.environ.get("NGRAM_CHUNK_TOKENS", "1048576")) # 1M default + num_chunks = (total_tokens + chunk_tokens - 1) // chunk_tokens + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in all_window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // chunk_tokens, num_chunks - 1) + chunk_windows[ci].append(ws) + + val_np = val_tokens.numpy() + ctx_tables = {n: np.zeros((buckets,), dtype=np.uint32) for n in range(min_order, max_order + 1)} + full_tables = {n: np.zeros((buckets,), dtype=np.uint32) for n in range(min_order, max_order + 1)} + mask = np.uint64(buckets - 1) + primes = np.array( + [np.uint64(36313), np.uint64(27191), np.uint64(51647), np.uint64(81929), + np.uint64(131071), np.uint64(174763), np.uint64(233017)], + dtype=np.uint64, + ) + + loss_sum = 0.0 + token_count = 0.0 + byte_count = 0.0 + + # Cubric 3D: per (order x entropy_bin x count_bin) adaptive alpha scaling + _NUM_ENT_BINS = 3 # low / mid / high entropy + _NUM_CNT_BINS = 3 # low / mid / high count + _ENT_EDGES = np.array([ent_center - 1.0, ent_center + 1.0]) # [2.0, 4.0] for center=3.0 + _CNT_EDGES = np.array([5.0, 50.0]) # low=<5, mid=5-50, high=>50 context count + _TOTAL_CELLS = _NUM_ENT_BINS * _NUM_CNT_BINS # 9 cells per order = 54 total + _cc = getattr(args, 'cubric_cadence', 0); _con = _cc > 0; _cfired = 0 + if _con: + # Warm-start: proven converged values from 4+ runs (orders 2-7) + # All 9 cells per order get the same warm-start, 3D cubric refines from there + _WARM = {2: 0.45, 3: 0.30, 4: 0.45, 5: 1.88, 6: 2.00, 7: 2.00, 8: 2.00, 9: 2.00} + _c_alpha_mult = {n: [_WARM.get(n, 1.0)] * _TOTAL_CELLS for n in range(min_order, max_order + 1)} + _c_hits = {n: [0] * _TOTAL_CELLS for n in range(min_order, max_order + 1)} + _c_beats = {n: [0] * _TOTAL_CELLS for n in range(min_order, max_order + 1)} + + base_model.eval() + compiled_logits = maybe_compile( + base_model.forward_logits, + enabled=args.compile_enabled, + fullgraph=False, + ) + t0 = time.perf_counter() + deadline = (t0 + max_seconds) if max_seconds > 0.0 else None + cutoff_hit = False + + if rank == 0: + print(f"ngram_eval:chunks={num_chunks} chunk_tokens={chunk_tokens} " + f"windows={len(all_window_starts)} shared_tables=True", flush=True) + + with torch.inference_mode(): + for ci in range(num_chunks): + if deadline is not None and time.perf_counter() >= deadline: + cutoff_hit = True + break + + windows = chunk_windows[ci] + if not windows: + continue + + # Distribute this chunk's windows across ranks + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + # --- Phase 1: SCORE this chunk's windows --- + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + logits_f = logits.float() + nll = F.cross_entropy( + logits_f.reshape(-1, logits_f.size(-1)), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + seg_len = wlen - s + if seg_len <= 0: + continue + + seg_nll = nll[i, s:wlen].to(torch.float64).cpu().numpy() + seg_model_p = np.exp(-seg_nll) + + if adaptive: + log_probs = F.log_softmax(logits_f[i, s:wlen], dim=-1) + probs_a = log_probs.exp() + entropy = -(probs_a * log_probs).sum(dim=-1).cpu().numpy() + sig = 1.0 / (1.0 + np.exp(-ent_scale * (entropy - ent_center))) + per_token_alpha = alpha_min + (alpha_max - alpha_min) * sig + # Bin entropy for 2D cubric: 0=low, 1=mid, 2=high + _ent_bins = np.digitize(entropy, _ENT_EDGES).astype(np.int32) + else: + per_token_alpha = np.full(seg_len, alpha) + _ent_bins = np.ones(seg_len, dtype=np.int32) # all mid + + global_j = np.arange(ws + s + 1, ws + wlen + 1, dtype=np.int64) + p_ng = np.zeros(seg_len, dtype=np.float64) + ng_matched = np.zeros(seg_len, dtype=np.bool_) + _ng_ord = np.zeros(seg_len, dtype=np.int32) + _ng_ctx_count = np.zeros(seg_len, dtype=np.float64) + tgt_np = val_np[global_j].astype(np.uint64) + + for n in range(max_order, min_order - 1, -1): + ctx_width = n - 1 + valid = (global_j >= ctx_width) & (~ng_matched) + if not valid.any(): + continue + v_idx = np.nonzero(valid)[0] + jv = global_j[v_idx] + ctx_hash = np.zeros(len(jv), dtype=np.uint64) + for k in range(ctx_width): + tok = val_np[jv - (ctx_width - k)].astype(np.uint64) + ctx_hash ^= tok * primes[k % len(primes)] + ctx_key = (ctx_hash & mask).astype(np.int64) + full_key = ((ctx_hash ^ (tgt_np[v_idx] * primes[ctx_width % len(primes)])) & mask).astype(np.int64) + ctx_counts = ctx_tables[n][ctx_key].astype(np.float64) + full_counts = full_tables[n][full_key].astype(np.float64) + has_data = ctx_counts >= float(min_count) + if has_data.any(): + p = np.minimum(full_counts, ctx_counts) / np.maximum(ctx_counts, 1.0) + p = np.clip(p, 0.0, 1.0) + hit_idx = v_idx[has_data] + p_ng[hit_idx] = p[has_data] + ng_matched[hit_idx] = True + _ng_ord[hit_idx] = n + _ng_ctx_count[hit_idx] = ctx_counts[has_data] + + # Mix where n-gram matched (PR #809 style or cubric 3D fallback) + if ng_matched.any(): + m_idx = np.nonzero(ng_matched)[0] + # Per-order entropy center shift (PR #809) + if adaptive and args.ngram_entropy_shift: + matched_ords = _ng_ord[m_idx].astype(np.float64) + shifted_centers = ent_center - 0.25 * (matched_ords - float(min_order)) + shifted_sig = 1.0 / (1.0 + np.exp(-ent_scale * (entropy[m_idx] - shifted_centers))) + per_token_alpha[m_idx] = alpha_min + (alpha_max - alpha_min) * shifted_sig + if _fixed_order_mults is not None: + # PR #809 fixed order multipliers (replaces cubric) + a = per_token_alpha[m_idx].copy() + mult_indices = _ng_ord[m_idx] - min_order + mult_indices = np.clip(mult_indices, 0, len(_fixed_order_mults) - 1) + a *= _fixed_order_mults[mult_indices] + np.clip(a, 0.0, 0.95, out=a) + elif _con: + a = per_token_alpha[m_idx].copy() + m_ent_bins = _ent_bins[m_idx] + m_cnt_bins = np.digitize(_ng_ctx_count[m_idx], _CNT_EDGES).astype(np.int32) + for n in range(min_order, max_order + 1): + om = _ng_ord[m_idx] == n + if not om.any(): + continue + for eb in range(_NUM_ENT_BINS): + for cb in range(_NUM_CNT_BINS): + cell = eb * _NUM_CNT_BINS + cb + mask_ecb = om & (m_ent_bins == eb) & (m_cnt_bins == cb) + if mask_ecb.any(): + _c_hits[n][cell] += int(mask_ecb.sum()) + _c_beats[n][cell] += int((p_ng[m_idx[mask_ecb]] > seg_model_p[m_idx[mask_ecb]]).sum()) + a[mask_ecb] *= _c_alpha_mult[n][cell] + np.clip(a, 0.0, 0.95, out=a) + else: + a = per_token_alpha[m_idx] + seg_model_p[m_idx] = (1.0 - a) * seg_model_p[m_idx] + a * p_ng[m_idx] + + seg_nll = -np.log(np.clip(seg_model_p, 1e-12, 1.0)) + loss_sum += float(seg_nll.sum()) + token_count += float(seg_len) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += float(tb.sum().item()) + + # --- Phase 2: SHARED UPDATE -- all ranks update with same chunk tokens --- + chunk_start = ci * chunk_tokens + chunk_end = min((ci + 1) * chunk_tokens, total_tokens) + _ngram_bulk_update(val_np, chunk_start, chunk_end + 1, + ctx_tables, full_tables, min_order, max_order, + primes, mask) + + # Cubric 2D c-step: adapt per (order x entropy_bin) + if _con: + # Collect all (order, ent_bin, cnt_bin) cells with enough data + all_rates = [] + for n in range(min_order, max_order + 1): + for cell in range(_TOTAL_CELLS): + if _c_hits[n][cell] >= 8: + all_rates.append(_c_beats[n][cell] / _c_hits[n][cell]) + if len(all_rates) >= 4: + avg_rate = sum(all_rates) / len(all_rates) + for n in range(min_order, max_order + 1): + for cell in range(_TOTAL_CELLS): + if _c_hits[n][cell] >= 8: + rate = _c_beats[n][cell] / _c_hits[n][cell] + if rate > avg_rate + 0.05: + _c_alpha_mult[n][cell] = min(_c_alpha_mult[n][cell] * 1.03, 2.0) + elif rate < avg_rate - 0.05: + _c_alpha_mult[n][cell] = max(_c_alpha_mult[n][cell] * 0.97, 0.3) + _cfired += 1 + if rank == 0 and _cfired % 8 == 0: + parts = [] + for n in range(min_order, max_order + 1): + m = _c_alpha_mult[n] + avg_m = sum(m) / len(m) + parts.append(f"o{n}:avg={avg_m:.2f}") + print(f"cubric3d:step={_cfired} {' '.join(parts)}", flush=True) + _c_hits = {n: [0] * _TOTAL_CELLS for n in range(min_order, max_order + 1)} + _c_beats = {n: [0] * _TOTAL_CELLS for n in range(min_order, max_order + 1)} + + # Progress + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1 or ci < 3): + elapsed = time.perf_counter() - t0 + cur_bpb = (loss_sum / max(token_count, 1.0)) / math.log(2.0) * (token_count / max(byte_count, 1.0)) if token_count > 0 else 0.0 + print( + f"ngram_eval:chunk [{ci+1}/{num_chunks}] bpb={cur_bpb:.6f} t={elapsed:.0f}s", + flush=True, + ) + + # All-reduce across ranks + _loss = torch.tensor(loss_sum, device=device, dtype=torch.float64) + _toks = torch.tensor(token_count, device=device, dtype=torch.float64) + _bytes = torch.tensor(byte_count, device=device, dtype=torch.float64) + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(_loss, op=dist.ReduceOp.SUM) + dist.all_reduce(_toks, op=dist.ReduceOp.SUM) + dist.all_reduce(_bytes, op=dist.ReduceOp.SUM) + loss_sum = _loss.item() + token_count = _toks.item() + byte_count = _bytes.item() + + coverage = token_count / max(total_scored_tokens, 1.0) + if cutoff_hit: + elapsed = time.perf_counter() - t0 + print( + f"ngram_eval:cutoff max_seconds={max_seconds:.1f} " + f"coverage={coverage*100:.2f}% elapsed={elapsed:.0f}s", + flush=True, + ) + + if _con and rank == 0: + print(f"cubric3d:final c_steps={_cfired} cells={_TOTAL_CELLS}x{max_order-min_order+1}={_TOTAL_CELLS*(max_order-min_order+1)}", flush=True) + for n in range(min_order, max_order + 1): + m = _c_alpha_mult[n] + row = " ".join(f"{m[cell]:.2f}" for cell in range(_TOTAL_CELLS)) + print(f" o{n}: [{row}]", flush=True) + val_loss = loss_sum / max(token_count, 1.0) + val_bpb = val_loss / math.log(2.0) * (token_count / max(byte_count, 1.0)) + base_model.train() + return val_loss, val_bpb, coverage + +# --- Sliding window evaluation --- + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = maybe_compile( + base_model.forward_logits, + enabled=args.compile_enabled, + fullgraph=args.compile_fullgraph, + ) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte + + +# --- Training --- + +def main() -> None: + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + # zeropower_via_newtonschulz5 runs eagerly with bmm -- do NOT compile + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size != 8: + raise ValueError( + f"Rascal 4k 12L brotli+mixed 8x requires WORLD_SIZE=8, got {world_size}. " + "Launch with: torchrun --standalone --nproc_per_node=8 4k_vocab_rascal_12l_brotli_mixed/train_gpt_4K_12L_brotli_mixed_8xgpu.py" + ) + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 100, console=False) + log0("condition_id:rascal_4k_12L_brotli_mixed_8x_seed444") + log0("run_label:salvage_v2 source_record:rascal_4k_8x_seed444_run20260427 axis:depth_12L+brotli+mixed_int") + log0("changed_fields:num_layers (11->12), compression (zstd->brotli+bshf), quant_policy (uniform_int6->mixed_int5_int6_int8)") + log0("expected_metric:final_sliding_window_exact comparator:0.8672_4k_8x_oversize_run prior_size:17766043_target:<16000000") + log0(f"condition:DATA_PATH={args.data_path}") + log0(f"condition:TOKENIZER_PATH={args.tokenizer_path}") + log0(f"condition:VOCAB_SIZE={args.vocab_size}") + log0(f"condition:SEED={args.seed}") + log0(f"condition:MAX_WALLCLOCK_SECONDS={args.max_wallclock_seconds}") + log0(f"condition:LOADER_MODE={args.loader_mode}") + log0(f"condition:COPRIME_MAX_LOADED_SHARDS={args.coprime_max_loaded_shards}") + log0(f"condition:COPRIME_SHARDS_PER_BATCH={args.coprime_shards_per_batch}") + log0(f"condition:COPRIME_SHARD_HOLD_STEPS={args.coprime_shard_hold_steps}") + log0(f"condition:SKIP_GPTQ={os.environ.get('SKIP_GPTQ', '1')}") + log0(f"condition:TRIGRAM={int(args.trigram_enabled)}") + log0(f"condition:NGRAM_EVAL_ORDER={args.ngram_eval_order}") + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + if args.ngram_eval_order >= 2: + log0(f"ngram_eval:order={args.ngram_eval_order} alpha={args.ngram_eval_alpha} min_count={args.ngram_eval_min_count} buckets={args.ngram_eval_buckets}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + CastedLinear._qat_enabled = args.qat_enabled + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + gated_attention=args.gated_attention, + value_residual=args.value_residual, + ).to(device).bfloat16() + # Banks stay FP32 (like CastedLinear weights), cast to BF16 in forward + base_model.qo_bank.data = base_model.qo_bank.data.float() + base_model.kv_bank.data = base_model.kv_bank.data.float() + base_model.mlp_up_bank.data = base_model.mlp_up_bank.data.float() + base_model.mlp_down_bank.data = base_model.mlp_down_bank.data.float() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + if args.complement_alpha > 0: + tracker = TrainNgramTracker(args.vocab_size, device, complement_alpha=args.complement_alpha) + base_model._ngram_tracker = tracker + log0(f"complementary_training:alpha={args.complement_alpha}") + else: + base_model._ngram_tracker = None + # No DDP -- Parallel Muon handles bank grad communication via reduce-scatter, + # and non-bank grads are manually all-reduced before Adam steps. + compiled_model = maybe_compile( + base_model, + enabled=args.compile_enabled, + fullgraph=args.compile_fullgraph, + mode=args.compile_mode, + ) + model = compiled_model + + # Optimizer split: + # - 4 parameter banks -> Muon (batched Newton-Schulz) + # - token embedding -> Adam + # - scalars/control tensors -> Adam + # - bigram proj, mtp heads, VE proj -> Adam (small matrix params not worth banking) + matrix_params = [ + base_model.qo_bank, base_model.kv_bank, + base_model.mlp_up_bank, base_model.mlp_down_bank, + ] + block_named_params = list(base_model.blocks.named_parameters()) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + scalar_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + scalar_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + # Non-bank params that need manual all-reduce (replicated across GPUs) + replicated_params = list(optimizer_tok.param_groups[0]["params"]) + for pg in optimizer_tok.param_groups[1:]: + replicated_params.extend(pg["params"]) + replicated_params.extend(scalar_params) + + optimizer_head = None + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + replicated_params.append(base_model.lm_head.weight) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if optimizer_head is not None: + optimizers.append(optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + compile_mode = args.compile_mode if args.compile_mode else "default" + log0( + f"compile:enabled={int(args.compile_enabled)} mode:{compile_mode} " + f"fullgraph={int(args.compile_fullgraph)}" + ) + log0(f"mlp_kernel_mode:{args.mlp_kernel_mode or 'eager'}") + log0( + f"scale_init:attn={args.attn_scale_init:.4f} mlp={args.mlp_scale_init:.4f} " + f"resid_mix=({args.resid_mix_x_init:.4f},{args.resid_mix_x0_init:.4f}) " + f"ln_scale={int(args.ln_scale)}" + ) + log0(f"seed:{args.seed}") + train_loader = build_train_loader(args, rank, world_size, device) + log0(train_loader.describe()) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + # GPTQ calibration reads training data — it must complete within the wallclock budget. + # We stop the training loop early (by GPTQ_RESERVE_MS) so GPTQ runs before the cap. + _skip_gptq = int(os.environ.get("SKIP_GPTQ", "1")) + _gptq_reserve_ms = float(os.environ.get("GPTQ_RESERVE_MS", "30000")) if (max_wallclock_ms is not None and not _skip_gptq) else 0.0 + effective_max_wallclock_ms = (max_wallclock_ms - _gptq_reserve_ms) if max_wallclock_ms is not None else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + # All-reduce all grads for warmup (simple, not optimized) + if distributed: + for p in base_model.parameters(): + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + train_loader = build_train_loader(args, rank, world_size, device) + log0(f"loader_reset:{train_loader.describe()}") + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + from collections import deque + lawa_queue: deque[dict[str, Tensor]] = deque(maxlen=args.lawa_k) + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms tok/s:{(step * args.train_batch_tokens) / max(training_time_ms / 1000.0, 1e-9):.0f}" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + if base_model._ngram_tracker is not None: + base_model._ngram_tracker.update(x, y) + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + # === 3-phase overlapped optimizer step === + # Phase 1: Launch async reduce-scatter for banks (biggest first) + optimizer_muon.launch_reduce_scatters() + # Phase 2: All-reduce non-bank grads + step Adam (while bank RS is in-flight) + if distributed: + for p in replicated_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + optimizer_tok.step() + optimizer_scalar.step() + if optimizer_head is not None: + optimizer_head.step() + # Phase 3: Wait for RS, local NS5, all-gather (banks processed last) + optimizer_muon.step() + zero_grad_all() + # EMA update + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + if args.lawa_enabled and step % args.lawa_freq == 0: + lawa_queue.append({name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()}) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms tok/s:{(step * args.train_batch_tokens) / max(approx_training_time_ms / 1000.0, 1e-9):.0f}" + ) + reached_cap = effective_max_wallclock_ms is not None and approx_training_time_ms >= effective_max_wallclock_ms + if distributed and effective_max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + # GPTQ calibration: reads training data — must complete within MAX_WALLCLOCK_SECONDS. + # Training loop stopped GPTQ_RESERVE_MS early so this runs inside the budget. + if _skip_gptq: + log0("gptq:SKIPPED (SKIP_GPTQ=1) — will use naive int6") + gptq_hessians: dict[str, Tensor] = {} + else: + log0("gptq:calibrating with training data...") + t_gptq = time.perf_counter() + gptq_hessians = gptq_calibrate(base_model, args.train_files, device, n_samples=256, seq_len=args.train_seq_len) + log0(f"gptq:calibrated {len(gptq_hessians)} layers in {time.perf_counter()-t_gptq:.1f}s") + # Apply weight averaging + if args.lawa_enabled and len(lawa_queue) > 1: + log0(f"lawa:applying LAWA averaging k={len(lawa_queue)}") + current_state = base_model.state_dict() + avg_state = {name: torch.zeros(t.shape, dtype=torch.float32, device='cpu') for name, t in current_state.items()} + for snap in lawa_queue: + for name in avg_state: + avg_state[name] += snap[name].float() + for name in avg_state: + avg_state[name] /= len(lawa_queue) + avg_state[name] = avg_state[name].to(dtype=current_state[name].dtype) + base_model.load_state_dict(avg_state, strict=True) + else: + log0("ema:applying EMA weights") + current_state = base_model.state_dict() + avg_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(avg_state, strict=True) + if args.post_ema_diagnostic: + torch.cuda.synchronize() + t_diag = time.perf_counter() + diag_val_loss, diag_val_bpb = eval_val( + args, compiled_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"DIAGNOSTIC post_ema val_loss:{diag_val_loss:.4f} val_bpb:{diag_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" + ) + else: + log0("diagnostic_eval:skipped POST_EMA_DIAGNOSTIC=0") + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k) + if excluded_mtp > 0: + log0(f"export_excluding_mtp_params:{excluded_mtp}") + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + # GPTQ quantization using Hessians collected from training data. + # salvage_v2 mixed-int policy: int5 for mlp_up_bank/mlp_down_bank (most quant-tolerant per + # 11-day collate), int6 for qo_bank/kv_bank/embed (attention + token embed kept at int6 for + # quant safety; matches seed for embed). See `mixed_quantize_int6_gptq` docstring. + quant_result, quant_meta = mixed_quantize_int6_gptq( + sd_cpu, + int6_cats={"qo", "kv", "attn_other", "mlp_other", "aux", "embed"}, + hessians=gptq_hessians, + int5_cats={"mlp_down", "mlp_up"}, + ) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = _compress_blob(quant_raw) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + log0(f"Serialized model mixed_int5_int6_int8+{_COMPRESSOR}: {quant_file_bytes} bytes") + log0(f"Total submission size mixed_int5_int6_int8+{_COMPRESSOR}: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(_decompress_blob(quant_blob_disk)), + map_location="cpu", + ) + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, rope_dims=args.rope_dims, ln_scale=args.ln_scale, + dtg=args.dtg_enabled, ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + gated_attention=args.gated_attention, value_residual=args.value_residual, + ).to(device).bfloat16() + eval_model.qo_bank.data = eval_model.qo_bank.data.float() + eval_model.kv_bank.data = eval_model.kv_bank.data.float() + eval_model.mlp_up_bank.data = eval_model.mlp_up_bank.data.float() + eval_model.mlp_down_bank.data = eval_model.mlp_down_bank.data.float() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, eval_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + del eval_model, deq_state, quant_state, sd_cpu + torch.cuda.empty_cache() + sw_seq_len = effective_eval_seq_len + if args.skip_final_eval: + log0("final_eval:skipped sliding/ngram by SKIP_FINAL_EVAL=1") + else: + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, base_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, base_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + if args.ngram_eval_order >= 2: + if distributed: + dist.barrier() + torch.cuda.synchronize() + t_ng = time.perf_counter() + ng_loss, ng_bpb, ng_coverage = eval_val_sliding_hashed_ngram( + args, + base_model, + rank, + world_size, + device, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + stride=args.eval_stride, + order=args.ngram_eval_order, + alpha=args.ngram_eval_alpha, + min_count=args.ngram_eval_min_count, + buckets=args.ngram_eval_buckets, + max_seconds=args.ngram_eval_max_seconds, + eval_seq_len=sw_seq_len, + ) + if rank == 0: + torch.cuda.synchronize() + ng_eval_ms = 1000.0 * (time.perf_counter() - t_ng) + if ng_coverage >= 0.999999: + log0( + f"final_sliding_window_ngram{args.ngram_eval_order} val_loss:{ng_loss:.4f} " + f"val_bpb:{ng_bpb:.4f} eval_time:{ng_eval_ms:.0f}ms" + ) + log0( + f"final_sliding_window_ngram{args.ngram_eval_order}_exact " + f"val_loss:{ng_loss:.8f} val_bpb:{ng_bpb:.8f}" + ) + else: + log0( + f"final_sliding_window_ngram{args.ngram_eval_order}_partial val_loss:{ng_loss:.4f} " + f"val_bpb:{ng_bpb:.4f} coverage:{ng_coverage:.4f} eval_time:{ng_eval_ms:.0f}ms" + ) + log0( + f"final_sliding_window_ngram{args.ngram_eval_order}_partial_exact " + f"val_loss:{ng_loss:.8f} val_bpb:{ng_bpb:.8f} coverage:{ng_coverage:.8f}" + ) + if distributed: + dist.barrier() + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() + +==================================================================================================== +condition_id:rascal_4k_12L_brotli_mixed_8x_seed444 +run_label:salvage_v2 source_record:rascal_4k_8x_seed444_run20260427 axis:depth_12L+brotli+mixed_int +changed_fields:num_layers (11->12), compression (zstd->brotli+bshf), quant_policy (uniform_int6->mixed_int5_int6_int8) +expected_metric:final_sliding_window_exact comparator:0.8672_4k_8x_oversize_run prior_size:17766043_target:<16000000 +condition:DATA_PATH=./data/datasets/fineweb10B_sp4096 +condition:TOKENIZER_PATH=./data/tokenizers/fineweb_4096_bpe.model +condition:VOCAB_SIZE=4096 +condition:SEED=300 +condition:MAX_WALLCLOCK_SECONDS=600.0 +condition:LOADER_MODE=coprime +condition:COPRIME_MAX_LOADED_SHARDS=143 +condition:COPRIME_SHARDS_PER_BATCH=1 +condition:COPRIME_SHARD_HOLD_STEPS=64 +condition:SKIP_GPTQ=1 +condition:TRIGRAM=0 +condition:NGRAM_EVAL_ORDER=0 +Running Python 3.12.13 | packaged by conda-forge | (main, Mar 5 2026, 16:50:00) [GCC 14.3.0] +Running PyTorch 2.11.0+cu130 +Mon Apr 27 03:54:52 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 580.95.05 Driver Version: 580.95.05 CUDA Version: 13.0 | ++-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:04:00.0 Off | 0 | +| N/A 40C P0 115W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:05:00.0 Off | 0 | +| N/A 35C P0 119W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:0A:00.0 Off | 0 | +| N/A 39C P0 119W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:0B:00.0 Off | 0 | +| N/A 35C P0 119W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:84:00.0 Off | 0 | +| N/A 39C P0 118W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:85:00.0 Off | 0 | +| N/A 34C P0 117W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:8A:00.0 Off | 0 | +| N/A 38C P0 120W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:8B:00.0 Off | 0 | +| N/A 34C P0 116W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| No running processes found | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_4096_bpe.model +train_loader:dataset:fineweb10B_sp4096 train_shards:143 +val_loader:shards pattern=./data/datasets/fineweb10B_sp4096/fineweb_val_*.bin tokens:45514752 +model_params:31321700 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_11 active_layers:[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11] +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +compile:enabled=1 mode:default fullgraph=1 +mlp_kernel_mode:eager +scale_init:attn=1.0000 mlp=1.0000 resid_mix=(1.0000,0.0000) ln_scale=1 +seed:300 +loader:coprime shards:143 blocks:6967965 seq_len:2048 shards_per_batch:1 cache:143 batch_stride:5 hold_steps:64 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +loader_reset:loader:coprime shards:143 blocks:6967965 seq_len:2048 shards_per_batch:1 cache:143 batch_stride:5 hold_steps:64 +step:0/20000 val_loss:8.3069 val_bpb:2.8323 train_time:0ms step_avg:0.01ms tok/s:0 +step:1/20000 train_loss:8.3049 train_time:357ms step_avg:356.99ms tok/s:2202934 +step:2/20000 train_loss:9.4872 train_time:407ms step_avg:203.60ms tok/s:3862598 +step:3/20000 train_loss:8.8307 train_time:508ms step_avg:169.43ms tok/s:4641597 +step:4/20000 train_loss:8.1811 train_time:610ms step_avg:152.38ms tok/s:5160983 +step:5/20000 train_loss:8.1419 train_time:711ms step_avg:142.27ms tok/s:5527751 +step:6/20000 train_loss:8.2182 train_time:814ms step_avg:135.59ms tok/s:5800185 +step:7/20000 train_loss:8.0845 train_time:915ms step_avg:130.75ms tok/s:6014567 +step:8/20000 train_loss:7.8136 train_time:1018ms step_avg:127.24ms tok/s:6180784 +step:9/20000 train_loss:7.4422 train_time:1119ms step_avg:124.38ms tok/s:6322879 +step:10/20000 train_loss:7.3104 train_time:1221ms step_avg:122.13ms tok/s:6439340 +step:500/20000 train_loss:3.1363 train_time:52838ms step_avg:105.68ms tok/s:7441968 +step:1000/20000 train_loss:2.9131 train_time:105746ms step_avg:105.75ms tok/s:7436986 +step:1500/20000 train_loss:2.8350 train_time:158889ms step_avg:105.93ms tok/s:7424371 +step:2000/20000 train_loss:2.8340 train_time:212131ms step_avg:106.07ms tok/s:7414599 +step:2500/20000 train_loss:2.8605 train_time:265160ms step_avg:106.06ms tok/s:7414703 +step:3000/20000 train_loss:2.7954 train_time:317918ms step_avg:105.97ms tok/s:7421073 +step:3500/20000 train_loss:2.7500 train_time:370945ms step_avg:105.98ms tok/s:7420279 +step:4000/20000 train_loss:2.6705 train_time:423988ms step_avg:106.00ms tok/s:7419378 +step:4000/20000 val_loss:2.6826 val_bpb:0.9147 train_time:424041ms step_avg:106.01ms tok/s:7418447 +step:4500/20000 train_loss:2.6108 train_time:477068ms step_avg:106.02ms tok/s:7418106 +swa:start step:5000 +step:5000/20000 train_loss:2.6237 train_time:530132ms step_avg:106.03ms tok/s:7417316 +late_qat:enabled step:5131 scale:0.1499 +step:5500/20000 train_loss:2.5543 train_time:583851ms step_avg:106.15ms tok/s:7408359 +step:5649/20000 val_loss:2.5868 val_bpb:0.8820 train_time:600112ms step_avg:106.23ms tok/s:7402879 +stopping_early: wallclock_cap train_time:600112ms step:5649/20000 +peak memory allocated: 25252 MiB reserved: 25750 MiB +gptq:SKIPPED (SKIP_GPTQ=1) — will use naive int6 +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:2.5842 val_bpb:0.8811 eval_time:1718ms +Serialized model: 119539322 bytes +Code size: 128464 bytes +Serialized model mixed_int5_int6_int8+brotli: 15465911 bytes +Total submission size mixed_int5_int6_int8+brotli: 15594375 bytes +final_int6_roundtrip val_loss:2.6727 val_bpb:0.9113 eval_time:5050ms +final_int6_roundtrip_exact val_loss:2.67270610 val_bpb:0.91128978 +final_sliding_window val_loss:2.5428 val_bpb:0.8670 stride:64 eval_time:65449ms +final_sliding_window_exact val_loss:2.54276697 val_bpb:0.86698133 diff --git a/records/track_10min_16mb/2026-04-27_Mikey/train_seed42.log b/records/track_10min_16mb/2026-04-27_Mikey/train_seed42.log new file mode 100644 index 0000000000..726b5eef07 --- /dev/null +++ b/records/track_10min_16mb/2026-04-27_Mikey/train_seed42.log @@ -0,0 +1,2801 @@ +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +from collections import OrderedDict +from pathlib import Path +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP + +os.environ.setdefault("RUN_ID", "rascal_4k_12L_brotli_mixed_8x_seed444_" + time.strftime("%Y%m%d_%H%M%S")) + +try: + import triton + import triton.language as tl +except ImportError: + triton = None + tl = None +try: + from flash_attn_interface import flash_attn_func as flash_attn_3_func +except ImportError: + flash_attn_3_func = None +# Compression: brotli-11 + byte-shuffle is the salvage_v2 preferred path (PR #1493 recipe). +# Falls back to zstd then zlib so this file still runs if brotli isn't installed. +_brotli_module = None +_zstandard_module = None +_zlib_module = None +try: + import brotli as _brotli_module + _COMPRESSOR = "brotli" +except ImportError: + try: + import zstandard as _zstandard_module + _COMPRESSOR = "zstd" + import warnings + warnings.warn("brotli not found — falling back to zstd (~1MB+ larger). pip install brotli") + except ImportError: + import zlib as _zlib_module + import warnings + _COMPRESSOR = "zlib" + warnings.warn("brotli/zstandard not found — falling back to zlib. Artifact will be much larger! pip install brotli") +# Backwards-compat shims so any remaining `zstandard.*` or `_zlib_module.*` references still work. +if _zstandard_module is not None: + zstandard = _zstandard_module +if _zlib_module is None: + import zlib as _zlib_module # always available; used by zlib fallback path +# --- Byte-shuffle (de-interleave) wrapper from PR #1493: improves brotli ratio on quantized payloads. --- +_BSHF_MAGIC = b"BSHF" +def _byte_shuffle(data: bytes, stride: int = 2) -> bytes: + if stride <= 1 or len(data) < stride: + return data + src = np.frombuffer(data, dtype=np.uint8) + n = len(src) + out = np.empty(n, dtype=np.uint8) + dest_off = 0 + for pos in range(stride): + chunk = src[pos::stride] + out[dest_off:dest_off + len(chunk)] = chunk + dest_off += len(chunk) + return _BSHF_MAGIC + bytes([stride]) + out.tobytes() +def _byte_unshuffle(data: bytes) -> bytes: + if len(data) < 5 or data[:4] != _BSHF_MAGIC: + return data + stride = data[4] + if stride < 2: + return data[5:] + payload = np.frombuffer(data, dtype=np.uint8, offset=5) + n = len(payload) + out = np.empty(n, dtype=np.uint8) + src_off = 0 + for pos in range(stride): + chunk_len = n // stride + (1 if pos < n % stride else 0) + out[pos::stride][:chunk_len] = payload[src_off:src_off + chunk_len] + src_off += chunk_len + return out.tobytes() +def _compress_blob(raw: bytes) -> bytes: + if _COMPRESSOR == "brotli": + return _brotli_module.compress(_byte_shuffle(raw, stride=2), quality=11) + elif _COMPRESSOR == "zstd": + return _zstandard_module.ZstdCompressor(level=22).compress(raw) + else: + return _zlib_module.compress(raw, 9) +def _decompress_blob(blob: bytes) -> bytes: + if _COMPRESSOR == "brotli": + return _byte_unshuffle(_brotli_module.decompress(blob)) + elif _COMPRESSOR == "zstd": + return _zstandard_module.ZstdDecompressor().decompress(blob) + else: + return _zlib_module.decompress(blob) + +if os.environ.get("TORCHDYNAMO_SUPPRESS_ERRORS", "0") == "1": + import torch._dynamo + torch._dynamo.config.suppress_errors = True +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp4096") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_4096_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 444)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 4096)) + num_layers = int(os.environ.get("NUM_LAYERS", 12)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + lawa_enabled = bool(int(os.environ.get("LAWA_ENABLED", "0"))) + lawa_k = int(os.environ.get("LAWA_K", 10)) + lawa_freq = int(os.environ.get("LAWA_FREQ", 100)) + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + trigram_enabled = bool(int(os.environ.get("TRIGRAM", "0"))) # TrigramHash (off by default, risky) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 11)) # XSA on ALL layers (our novel contribution) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + gated_attention = bool(int(os.environ.get("GATED_ATTENTION", "0"))) + value_residual = bool(int(os.environ.get("VALUE_RESIDUAL", "0"))) # VRL with sigmoid gates (off by default, risky) + attn_scale_init = float(os.environ.get("ATTN_SCALE_INIT", 1.0)) + mlp_scale_init = float(os.environ.get("MLP_SCALE_INIT", 1.0)) + resid_mix_x_init = float(os.environ.get("RESID_MIX_X_INIT", 1.0)) + resid_mix_x0_init = float(os.environ.get("RESID_MIX_X0_INIT", 0.0)) + complement_alpha = float(os.environ.get("COMPLEMENT_ALPHA", "0")) + ngram_eval_order = int(os.environ.get("NGRAM_EVAL_ORDER", 0)) + ngram_eval_min_order = int(os.environ.get("NGRAM_EVAL_MIN_ORDER", 2)) + ngram_eval_alpha = float(os.environ.get("NGRAM_EVAL_ALPHA", 0.30)) + ngram_eval_adaptive = bool(int(os.environ.get("NGRAM_EVAL_ADAPTIVE", "1"))) + ngram_eval_alpha_min = float(os.environ.get("NGRAM_EVAL_ALPHA_MIN", 0.05)) + ngram_eval_alpha_max = float(os.environ.get("NGRAM_EVAL_ALPHA_MAX", 0.60)) + ngram_eval_entropy_center = float(os.environ.get("NGRAM_EVAL_ENTROPY_CENTER", 4.0)) + ngram_eval_entropy_scale = float(os.environ.get("NGRAM_EVAL_ENTROPY_SCALE", 2.0)) + ngram_eval_min_count = int(os.environ.get("NGRAM_EVAL_MIN_COUNT", 2)) + ngram_eval_buckets = int(os.environ.get("NGRAM_EVAL_BUCKETS", 4_194_304)) + ngram_eval_max_seconds = float(os.environ.get("NGRAM_EVAL_MAX_SECONDS", 0.0)) + ngram_entropy_shift = bool(int(os.environ.get("NGRAM_ENTROPY_SHIFT", "0"))) + ngram_order_mults_str = os.environ.get("NGRAM_ORDER_MULTS", "") + cubric_cadence = int(os.environ.get("CUBRIC_CADENCE", 0)) + skip_final_eval = bool(int(os.environ.get("SKIP_FINAL_EVAL", "0"))) + post_ema_diagnostic = bool(int(os.environ.get("POST_EMA_DIAGNOSTIC", "1"))) + compile_enabled = bool(int(os.environ.get("COMPILE_ENABLED", "1"))) + compile_mode = os.environ.get("COMPILE_MODE", "").strip() + compile_fullgraph = bool(int(os.environ.get("COMPILE_FULLGRAPH", "1"))) + mlp_kernel_mode = os.environ.get("MLP_KERNEL_MODE", "").strip().lower() + loader_mode = os.environ.get("LOADER_MODE", "coprime").strip().lower() + coprime_max_loaded_shards = int(os.environ.get("COPRIME_MAX_LOADED_SHARDS", 143)) + coprime_shards_per_batch = int(os.environ.get("COPRIME_SHARDS_PER_BATCH", 1)) + coprime_shard_hold_steps = int(os.environ.get("COPRIME_SHARD_HOLD_STEPS", 64)) + + +def maybe_compile(fn_or_module, *, enabled: bool, fullgraph: bool, mode: str = ""): + if not enabled: + return fn_or_module + kwargs = dict(dynamic=False, fullgraph=fullgraph) + if mode: + kwargs["mode"] = mode + return torch.compile(fn_or_module, **kwargs) + + +if triton is not None: + @triton.jit + def _leaky_relu_sq_forward_kernel(x_ptr, y_ptr, n_elements, BLOCK_SIZE: tl.constexpr): + pid = tl.program_id(0) + offsets = pid * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE) + mask = offsets < n_elements + x = tl.load(x_ptr + offsets, mask=mask, other=0.0).to(tl.float32) + a = tl.where(x >= 0, x, 0.5 * x) + y = a * a + tl.store(y_ptr + offsets, y, mask=mask) + + @triton.jit + def _leaky_relu_sq_backward_kernel(x_ptr, grad_out_ptr, grad_in_ptr, n_elements, BLOCK_SIZE: tl.constexpr): + pid = tl.program_id(0) + offsets = pid * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE) + mask = offsets < n_elements + x = tl.load(x_ptr + offsets, mask=mask, other=0.0).to(tl.float32) + grad_out = tl.load(grad_out_ptr + offsets, mask=mask, other=0.0).to(tl.float32) + a = tl.where(x >= 0, x, 0.5 * x) + slope = tl.where(x >= 0, 1.0, 0.5) + grad_in = grad_out * (2.0 * a * slope) + tl.store(grad_in_ptr + offsets, grad_in, mask=mask) + + +class TritonLeakyReluSqFn(torch.autograd.Function): + @staticmethod + def forward(ctx, x: Tensor) -> Tensor: + if triton is None or not x.is_cuda: + a = F.leaky_relu(x, negative_slope=0.5) + ctx.save_for_backward(x) + return a.square() + x_contig = x.contiguous() + y = torch.empty_like(x_contig) + n_elements = x_contig.numel() + grid = lambda meta: (triton.cdiv(n_elements, meta["BLOCK_SIZE"]),) + _leaky_relu_sq_forward_kernel[grid](x_contig, y, n_elements, BLOCK_SIZE=1024) + ctx.save_for_backward(x_contig) + return y + + @staticmethod + def backward(ctx, grad_out: Tensor) -> tuple[Tensor]: + (x,) = ctx.saved_tensors + if triton is None or not grad_out.is_cuda: + a = F.leaky_relu(x, negative_slope=0.5) + slope = torch.where(x >= 0, torch.ones_like(x), torch.full_like(x, 0.5)) + return (grad_out * (2.0 * a * slope),) + grad_out_contig = grad_out.contiguous() + grad_in = torch.empty_like(grad_out_contig) + n_elements = grad_out_contig.numel() + grid = lambda meta: (triton.cdiv(n_elements, meta["BLOCK_SIZE"]),) + _leaky_relu_sq_backward_kernel[grid](x, grad_out_contig, grad_in, n_elements, BLOCK_SIZE=1024) + return (grad_in,) + + +def leaky_relu_sq(x: Tensor, kernel_mode: str = "") -> Tensor: + if kernel_mode == "triton_act": + return TritonLeakyReluSqFn.apply(x) + a = F.leaky_relu(x, negative_slope=0.5) + return a.square() + +class TrainNgramTracker: + """Complementary training: track bigram stats, downweight tokens n-grams can predict.""" + def __init__(self, vocab_size: int, device: torch.device, complement_alpha: float = 0.5): + self.V = vocab_size + self.alpha = complement_alpha + self.bi_counts = torch.zeros(vocab_size, vocab_size, device=device, dtype=torch.float32) + self.bi_totals = torch.zeros(vocab_size, device=device, dtype=torch.float32) + @torch.no_grad() + def update(self, x: Tensor, y: Tensor): + xf = x.reshape(-1) + yf = y.reshape(-1) + ones = torch.ones(xf.numel(), device=xf.device, dtype=torch.float32) + self.bi_counts.reshape(-1).scatter_add_(0, xf * self.V + yf, ones) + self.bi_totals.scatter_add_(0, xf, ones) + def get_weights(self, x: Tensor, y: Tensor) -> Tensor: + xf = x.reshape(-1) + yf = y.reshape(-1) + total = self.bi_totals[xf] + count = self.bi_counts.reshape(-1)[xf * self.V + yf] + ngram_prob = count / (total + 1) + return (1.0 - self.alpha * ngram_prob).clamp(min=0.1) + +# --- Batched Newton-Schulz orthogonalization --- + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 5, eps: float = 1e-7) -> Tensor: + """Batched Newton-Schulz orthogonalization. G: (B,M,N) or (M,N).""" + a, b, c = (3.4445, -4.7750, 2.0315) + was_2d = G.ndim == 2 + if was_2d: + G = G.unsqueeze(0) + X = G.bfloat16() + transposed = X.size(-2) > X.size(-1) + if transposed: + X = X.mT + X = X / (X.norm(dim=(-2, -1), keepdim=True) + eps) + for _ in range(steps): + A = X @ X.mT + B = b * A + c * (A @ A) + X = a * X + B @ X + if transposed: + X = X.mT + if was_2d: + X = X.squeeze(0) + return X + +# --- Parallel Muon optimizer --- + +class Muon(torch.optim.Optimizer): + """Parallel Muon: post-backward reduce-scatter -> local NS5 -> all-gather. + + No DDP for bank params. After backward, this optimizer: + 1. Launches async reduce-scatter for all banks (biggest first) + 2. Returns control so Adam can step on small params while RS is in-flight + 3. Waits for each RS, runs local NS5 on the shard, launches async all-gather + 4. Each all-gather overlaps with next bank's NS5 + """ + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + self._built = False + + def _build(self): + self._distributed = dist.is_available() and dist.is_initialized() + self._world_size = dist.get_world_size() if self._distributed else 1 + self._rank = dist.get_rank() if self._distributed else 0 + ws = self._world_size + + self._bank_meta = [] + for group in self.param_groups: + for p in group["params"]: + B = p.shape[0] + padded_B = ((B + ws - 1) // ws) * ws + shard_B = padded_B // ws + tail = p.shape[1:] + dev = p.device + self._bank_meta.append({ + 'p': p, + 'B': B, + 'padded_grad': torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + 'shard': torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + 'shard_mom': torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + 'full_update': torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + 'scale': max(1, p.shape[-2] / p.shape[-1]) ** 0.5, + }) + # Sort by size descending -- launch biggest reduce-scatters first + self._bank_meta.sort(key=lambda m: -m['p'].numel()) + self._built = True + + def launch_reduce_scatters(self): + """Phase 1: launch async reduce-scatter for all banks. Call right after backward.""" + if not self._built: + self._build() + if not self._distributed: + return + self._rs_futures = [] + for m in self._bank_meta: + p = m['p'] + if p.grad is None: + self._rs_futures.append(None) + continue + pg = m['padded_grad'] + pg[:m['B']].copy_(p.grad.bfloat16()) + if pg.shape[0] > m['B']: + pg[m['B']:].zero_() + fut = dist.reduce_scatter_tensor(m['shard'], pg, op=dist.ReduceOp.AVG, async_op=True) + self._rs_futures.append(fut) + + @torch.no_grad() + def step(self, closure=None): + """Phase 3: wait for RS, local NS5, all-gather. Call AFTER Adam steps.""" + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + if not self._built: + self._build() + + for group in self.param_groups: + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + wd = group.get("weight_decay", 0.0) + + prev_ag_handle = None + prev_m = None + + sharded = self._distributed and hasattr(self, '_rs_futures') + + for i, m in enumerate(self._bank_meta): + p = m['p'] + if p.grad is None: + continue + + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m['p'] + upd = prev_m['full_update'][:prev_m['B']] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m['scale']) + + if sharded and self._rs_futures[i] is not None: + self._rs_futures[i].wait() + g = m['shard'] + buf = m['shard_mom'] + else: + g = p.grad.bfloat16() + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + + buf.mul_(momentum).add_(g) + if nesterov: + update = g.add(buf, alpha=momentum) + else: + update = buf + + update = zeropower_via_newtonschulz5(update, steps=backend_steps) + + if sharded: + prev_ag_handle = dist.all_gather_into_tensor( + m['full_update'], update, async_op=True) + prev_m = m + else: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + p.add_(update.to(dtype=p.dtype), alpha=-lr * m['scale']) + + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m['p'] + upd = prev_m['full_update'][:prev_m['B']] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m['scale']) + + if hasattr(self, '_rs_futures'): + del self._rs_futures + + return loss + +# --- Tokenizer evaluation helpers --- + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) + +# --- Quantization helpers --- + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale,attn_gate,vr_lambda", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if "f1_corr_in" in name or "f1_corr_out" in name: + return "aux" + if "qo_bank" in name or "kv_bank" in name: + return "attn" + if "mlp_up_bank" in name or "mlp_down_bank" in name: + return "mlp" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +# GPTQ: Hessian-aware quantization with column-wise error compensation +def _find_best_row_scales(W: Tensor, clip_range: int = 31) -> Tensor: + t32 = W.float() + best_s = t32.abs().amax(dim=1) / clip_range + best_s = best_s.clamp_min(1.0 / clip_range) + best_err = torch.full((t32.shape[0],), float('inf')) + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range) + q = torch.clamp(torch.round(t32 / s[:, None]), -clip_range, clip_range) + recon = q * s[:, None] + err = (t32 - recon).pow(2).mean(dim=1) + improved = err < best_err + best_s[improved] = s[improved] + best_err[improved] = err[improved] + return best_s +def gptq_quantize_weight(W: Tensor, H: Tensor, clip_range: int = 31, + block_size: int = 64, percdamp: float = 0.002) -> tuple[Tensor, Tensor]: + """GPTQ: quantize weight matrix W using Hessian H = X^T X for error compensation. + Returns (quantized_int8, scale_fp16) in int6 range [-clip_range, clip_range].""" + W = W.float().clone() + rows, cols = W.shape + row_scale = _find_best_row_scales(W, clip_range) + H = H.float().clone() + damp = percdamp * H.diag().mean() + H.diagonal().add_(damp) + perm = torch.argsort(H.diag()) + invperm = torch.argsort(perm) + W = W[:, perm] + H = H[perm][:, perm] + try: + L = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(L) + except torch._C._LinAlgError: + Hinv = torch.diag(1.0 / H.diag().clamp_min(1e-6)) + Q = torch.zeros(rows, cols, dtype=torch.int8) + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + W_block = W[:, i1:i2].clone() + Hinv_block = Hinv[i1:i2, i1:i2] + Err = torch.zeros_like(W_block) + for j in range(i2 - i1): + w_col = W_block[:, j] + h_inv_jj = Hinv_block[j, j].clamp_min(1e-8) + q_col = torch.clamp(torch.round(w_col / row_scale), -clip_range, clip_range) + deq_col = q_col * row_scale + Q[:, i1 + j] = q_col.to(torch.int8) + err = (w_col - deq_col) / h_inv_jj + Err[:, j] = err + if j + 1 < i2 - i1: + W_block[:, j + 1:] -= err.unsqueeze(1) * Hinv_block[j, j + 1:].unsqueeze(0) + if i2 < cols: + W[:, i2:] -= Err @ Hinv[i1:i2, i2:] + Q = Q[:, invperm] + return Q, row_scale.to(torch.float16) +def gptq_calibrate(model: nn.Module, train_pattern: str, device: torch.device, + n_samples: int = 256, seq_len: int = 2048) -> dict[str, Tensor]: + """Collect Hessian H = X^T X for each linear layer using training data.""" + hessians: dict[str, Tensor] = {} + n_seen: dict[str, int] = {} + hooks = [] + def make_hook(name: str): + def hook_fn(module, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + if name not in hessians: + hessians[name] = torch.zeros(x.shape[1], x.shape[1], device=x.device, dtype=torch.float32) + n_seen[name] = 0 + hessians[name].addmm_(x.t(), x) + n_seen[name] += x.shape[0] + return hook_fn + for name, module in model.named_modules(): + if isinstance(module, (nn.Linear, CastedLinear)): + hooks.append(module.register_forward_hook(make_hook(name))) + stream = TokenStream(train_pattern) + model.eval() + with torch.no_grad(): + for _ in range(n_samples): + tokens = stream.take(seq_len + 1).to(device=device, dtype=torch.int64) + x = tokens[:-1].unsqueeze(0) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + model.forward_logits(x) + for h in hooks: + h.remove() + for name in hessians: + hessians[name] /= max(n_seen[name], 1) + model.train() + return hessians +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale +def quantize_int5_per_row(t: Tensor, clip_range: int = 15) -> tuple[Tensor, Tensor]: + """int5 (signed, range [-15, 15]) per-row quant. Modeled on quantize_int6_per_row. + Returned int8 tensor holds values in [-15, 15] — leaves the high 3 bits as zeros, which + brotli compresses very efficiently. Dequant path is identical (q.float() * scale).""" + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale +def _classify_param_fine(name: str) -> str: + """Finer classifier than _classify_param, splitting attention/MLP banks for mixed-int policy. + Returns one of: embed, qo (Q+O bank), kv (K+V bank), mlp_up (mlp_fc), mlp_down (mlp_proj), + aux, attn_other, mlp_other, other. Categories `qo`/`kv` map to attention; `mlp_up`/`mlp_down` + map to MLP. The split lets us route mlp banks to int5 while keeping attn at int6.""" + if "tok_emb" in name or "lm_head" in name: + return "embed" + if "f1_corr_in" in name or "f1_corr_out" in name: + return "aux" + if "qo_bank" in name: + return "qo" + if "kv_bank" in name: + return "kv" + if "mlp_up_bank" in name: + return "mlp_up" + if "mlp_down_bank" in name: + return "mlp_down" + if ".mlp." in name: + return "mlp_other" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn_other" + return "other" +# salvage_v2 mixed-int policy (applied at the Tensor level via _classify_param_fine): +# - mlp_down_bank (mlp_proj, MOST quant-tolerant per 11-day collate) -> int5 (clip_range 15) +# - mlp_up_bank (mlp_fc, also tolerant) -> int5 +# - qo_bank, kv_bank (attention; LEAST quant-tolerant) -> int6 (clip_range 31) +# - tok_emb / lm_head (embed) -> int6 (matches seed; keeping +# attn/embed at int6 for +# quant safety on this +# first salvage attempt) +# Bytes savings: int5 keeps the int8 storage container (no bit-packing) but the high 3 bits are +# forced zero, giving brotli a compressible pattern. Combined with the byte-shuffle wrapper, +# expected savings vs uniform-int6+zstd is roughly the int5 bit ratio (5/6 = -17%) APPLIED only +# to MLP banks (which are ~60-65% of the model parameter mass for a 10L Rascal at mlp_mult=3.0). +# Net est: ~0.62 * 0.17 = ~10-11% blob shrink from int5 alone, plus brotli ~5-8% over zstd. +DEFAULT_INT5_CATS = {"mlp_down", "mlp_up"} +DEFAULT_INT6_CATS = {"qo", "kv", "attn_other", "mlp_other", "aux"} +def mixed_quantize_int6_gptq(state_dict: dict[str, Tensor], int6_cats: set[str], + hessians: dict[str, Tensor], + int5_cats: set[str] | None = None) -> tuple[dict, dict]: + """Mixed-int (int5/int6/int8) quant with GPTQ for matrix categories when Hessian available. + `int6_cats` and `int5_cats` use FINE-grained category names from `_classify_param_fine` + (qo, kv, mlp_up, mlp_down, attn_other, mlp_other, aux, embed). For backwards-compat with + the old uniform-int6 caller, the legacy coarse names {'mlp','attn','aux','embed'} are also + accepted in `int6_cats` and expand to their fine-grained children. `int5_cats` always uses + fine names. If `int5_cats` is None, defaults to DEFAULT_INT5_CATS (the salvage_v2 policy).""" + if int5_cats is None: + int5_cats = set(DEFAULT_INT5_CATS) + # Expand legacy coarse names so the existing call signature keeps working. + _LEGACY = { + "mlp": {"mlp_up", "mlp_down", "mlp_other"}, + "attn": {"qo", "kv", "attn_other"}, + "aux": {"aux"}, + "embed": {"embed"}, + } + expanded_int6: set[str] = set() + for c in int6_cats: + expanded_int6.update(_LEGACY.get(c, {c})) + # int5 categories take precedence over int6 for the same fine name. + expanded_int6 -= int5_cats + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + gptq_count, naive_count, int5_count = 0, 0, 0 + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param_fine(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int5_cats and t.ndim == 2: + module_name = name.rsplit(".weight", 1)[0] if name.endswith(".weight") else name + H = hessians.get(module_name) + if H is not None and H.shape[0] == t.shape[1]: + q, s = gptq_quantize_weight(t, H.cpu(), clip_range=15) + gptq_count += 1 + else: + q, s = quantize_int5_per_row(t) + naive_count += 1 + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int5"} + int5_count += 1 + elif cat in int5_cats and t.ndim >= 1: + t_2d = t.reshape(-1, t.shape[-1]) if t.ndim > 2 else t + q, s = quantize_int5_per_row(t_2d) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int5"} + naive_count += 1 + int5_count += 1 + elif cat in expanded_int6 and t.ndim == 2: + module_name = name.rsplit(".weight", 1)[0] if name.endswith(".weight") else name + H = hessians.get(module_name) + if H is not None and H.shape[0] == t.shape[1]: + q, s = gptq_quantize_weight(t, H.cpu()) + gptq_count += 1 + else: + q, s = quantize_int6_per_row(t) + naive_count += 1 + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + elif cat in expanded_int6 and t.ndim >= 1: + t_2d = t.reshape(-1, t.shape[-1]) if t.ndim > 2 else t + q, s = quantize_int6_per_row(t_2d) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + naive_count += 1 + else: + t_q = t.reshape(-1, t.shape[-1]) if t.ndim > 2 else t + q, s = quantize_float_tensor(t_q) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + print(f"gptq_quantize: {gptq_count} GPTQ layers, {naive_count} naive layers, {int5_count} int5 layers", flush=True) + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + val = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + val = (q.float() * float(s.item())).to(orig_dtype) + out[name] = val.reshape(orig.shape) if val.shape != orig.shape else val + return out + +# --- Data loading --- + +SHARD_HEADER_DTYPE = np.dtype(" dict[str, int]: + header = np.fromfile(file, dtype=SHARD_HEADER_DTYPE, count=SHARD_HEADER_WORDS) + if header.size != SHARD_HEADER_WORDS or int(header[0]) != SHARD_MAGIC or int(header[1]) != SHARD_VERSION: + raise ValueError(f"Unexpected shard header for {file}") + return {"num_tokens": int(header[2])} + +def load_data_shard(file: Path) -> Tensor: + header = read_data_shard_header(file) + num_tokens = header["num_tokens"] + expected_size = SHARD_HEADER_BYTES + num_tokens * SHARD_TOKEN_DTYPE.itemsize + if file.stat().st_size != expected_size: + raise ValueError(f"Shard size mismatch for {file}: expected {expected_size} bytes") + tokens_np = np.fromfile(file, dtype=SHARD_TOKEN_DTYPE, count=num_tokens, offset=SHARD_HEADER_BYTES) + if tokens_np.size != num_tokens: + raise ValueError(f"Short read for {file}") + return torch.from_numpy(tokens_np.astype(np.uint16, copy=False)) + +def choose_coprime_stride(modulus: int, salt: int) -> int: + if modulus <= 1: + return 1 + candidate = abs(salt) % modulus + if candidate == 0: + candidate = 1 + while math.gcd(candidate, modulus) != 1: + candidate += 1 + if candidate >= modulus: + candidate = 1 + return candidate + +class TokenStream: + def __init__(self, pattern: str): + self.files = [Path(p) for p in sorted(glob.glob(pattern))] + if not self.files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + self.file_idx = 0 + self.tokens = load_data_shard(self.files[0]) + self.pos = 0 + def _advance_file(self) -> None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def describe(self) -> str: + return f"loader:sequential shards:{len(self.stream.files)}" + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +class CoprimeDistributedTokenLoader: + """Shard-aware block sampler with deterministic coprime walks.""" + def __init__( + self, + pattern: str, + rank: int, + world_size: int, + device: torch.device, + seq_len: int, + seed: int, + max_loaded_shards: int, + shards_per_batch: int, + shard_hold_steps: int, + ): + self.rank = rank + self.world_size = world_size + self.device = device + self.seq_len = seq_len + self.seed = seed + self.token_offsets = torch.arange(seq_len + 1, dtype=torch.int64) + self.cache: OrderedDict[Path, Tensor] = OrderedDict() + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + self.shards: list[dict[str, int | Path]] = [] + for shard_idx, file in enumerate(files): + header = read_data_shard_header(file) + num_blocks = (header["num_tokens"] - 1) // seq_len + if num_blocks <= 0: + continue + self.shards.append( + { + "file": file, + "num_blocks": num_blocks, + "offset": (seed * 131 + shard_idx * 17) % num_blocks, + "stride": choose_coprime_stride(num_blocks, seed * 29 + shard_idx * 7 + 1), + } + ) + if not self.shards: + raise ValueError(f"No usable shards found for seq_len={seq_len}") + self.num_shards = len(self.shards) + self.max_loaded_shards = max(1, min(max_loaded_shards, self.num_shards)) + self.shards_per_batch = max(1, min(shards_per_batch, self.num_shards)) + self.shard_hold_steps = max(1, shard_hold_steps) + self.batch_shard_stride = choose_coprime_stride(self.num_shards, seed * 41 + 3) + self.batch_idx = 0 + self.shard_visits = [0 for _ in range(self.num_shards)] + def _get_tokens(self, file: Path) -> Tensor: + cached = self.cache.get(file) + if cached is not None: + self.cache.move_to_end(file) + return cached + # CPU advanced indexing is not implemented for uint16, so cache coprime-loader + # shards in int32 and cast to int64 only after batch assembly. + tokens = load_data_shard(file).to(dtype=torch.int32) + if len(self.cache) >= self.max_loaded_shards: + self.cache.popitem(last=False) + self.cache[file] = tokens + return tokens + def _sample_sequences(self, shard_idx: int, count: int) -> Tensor: + shard = self.shards[shard_idx] + num_blocks = int(shard["num_blocks"]) + offset = int(shard["offset"]) + stride = int(shard["stride"]) + visits = self.shard_visits[shard_idx] + block_ids = ( + offset + + (visits + torch.arange(count, dtype=torch.int64)) * stride + ) % num_blocks + self.shard_visits[shard_idx] += count + token_starts = block_ids * self.seq_len + gather_idx = token_starts.unsqueeze(1) + self.token_offsets.unsqueeze(0) + tokens = self._get_tokens(shard["file"]) + return tokens[gather_idx] + def describe(self) -> str: + total_blocks = sum(int(shard["num_blocks"]) for shard in self.shards) + return ( + f"loader:coprime shards:{self.num_shards} blocks:{total_blocks} " + f"seq_len:{self.seq_len} shards_per_batch:{self.shards_per_batch} " + f"cache:{self.max_loaded_shards} batch_stride:{self.batch_shard_stride} " + f"hold_steps:{self.shard_hold_steps}" + ) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + if seq_len != self.seq_len: + raise ValueError(f"Coprime loader was built for seq_len={self.seq_len}, got {seq_len}") + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + if local_tokens % seq_len != 0: + raise ValueError( + f"TRAIN_BATCH_TOKENS={global_tokens} does not divide into full local sequences " + f"for WORLD_SIZE={self.world_size}, GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_seqs = local_tokens // seq_len + active_shards = min(self.shards_per_batch, self.num_shards, local_seqs) + if active_shards <= 0: + raise ValueError(f"No active shards available for local_seqs={local_seqs}") + seqs_per_shard = local_seqs // active_shards + seq_remainder = local_seqs % active_shards + hold_idx = self.batch_idx // self.shard_hold_steps + shard_start = ((hold_idx * self.world_size) + self.rank) * self.batch_shard_stride + chunks: list[Tensor] = [] + for shard_slot in range(active_shards): + count = seqs_per_shard + (1 if shard_slot < seq_remainder else 0) + if count <= 0: + continue + shard_idx = (shard_start + shard_slot * self.batch_shard_stride) % self.num_shards + chunks.append(self._sample_sequences(shard_idx, count)) + self.batch_idx += 1 + local = chunks[0] if len(chunks) == 1 else torch.cat(chunks, dim=0) + local = local.to(dtype=torch.int64) + x = local[:, :-1] + y = local[:, 1:] + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +def build_train_loader(args: Hyperparameters, rank: int, world_size: int, device: torch.device): + if args.loader_mode == "sequential": + return DistributedTokenLoader(args.train_files, rank, world_size, device) + if args.loader_mode == "coprime": + return CoprimeDistributedTokenLoader( + args.train_files, + rank, + world_size, + device, + seq_len=args.train_seq_len, + seed=args.seed, + max_loaded_shards=args.coprime_max_loaded_shards, + shards_per_batch=args.coprime_shards_per_batch, + shard_hold_steps=args.coprime_shard_hold_steps, + ) + raise ValueError(f"Unknown LOADER_MODE={args.loader_mode!r}") + +# --- Transformer modules --- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + # No CastedLinear -- weights come from banks + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 # set by GPT.__init__ for partial RoPE + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False # set by GPT.__init__ for deep layers only + # Gated attention and value residual (non-banked small params) + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + self.value_residual = value_residual + if value_residual: + self.vrl_alpha = nn.Parameter(torch.zeros(1, dtype=torch.float32)) # sigmoid gate (PR #569 style) + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + """Efficient XSA: subtract self-value projection via GQA-aware reshape (no repeat_interleave). + y: [B, T, H, D], v: [B, T, Hkv, D]. H must be divisible by Hkv.""" + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) # [B, T, Hkv, group, D] + vn = F.normalize(v, dim=-1).unsqueeze(-2) # [B, T, Hkv, 1, D] -- broadcast ready + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + bsz, seqlen, dim = x.shape + q = F.linear(x, q_w.to(x.dtype)).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = F.linear(x, k_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = F.linear(x, v_w.to(x.dtype)) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v if self.value_residual else None + if self.value_residual and v0 is not None: + alpha = torch.sigmoid(self.vrl_alpha.to(dtype=v.dtype)) + v = v + alpha * v0 # sigmoid-gated residual (PR #569 style) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if flash_attn_3_func is not None: + q_attn, k_attn, v_attn = q, k, v + if q_attn.dtype not in (torch.float16, torch.bfloat16): + q_attn = q_attn.to(torch.bfloat16) + k_attn = k_attn.to(torch.bfloat16) + v_attn = v_attn.to(torch.bfloat16) + y = flash_attn_3_func(q_attn, k_attn, v_attn, causal=True) + else: + qh = q.transpose(1, 2) + kh = k.transpose(1, 2) + vh = v.transpose(1, 2) + if self.num_heads != self.num_kv_heads: + repeat = self.num_heads // self.num_kv_heads + kh = kh.repeat_interleave(repeat, dim=1) + vh = vh.repeat_interleave(repeat, dim=1) + y = F.scaled_dot_product_attention(qh, kh, vh, is_causal=True).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + # gate shape: (bsz, seqlen, num_heads) -> (bsz, seqlen, num_heads, 1) for B,T,H,D layout + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) + y = y * gate + y = y.reshape(bsz, seqlen, dim) + return F.linear(y, out_w.to(x.dtype)), raw_v + +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int, trigram: bool = False): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self._trigram = trigram + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def trigram_hash(self, tokens: Tensor) -> Tensor: + """Hash (t-2, t-1, t) trigrams into same embedding table. Zero extra params.""" + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., :2] = mod + out[..., 2:] = (36313 * t[..., 2:] ^ 27191 * t[..., 1:-1] ^ 51497 * t[..., :-2]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self._trigram: + h = h + self.embed(self.trigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + # No CastedLinear -- weights come from banks + self.kernel_mode = os.environ.get("MLP_KERNEL_MODE", "").strip().lower() + def forward(self, x: Tensor, up_w: Tensor, down_w: Tensor) -> Tensor: + x = F.linear(x, up_w.to(x.dtype)) + x = leaky_relu_sq(x, kernel_mode=self.kernel_mode) + return F.linear(x, down_w.to(x.dtype)) + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + layer_idx: int = 0, + ln_scale: bool = False, + dtg: bool = False, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=gated_attention, value_residual=value_residual) + self.mlp = MLP(dim, mlp_mult) + attn_scale_init = float(os.environ.get("ATTN_SCALE_INIT", "1.0")) + mlp_scale_init = float(os.environ.get("MLP_SCALE_INIT", "1.0")) + resid_mix_x_init = float(os.environ.get("RESID_MIX_X_INIT", "1.0")) + resid_mix_x0_init = float(os.environ.get("RESID_MIX_X0_INIT", "0.0")) + self.attn_scale = nn.Parameter(torch.full((dim,), attn_scale_init, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.full((dim,), mlp_scale_init, dtype=torch.float32)) + self.resid_mix = nn.Parameter( + torch.stack( + ( + torch.full((dim,), resid_mix_x_init, dtype=torch.float32), + torch.full((dim,), resid_mix_x0_init, dtype=torch.float32), + ) + ) + ) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, up_w: Tensor, down_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out, raw_v = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, q_w, k_w, v_w, out_w, v_embed=v_embed, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor, up_w, down_w) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out, raw_v + +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) # kv_dim for value projection + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.value_residual = value_residual + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim, trigram=bool(int(os.environ.get("TRIGRAM", "0")))) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + # Parameter banks: contiguous 3D tensors for batched optimizer + head_dim = model_dim // num_heads + kv_dim = num_kv_heads * head_dim + mlp_dim = int(mlp_mult * model_dim) + self.num_layers = num_layers + self.qo_bank = nn.Parameter(torch.empty(2 * num_layers, model_dim, model_dim)) + self.kv_bank = nn.Parameter(torch.empty(2 * num_layers, kv_dim, model_dim)) + self.mlp_up_bank = nn.Parameter(torch.empty(num_layers, mlp_dim, model_dim)) + self.mlp_down_bank = nn.Parameter(torch.empty(num_layers, model_dim, mlp_dim)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + layer_idx=i, + ln_scale=ln_scale, + dtg=dtg, + gated_attention=gated_attention, + value_residual=value_residual, + ) + for i in range(num_layers) + ] + ) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim_ve = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim_ve) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() # keep empty for compat + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + n = self.num_layers + proj_scale = 1.0 / math.sqrt(2 * n) + # Init banks: orthogonal, with proj layers scaled down and out/down zero-init + for i in range(n): + nn.init.orthogonal_(self.qo_bank.data[i], gain=1.0) # Q + nn.init.zeros_(self.qo_bank.data[n + i]) # Out (zero init) + nn.init.orthogonal_(self.kv_bank.data[i], gain=1.0) # K + nn.init.orthogonal_(self.kv_bank.data[n + i], gain=1.0) # V + nn.init.orthogonal_(self.mlp_up_bank.data[i], gain=1.0) # MLP up + nn.init.zeros_(self.mlp_down_bank.data[i]) # MLP down (zero init) + # Scale proj layers (out_proj and mlp_down are "proj" layers) + self.qo_bank.data[n + i].mul_(proj_scale) + self.mlp_down_bank.data[i].mul_(proj_scale) + # Init remaining nn.Linear modules (bigram proj, mtp heads, lm_head) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + """Get value embedding for a specific layer using shared table + per-layer scale.""" + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if hasattr(self, '_ngram_tracker') and self._ngram_tracker is not None and self.training: + per_tok_loss = F.cross_entropy(logits.float(), targets, reduction="none") + weights = self._ngram_tracker.get_weights(input_ids, target_ids) + main_loss = (per_tok_loss * weights).mean() + else: + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + def forward_logits(self, input_ids: Tensor) -> Tensor: + """Return logits (bsz, seq_len, vocab) without computing loss.""" + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + +# --- N-gram bulk update and hashed n-gram sliding eval --- + +def _ngram_bulk_update(val_np, start, end, ctx_tables, full_tables, + min_order, max_order, primes, mask): + """Bulk update n-gram tables with a contiguous range of tokens. + All ranks call this with the SAME token range -> identical tables everywhere.""" + t = val_np[start:end].astype(np.uint64) + n = len(t) + for order in range(min_order, max_order + 1): + if n < order: + continue + ctx_width = order - 1 + ctx_hash = np.zeros(n - order + 1, dtype=np.uint64) + for k in range(ctx_width): + ctx_hash ^= t[k:n - order + 1 + k] * primes[k % len(primes)] + ctx_key = (ctx_hash & mask).astype(np.int64) + tgt = t[order - 1:] + full_key = ((ctx_hash ^ (tgt * primes[ctx_width % len(primes)])) & mask).astype(np.int64) + ctx_tables[order] += np.bincount(ctx_key, minlength=len(ctx_tables[order])).astype(np.uint32) + full_tables[order] += np.bincount(full_key, minlength=len(full_tables[order])).astype(np.uint32) + +def eval_val_sliding_hashed_ngram( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + order: int, + alpha: float, + min_count: int, + buckets: int, + max_seconds: float = 0.0, + batch_seqs: int = 128, + eval_seq_len: int | None = None, +) -> tuple[float, float, float]: + """Score-first sliding eval with chunk-based SHARED n-gram tables + cubric. + + Key design: all ranks share identical n-gram tables via bulk chunk updates. + Each chunk's windows are distributed across ranks for scoring, then ALL ranks + update tables with the same contiguous token range. Every rank sees the full + n-gram picture (not 1/world_size like per-segment updates). + + Legal: entire chunk scored before its tokens update the tables. + """ + min_order = max(args.ngram_eval_min_order, 2) + max_order = max(order, min_order) + adaptive = args.ngram_eval_adaptive + alpha_min = args.ngram_eval_alpha_min + alpha_max = args.ngram_eval_alpha_max + ent_center = args.ngram_eval_entropy_center + ent_scale = args.ngram_eval_entropy_scale + + # Parse fixed per-order multipliers (PR #809 style) + _fixed_order_mults = None + if args.ngram_order_mults_str: + _fixed_order_mults = np.array([float(x) for x in args.ngram_order_mults_str.split(",")], dtype=np.float64) + + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + + # Build all windows and total scored tokens + all_window_starts = [ws for ws in range(0, total_tokens, stride) if min(ws + seq_len, total_tokens) - ws >= 1] + total_scored_tokens = 0.0 + for ws in all_window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + total_scored_tokens += float(max(wlen - s, 0)) + + # Group windows into chunks by scored position -- all ranks share this grouping + chunk_tokens = int(os.environ.get("NGRAM_CHUNK_TOKENS", "1048576")) # 1M default + num_chunks = (total_tokens + chunk_tokens - 1) // chunk_tokens + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in all_window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // chunk_tokens, num_chunks - 1) + chunk_windows[ci].append(ws) + + val_np = val_tokens.numpy() + ctx_tables = {n: np.zeros((buckets,), dtype=np.uint32) for n in range(min_order, max_order + 1)} + full_tables = {n: np.zeros((buckets,), dtype=np.uint32) for n in range(min_order, max_order + 1)} + mask = np.uint64(buckets - 1) + primes = np.array( + [np.uint64(36313), np.uint64(27191), np.uint64(51647), np.uint64(81929), + np.uint64(131071), np.uint64(174763), np.uint64(233017)], + dtype=np.uint64, + ) + + loss_sum = 0.0 + token_count = 0.0 + byte_count = 0.0 + + # Cubric 3D: per (order x entropy_bin x count_bin) adaptive alpha scaling + _NUM_ENT_BINS = 3 # low / mid / high entropy + _NUM_CNT_BINS = 3 # low / mid / high count + _ENT_EDGES = np.array([ent_center - 1.0, ent_center + 1.0]) # [2.0, 4.0] for center=3.0 + _CNT_EDGES = np.array([5.0, 50.0]) # low=<5, mid=5-50, high=>50 context count + _TOTAL_CELLS = _NUM_ENT_BINS * _NUM_CNT_BINS # 9 cells per order = 54 total + _cc = getattr(args, 'cubric_cadence', 0); _con = _cc > 0; _cfired = 0 + if _con: + # Warm-start: proven converged values from 4+ runs (orders 2-7) + # All 9 cells per order get the same warm-start, 3D cubric refines from there + _WARM = {2: 0.45, 3: 0.30, 4: 0.45, 5: 1.88, 6: 2.00, 7: 2.00, 8: 2.00, 9: 2.00} + _c_alpha_mult = {n: [_WARM.get(n, 1.0)] * _TOTAL_CELLS for n in range(min_order, max_order + 1)} + _c_hits = {n: [0] * _TOTAL_CELLS for n in range(min_order, max_order + 1)} + _c_beats = {n: [0] * _TOTAL_CELLS for n in range(min_order, max_order + 1)} + + base_model.eval() + compiled_logits = maybe_compile( + base_model.forward_logits, + enabled=args.compile_enabled, + fullgraph=False, + ) + t0 = time.perf_counter() + deadline = (t0 + max_seconds) if max_seconds > 0.0 else None + cutoff_hit = False + + if rank == 0: + print(f"ngram_eval:chunks={num_chunks} chunk_tokens={chunk_tokens} " + f"windows={len(all_window_starts)} shared_tables=True", flush=True) + + with torch.inference_mode(): + for ci in range(num_chunks): + if deadline is not None and time.perf_counter() >= deadline: + cutoff_hit = True + break + + windows = chunk_windows[ci] + if not windows: + continue + + # Distribute this chunk's windows across ranks + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + # --- Phase 1: SCORE this chunk's windows --- + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + logits_f = logits.float() + nll = F.cross_entropy( + logits_f.reshape(-1, logits_f.size(-1)), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + seg_len = wlen - s + if seg_len <= 0: + continue + + seg_nll = nll[i, s:wlen].to(torch.float64).cpu().numpy() + seg_model_p = np.exp(-seg_nll) + + if adaptive: + log_probs = F.log_softmax(logits_f[i, s:wlen], dim=-1) + probs_a = log_probs.exp() + entropy = -(probs_a * log_probs).sum(dim=-1).cpu().numpy() + sig = 1.0 / (1.0 + np.exp(-ent_scale * (entropy - ent_center))) + per_token_alpha = alpha_min + (alpha_max - alpha_min) * sig + # Bin entropy for 2D cubric: 0=low, 1=mid, 2=high + _ent_bins = np.digitize(entropy, _ENT_EDGES).astype(np.int32) + else: + per_token_alpha = np.full(seg_len, alpha) + _ent_bins = np.ones(seg_len, dtype=np.int32) # all mid + + global_j = np.arange(ws + s + 1, ws + wlen + 1, dtype=np.int64) + p_ng = np.zeros(seg_len, dtype=np.float64) + ng_matched = np.zeros(seg_len, dtype=np.bool_) + _ng_ord = np.zeros(seg_len, dtype=np.int32) + _ng_ctx_count = np.zeros(seg_len, dtype=np.float64) + tgt_np = val_np[global_j].astype(np.uint64) + + for n in range(max_order, min_order - 1, -1): + ctx_width = n - 1 + valid = (global_j >= ctx_width) & (~ng_matched) + if not valid.any(): + continue + v_idx = np.nonzero(valid)[0] + jv = global_j[v_idx] + ctx_hash = np.zeros(len(jv), dtype=np.uint64) + for k in range(ctx_width): + tok = val_np[jv - (ctx_width - k)].astype(np.uint64) + ctx_hash ^= tok * primes[k % len(primes)] + ctx_key = (ctx_hash & mask).astype(np.int64) + full_key = ((ctx_hash ^ (tgt_np[v_idx] * primes[ctx_width % len(primes)])) & mask).astype(np.int64) + ctx_counts = ctx_tables[n][ctx_key].astype(np.float64) + full_counts = full_tables[n][full_key].astype(np.float64) + has_data = ctx_counts >= float(min_count) + if has_data.any(): + p = np.minimum(full_counts, ctx_counts) / np.maximum(ctx_counts, 1.0) + p = np.clip(p, 0.0, 1.0) + hit_idx = v_idx[has_data] + p_ng[hit_idx] = p[has_data] + ng_matched[hit_idx] = True + _ng_ord[hit_idx] = n + _ng_ctx_count[hit_idx] = ctx_counts[has_data] + + # Mix where n-gram matched (PR #809 style or cubric 3D fallback) + if ng_matched.any(): + m_idx = np.nonzero(ng_matched)[0] + # Per-order entropy center shift (PR #809) + if adaptive and args.ngram_entropy_shift: + matched_ords = _ng_ord[m_idx].astype(np.float64) + shifted_centers = ent_center - 0.25 * (matched_ords - float(min_order)) + shifted_sig = 1.0 / (1.0 + np.exp(-ent_scale * (entropy[m_idx] - shifted_centers))) + per_token_alpha[m_idx] = alpha_min + (alpha_max - alpha_min) * shifted_sig + if _fixed_order_mults is not None: + # PR #809 fixed order multipliers (replaces cubric) + a = per_token_alpha[m_idx].copy() + mult_indices = _ng_ord[m_idx] - min_order + mult_indices = np.clip(mult_indices, 0, len(_fixed_order_mults) - 1) + a *= _fixed_order_mults[mult_indices] + np.clip(a, 0.0, 0.95, out=a) + elif _con: + a = per_token_alpha[m_idx].copy() + m_ent_bins = _ent_bins[m_idx] + m_cnt_bins = np.digitize(_ng_ctx_count[m_idx], _CNT_EDGES).astype(np.int32) + for n in range(min_order, max_order + 1): + om = _ng_ord[m_idx] == n + if not om.any(): + continue + for eb in range(_NUM_ENT_BINS): + for cb in range(_NUM_CNT_BINS): + cell = eb * _NUM_CNT_BINS + cb + mask_ecb = om & (m_ent_bins == eb) & (m_cnt_bins == cb) + if mask_ecb.any(): + _c_hits[n][cell] += int(mask_ecb.sum()) + _c_beats[n][cell] += int((p_ng[m_idx[mask_ecb]] > seg_model_p[m_idx[mask_ecb]]).sum()) + a[mask_ecb] *= _c_alpha_mult[n][cell] + np.clip(a, 0.0, 0.95, out=a) + else: + a = per_token_alpha[m_idx] + seg_model_p[m_idx] = (1.0 - a) * seg_model_p[m_idx] + a * p_ng[m_idx] + + seg_nll = -np.log(np.clip(seg_model_p, 1e-12, 1.0)) + loss_sum += float(seg_nll.sum()) + token_count += float(seg_len) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += float(tb.sum().item()) + + # --- Phase 2: SHARED UPDATE -- all ranks update with same chunk tokens --- + chunk_start = ci * chunk_tokens + chunk_end = min((ci + 1) * chunk_tokens, total_tokens) + _ngram_bulk_update(val_np, chunk_start, chunk_end + 1, + ctx_tables, full_tables, min_order, max_order, + primes, mask) + + # Cubric 2D c-step: adapt per (order x entropy_bin) + if _con: + # Collect all (order, ent_bin, cnt_bin) cells with enough data + all_rates = [] + for n in range(min_order, max_order + 1): + for cell in range(_TOTAL_CELLS): + if _c_hits[n][cell] >= 8: + all_rates.append(_c_beats[n][cell] / _c_hits[n][cell]) + if len(all_rates) >= 4: + avg_rate = sum(all_rates) / len(all_rates) + for n in range(min_order, max_order + 1): + for cell in range(_TOTAL_CELLS): + if _c_hits[n][cell] >= 8: + rate = _c_beats[n][cell] / _c_hits[n][cell] + if rate > avg_rate + 0.05: + _c_alpha_mult[n][cell] = min(_c_alpha_mult[n][cell] * 1.03, 2.0) + elif rate < avg_rate - 0.05: + _c_alpha_mult[n][cell] = max(_c_alpha_mult[n][cell] * 0.97, 0.3) + _cfired += 1 + if rank == 0 and _cfired % 8 == 0: + parts = [] + for n in range(min_order, max_order + 1): + m = _c_alpha_mult[n] + avg_m = sum(m) / len(m) + parts.append(f"o{n}:avg={avg_m:.2f}") + print(f"cubric3d:step={_cfired} {' '.join(parts)}", flush=True) + _c_hits = {n: [0] * _TOTAL_CELLS for n in range(min_order, max_order + 1)} + _c_beats = {n: [0] * _TOTAL_CELLS for n in range(min_order, max_order + 1)} + + # Progress + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1 or ci < 3): + elapsed = time.perf_counter() - t0 + cur_bpb = (loss_sum / max(token_count, 1.0)) / math.log(2.0) * (token_count / max(byte_count, 1.0)) if token_count > 0 else 0.0 + print( + f"ngram_eval:chunk [{ci+1}/{num_chunks}] bpb={cur_bpb:.6f} t={elapsed:.0f}s", + flush=True, + ) + + # All-reduce across ranks + _loss = torch.tensor(loss_sum, device=device, dtype=torch.float64) + _toks = torch.tensor(token_count, device=device, dtype=torch.float64) + _bytes = torch.tensor(byte_count, device=device, dtype=torch.float64) + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(_loss, op=dist.ReduceOp.SUM) + dist.all_reduce(_toks, op=dist.ReduceOp.SUM) + dist.all_reduce(_bytes, op=dist.ReduceOp.SUM) + loss_sum = _loss.item() + token_count = _toks.item() + byte_count = _bytes.item() + + coverage = token_count / max(total_scored_tokens, 1.0) + if cutoff_hit: + elapsed = time.perf_counter() - t0 + print( + f"ngram_eval:cutoff max_seconds={max_seconds:.1f} " + f"coverage={coverage*100:.2f}% elapsed={elapsed:.0f}s", + flush=True, + ) + + if _con and rank == 0: + print(f"cubric3d:final c_steps={_cfired} cells={_TOTAL_CELLS}x{max_order-min_order+1}={_TOTAL_CELLS*(max_order-min_order+1)}", flush=True) + for n in range(min_order, max_order + 1): + m = _c_alpha_mult[n] + row = " ".join(f"{m[cell]:.2f}" for cell in range(_TOTAL_CELLS)) + print(f" o{n}: [{row}]", flush=True) + val_loss = loss_sum / max(token_count, 1.0) + val_bpb = val_loss / math.log(2.0) * (token_count / max(byte_count, 1.0)) + base_model.train() + return val_loss, val_bpb, coverage + +# --- Sliding window evaluation --- + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = maybe_compile( + base_model.forward_logits, + enabled=args.compile_enabled, + fullgraph=args.compile_fullgraph, + ) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte + + +# --- Training --- + +def main() -> None: + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + # zeropower_via_newtonschulz5 runs eagerly with bmm -- do NOT compile + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size != 8: + raise ValueError( + f"Rascal 4k 12L brotli+mixed 8x requires WORLD_SIZE=8, got {world_size}. " + "Launch with: torchrun --standalone --nproc_per_node=8 4k_vocab_rascal_12l_brotli_mixed/train_gpt_4K_12L_brotli_mixed_8xgpu.py" + ) + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 100, console=False) + log0("condition_id:rascal_4k_12L_brotli_mixed_8x_seed444") + log0("run_label:salvage_v2 source_record:rascal_4k_8x_seed444_run20260427 axis:depth_12L+brotli+mixed_int") + log0("changed_fields:num_layers (11->12), compression (zstd->brotli+bshf), quant_policy (uniform_int6->mixed_int5_int6_int8)") + log0("expected_metric:final_sliding_window_exact comparator:0.8672_4k_8x_oversize_run prior_size:17766043_target:<16000000") + log0(f"condition:DATA_PATH={args.data_path}") + log0(f"condition:TOKENIZER_PATH={args.tokenizer_path}") + log0(f"condition:VOCAB_SIZE={args.vocab_size}") + log0(f"condition:SEED={args.seed}") + log0(f"condition:MAX_WALLCLOCK_SECONDS={args.max_wallclock_seconds}") + log0(f"condition:LOADER_MODE={args.loader_mode}") + log0(f"condition:COPRIME_MAX_LOADED_SHARDS={args.coprime_max_loaded_shards}") + log0(f"condition:COPRIME_SHARDS_PER_BATCH={args.coprime_shards_per_batch}") + log0(f"condition:COPRIME_SHARD_HOLD_STEPS={args.coprime_shard_hold_steps}") + log0(f"condition:SKIP_GPTQ={os.environ.get('SKIP_GPTQ', '1')}") + log0(f"condition:TRIGRAM={int(args.trigram_enabled)}") + log0(f"condition:NGRAM_EVAL_ORDER={args.ngram_eval_order}") + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + if args.ngram_eval_order >= 2: + log0(f"ngram_eval:order={args.ngram_eval_order} alpha={args.ngram_eval_alpha} min_count={args.ngram_eval_min_count} buckets={args.ngram_eval_buckets}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + CastedLinear._qat_enabled = args.qat_enabled + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + gated_attention=args.gated_attention, + value_residual=args.value_residual, + ).to(device).bfloat16() + # Banks stay FP32 (like CastedLinear weights), cast to BF16 in forward + base_model.qo_bank.data = base_model.qo_bank.data.float() + base_model.kv_bank.data = base_model.kv_bank.data.float() + base_model.mlp_up_bank.data = base_model.mlp_up_bank.data.float() + base_model.mlp_down_bank.data = base_model.mlp_down_bank.data.float() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + if args.complement_alpha > 0: + tracker = TrainNgramTracker(args.vocab_size, device, complement_alpha=args.complement_alpha) + base_model._ngram_tracker = tracker + log0(f"complementary_training:alpha={args.complement_alpha}") + else: + base_model._ngram_tracker = None + # No DDP -- Parallel Muon handles bank grad communication via reduce-scatter, + # and non-bank grads are manually all-reduced before Adam steps. + compiled_model = maybe_compile( + base_model, + enabled=args.compile_enabled, + fullgraph=args.compile_fullgraph, + mode=args.compile_mode, + ) + model = compiled_model + + # Optimizer split: + # - 4 parameter banks -> Muon (batched Newton-Schulz) + # - token embedding -> Adam + # - scalars/control tensors -> Adam + # - bigram proj, mtp heads, VE proj -> Adam (small matrix params not worth banking) + matrix_params = [ + base_model.qo_bank, base_model.kv_bank, + base_model.mlp_up_bank, base_model.mlp_down_bank, + ] + block_named_params = list(base_model.blocks.named_parameters()) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + scalar_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + scalar_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + # Non-bank params that need manual all-reduce (replicated across GPUs) + replicated_params = list(optimizer_tok.param_groups[0]["params"]) + for pg in optimizer_tok.param_groups[1:]: + replicated_params.extend(pg["params"]) + replicated_params.extend(scalar_params) + + optimizer_head = None + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + replicated_params.append(base_model.lm_head.weight) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if optimizer_head is not None: + optimizers.append(optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + compile_mode = args.compile_mode if args.compile_mode else "default" + log0( + f"compile:enabled={int(args.compile_enabled)} mode:{compile_mode} " + f"fullgraph={int(args.compile_fullgraph)}" + ) + log0(f"mlp_kernel_mode:{args.mlp_kernel_mode or 'eager'}") + log0( + f"scale_init:attn={args.attn_scale_init:.4f} mlp={args.mlp_scale_init:.4f} " + f"resid_mix=({args.resid_mix_x_init:.4f},{args.resid_mix_x0_init:.4f}) " + f"ln_scale={int(args.ln_scale)}" + ) + log0(f"seed:{args.seed}") + train_loader = build_train_loader(args, rank, world_size, device) + log0(train_loader.describe()) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + # GPTQ calibration reads training data — it must complete within the wallclock budget. + # We stop the training loop early (by GPTQ_RESERVE_MS) so GPTQ runs before the cap. + _skip_gptq = int(os.environ.get("SKIP_GPTQ", "1")) + _gptq_reserve_ms = float(os.environ.get("GPTQ_RESERVE_MS", "30000")) if (max_wallclock_ms is not None and not _skip_gptq) else 0.0 + effective_max_wallclock_ms = (max_wallclock_ms - _gptq_reserve_ms) if max_wallclock_ms is not None else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + # All-reduce all grads for warmup (simple, not optimized) + if distributed: + for p in base_model.parameters(): + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + train_loader = build_train_loader(args, rank, world_size, device) + log0(f"loader_reset:{train_loader.describe()}") + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + from collections import deque + lawa_queue: deque[dict[str, Tensor]] = deque(maxlen=args.lawa_k) + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms tok/s:{(step * args.train_batch_tokens) / max(training_time_ms / 1000.0, 1e-9):.0f}" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + if base_model._ngram_tracker is not None: + base_model._ngram_tracker.update(x, y) + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + # === 3-phase overlapped optimizer step === + # Phase 1: Launch async reduce-scatter for banks (biggest first) + optimizer_muon.launch_reduce_scatters() + # Phase 2: All-reduce non-bank grads + step Adam (while bank RS is in-flight) + if distributed: + for p in replicated_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + optimizer_tok.step() + optimizer_scalar.step() + if optimizer_head is not None: + optimizer_head.step() + # Phase 3: Wait for RS, local NS5, all-gather (banks processed last) + optimizer_muon.step() + zero_grad_all() + # EMA update + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + if args.lawa_enabled and step % args.lawa_freq == 0: + lawa_queue.append({name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()}) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms tok/s:{(step * args.train_batch_tokens) / max(approx_training_time_ms / 1000.0, 1e-9):.0f}" + ) + reached_cap = effective_max_wallclock_ms is not None and approx_training_time_ms >= effective_max_wallclock_ms + if distributed and effective_max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + # GPTQ calibration: reads training data — must complete within MAX_WALLCLOCK_SECONDS. + # Training loop stopped GPTQ_RESERVE_MS early so this runs inside the budget. + if _skip_gptq: + log0("gptq:SKIPPED (SKIP_GPTQ=1) — will use naive int6") + gptq_hessians: dict[str, Tensor] = {} + else: + log0("gptq:calibrating with training data...") + t_gptq = time.perf_counter() + gptq_hessians = gptq_calibrate(base_model, args.train_files, device, n_samples=256, seq_len=args.train_seq_len) + log0(f"gptq:calibrated {len(gptq_hessians)} layers in {time.perf_counter()-t_gptq:.1f}s") + # Apply weight averaging + if args.lawa_enabled and len(lawa_queue) > 1: + log0(f"lawa:applying LAWA averaging k={len(lawa_queue)}") + current_state = base_model.state_dict() + avg_state = {name: torch.zeros(t.shape, dtype=torch.float32, device='cpu') for name, t in current_state.items()} + for snap in lawa_queue: + for name in avg_state: + avg_state[name] += snap[name].float() + for name in avg_state: + avg_state[name] /= len(lawa_queue) + avg_state[name] = avg_state[name].to(dtype=current_state[name].dtype) + base_model.load_state_dict(avg_state, strict=True) + else: + log0("ema:applying EMA weights") + current_state = base_model.state_dict() + avg_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(avg_state, strict=True) + if args.post_ema_diagnostic: + torch.cuda.synchronize() + t_diag = time.perf_counter() + diag_val_loss, diag_val_bpb = eval_val( + args, compiled_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"DIAGNOSTIC post_ema val_loss:{diag_val_loss:.4f} val_bpb:{diag_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" + ) + else: + log0("diagnostic_eval:skipped POST_EMA_DIAGNOSTIC=0") + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k) + if excluded_mtp > 0: + log0(f"export_excluding_mtp_params:{excluded_mtp}") + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + # GPTQ quantization using Hessians collected from training data. + # salvage_v2 mixed-int policy: int5 for mlp_up_bank/mlp_down_bank (most quant-tolerant per + # 11-day collate), int6 for qo_bank/kv_bank/embed (attention + token embed kept at int6 for + # quant safety; matches seed for embed). See `mixed_quantize_int6_gptq` docstring. + quant_result, quant_meta = mixed_quantize_int6_gptq( + sd_cpu, + int6_cats={"qo", "kv", "attn_other", "mlp_other", "aux", "embed"}, + hessians=gptq_hessians, + int5_cats={"mlp_down", "mlp_up"}, + ) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = _compress_blob(quant_raw) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + log0(f"Serialized model mixed_int5_int6_int8+{_COMPRESSOR}: {quant_file_bytes} bytes") + log0(f"Total submission size mixed_int5_int6_int8+{_COMPRESSOR}: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(_decompress_blob(quant_blob_disk)), + map_location="cpu", + ) + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, rope_dims=args.rope_dims, ln_scale=args.ln_scale, + dtg=args.dtg_enabled, ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + gated_attention=args.gated_attention, value_residual=args.value_residual, + ).to(device).bfloat16() + eval_model.qo_bank.data = eval_model.qo_bank.data.float() + eval_model.kv_bank.data = eval_model.kv_bank.data.float() + eval_model.mlp_up_bank.data = eval_model.mlp_up_bank.data.float() + eval_model.mlp_down_bank.data = eval_model.mlp_down_bank.data.float() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, eval_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + del eval_model, deq_state, quant_state, sd_cpu + torch.cuda.empty_cache() + sw_seq_len = effective_eval_seq_len + if args.skip_final_eval: + log0("final_eval:skipped sliding/ngram by SKIP_FINAL_EVAL=1") + else: + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, base_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, base_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + if args.ngram_eval_order >= 2: + if distributed: + dist.barrier() + torch.cuda.synchronize() + t_ng = time.perf_counter() + ng_loss, ng_bpb, ng_coverage = eval_val_sliding_hashed_ngram( + args, + base_model, + rank, + world_size, + device, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + stride=args.eval_stride, + order=args.ngram_eval_order, + alpha=args.ngram_eval_alpha, + min_count=args.ngram_eval_min_count, + buckets=args.ngram_eval_buckets, + max_seconds=args.ngram_eval_max_seconds, + eval_seq_len=sw_seq_len, + ) + if rank == 0: + torch.cuda.synchronize() + ng_eval_ms = 1000.0 * (time.perf_counter() - t_ng) + if ng_coverage >= 0.999999: + log0( + f"final_sliding_window_ngram{args.ngram_eval_order} val_loss:{ng_loss:.4f} " + f"val_bpb:{ng_bpb:.4f} eval_time:{ng_eval_ms:.0f}ms" + ) + log0( + f"final_sliding_window_ngram{args.ngram_eval_order}_exact " + f"val_loss:{ng_loss:.8f} val_bpb:{ng_bpb:.8f}" + ) + else: + log0( + f"final_sliding_window_ngram{args.ngram_eval_order}_partial val_loss:{ng_loss:.4f} " + f"val_bpb:{ng_bpb:.4f} coverage:{ng_coverage:.4f} eval_time:{ng_eval_ms:.0f}ms" + ) + log0( + f"final_sliding_window_ngram{args.ngram_eval_order}_partial_exact " + f"val_loss:{ng_loss:.8f} val_bpb:{ng_bpb:.8f} coverage:{ng_coverage:.8f}" + ) + if distributed: + dist.barrier() + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() + +==================================================================================================== +condition_id:rascal_4k_12L_brotli_mixed_8x_seed444 +run_label:salvage_v2 source_record:rascal_4k_8x_seed444_run20260427 axis:depth_12L+brotli+mixed_int +changed_fields:num_layers (11->12), compression (zstd->brotli+bshf), quant_policy (uniform_int6->mixed_int5_int6_int8) +expected_metric:final_sliding_window_exact comparator:0.8672_4k_8x_oversize_run prior_size:17766043_target:<16000000 +condition:DATA_PATH=./data/datasets/fineweb10B_sp4096 +condition:TOKENIZER_PATH=./data/tokenizers/fineweb_4096_bpe.model +condition:VOCAB_SIZE=4096 +condition:SEED=42 +condition:MAX_WALLCLOCK_SECONDS=600.0 +condition:LOADER_MODE=coprime +condition:COPRIME_MAX_LOADED_SHARDS=143 +condition:COPRIME_SHARDS_PER_BATCH=1 +condition:COPRIME_SHARD_HOLD_STEPS=64 +condition:SKIP_GPTQ=1 +condition:TRIGRAM=0 +condition:NGRAM_EVAL_ORDER=0 +Running Python 3.12.13 | packaged by conda-forge | (main, Mar 5 2026, 16:50:00) [GCC 14.3.0] +Running PyTorch 2.11.0+cu130 +Mon Apr 27 03:41:14 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 580.95.05 Driver Version: 580.95.05 CUDA Version: 13.0 | ++-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:04:00.0 Off | 0 | +| N/A 41C P0 117W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:05:00.0 Off | 0 | +| N/A 36C P0 118W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:0A:00.0 Off | 0 | +| N/A 40C P0 119W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:0B:00.0 Off | 0 | +| N/A 36C P0 120W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:84:00.0 Off | 0 | +| N/A 39C P0 117W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:85:00.0 Off | 0 | +| N/A 34C P0 116W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:8A:00.0 Off | 0 | +| N/A 38C P0 121W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:8B:00.0 Off | 0 | +| N/A 35C P0 117W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| No running processes found | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_4096_bpe.model +train_loader:dataset:fineweb10B_sp4096 train_shards:143 +val_loader:shards pattern=./data/datasets/fineweb10B_sp4096/fineweb_val_*.bin tokens:45514752 +model_params:31321700 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_11 active_layers:[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11] +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +compile:enabled=1 mode:default fullgraph=1 +mlp_kernel_mode:eager +scale_init:attn=1.0000 mlp=1.0000 resid_mix=(1.0000,0.0000) ln_scale=1 +seed:42 +loader:coprime shards:143 blocks:6967965 seq_len:2048 shards_per_batch:1 cache:143 batch_stride:9 hold_steps:64 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +loader_reset:loader:coprime shards:143 blocks:6967965 seq_len:2048 shards_per_batch:1 cache:143 batch_stride:9 hold_steps:64 +step:0/20000 val_loss:8.3101 val_bpb:2.8334 train_time:0ms step_avg:0.01ms tok/s:0 +step:1/20000 train_loss:8.3103 train_time:356ms step_avg:355.82ms tok/s:2210168 +step:2/20000 train_loss:9.5390 train_time:407ms step_avg:203.42ms tok/s:3866108 +step:3/20000 train_loss:8.8566 train_time:508ms step_avg:169.26ms tok/s:4646338 +step:4/20000 train_loss:8.1436 train_time:610ms step_avg:152.38ms tok/s:5161001 +step:5/20000 train_loss:8.0018 train_time:711ms step_avg:142.29ms tok/s:5526982 +step:6/20000 train_loss:8.0293 train_time:813ms step_avg:135.49ms tok/s:5804497 +step:7/20000 train_loss:7.8916 train_time:915ms step_avg:130.71ms tok/s:6016846 +step:8/20000 train_loss:7.7694 train_time:1017ms step_avg:127.12ms tok/s:6186402 +step:9/20000 train_loss:7.5519 train_time:1119ms step_avg:124.29ms tok/s:6327601 +step:10/20000 train_loss:7.4294 train_time:1220ms step_avg:122.02ms tok/s:6445305 +step:500/20000 train_loss:3.1321 train_time:52909ms step_avg:105.82ms tok/s:7431948 +step:1000/20000 train_loss:2.9502 train_time:105802ms step_avg:105.80ms tok/s:7433032 +step:1500/20000 train_loss:2.8720 train_time:158682ms step_avg:105.79ms tok/s:7434016 +step:2000/20000 train_loss:2.8296 train_time:211642ms step_avg:105.82ms tok/s:7431713 +step:2500/20000 train_loss:2.7596 train_time:264621ms step_avg:105.85ms tok/s:7429786 +step:3000/20000 train_loss:2.6616 train_time:317371ms step_avg:105.79ms tok/s:7433875 +step:3500/20000 train_loss:2.7652 train_time:370413ms step_avg:105.83ms tok/s:7430917 +step:4000/20000 train_loss:2.7392 train_time:423442ms step_avg:105.86ms tok/s:7428947 +step:4000/20000 val_loss:2.6776 val_bpb:0.9130 train_time:423496ms step_avg:105.87ms tok/s:7428002 +step:4500/20000 train_loss:2.5911 train_time:476437ms step_avg:105.87ms tok/s:7427930 +swa:start step:5000 +step:5000/20000 train_loss:2.5953 train_time:529462ms step_avg:105.89ms tok/s:7426704 +late_qat:enabled step:5138 scale:0.1499 +step:5500/20000 train_loss:2.5660 train_time:583115ms step_avg:106.02ms tok/s:7417709 +step:5656/20000 val_loss:2.5812 val_bpb:0.8801 train_time:600149ms step_avg:106.11ms tok/s:7411597 +stopping_early: wallclock_cap train_time:600149ms step:5656/20000 +peak memory allocated: 25252 MiB reserved: 25750 MiB +gptq:SKIPPED (SKIP_GPTQ=1) — will use naive int6 +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:2.5786 val_bpb:0.8792 eval_time:1713ms +Serialized model: 119539322 bytes +Code size: 128464 bytes +Serialized model mixed_int5_int6_int8+brotli: 15511273 bytes +Total submission size mixed_int5_int6_int8+brotli: 15639737 bytes +final_int6_roundtrip val_loss:2.6638 val_bpb:0.9082 eval_time:5035ms +final_int6_roundtrip_exact val_loss:2.66376216 val_bpb:0.90824024 +final_sliding_window val_loss:2.5371 val_bpb:0.8650 stride:64 eval_time:65261ms +final_sliding_window_exact val_loss:2.53706469 val_bpb:0.86503709 diff --git a/records/track_10min_16mb/2026-04-27_Mikey/train_seed444.log b/records/track_10min_16mb/2026-04-27_Mikey/train_seed444.log new file mode 100644 index 0000000000..195a1048b3 --- /dev/null +++ b/records/track_10min_16mb/2026-04-27_Mikey/train_seed444.log @@ -0,0 +1,2801 @@ +from __future__ import annotations +import copy +import glob +import io +import math +import os +import random +import subprocess +import sys +import time +import uuid +from collections import OrderedDict +from pathlib import Path +import numpy as np +import sentencepiece as spm +import torch +import torch.distributed as dist +import torch.nn.functional as F +from torch import Tensor, nn +from torch.nn.parallel import DistributedDataParallel as DDP + +os.environ.setdefault("RUN_ID", "rascal_4k_12L_brotli_mixed_8x_seed444_" + time.strftime("%Y%m%d_%H%M%S")) + +try: + import triton + import triton.language as tl +except ImportError: + triton = None + tl = None +try: + from flash_attn_interface import flash_attn_func as flash_attn_3_func +except ImportError: + flash_attn_3_func = None +# Compression: brotli-11 + byte-shuffle is the salvage_v2 preferred path (PR #1493 recipe). +# Falls back to zstd then zlib so this file still runs if brotli isn't installed. +_brotli_module = None +_zstandard_module = None +_zlib_module = None +try: + import brotli as _brotli_module + _COMPRESSOR = "brotli" +except ImportError: + try: + import zstandard as _zstandard_module + _COMPRESSOR = "zstd" + import warnings + warnings.warn("brotli not found — falling back to zstd (~1MB+ larger). pip install brotli") + except ImportError: + import zlib as _zlib_module + import warnings + _COMPRESSOR = "zlib" + warnings.warn("brotli/zstandard not found — falling back to zlib. Artifact will be much larger! pip install brotli") +# Backwards-compat shims so any remaining `zstandard.*` or `_zlib_module.*` references still work. +if _zstandard_module is not None: + zstandard = _zstandard_module +if _zlib_module is None: + import zlib as _zlib_module # always available; used by zlib fallback path +# --- Byte-shuffle (de-interleave) wrapper from PR #1493: improves brotli ratio on quantized payloads. --- +_BSHF_MAGIC = b"BSHF" +def _byte_shuffle(data: bytes, stride: int = 2) -> bytes: + if stride <= 1 or len(data) < stride: + return data + src = np.frombuffer(data, dtype=np.uint8) + n = len(src) + out = np.empty(n, dtype=np.uint8) + dest_off = 0 + for pos in range(stride): + chunk = src[pos::stride] + out[dest_off:dest_off + len(chunk)] = chunk + dest_off += len(chunk) + return _BSHF_MAGIC + bytes([stride]) + out.tobytes() +def _byte_unshuffle(data: bytes) -> bytes: + if len(data) < 5 or data[:4] != _BSHF_MAGIC: + return data + stride = data[4] + if stride < 2: + return data[5:] + payload = np.frombuffer(data, dtype=np.uint8, offset=5) + n = len(payload) + out = np.empty(n, dtype=np.uint8) + src_off = 0 + for pos in range(stride): + chunk_len = n // stride + (1 if pos < n % stride else 0) + out[pos::stride][:chunk_len] = payload[src_off:src_off + chunk_len] + src_off += chunk_len + return out.tobytes() +def _compress_blob(raw: bytes) -> bytes: + if _COMPRESSOR == "brotli": + return _brotli_module.compress(_byte_shuffle(raw, stride=2), quality=11) + elif _COMPRESSOR == "zstd": + return _zstandard_module.ZstdCompressor(level=22).compress(raw) + else: + return _zlib_module.compress(raw, 9) +def _decompress_blob(blob: bytes) -> bytes: + if _COMPRESSOR == "brotli": + return _byte_unshuffle(_brotli_module.decompress(blob)) + elif _COMPRESSOR == "zstd": + return _zstandard_module.ZstdDecompressor().decompress(blob) + else: + return _zlib_module.decompress(blob) + +if os.environ.get("TORCHDYNAMO_SUPPRESS_ERRORS", "0") == "1": + import torch._dynamo + torch._dynamo.config.suppress_errors = True +class Hyperparameters: + data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp4096") + train_files = os.path.join(data_path, "fineweb_train_*.bin") + val_files = os.path.join(data_path, "fineweb_val_*.bin") + tokenizer_path = os.environ.get("TOKENIZER_PATH", "./data/tokenizers/fineweb_4096_bpe.model") + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + seed = int(os.environ.get("SEED", 444)) + val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + iterations = int(os.environ.get("ITERATIONS", 20000)) + warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 3500)) + warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786_432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) + vocab_size = int(os.environ.get("VOCAB_SIZE", 4096)) + num_layers = int(os.environ.get("NUM_LAYERS", 12)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + model_dim = int(os.environ.get("MODEL_DIM", 512)) + num_heads = int(os.environ.get("NUM_HEADS", 8)) + mlp_mult = float(os.environ.get("MLP_MULT", 3.0)) + tie_embeddings = bool(int(os.environ.get("TIE_EMBEDDINGS", "1"))) + rope_base = float(os.environ.get("ROPE_BASE", 10000.0)) + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 30.0)) + embed_lr = float(os.environ.get("EMBED_LR", 0.6)) + head_lr = float(os.environ.get("HEAD_LR", 0.008)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.035)) + tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.025)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.025)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.99)) + muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) + muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.92)) + muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 1500)) + beta1 = float(os.environ.get("BETA1", 0.9)) + beta2 = float(os.environ.get("BETA2", 0.95)) + adam_eps = float(os.environ.get("ADAM_EPS", 1e-8)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + mtp_num_heads = int(os.environ.get("MTP_NUM_HEADS", 0)) + mtp_loss_weight = float(os.environ.get("MTP_LOSS_WEIGHT", 0.2)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + swa_enabled = bool(int(os.environ.get("SWA_ENABLED", "1"))) + swa_every = int(os.environ.get("SWA_EVERY", 50)) + lawa_enabled = bool(int(os.environ.get("LAWA_ENABLED", "0"))) + lawa_k = int(os.environ.get("LAWA_K", 10)) + lawa_freq = int(os.environ.get("LAWA_FREQ", 100)) + muon_wd = float(os.environ.get("MUON_WD", 0.04)) + adam_wd = float(os.environ.get("ADAM_WD", 0.04)) + qat_enabled = bool(int(os.environ.get("QAT_ENABLED", "0"))) + bigram_vocab_size = int(os.environ.get("BIGRAM_VOCAB_SIZE", 2048)) + bigram_dim = int(os.environ.get("BIGRAM_DIM", 128)) + trigram_enabled = bool(int(os.environ.get("TRIGRAM", "0"))) # TrigramHash (off by default, risky) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 11)) # XSA on ALL layers (our novel contribution) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + dtg_enabled = bool(int(os.environ.get("DTG_ENABLED", "0"))) + late_qat_threshold = float(os.environ.get("LATE_QAT_THRESHOLD", 0.15)) + ve_enabled = bool(int(os.environ.get("VE_ENABLED", "1"))) + ve_dim = int(os.environ.get("VE_DIM", 128)) + ve_layers = os.environ.get("VE_LAYERS", "9,10") + gated_attention = bool(int(os.environ.get("GATED_ATTENTION", "0"))) + value_residual = bool(int(os.environ.get("VALUE_RESIDUAL", "0"))) # VRL with sigmoid gates (off by default, risky) + attn_scale_init = float(os.environ.get("ATTN_SCALE_INIT", 1.0)) + mlp_scale_init = float(os.environ.get("MLP_SCALE_INIT", 1.0)) + resid_mix_x_init = float(os.environ.get("RESID_MIX_X_INIT", 1.0)) + resid_mix_x0_init = float(os.environ.get("RESID_MIX_X0_INIT", 0.0)) + complement_alpha = float(os.environ.get("COMPLEMENT_ALPHA", "0")) + ngram_eval_order = int(os.environ.get("NGRAM_EVAL_ORDER", 0)) + ngram_eval_min_order = int(os.environ.get("NGRAM_EVAL_MIN_ORDER", 2)) + ngram_eval_alpha = float(os.environ.get("NGRAM_EVAL_ALPHA", 0.30)) + ngram_eval_adaptive = bool(int(os.environ.get("NGRAM_EVAL_ADAPTIVE", "1"))) + ngram_eval_alpha_min = float(os.environ.get("NGRAM_EVAL_ALPHA_MIN", 0.05)) + ngram_eval_alpha_max = float(os.environ.get("NGRAM_EVAL_ALPHA_MAX", 0.60)) + ngram_eval_entropy_center = float(os.environ.get("NGRAM_EVAL_ENTROPY_CENTER", 4.0)) + ngram_eval_entropy_scale = float(os.environ.get("NGRAM_EVAL_ENTROPY_SCALE", 2.0)) + ngram_eval_min_count = int(os.environ.get("NGRAM_EVAL_MIN_COUNT", 2)) + ngram_eval_buckets = int(os.environ.get("NGRAM_EVAL_BUCKETS", 4_194_304)) + ngram_eval_max_seconds = float(os.environ.get("NGRAM_EVAL_MAX_SECONDS", 0.0)) + ngram_entropy_shift = bool(int(os.environ.get("NGRAM_ENTROPY_SHIFT", "0"))) + ngram_order_mults_str = os.environ.get("NGRAM_ORDER_MULTS", "") + cubric_cadence = int(os.environ.get("CUBRIC_CADENCE", 0)) + skip_final_eval = bool(int(os.environ.get("SKIP_FINAL_EVAL", "0"))) + post_ema_diagnostic = bool(int(os.environ.get("POST_EMA_DIAGNOSTIC", "1"))) + compile_enabled = bool(int(os.environ.get("COMPILE_ENABLED", "1"))) + compile_mode = os.environ.get("COMPILE_MODE", "").strip() + compile_fullgraph = bool(int(os.environ.get("COMPILE_FULLGRAPH", "1"))) + mlp_kernel_mode = os.environ.get("MLP_KERNEL_MODE", "").strip().lower() + loader_mode = os.environ.get("LOADER_MODE", "coprime").strip().lower() + coprime_max_loaded_shards = int(os.environ.get("COPRIME_MAX_LOADED_SHARDS", 143)) + coprime_shards_per_batch = int(os.environ.get("COPRIME_SHARDS_PER_BATCH", 1)) + coprime_shard_hold_steps = int(os.environ.get("COPRIME_SHARD_HOLD_STEPS", 64)) + + +def maybe_compile(fn_or_module, *, enabled: bool, fullgraph: bool, mode: str = ""): + if not enabled: + return fn_or_module + kwargs = dict(dynamic=False, fullgraph=fullgraph) + if mode: + kwargs["mode"] = mode + return torch.compile(fn_or_module, **kwargs) + + +if triton is not None: + @triton.jit + def _leaky_relu_sq_forward_kernel(x_ptr, y_ptr, n_elements, BLOCK_SIZE: tl.constexpr): + pid = tl.program_id(0) + offsets = pid * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE) + mask = offsets < n_elements + x = tl.load(x_ptr + offsets, mask=mask, other=0.0).to(tl.float32) + a = tl.where(x >= 0, x, 0.5 * x) + y = a * a + tl.store(y_ptr + offsets, y, mask=mask) + + @triton.jit + def _leaky_relu_sq_backward_kernel(x_ptr, grad_out_ptr, grad_in_ptr, n_elements, BLOCK_SIZE: tl.constexpr): + pid = tl.program_id(0) + offsets = pid * BLOCK_SIZE + tl.arange(0, BLOCK_SIZE) + mask = offsets < n_elements + x = tl.load(x_ptr + offsets, mask=mask, other=0.0).to(tl.float32) + grad_out = tl.load(grad_out_ptr + offsets, mask=mask, other=0.0).to(tl.float32) + a = tl.where(x >= 0, x, 0.5 * x) + slope = tl.where(x >= 0, 1.0, 0.5) + grad_in = grad_out * (2.0 * a * slope) + tl.store(grad_in_ptr + offsets, grad_in, mask=mask) + + +class TritonLeakyReluSqFn(torch.autograd.Function): + @staticmethod + def forward(ctx, x: Tensor) -> Tensor: + if triton is None or not x.is_cuda: + a = F.leaky_relu(x, negative_slope=0.5) + ctx.save_for_backward(x) + return a.square() + x_contig = x.contiguous() + y = torch.empty_like(x_contig) + n_elements = x_contig.numel() + grid = lambda meta: (triton.cdiv(n_elements, meta["BLOCK_SIZE"]),) + _leaky_relu_sq_forward_kernel[grid](x_contig, y, n_elements, BLOCK_SIZE=1024) + ctx.save_for_backward(x_contig) + return y + + @staticmethod + def backward(ctx, grad_out: Tensor) -> tuple[Tensor]: + (x,) = ctx.saved_tensors + if triton is None or not grad_out.is_cuda: + a = F.leaky_relu(x, negative_slope=0.5) + slope = torch.where(x >= 0, torch.ones_like(x), torch.full_like(x, 0.5)) + return (grad_out * (2.0 * a * slope),) + grad_out_contig = grad_out.contiguous() + grad_in = torch.empty_like(grad_out_contig) + n_elements = grad_out_contig.numel() + grid = lambda meta: (triton.cdiv(n_elements, meta["BLOCK_SIZE"]),) + _leaky_relu_sq_backward_kernel[grid](x, grad_out_contig, grad_in, n_elements, BLOCK_SIZE=1024) + return (grad_in,) + + +def leaky_relu_sq(x: Tensor, kernel_mode: str = "") -> Tensor: + if kernel_mode == "triton_act": + return TritonLeakyReluSqFn.apply(x) + a = F.leaky_relu(x, negative_slope=0.5) + return a.square() + +class TrainNgramTracker: + """Complementary training: track bigram stats, downweight tokens n-grams can predict.""" + def __init__(self, vocab_size: int, device: torch.device, complement_alpha: float = 0.5): + self.V = vocab_size + self.alpha = complement_alpha + self.bi_counts = torch.zeros(vocab_size, vocab_size, device=device, dtype=torch.float32) + self.bi_totals = torch.zeros(vocab_size, device=device, dtype=torch.float32) + @torch.no_grad() + def update(self, x: Tensor, y: Tensor): + xf = x.reshape(-1) + yf = y.reshape(-1) + ones = torch.ones(xf.numel(), device=xf.device, dtype=torch.float32) + self.bi_counts.reshape(-1).scatter_add_(0, xf * self.V + yf, ones) + self.bi_totals.scatter_add_(0, xf, ones) + def get_weights(self, x: Tensor, y: Tensor) -> Tensor: + xf = x.reshape(-1) + yf = y.reshape(-1) + total = self.bi_totals[xf] + count = self.bi_counts.reshape(-1)[xf * self.V + yf] + ngram_prob = count / (total + 1) + return (1.0 - self.alpha * ngram_prob).clamp(min=0.1) + +# --- Batched Newton-Schulz orthogonalization --- + +def zeropower_via_newtonschulz5(G: Tensor, steps: int = 5, eps: float = 1e-7) -> Tensor: + """Batched Newton-Schulz orthogonalization. G: (B,M,N) or (M,N).""" + a, b, c = (3.4445, -4.7750, 2.0315) + was_2d = G.ndim == 2 + if was_2d: + G = G.unsqueeze(0) + X = G.bfloat16() + transposed = X.size(-2) > X.size(-1) + if transposed: + X = X.mT + X = X / (X.norm(dim=(-2, -1), keepdim=True) + eps) + for _ in range(steps): + A = X @ X.mT + B = b * A + c * (A @ A) + X = a * X + B @ X + if transposed: + X = X.mT + if was_2d: + X = X.squeeze(0) + return X + +# --- Parallel Muon optimizer --- + +class Muon(torch.optim.Optimizer): + """Parallel Muon: post-backward reduce-scatter -> local NS5 -> all-gather. + + No DDP for bank params. After backward, this optimizer: + 1. Launches async reduce-scatter for all banks (biggest first) + 2. Returns control so Adam can step on small params while RS is in-flight + 3. Waits for each RS, runs local NS5 on the shard, launches async all-gather + 4. Each all-gather overlaps with next bank's NS5 + """ + def __init__(self, params, lr: float, momentum: float, backend_steps: int, + nesterov: bool = True, weight_decay: float = 0.0): + super().