diff --git a/train_gpt.py b/train_gpt.py index 651beb2b89..873d9d683e 100644 --- a/train_gpt.py +++ b/train_gpt.py @@ -1,186 +1,176 @@ -""" -The `train_gpt.py` and `train_gpt_mlx.py` scripts are intended as good launching-off points for new participants, not SOTA configs. We'll accept PRs that tune, improve, or simplify these scripts without significantly increasing complexity, but competitive submissions should stay in the `/records` folder. - -Hard stop: To keep readable for newcomers, let's make sure `train_gpt.py` and `train_gpt_mlx.py` never are longer than 1500 lines. -""" - -from __future__ import annotations - -import copy -import glob -import io -import math -import os -import random -import subprocess -import sys -import time -import uuid -import zlib +import base64, collections, copy, fcntl, glob, io, json, lzma, math, os from pathlib import Path +import random, re, subprocess, sys, time, uuid, numpy as np, sentencepiece as spm, torch, torch.distributed as dist, torch.nn.functional as F +from torch import nn +from flash_attn_interface import ( + flash_attn_func as flash_attn_3_func, + flash_attn_varlen_func, +) +from concurrent.futures import ThreadPoolExecutor +import triton +import triton.language as tl +from triton.tools.tensor_descriptor import TensorDescriptor -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 - -# ----------------------------- -# HYPERPARAMETERS -# ----------------------------- -# Default Simple Baseline run: -# - 9 transformer blocks at width 512 -# - 8 attention heads with 4 KV heads (GQA) and 2x MLP expansion -# - vocab size 1024, sequence length 1024, tied embeddings -# - 524,288 train tokens per step for 20,000 iterations with a ~10 minute cap class Hyperparameters: - # Data paths are shard globs produced by the existing preprocessing pipeline. - data_path = os.environ.get("DATA_PATH", "./data/datasets/fineweb10B_sp1024") - 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_1024_bpe.model") - run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) + data_dir = os.environ.get("DATA_DIR", "./data/") seed = int(os.environ.get("SEED", 1337)) - - # Validation cadence and batch size. Validation always uses the full fineweb_val split. - val_batch_size = int(os.environ.get("VAL_BATCH_SIZE", 524_288)) - val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 1000)) - train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 200)) - - # Training length. + run_id = os.environ.get("RUN_ID", str(uuid.uuid4())) iterations = int(os.environ.get("ITERATIONS", 20000)) - warmdown_iters = int(os.environ.get("WARMDOWN_ITERS", 1200)) + warmdown_frac = float(os.environ.get("WARMDOWN_FRAC", 0.75)) warmup_steps = int(os.environ.get("WARMUP_STEPS", 20)) - train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 524_288)) - train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 1024)) - max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 600.0)) - qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 1.5)) - - # Model shape. - vocab_size = int(os.environ.get("VOCAB_SIZE", 1024)) - num_layers = int(os.environ.get("NUM_LAYERS", 9)) - num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) + train_batch_tokens = int(os.environ.get("TRAIN_BATCH_TOKENS", 786432)) + train_seq_len = int(os.environ.get("TRAIN_SEQ_LEN", 2048)) + train_log_every = int(os.environ.get("TRAIN_LOG_EVERY", 500)) + max_wallclock_seconds = float(os.environ.get("MAX_WALLCLOCK_SECONDS", 