__init__( + params, + dict(lr=lr, momentum=momentum, backend_steps=backend_steps, + nesterov=nesterov, weight_decay=weight_decay), + ) + self._built = False + + def _build(self): + self._distributed = dist.is_available() and dist.is_initialized() + self._world_size = dist.get_world_size() if self._distributed else 1 + self._rank = dist.get_rank() if self._distributed else 0 + ws = self._world_size + + self._bank_meta = [] + for group in self.param_groups: + for p in group["params"]: + B = p.shape[0] + padded_B = ((B + ws - 1) // ws) * ws + shard_B = padded_B // ws + tail = p.shape[1:] + dev = p.device + self._bank_meta.append({ + 'p': p, + 'B': B, + 'padded_grad': torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + 'shard': torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + 'shard_mom': torch.zeros(shard_B, *tail, device=dev, dtype=torch.bfloat16), + 'full_update': torch.zeros(padded_B, *tail, device=dev, dtype=torch.bfloat16), + 'scale': max(1, p.shape[-2] / p.shape[-1]) ** 0.5, + }) + # Sort by size descending -- launch biggest reduce-scatters first + self._bank_meta.sort(key=lambda m: -m['p'].numel()) + self._built = True + + def launch_reduce_scatters(self): + """Phase 1: launch async reduce-scatter for all banks. Call right after backward.""" + if not self._built: + self._build() + if not self._distributed: + return + self._rs_futures = [] + for m in self._bank_meta: + p = m['p'] + if p.grad is None: + self._rs_futures.append(None) + continue + pg = m['padded_grad'] + pg[:m['B']].copy_(p.grad.bfloat16()) + if pg.shape[0] > m['B']: + pg[m['B']:].zero_() + fut = dist.reduce_scatter_tensor(m['shard'], pg, op=dist.ReduceOp.AVG, async_op=True) + self._rs_futures.append(fut) + + @torch.no_grad() + def step(self, closure=None): + """Phase 3: wait for RS, local NS5, all-gather. Call AFTER Adam steps.""" + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + if not self._built: + self._build() + + for group in self.param_groups: + lr = group["lr"] + momentum = group["momentum"] + backend_steps = group["backend_steps"] + nesterov = group["nesterov"] + wd = group.get("weight_decay", 0.0) + + prev_ag_handle = None + prev_m = None + + sharded = self._distributed and hasattr(self, '_rs_futures') + + for i, m in enumerate(self._bank_meta): + p = m['p'] + if p.grad is None: + continue + + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m['p'] + upd = prev_m['full_update'][:prev_m['B']] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m['scale']) + + if sharded and self._rs_futures[i] is not None: + self._rs_futures[i].wait() + g = m['shard'] + buf = m['shard_mom'] + else: + g = p.grad.bfloat16() + state = self.state[p] + if "momentum_buffer" not in state: + state["momentum_buffer"] = torch.zeros_like(g) + buf = state["momentum_buffer"] + + buf.mul_(momentum).add_(g) + if nesterov: + update = g.add(buf, alpha=momentum) + else: + update = buf + + update = zeropower_via_newtonschulz5(update, steps=backend_steps) + + if sharded: + prev_ag_handle = dist.all_gather_into_tensor( + m['full_update'], update, async_op=True) + prev_m = m + else: + if wd > 0.0: + p.data.mul_(1.0 - lr * wd) + p.add_(update.to(dtype=p.dtype), alpha=-lr * m['scale']) + + if prev_ag_handle is not None: + prev_ag_handle.wait() + pp = prev_m['p'] + upd = prev_m['full_update'][:prev_m['B']] + if wd > 0.0: + pp.data.mul_(1.0 - lr * wd) + pp.add_(upd.to(dtype=pp.dtype), alpha=-lr * prev_m['scale']) + + if hasattr(self, '_rs_futures'): + del self._rs_futures + + return loss + +# --- Tokenizer evaluation helpers --- + +def build_sentencepiece_luts( + sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device +) -> tuple[Tensor, Tensor, Tensor]: + sp_vocab_size = int(sp.vocab_size()) + table_size = max(sp_vocab_size, vocab_size) + base_bytes_np = np.zeros((table_size,), dtype=np.int16) + has_leading_space_np = np.zeros((table_size,), dtype=np.bool_) + is_boundary_token_np = np.ones((table_size,), dtype=np.bool_) + for token_id in range(sp_vocab_size): + if sp.is_control(token_id) or sp.is_unknown(token_id) or sp.is_unused(token_id): + continue + is_boundary_token_np[token_id] = False + if sp.is_byte(token_id): + base_bytes_np[token_id] = 1 + continue + piece = sp.id_to_piece(token_id) + if piece.startswith("\u2581"): + has_leading_space_np[token_id] = True + piece = piece[1:] + base_bytes_np[token_id] = len(piece.encode("utf-8")) + return ( + torch.tensor(base_bytes_np, dtype=torch.int16, device=device), + torch.tensor(has_leading_space_np, dtype=torch.bool, device=device), + torch.tensor(is_boundary_token_np, dtype=torch.bool, device=device), + ) +def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() + usable = ((tokens.numel() - 1) // seq_len) * seq_len + if usable <= 0: + raise ValueError(f"Validation split is too short for TRAIN_SEQ_LEN={seq_len}") + return tokens[: usable + 1] +def eval_val( + args: Hyperparameters, + model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + grad_accum_steps: int, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + "VAL_BATCH_SIZE must provide at least one sequence per rank; " + f"got VAL_BATCH_SIZE={args.val_batch_size}, WORLD_SIZE={world_size}, " + f"GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_tokens.numel() - 1) // seq_len + seq_start = (total_seqs * rank) // world_size + seq_end = (total_seqs * (rank + 1)) // world_size + val_loss_sum = torch.zeros((), device=device, dtype=torch.float64) + val_token_count = torch.zeros((), device=device, dtype=torch.float64) + val_byte_count = torch.zeros((), device=device, dtype=torch.float64) + model.eval() + with torch.inference_mode(): + for batch_seq_start in range(seq_start, seq_end, local_batch_seqs): + batch_seq_end = min(batch_seq_start + local_batch_seqs, seq_end) + raw_start = batch_seq_start * seq_len + raw_end = batch_seq_end * seq_len + 1 + local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + batch_loss = model(x, y).detach() + batch_token_count = float(y.numel()) + val_loss_sum += batch_loss.to(torch.float64) * batch_token_count + val_token_count += batch_token_count + prev_ids = x.reshape(-1) + tgt_ids = y.reshape(-1) + token_bytes = base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += (has_leading_space_lut[tgt_ids] & ~is_boundary_token_lut[prev_ids]).to(dtype=torch.int16) + val_byte_count += token_bytes.to(torch.float64).sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(val_loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(val_token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(val_byte_count, op=dist.ReduceOp.SUM) + val_loss = val_loss_sum / val_token_count + bits_per_token = val_loss.item() / math.log(2.0) + tokens_per_byte = val_token_count.item() / val_byte_count.item() + model.train() + return float(val_loss.item()), float(bits_per_token * tokens_per_byte) + +# --- Quantization helpers --- + +CONTROL_TENSOR_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "CONTROL_TENSOR_NAME_PATTERNS", + "attn_scale,attn_scales,mlp_scale,mlp_scales,resid_mix,resid_mixes,q_gain,skip_weight,skip_weights,smear,dtg_gate,ve_layer_scales,ve_shared.scale,attn_gate,vr_lambda", + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_FP32_NAME_PATTERNS = tuple( + pattern + for pattern in os.environ.get( + "INT8_KEEP_FLOAT_FP32_NAME_PATTERNS", + ",".join(CONTROL_TENSOR_NAME_PATTERNS), + ).split(",") + if pattern +) +INT8_KEEP_FLOAT_MAX_NUMEL = 65_536 +INT8_KEEP_FLOAT_STORE_DTYPE = torch.float16 +INT8_PER_ROW_SCALE_DTYPE = torch.float16 +INT8_CLIP_PERCENTILE = 99.99984 +INT8_CLIP_Q = INT8_CLIP_PERCENTILE / 100.0 +def tensor_nbytes(t: Tensor) -> int: + return int(t.numel()) * int(t.element_size()) +def keep_float_tensor(name: str, t: Tensor, passthrough_orig_dtypes: dict[str, str]) -> Tensor: + if any(pattern in name for pattern in INT8_KEEP_FLOAT_FP32_NAME_PATTERNS): + return t.float().contiguous() + if t.dtype in {torch.float32, torch.bfloat16}: + passthrough_orig_dtypes[name] = str(t.dtype).removeprefix("torch.") + return t.to(dtype=INT8_KEEP_FLOAT_STORE_DTYPE).contiguous() + return t +def quantize_float_tensor(t: Tensor) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + clip_abs = ( + torch.quantile(t32.abs(), INT8_CLIP_Q, dim=1) + if t32.numel() + else torch.empty((t32.shape[0],), dtype=torch.float32) + ) + clipped = torch.maximum(torch.minimum(t32, clip_abs[:, None]), -clip_abs[:, None]) + scale = (clip_abs / 127.0).clamp_min(1.0 / 127.0) + q = torch.clamp(torch.round(clipped / scale[:, None]), -127, 127).to(torch.int8).contiguous() + return q, scale.to(dtype=INT8_PER_ROW_SCALE_DTYPE).contiguous() + clip_abs = float(torch.quantile(t32.abs().flatten(), INT8_CLIP_Q).item()) if t32.numel() else 0.0 + scale = torch.tensor(clip_abs / 127.0 if clip_abs > 0 else 1.0, dtype=torch.float32) + q = torch.clamp(torch.round(torch.clamp(t32, -clip_abs, clip_abs) / scale), -127, 127).to(torch.int8).contiguous() + return q, scale +def _classify_param(name: str) -> str: + if "tok_emb" in name or "lm_head" in name: + return "embed" + if "f1_corr_in" in name or "f1_corr_out" in name: + return "aux" + if "qo_bank" in name or "kv_bank" in name: + return "attn" + if "mlp_up_bank" in name or "mlp_down_bank" in name: + return "mlp" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" +# GPTQ: Hessian-aware quantization with column-wise error compensation +def _find_best_row_scales(W: Tensor, clip_range: int = 31) -> Tensor: + t32 = W.float() + best_s = t32.abs().amax(dim=1) / clip_range + best_s = best_s.clamp_min(1.0 / clip_range) + best_err = torch.full((t32.shape[0],), float('inf')) + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range) + q = torch.clamp(torch.round(t32 / s[:, None]), -clip_range, clip_range) + recon = q * s[:, None] + err = (t32 - recon).pow(2).mean(dim=1) + improved = err < best_err + best_s[improved] = s[improved] + best_err[improved] = err[improved] + return best_s +def gptq_quantize_weight(W: Tensor, H: Tensor, clip_range: int = 31, + block_size: int = 64, percdamp: float = 0.002) -> tuple[Tensor, Tensor]: + """GPTQ: quantize weight matrix W using Hessian H = X^T X for error compensation. + Returns (quantized_int8, scale_fp16) in int6 range [-clip_range, clip_range].""" + W = W.float().clone() + rows, cols = W.shape + row_scale = _find_best_row_scales(W, clip_range) + H = H.float().clone() + damp = percdamp * H.diag().mean() + H.diagonal().add_(damp) + perm = torch.argsort(H.diag()) + invperm = torch.argsort(perm) + W = W[:, perm] + H = H[perm][:, perm] + try: + L = torch.linalg.cholesky(H) + Hinv = torch.cholesky_inverse(L) + except torch._C._LinAlgError: + Hinv = torch.diag(1.0 / H.diag().clamp_min(1e-6)) + Q = torch.zeros(rows, cols, dtype=torch.int8) + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + W_block = W[:, i1:i2].clone() + Hinv_block = Hinv[i1:i2, i1:i2] + Err = torch.zeros_like(W_block) + for j in range(i2 - i1): + w_col = W_block[:, j] + h_inv_jj = Hinv_block[j, j].clamp_min(1e-8) + q_col = torch.clamp(torch.round(w_col / row_scale), -clip_range, clip_range) + deq_col = q_col * row_scale + Q[:, i1 + j] = q_col.to(torch.int8) + err = (w_col - deq_col) / h_inv_jj + Err[:, j] = err + if j + 1 < i2 - i1: + W_block[:, j + 1:] -= err.unsqueeze(1) * Hinv_block[j, j + 1:].unsqueeze(0) + if i2 < cols: + W[:, i2:] -= Err @ Hinv[i1:i2, i2:] + Q = Q[:, invperm] + return Q, row_scale.to(torch.float16) +def gptq_calibrate(model: nn.Module, train_pattern: str, device: torch.device, + n_samples: int = 256, seq_len: int = 2048) -> dict[str, Tensor]: + """Collect Hessian H = X^T X for each linear layer using training data.""" + hessians: dict[str, Tensor] = {} + n_seen: dict[str, int] = {} + hooks = [] + def make_hook(name: str): + def hook_fn(module, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + if name not in hessians: + hessians[name] = torch.zeros(x.shape[1], x.shape[1], device=x.device, dtype=torch.float32) + n_seen[name] = 0 + hessians[name].addmm_(x.t(), x) + n_seen[name] += x.shape[0] + return hook_fn + for name, module in model.named_modules(): + if isinstance(module, (nn.Linear, CastedLinear)): + hooks.append(module.register_forward_hook(make_hook(name))) + stream = TokenStream(train_pattern) + model.eval() + with torch.no_grad(): + for _ in range(n_samples): + tokens = stream.take(seq_len + 1).to(device=device, dtype=torch.int64) + x = tokens[:-1].unsqueeze(0) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + model.forward_logits(x) + for h in hooks: + h.remove() + for name in hessians: + hessians[name] /= max(n_seen[name], 1) + model.train() + return hessians +def quantize_int6_per_row(t: Tensor, clip_range: int = 31) -> tuple[Tensor, Tensor]: + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale +def quantize_int5_per_row(t: Tensor, clip_range: int = 15) -> tuple[Tensor, Tensor]: + """int5 (signed, range [-15, 15]) per-row quant. Modeled on quantize_int6_per_row. + Returned int8 tensor holds values in [-15, 15] — leaves the high 3 bits as zeros, which + brotli compresses very efficiently. Dequant path is identical (q.float() * scale).""" + t32 = t.float() + if t32.ndim == 2: + best_q, best_s, best_err = None, None, float('inf') + for pct in [0.9990, 0.9995, 0.9999, 0.99999, 1.0]: + if pct < 1.0: + row_clip = torch.quantile(t32.abs(), pct, dim=1) + else: + row_clip = t32.abs().amax(dim=1) + s = (row_clip / clip_range).clamp_min(1.0 / clip_range).to(torch.float16) + q = torch.clamp(torch.round(t32 / s.float()[:, None]), -clip_range, clip_range).to(torch.int8) + recon = q.float() * s.float()[:, None] + err = (t32 - recon).pow(2).mean().item() + if err < best_err: + best_q, best_s, best_err = q, s, err + return best_q, best_s + amax = t32.abs().max().item() + scale = torch.tensor(amax / clip_range if amax > 0 else 1.0, dtype=torch.float16) + q = torch.clamp(torch.round(t32 / scale.float()), -clip_range, clip_range).to(torch.int8) + return q, scale +def _classify_param_fine(name: str) -> str: + """Finer classifier than _classify_param, splitting attention/MLP banks for mixed-int policy. + Returns one of: embed, qo (Q+O bank), kv (K+V bank), mlp_up (mlp_fc), mlp_down (mlp_proj), + aux, attn_other, mlp_other, other. Categories `qo`/`kv` map to attention; `mlp_up`/`mlp_down` + map to MLP. The split lets us route mlp banks to int5 while keeping attn at int6.""" + if "tok_emb" in name or "lm_head" in name: + return "embed" + if "f1_corr_in" in name or "f1_corr_out" in name: + return "aux" + if "qo_bank" in name: + return "qo" + if "kv_bank" in name: + return "kv" + if "mlp_up_bank" in name: + return "mlp_up" + if "mlp_down_bank" in name: + return "mlp_down" + if ".mlp." in name: + return "mlp_other" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn_other" + return "other" +# salvage_v2 mixed-int policy (applied at the Tensor level via _classify_param_fine): +# - mlp_down_bank (mlp_proj, MOST quant-tolerant per 11-day collate) -> int5 (clip_range 15) +# - mlp_up_bank (mlp_fc, also tolerant) -> int5 +# - qo_bank, kv_bank (attention; LEAST quant-tolerant) -> int6 (clip_range 31) +# - tok_emb / lm_head (embed) -> int6 (matches seed; keeping +# attn/embed at int6 for +# quant safety on this +# first salvage attempt) +# Bytes savings: int5 keeps the int8 storage container (no bit-packing) but the high 3 bits are +# forced zero, giving brotli a compressible pattern. Combined with the byte-shuffle wrapper, +# expected savings vs uniform-int6+zstd is roughly the int5 bit ratio (5/6 = -17%) APPLIED only +# to MLP banks (which are ~60-65% of the model parameter mass for a 10L Rascal at mlp_mult=3.0). +# Net est: ~0.62 * 0.17 = ~10-11% blob shrink from int5 alone, plus brotli ~5-8% over zstd. +DEFAULT_INT5_CATS = {"mlp_down", "mlp_up"} +DEFAULT_INT6_CATS = {"qo", "kv", "attn_other", "mlp_other", "aux"} +def mixed_quantize_int6_gptq(state_dict: dict[str, Tensor], int6_cats: set[str], + hessians: dict[str, Tensor], + int5_cats: set[str] | None = None) -> tuple[dict, dict]: + """Mixed-int (int5/int6/int8) quant with GPTQ for matrix categories when Hessian available. + `int6_cats` and `int5_cats` use FINE-grained category names from `_classify_param_fine` + (qo, kv, mlp_up, mlp_down, attn_other, mlp_other, aux, embed). For backwards-compat with + the old uniform-int6 caller, the legacy coarse names {'mlp','attn','aux','embed'} are also + accepted in `int6_cats` and expand to their fine-grained children. `int5_cats` always uses + fine names. If `int5_cats` is None, defaults to DEFAULT_INT5_CATS (the salvage_v2 policy).""" + if int5_cats is None: + int5_cats = set(DEFAULT_INT5_CATS) + # Expand legacy coarse names so the existing call signature keeps working. + _LEGACY = { + "mlp": {"mlp_up", "mlp_down", "mlp_other"}, + "attn": {"qo", "kv", "attn_other"}, + "aux": {"aux"}, + "embed": {"embed"}, + } + expanded_int6: set[str] = set() + for c in int6_cats: + expanded_int6.update(_LEGACY.get(c, {c})) + # int5 categories take precedence over int6 for the same fine name. + expanded_int6 -= int5_cats + result: dict[str, Tensor] = {} + meta: dict[str, object] = {} + gptq_count, naive_count, int5_count = 0, 0, 0 + for name, tensor in state_dict.items(): + t = tensor.detach().cpu().contiguous() + cat = _classify_param_fine(name) + if not t.is_floating_point() or t.numel() <= 65536: + result[name] = t.to(torch.float16) if t.is_floating_point() else t + meta[name] = "passthrough" + continue + if any(p in name for p in CONTROL_TENSOR_NAME_PATTERNS): + result[name] = t.float() + meta[name] = "passthrough_ctrl" + continue + if cat in int5_cats and t.ndim == 2: + module_name = name.rsplit(".weight", 1)[0] if name.endswith(".weight") else name + H = hessians.get(module_name) + if H is not None and H.shape[0] == t.shape[1]: + q, s = gptq_quantize_weight(t, H.cpu(), clip_range=15) + gptq_count += 1 + else: + q, s = quantize_int5_per_row(t) + naive_count += 1 + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int5"} + int5_count += 1 + elif cat in int5_cats and t.ndim >= 1: + t_2d = t.reshape(-1, t.shape[-1]) if t.ndim > 2 else t + q, s = quantize_int5_per_row(t_2d) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int5"} + naive_count += 1 + int5_count += 1 + elif cat in expanded_int6 and t.ndim == 2: + module_name = name.rsplit(".weight", 1)[0] if name.endswith(".weight") else name + H = hessians.get(module_name) + if H is not None and H.shape[0] == t.shape[1]: + q, s = gptq_quantize_weight(t, H.cpu()) + gptq_count += 1 + else: + q, s = quantize_int6_per_row(t) + naive_count += 1 + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + elif cat in expanded_int6 and t.ndim >= 1: + t_2d = t.reshape(-1, t.shape[-1]) if t.ndim > 2 else t + q, s = quantize_int6_per_row(t_2d) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int6"} + naive_count += 1 + else: + t_q = t.reshape(-1, t.shape[-1]) if t.ndim > 2 else t + q, s = quantize_float_tensor(t_q) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = {"type": "int8"} + print(f"gptq_quantize: {gptq_count} GPTQ layers, {naive_count} naive layers, {int5_count} int5 layers", flush=True) + return result, meta +def dequantize_mixed_int6(result: dict[str, Tensor], meta: dict[str, object], + template_sd: dict[str, Tensor]) -> dict[str, Tensor]: + out: dict[str, Tensor] = {} + for name, orig in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if info in ("passthrough", "passthrough_ctrl", "passthrough_fp16"): + t = result[name] + if t.dtype == torch.float16 and orig_dtype in (torch.float32, torch.bfloat16): + t = t.to(orig_dtype) + out[name] = t + continue + q, s = result[name + ".q"], result[name + ".scale"] + if s.ndim > 0: + val = (q.float() * s.float().view(q.shape[0], *([1] * (q.ndim - 1)))).to(orig_dtype) + else: + val = (q.float() * float(s.item())).to(orig_dtype) + out[name] = val.reshape(orig.shape) if val.shape != orig.shape else val + return out + +# --- Data loading --- + +SHARD_HEADER_DTYPE = np.dtype(" dict[str, int]: + header = np.fromfile(file, dtype=SHARD_HEADER_DTYPE, count=SHARD_HEADER_WORDS) + if header.size != SHARD_HEADER_WORDS or int(header[0]) != SHARD_MAGIC or int(header[1]) != SHARD_VERSION: + raise ValueError(f"Unexpected shard header for {file}") + return {"num_tokens": int(header[2])} + +def load_data_shard(file: Path) -> Tensor: + header = read_data_shard_header(file) + num_tokens = header["num_tokens"] + expected_size = SHARD_HEADER_BYTES + num_tokens * SHARD_TOKEN_DTYPE.itemsize + if file.stat().st_size != expected_size: + raise ValueError(f"Shard size mismatch for {file}: expected {expected_size} bytes") + tokens_np = np.fromfile(file, dtype=SHARD_TOKEN_DTYPE, count=num_tokens, offset=SHARD_HEADER_BYTES) + if tokens_np.size != num_tokens: + raise ValueError(f"Short read for {file}") + return torch.from_numpy(tokens_np.astype(np.uint16, copy=False)) + +def choose_coprime_stride(modulus: int, salt: int) -> int: + if modulus <= 1: + return 1 + candidate = abs(salt) % modulus + if candidate == 0: + candidate = 1 + while math.gcd(candidate, modulus) != 1: + candidate += 1 + if candidate >= modulus: + candidate = 1 + return candidate + +class TokenStream: + def __init__(self, pattern: str): + self.files = [Path(p) for p in sorted(glob.glob(pattern))] + if not self.files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + self.file_idx = 0 + self.tokens = load_data_shard(self.files[0]) + self.pos = 0 + def _advance_file(self) -> None: + self.file_idx = (self.file_idx + 1) % len(self.files) + self.tokens = load_data_shard(self.files[self.file_idx]) + self.pos = 0 + def take(self, n: int) -> Tensor: + chunks: list[Tensor] = [] + remaining = n + while remaining > 0: + avail = self.tokens.numel() - self.pos + if avail <= 0: + self._advance_file() + continue + k = min(remaining, avail) + chunks.append(self.tokens[self.pos : self.pos + k]) + self.pos += k + remaining -= k + return chunks[0] if len(chunks) == 1 else torch.cat(chunks) +class DistributedTokenLoader: + def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): + self.rank = rank + self.world_size = world_size + self.device = device + self.stream = TokenStream(pattern) + def describe(self) -> str: + return f"loader:sequential shards:{len(self.stream.files)}" + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + per_rank_span = local_tokens + 1 + chunk = self.stream.take(per_rank_span * self.world_size) + start = self.rank * per_rank_span + local = chunk[start : start + per_rank_span].to(dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +class CoprimeDistributedTokenLoader: + """Shard-aware block sampler with deterministic coprime walks.""" + def __init__( + self, + pattern: str, + rank: int, + world_size: int, + device: torch.device, + seq_len: int, + seed: int, + max_loaded_shards: int, + shards_per_batch: int, + shard_hold_steps: int, + ): + self.rank = rank + self.world_size = world_size + self.device = device + self.seq_len = seq_len + self.seed = seed + self.token_offsets = torch.arange(seq_len + 1, dtype=torch.int64) + self.cache: OrderedDict[Path, Tensor] = OrderedDict() + files = [Path(p) for p in sorted(glob.glob(pattern))] + if not files: + raise FileNotFoundError(f"No files found for pattern: {pattern}") + self.shards: list[dict[str, int | Path]] = [] + for shard_idx, file in enumerate(files): + header = read_data_shard_header(file) + num_blocks = (header["num_tokens"] - 1) // seq_len + if num_blocks <= 0: + continue + self.shards.append( + { + "file": file, + "num_blocks": num_blocks, + "offset": (seed * 131 + shard_idx * 17) % num_blocks, + "stride": choose_coprime_stride(num_blocks, seed * 29 + shard_idx * 7 + 1), + } + ) + if not self.shards: + raise ValueError(f"No usable shards found for seq_len={seq_len}") + self.num_shards = len(self.shards) + self.max_loaded_shards = max(1, min(max_loaded_shards, self.num_shards)) + self.shards_per_batch = max(1, min(shards_per_batch, self.num_shards)) + self.shard_hold_steps = max(1, shard_hold_steps) + self.batch_shard_stride = choose_coprime_stride(self.num_shards, seed * 41 + 3) + self.batch_idx = 0 + self.shard_visits = [0 for _ in range(self.num_shards)] + def _get_tokens(self, file: Path) -> Tensor: + cached = self.cache.get(file) + if cached is not None: + self.cache.move_to_end(file) + return cached + # CPU advanced indexing is not implemented for uint16, so cache coprime-loader + # shards in int32 and cast to int64 only after batch assembly. + tokens = load_data_shard(file).to(dtype=torch.int32) + if len(self.cache) >= self.max_loaded_shards: + self.cache.popitem(last=False) + self.cache[file] = tokens + return tokens + def _sample_sequences(self, shard_idx: int, count: int) -> Tensor: + shard = self.shards[shard_idx] + num_blocks = int(shard["num_blocks"]) + offset = int(shard["offset"]) + stride = int(shard["stride"]) + visits = self.shard_visits[shard_idx] + block_ids = ( + offset + + (visits + torch.arange(count, dtype=torch.int64)) * stride + ) % num_blocks + self.shard_visits[shard_idx] += count + token_starts = block_ids * self.seq_len + gather_idx = token_starts.unsqueeze(1) + self.token_offsets.unsqueeze(0) + tokens = self._get_tokens(shard["file"]) + return tokens[gather_idx] + def describe(self) -> str: + total_blocks = sum(int(shard["num_blocks"]) for shard in self.shards) + return ( + f"loader:coprime shards:{self.num_shards} blocks:{total_blocks} " + f"seq_len:{self.seq_len} shards_per_batch:{self.shards_per_batch} " + f"cache:{self.max_loaded_shards} batch_stride:{self.batch_shard_stride} " + f"hold_steps:{self.shard_hold_steps}" + ) + def next_batch(self, global_tokens: int, seq_len: int, grad_accum_steps: int) -> tuple[Tensor, Tensor]: + if seq_len != self.seq_len: + raise ValueError(f"Coprime loader was built for seq_len={self.seq_len}, got {seq_len}") + local_tokens = global_tokens // (self.world_size * grad_accum_steps) + if local_tokens % seq_len != 0: + raise ValueError( + f"TRAIN_BATCH_TOKENS={global_tokens} does not divide into full local sequences " + f"for WORLD_SIZE={self.world_size}, GRAD_ACCUM_STEPS={grad_accum_steps}, seq_len={seq_len}" + ) + local_seqs = local_tokens // seq_len + active_shards = min(self.shards_per_batch, self.num_shards, local_seqs) + if active_shards <= 0: + raise ValueError(f"No active shards available for local_seqs={local_seqs}") + seqs_per_shard = local_seqs // active_shards + seq_remainder = local_seqs % active_shards + hold_idx = self.batch_idx // self.shard_hold_steps + shard_start = ((hold_idx * self.world_size) + self.rank) * self.batch_shard_stride + chunks: list[Tensor] = [] + for shard_slot in range(active_shards): + count = seqs_per_shard + (1 if shard_slot < seq_remainder else 0) + if count <= 0: + continue + shard_idx = (shard_start + shard_slot * self.batch_shard_stride) % self.num_shards + chunks.append(self._sample_sequences(shard_idx, count)) + self.batch_idx += 1 + local = chunks[0] if len(chunks) == 1 else torch.cat(chunks, dim=0) + local = local.to(dtype=torch.int64) + x = local[:, :-1] + y = local[:, 1:] + return x.to(self.device, non_blocking=True), y.to(self.device, non_blocking=True) + +def build_train_loader(args: Hyperparameters, rank: int, world_size: int, device: torch.device): + if args.loader_mode == "sequential": + return DistributedTokenLoader(args.train_files, rank, world_size, device) + if args.loader_mode == "coprime": + return CoprimeDistributedTokenLoader( + args.train_files, + rank, + world_size, + device, + seq_len=args.train_seq_len, + seed=args.seed, + max_loaded_shards=args.coprime_max_loaded_shards, + shards_per_batch=args.coprime_shards_per_batch, + shard_hold_steps=args.coprime_shard_hold_steps, + ) + raise ValueError(f"Unknown LOADER_MODE={args.loader_mode!r}") + +# --- Transformer modules --- + +class RMSNorm(nn.Module): + def __init__(self, eps: float | None = None): + super().__init__() + self.eps = eps + def forward(self, x: Tensor) -> Tensor: + return F.rms_norm(x, (x.size(-1),), eps=self.eps) +class CastedLinear(nn.Linear): + _qat_enabled: bool = False + def forward(self, x: Tensor) -> Tensor: + w = self.weight.to(x.dtype) + if CastedLinear._qat_enabled and self.training and w.ndim == 2: + with torch.no_grad(): + w32 = self.weight.float() + row_max = w32.abs().amax(dim=1) + scale = (row_max / 31.0).clamp_min(1.0 / 31.0) + w_q = (torch.clamp(torch.round(w32 / scale[:, None]), -32, 31) * scale[:, None]).to(x.dtype) + w = w + (w_q - w).detach() + bias = self.bias.to(x.dtype) if self.bias is not None else None + return F.linear(x, w, bias) +def restore_low_dim_params_to_fp32(module: nn.Module) -> None: + with torch.no_grad(): + for name, param in module.named_parameters(): + if (param.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS)) and param.dtype != torch.float32: + param.data = param.data.float() +class Rotary(nn.Module): + def __init__(self, dim: int, base: float = 10000.0, train_seq_len: int = 1024, rope_dims: int = 0): + super().__init__() + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.rope_dims = rope_dims if rope_dims > 0 else dim + inv_freq = 1.0 / (base ** (torch.arange(0, self.rope_dims, 2, dtype=torch.float32) / self.rope_dims)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self._seq_len_cached = 0 + self._cos_cached: Tensor | None = None + self._sin_cached: Tensor | None = None + def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + if ( + self._cos_cached is None + or self._sin_cached is None + or self._seq_len_cached != seq_len + or self._cos_cached.device != device + ): + rd = self.rope_dims + if seq_len > self.train_seq_len: + scale = seq_len / self.train_seq_len + new_base = self.base * (scale ** (rd / (rd - 2))) + inv_freq = 1.0 / (new_base ** (torch.arange(0, rd, 2, dtype=torch.float32, device=device) / rd)) + else: + inv_freq = self.inv_freq.to(device) + t = torch.arange(seq_len, device=device, dtype=inv_freq.dtype) + freqs = torch.outer(t, inv_freq) + self._cos_cached = freqs.cos()[None, :, None, :] + self._sin_cached = freqs.sin()[None, :, None, :] + self._seq_len_cached = seq_len + return self._cos_cached.to(dtype=dtype), self._sin_cached.to(dtype=dtype) +def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor, rope_dims: int = 0) -> Tensor: + if rope_dims > 0 and rope_dims < x.size(-1): + x_rope, x_pass = x[..., :rope_dims], x[..., rope_dims:] + half = rope_dims // 2 + x1, x2 = x_rope[..., :half], x_rope[..., half:] + x_rope = torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + return torch.cat((x_rope, x_pass), dim=-1) + half = x.size(-1) // 2 + x1, x2 = x[..., :half], x[..., half:] + return torch.cat((x1 * cos + x2 * sin, x1 * (-sin) + x2 * cos), dim=-1) + +class CausalSelfAttention(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + rope_base: float, + qk_gain_init: float, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + if dim % num_heads != 0: + raise ValueError("model_dim must be divisible by num_heads") + if num_heads % num_kv_heads != 0: + raise ValueError("num_heads must be divisible by num_kv_heads") + self.num_heads = num_heads + self.num_kv_heads = num_kv_heads + self.head_dim = dim // num_heads + if self.head_dim % 2 != 0: + raise ValueError("head_dim must be even for RoPE") + # No CastedLinear -- weights come from banks + self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) + self.rope_dims = 0 # set by GPT.__init__ for partial RoPE + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=1024) + self.use_xsa = False # set by GPT.