6e2)) + val_batch_tokens = int(os.environ.get("VAL_BATCH_TOKENS", 524288)) + eval_seq_len = int(os.environ.get("EVAL_SEQ_LEN", 2048)) + val_loss_every = int(os.environ.get("VAL_LOSS_EVERY", 4000)) + sliding_window_enabled = bool(int(os.environ.get("SLIDING_WINDOW_ENABLED", "0"))) + vocab_size = int(os.environ.get("VOCAB_SIZE", 8192)) + num_layers = int(os.environ.get("NUM_LAYERS", 11)) + xsa_last_n = int(os.environ.get("XSA_LAST_N", 11)) model_dim = int(os.environ.get("MODEL_DIM", 512)) + embedding_dim = int(os.environ.get("EMBEDDING_DIM", 512)) + num_kv_heads = int(os.environ.get("NUM_KV_HEADS", 4)) num_heads = int(os.environ.get("NUM_HEADS", 8)) - mlp_mult = int(os.environ.get("MLP_MULT", 2)) + mlp_mult = float(os.environ.get("MLP_MULT", 4.0)) + skip_gates_enabled = bool(int(os.environ.get("SKIP_GATES_ENABLED", "1"))) 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)) - - # Optimizer hyperparameters. + logit_softcap = float(os.environ.get("LOGIT_SOFTCAP", 3e1)) + rope_base = float(os.environ.get("ROPE_BASE", 1e4)) + rope_dims = int(os.environ.get("ROPE_DIMS", 16)) + rope_train_seq_len = int(os.environ.get("ROPE_TRAIN_SEQ_LEN", 2048)) + rope_yarn = bool(int(os.environ.get("ROPE_YARN", "0"))) + ln_scale = bool(int(os.environ.get("LN_SCALE", "1"))) + qk_gain_init = float(os.environ.get("QK_GAIN_INIT", 5.0)) + num_loops = int(os.environ.get("NUM_LOOPS", 2)) + loop_start = int(os.environ.get("LOOP_START", 3)) + loop_end = int(os.environ.get("LOOP_END", 5)) + enable_looping_at = float(os.environ.get("ENABLE_LOOPING_AT", 0.35)) + parallel_start_layer = int(os.environ.get("PARALLEL_START_LAYER", 8)) + parallel_final_lane = os.environ.get("PARALLEL_FINAL_LANE", "mean") + min_lr = float(os.environ.get("MIN_LR", 0.0)) 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", 0.05)) + tied_embed_lr = float(os.environ.get("TIED_EMBED_LR", 0.03)) tied_embed_init_std = float(os.environ.get("TIED_EMBED_INIT_STD", 0.005)) - matrix_lr = float(os.environ.get("MATRIX_LR", 0.04)) - scalar_lr = float(os.environ.get("SCALAR_LR", 0.04)) - muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.95)) + matrix_lr = float(os.environ.get("MATRIX_LR", 0.026)) + scalar_lr = float(os.environ.get("SCALAR_LR", 0.02)) + muon_momentum = float(os.environ.get("MUON_MOMENTUM", 0.97)) muon_backend_steps = int(os.environ.get("MUON_BACKEND_STEPS", 5)) - muon_momentum_warmup_start = float(os.environ.get("MUON_MOMENTUM_WARMUP_START", 0.85)) - muon_momentum_warmup_steps = int(os.environ.get("MUON_MOMENTUM_WARMUP_STEPS", 500)) + 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)) + muon_row_normalize = bool(int(os.environ.get("MUON_ROW_NORMALIZE", "1"))) 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", 0.0)) - -# ----------------------------- -# MUON OPTIMIZER -# ----------------------------- -# -# As borrowed from modded-nanogpt -# Background on Muon: https://kellerjordan.github.io/posts/muon/ - -def zeropower_via_newtonschulz5(G: Tensor, steps: int = 10, eps: float = 1e-7) -> Tensor: - # Orthogonalize a 2D update matrix with a fast Newton-Schulz iteration. - # Muon uses this to normalize matrix-shaped gradients before applying them. - a, b, c = (3.4445, -4.7750, 2.0315) - X = G.bfloat16() - X /= X.norm() + eps - transposed = G.size(0) > G.size(1) - if transposed: - X = X.T - for _ in range(steps): - A = X @ X.T - B = b * A + c * A @ A - X = a * X + B @ X - return X.T if transposed else X - + adam_eps = float(os.environ.get("ADAM_EPS", 1e-08)) + grad_clip_norm = float(os.environ.get("GRAD_CLIP_NORM", 0.3)) + eval_stride = int(os.environ.get("EVAL_STRIDE", 64)) + muon_beta2 = float(os.environ.get("MUON_BETA2", 0.95)) + adam_wd = float(os.environ.get("ADAM_WD", 0.02)) + muon_wd = float(os.environ.get("MUON_WD", 0.095)) + embed_wd = float(os.environ.get("EMBED_WD", 0.085)) + ema_decay = float(os.environ.get("EMA_DECAY", 0.9965)) + ttt_enabled = bool(int(os.environ.get("TTT_ENABLED", "1"))) + ttt_lora_rank = int(os.environ.get("TTT_LORA_RANK", 96)) + ttt_lora_lr = float(os.environ.get("TTT_LORA_LR", 0.0001)) + ttt_chunk_size = int(os.environ.get("TTT_CHUNK_SIZE", 48)) + ttt_eval_seq_len = int(os.environ.get("TTT_EVAL_SEQ_LEN", 2048)) + ttt_batch_size = int(os.environ.get("TTT_BATCH_SIZE", 64)) + ttt_grad_steps = int(os.environ.get("TTT_GRAD_STEPS", 1)) + ttt_weight_decay = float(os.environ.get("TTT_WEIGHT_DECAY", 0.5)) + ttt_beta1 = float(os.environ.get("TTT_BETA1", 0)) + ttt_beta2 = float(os.environ.get("TTT_BETA2", 0.999)) + ttt_k_lora = bool(int(os.environ.get("TTT_K_LORA", "1"))) + ttt_mlp_lora = bool(int(os.environ.get("TTT_MLP_LORA", "1"))) + ttt_o_lora = bool(int(os.environ.get("TTT_O_LORA", "1"))) + ttt_optimizer = os.environ.get("TTT_OPTIMIZER", "adam") + ttt_eval_batches = os.environ.get("TTT_EVAL_BATCHES", "") + ttt_output_dir = os.environ.get("TTT_OUTPUT_DIR", "") + val_doc_fraction = float(os.environ.get("VAL_DOC_FRACTION", 1.0)) + compressor = os.environ.get("COMPRESSOR", "brotli") + gptq_calibration_batches = int(os.environ.get("GPTQ_CALIBRATION_BATCHES", 64)) + gptq_reserve_seconds = float(os.environ.get("GPTQ_RESERVE_SECONDS", 13.0)) + matrix_bits = int(os.environ.get("MATRIX_BITS", 6)) + embed_bits = int(os.environ.get("EMBED_BITS", 7)) + matrix_clip_sigmas = float(os.environ.get("MATRIX_CLIP_SIGMAS", 12.85)) + embed_clip_sigmas = float(os.environ.get("EMBED_CLIP_SIGMAS", 15.0)) + mlp_clip_sigmas = float(os.environ.get("MLP_CLIP_SIGMAS", 12.0)) + attn_clip_sigmas = float(os.environ.get("ATTN_CLIP_SIGMAS", 13.0)) + 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")) + is_main_process = rank == 0 + grad_accum_steps = 8 // world_size + datasets_dir = os.path.join(data_dir, "datasets", f"fineweb10B_sp{vocab_size}") + train_files = os.path.join(datasets_dir, "fineweb_train_*.bin") + val_files = os.path.join(datasets_dir, "fineweb_val_*.bin") + tokenizer_path = os.path.join( + data_dir, "tokenizers", f"fineweb_{vocab_size}_bpe.model" + ) + artifact_dir = os.environ.get("ARTIFACT_DIR", "") + eval_only_path = os.environ.get("EVAL_ONLY_PATH", "") + logfile = ( + os.path.join(artifact_dir, f"{run_id}.txt") + if artifact_dir + else f"logs/{run_id}.txt" + ) + model_path = ( + os.path.join(artifact_dir, "final_model.pt") + if artifact_dir + else "final_model.pt" + ) + quantized_model_path = ( + os.path.join(artifact_dir, "final_model.int6.ptz") + if artifact_dir + else "final_model.int6.ptz" + ) -class Muon(torch.optim.Optimizer): - def __init__(self, params, lr: float, momentum: float, backend_steps: int, nesterov: bool = True): - super().