__init__ for deep layers only + # Gated attention and value residual (non-banked small params) + self.gated_attention = gated_attention + if gated_attention: + self.attn_gate = nn.Linear(dim, num_heads, bias=True) + nn.init.zeros_(self.attn_gate.weight) + nn.init.constant_(self.attn_gate.bias, 4.0) + self.value_residual = value_residual + if value_residual: + self.vrl_alpha = nn.Parameter(torch.zeros(1, dtype=torch.float32)) # sigmoid gate (PR #569 style) + def _xsa_efficient(self, y: Tensor, v: Tensor) -> Tensor: + """Efficient XSA: subtract self-value projection via GQA-aware reshape (no repeat_interleave). + y: [B, T, H, D], v: [B, T, Hkv, D]. H must be divisible by Hkv.""" + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) # [B, T, Hkv, group, D] + vn = F.normalize(v, dim=-1).unsqueeze(-2) # [B, T, Hkv, 1, D] -- broadcast ready + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + def forward(self, x: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + bsz, seqlen, dim = x.shape + q = F.linear(x, q_w.to(x.dtype)).reshape(bsz, seqlen, self.num_heads, self.head_dim) + k = F.linear(x, k_w.to(x.dtype)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + v = F.linear(x, v_w.to(x.dtype)) + if v_embed is not None: + v = v + v_embed + v = v.reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) + raw_v = v if self.value_residual else None + if self.value_residual and v0 is not None: + alpha = torch.sigmoid(self.vrl_alpha.to(dtype=v.dtype)) + v = v + alpha * v0 # sigmoid-gated residual (PR #569 style) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = self.rotary(seqlen, x.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, self.rope_dims) + k = apply_rotary_emb(k, cos, sin, self.rope_dims) + q = q * self.q_gain.to(dtype=q.dtype)[None, None, :, None] + if flash_attn_3_func is not None: + q_attn, k_attn, v_attn = q, k, v + if q_attn.dtype not in (torch.float16, torch.bfloat16): + q_attn = q_attn.to(torch.bfloat16) + k_attn = k_attn.to(torch.bfloat16) + v_attn = v_attn.to(torch.bfloat16) + y = flash_attn_3_func(q_attn, k_attn, v_attn, causal=True) + else: + qh = q.transpose(1, 2) + kh = k.transpose(1, 2) + vh = v.transpose(1, 2) + if self.num_heads != self.num_kv_heads: + repeat = self.num_heads // self.num_kv_heads + kh = kh.repeat_interleave(repeat, dim=1) + vh = vh.repeat_interleave(repeat, dim=1) + y = F.scaled_dot_product_attention(qh, kh, vh, is_causal=True).transpose(1, 2) + if self.use_xsa: + y = self._xsa_efficient(y, v) + if self.gated_attention: + # gate shape: (bsz, seqlen, num_heads) -> (bsz, seqlen, num_heads, 1) for B,T,H,D layout + gate = torch.sigmoid(self.attn_gate(x)).unsqueeze(-1) + y = y * gate + y = y.reshape(bsz, seqlen, dim) + return F.linear(y, out_w.to(x.dtype)), raw_v + +class SmearGate(nn.Module): + def __init__(self, dim: int): + super().__init__() + self.gate = nn.Parameter(torch.zeros(dim, dtype=torch.float32)) + def forward(self, x: Tensor) -> Tensor: + g = torch.sigmoid(self.gate.to(dtype=x.dtype))[None, None, :] + x_prev = torch.cat([torch.zeros_like(x[:, :1]), x[:, :-1]], dim=1) + return (1 - g) * x + g * x_prev + +class BigramHashEmbedding(nn.Module): + def __init__(self, bigram_vocab_size: int, bigram_dim: int, model_dim: int, trigram: bool = False): + super().__init__() + self.bigram_vocab_size = bigram_vocab_size + self._trigram = trigram + self.embed = nn.Embedding(bigram_vocab_size, bigram_dim) + nn.init.zeros_(self.embed.weight) + self.proj = CastedLinear(bigram_dim, model_dim, bias=False) if bigram_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.05, dtype=torch.float32)) + def bigram_hash(self, tokens: Tensor) -> Tensor: + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., 0] = mod + out[..., 1:] = torch.bitwise_xor(36313 * t[..., 1:], 27191 * t[..., :-1]) % mod + return out.long() + def trigram_hash(self, tokens: Tensor) -> Tensor: + """Hash (t-2, t-1, t) trigrams into same embedding table. Zero extra params.""" + t = tokens.to(torch.int32) + mod = self.bigram_vocab_size - 1 + out = torch.empty_like(t) + out[..., :2] = mod + out[..., 2:] = (36313 * t[..., 2:] ^ 27191 * t[..., 1:-1] ^ 51497 * t[..., :-2]) % mod + return out.long() + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(self.bigram_hash(token_ids)) + if self._trigram: + h = h + self.embed(self.trigram_hash(token_ids)) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class ValueEmbedding(nn.Module): + """Reinject token identity into attention values at specific layers. + Each table maps vocab tokens to a low-dim embedding, projected to model_dim.""" + def __init__(self, vocab_size: int, ve_dim: int, model_dim: int): + super().__init__() + self.embed = nn.Embedding(vocab_size, ve_dim) + nn.init.normal_(self.embed.weight, std=0.01) + self.proj = CastedLinear(ve_dim, model_dim, bias=False) if ve_dim != model_dim else None + if self.proj is not None: + nn.init.zeros_(self.proj.weight) + self.scale = nn.Parameter(torch.tensor(0.1, dtype=torch.float32)) + def forward(self, token_ids: Tensor) -> Tensor: + h = self.embed(token_ids) + if self.proj is not None: + h = self.proj(h) + return h * self.scale.to(dtype=h.dtype) + +class MLP(nn.Module): + def __init__(self, dim: int, mlp_mult: int): + super().__init__() + # No CastedLinear -- weights come from banks + self.kernel_mode = os.environ.get("MLP_KERNEL_MODE", "").strip().lower() + def forward(self, x: Tensor, up_w: Tensor, down_w: Tensor) -> Tensor: + x = F.linear(x, up_w.to(x.dtype)) + x = leaky_relu_sq(x, kernel_mode=self.kernel_mode) + return F.linear(x, down_w.to(x.dtype)) + +class Block(nn.Module): + def __init__( + self, + dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + rope_base: float, + qk_gain_init: float, + layer_idx: int = 0, + ln_scale: bool = False, + dtg: bool = False, + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self.attn_norm = RMSNorm() + self.mlp_norm = RMSNorm() + self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init, + gated_attention=gated_attention, value_residual=value_residual) + self.mlp = MLP(dim, mlp_mult) + attn_scale_init = float(os.environ.get("ATTN_SCALE_INIT", "1.0")) + mlp_scale_init = float(os.environ.get("MLP_SCALE_INIT", "1.0")) + resid_mix_x_init = float(os.environ.get("RESID_MIX_X_INIT", "1.0")) + resid_mix_x0_init = float(os.environ.get("RESID_MIX_X0_INIT", "0.0")) + self.attn_scale = nn.Parameter(torch.full((dim,), attn_scale_init, dtype=torch.float32)) + self.mlp_scale = nn.Parameter(torch.full((dim,), mlp_scale_init, dtype=torch.float32)) + self.resid_mix = nn.Parameter( + torch.stack( + ( + torch.full((dim,), resid_mix_x_init, dtype=torch.float32), + torch.full((dim,), resid_mix_x0_init, dtype=torch.float32), + ) + ) + ) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 + if dtg: + self.dtg_gate = nn.Linear(dim, 1, bias=True) + nn.init.zeros_(self.dtg_gate.weight) + nn.init.constant_(self.dtg_gate.bias, 2.0) + else: + self.dtg_gate = None + def forward(self, x: Tensor, x0: Tensor, q_w: Tensor, k_w: Tensor, v_w: Tensor, out_w: Tensor, up_w: Tensor, down_w: Tensor, v_embed: Tensor | None = None, v0: Tensor | None = None) -> tuple[Tensor, Tensor | None]: + mix = self.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out, raw_v = self.attn(self.attn_norm(x_in) * self.ln_scale_factor, q_w, k_w, v_w, out_w, v_embed=v_embed, v0=v0) + x_out = x_in + self.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + x_out = x_out + self.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * self.mlp(self.mlp_norm(x_out) * self.ln_scale_factor, up_w, down_w) + if self.dtg_gate is not None: + gate = torch.sigmoid(self.dtg_gate(x_in.detach())) + x_out = x_in + gate * (x_out - x_in) + return x_out, raw_v + +class GPT(nn.Module): + def __init__( + self, + vocab_size: int, + num_layers: int, + model_dim: int, + num_heads: int, + num_kv_heads: int, + mlp_mult: int, + tie_embeddings: bool, + tied_embed_init_std: float, + logit_softcap: float, + rope_base: float, + qk_gain_init: float, + mtp_num_heads: int = 0, + mtp_loss_weight: float = 0.1, + bigram_vocab_size: int = 0, + bigram_dim: int = 128, + xsa_last_n: int = 0, + rope_dims: int = 0, + ln_scale: bool = False, + dtg: bool = False, + ve_enabled: bool = False, + ve_dim: int = 128, + ve_layers: str = "9,10", + gated_attention: bool = False, + value_residual: bool = False, + ): + super().__init__() + self._ve_target_dim = num_kv_heads * (model_dim // num_heads) # kv_dim for value projection + if logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {logit_softcap}") + self.tie_embeddings = tie_embeddings + self.tied_embed_init_std = tied_embed_init_std + self.logit_softcap = logit_softcap + self.value_residual = value_residual + self.mtp_num_heads = mtp_num_heads + self.mtp_loss_weight = mtp_loss_weight + self.tok_emb = nn.Embedding(vocab_size, model_dim) + self.bigram = BigramHashEmbedding(bigram_vocab_size, bigram_dim, model_dim, trigram=bool(int(os.environ.get("TRIGRAM", "0")))) if bigram_vocab_size > 0 else None + self.smear = SmearGate(model_dim) + self.num_encoder_layers = num_layers // 2 + self.num_decoder_layers = num_layers - self.num_encoder_layers + self.num_skip_weights = min(self.num_encoder_layers, self.num_decoder_layers) + self.skip_weights = nn.Parameter(torch.ones(self.num_skip_weights, model_dim, dtype=torch.float32)) + # Parameter banks: contiguous 3D tensors for batched optimizer + head_dim = model_dim // num_heads + kv_dim = num_kv_heads * head_dim + mlp_dim = int(mlp_mult * model_dim) + self.num_layers = num_layers + self.qo_bank = nn.Parameter(torch.empty(2 * num_layers, model_dim, model_dim)) + self.kv_bank = nn.Parameter(torch.empty(2 * num_layers, kv_dim, model_dim)) + self.mlp_up_bank = nn.Parameter(torch.empty(num_layers, mlp_dim, model_dim)) + self.mlp_down_bank = nn.Parameter(torch.empty(num_layers, model_dim, mlp_dim)) + self.blocks = nn.ModuleList( + [ + Block( + model_dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + layer_idx=i, + ln_scale=ln_scale, + dtg=dtg, + gated_attention=gated_attention, + value_residual=value_residual, + ) + for i in range(num_layers) + ] + ) + if rope_dims > 0: + head_dim = model_dim // num_heads + for block in self.blocks: + block.attn.rope_dims = rope_dims + block.attn.rotary = Rotary(head_dim, base=rope_base, train_seq_len=1024, rope_dims=rope_dims) + self.ve_layer_indices = [int(x) for x in ve_layers.split(",") if x.strip()] if ve_enabled else [] + kv_dim_ve = self._ve_target_dim + if self.ve_layer_indices: + self.ve_shared = ValueEmbedding(vocab_size, ve_dim, kv_dim_ve) + self.ve_layer_scales = nn.ParameterList( + [nn.Parameter(torch.ones(1, dtype=torch.float32)) for _ in self.ve_layer_indices] + ) + else: + self.ve_shared = None + self.ve_layer_scales = nn.ParameterList() + self.value_embeds = nn.ModuleList() # keep empty for compat + self.final_norm = RMSNorm() + self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + if self.lm_head is not None: + self.lm_head._zero_init = True + self.mtp_heads = nn.ModuleList( + [CastedLinear(model_dim, vocab_size, bias=False) for _ in range(mtp_num_heads)] + ) + for head in self.mtp_heads: + head._zero_init = True + if xsa_last_n > 0: + for i in range(max(0, num_layers - xsa_last_n), num_layers): + self.blocks[i].attn.use_xsa = True + self._init_weights() + def _init_weights(self) -> None: + if self.tie_embeddings: + nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) + n = self.num_layers + proj_scale = 1.0 / math.sqrt(2 * n) + # Init banks: orthogonal, with proj layers scaled down and out/down zero-init + for i in range(n): + nn.init.orthogonal_(self.qo_bank.data[i], gain=1.0) # Q + nn.init.zeros_(self.qo_bank.data[n + i]) # Out (zero init) + nn.init.orthogonal_(self.kv_bank.data[i], gain=1.0) # K + nn.init.orthogonal_(self.kv_bank.data[n + i], gain=1.0) # V + nn.init.orthogonal_(self.mlp_up_bank.data[i], gain=1.0) # MLP up + nn.init.zeros_(self.mlp_down_bank.data[i]) # MLP down (zero init) + # Scale proj layers (out_proj and mlp_down are "proj" layers) + self.qo_bank.data[n + i].mul_(proj_scale) + self.mlp_down_bank.data[i].mul_(proj_scale) + # Init remaining nn.Linear modules (bigram proj, mtp heads, lm_head) + for name, module in self.named_modules(): + if isinstance(module, nn.Linear): + if getattr(module, "_zero_init", False): + nn.init.zeros_(module.weight) + elif module.weight.ndim == 2 and module.weight.shape[0] >= 64 and module.weight.shape[1] >= 64: + nn.init.orthogonal_(module.weight, gain=1.0) + def _get_ve(self, layer_idx: int, input_ids: Tensor, ve_cache: dict | None = None) -> Tensor | None: + """Get value embedding for a specific layer using shared table + per-layer scale.""" + if self.ve_shared is None or layer_idx not in self.ve_layer_indices: + return None + if ve_cache is not None and 've' not in ve_cache: + ve_cache['ve'] = self.ve_shared(input_ids) + ve_base = ve_cache['ve'] if ve_cache is not None else self.ve_shared(input_ids) + ve_idx = self.ve_layer_indices.index(layer_idx) + return ve_base * self.ve_layer_scales[ve_idx].to(dtype=ve_base.dtype) + def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + x_flat = x.reshape(-1, x.size(-1)) + targets = target_ids.reshape(-1) + if self.tie_embeddings: + logits_proj = F.linear(x_flat, self.tok_emb.weight) + else: + if self.lm_head is None: + raise RuntimeError("lm_head is required when tie_embeddings=False") + logits_proj = self.lm_head(x_flat) + logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + if hasattr(self, '_ngram_tracker') and self._ngram_tracker is not None and self.training: + per_tok_loss = F.cross_entropy(logits.float(), targets, reduction="none") + weights = self._ngram_tracker.get_weights(input_ids, target_ids) + main_loss = (per_tok_loss * weights).mean() + else: + main_loss = F.cross_entropy(logits.float(), targets, reduction="mean") + if self.training and self.mtp_num_heads > 0 and self.mtp_loss_weight > 0.0: + _, seqlen, dim = x.shape + mtp_loss_sum = x.new_zeros(()) + mtp_loss_count = 0 + for k, mtp_head in enumerate(self.mtp_heads): + valid_t = seqlen - (k + 1) + if valid_t <= 0: + continue + mtp_hidden = x[:, :valid_t, :].reshape(-1, dim) + mtp_targets = target_ids[:, k + 1 :].reshape(-1) + mtp_logits_proj = mtp_head(mtp_hidden) + mtp_logits = self.logit_softcap * torch.tanh(mtp_logits_proj / self.logit_softcap) + mtp_loss_sum = mtp_loss_sum + F.cross_entropy(mtp_logits.float(), mtp_targets, reduction="mean") + mtp_loss_count += 1 + if mtp_loss_count > 0: + main_loss = main_loss + self.mtp_loss_weight * (mtp_loss_sum / mtp_loss_count) + return main_loss + def forward_logits(self, input_ids: Tensor) -> Tensor: + """Return logits (bsz, seq_len, vocab) without computing loss.""" + n = self.num_layers + x = self.tok_emb(input_ids) + if self.bigram is not None: + x = x + self.bigram(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + x = self.smear(x) + x0 = x + v0 = None + skips: list[Tensor] = [] + ve_cache: dict = {} + for i in range(self.num_encoder_layers): + ve = self._get_ve(i, input_ids, ve_cache) + x, raw_v = self.blocks[i](x, x0, + self.qo_bank[i], self.kv_bank[i], self.kv_bank[n + i], + self.qo_bank[n + i], self.mlp_up_bank[i], self.mlp_down_bank[i], + v_embed=ve, v0=v0) + if v0 is None and raw_v is not None: + v0 = raw_v + skips.append(x) + for i in range(self.num_decoder_layers): + bi = self.num_encoder_layers + i + if skips: + x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() + ve = self._get_ve(bi, input_ids, ve_cache) + x, _ = self.blocks[bi](x, x0, + self.qo_bank[bi], self.kv_bank[bi], self.kv_bank[n + bi], + self.qo_bank[n + bi], self.mlp_up_bank[bi], self.mlp_down_bank[bi], + v_embed=ve, v0=v0) + x = self.final_norm(x) + if self.tie_embeddings: + logits_proj = F.linear(x, self.tok_emb.weight) + else: + logits_proj = self.lm_head(x) + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) + +# --- N-gram bulk update and hashed n-gram sliding eval --- + +def _ngram_bulk_update(val_np, start, end, ctx_tables, full_tables, + min_order, max_order, primes, mask): + """Bulk update n-gram tables with a contiguous range of tokens. + All ranks call this with the SAME token range -> identical tables everywhere.""" + t = val_np[start:end].astype(np.uint64) + n = len(t) + for order in range(min_order, max_order + 1): + if n < order: + continue + ctx_width = order - 1 + ctx_hash = np.zeros(n - order + 1, dtype=np.uint64) + for k in range(ctx_width): + ctx_hash ^= t[k:n - order + 1 + k] * primes[k % len(primes)] + ctx_key = (ctx_hash & mask).astype(np.int64) + tgt = t[order - 1:] + full_key = ((ctx_hash ^ (tgt * primes[ctx_width % len(primes)])) & mask).astype(np.int64) + ctx_tables[order] += np.bincount(ctx_key, minlength=len(ctx_tables[order])).astype(np.uint32) + full_tables[order] += np.bincount(full_key, minlength=len(full_tables[order])).astype(np.uint32) + +def eval_val_sliding_hashed_ngram( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + order: int, + alpha: float, + min_count: int, + buckets: int, + max_seconds: float = 0.0, + batch_seqs: int = 128, + eval_seq_len: int | None = None, +) -> tuple[float, float, float]: + """Score-first sliding eval with chunk-based SHARED n-gram tables + cubric. + + Key design: all ranks share identical n-gram tables via bulk chunk updates. + Each chunk's windows are distributed across ranks for scoring, then ALL ranks + update tables with the same contiguous token range. Every rank sees the full + n-gram picture (not 1/world_size like per-segment updates). + + Legal: entire chunk scored before its tokens update the tables. + """ + min_order = max(args.ngram_eval_min_order, 2) + max_order = max(order, min_order) + adaptive = args.ngram_eval_adaptive + alpha_min = args.ngram_eval_alpha_min + alpha_max = args.ngram_eval_alpha_max + ent_center = args.ngram_eval_entropy_center + ent_scale = args.ngram_eval_entropy_scale + + # Parse fixed per-order multipliers (PR #809 style) + _fixed_order_mults = None + if args.ngram_order_mults_str: + _fixed_order_mults = np.array([float(x) for x in args.ngram_order_mults_str.split(",")], dtype=np.float64) + + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + + # Build all windows and total scored tokens + all_window_starts = [ws for ws in range(0, total_tokens, stride) if min(ws + seq_len, total_tokens) - ws >= 1] + total_scored_tokens = 0.0 + for ws in all_window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + total_scored_tokens += float(max(wlen - s, 0)) + + # Group windows into chunks by scored position -- all ranks share this grouping + chunk_tokens = int(os.environ.get("NGRAM_CHUNK_TOKENS", "1048576")) # 1M default + num_chunks = (total_tokens + chunk_tokens - 1) // chunk_tokens + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in all_window_starts: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_start = ws + s + ci = min(scored_start // chunk_tokens, num_chunks - 1) + chunk_windows[ci].append(ws) + + val_np = val_tokens.numpy() + ctx_tables = {n: np.zeros((buckets,), dtype=np.uint32) for n in range(min_order, max_order + 1)} + full_tables = {n: np.zeros((buckets,), dtype=np.uint32) for n in range(min_order, max_order + 1)} + mask = np.uint64(buckets - 1) + primes = np.array( + [np.uint64(36313), np.uint64(27191), np.uint64(51647), np.uint64(81929), + np.uint64(131071), np.uint64(174763), np.uint64(233017)], + dtype=np.uint64, + ) + + loss_sum = 0.0 + token_count = 0.0 + byte_count = 0.0 + + # Cubric 3D: per (order x entropy_bin x count_bin) adaptive alpha scaling + _NUM_ENT_BINS = 3 # low / mid / high entropy + _NUM_CNT_BINS = 3 # low / mid / high count + _ENT_EDGES = np.array([ent_center - 1.0, ent_center + 1.0]) # [2.0, 4.0] for center=3.0 + _CNT_EDGES = np.array([5.0, 50.0]) # low=<5, mid=5-50, high=>50 context count + _TOTAL_CELLS = _NUM_ENT_BINS * _NUM_CNT_BINS # 9 cells per order = 54 total + _cc = getattr(args, 'cubric_cadence', 0); _con = _cc > 0; _cfired = 0 + if _con: + # Warm-start: proven converged values from 4+ runs (orders 2-7) + # All 9 cells per order get the same warm-start, 3D cubric refines from there + _WARM = {2: 0.45, 3: 0.30, 4: 0.45, 5: 1.88, 6: 2.00, 7: 2.00, 8: 2.00, 9: 2.00} + _c_alpha_mult = {n: [_WARM.get(n, 1.0)] * _TOTAL_CELLS for n in range(min_order, max_order + 1)} + _c_hits = {n: [0] * _TOTAL_CELLS for n in range(min_order, max_order + 1)} + _c_beats = {n: [0] * _TOTAL_CELLS for n in range(min_order, max_order + 1)} + + base_model.eval() + compiled_logits = maybe_compile( + base_model.forward_logits, + enabled=args.compile_enabled, + fullgraph=False, + ) + t0 = time.perf_counter() + deadline = (t0 + max_seconds) if max_seconds > 0.0 else None + cutoff_hit = False + + if rank == 0: + print(f"ngram_eval:chunks={num_chunks} chunk_tokens={chunk_tokens} " + f"windows={len(all_window_starts)} shared_tables=True", flush=True) + + with torch.inference_mode(): + for ci in range(num_chunks): + if deadline is not None and time.perf_counter() >= deadline: + cutoff_hit = True + break + + windows = chunk_windows[ci] + if not windows: + continue + + # Distribute this chunk's windows across ranks + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + + # --- Phase 1: SCORE this chunk's windows --- + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + logits_f = logits.float() + nll = F.cross_entropy( + logits_f.reshape(-1, logits_f.size(-1)), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + seg_len = wlen - s + if seg_len <= 0: + continue + + seg_nll = nll[i, s:wlen].to(torch.float64).cpu().numpy() + seg_model_p = np.exp(-seg_nll) + + if adaptive: + log_probs = F.log_softmax(logits_f[i, s:wlen], dim=-1) + probs_a = log_probs.exp() + entropy = -(probs_a * log_probs).sum(dim=-1).cpu().numpy() + sig = 1.0 / (1.0 + np.exp(-ent_scale * (entropy - ent_center))) + per_token_alpha = alpha_min + (alpha_max - alpha_min) * sig + # Bin entropy for 2D cubric: 0=low, 1=mid, 2=high + _ent_bins = np.digitize(entropy, _ENT_EDGES).astype(np.int32) + else: + per_token_alpha = np.full(seg_len, alpha) + _ent_bins = np.ones(seg_len, dtype=np.int32) # all mid + + global_j = np.arange(ws + s + 1, ws + wlen + 1, dtype=np.int64) + p_ng = np.zeros(seg_len, dtype=np.float64) + ng_matched = np.zeros(seg_len, dtype=np.bool_) + _ng_ord = np.zeros(seg_len, dtype=np.int32) + _ng_ctx_count = np.zeros(seg_len, dtype=np.float64) + tgt_np = val_np[global_j].astype(np.uint64) + + for n in range(max_order, min_order - 1, -1): + ctx_width = n - 1 + valid = (global_j >= ctx_width) & (~ng_matched) + if not valid.any(): + continue + v_idx = np.nonzero(valid)[0] + jv = global_j[v_idx] + ctx_hash = np.zeros(len(jv), dtype=np.uint64) + for k in range(ctx_width): + tok = val_np[jv - (ctx_width - k)].astype(np.uint64) + ctx_hash ^= tok * primes[k % len(primes)] + ctx_key = (ctx_hash & mask).astype(np.int64) + full_key = ((ctx_hash ^ (tgt_np[v_idx] * primes[ctx_width % len(primes)])) & mask).astype(np.int64) + ctx_counts = ctx_tables[n][ctx_key].astype(np.float64) + full_counts = full_tables[n][full_key].astype(np.float64) + has_data = ctx_counts >= float(min_count) + if has_data.any(): + p = np.minimum(full_counts, ctx_counts) / np.maximum(ctx_counts, 1.0) + p = np.clip(p, 0.0, 1.0) + hit_idx = v_idx[has_data] + p_ng[hit_idx] = p[has_data] + ng_matched[hit_idx] = True + _ng_ord[hit_idx] = n + _ng_ctx_count[hit_idx] = ctx_counts[has_data] + + # Mix where n-gram matched (PR #809 style or cubric 3D fallback) + if ng_matched.any(): + m_idx = np.nonzero(ng_matched)[0] + # Per-order entropy center shift (PR #809) + if adaptive and args.ngram_entropy_shift: + matched_ords = _ng_ord[m_idx].astype(np.float64) + shifted_centers = ent_center - 0.25 * (matched_ords - float(min_order)) + shifted_sig = 1.0 / (1.0 + np.exp(-ent_scale * (entropy[m_idx] - shifted_centers))) + per_token_alpha[m_idx] = alpha_min + (alpha_max - alpha_min) * shifted_sig + if _fixed_order_mults is not None: + # PR #809 fixed order multipliers (replaces cubric) + a = per_token_alpha[m_idx].copy() + mult_indices = _ng_ord[m_idx] - min_order + mult_indices = np.clip(mult_indices, 0, len(_fixed_order_mults) - 1) + a *= _fixed_order_mults[mult_indices] + np.clip(a, 0.0, 0.95, out=a) + elif _con: + a = per_token_alpha[m_idx].copy() + m_ent_bins = _ent_bins[m_idx] + m_cnt_bins = np.digitize(_ng_ctx_count[m_idx], _CNT_EDGES).astype(np.int32) + for n in range(min_order, max_order + 1): + om = _ng_ord[m_idx] == n + if not om.any(): + continue + for eb in range(_NUM_ENT_BINS): + for cb in range(_NUM_CNT_BINS): + cell = eb * _NUM_CNT_BINS + cb + mask_ecb = om & (m_ent_bins == eb) & (m_cnt_bins == cb) + if mask_ecb.any(): + _c_hits[n][cell] += int(mask_ecb.sum()) + _c_beats[n][cell] += int((p_ng[m_idx[mask_ecb]] > seg_model_p[m_idx[mask_ecb]]).sum()) + a[mask_ecb] *= _c_alpha_mult[n][cell] + np.clip(a, 0.0, 0.95, out=a) + else: + a = per_token_alpha[m_idx] + seg_model_p[m_idx] = (1.0 - a) * seg_model_p[m_idx] + a * p_ng[m_idx] + + seg_nll = -np.log(np.clip(seg_model_p, 1e-12, 1.0)) + loss_sum += float(seg_nll.sum()) + token_count += float(seg_len) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += float(tb.sum().item()) + + # --- Phase 2: SHARED UPDATE -- all ranks update with same chunk tokens --- + chunk_start = ci * chunk_tokens + chunk_end = min((ci + 1) * chunk_tokens, total_tokens) + _ngram_bulk_update(val_np, chunk_start, chunk_end + 1, + ctx_tables, full_tables, min_order, max_order, + primes, mask) + + # Cubric 2D c-step: adapt per (order x entropy_bin) + if _con: + # Collect all (order, ent_bin, cnt_bin) cells with enough data + all_rates = [] + for n in range(min_order, max_order + 1): + for cell in range(_TOTAL_CELLS): + if _c_hits[n][cell] >= 8: + all_rates.append(_c_beats[n][cell] / _c_hits[n][cell]) + if len(all_rates) >= 4: + avg_rate = sum(all_rates) / len(all_rates) + for n in range(min_order, max_order + 1): + for cell in range(_TOTAL_CELLS): + if _c_hits[n][cell] >= 8: + rate = _c_beats[n][cell] / _c_hits[n][cell] + if rate > avg_rate + 0.05: + _c_alpha_mult[n][cell] = min(_c_alpha_mult[n][cell] * 1.03, 2.0) + elif rate < avg_rate - 0.05: + _c_alpha_mult[n][cell] = max(_c_alpha_mult[n][cell] * 0.97, 0.3) + _cfired += 1 + if rank == 0 and _cfired % 8 == 0: + parts = [] + for n in range(min_order, max_order + 1): + m = _c_alpha_mult[n] + avg_m = sum(m) / len(m) + parts.append(f"o{n}:avg={avg_m:.2f}") + print(f"cubric3d:step={_cfired} {' '.join(parts)}", flush=True) + _c_hits = {n: [0] * _TOTAL_CELLS for n in range(min_order, max_order + 1)} + _c_beats = {n: [0] * _TOTAL_CELLS for n in range(min_order, max_order + 1)} + + # Progress + if rank == 0 and (ci % 10 == 0 or ci == num_chunks - 1 or ci < 3): + elapsed = time.perf_counter() - t0 + cur_bpb = (loss_sum / max(token_count, 1.0)) / math.log(2.0) * (token_count / max(byte_count, 1.0)) if token_count > 0 else 0.0 + print( + f"ngram_eval:chunk [{ci+1}/{num_chunks}] bpb={cur_bpb:.6f} t={elapsed:.0f}s", + flush=True, + ) + + # All-reduce across ranks + _loss = torch.tensor(loss_sum, device=device, dtype=torch.float64) + _toks = torch.tensor(token_count, device=device, dtype=torch.float64) + _bytes = torch.tensor(byte_count, device=device, dtype=torch.float64) + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(_loss, op=dist.ReduceOp.SUM) + dist.all_reduce(_toks, op=dist.ReduceOp.SUM) + dist.all_reduce(_bytes, op=dist.ReduceOp.SUM) + loss_sum = _loss.item() + token_count = _toks.item() + byte_count = _bytes.item() + + coverage = token_count / max(total_scored_tokens, 1.0) + if cutoff_hit: + elapsed = time.perf_counter() - t0 + print( + f"ngram_eval:cutoff max_seconds={max_seconds:.1f} " + f"coverage={coverage*100:.2f}% elapsed={elapsed:.0f}s", + flush=True, + ) + + if _con and rank == 0: + print(f"cubric3d:final c_steps={_cfired} cells={_TOTAL_CELLS}x{max_order-min_order+1}={_TOTAL_CELLS*(max_order-min_order+1)}", flush=True) + for n in range(min_order, max_order + 1): + m = _c_alpha_mult[n] + row = " ".join(f"{m[cell]:.2f}" for cell in range(_TOTAL_CELLS)) + print(f" o{n}: [{row}]", flush=True) + val_loss = loss_sum / max(token_count, 1.0) + val_bpb = val_loss / math.log(2.0) * (token_count / max(byte_count, 1.0)) + base_model.train() + return val_loss, val_bpb, coverage + +# --- Sliding window evaluation --- + +def eval_val_sliding( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + """Sliding window evaluation: each token scored with maximum context.""" + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + window_starts = [ws for ws in range(0, total_tokens, stride) + if min(ws + seq_len, total_tokens) - ws >= 1] + total_windows = len(window_starts) + my_s = (total_windows * rank) // world_size + my_e = (total_windows * (rank + 1)) // world_size + my_windows = window_starts[my_s:my_e] + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + base_model.eval() + compiled_logits = maybe_compile( + base_model.forward_logits, + enabled=args.compile_enabled, + fullgraph=args.compile_fullgraph, + ) + with torch.inference_mode(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk[:-1] + y_batch[i, :wlen] = chunk[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else max(wlen - stride, 0) + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.train() + return val_loss, bits_per_token * tokens_per_byte + + +# --- Training --- + +def main() -> None: + code = Path(__file__).read_text(encoding="utf-8") + args = Hyperparameters() + # zeropower_via_newtonschulz5 runs eagerly with bmm -- do NOT compile + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + rank = int(os.environ.get("RANK", "0")) + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + if world_size != 8: + raise ValueError( + f"Rascal 4k 12L brotli+mixed 8x requires WORLD_SIZE=8, got {world_size}. " + "Launch with: torchrun --standalone --nproc_per_node=8 4k_vocab_rascal_12l_brotli_mixed/train_gpt_4K_12L_brotli_mixed_8xgpu.py" + ) + grad_accum_steps = 8 // world_size + grad_scale = 1.0 / grad_accum_steps + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + master_process = rank == 0 + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + from torch.backends.cuda import enable_cudnn_sdp, enable_flash_sdp, enable_math_sdp, enable_mem_efficient_sdp + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + logfile = None + if master_process: + os.makedirs("logs", exist_ok=True) + logfile = f"logs/{args.run_id}.txt" + print(logfile) + def log0(msg: str, console: bool = True) -> None: + if not master_process: + return + if console: + print(msg) + if logfile is not None: + with open(logfile, "a", encoding="utf-8") as f: + print(msg, file=f) + log0(code, console=False) + log0("=" * 100, console=False) + log0("condition_id:rascal_4k_12L_brotli_mixed_8x_seed444") + log0("run_label:salvage_v2 source_record:rascal_4k_8x_seed444_run20260427 axis:depth_12L+brotli+mixed_int") + log0("changed_fields:num_layers (11->12), compression (zstd->brotli+bshf), quant_policy (uniform_int6->mixed_int5_int6_int8)") + log0("expected_metric:final_sliding_window_exact comparator:0.8672_4k_8x_oversize_run prior_size:17766043_target:<16000000") + log0(f"condition:DATA_PATH={args.data_path}") + log0(f"condition:TOKENIZER_PATH={args.tokenizer_path}") + log0(f"condition:VOCAB_SIZE={args.vocab_size}") + log0(f"condition:SEED={args.seed}") + log0(f"condition:MAX_WALLCLOCK_SECONDS={args.max_wallclock_seconds}") + log0(f"condition:LOADER_MODE={args.loader_mode}") + log0(f"condition:COPRIME_MAX_LOADED_SHARDS={args.coprime_max_loaded_shards}") + log0(f"condition:COPRIME_SHARDS_PER_BATCH={args.coprime_shards_per_batch}") + log0(f"condition:COPRIME_SHARD_HOLD_STEPS={args.coprime_shard_hold_steps}") + log0(f"condition:SKIP_GPTQ={os.environ.get('SKIP_GPTQ', '1')}") + log0(f"condition:TRIGRAM={int(args.trigram_enabled)}") + log0(f"condition:NGRAM_EVAL_ORDER={args.ngram_eval_order}") + log0(f"Running Python {sys.version}", console=False) + log0(f"Running PyTorch {torch.__version__}", console=False) + log0( + subprocess.run(["nvidia-smi"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True, check=False).stdout, + console=False, + ) + log0("=" * 100, console=False) + random.seed(args.seed) + np.random.seed(args.seed) + torch.manual_seed(args.seed) + torch.cuda.manual_seed_all(args.seed) + if not args.tokenizer_path.endswith(".model"): + raise ValueError(f"Script only setup for SentencePiece .model file: {args.tokenizer_path}") + sp = spm.SentencePieceProcessor(model_file=args.tokenizer_path) + if int(sp.vocab_size()) != args.vocab_size: + raise ValueError( + f"VOCAB_SIZE={args.vocab_size} does not match tokenizer vocab_size={int(sp.vocab_size())}" + ) + dataset_dir = Path(args.data_path).resolve() + actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) + effective_eval_seq_len = args.eval_seq_len if args.eval_seq_len > 0 else args.train_seq_len + val_seq_len = max(args.train_seq_len, effective_eval_seq_len) + val_tokens = load_validation_tokens(args.val_files, val_seq_len) + base_bytes_lut, has_leading_space_lut, is_boundary_token_lut = build_sentencepiece_luts( + sp, args.vocab_size, device + ) + log0(f"val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path={args.tokenizer_path}") + if args.ngram_eval_order >= 2: + log0(f"ngram_eval:order={args.ngram_eval_order} alpha={args.ngram_eval_alpha} min_count={args.ngram_eval_min_count} buckets={args.ngram_eval_buckets}") + log0(f"train_loader:dataset:{dataset_dir.name} train_shards:{actual_train_files}") + log0(f"val_loader:shards pattern={args.val_files} tokens:{val_tokens.numel() - 1}") + CastedLinear._qat_enabled = args.qat_enabled + base_model = GPT( + vocab_size=args.vocab_size, + num_layers=args.num_layers, + model_dim=args.model_dim, + num_heads=args.num_heads, + num_kv_heads=args.num_kv_heads, + mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, + tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, + rope_base=args.rope_base, + qk_gain_init=args.qk_gain_init, + mtp_num_heads=args.mtp_num_heads, + mtp_loss_weight=args.mtp_loss_weight, + bigram_vocab_size=args.bigram_vocab_size, + bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, + rope_dims=args.rope_dims, + ln_scale=args.ln_scale, + dtg=args.dtg_enabled, + ve_enabled=args.ve_enabled, + ve_dim=args.ve_dim, + ve_layers=args.ve_layers, + gated_attention=args.gated_attention, + value_residual=args.value_residual, + ).to(device).bfloat16() + # Banks stay FP32 (like CastedLinear weights), cast to BF16 in forward + base_model.qo_bank.data = base_model.qo_bank.data.float() + base_model.kv_bank.data = base_model.kv_bank.data.float() + base_model.mlp_up_bank.data = base_model.mlp_up_bank.data.float() + base_model.mlp_down_bank.data = base_model.mlp_down_bank.data.float() + for module in base_model.modules(): + if isinstance(module, CastedLinear): + module.float() + restore_low_dim_params_to_fp32(base_model) + if args.complement_alpha > 0: + tracker = TrainNgramTracker(args.vocab_size, device, complement_alpha=args.complement_alpha) + base_model._ngram_tracker = tracker + log0(f"complementary_training:alpha={args.complement_alpha}") + else: + base_model._ngram_tracker = None + # No DDP -- Parallel Muon handles bank grad communication via reduce-scatter, + # and non-bank grads are manually all-reduced before Adam steps. + compiled_model = maybe_compile( + base_model, + enabled=args.compile_enabled, + fullgraph=args.compile_fullgraph, + mode=args.compile_mode, + ) + model = compiled_model + + # Optimizer split: + # - 4 parameter banks -> Muon (batched Newton-Schulz) + # - token embedding -> Adam + # - scalars/control tensors -> Adam + # - bigram proj, mtp heads, VE proj -> Adam (small matrix params not worth banking) + matrix_params = [ + base_model.qo_bank, base_model.kv_bank, + base_model.mlp_up_bank, base_model.mlp_down_bank, + ] + block_named_params = list(base_model.blocks.named_parameters()) + scalar_params = [ + p + for name, p in block_named_params + if p.ndim < 2 or any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) + ] + if base_model.skip_weights.numel() > 0: + scalar_params.append(base_model.skip_weights) + scalar_params.append(base_model.smear.gate) + if base_model.bigram is not None: + scalar_params.append(base_model.bigram.scale) + token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr + tok_params = [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}] + if base_model.bigram is not None: + tok_params.append({"params": [base_model.bigram.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.bigram.proj is not None: + scalar_params.append(base_model.bigram.proj.weight) + if base_model.ve_shared is not None: + tok_params.append({"params": [base_model.ve_shared.embed.weight], "lr": token_lr, "base_lr": token_lr}) + if base_model.ve_shared.proj is not None: + scalar_params.append(base_model.ve_shared.proj.weight) + scalar_params.append(base_model.ve_shared.scale) + for s in base_model.ve_layer_scales: + scalar_params.append(s) + optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + optimizer_muon = Muon( + matrix_params, + lr=args.matrix_lr, + momentum=args.muon_momentum, + backend_steps=args.muon_backend_steps, + weight_decay=args.muon_wd, + ) + for group in optimizer_muon.param_groups: + group["base_lr"] = args.matrix_lr + optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + weight_decay=args.adam_wd, + fused=True, + ) + # Non-bank params that need manual all-reduce (replicated across GPUs) + replicated_params = list(optimizer_tok.param_groups[0]["params"]) + for pg in optimizer_tok.param_groups[1:]: + replicated_params.extend(pg["params"]) + replicated_params.extend(scalar_params) + + optimizer_head = None + if base_model.lm_head is not None: + optimizer_head = torch.optim.Adam( + [{"params": [base_model.lm_head.weight], "lr": args.head_lr, "base_lr": args.head_lr}], + betas=(args.beta1, args.beta2), + eps=args.adam_eps, + fused=True, + ) + replicated_params.append(base_model.lm_head.weight) + optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] + if optimizer_head is not None: + optimizers.append(optimizer_head) + n_params = sum(p.numel() for p in base_model.parameters()) + mtp_params = sum(p.numel() for p in base_model.mtp_heads.parameters()) + log0(f"model_params:{n_params}") + log0(f"mtp_num_heads:{args.mtp_num_heads} mtp_loss_weight:{args.mtp_loss_weight} mtp_params:{mtp_params}") + xsa_layers = [i for i, b in enumerate(base_model.blocks) if b.attn.use_xsa] + log0(f"XSA:last_{args.xsa_last_n} active_layers:{xsa_layers}") + log0(f"world_size:{world_size} grad_accum_steps:{grad_accum_steps}") + log0("sdp_backends:cudnn=False flash=True mem_efficient=False math=False") + log0(f"attention_mode:gqa num_heads:{args.num_heads} num_kv_heads:{args.num_kv_heads}") + log0( + f"tie_embeddings:{args.tie_embeddings} embed_lr:{token_lr} " + f"head_lr:{args.head_lr if base_model.lm_head is not None else 0.0} " + f"matrix_lr:{args.matrix_lr} scalar_lr:{args.scalar_lr}" + ) + log0( + f"train_batch_tokens:{args.train_batch_tokens} train_seq_len:{args.train_seq_len} " + f"iterations:{args.iterations} warmup_steps:{args.warmup_steps} " + f"max_wallclock_seconds:{args.max_wallclock_seconds:.3f}" + ) + compile_mode = args.compile_mode if args.compile_mode else "default" + log0( + f"compile:enabled={int(args.compile_enabled)} mode:{compile_mode} " + f"fullgraph={int(args.compile_fullgraph)}" + ) + log0(f"mlp_kernel_mode:{args.mlp_kernel_mode or 'eager'}") + log0( + f"scale_init:attn={args.attn_scale_init:.4f} mlp={args.mlp_scale_init:.4f} " + f"resid_mix=({args.resid_mix_x_init:.4f},{args.resid_mix_x0_init:.4f}) " + f"ln_scale={int(args.ln_scale)}" + ) + log0(f"seed:{args.seed}") + train_loader = build_train_loader(args, rank, world_size, device) + log0(train_loader.describe()) + def zero_grad_all() -> None: + for opt in optimizers: + opt.zero_grad(set_to_none=True) + max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + # GPTQ calibration reads training data — it must complete within the wallclock budget. + # We stop the training loop early (by GPTQ_RESERVE_MS) so GPTQ runs before the cap. + _skip_gptq = int(os.environ.get("SKIP_GPTQ", "1")) + _gptq_reserve_ms = float(os.environ.get("GPTQ_RESERVE_MS", "30000")) if (max_wallclock_ms is not None and not _skip_gptq) else 0.0 + effective_max_wallclock_ms = (max_wallclock_ms - _gptq_reserve_ms) if max_wallclock_ms is not None else None + def lr_mul(step: int, elapsed_ms: float) -> float: + if args.warmdown_iters <= 0: + return 1.0 + if max_wallclock_ms is None: + warmdown_start = max(args.iterations - args.warmdown_iters, 0) + return max((args.iterations - step) / max(args.warmdown_iters, 1), 0.0) if warmdown_start <= step < args.iterations else 1.0 + step_ms = elapsed_ms / max(step, 1) + warmdown_ms = args.warmdown_iters * step_ms + remaining_ms = max(max_wallclock_ms - elapsed_ms, 0.0) + return remaining_ms / max(warmdown_ms, 1e-9) if remaining_ms <= warmdown_ms else 1.0 + if args.warmup_steps > 0: + initial_model_state = {name: tensor.detach().cpu().clone() for name, tensor in base_model.state_dict().items()} + initial_optimizer_states = [copy.deepcopy(opt.state_dict()) for opt in optimizers] + model.train() + for warmup_step in range(args.warmup_steps): + zero_grad_all() + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + warmup_loss = model(x, y) + (warmup_loss * grad_scale).backward() + # All-reduce all grads for warmup (simple, not optimized) + if distributed: + for p in base_model.parameters(): + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + for opt in optimizers: + opt.step() + zero_grad_all() + if args.warmup_steps <= 20 or (warmup_step + 1) % 10 == 0 or warmup_step + 1 == args.warmup_steps: + log0(f"warmup_step:{warmup_step + 1}/{args.warmup_steps}") + base_model.load_state_dict(initial_model_state, strict=True) + for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + opt.load_state_dict(state) + zero_grad_all() + train_loader = build_train_loader(args, rank, world_size, device) + log0(f"loader_reset:{train_loader.describe()}") + swa_state: dict[str, Tensor] | None = None + swa_count = 0 + from collections import deque + lawa_queue: deque[dict[str, Tensor]] = deque(maxlen=args.lawa_k) + ema_state = {name: t.detach().float().clone() for name, t in base_model.state_dict().items()} + ema_decay = 0.997 + training_time_ms = 0.0 + stop_after_step: int | None = None + torch.cuda.synchronize() + t0 = time.perf_counter() + step = 0 + while True: + last_step = step == args.iterations or (stop_after_step is not None and step >= stop_after_step) + should_validate = last_step or (args.val_loss_every > 0 and step % args.val_loss_every == 0) + if should_validate: + torch.cuda.synchronize() + training_time_ms += 1000.0 * (time.perf_counter() - t0) + val_loss, val_bpb = eval_val( + args, + model, + rank, + world_size, + device, + grad_accum_steps, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + ) + log0( + f"step:{step}/{args.iterations} val_loss:{val_loss:.4f} val_bpb:{val_bpb:.4f} " + f"train_time:{training_time_ms:.0f}ms step_avg:{training_time_ms / max(step, 1):.2f}ms tok/s:{(step * args.train_batch_tokens) / max(training_time_ms / 1000.0, 1e-9):.0f}" + ) + torch.cuda.synchronize() + t0 = time.perf_counter() + if last_step: + if stop_after_step is not None and step < args.iterations: + log0( + f"stopping_early: wallclock_cap train_time:{training_time_ms:.0f}ms " + f"step:{step}/{args.iterations}" + ) + break + elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + scale = lr_mul(step, elapsed_ms) + if args.late_qat_threshold > 0 and scale < args.late_qat_threshold and not CastedLinear._qat_enabled: + CastedLinear._qat_enabled = True + log0(f"late_qat:enabled step:{step} scale:{scale:.4f}") + zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(grad_accum_steps): + x, y = train_loader.next_batch(args.train_batch_tokens, args.train_seq_len, grad_accum_steps) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y) + train_loss += loss.detach() + (loss * grad_scale).backward() + if base_model._ngram_tracker is not None: + base_model._ngram_tracker.update(x, y) + train_loss /= grad_accum_steps + frac = min(step / args.muon_momentum_warmup_steps, 1.0) if args.muon_momentum_warmup_steps > 0 else 1.0 + muon_momentum = (1 - frac) * args.muon_momentum_warmup_start + frac * args.muon_momentum + for group in optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * scale + if args.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), args.grad_clip_norm) + # === 3-phase overlapped optimizer step === + # Phase 1: Launch async reduce-scatter for banks (biggest first) + optimizer_muon.launch_reduce_scatters() + # Phase 2: All-reduce non-bank grads + step Adam (while bank RS is in-flight) + if distributed: + for p in replicated_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + optimizer_tok.step() + optimizer_scalar.step() + if optimizer_head is not None: + optimizer_head.step() + # Phase 3: Wait for RS, local NS5, all-gather (banks processed last) + optimizer_muon.step() + zero_grad_all() + # EMA update + with torch.no_grad(): + for name, t in base_model.state_dict().items(): + ema_state[name].mul_(ema_decay).add_(t.detach().float(), alpha=1.0 - ema_decay) + step += 1 + approx_training_time_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) + if args.swa_enabled and scale < 0.2 and step % args.swa_every == 0: + if swa_state is None: + swa_state = {name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()} + swa_count = 1 + log0(f"swa:start step:{step}") + else: + for name, t in base_model.state_dict().items(): + swa_state[name] += t.detach().cpu() + swa_count += 1 + if args.lawa_enabled and step % args.lawa_freq == 0: + lawa_queue.append({name: t.detach().cpu().clone() for name, t in base_model.state_dict().items()}) + should_log_train = ( + args.train_log_every > 0 + and (step <= 10 or step % args.train_log_every == 0 or stop_after_step is not None) + ) + if should_log_train: + log0( + f"step:{step}/{args.iterations} train_loss:{train_loss.item():.4f} " + f"train_time:{approx_training_time_ms:.0f}ms step_avg:{approx_training_time_ms / step:.2f}ms tok/s:{(step * args.train_batch_tokens) / max(approx_training_time_ms / 1000.0, 1e-9):.0f}" + ) + reached_cap = effective_max_wallclock_ms is not None and approx_training_time_ms >= effective_max_wallclock_ms + if distributed and effective_max_wallclock_ms is not None: + reached_cap_tensor = torch.tensor(int(reached_cap), device=device) + dist.all_reduce(reached_cap_tensor, op=dist.ReduceOp.MAX) + reached_cap = bool(reached_cap_tensor.item()) + if stop_after_step is None and reached_cap: + stop_after_step = step + log0( + f"peak memory allocated: {torch.cuda.max_memory_allocated() // 1024 // 1024} MiB " + f"reserved: {torch.cuda.max_memory_reserved() // 1024 // 1024} MiB" + ) + # GPTQ calibration: reads training data — must complete within MAX_WALLCLOCK_SECONDS. + # Training loop stopped GPTQ_RESERVE_MS early so this runs inside the budget. + if _skip_gptq: + log0("gptq:SKIPPED (SKIP_GPTQ=1) — will use naive int6") + gptq_hessians: dict[str, Tensor] = {} + else: + log0("gptq:calibrating with training data...") + t_gptq = time.perf_counter() + gptq_hessians = gptq_calibrate(base_model, args.train_files, device, n_samples=256, seq_len=args.train_seq_len) + log0(f"gptq:calibrated {len(gptq_hessians)} layers in {time.perf_counter()-t_gptq:.1f}s") + # Apply weight averaging + if args.lawa_enabled and len(lawa_queue) > 1: + log0(f"lawa:applying LAWA averaging k={len(lawa_queue)}") + current_state = base_model.state_dict() + avg_state = {name: torch.zeros(t.shape, dtype=torch.float32, device='cpu') for name, t in current_state.items()} + for snap in lawa_queue: + for name in avg_state: + avg_state[name] += snap[name].float() + for name in avg_state: + avg_state[name] /= len(lawa_queue) + avg_state[name] = avg_state[name].to(dtype=current_state[name].dtype) + base_model.load_state_dict(avg_state, strict=True) + else: + log0("ema:applying EMA weights") + current_state = base_model.state_dict() + avg_state = {name: t.to(dtype=current_state[name].dtype) for name, t in ema_state.items()} + base_model.load_state_dict(avg_state, strict=True) + if args.post_ema_diagnostic: + torch.cuda.synchronize() + t_diag = time.perf_counter() + diag_val_loss, diag_val_bpb = eval_val( + args, compiled_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"DIAGNOSTIC post_ema val_loss:{diag_val_loss:.4f} val_bpb:{diag_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_diag):.0f}ms" + ) + else: + log0("diagnostic_eval:skipped POST_EMA_DIAGNOSTIC=0") + full_state_dict = base_model.state_dict() + export_sd = {k: v for k, v in full_state_dict.items() if "mtp_heads" not in k} + excluded_mtp = sum(int(t.numel()) for k, t in full_state_dict.items() if "mtp_heads" in k) + if excluded_mtp > 0: + log0(f"export_excluding_mtp_params:{excluded_mtp}") + if master_process: + torch.save(export_sd, "final_model.pt") + model_bytes = os.path.getsize("final_model.pt") + code_bytes = len(code.encode("utf-8")) + log0(f"Serialized model: {model_bytes} bytes") + log0(f"Code size: {code_bytes} bytes") + sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} + # GPTQ quantization using Hessians collected from training data. + # salvage_v2 mixed-int policy: int5 for mlp_up_bank/mlp_down_bank (most quant-tolerant per + # 11-day collate), int6 for qo_bank/kv_bank/embed (attention + token embed kept at int6 for + # quant safety; matches seed for embed). See `mixed_quantize_int6_gptq` docstring. + quant_result, quant_meta = mixed_quantize_int6_gptq( + sd_cpu, + int6_cats={"qo", "kv", "attn_other", "mlp_other", "aux", "embed"}, + hessians=gptq_hessians, + int5_cats={"mlp_down", "mlp_up"}, + ) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = _compress_blob(quant_raw) + if master_process: + with open("final_model.int6.ptz", "wb") as f: + f.write(quant_blob) + quant_file_bytes = len(quant_blob) + log0(f"Serialized model mixed_int5_int6_int8+{_COMPRESSOR}: {quant_file_bytes} bytes") + log0(f"Total submission size mixed_int5_int6_int8+{_COMPRESSOR}: {quant_file_bytes + code_bytes} bytes") + if distributed: + dist.barrier() + with open("final_model.int6.ptz", "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(_decompress_blob(quant_blob_disk)), + map_location="cpu", + ) + deq_state = dequantize_mixed_int6(quant_state["w"], quant_state["m"], sd_cpu) + eval_model = GPT( + vocab_size=args.vocab_size, num_layers=args.num_layers, model_dim=args.model_dim, + num_heads=args.num_heads, num_kv_heads=args.num_kv_heads, mlp_mult=args.mlp_mult, + tie_embeddings=args.tie_embeddings, tied_embed_init_std=args.tied_embed_init_std, + logit_softcap=args.logit_softcap, rope_base=args.rope_base, qk_gain_init=args.qk_gain_init, + mtp_num_heads=0, mtp_loss_weight=0.0, + bigram_vocab_size=args.bigram_vocab_size, bigram_dim=args.bigram_dim, + xsa_last_n=args.xsa_last_n, rope_dims=args.rope_dims, ln_scale=args.ln_scale, + dtg=args.dtg_enabled, ve_enabled=args.ve_enabled, ve_dim=args.ve_dim, ve_layers=args.ve_layers, + gated_attention=args.gated_attention, value_residual=args.value_residual, + ).to(device).bfloat16() + eval_model.qo_bank.data = eval_model.qo_bank.data.float() + eval_model.kv_bank.data = eval_model.kv_bank.data.float() + eval_model.mlp_up_bank.data = eval_model.mlp_up_bank.data.float() + eval_model.mlp_down_bank.data = eval_model.mlp_down_bank.data.float() + for m in eval_model.modules(): + if isinstance(m, CastedLinear): + m.float() + restore_low_dim_params_to_fp32(eval_model) + eval_model.load_state_dict(deq_state, strict=True) + torch.cuda.synchronize() + t_qeval = time.perf_counter() + q_val_loss, q_val_bpb = eval_val( + args, eval_model, rank, world_size, device, grad_accum_steps, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + ) + torch.cuda.synchronize() + log0( + f"final_int6_roundtrip val_loss:{q_val_loss:.4f} val_bpb:{q_val_bpb:.4f} " + f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" + ) + log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + del eval_model, deq_state, quant_state, sd_cpu + torch.cuda.empty_cache() + sw_seq_len = effective_eval_seq_len + if args.skip_final_eval: + log0("final_eval:skipped sliding/ngram by SKIP_FINAL_EVAL=1") + else: + if args.eval_stride > 0 and args.eval_stride < sw_seq_len: + torch.cuda.synchronize() + t_slide = time.perf_counter() + sw_val_loss, sw_val_bpb = eval_val_sliding( + args, base_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_sliding_window val_loss:{sw_val_loss:.4f} val_bpb:{sw_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_slide):.0f}ms" + ) + log0(f"final_sliding_window_exact val_loss:{sw_val_loss:.8f} val_bpb:{sw_val_bpb:.8f}") + if args.eval_stride != 64 and 64 < sw_seq_len: + torch.cuda.synchronize() + t_slide64 = time.perf_counter() + sw64_val_loss, sw64_val_bpb = eval_val_sliding( + args, base_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=64, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_sliding_window_s64 val_loss:{sw64_val_loss:.4f} val_bpb:{sw64_val_bpb:.4f} " + f"stride:64 eval_time:{1000.0 * (time.perf_counter() - t_slide64):.0f}ms" + ) + log0(f"final_sliding_window_s64_exact val_loss:{sw64_val_loss:.8f} val_bpb:{sw64_val_bpb:.8f}") + if args.ngram_eval_order >= 2: + if distributed: + dist.barrier() + torch.cuda.synchronize() + t_ng = time.perf_counter() + ng_loss, ng_bpb, ng_coverage = eval_val_sliding_hashed_ngram( + args, + base_model, + rank, + world_size, + device, + val_tokens, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + stride=args.eval_stride, + order=args.ngram_eval_order, + alpha=args.ngram_eval_alpha, + min_count=args.ngram_eval_min_count, + buckets=args.ngram_eval_buckets, + max_seconds=args.ngram_eval_max_seconds, + eval_seq_len=sw_seq_len, + ) + if rank == 0: + torch.cuda.synchronize() + ng_eval_ms = 1000.0 * (time.perf_counter() - t_ng) + if ng_coverage >= 0.999999: + log0( + f"final_sliding_window_ngram{args.ngram_eval_order} val_loss:{ng_loss:.4f} " + f"val_bpb:{ng_bpb:.4f} eval_time:{ng_eval_ms:.0f}ms" + ) + log0( + f"final_sliding_window_ngram{args.ngram_eval_order}_exact " + f"val_loss:{ng_loss:.8f} val_bpb:{ng_bpb:.8f}" + ) + else: + log0( + f"final_sliding_window_ngram{args.ngram_eval_order}_partial val_loss:{ng_loss:.4f} " + f"val_bpb:{ng_bpb:.4f} coverage:{ng_coverage:.4f} eval_time:{ng_eval_ms:.0f}ms" + ) + log0( + f"final_sliding_window_ngram{args.ngram_eval_order}_partial_exact " + f"val_loss:{ng_loss:.8f} val_bpb:{ng_bpb:.8f} coverage:{ng_coverage:.8f}" + ) + if distributed: + dist.barrier() + if distributed: + dist.destroy_process_group() +if __name__ == "__main__": + main() + +==================================================================================================== +condition_id:rascal_4k_12L_brotli_mixed_8x_seed444 +run_label:salvage_v2 source_record:rascal_4k_8x_seed444_run20260427 axis:depth_12L+brotli+mixed_int +changed_fields:num_layers (11->12), compression (zstd->brotli+bshf), quant_policy (uniform_int6->mixed_int5_int6_int8) +expected_metric:final_sliding_window_exact comparator:0.8672_4k_8x_oversize_run prior_size:17766043_target:<16000000 +condition:DATA_PATH=./data/datasets/fineweb10B_sp4096 +condition:TOKENIZER_PATH=./data/tokenizers/fineweb_4096_bpe.model +condition:VOCAB_SIZE=4096 +condition:SEED=444 +condition:MAX_WALLCLOCK_SECONDS=600.0 +condition:LOADER_MODE=coprime +condition:COPRIME_MAX_LOADED_SHARDS=143 +condition:COPRIME_SHARDS_PER_BATCH=1 +condition:COPRIME_SHARD_HOLD_STEPS=64 +condition:SKIP_GPTQ=1 +condition:TRIGRAM=0 +condition:NGRAM_EVAL_ORDER=0 +Running Python 3.12.13 | packaged by conda-forge | (main, Mar 5 2026, 16:50:00) [GCC 14.3.0] +Running PyTorch 2.11.0+cu130 +Mon Apr 27 03:26:40 2026 ++-----------------------------------------------------------------------------------------+ +| NVIDIA-SMI 580.95.05 Driver Version: 580.95.05 CUDA Version: 13.0 | ++-----------------------------------------+------------------------+----------------------+ +| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | +| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | +| | | MIG M. | +|=========================================+========================+======================| +| 0 NVIDIA H100 80GB HBM3 On | 00000000:04:00.0 Off | 0 | +| N/A 41C P0 116W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 1 NVIDIA H100 80GB HBM3 On | 00000000:05:00.0 Off | 0 | +| N/A 36C P0 119W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 2 NVIDIA H100 80GB HBM3 On | 00000000:0A:00.0 Off | 0 | +| N/A 41C P0 121W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 3 NVIDIA H100 80GB HBM3 On | 00000000:0B:00.0 Off | 0 | +| N/A 36C P0 121W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 4 NVIDIA H100 80GB HBM3 On | 00000000:84:00.0 Off | 0 | +| N/A 40C P0 118W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 5 NVIDIA H100 80GB HBM3 On | 00000000:85:00.0 Off | 0 | +| N/A 35C P0 116W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 6 NVIDIA H100 80GB HBM3 On | 00000000:8A:00.0 Off | 0 | +| N/A 39C P0 121W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ +| 7 NVIDIA H100 80GB HBM3 On | 00000000:8B:00.0 Off | 0 | +| N/A 35C P0 116W / 700W | 1505MiB / 81559MiB | 0% Default | +| | | Disabled | ++-----------------------------------------+------------------------+----------------------+ + ++-----------------------------------------------------------------------------------------+ +| Processes: | +| GPU GI CI PID Type Process name GPU Memory | +| ID ID Usage | +|=========================================================================================| +| No running processes found | ++-----------------------------------------------------------------------------------------+ + +==================================================================================================== +val_bpb:enabled tokenizer_kind=sentencepiece tokenizer_path=./data/tokenizers/fineweb_4096_bpe.model +train_loader:dataset:fineweb10B_sp4096 train_shards:143 +val_loader:shards pattern=./data/datasets/fineweb10B_sp4096/fineweb_val_*.bin tokens:45514752 +model_params:31321700 +mtp_num_heads:0 mtp_loss_weight:0.2 mtp_params:0 +XSA:last_11 active_layers:[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11] +world_size:8 grad_accum_steps:1 +sdp_backends:cudnn=False flash=True mem_efficient=False math=False +attention_mode:gqa num_heads:8 num_kv_heads:4 +tie_embeddings:True embed_lr:0.035 head_lr:0.0 matrix_lr:0.025 scalar_lr:0.025 +train_batch_tokens:786432 train_seq_len:2048 iterations:20000 warmup_steps:20 max_wallclock_seconds:600.000 +compile:enabled=1 mode:default fullgraph=1 +mlp_kernel_mode:eager +scale_init:attn=1.0000 mlp=1.0000 resid_mix=(1.0000,0.0000) ln_scale=1 +seed:444 +loader:coprime shards:143 blocks:6967965 seq_len:2048 shards_per_batch:1 cache:143 batch_stride:46 hold_steps:64 +warmup_step:1/20 +warmup_step:2/20 +warmup_step:3/20 +warmup_step:4/20 +warmup_step:5/20 +warmup_step:6/20 +warmup_step:7/20 +warmup_step:8/20 +warmup_step:9/20 +warmup_step:10/20 +warmup_step:11/20 +warmup_step:12/20 +warmup_step:13/20 +warmup_step:14/20 +warmup_step:15/20 +warmup_step:16/20 +warmup_step:17/20 +warmup_step:18/20 +warmup_step:19/20 +warmup_step:20/20 +loader_reset:loader:coprime shards:143 blocks:6967965 seq_len:2048 shards_per_batch:1 cache:143 batch_stride:46 hold_steps:64 +step:0/20000 val_loss:8.3098 val_bpb:2.8333 train_time:0ms step_avg:0.01ms tok/s:0 +step:1/20000 train_loss:8.3105 train_time:360ms step_avg:360.42ms tok/s:2181977 +step:2/20000 train_loss:9.4377 train_time:411ms step_avg:205.53ms tok/s:3826435 +step:3/20000 train_loss:8.8185 train_time:512ms step_avg:170.64ms tok/s:4608598 +step:4/20000 train_loss:7.9978 train_time:614ms step_avg:153.57ms tok/s:5121109 +step:5/20000 train_loss:8.1092 train_time:715ms step_avg:143.09ms tok/s:5496220 +step:6/20000 train_loss:8.1789 train_time:817ms step_avg:136.19ms tok/s:5774519 +step:7/20000 train_loss:7.9287 train_time:919ms step_avg:131.25ms tok/s:5991643 +step:8/20000 train_loss:7.7197 train_time:1020ms step_avg:127.54ms tok/s:6166189 +step:9/20000 train_loss:7.5380 train_time:1121ms step_avg:124.60ms tok/s:6311749 +step:10/20000 train_loss:7.5174 train_time:1223ms step_avg:122.31ms tok/s:6429901 +step:500/20000 train_loss:3.0750 train_time:52893ms step_avg:105.79ms tok/s:7434165 +step:1000/20000 train_loss:2.9603 train_time:105882ms step_avg:105.88ms tok/s:7427432 +step:1500/20000 train_loss:2.8906 train_time:158929ms step_avg:105.95ms tok/s:7422479 +step:2000/20000 train_loss:2.8656 train_time:211956ms step_avg:105.98ms tok/s:7420719 +step:2500/20000 train_loss:2.7443 train_time:265011ms step_avg:106.00ms tok/s:7418864 +step:3000/20000 train_loss:2.7167 train_time:317872ms step_avg:105.96ms tok/s:7422150 +step:3500/20000 train_loss:2.7207 train_time:370962ms step_avg:105.99ms tok/s:7419927 +step:4000/20000 train_loss:2.6442 train_time:424059ms step_avg:106.01ms tok/s:7418130 +step:4000/20000 val_loss:2.6752 val_bpb:0.9121 train_time:424113ms step_avg:106.03ms tok/s:7417191 +step:4500/20000 train_loss:2.6040 train_time:477126ms step_avg:106.03ms tok/s:7417210 +swa:start step:5000 +step:5000/20000 train_loss:2.6335 train_time:530182ms step_avg:106.04ms tok/s:7416616 +late_qat:enabled step:5131 scale:0.1497 +step:5500/20000 train_loss:2.5906 train_time:583846ms step_avg:106.15ms tok/s:7408420 +step:5649/20000 val_loss:2.5794 val_bpb:0.8795 train_time:600083ms step_avg:106.23ms tok/s:7403235 +stopping_early: wallclock_cap train_time:600083ms step:5649/20000 +peak memory allocated: 25261 MiB reserved: 25806 MiB +gptq:SKIPPED (SKIP_GPTQ=1) — will use naive int6 +ema:applying EMA weights +DIAGNOSTIC post_ema val_loss:2.5768 val_bpb:0.8786 eval_time:1717ms +Serialized model: 119539322 bytes +Code size: 128464 bytes +Serialized model mixed_int5_int6_int8+brotli: 15525048 bytes +Total submission size mixed_int5_int6_int8+brotli: 15653512 bytes +final_int6_roundtrip val_loss:2.6681 val_bpb:0.9097 eval_time:5059ms +final_int6_roundtrip_exact val_loss:2.66809399 val_bpb:0.90971723 +final_sliding_window val_loss:2.5352 val_bpb:0.8644 stride:64 eval_time:77473ms +final_sliding_window_exact val_loss:2.53522743 val_bpb:0.86441066 From d8e98d5b61b5edf53be08c5159a8dec87c3b572e Mon Sep 17 00:00:00 2001 From: Octavian Date: Mon, 27 Apr 2026 11:20:34 -0500 Subject: [PATCH 2/3] Mikey: val_bpb 0.87980 (3-seed mean), 15.6MB legal Co-Authored-By: Claude Opus 4.7 (1M context) --- .../2026-04-27_Mikey/README.md | 12 +- .../2026-04-27_Mikey/submission.json | 10 +- .../2026-04-27_Mikey/train_gpt.py | 194 +++++++++-- .../2026-04-27_Mikey/train_seed300.log | 326 ++++++++++++----- .../2026-04-27_Mikey/train_seed42.log | 328 +++++++++++++----- .../2026-04-27_Mikey/train_seed444.log | 328 +++++++++++++----- 6 files changed, 910 insertions(+), 288 deletions(-) diff --git a/records/track_10min_16mb/2026-04-27_Mikey/README.md b/records/track_10min_16mb/2026-04-27_Mikey/README.md index 51e267444d..546aa58a37 100644 --- a/records/track_10min_16mb/2026-04-27_Mikey/README.md +++ b/records/track_10min_16mb/2026-04-27_Mikey/README.md @@ -1,11 +1,11 @@ # Mikey -| seed | val_bpb (sliding) | bytes | -|------|-------------------|-------| -| 42 | 0.86503709 | 15,639,737 | -| 300 | 0.86698133 | 15,594,375 | -| 444 | 0.86441066 | 15,653,512 | -| **mean** | **0.86547636** | — | +| seed | val_bpb | bytes | +|------|---------|-------| +| 42 | 0.87994906 | 15,626,464 | +| 300 | 0.88084235 | 15,601,472 | +| 444 | 0.87861866 | 15,671,210 | +| **mean** | **0.87980336** | — | ``` torchrun --standalone --nproc_per_node=8 train_gpt.py diff --git a/records/track_10min_16mb/2026-04-27_Mikey/submission.json b/records/track_10min_16mb/2026-04-27_Mikey/submission.json index 1970614510..4f67c3d067 100644 --- a/records/track_10min_16mb/2026-04-27_Mikey/submission.json +++ b/records/track_10min_16mb/2026-04-27_Mikey/submission.json @@ -4,13 +4,13 @@ "name": "Mikey", "date": "2026-04-27", "track": "10min_16mb", - "val_bpb": 0.86547636, - "val_bpb_std": 0.00109, + "val_bpb": 0.87980336, + "val_bpb_std": 0.00112, "seeds": [42, 300, 444], "seed_results": { - "42": {"val_bpb": 0.86503709, "artifact_bytes": 15639737}, - "300": {"val_bpb": 0.86698133, "artifact_bytes": 15594375}, - "444": {"val_bpb": 0.86441066, "artifact_bytes": 15653512} + "42": {"val_bpb": 0.87994906, "artifact_bytes": 15626464}, + "300": {"val_bpb": 0.88084235, "artifact_bytes": 15601472}, + "444": {"val_bpb": 0.87861866, "artifact_bytes": 15671210} }, "hardware": "8xH100 80GB SXM", "pytorch_version": "2.11.0+cu130", diff --git a/records/track_10min_16mb/2026-04-27_Mikey/train_gpt.py b/records/track_10min_16mb/2026-04-27_Mikey/train_gpt.py index cd29d607f9..fa88d152b9 100644 --- a/records/track_10min_16mb/2026-04-27_Mikey/train_gpt.py +++ b/records/track_10min_16mb/2026-04-27_Mikey/train_gpt.py @@ -31,7 +31,7 @@ from flash_attn_interface import flash_attn_func as flash_attn_3_func except ImportError: flash_attn_3_func = None -# Compression: brotli-11 + byte-shuffle is the salvage_v2 preferred path (PR #1493 recipe). +# Compression: brotli-11 + byte-shuffle. # Falls back to zstd then zlib so this file still runs if brotli isn't installed. _brotli_module = None _zstandard_module = None @@ -55,7 +55,7 @@ zstandard = _zstandard_module if _zlib_module is None: import zlib as _zlib_module # always available; used by zlib fallback path -# --- Byte-shuffle (de-interleave) wrapper from PR #1493: improves brotli ratio on quantized payloads. --- +# --- Byte-shuffle (de-interleave) wrapper: improves brotli ratio on quantized payloads. --- _BSHF_MAGIC = b"BSHF" def _byte_shuffle(data: bytes, stride: int = 2) -> bytes: if stride <= 1 or len(data) < stride: @@ -186,6 +186,11 @@ class Hyperparameters: ngram_eval_max_seconds = float(os.environ.get("NGRAM_EVAL_MAX_SECONDS", 0.0)) ngram_entropy_shift = bool(int(os.environ.get("NGRAM_ENTROPY_SHIFT", "0"))) ngram_order_mults_str = os.environ.get("NGRAM_ORDER_MULTS", "") + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.005)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) cubric_cadence = int(os.environ.get("CUBRIC_CADENCE", 0)) skip_final_eval = bool(int(os.environ.get("SKIP_FINAL_EVAL", "0"))) post_ema_diagnostic = bool(int(os.environ.get("POST_EMA_DIAGNOSTIC", "1"))) @@ -768,14 +773,12 @@ def _classify_param_fine(name: str) -> str: if ".attn." in name or (".proj." in name and ".mlp." not in name): return "attn_other" return "other" -# salvage_v2 mixed-int policy (applied at the Tensor level via _classify_param_fine): -# - mlp_down_bank (mlp_proj, MOST quant-tolerant per 11-day collate) -> int5 (clip_range 15) -# - mlp_up_bank (mlp_fc, also tolerant) -> int5 +# Mixed-int policy (applied at the Tensor level via _classify_param_fine): +# - mlp_down_bank (mlp_proj, most quant-tolerant) -> int5 (clip_range 15) +# - mlp_up_bank (mlp_fc, also tolerant) -> int5 # - qo_bank, kv_bank (attention; LEAST quant-tolerant) -> int6 (clip_range 31) -# - tok_emb / lm_head (embed) -> int6 (matches seed; keeping -# attn/embed at int6 for -# quant safety on this -# first salvage attempt) +# - tok_emb / lm_head (embed) -> int6 (attn/embed at int6 for +# quant safety) # Bytes savings: int5 keeps the int8 storage container (no bit-packing) but the high 3 bits are # forced zero, giving brotli a compressible pattern. Combined with the byte-shuffle wrapper, # expected savings vs uniform-int6+zstd is roughly the int5 bit ratio (5/6 = -17%) APPLIED only @@ -791,7 +794,7 @@ def mixed_quantize_int6_gptq(state_dict: dict[str, Tensor], int6_cats: set[str], (qo, kv, mlp_up, mlp_down, attn_other, mlp_other, aux, embed). For backwards-compat with the old uniform-int6 caller, the legacy coarse names {'mlp','attn','aux','embed'} are also accepted in `int6_cats` and expand to their fine-grained children. `int5_cats` always uses - fine names. If `int5_cats` is None, defaults to DEFAULT_INT5_CATS (the salvage_v2 policy).""" + fine names. If `int5_cats` is None, defaults to DEFAULT_INT5_CATS.""" if int5_cats is None: int5_cats = set(DEFAULT_INT5_CATS) # Expand legacy coarse names so the existing call signature keeps working. @@ -1653,7 +1656,7 @@ def eval_val_sliding_hashed_ngram( batch_seqs: int = 128, eval_seq_len: int | None = None, ) -> tuple[float, float, float]: - """Score-first sliding eval with chunk-based SHARED n-gram tables + cubric. + """Sliding eval with chunk-based SHARED n-gram tables + cubric. Key design: all ranks share identical n-gram tables via bulk chunk updates. Each chunk's windows are distributed across ranks for scoring, then ALL ranks @@ -2035,6 +2038,130 @@ def eval_val_sliding( return val_loss, bits_per_token * tokens_per_byte +def eval_val_ttt( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + context_size = seq_len - stride + window_starts = [ws for ws in range(0, total_tokens, stride) + if ws + context_size < total_tokens] + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + s = 0 if ws == 0 else context_size + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + compiled_logits = maybe_compile( + base_model.forward_logits, + enabled=args.compile_enabled, + fullgraph=args.compile_fullgraph, + ) + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + ttt_params = [p for p in base_model.parameters()] + for p in ttt_params: + p.requires_grad_(True) + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + base_model.eval() + with torch.no_grad(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else context_size + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + is_last_chunk = ci == num_chunks - 1 + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + for pg in optimizer.param_groups: + pg["lr"] = cos_lr + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, batch_seqs): + be = min(bs + batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1 and dist.is_available() and dist.is_initialized(): + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, 1.0) + optimizer.step() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.eval() + return val_loss, bits_per_token * tokens_per_byte + + # --- Training --- def main() -> None: @@ -2082,10 +2209,6 @@ def log0(msg: str, console: bool = True) -> None: print(msg, file=f) log0(code, console=False) log0("=" * 100, console=False) - log0("condition_id:mikey_8x_seed444") - log0("run_label:salvage_v2 source_record:mikey_origin_run_20260427 axis:depth_12L+brotli+mixed_int") - log0("changed_fields:num_layers (11->12), compression (zstd->brotli+bshf), quant_policy (uniform_int6->mixed_int5_int6_int8)") - log0("expected_metric:final_sliding_window_exact comparator:0.8672_4k_8x_oversize_run prior_size:17766043_target:<16000000") log0(f"condition:DATA_PATH={args.data_path}") log0(f"condition:TOKENIZER_PATH={args.tokenizer_path}") log0(f"condition:VOCAB_SIZE={args.vocab_size}") @@ -2095,6 +2218,11 @@ def log0(msg: str, console: bool = True) -> None: log0(f"condition:COPRIME_MAX_LOADED_SHARDS={args.coprime_max_loaded_shards}") log0(f"condition:COPRIME_SHARDS_PER_BATCH={args.coprime_shards_per_batch}") log0(f"condition:COPRIME_SHARD_HOLD_STEPS={args.coprime_shard_hold_steps}") + log0(f"condition:TTT_ENABLED={int(args.ttt_enabled)}") + log0(f"condition:TTT_LR={args.ttt_lr}") + log0(f"condition:TTT_EPOCHS={args.ttt_epochs}") + log0(f"condition:TTT_MOMENTUM={args.ttt_momentum}") + log0(f"condition:TTT_CHUNK_TOKENS={args.ttt_chunk_tokens}") log0(f"condition:SKIP_GPTQ={os.environ.get('SKIP_GPTQ', '1')}") log0(f"condition:TRIGRAM={int(args.trigram_enabled)}") log0(f"condition:NGRAM_EVAL_ORDER={args.ngram_eval_order}") @@ -2507,9 +2635,9 @@ def lr_mul(step: int, elapsed_ms: float) -> float: log0(f"Code size: {code_bytes} bytes") sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} # GPTQ quantization using Hessians collected from training data. - # salvage_v2 mixed-int policy: int5 for mlp_up_bank/mlp_down_bank (most quant-tolerant per - # 11-day collate), int6 for qo_bank/kv_bank/embed (attention + token embed kept at int6 for - # quant safety; matches seed for embed). See `mixed_quantize_int6_gptq` docstring. + # Mixed-int policy: int5 for mlp_up_bank/mlp_down_bank (most quant-tolerant), + # int6 for