__init__( - params, - dict(lr=lr, momentum=momentum, backend_steps=backend_steps, nesterov=nesterov), - ) - @torch.no_grad() - def step(self, closure=None): - loss = None - if closure is not None: - with torch.enable_grad(): - loss = closure() +_logger_hparams = None - distributed = dist.is_available() and dist.is_initialized() - world_size = dist.get_world_size() if distributed else 1 - rank = dist.get_rank() if distributed else 0 - for group in self.param_groups: - params = group["params"] - if not params: - continue - lr = group["lr"] - momentum = group["momentum"] - backend_steps = group["backend_steps"] - nesterov = group["nesterov"] +def set_logging_hparams(h): + global _logger_hparams + _logger_hparams = h - total_params = sum(int(p.numel()) for p in params) - updates_flat = torch.zeros(total_params, device=params[0].device, dtype=torch.bfloat16) - curr = 0 - for i, p in enumerate(params): - if i % world_size == rank and p.grad is not None: - g = p.grad - 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: - g = g.add(buf, alpha=momentum) - g = zeropower_via_newtonschulz5(g, steps=backend_steps) - # Scale correction from Muon reference implementations. - g *= max(1, g.size(0) / g.size(1)) ** 0.5 - updates_flat[curr : curr + p.numel()] = g.reshape(-1) - curr += p.numel() - - if distributed: - dist.all_reduce(updates_flat, op=dist.ReduceOp.SUM) - - curr = 0 - for p in params: - g = updates_flat[curr : curr + p.numel()].view_as(p).to(dtype=p.dtype) - p.add_(g, alpha=-lr) - curr += p.numel() +def log(msg, console=True): + if _logger_hparams is None: + print(msg) + return + if _logger_hparams.is_main_process: + if console: + print(msg) + if _logger_hparams.logfile is not None: + with open(_logger_hparams.logfile, "a", encoding="utf-8") as f: + print(msg, file=f) - return loss +class ValidationData: + def __init__(self, h, device): + self.sp = spm.SentencePieceProcessor(model_file=h.tokenizer_path) + if int(self.sp.vocab_size()) != h.vocab_size: + raise ValueError( + f"VOCAB_SIZE={h.vocab_size} does not match tokenizer vocab_size={int(self.sp.vocab_size())}" + ) + self.val_tokens = load_validation_tokens(h.val_files, h.eval_seq_len) + ( + self.base_bytes_lut, + self.has_leading_space_lut, + self.is_boundary_token_lut, + ) = build_sentencepiece_luts(self.sp, h.vocab_size, device) -# ----------------------------- -# TOKENIZER-AGNOSTIC EVALUATION SETUP -# ----------------------------- -# -# It's common for small models have a large fraction of their parameters be embeddings, since the 2 * d_model * d_vocab vectors can be gigantic. -# Instead of locking the tokenizer, we let you bring your own and calculate our validation metrics on the average compression of the validation set. -# We calculate BPB (bits-per-byte) instead of validation loss, so we need methods to count the number of bits per token in the tokenizer. -# Note: Submissions that edit the tokenizer will be examined more carefully, since screwing this up might unjustly improve your score. -def build_sentencepiece_luts( - sp: spm.SentencePieceProcessor, vocab_size: int, device: torch.device -) -> tuple[Tensor, Tensor, Tensor]: +def build_sentencepiece_luts(sp, vocab_size, device): sp_vocab_size = int(sp.vocab_size()) + assert ( + sp.piece_to_id("▁") != sp.unk_id() + ), "Tokenizer must have '▁' (space) as its own token for correct BPB byte counting" 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_) @@ -204,362 +194,435 @@ def build_sentencepiece_luts( ) -def load_validation_tokens(pattern: str, seq_len: int) -> Tensor: +def load_validation_tokens(pattern, seq_len): files = [Path(p) for p in sorted(glob.glob(pattern))] if not files: raise FileNotFoundError(f"No files found for pattern: {pattern}") - # The export pipeline writes the fixed first-50k-doc validation set to fineweb_val_*. tokens = torch.cat([load_data_shard(file) for file in files]).contiguous() - usable = ((tokens.numel() - 1) // seq_len) * seq_len + 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, -) -> tuple[float, float]: - # Validation computes two metrics: - # - val_loss: token cross-entropy (natural log) - # - val_bpb: tokenizer-agnostic compression metric used by the challenge - local_batch_tokens = args.val_batch_size // (world_size * grad_accum_steps) - if local_batch_tokens < args.train_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}, TRAIN_SEQ_LEN={args.train_seq_len}" - ) - local_batch_seqs = local_batch_tokens // args.train_seq_len - total_seqs = (val_tokens.numel() - 1) // args.train_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 * args.train_seq_len - raw_end = batch_seq_end * args.train_seq_len + 1 - local = val_tokens[raw_start:raw_end].to(device=device, dtype=torch.int64, non_blocking=True) - x = local[:-1].reshape(-1, args.train_seq_len) - y = local[1:].reshape(-1, args.train_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) - -# ----------------------------- -# POST-TRAINING QUANTIZATION -# ----------------------------- -# -# It's silly to export our model, which is trained in bf16 and fp32, at that same precision. -# Instead, we get approximately the same model (with a small hit) by quantizing the model to int8 & zlib compressing. -# We can then decompress the model and run in higher precision for evaluation, after closing in under the size limit. - -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", - ).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: - # Matrices get one scale per row, which usually tracks output-channel - # ranges much better than a single tensor-wide scale. - 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() - - # Vectors / scalars use a simpler per-tensor scale. - 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 quantize_state_dict_int8(state_dict: dict[str, Tensor]): - # Single supported clean-script export format: - # - per-row int8 for 2D float tensors - # - per-tensor int8 for other float tensors - # - exact passthrough for non-floats - # - passthrough for small float tensors, stored as fp16 to save bytes - quantized: dict[str, Tensor] = {} - scales: dict[str, Tensor] = {} - dtypes: dict[str, str] = {} - passthrough: dict[str, Tensor] = {} - passthrough_orig_dtypes: dict[str, str] = {} - qmeta: dict[str, dict[str, object]] = {} - stats = dict.fromkeys( - ("param_count", "num_tensors", "num_float_tensors", "num_nonfloat_tensors", "baseline_tensor_bytes", "int8_payload_bytes"), - 0, - ) - - for name, tensor in state_dict.items(): - t = tensor.detach().to("cpu").contiguous() - stats["param_count"] += int(t.numel()) - stats["num_tensors"] += 1 - stats["baseline_tensor_bytes"] += tensor_nbytes(t) - - if not t.is_floating_point(): - stats["num_nonfloat_tensors"] += 1 - passthrough[name] = t - stats["int8_payload_bytes"] += tensor_nbytes(t) - continue - - # Small float tensors are cheap enough to keep directly. We still downcast - # fp32/bf16 passthrough tensors to fp16 so metadata does not dominate size. - if t.numel() <= INT8_KEEP_FLOAT_MAX_NUMEL: - kept = keep_float_tensor(name, t, passthrough_orig_dtypes) - passthrough[name] = kept - stats["int8_payload_bytes"] += tensor_nbytes(kept) - continue - - stats["num_float_tensors"] += 1 - q, s = quantize_float_tensor(t) - if s.ndim > 0: - qmeta[name] = {"scheme": "per_row", "axis": 