qo_bank/kv_bank/embed (attention + token embed kept at int6 for + # quant safety). See `mixed_quantize_int6_gptq` docstring. quant_result, quant_meta = mixed_quantize_int6_gptq( sd_cpu, int6_cats={"qo", "kv", "attn_other", "mlp_other", "aux", "embed"}, @@ -2567,8 +2695,6 @@ def lr_mul(step: int, elapsed_ms: float) -> float: f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" ) log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") - del eval_model, deq_state, quant_state, sd_cpu - torch.cuda.empty_cache() sw_seq_len = effective_eval_seq_len if args.skip_final_eval: log0("final_eval:skipped sliding/ngram by SKIP_FINAL_EVAL=1") @@ -2577,7 +2703,7 @@ def lr_mul(step: int, elapsed_ms: float) -> float: torch.cuda.synchronize() t_slide = time.perf_counter() sw_val_loss, sw_val_bpb = eval_val_sliding( - args, base_model, rank, world_size, device, + args, eval_model, rank, world_size, device, val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, stride=args.eval_stride, eval_seq_len=sw_seq_len, @@ -2592,7 +2718,7 @@ def lr_mul(step: int, elapsed_ms: float) -> float: torch.cuda.synchronize() t_slide64 = time.perf_counter() sw64_val_loss, sw64_val_bpb = eval_val_sliding( - args, base_model, rank, world_size, device, + args, eval_model, rank, world_size, device, val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, stride=64, eval_seq_len=sw_seq_len, @@ -2610,7 +2736,7 @@ def lr_mul(step: int, elapsed_ms: float) -> float: t_ng = time.perf_counter() ng_loss, ng_bpb, ng_coverage = eval_val_sliding_hashed_ngram( args, - base_model, + eval_model, rank, world_size, device, @@ -2649,6 +2775,28 @@ def lr_mul(step: int, elapsed_ms: float) -> float: ) if distributed: dist.barrier() + if args.ttt_enabled and args.eval_stride > 0 and args.eval_stride < sw_seq_len: + if distributed: + dist.barrier() + torch.cuda.synchronize() + t_ttt = time.perf_counter() + log0(f"ttt:start chunks:{(val_tokens.numel() - 1 + args.ttt_chunk_tokens - 1) // args.ttt_chunk_tokens} ttt_lr:{args.ttt_lr} ttt_epochs:{args.ttt_epochs} chunk_tokens:{args.ttt_chunk_tokens}") + ttt_val_loss, ttt_val_bpb = eval_val_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_sliding_window_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" + ) + log0(f"final_sliding_window_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + del eval_model, deq_state, quant_state, sd_cpu + torch.cuda.empty_cache() if distributed: dist.destroy_process_group() if __name__ == "__main__": diff --git a/records/track_10min_16mb/2026-04-27_Mikey/train_seed300.log b/records/track_10min_16mb/2026-04-27_Mikey/train_seed300.log index 0097b6c4cc..8e199ca6c3 100644 --- a/records/track_10min_16mb/2026-04-27_Mikey/train_seed300.log +++ b/records/track_10min_16mb/2026-04-27_Mikey/train_seed300.log @@ -19,7 +19,7 @@ import torch.nn.functional as F from torch import Tensor, nn from torch.nn.parallel import DistributedDataParallel as DDP -os.environ.setdefault("RUN_ID", "rascal_4k_12L_brotli_mixed_8x_seed444_" + time.strftime("%Y%m%d_%H%M%S")) +os.environ.setdefault("RUN_ID", "mikey_8x_seed444_" + time.strftime("%Y%m%d_%H%M%S")) try: import triton @@ -31,7 +31,7 @@ try: from flash_attn_interface import flash_attn_func as flash_attn_3_func except ImportError: flash_attn_3_func = None -# Compression: brotli-11 + byte-shuffle is the salvage_v2 preferred path (PR #1493 recipe). +# Compression: brotli-11 + byte-shuffle. # Falls back to zstd then zlib so this file still runs if brotli isn't installed. _brotli_module = None _zstandard_module = None @@ -55,7 +55,7 @@ if _zstandard_module is not None: zstandard = _zstandard_module if _zlib_module is None: import zlib as _zlib_module # always available; used by zlib fallback path -# --- Byte-shuffle (de-interleave) wrapper from PR #1493: improves brotli ratio on quantized payloads. --- +# --- Byte-shuffle (de-interleave) wrapper: improves brotli ratio on quantized payloads. --- _BSHF_MAGIC = b"BSHF" def _byte_shuffle(data: bytes, stride: int = 2) -> bytes: if stride <= 1 or len(data) < stride: @@ -186,6 +186,11 @@ class Hyperparameters: ngram_eval_max_seconds = float(os.environ.get("NGRAM_EVAL_MAX_SECONDS", 0.0)) ngram_entropy_shift = bool(int(os.environ.get("NGRAM_ENTROPY_SHIFT", "0"))) ngram_order_mults_str = os.environ.get("NGRAM_ORDER_MULTS", "") + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.005)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) cubric_cadence = int(os.environ.get("CUBRIC_CADENCE", 0)) skip_final_eval = bool(int(os.environ.get("SKIP_FINAL_EVAL", "0"))) post_ema_diagnostic = bool(int(os.environ.get("POST_EMA_DIAGNOSTIC", "1"))) @@ -768,14 +773,12 @@ def _classify_param_fine(name: str) -> str: if ".attn." in name or (".proj." in name and ".mlp." not in name): return "attn_other" return "other" -# salvage_v2 mixed-int policy (applied at the Tensor level via _classify_param_fine): -# - mlp_down_bank (mlp_proj, MOST quant-tolerant per 11-day collate) -> int5 (clip_range 15) -# - mlp_up_bank (mlp_fc, also tolerant) -> int5 +# Mixed-int policy (applied at the Tensor level via _classify_param_fine): +# - mlp_down_bank (mlp_proj, most quant-tolerant) -> int5 (clip_range 15) +# - mlp_up_bank (mlp_fc, also tolerant) -> int5 # - qo_bank, kv_bank (attention; LEAST quant-tolerant) -> int6 (clip_range 31) -# - tok_emb / lm_head (embed) -> int6 (matches seed; keeping -# attn/embed at int6 for -# quant safety on this -# first salvage attempt) +# - tok_emb / lm_head (embed) -> int6 (attn/embed at int6 for +# quant safety) # Bytes savings: int5 keeps the int8 storage container (no bit-packing) but the high 3 bits are # forced zero, giving brotli a compressible pattern. Combined with the byte-shuffle wrapper, # expected savings vs uniform-int6+zstd is roughly the int5 bit ratio (5/6 = -17%) APPLIED only @@ -791,7 +794,7 @@ def mixed_quantize_int6_gptq(state_dict: dict[str, Tensor], int6_cats: set[str], (qo, kv, mlp_up, mlp_down, attn_other, mlp_other, aux, embed). For backwards-compat with the old uniform-int6 caller, the legacy coarse names {'mlp','attn','aux','embed'} are also accepted in `int6_cats` and expand to their fine-grained children. `int5_cats` always uses - fine names. If `int5_cats` is None, defaults to DEFAULT_INT5_CATS (the salvage_v2 policy).""" + fine names. If `int5_cats` is None, defaults to DEFAULT_INT5_CATS.""" if int5_cats is None: int5_cats = set(DEFAULT_INT5_CATS) # Expand legacy coarse names so the existing call signature keeps working. @@ -1653,7 +1656,7 @@ def eval_val_sliding_hashed_ngram( batch_seqs: int = 128, eval_seq_len: int | None = None, ) -> tuple[float, float, float]: - """Score-first sliding eval with chunk-based SHARED n-gram tables + cubric. + """Sliding eval with chunk-based SHARED n-gram tables + cubric. Key design: all ranks share identical n-gram tables via bulk chunk updates. Each chunk's windows are distributed across ranks for scoring, then ALL ranks @@ -2035,6 +2038,130 @@ def eval_val_sliding( return val_loss, bits_per_token * tokens_per_byte +def eval_val_ttt( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + context_size = seq_len - stride + window_starts = [ws for ws in range(0, total_tokens, stride) + if ws + context_size < total_tokens] + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + s = 0 if ws == 0 else context_size + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + compiled_logits = maybe_compile( + base_model.forward_logits, + enabled=args.compile_enabled, + fullgraph=args.compile_fullgraph, + ) + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + ttt_params = [p for p in base_model.parameters()] + for p in ttt_params: + p.requires_grad_(True) + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + base_model.eval() + with torch.no_grad(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else context_size + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + is_last_chunk = ci == num_chunks - 1 + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + for pg in optimizer.param_groups: + pg["lr"] = cos_lr + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, batch_seqs): + be = min(bs + batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1 and dist.is_available() and dist.is_initialized(): + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, 1.0) + optimizer.step() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.eval() + return val_loss, bits_per_token * tokens_per_byte + + # --- Training --- def main() -> None: @@ -2047,8 +2174,8 @@ def main() -> None: local_rank = int(os.environ.get("LOCAL_RANK", "0")) if world_size != 8: raise ValueError( - f"Rascal 4k 12L brotli+mixed 8x requires WORLD_SIZE=8, got {world_size}. " - "Launch with: torchrun --standalone --nproc_per_node=8 4k_vocab_rascal_12l_brotli_mixed/train_gpt_4K_12L_brotli_mixed_8xgpu.py" + f"Mikey 8x requires WORLD_SIZE=8, got {world_size}. " + "Launch with: torchrun --standalone --nproc_per_node=8 Mikey/train_gpt_8xgpu.py" ) grad_accum_steps = 8 // world_size grad_scale = 1.0 / grad_accum_steps @@ -2082,10 +2209,6 @@ def main() -> None: print(msg, file=f) log0(code, console=False) log0("=" * 100, console=False) - log0("condition_id:rascal_4k_12L_brotli_mixed_8x_seed444") - log0("run_label:salvage_v2 source_record:rascal_4k_8x_seed444_run20260427 axis:depth_12L+brotli+mixed_int") - log0("changed_fields:num_layers (11->12), compression (zstd->brotli+bshf), quant_policy (uniform_int6->mixed_int5_int6_int8)") - log0("expected_metric:final_sliding_window_exact comparator:0.8672_4k_8x_oversize_run prior_size:17766043_target:<16000000") log0(f"condition:DATA_PATH={args.data_path}") log0(f"condition:TOKENIZER_PATH={args.tokenizer_path}") log0(f"condition:VOCAB_SIZE={args.vocab_size}") @@ -2095,6 +2218,11 @@ def main() -> None: log0(f"condition:COPRIME_MAX_LOADED_SHARDS={args.coprime_max_loaded_shards}") log0(f"condition:COPRIME_SHARDS_PER_BATCH={args.coprime_shards_per_batch}") log0(f"condition:COPRIME_SHARD_HOLD_STEPS={args.coprime_shard_hold_steps}") + log0(f"condition:TTT_ENABLED={int(args.ttt_enabled)}") + log0(f"condition:TTT_LR={args.ttt_lr}") + log0(f"condition:TTT_EPOCHS={args.ttt_epochs}") + log0(f"condition:TTT_MOMENTUM={args.ttt_momentum}") + log0(f"condition:TTT_CHUNK_TOKENS={args.ttt_chunk_tokens}") log0(f"condition:SKIP_GPTQ={os.environ.get('SKIP_GPTQ', '1')}") log0(f"condition:TRIGRAM={int(args.trigram_enabled)}") log0(f"condition:NGRAM_EVAL_ORDER={args.ngram_eval_order}") @@ -2507,9 +2635,9 @@ def main() -> None: log0(f"Code size: {code_bytes} bytes") sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} # GPTQ quantization using Hessians collected from training data. - # salvage_v2 mixed-int policy: int5 for mlp_up_bank/mlp_down_bank (most quant-tolerant per - # 11-day collate), int6 for qo_bank/kv_bank/embed (attention + token embed kept at int6 for - # quant safety; matches seed for embed). See `mixed_quantize_int6_gptq` docstring. + # Mixed-int policy: int5 for mlp_up_bank/mlp_down_bank (most quant-tolerant), + # int6 for qo_bank/kv_bank/embed (attention + token embed kept at int6 for + # quant safety). See `mixed_quantize_int6_gptq` docstring. quant_result, quant_meta = mixed_quantize_int6_gptq( sd_cpu, int6_cats={"qo", "kv", "attn_other", "mlp_other", "aux", "embed"}, @@ -2567,8 +2695,6 @@ def main() -> None: f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" ) log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") - del eval_model, deq_state, quant_state, sd_cpu - torch.cuda.empty_cache() sw_seq_len = effective_eval_seq_len if args.skip_final_eval: log0("final_eval:skipped sliding/ngram by SKIP_FINAL_EVAL=1") @@ -2577,7 +2703,7 @@ def main() -> None: torch.cuda.synchronize() t_slide = time.perf_counter() sw_val_loss, sw_val_bpb = eval_val_sliding( - args, base_model, rank, world_size, device, + args, eval_model, rank, world_size, device, val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, stride=args.eval_stride, eval_seq_len=sw_seq_len, @@ -2592,7 +2718,7 @@ def main() -> None: torch.cuda.synchronize() t_slide64 = time.perf_counter() sw64_val_loss, sw64_val_bpb = eval_val_sliding( - args, base_model, rank, world_size, device, + args, eval_model, rank, world_size, device, val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, stride=64, eval_seq_len=sw_seq_len, @@ -2610,7 +2736,7 @@ def main() -> None: t_ng = time.perf_counter() ng_loss, ng_bpb, ng_coverage = eval_val_sliding_hashed_ngram( args, - base_model, + eval_model, rank, world_size, device, @@ -2649,16 +2775,33 @@ def main() -> None: ) if distributed: dist.barrier() + if args.ttt_enabled and args.eval_stride > 0 and args.eval_stride < sw_seq_len: + if distributed: + dist.barrier() + torch.cuda.synchronize() + t_ttt = time.perf_counter() + log0(f"ttt:start chunks:{(val_tokens.numel() - 1 + args.ttt_chunk_tokens - 1) // args.ttt_chunk_tokens} ttt_lr:{args.ttt_lr} ttt_epochs:{args.ttt_epochs} chunk_tokens:{args.ttt_chunk_tokens}") + ttt_val_loss, ttt_val_bpb = eval_val_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_sliding_window_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" + ) + log0(f"final_sliding_window_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + del eval_model, deq_state, quant_state, sd_cpu + torch.cuda.empty_cache() if distributed: dist.destroy_process_group() if __name__ == "__main__": main() - ==================================================================================================== -condition_id:rascal_4k_12L_brotli_mixed_8x_seed444 -run_label:salvage_v2 source_record:rascal_4k_8x_seed444_run20260427 axis:depth_12L+brotli+mixed_int -changed_fields:num_layers (11->12), compression (zstd->brotli+bshf), quant_policy (uniform_int6->mixed_int5_int6_int8) -expected_metric:final_sliding_window_exact comparator:0.8672_4k_8x_oversize_run prior_size:17766043_target:<16000000 condition:DATA_PATH=./data/datasets/fineweb10B_sp4096 condition:TOKENIZER_PATH=./data/tokenizers/fineweb_4096_bpe.model condition:VOCAB_SIZE=4096 @@ -2668,49 +2811,54 @@ condition:LOADER_MODE=coprime condition:COPRIME_MAX_LOADED_SHARDS=143 condition:COPRIME_SHARDS_PER_BATCH=1 condition:COPRIME_SHARD_HOLD_STEPS=64 +condition:TTT_ENABLED=1 +condition:TTT_LR=0.005 +condition:TTT_EPOCHS=3 +condition:TTT_MOMENTUM=0.9 +condition:TTT_CHUNK_TOKENS=32768 condition:SKIP_GPTQ=1 condition:TRIGRAM=0 condition:NGRAM_EVAL_ORDER=0 Running Python 3.12.13 | packaged by conda-forge | (main, Mar 5 2026, 16:50:00) [GCC 14.3.0] Running PyTorch 2.11.0+cu130 -Mon Apr 27 03:54:52 2026 +Mon Apr 27 15:40:37 2026 +-----------------------------------------------------------------------------------------+ -| NVIDIA-SMI 580.95.05 Driver Version: 580.95.05 CUDA Version: 13.0 | +| NVIDIA-SMI 580.126.09 Driver Version: 580.126.09 CUDA Version: 13.0 | +-----------------------------------------+------------------------+----------------------+ | GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | | | | MIG M. | |=========================================+========================+======================| -| 0 NVIDIA H100 80GB HBM3 On | 00000000:04:00.0 Off | 0 | -| N/A 40C P0 115W / 700W | 1505MiB / 81559MiB | 0% Default | +| 0 NVIDIA H100 80GB HBM3 On | 00000000:0A:00.0 Off | 0 | +| N/A 41C P0 125W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ -| 1 NVIDIA H100 80GB HBM3 On | 00000000:05:00.0 Off | 0 | -| N/A 35C P0 119W / 700W | 1505MiB / 81559MiB | 0% Default | +| 1 NVIDIA H100 80GB HBM3 On | 00000000:18:00.0 Off | 0 | +| N/A 34C P0 119W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ -| 2 NVIDIA H100 80GB HBM3 On | 00000000:0A:00.0 Off | 0 | -| N/A 39C P0 119W / 700W | 1505MiB / 81559MiB | 0% Default | +| 2 NVIDIA H100 80GB HBM3 On | 00000000:3F:00.0 Off | 0 | +| N/A 35C P0 119W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ -| 3 NVIDIA H100 80GB HBM3 On | 00000000:0B:00.0 Off | 0 | -| N/A 35C P0 119W / 700W | 1505MiB / 81559MiB | 0% Default | +| 3 NVIDIA H100 80GB HBM3 On | 00000000:48:00.0 Off | 0 | +| N/A 41C P0 127W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ -| 4 NVIDIA H100 80GB HBM3 On | 00000000:84:00.0 Off | 0 | -| N/A 39C P0 118W / 700W | 1505MiB / 81559MiB | 0% Default | +| 4 NVIDIA H100 80GB HBM3 On | 00000000:87:00.0 Off | 0 | +| N/A 41C P0 124W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ -| 5 NVIDIA H100 80GB HBM3 On | 00000000:85:00.0 Off | 0 | -| N/A 34C P0 117W / 700W | 1505MiB / 81559MiB | 0% Default | +| 5 NVIDIA H100 80GB HBM3 On | 00000000:90:00.0 Off | 0 | +| N/A 33C P0 119W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ -| 6 NVIDIA H100 80GB HBM3 On | 00000000:8A:00.0 Off | 0 | -| N/A 38C P0 120W / 700W | 1505MiB / 81559MiB | 0% Default | +| 6 NVIDIA H100 80GB HBM3 On | 00000000:BE:00.0 Off | 0 | +| N/A 34C P0 121W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ -| 7 NVIDIA H100 80GB HBM3 On | 00000000:8B:00.0 Off | 0 | -| N/A 34C P0 116W / 700W | 1505MiB / 81559MiB | 0% Default | +| 7 NVIDIA H100 80GB HBM3 On | 00000000:C7:00.0 Off | 0 | +| N/A 41C P0 124W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ @@ -2719,7 +2867,14 @@ Mon Apr 27 03:54:52 2026 | GPU GI CI PID Type Process name GPU Memory | | ID ID Usage | |=========================================================================================| -| No running processes found | +| 0 N/A N/A 14524 C /venv/main/bin/python3 1496MiB | +| 1 N/A N/A 14525 C /venv/main/bin/python3 1496MiB | +| 2 N/A N/A 14526 C /venv/main/bin/python3 1496MiB | +| 3 N/A N/A 14527 C /venv/main/bin/python3 1496MiB | +| 4 N/A N/A 14528 C /venv/main/bin/python3 1496MiB | +| 5 N/A N/A 14529 C /venv/main/bin/python3 1496MiB | +| 6 N/A N/A 14530 C /venv/main/bin/python3 1496MiB | +| 7 N/A N/A 14531 C /venv/main/bin/python3 1496MiB | +-----------------------------------------------------------------------------------------+ ==================================================================================================== @@ -2761,41 +2916,44 @@ warmup_step:19/20 warmup_step:20/20 loader_reset:loader:coprime shards:143 blocks:6967965 seq_len:2048 shards_per_batch:1 cache:143 batch_stride:5 hold_steps:64 step:0/20000 val_loss:8.3069 val_bpb:2.8323 train_time:0ms step_avg:0.01ms tok/s:0 -step:1/20000 train_loss:8.3049 train_time:357ms step_avg:356.99ms tok/s:2202934 -step:2/20000 train_loss:9.4872 train_time:407ms step_avg:203.60ms tok/s:3862598 -step:3/20000 train_loss:8.8307 train_time:508ms step_avg:169.43ms tok/s:4641597 -step:4/20000 train_loss:8.1811 train_time:610ms step_avg:152.38ms tok/s:5160983 -step:5/20000 train_loss:8.1419 train_time:711ms step_avg:142.27ms tok/s:5527751 -step:6/20000 train_loss:8.2182 train_time:814ms step_avg:135.59ms tok/s:5800185 -step:7/20000 train_loss:8.0845 train_time:915ms step_avg:130.75ms tok/s:6014567 -step:8/20000 train_loss:7.8136 train_time:1018ms step_avg:127.24ms tok/s:6180784 -step:9/20000 train_loss:7.4422 train_time:1119ms step_avg:124.38ms tok/s:6322879 -step:10/20000 train_loss:7.3104 train_time:1221ms step_avg:122.13ms tok/s:6439340 -step:500/20000 train_loss:3.1363 train_time:52838ms step_avg:105.68ms tok/s:7441968 -step:1000/20000 train_loss:2.9131 train_time:105746ms step_avg:105.75ms tok/s:7436986 -step:1500/20000 train_loss:2.8350 train_time:158889ms step_avg:105.93ms tok/s:7424371 -step:2000/20000 train_loss:2.8340 train_time:212131ms step_avg:106.07ms tok/s:7414599 -step:2500/20000 train_loss:2.8605 train_time:265160ms step_avg:106.06ms tok/s:7414703 -step:3000/20000 train_loss:2.7954 train_time:317918ms step_avg:105.97ms tok/s:7421073 -step:3500/20000 train_loss:2.7500 train_time:370945ms step_avg:105.98ms tok/s:7420279 -step:4000/20000 train_loss:2.6705 train_time:423988ms step_avg:106.00ms tok/s:7419378 -step:4000/20000 val_loss:2.6826 val_bpb:0.9147 train_time:424041ms step_avg:106.01ms tok/s:7418447 -step:4500/20000 train_loss:2.6108 train_time:477068ms step_avg:106.02ms tok/s:7418106 -swa:start step:5000 -step:5000/20000 train_loss:2.6237 train_time:530132ms step_avg:106.03ms tok/s:7417316 -late_qat:enabled step:5131 scale:0.1499 -step:5500/20000 train_loss:2.5543 train_time:583851ms step_avg:106.15ms tok/s:7408359 -step:5649/20000 val_loss:2.5868 val_bpb:0.8820 train_time:600112ms step_avg:106.23ms tok/s:7402879 -stopping_early: wallclock_cap train_time:600112ms step:5649/20000 +step:1/20000 train_loss:8.3049 train_time:316ms step_avg:315.96ms tok/s:2489056 +step:2/20000 train_loss:9.4873 train_time:366ms step_avg:182.77ms tok/s:4302780 +step:3/20000 train_loss:8.8165 train_time:466ms step_avg:155.44ms tok/s:5059468 +step:4/20000 train_loss:8.2198 train_time:569ms step_avg:142.37ms tok/s:5523932 +step:5/20000 train_loss:8.2946 train_time:670ms step_avg:134.06ms tok/s:5866094 +step:6/20000 train_loss:8.3620 train_time:772ms step_avg:128.59ms tok/s:6115647 +step:7/20000 train_loss:8.2777 train_time:873ms step_avg:124.69ms tok/s:6306849 +step:8/20000 train_loss:7.9918 train_time:974ms step_avg:121.80ms tok/s:6456800 +step:9/20000 train_loss:7.5262 train_time:1077ms step_avg:119.62ms tok/s:6574604 +step:10/20000 train_loss:7.2360 train_time:1179ms step_avg:117.89ms tok/s:6670845 +step:500/20000 train_loss:3.1472 train_time:52910ms step_avg:105.82ms tok/s:7431795 +step:1000/20000 train_loss:2.9160 train_time:106013ms step_avg:106.01ms tok/s:7418281 +step:1500/20000 train_loss:2.8381 train_time:159182ms step_avg:106.12ms tok/s:7410703 +step:2000/20000 train_loss:2.8330 train_time:212333ms step_avg:106.17ms tok/s:7407545 +step:2500/20000 train_loss:2.8640 train_time:265602ms step_avg:106.24ms tok/s:7402352 +step:3000/20000 train_loss:2.7969 train_time:318632ms step_avg:106.21ms tok/s:7404451 +step:3500/20000 train_loss:2.7451 train_time:371825ms step_avg:106.24ms tok/s:7402710 +step:4000/20000 train_loss:2.6687 train_time:425013ms step_avg:106.25ms tok/s:7401486 +step:4000/20000 val_loss:2.6803 val_bpb:0.9139 train_time:425067ms step_avg:106.27ms tok/s:7400552 +step:4500/20000 train_loss:2.6089 train_time:478196ms step_avg:106.27ms tok/s:7400612 +swa:start step:4950 +step:5000/20000 train_loss:2.6225 train_time:531437ms step_avg:106.29ms tok/s:7399113 +late_qat:enabled step:5119 scale:0.1500 +step:5500/20000 train_loss:2.5525 train_time:585128ms step_avg:106.39ms tok/s:7392188 +step:5637/20000 val_loss:2.5851 val_bpb:0.8814 train_time:600100ms step_avg:106.46ms tok/s:7387296 +stopping_early: wallclock_cap train_time:600100ms step:5637/20000 peak memory allocated: 25252 MiB reserved: 25750 MiB gptq:SKIPPED (SKIP_GPTQ=1) — will use naive int6 ema:applying EMA weights -DIAGNOSTIC post_ema val_loss:2.5842 val_bpb:0.8811 eval_time:1718ms +DIAGNOSTIC post_ema val_loss:2.5825 val_bpb:0.8805 eval_time:1723ms Serialized model: 119539322 bytes -Code size: 128464 bytes -Serialized model mixed_int5_int6_int8+brotli: 15465911 bytes -Total submission size mixed_int5_int6_int8+brotli: 15594375 bytes -final_int6_roundtrip val_loss:2.6727 val_bpb:0.9113 eval_time:5050ms -final_int6_roundtrip_exact val_loss:2.67270610 val_bpb:0.91128978 -final_sliding_window val_loss:2.5428 val_bpb:0.8670 stride:64 eval_time:65449ms -final_sliding_window_exact val_loss:2.54276697 val_bpb:0.86698133 +Code size: 136521 bytes +Serialized model mixed_int5_int6_int8+brotli: 15464951 bytes +Total submission size mixed_int5_int6_int8+brotli: 15601472 bytes +final_int6_roundtrip val_loss:2.6770 val_bpb:0.9127 eval_time:5017ms +final_int6_roundtrip_exact val_loss:2.67696460 val_bpb:0.91274177 +final_sliding_window val_loss:2.6330 val_bpb:0.8977 stride:64 eval_time:64361ms +final_sliding_window_exact val_loss:2.63300498 val_bpb:0.89774887 +ttt:start chunks:1389 ttt_lr:0.005 ttt_epochs:3 chunk_tokens:32768 +final_sliding_window_ttt val_loss:2.5834 val_bpb:0.8808 stride:64 eval_time:243221ms +final_sliding_window_ttt_exact val_loss:2.58340734 val_bpb:0.88084235 diff --git a/records/track_10min_16mb/2026-04-27_Mikey/train_seed42.log b/records/track_10min_16mb/2026-04-27_Mikey/train_seed42.log index 726b5eef07..dc05d9ad2c 100644 --- a/records/track_10min_16mb/2026-04-27_Mikey/train_seed42.log +++ b/records/track_10min_16mb/2026-04-27_Mikey/train_seed42.log @@ -19,7 +19,7 @@ import torch.nn.functional as F from torch import Tensor, nn from torch.nn.parallel import DistributedDataParallel as DDP -os.environ.setdefault("RUN_ID", "rascal_4k_12L_brotli_mixed_8x_seed444_" + time.strftime("%Y%m%d_%H%M%S")) +os.environ.setdefault("RUN_ID", "mikey_8x_seed444_" + time.strftime("%Y%m%d_%H%M%S")) try: import triton @@ -31,7 +31,7 @@ try: from flash_attn_interface import flash_attn_func as flash_attn_3_func except ImportError: flash_attn_3_func = None -# Compression: brotli-11 + byte-shuffle is the salvage_v2 preferred path (PR #1493 recipe). +# Compression: brotli-11 + byte-shuffle. # Falls back to zstd then zlib so this file still runs if brotli isn't installed. _brotli_module = None _zstandard_module = None @@ -55,7 +55,7 @@ if _zstandard_module is not None: zstandard = _zstandard_module if _zlib_module is None: import zlib as _zlib_module # always available; used by zlib fallback path -# --- Byte-shuffle (de-interleave) wrapper from PR #1493: improves brotli ratio on quantized payloads. --- +# --- Byte-shuffle (de-interleave) wrapper: improves brotli ratio on quantized payloads. --- _BSHF_MAGIC = b"BSHF" def _byte_shuffle(data: bytes, stride: int = 2) -> bytes: if stride <= 1 or len(data) < stride: @@ -186,6 +186,11 @@ class Hyperparameters: ngram_eval_max_seconds = float(os.environ.get("NGRAM_EVAL_MAX_SECONDS", 0.0)) ngram_entropy_shift = bool(int(os.environ.get("NGRAM_ENTROPY_SHIFT", "0"))) ngram_order_mults_str = os.environ.get("NGRAM_ORDER_MULTS", "") + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.005)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) cubric_cadence = int(os.environ.get("CUBRIC_CADENCE", 0)) skip_final_eval = bool(int(os.environ.get("SKIP_FINAL_EVAL", "0"))) post_ema_diagnostic = bool(int(os.environ.get("POST_EMA_DIAGNOSTIC", "1"))) @@ -768,14 +773,12 @@ def _classify_param_fine(name: str) -> str: if ".attn." in name or (".proj." in name and ".mlp." not in name): return "attn_other" return "other" -# salvage_v2 mixed-int policy (applied at the Tensor level via _classify_param_fine): -# - mlp_down_bank (mlp_proj, MOST quant-tolerant per 11-day collate) -> int5 (clip_range 15) -# - mlp_up_bank (mlp_fc, also tolerant) -> int5 +# Mixed-int policy (applied at the Tensor level via _classify_param_fine): +# - mlp_down_bank (mlp_proj, most quant-tolerant) -> int5 (clip_range 15) +# - mlp_up_bank (mlp_fc, also tolerant) -> int5 # - qo_bank, kv_bank (attention; LEAST quant-tolerant) -> int6 (clip_range 31) -# - tok_emb / lm_head (embed) -> int6 (matches seed; keeping -# attn/embed at int6 for -# quant safety on this -# first salvage attempt) +# - tok_emb / lm_head (embed) -> int6 (attn/embed at int6 for +# quant safety) # Bytes savings: int5 keeps the int8 storage container (no bit-packing) but the high 3 bits are # forced zero, giving brotli a compressible pattern. Combined with the byte-shuffle wrapper, # expected savings vs uniform-int6+zstd is roughly the int5 bit ratio (5/6 = -17%) APPLIED only @@ -791,7 +794,7 @@ def mixed_quantize_int6_gptq(state_dict: dict[str, Tensor], int6_cats: set[str], (qo, kv, mlp_up, mlp_down, attn_other, mlp_other, aux, embed). For backwards-compat with the old uniform-int6 caller, the legacy coarse names {'mlp','attn','aux','embed'} are also accepted in `int6_cats` and expand to their fine-grained children. `int5_cats` always uses - fine names. If `int5_cats` is None, defaults to DEFAULT_INT5_CATS (the salvage_v2 policy).""" + fine names. If `int5_cats` is None, defaults to DEFAULT_INT5_CATS.""" if int5_cats is None: int5_cats = set(DEFAULT_INT5_CATS) # Expand legacy coarse names so the existing call signature keeps working. @@ -1653,7 +1656,7 @@ def eval_val_sliding_hashed_ngram( batch_seqs: int = 128, eval_seq_len: int | None = None, ) -> tuple[float, float, float]: - """Score-first sliding eval with chunk-based SHARED n-gram tables + cubric. + """Sliding eval with chunk-based SHARED n-gram tables + cubric. Key design: all ranks share identical n-gram tables via bulk chunk updates. Each chunk's windows are distributed across ranks for scoring, then ALL ranks @@ -2035,6 +2038,130 @@ def eval_val_sliding( return val_loss, bits_per_token * tokens_per_byte +def eval_val_ttt( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + context_size = seq_len - stride + window_starts = [ws for ws in range(0, total_tokens, stride) + if ws + context_size < total_tokens] + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + s = 0 if ws == 0 else context_size + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + compiled_logits = maybe_compile( + base_model.forward_logits, + enabled=args.compile_enabled, + fullgraph=args.compile_fullgraph, + ) + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + ttt_params = [p for p in base_model.parameters()] + for p in ttt_params: + p.requires_grad_(True) + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + base_model.eval() + with torch.no_grad(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else context_size + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + is_last_chunk = ci == num_chunks - 1 + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + for pg in optimizer.param_groups: + pg["lr"] = cos_lr + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, batch_seqs): + be = min(bs + batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1 and dist.is_available() and dist.is_initialized(): + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, 1.0) + optimizer.step() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.eval() + return val_loss, bits_per_token * tokens_per_byte + + # --- Training --- def main() -> None: @@ -2047,8 +2174,8 @@ def main() -> None: local_rank = int(os.environ.get("LOCAL_RANK", "0")) if world_size != 8: raise ValueError( - f"Rascal 4k 12L brotli+mixed 8x requires WORLD_SIZE=8, got {world_size}. " - "Launch with: torchrun --standalone --nproc_per_node=8 4k_vocab_rascal_12l_brotli_mixed/train_gpt_4K_12L_brotli_mixed_8xgpu.py" + f"Mikey 8x requires WORLD_SIZE=8, got {world_size}. " + "Launch with: torchrun --standalone --nproc_per_node=8 Mikey/train_gpt_8xgpu.py" ) grad_accum_steps = 8 // world_size grad_scale = 1.0 / grad_accum_steps @@ -2082,10 +2209,6 @@ def main() -> None: print(msg, file=f) log0(code, console=False) log0("=" * 100, console=False) - log0("condition_id:rascal_4k_12L_brotli_mixed_8x_seed444") - log0("run_label:salvage_v2 source_record:rascal_4k_8x_seed444_run20260427 axis:depth_12L+brotli+mixed_int") - log0("changed_fields:num_layers (11->12), compression (zstd->brotli+bshf), quant_policy (uniform_int6->mixed_int5_int6_int8)") - log0("expected_metric:final_sliding_window_exact comparator:0.8672_4k_8x_oversize_run prior_size:17766043_target:<16000000") log0(f"condition:DATA_PATH={args.data_path}") log0(f"condition:TOKENIZER_PATH={args.tokenizer_path}") log0(f"condition:VOCAB_SIZE={args.vocab_size}") @@ -2095,6 +2218,11 @@ def main() -> None: log0(f"condition:COPRIME_MAX_LOADED_SHARDS={args.coprime_max_loaded_shards}") log0(f"condition:COPRIME_SHARDS_PER_BATCH={args.coprime_shards_per_batch}") log0(f"condition:COPRIME_SHARD_HOLD_STEPS={args.coprime_shard_hold_steps}") + log0(f"condition:TTT_ENABLED={int(args.ttt_enabled)}") + log0(f"condition:TTT_LR={args.ttt_lr}") + log0(f"condition:TTT_EPOCHS={args.ttt_epochs}") + log0(f"condition:TTT_MOMENTUM={args.ttt_momentum}") + log0(f"condition:TTT_CHUNK_TOKENS={args.ttt_chunk_tokens}") log0(f"condition:SKIP_GPTQ={os.environ.get('SKIP_GPTQ', '1')}") log0(f"condition:TRIGRAM={int(args.trigram_enabled)}") log0(f"condition:NGRAM_EVAL_ORDER={args.ngram_eval_order}") @@ -2507,9 +2635,9 @@ def main() -> None: log0(f"Code size: {code_bytes} bytes") sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} # GPTQ quantization using Hessians collected from training data. - # salvage_v2 mixed-int policy: int5 for mlp_up_bank/mlp_down_bank (most quant-tolerant per - # 11-day collate), int6 for qo_bank/kv_bank/embed (attention + token embed kept at int6 for - # quant safety; matches seed for embed). See `mixed_quantize_int6_gptq` docstring. + # Mixed-int policy: int5 for mlp_up_bank/mlp_down_bank (most quant-tolerant), + # int6 for qo_bank/kv_bank/embed (attention + token embed kept at int6 for + # quant safety). See `mixed_quantize_int6_gptq` docstring. quant_result, quant_meta = mixed_quantize_int6_gptq( sd_cpu, int6_cats={"qo", "kv", "attn_other", "mlp_other", "aux", "embed"}, @@ -2567,8 +2695,6 @@ def main() -> None: f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" ) log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") - del eval_model, deq_state, quant_state, sd_cpu - torch.cuda.empty_cache() sw_seq_len = effective_eval_seq_len if args.skip_final_eval: log0("final_eval:skipped sliding/ngram by SKIP_FINAL_EVAL=1") @@ -2577,7 +2703,7 @@ def main() -> None: torch.cuda.synchronize() t_slide = time.perf_counter() sw_val_loss, sw_val_bpb = eval_val_sliding( - args, base_model, rank, world_size, device, + args, eval_model, rank, world_size, device, val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, stride=args.eval_stride, eval_seq_len=sw_seq_len, @@ -2592,7 +2718,7 @@ def main() -> None: torch.cuda.synchronize() t_slide64 = time.perf_counter() sw64_val_loss, sw64_val_bpb = eval_val_sliding( - args, base_model, rank, world_size, device, + args, eval_model, rank, world_size, device, val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, stride=64, eval_seq_len=sw_seq_len, @@ -2610,7 +2736,7 @@ def main() -> None: t_ng = time.perf_counter() ng_loss, ng_bpb, ng_coverage = eval_val_sliding_hashed_ngram( args, - base_model, + eval_model, rank, world_size, device, @@ -2649,16 +2775,33 @@ def main() -> None: ) if distributed: dist.barrier() + if args.ttt_enabled and args.eval_stride > 0 and args.eval_stride < sw_seq_len: + if distributed: + dist.barrier() + torch.cuda.synchronize() + t_ttt = time.perf_counter() + log0(f"ttt:start chunks:{(val_tokens.numel() - 1 + args.ttt_chunk_tokens - 1) // args.ttt_chunk_tokens} ttt_lr:{args.ttt_lr} ttt_epochs:{args.ttt_epochs} chunk_tokens:{args.ttt_chunk_tokens}") + ttt_val_loss, ttt_val_bpb = eval_val_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_sliding_window_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" + ) + log0(f"final_sliding_window_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + del eval_model, deq_state, quant_state, sd_cpu + torch.cuda.empty_cache() if distributed: dist.destroy_process_group() if __name__ == "__main__": main() - ==================================================================================================== -condition_id:rascal_4k_12L_brotli_mixed_8x_seed444 -run_label:salvage_v2 source_record:rascal_4k_8x_seed444_run20260427 axis:depth_12L+brotli+mixed_int -changed_fields:num_layers (11->12), compression (zstd->brotli+bshf), quant_policy (uniform_int6->mixed_int5_int6_int8) -expected_metric:final_sliding_window_exact comparator:0.8672_4k_8x_oversize_run prior_size:17766043_target:<16000000 condition:DATA_PATH=./data/datasets/fineweb10B_sp4096 condition:TOKENIZER_PATH=./data/tokenizers/fineweb_4096_bpe.model condition:VOCAB_SIZE=4096 @@ -2668,49 +2811,54 @@ condition:LOADER_MODE=coprime condition:COPRIME_MAX_LOADED_SHARDS=143 condition:COPRIME_SHARDS_PER_BATCH=1 condition:COPRIME_SHARD_HOLD_STEPS=64 +condition:TTT_ENABLED=1 +condition:TTT_LR=0.005 +condition:TTT_EPOCHS=3 +condition:TTT_MOMENTUM=0.9 +condition:TTT_CHUNK_TOKENS=32768 condition:SKIP_GPTQ=1 condition:TRIGRAM=0 condition:NGRAM_EVAL_ORDER=0 Running Python 3.12.13 | packaged by conda-forge | (main, Mar 5 2026, 16:50:00) [GCC 14.3.0] Running PyTorch 2.11.0+cu130 -Mon Apr 27 03:41:14 2026 +Mon Apr 27 15:21:20 2026 +-----------------------------------------------------------------------------------------+ -| NVIDIA-SMI 580.95.05 Driver Version: 580.95.05 CUDA Version: 13.0 | +| NVIDIA-SMI 580.126.09 Driver Version: 580.126.09 CUDA Version: 13.0 | +-----------------------------------------+------------------------+----------------------+ | GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | | | | MIG M. | |=========================================+========================+======================| -| 0 NVIDIA H100 80GB HBM3 On | 00000000:04:00.0 Off | 0 | -| N/A 41C P0 117W / 700W | 1505MiB / 81559MiB | 0% Default | +| 0 NVIDIA H100 80GB HBM3 On | 00000000:0A:00.0 Off | 0 | +| N/A 36C P0 122W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ -| 1 NVIDIA H100 80GB HBM3 On | 00000000:05:00.0 Off | 0 | -| N/A 36C P0 118W / 700W | 1505MiB / 81559MiB | 0% Default | +| 1 NVIDIA H100 80GB HBM3 On | 00000000:18:00.0 Off | 0 | +| N/A 32C P0 120W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ -| 2 NVIDIA H100 80GB HBM3 On | 00000000:0A:00.0 Off | 0 | -| N/A 40C P0 119W / 700W | 1505MiB / 81559MiB | 0% Default | +| 2 NVIDIA H100 80GB HBM3 On | 00000000:3F:00.0 Off | 0 | +| N/A 33C P0 121W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ -| 3 NVIDIA H100 80GB HBM3 On | 00000000:0B:00.0 Off | 0 | -| N/A 36C P0 120W / 700W | 1505MiB / 81559MiB | 0% Default | +| 3 NVIDIA H100 80GB HBM3 On | 00000000:48:00.0 Off | 0 | +| N/A 36C P0 122W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ -| 4 NVIDIA H100 80GB HBM3 On | 00000000:84:00.0 Off | 0 | -| N/A 39C P0 117W / 700W | 1505MiB / 81559MiB | 0% Default | +| 4 NVIDIA H100 80GB HBM3 On | 00000000:87:00.0 Off | 0 | +| N/A 36C P0 121W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ -| 5 NVIDIA H100 80GB HBM3 On | 00000000:85:00.0 Off | 0 | -| N/A 34C P0 116W / 700W | 1505MiB / 81559MiB | 0% Default | +| 5 NVIDIA H100 80GB HBM3 On | 00000000:90:00.0 Off | 0 | +| N/A 32C P0 118W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ -| 6 NVIDIA H100 80GB HBM3 On | 00000000:8A:00.0 Off | 0 | -| N/A 38C P0 121W / 700W | 1505MiB / 81559MiB | 0% Default | +| 6 NVIDIA H100 80GB HBM3 On | 00000000:BE:00.0 Off | 0 | +| N/A 32C P0 120W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ -| 7 NVIDIA H100 80GB HBM3 On | 00000000:8B:00.0 Off | 0 | -| N/A 35C P0 117W / 700W | 1505MiB / 81559MiB | 0% Default | +| 7 NVIDIA H100 80GB HBM3 On | 00000000:C7:00.0 Off | 0 | +| N/A 36C P0 122W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ @@ -2719,7 +2867,14 @@ Mon Apr 27 03:41:14 2026 | GPU GI CI PID Type Process name GPU Memory | | ID ID Usage | |=========================================================================================| -| No running processes found | +| 0 N/A N/A 3021 C /venv/main/bin/python3 1496MiB | +| 1 N/A N/A 3022 C /venv/main/bin/python3 1496MiB | +| 2 N/A N/A 3023 C /venv/main/bin/python3 1496MiB | +| 3 N/A N/A 3024 C /venv/main/bin/python3 1496MiB | +| 4 N/A N/A 3025 C /venv/main/bin/python3 1496MiB | +| 5 N/A N/A 3026 C /venv/main/bin/python3 1496MiB | +| 6 N/A N/A 3027 C /venv/main/bin/python3 1496MiB | +| 7 N/A N/A 3028 C /venv/main/bin/python3 1496MiB | +-----------------------------------------------------------------------------------------+ ==================================================================================================== @@ -2761,41 +2916,44 @@ warmup_step:19/20 warmup_step:20/20 loader_reset:loader:coprime shards:143 blocks:6967965 seq_len:2048 shards_per_batch:1 cache:143 batch_stride:9 hold_steps:64 step:0/20000 val_loss:8.3101 val_bpb:2.8334 train_time:0ms step_avg:0.01ms tok/s:0 -step:1/20000 train_loss:8.3103 train_time:356ms step_avg:355.82ms tok/s:2210168 -step:2/20000 train_loss:9.5390 train_time:407ms step_avg:203.42ms tok/s:3866108 -step:3/20000 train_loss:8.8566 train_time:508ms step_avg:169.26ms tok/s:4646338 -step:4/20000 train_loss:8.1436 train_time:610ms step_avg:152.38ms tok/s:5161001 -step:5/20000 train_loss:8.0018 train_time:711ms step_avg:142.29ms tok/s:5526982 -step:6/20000 train_loss:8.0293 train_time:813ms step_avg:135.49ms tok/s:5804497 -step:7/20000 train_loss:7.8916 train_time:915ms step_avg:130.71ms tok/s:6016846 -step:8/20000 train_loss:7.7694 train_time:1017ms step_avg:127.12ms tok/s:6186402 -step:9/20000 train_loss:7.5519 train_time:1119ms step_avg:124.29ms tok/s:6327601 -step:10/20000 train_loss:7.4294 train_time:1220ms step_avg:122.02ms tok/s:6445305 -step:500/20000 train_loss:3.1321 train_time:52909ms step_avg:105.82ms tok/s:7431948 -step:1000/20000 train_loss:2.9502 train_time:105802ms step_avg:105.80ms tok/s:7433032 -step:1500/20000 train_loss:2.8720 train_time:158682ms step_avg:105.79ms tok/s:7434016 -step:2000/20000 train_loss:2.8296 train_time:211642ms step_avg:105.82ms tok/s:7431713 -step:2500/20000 train_loss:2.7596 train_time:264621ms step_avg:105.85ms tok/s:7429786 -step:3000/20000 train_loss:2.6616 train_time:317371ms step_avg:105.79ms tok/s:7433875 -step:3500/20000 train_loss:2.7652 train_time:370413ms step_avg:105.83ms tok/s:7430917 -step:4000/20000 train_loss:2.7392 train_time:423442ms step_avg:105.86ms tok/s:7428947 -step:4000/20000 val_loss:2.6776 val_bpb:0.9130 train_time:423496ms step_avg:105.87ms tok/s:7428002 -step:4500/20000 train_loss:2.5911 train_time:476437ms step_avg:105.87ms tok/s:7427930 -swa:start step:5000 -step:5000/20000 train_loss:2.5953 train_time:529462ms step_avg:105.89ms tok/s:7426704 -late_qat:enabled step:5138 scale:0.1499 -step:5500/20000 train_loss:2.5660 train_time:583115ms step_avg:106.02ms tok/s:7417709 -step:5656/20000 val_loss:2.5812 val_bpb:0.8801 train_time:600149ms step_avg:106.11ms tok/s:7411597 -stopping_early: wallclock_cap train_time:600149ms step:5656/20000 -peak memory allocated: 25252 MiB reserved: 25750 MiB +step:1/20000 train_loss:8.3103 train_time:382ms step_avg:381.50ms tok/s:2061406 +step:2/20000 train_loss:9.5390 train_time:434ms step_avg:217.24ms tok/s:3620088 +step:3/20000 train_loss:8.8728 train_time:536ms step_avg:178.64ms tok/s:4402315 +step:4/20000 train_loss:8.1470 train_time:638ms step_avg:159.38ms tok/s:4934464 +step:5/20000 train_loss:7.9047 train_time:739ms step_avg:147.80ms tok/s:5320960 +step:6/20000 train_loss:7.9601 train_time:840ms step_avg:140.08ms tok/s:5614244 +step:7/20000 train_loss:7.8922 train_time:942ms step_avg:134.57ms tok/s:5843962 +step:8/20000 train_loss:7.8392 train_time:1044ms step_avg:130.49ms tok/s:6026625 +step:9/20000 train_loss:7.6415 train_time:1146ms step_avg:127.29ms tok/s:6178393 +step:10/20000 train_loss:7.5231 train_time:1247ms step_avg:124.75ms tok/s:6304280 +step:500/20000 train_loss:3.1399 train_time:52985ms step_avg:105.97ms tok/s:7421266 +step:1000/20000 train_loss:2.9554 train_time:106291ms step_avg:106.29ms tok/s:7398890 +step:1500/20000 train_loss:2.8761 train_time:159735ms step_avg:106.49ms tok/s:7385049 +step:2000/20000 train_loss:2.8274 train_time:213190ms step_avg:106.59ms tok/s:7377772 +step:2500/20000 train_loss:2.7613 train_time:266640ms step_avg:106.66ms tok/s:7373528 +step:3000/20000 train_loss:2.6630 train_time:319828ms step_avg:106.61ms tok/s:7376771 +step:3500/20000 train_loss:2.7620 train_time:373297ms step_avg:106.66ms tok/s:7373515 +step:4000/20000 train_loss:2.7378 train_time:426639ms step_avg:106.66ms tok/s:7373276 +step:4000/20000 val_loss:2.6772 val_bpb:0.9128 train_time:426693ms step_avg:106.67ms tok/s:7372349 +step:4500/20000 train_loss:2.5934 train_time:480029ms step_avg:106.67ms tok/s:7372352 +swa:start step:4950 +step:5000/20000 train_loss:2.5953 train_time:533610ms step_avg:106.72ms tok/s:7368978 +late_qat:enabled step:5096 scale:0.1500 +step:5500/20000 train_loss:2.5692 train_time:587573ms step_avg:106.83ms tok/s:7361421 +step:5614/20000 val_loss:2.5832 val_bpb:0.8808 train_time:600090ms step_avg:106.89ms tok/s:7357278 +stopping_early: wallclock_cap train_time:600090ms step:5614/20000 +peak memory allocated: 25263 MiB reserved: 25784 MiB gptq:SKIPPED (SKIP_GPTQ=1) — will use naive int6 ema:applying EMA weights -DIAGNOSTIC post_ema val_loss:2.5786 val_bpb:0.8792 eval_time:1713ms +DIAGNOSTIC post_ema val_loss:2.5806 val_bpb:0.8799 eval_time:1729ms Serialized model: 119539322 bytes -Code size: 128464 bytes -Serialized model mixed_int5_int6_int8+brotli: 15511273 bytes -Total submission size mixed_int5_int6_int8+brotli: 15639737 bytes -final_int6_roundtrip val_loss:2.6638 val_bpb:0.9082 eval_time:5035ms -final_int6_roundtrip_exact val_loss:2.66376216 val_bpb:0.90824024 -final_sliding_window val_loss:2.5371 val_bpb:0.8650 stride:64 eval_time:65261ms -final_sliding_window_exact val_loss:2.53706469 val_bpb:0.86503709 +Code size: 136521 bytes +Serialized model mixed_int5_int6_int8+brotli: 15489943 bytes +Total submission size mixed_int5_int6_int8+brotli: 15626464 bytes +final_int6_roundtrip val_loss:2.6648 val_bpb:0.9086 eval_time:5025ms +final_int6_roundtrip_exact val_loss:2.66482791 val_bpb:0.90860363 +final_sliding_window val_loss:2.6217 val_bpb:0.8939 stride:64 eval_time:73078ms +final_sliding_window_exact val_loss:2.62167670 val_bpb:0.89388638 +ttt:start chunks:1389 ttt_lr:0.005 ttt_epochs:3 chunk_tokens:32768 +final_sliding_window_ttt val_loss:2.5808 val_bpb:0.8799 stride:64 eval_time:272950ms +final_sliding_window_ttt_exact val_loss:2.58078745 val_bpb:0.87994906 diff --git a/records/track_10min_16mb/2026-04-27_Mikey/train_seed444.log b/records/track_10min_16mb/2026-04-27_Mikey/train_seed444.log index 195a1048b3..c160f1b8df 100644 --- a/records/track_10min_16mb/2026-04-27_Mikey/train_seed444.log +++ b/records/track_10min_16mb/2026-04-27_Mikey/train_seed444.log @@ -19,7 +19,7 @@ import torch.nn.functional as F from torch import Tensor, nn from torch.nn.parallel import DistributedDataParallel as DDP -os.environ.setdefault("RUN_ID", "rascal_4k_12L_brotli_mixed_8x_seed444_" + time.strftime("%Y%m%d_%H%M%S")) +os.environ.setdefault("RUN_ID", "mikey_8x_seed444_" + time.strftime("%Y%m%d_%H%M%S")) try: import triton @@ -31,7 +31,7 @@ try: from flash_attn_interface import flash_attn_func as flash_attn_3_func except ImportError: flash_attn_3_func = None -# Compression: brotli-11 + byte-shuffle is the salvage_v2 preferred path (PR #1493 recipe). +# Compression: brotli-11 + byte-shuffle. # Falls back to zstd then zlib so this file still runs if brotli isn't installed. _brotli_module = None _zstandard_module = None @@ -55,7 +55,7 @@ if _zstandard_module is not None: zstandard = _zstandard_module if _zlib_module is None: import zlib as _zlib_module # always available; used by zlib fallback path -# --- Byte-shuffle (de-interleave) wrapper from PR #1493: improves brotli ratio on quantized payloads. --- +# --- Byte-shuffle (de-interleave) wrapper: improves brotli ratio on quantized payloads. --- _BSHF_MAGIC = b"BSHF" def _byte_shuffle(data: bytes, stride: int = 2) -> bytes: if stride <= 1 or len(data) < stride: @@ -186,6 +186,11 @@ class Hyperparameters: ngram_eval_max_seconds = float(os.environ.get("NGRAM_EVAL_MAX_SECONDS", 0.0)) ngram_entropy_shift = bool(int(os.environ.get("NGRAM_ENTROPY_SHIFT", "0"))) ngram_order_mults_str = os.environ.get("NGRAM_ORDER_MULTS", "") + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + ttt_lr = float(os.environ.get("TTT_LR", 0.005)) + ttt_epochs = int(os.environ.get("TTT_EPOCHS", 3)) + ttt_momentum = float(os.environ.get("TTT_MOMENTUM", 0.9)) + ttt_chunk_tokens = int(os.environ.get("TTT_CHUNK_TOKENS", 32768)) cubric_cadence = int(os.environ.get("CUBRIC_CADENCE", 0)) skip_final_eval = bool(int(os.environ.get("SKIP_FINAL_EVAL", "0"))) post_ema_diagnostic = bool(int(os.environ.get("POST_EMA_DIAGNOSTIC", "1"))) @@ -768,14 +773,12 @@ def _classify_param_fine(name: str) -> str: if ".attn." in name or (".proj." in name and ".mlp." not in name): return "attn_other" return "other" -# salvage_v2 mixed-int policy (applied at the Tensor level via _classify_param_fine): -# - mlp_down_bank (mlp_proj, MOST quant-tolerant per 11-day collate) -> int5 (clip_range 15) -# - mlp_up_bank (mlp_fc, also tolerant) -> int5 +# Mixed-int policy (applied at the Tensor level via _classify_param_fine): +# - mlp_down_bank (mlp_proj, most quant-tolerant) -> int5 (clip_range 15) +# - mlp_up_bank (mlp_fc, also tolerant) -> int5 # - qo_bank, kv_bank (attention; LEAST quant-tolerant) -> int6 (clip_range 31) -# - tok_emb / lm_head (embed) -> int6 (matches seed; keeping -# attn/embed at int6 for -# quant safety on this -# first salvage attempt) +# - tok_emb / lm_head (embed) -> int6 (attn/embed at int6 for +# quant safety) # Bytes savings: int5 keeps the int8 storage container (no bit-packing) but the high 3 bits are # forced zero, giving brotli a compressible pattern. Combined with the byte-shuffle wrapper, # expected savings vs uniform-int6+zstd is roughly the int5 bit ratio (5/6 = -17%) APPLIED only @@ -791,7 +794,7 @@ def mixed_quantize_int6_gptq(state_dict: dict[str, Tensor], int6_cats: set[str], (qo, kv, mlp_up, mlp_down, attn_other, mlp_other, aux, embed). For backwards-compat with the old uniform-int6 caller, the legacy coarse names {'mlp','attn','aux','embed'} are also accepted in `int6_cats` and expand to their fine-grained children. `int5_cats` always uses - fine names. If `int5_cats` is None, defaults to DEFAULT_INT5_CATS (the salvage_v2 policy).""" + fine names. If `int5_cats` is None, defaults to DEFAULT_INT5_CATS.""" if int5_cats is None: int5_cats = set(DEFAULT_INT5_CATS) # Expand legacy coarse names so the existing call signature keeps working. @@ -1653,7 +1656,7 @@ def eval_val_sliding_hashed_ngram( batch_seqs: int = 128, eval_seq_len: int | None = None, ) -> tuple[float, float, float]: - """Score-first sliding eval with chunk-based SHARED n-gram tables + cubric. + """Sliding eval with chunk-based SHARED n-gram tables + cubric. Key design: all ranks share identical n-gram tables via bulk chunk updates. Each chunk's windows are distributed across ranks for scoring, then ALL ranks @@ -2035,6 +2038,130 @@ def eval_val_sliding( return val_loss, bits_per_token * tokens_per_byte +def eval_val_ttt( + args: Hyperparameters, + base_model: nn.Module, + rank: int, + world_size: int, + device: torch.device, + val_tokens: Tensor, + base_bytes_lut: Tensor, + has_leading_space_lut: Tensor, + is_boundary_token_lut: Tensor, + stride: int, + batch_seqs: int = 32, + eval_seq_len: int | None = None, +) -> tuple[float, float]: + seq_len = eval_seq_len or args.train_seq_len + total_tokens = val_tokens.numel() - 1 + ttt_chunk = args.ttt_chunk_tokens + context_size = seq_len - stride + window_starts = [ws for ws in range(0, total_tokens, stride) + if ws + context_size < total_tokens] + num_chunks = (total_tokens + ttt_chunk - 1) // ttt_chunk + chunk_windows: list[list[int]] = [[] for _ in range(num_chunks)] + for ws in window_starts: + s = 0 if ws == 0 else context_size + scored_start = ws + s + ci = min(scored_start // ttt_chunk, num_chunks - 1) + chunk_windows[ci].append(ws) + compiled_logits = maybe_compile( + base_model.forward_logits, + enabled=args.compile_enabled, + fullgraph=args.compile_fullgraph, + ) + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + byte_count = torch.zeros((), device=device, dtype=torch.float64) + ttt_params = [p for p in base_model.parameters()] + for p in ttt_params: + p.requires_grad_(True) + optimizer = torch.optim.SGD(ttt_params, lr=args.ttt_lr, momentum=args.ttt_momentum) + for ci in range(num_chunks): + windows = chunk_windows[ci] + if not windows: + continue + chunk_start = ci * ttt_chunk + chunk_end = min((ci + 1) * ttt_chunk, total_tokens) + my_s = (len(windows) * rank) // world_size + my_e = (len(windows) * (rank + 1)) // world_size + my_windows = windows[my_s:my_e] + base_model.eval() + with torch.no_grad(): + for bi in range(0, len(my_windows), batch_seqs): + batch_ws = my_windows[bi:bi + batch_seqs] + bsz = len(batch_ws) + x_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + y_batch = torch.zeros(bsz, seq_len, dtype=torch.int64, device=device) + wlens: list[int] = [] + for i, ws in enumerate(batch_ws): + end = min(ws + seq_len, total_tokens) + wlen = end - ws + wlens.append(wlen) + chunk_tok = val_tokens[ws:end + 1].to(dtype=torch.int64, device=device) + x_batch[i, :wlen] = chunk_tok[:-1] + y_batch[i, :wlen] = chunk_tok[1:] + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = compiled_logits(x_batch) + nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_batch.reshape(-1), + reduction="none", + ).reshape(bsz, seq_len) + for i, ws in enumerate(batch_ws): + wlen = wlens[i] + s = 0 if ws == 0 else context_size + scored_nll = nll[i, s:wlen].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(wlen - s) + tgt = y_batch[i, s:wlen] + prev = x_batch[i, s:wlen] + tb = base_bytes_lut[tgt].to(torch.float64) + tb += (has_leading_space_lut[tgt] & ~is_boundary_token_lut[prev]).to(torch.float64) + byte_count += tb.sum() + is_last_chunk = ci == num_chunks - 1 + if not is_last_chunk and args.ttt_epochs > 0: + base_model.train() + chunk_seqs = (chunk_end - chunk_start) // seq_len + if chunk_seqs > 0: + cos_lr = args.ttt_lr * 0.5 * (1.0 + math.cos(math.pi * ci / max(num_chunks - 1, 1))) + for pg in optimizer.param_groups: + pg["lr"] = cos_lr + my_seq_s = (chunk_seqs * rank) // world_size + my_seq_e = (chunk_seqs * (rank + 1)) // world_size + my_chunk_seqs = my_seq_e - my_seq_s + for _ep in range(args.ttt_epochs): + for bs in range(0, my_chunk_seqs, batch_seqs): + be = min(bs + batch_seqs, my_chunk_seqs) + actual_bs = my_seq_s + bs + start_tok = chunk_start + actual_bs * seq_len + end_tok = chunk_start + (my_seq_s + be) * seq_len + 1 + if end_tok > val_tokens.numel(): + continue + local = val_tokens[start_tok:end_tok].to(device=device, dtype=torch.int64) + x = local[:-1].reshape(-1, seq_len) + y = local[1:].reshape(-1, seq_len) + optimizer.zero_grad(set_to_none=True) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + loss = base_model(x, y) + loss.backward() + if world_size > 1 and dist.is_available() and dist.is_initialized(): + for p in ttt_params: + if p.grad is not None: + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG) + torch.nn.utils.clip_grad_norm_(ttt_params, 1.0) + optimizer.step() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_count, op=dist.ReduceOp.SUM) + val_loss = (loss_sum / token_count).item() + bits_per_token = val_loss / math.log(2.0) + tokens_per_byte = token_count.item() / byte_count.item() + base_model.eval() + return val_loss, bits_per_token * tokens_per_byte + + # --- Training --- def main() -> None: @@ -2047,8 +2174,8 @@ def main() -> None: local_rank = int(os.environ.get("LOCAL_RANK", "0")) if world_size != 8: raise ValueError( - f"Rascal 4k 12L brotli+mixed 8x requires WORLD_SIZE=8, got {world_size}. " - "Launch with: torchrun --standalone --nproc_per_node=8 4k_vocab_rascal_12l_brotli_mixed/train_gpt_4K_12L_brotli_mixed_8xgpu.py" + f"Mikey 8x requires WORLD_SIZE=8, got {world_size}. " + "Launch with: torchrun --standalone --nproc_per_node=8 Mikey/train_gpt_8xgpu.py" ) grad_accum_steps = 8 // world_size grad_scale = 1.0 / grad_accum_steps @@ -2082,10 +2209,6 @@ def main() -> None: print(msg, file=f) log0(code, console=False) log0("=" * 100, console=False) - log0("condition_id:rascal_4k_12L_brotli_mixed_8x_seed444") - log0("run_label:salvage_v2 source_record:rascal_4k_8x_seed444_run20260427 axis:depth_12L+brotli+mixed_int") - log0("changed_fields:num_layers (11->12), compression (zstd->brotli+bshf), quant_policy (uniform_int6->mixed_int5_int6_int8)") - log0("expected_metric:final_sliding_window_exact comparator:0.8672_4k_8x_oversize_run prior_size:17766043_target:<16000000") log0(f"condition:DATA_PATH={args.data_path}") log0(f"condition:TOKENIZER_PATH={args.tokenizer_path}") log0(f"condition:VOCAB_SIZE={args.vocab_size}") @@ -2095,6 +2218,11 @@ def main() -> None: log0(f"condition:COPRIME_MAX_LOADED_SHARDS={args.coprime_max_loaded_shards}") log0(f"condition:COPRIME_SHARDS_PER_BATCH={args.coprime_shards_per_batch}") log0(f"condition:COPRIME_SHARD_HOLD_STEPS={args.coprime_shard_hold_steps}") + log0(f"condition:TTT_ENABLED={int(args.ttt_enabled)}") + log0(f"condition:TTT_LR={args.ttt_lr}") + log0(f"condition:TTT_EPOCHS={args.ttt_epochs}") + log0(f"condition:TTT_MOMENTUM={args.ttt_momentum}") + log0(f"condition:TTT_CHUNK_TOKENS={args.ttt_chunk_tokens}") log0(f"condition:SKIP_GPTQ={os.environ.get('SKIP_GPTQ', '1')}") log0(f"condition:TRIGRAM={int(args.trigram_enabled)}") log0(f"condition:NGRAM_EVAL_ORDER={args.ngram_eval_order}") @@ -2507,9 +2635,9 @@ def main() -> None: log0(f"Code size: {code_bytes} bytes") sd_cpu = {k: v.detach().cpu() for k, v in export_sd.items()} # GPTQ quantization using Hessians collected from training data. - # salvage_v2 mixed-int policy: int5 for mlp_up_bank/mlp_down_bank (most quant-tolerant per - # 11-day collate), int6 for qo_bank/kv_bank/embed (attention + token embed kept at int6 for - # quant safety; matches seed for embed). See `mixed_quantize_int6_gptq` docstring. + # Mixed-int policy: int5 for mlp_up_bank/mlp_down_bank (most quant-tolerant), + # int6 for qo_bank/kv_bank/embed (attention + token embed kept at int6 for + # quant safety). See `mixed_quantize_int6_gptq` docstring. quant_result, quant_meta = mixed_quantize_int6_gptq( sd_cpu, int6_cats={"qo", "kv", "attn_other", "mlp_other", "aux", "embed"}, @@ -2567,8 +2695,6 @@ def main() -> None: f"eval_time:{1000.0 * (time.perf_counter() - t_qeval):.0f}ms" ) log0(f"final_int6_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") - del eval_model, deq_state, quant_state, sd_cpu - torch.cuda.empty_cache() sw_seq_len = effective_eval_seq_len if args.skip_final_eval: log0("final_eval:skipped sliding/ngram by SKIP_FINAL_EVAL=1") @@ -2577,7 +2703,7 @@ def main() -> None: torch.cuda.synchronize() t_slide = time.perf_counter() sw_val_loss, sw_val_bpb = eval_val_sliding( - args, base_model, rank, world_size, device, + args, eval_model, rank, world_size, device, val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, stride=args.eval_stride, eval_seq_len=sw_seq_len, @@ -2592,7 +2718,7 @@ def main() -> None: torch.cuda.synchronize() t_slide64 = time.perf_counter() sw64_val_loss, sw64_val_bpb = eval_val_sliding( - args, base_model, rank, world_size, device, + args, eval_model, rank, world_size, device, val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, stride=64, eval_seq_len=sw_seq_len, @@ -2610,7 +2736,7 @@ def main() -> None: t_ng = time.perf_counter() ng_loss, ng_bpb, ng_coverage = eval_val_sliding_hashed_ngram( args, - base_model, + eval_model, rank, world_size, device, @@ -2649,16 +2775,33 @@ def main() -> None: ) if distributed: dist.barrier() + if args.ttt_enabled and args.eval_stride > 0 and args.eval_stride < sw_seq_len: + if distributed: + dist.barrier() + torch.cuda.synchronize() + t_ttt = time.perf_counter() + log0(f"ttt:start chunks:{(val_tokens.numel() - 1 + args.ttt_chunk_tokens - 1) // args.ttt_chunk_tokens} ttt_lr:{args.ttt_lr} ttt_epochs:{args.ttt_epochs} chunk_tokens:{args.ttt_chunk_tokens}") + ttt_val_loss, ttt_val_bpb = eval_val_ttt( + args, eval_model, rank, world_size, device, + val_tokens, base_bytes_lut, has_leading_space_lut, is_boundary_token_lut, + stride=args.eval_stride, + eval_seq_len=sw_seq_len, + ) + torch.cuda.synchronize() + log0( + f"final_sliding_window_ttt val_loss:{ttt_val_loss:.4f} val_bpb:{ttt_val_bpb:.4f} " + f"stride:{args.eval_stride} eval_time:{1000.0 * (time.perf_counter() - t_ttt):.0f}ms" + ) + log0(f"final_sliding_window_ttt_exact val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f}") + if distributed: + dist.barrier() + del eval_model, deq_state, quant_state, sd_cpu + torch.cuda.empty_cache() if distributed: dist.destroy_process_group() if __name__ == "__main__": main() - ==================================================================================================== -condition_id:rascal_4k_12L_brotli_mixed_8x_seed444 -run_label:salvage_v2 source_record:rascal_4k_8x_seed444_run20260427 axis:depth_12L+brotli+mixed_int -changed_fields:num_layers (11->12), compression (zstd->brotli+bshf), quant_policy (uniform_int6->mixed_int5_int6_int8) -expected_metric:final_sliding_window_exact comparator:0.8672_4k_8x_oversize_run prior_size:17766043_target:<16000000 condition:DATA_PATH=./data/datasets/fineweb10B_sp4096 condition:TOKENIZER_PATH=./data/tokenizers/fineweb_4096_bpe.model condition:VOCAB_SIZE=4096 @@ -2668,49 +2811,54 @@ condition:LOADER_MODE=coprime condition:COPRIME_MAX_LOADED_SHARDS=143 condition:COPRIME_SHARDS_PER_BATCH=1 condition:COPRIME_SHARD_HOLD_STEPS=64 +condition:TTT_ENABLED=1 +condition:TTT_LR=0.005 +condition:TTT_EPOCHS=3 +condition:TTT_MOMENTUM=0.9 +condition:TTT_CHUNK_TOKENS=32768 condition:SKIP_GPTQ=1 condition:TRIGRAM=0 condition:NGRAM_EVAL_ORDER=0 Running Python 3.12.13 | packaged by conda-forge | (main, Mar 5 2026, 16:50:00) [GCC 14.3.0] Running PyTorch 2.11.0+cu130 -Mon Apr 27 03:26:40 2026 +Mon Apr 27 15:57:51 2026 +-----------------------------------------------------------------------------------------+ -| NVIDIA-SMI 580.95.05 Driver Version: 580.95.05 CUDA Version: 13.0 | +| NVIDIA-SMI 580.126.09 Driver Version: 580.126.09 CUDA Version: 13.0 | +-----------------------------------------+------------------------+----------------------+ | GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC | | Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. | | | | MIG M. | |=========================================+========================+======================| -| 0 NVIDIA H100 80GB HBM3 On | 00000000:04:00.0 Off | 0 | -| N/A 41C P0 116W / 700W | 1505MiB / 81559MiB | 0% Default | +| 0 NVIDIA H100 80GB HBM3 On | 00000000:0A:00.0 Off | 0 | +| N/A 43C P0 126W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ -| 1 NVIDIA H100 80GB HBM3 On | 00000000:05:00.0 Off | 0 | -| N/A 36C P0 119W / 700W | 1505MiB / 81559MiB | 0% Default | +| 1 NVIDIA H100 80GB HBM3 On | 00000000:18:00.0 Off | 0 | +| N/A 35C P0 120W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ -| 2 NVIDIA H100 80GB HBM3 On | 00000000:0A:00.0 Off | 0 | -| N/A 41C P0 121W / 700W | 1505MiB / 81559MiB | 0% Default | +| 2 NVIDIA H100 80GB HBM3 On | 00000000:3F:00.0 Off | 0 | +| N/A 36C P0 121W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ -| 3 NVIDIA H100 80GB HBM3 On | 00000000:0B:00.0 Off | 0 | -| N/A 36C P0 121W / 700W | 1505MiB / 81559MiB | 0% Default | +| 3 NVIDIA H100 80GB HBM3 On | 00000000:48:00.0 Off | 0 | +| N/A 44C P0 128W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ -| 4 NVIDIA H100 80GB HBM3 On | 00000000:84:00.0 Off | 0 | -| N/A 40C P0 118W / 700W | 1505MiB / 81559MiB | 0% Default | +| 4 NVIDIA H100 80GB HBM3 On | 00000000:87:00.0 Off | 0 | +| N/A 44C P0 126W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ -| 5 NVIDIA H100 80GB HBM3 On | 00000000:85:00.0 Off | 0 | -| N/A 35C P0 116W / 700W | 1505MiB / 81559MiB | 0% Default | +| 5 NVIDIA H100 80GB HBM3 On | 00000000:90:00.0 Off | 0 | +| N/A 35C P0 120W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ -| 6 NVIDIA H100 80GB HBM3 On | 00000000:8A:00.0 Off | 0 | -| N/A 39C P0 121W / 700W | 1505MiB / 81559MiB | 0% Default | +| 6 NVIDIA H100 80GB HBM3 On | 00000000:BE:00.0 Off | 0 | +| N/A 35C P0 123W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ -| 7 NVIDIA H100 80GB HBM3 On | 00000000:8B:00.0 Off | 0 | -| N/A 35C P0 116W / 700W | 1505MiB / 81559MiB | 0% Default | +| 7 NVIDIA H100 80GB HBM3 On | 00000000:C7:00.0 Off | 0 | +| N/A 44C P0 126W / 700W | 1505MiB / 81559MiB | 0% Default | | | | Disabled | +-----------------------------------------+------------------------+----------------------+ @@ -2719,7 +2867,14 @@ Mon Apr 27 03:26:40 2026 | GPU GI CI PID Type Process name GPU Memory | | ID ID Usage | |=========================================================================================| -| No running processes found | +| 0 N/A N/A 16132 C /venv/main/bin/python3 1496MiB | +| 1 N/A N/A 16133 C /venv/main/bin/python3 1496MiB | +| 2 N/A N/A 16134 C /venv/main/bin/python3 1496MiB | +| 3 N/A N/A 16135 C /venv/main/bin/python3 1496MiB | +| 4 N/A N/A 16136 C /venv/main/bin/python3 1496MiB | +| 5 N/A N/A 16137 C /venv/main/bin/python3 1496MiB | +| 6 N/A N/A 16138 C /venv/main/bin/python3 1496MiB | +| 7 N/A N/A 16139 C /venv/main/bin/python3 1496MiB | +-----------------------------------------------------------------------------------------+ ==================================================================================================== @@ -2761,41 +2916,44 @@ warmup_step:19/20 warmup_step:20/20 loader_reset:loader:coprime shards:143 blocks:6967965 seq_len:2048 shards_per_batch:1 cache:143 batch_stride:46 hold_steps:64 step:0/20000 val_loss:8.3098 val_bpb:2.8333 train_time:0ms step_avg:0.01ms tok/s:0 -step:1/20000 train_loss:8.3105 train_time:360ms step_avg:360.42ms tok/s:2181977 -step:2/20000 train_loss:9.4377 train_time:411ms step_avg:205.53ms tok/s:3826435 -step:3/20000 train_loss:8.8185 train_time:512ms step_avg:170.64ms tok/s:4608598 -step:4/20000 train_loss:7.9978 train_time:614ms step_avg:153.57ms tok/s:5121109 -step:5/20000 train_loss:8.1092 train_time:715ms step_avg:143.09ms tok/s:5496220 -step:6/20000 train_loss:8.1789 train_time:817ms step_avg:136.19ms tok/s:5774519 -step:7/20000 train_loss:7.9287 train_time:919ms step_avg:131.25ms tok/s:5991643 -step:8/20000 train_loss:7.7197 train_time:1020ms step_avg:127.54ms tok/s:6166189 -step:9/20000 train_loss:7.5380 train_time:1121ms step_avg:124.60ms tok/s:6311749 -step:10/20000 train_loss:7.5174 train_time:1223ms step_avg:122.31ms tok/s:6429901 -step:500/20000 train_loss:3.0750 train_time:52893ms step_avg:105.79ms tok/s:7434165 -step:1000/20000 train_loss:2.9603 train_time:105882ms step_avg:105.88ms tok/s:7427432 -step:1500/20000 train_loss:2.8906 train_time:158929ms step_avg:105.95ms tok/s:7422479 -step:2000/20000 train_loss:2.8656 train_time:211956ms step_avg:105.98ms tok/s:7420719 -step:2500/20000 train_loss:2.7443 train_time:265011ms step_avg:106.00ms tok/s:7418864 -step:3000/20000 train_loss:2.7167 train_time:317872ms step_avg:105.96ms tok/s:7422150 -step:3500/20000 train_loss:2.7207 train_time:370962ms step_avg:105.99ms tok/s:7419927 -step:4000/20000 train_loss:2.6442 train_time:424059ms step_avg:106.01ms tok/s:7418130 -step:4000/20000 val_loss:2.6752 val_bpb:0.9121 train_time:424113ms step_avg:106.03ms tok/s:7417191 -step:4500/20000 train_loss:2.6040 train_time:477126ms step_avg:106.03ms tok/s:7417210 -swa:start step:5000 -step:5000/20000 train_loss:2.6335 train_time:530182ms step_avg:106.04ms tok/s:7416616 -late_qat:enabled step:5131 scale:0.1497 -step:5500/20000 train_loss:2.5906 train_time:583846ms step_avg:106.15ms tok/s:7408420 -step:5649/20000 val_loss:2.5794 val_bpb:0.8795 train_time:600083ms step_avg:106.23ms tok/s:7403235 -stopping_early: wallclock_cap train_time:600083ms step:5649/20000 -peak memory allocated: 25261 MiB reserved: 25806 MiB +step:1/20000 train_loss:8.3105 train_time:322ms step_avg:321.89ms tok/s:2443158 +step:2/20000 train_loss:9.4377 train_time:371ms step_avg:185.53ms tok/s:4238774 +step:3/20000 train_loss:8.7992 train_time:473ms step_avg:157.60ms tok/s:4990160 +step:4/20000 train_loss:8.0059 train_time:575ms step_avg:143.72ms tok/s:5472043 +step:5/20000 train_loss:8.2968 train_time:677ms step_avg:135.38ms tok/s:5809041 +step:6/20000 train_loss:8.3576 train_time:778ms step_avg:129.75ms tok/s:6061346 +step:7/20000 train_loss:8.0767 train_time:881ms step_avg:125.83ms tok/s:6250092 +step:8/20000 train_loss:7.8084 train_time:982ms step_avg:122.81ms tok/s:6403519 +step:9/20000 train_loss:7.5402 train_time:1085ms step_avg:120.52ms tok/s:6525534 +step:10/20000 train_loss:7.3918 train_time:1187ms step_avg:118.72ms tok/s:6624388 +step:500/20000 train_loss:3.0773 train_time:53059ms step_avg:106.12ms tok/s:7410954 +step:1000/20000 train_loss:2.9536 train_time:106186ms step_avg:106.19ms tok/s:7406143 +step:1500/20000 train_loss:2.8885 train_time:159472ms step_avg:106.31ms tok/s:7397200 +step:2000/20000 train_loss:2.8663 train_time:212777ms step_avg:106.39ms tok/s:7392085 +step:2500/20000 train_loss:2.7405 train_time:266121ms step_avg:106.45ms tok/s:7387921 +step:3000/20000 train_loss:2.7177 train_time:319334ms step_avg:106.44ms tok/s:7388168 +step:3500/20000 train_loss:2.7149 train_time:372784ms step_avg:106.51ms tok/s:7383665 +step:4000/20000 train_loss:2.6411 train_time:426299ms step_avg:106.57ms tok/s:7379154 +step:4000/20000 val_loss:2.6727 val_bpb:0.9113 train_time:426354ms step_avg:106.59ms tok/s:7378213 +step:4500/20000 train_loss:2.6005 train_time:479672ms step_avg:106.59ms tok/s:7377847 +swa:start step:4950 +step:5000/20000 train_loss:2.6263 train_time:533084ms step_avg:106.62ms tok/s:7376248 +late_qat:enabled step:5102 scale:0.1498 +step:5500/20000 train_loss:2.5891 train_time:586917ms step_avg:106.71ms tok/s:7369659 +step:5621/20000 val_loss:2.5784 val_bpb:0.8791 train_time:600090ms step_avg:106.76ms tok/s:7366451 +stopping_early: wallclock_cap train_time:600090ms step:5621/20000 +peak memory allocated: 25252 MiB reserved: 25750 MiB gptq:SKIPPED (SKIP_GPTQ=1) — will use naive int6 ema:applying EMA weights -DIAGNOSTIC post_ema val_loss:2.5768 val_bpb:0.8786 eval_time:1717ms +DIAGNOSTIC post_ema val_loss:2.5758 val_bpb:0.8782 eval_time:1735ms Serialized model: 119539322 bytes -Code size: 128464 bytes -Serialized model mixed_int5_int6_int8+brotli: 15525048 bytes -Total submission size mixed_int5_int6_int8+brotli: 15653512 bytes -final_int6_roundtrip val_loss:2.6681 val_bpb:0.9097 eval_time:5059ms -final_int6_roundtrip_exact val_loss:2.66809399 val_bpb:0.90971723 -final_sliding_window val_loss:2.5352 val_bpb:0.8644 stride:64 eval_time:77473ms -final_sliding_window_exact val_loss:2.53522743 val_bpb:0.86441066 +Code size: 136521 bytes +Serialized model mixed_int5_int6_int8+brotli: 15534689 bytes +Total submission size mixed_int5_int6_int8+brotli: 15671210 bytes +final_int6_roundtrip val_loss:2.6617 val_bpb:0.9075 eval_time:5017ms +final_int6_roundtrip_exact val_loss:2.66172888 val_bpb:0.90754698 +final_sliding_window val_loss:2.6180 val_bpb:0.8926 stride:64 eval_time:64410ms +final_sliding_window_exact val_loss:2.61804066 val_bpb:0.89264664 +ttt:start chunks:1389 ttt_lr:0.005 ttt_epochs:3 chunk_tokens:32768 +final_sliding_window_ttt val_loss:2.5769 val_bpb:0.8786 stride:64 eval_time:262171ms +final_sliding_window_ttt_exact val_loss:2.57688551 val_bpb:0.87861866 From 43937b85cdb15ce94e693eb5ae576ca17f0b3ea0 Mon Sep 17 00:00:00 2001 From: Octavian Date: Mon, 27 Apr 2026 11:24:52 -0500 Subject: [PATCH 3/3] README: add Mikey 0.87980 to leaderboard Co-Authored-By: Claude Opus 4.7 (1M context) --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index 1445065db7..7d6ea96976 100644 --- a/README.md +++ b/README.md @@ -30,6 +30,7 @@ Happy training! | Run | Score | Author | Summary | Date | Info | |-----|------:|--------|---------|------|------| +| Mikey | 0.87980 | newjordan | On PR #1848: 12L SP4096 + brotli + mixed-int + score-first TTT | 2026-04-27 | [info](records/track_10min_16mb/2026-04-27_Mikey/README.md) | | SP8192 + 3-Layer Recurrence + Parallel Residuals + Legal TTT | 1.0810 | bigbag | On PR #1493: 3-layer recurrence, parallel residuals, QK-Gain 5.25, and legal score-first TTT on the PR #1394 stack | 2026-04-09 | [info](records/track_10min_16mb/2026-04-09_SP8192_3LayerRecur_ParResid_QK525_LegalTTT/README.md) | | SP8192 + Parallel Residuals + Score-First TTT | 1.0822 | aryanbhosale | On PR #1477: parallel residuals on the PR #1413 SP8192 + legal score-first TTT stack | 2026-04-08 | [info](records/track_10min_16mb/2026-04-08_SP8192_ParallelResid_ScoreFirstTTT/README.md) | | SP8192 + QK-Gain 5 + Legal Score-First TTT | 1.0828 | dexhunter | On PR #1413: QK-Gain 5.0 + legal score-first TTT on the PR #1394 SP8192 stack | 2026-04-06 | [info](records/track_10min_16mb/2026-04-06_SP8192_QK5_LegalTTT_1.0828/README.md) |