0} - quantized[name] = q - scales[name] = s - dtypes[name] = str(t.dtype).removeprefix("torch.") - stats["int8_payload_bytes"] += tensor_nbytes(q) + tensor_nbytes(s) - - obj: dict[str, object] = { - "__quant_format__": "int8_clean_per_row_v1", - "quantized": quantized, - "scales": scales, - "dtypes": dtypes, - "passthrough": passthrough, - } - if qmeta: - obj["qmeta"] = qmeta - if passthrough_orig_dtypes: - obj["passthrough_orig_dtypes"] = passthrough_orig_dtypes - return obj, stats - -def dequantize_state_dict_int8(obj: dict[str, object]) -> dict[str, Tensor]: - out: dict[str, Tensor] = {} - qmeta = obj.get("qmeta", {}) - passthrough_orig_dtypes = obj.get("passthrough_orig_dtypes", {}) - for name, q in obj["quantized"].items(): - dtype = getattr(torch, obj["dtypes"][name]) - s = obj["scales"][name] - if qmeta.get(name, {}).get("scheme") == "per_row" or s.ndim > 0: - s = s.to(dtype=torch.float32) - # Broadcast the saved row scale back across trailing dimensions. - out[name] = (q.float() * s.view(q.shape[0], *([1] * (q.ndim - 1)))).to(dtype=dtype).contiguous() - else: - scale = float(s.item()) - out[name] = (q.float() * scale).to(dtype=dtype).contiguous() - for name, t in obj["passthrough"].items(): - # Restore small tensors, undoing the temporary fp16 storage cast if needed. - out_t = t.detach().to("cpu").contiguous() - orig_dtype = passthrough_orig_dtypes.get(name) - if isinstance(orig_dtype, str): - out_t = out_t.to(dtype=getattr(torch, orig_dtype)).contiguous() - out[name] = out_t - return out - - -# ----------------------------- -# DATA LOADING -# ----------------------------- - -def load_data_shard(file: Path) -> Tensor: +def load_data_shard(file): header_bytes = 256 * np.dtype(" 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: - # Each call consumes a contiguous chunk from the shared token stream, then slices out - # one disjoint span per rank. The extra "+1" token lets us build (x, y) by shifting. - def __init__(self, pattern: str, rank: int, world_size: int, device: torch.device): - self.rank = rank - self.world_size = world_size +_SHARD_HEADER_BYTES = 256 * np.dtype(" 0: + pos = start + while pos < end: + seg_starts.append(pos) + pos += max_doc_len + else: + seg_starts.append(start) + boundaries = seg_starts + [total_len] + padded_len = get_next_multiple_of_n(len(boundaries), bucket_size) + cu = torch.full((padded_len,), total_len, dtype=torch.int32, device=device) + cu[: len(boundaries)] = torch.tensor(boundaries, dtype=torch.int32, device=device) + seg_ends = seg_starts[1:] + [total_len] + max_seqlen = max(end - start for start, end in zip(seg_starts, seg_ends)) + return cu, max_seqlen + +class DocumentPackingLoader: + _shard_pool = ThreadPoolExecutor(1) + + def __init__(self, h, device, cu_bucket_size=64): + self.rank = h.rank + self.world_size = h.world_size + self.device = device + self.cu_bucket_size = cu_bucket_size + self.max_seq_len = h.train_seq_len + all_files = [Path(p) for p in sorted(glob.glob(h.train_files))] + if not all_files: + raise FileNotFoundError(f"No files found for pattern: {h.train_files}") + self.files = all_files + self.file_iter = iter(self.files) + self._init_shard(load_data_shard(next(self.file_iter))) + self._next_shard = self._submit_next_shard() + self._batch_pool = ThreadPoolExecutor(1) + self._next_batch = None + + def _init_shard(self, tokens): + global BOS_ID + self.tokens = tokens + self.shard_size = tokens.numel() + if BOS_ID is None: + BOS_ID = 1 + self.bos_idx = ( + (tokens == BOS_ID).nonzero(as_tuple=True)[0].to(torch.int64).cpu().numpy() + ) + if self.bos_idx.size == 0: + self.bos_idx = np.array([0], dtype=np.int64) + self.cursor = int(self.bos_idx[0]) + + def _submit_next_shard(self): + try: + path = next(self.file_iter) + return self._shard_pool.submit(load_data_shard, path) + except StopIteration: + return None + + def _advance_shard(self): + if self._next_shard is None: + self.file_iter = iter(self.files) + self._next_shard = self._shard_pool.submit( + load_data_shard, next(self.file_iter) + ) + self._init_shard(self._next_shard.result()) + self._next_shard = self._submit_next_shard() + + def _local_doc_starts(self, local_start, total_len): + lo = np.searchsorted(self.bos_idx, local_start, side="left") + hi = np.searchsorted(self.bos_idx, local_start + total_len, side="left") + return (self.bos_idx[lo:hi] - local_start).tolist() + + def _prepare_batch(self, num_tokens_local, max_seq_len): + per_rank_span = num_tokens_local + 1 + global_span = per_rank_span * self.world_size + while self.cursor + global_span > self.shard_size: + self._advance_shard() + local_start = self.cursor + self.rank * per_rank_span + buf = self.tokens[local_start : local_start + per_rank_span] + inputs = buf[:-1].to(dtype=torch.int64).pin_memory() + targets = buf[1:].to(dtype=torch.int64).pin_memory() + starts = self._local_doc_starts(local_start, inputs.numel()) + cu_seqlens, max_seqlen = _build_cu_seqlens( + starts, inputs.numel(), inputs.device, max_seq_len, self.cu_bucket_size + ) + cu_seqlens = cu_seqlens.pin_memory() + self.cursor += global_span + return inputs, targets, cu_seqlens, max_seqlen + + def next_batch(self, global_tokens, grad_accum_steps): + num_tokens_local = global_tokens // (self.world_size * grad_accum_steps) + if self._next_batch is not None: + inputs, targets, cu_seqlens, max_seqlen = self._next_batch.result() + else: + inputs, targets, cu_seqlens, max_seqlen = self._prepare_batch( + num_tokens_local, self.max_seq_len + ) + self._next_batch = self._batch_pool.submit( + self._prepare_batch, num_tokens_local, self.max_seq_len + ) + return ( + inputs[None].to(self.device, non_blocking=True), + targets[None].to(self.device, non_blocking=True), + cu_seqlens.to(self.device, non_blocking=True), + max_seqlen, + ) + + +class ShuffledSequenceLoader: + def __init__(self, h, device): + self.world_size = h.world_size + self.seq_len = h.train_seq_len self.device = device - self.stream = TokenStream(pattern) - - 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) - -# ----------------------------- -# TRANSFORMER MODULES -# ----------------------------- + all_files = [Path(p) for p in sorted(glob.glob(h.train_files))] + if not all_files: + raise FileNotFoundError(f"No files found for pattern: {h.train_files}") + self.files = all_files[h.rank :: h.world_size] + self.rng = np.random.Generator(np.random.PCG64(h.rank)) + self.num_tokens = [_read_num_tokens(f) for f in self.files] + self.start_inds = [[] for _ in self.files] + for si in range(len(self.files)): + self._reset_shard(si) + + def _reset_shard(self, si): + max_phase = min( + self.seq_len - 1, max(0, self.num_tokens[si] - self.seq_len - 1) + ) + phase = int(self.rng.integers(max_phase + 1)) if max_phase > 0 else 0 + num_sequences = (self.num_tokens[si] - 1 - phase) // self.seq_len + sequence_order = self.rng.permutation(num_sequences) + self.start_inds[si] = (phase + sequence_order * self.seq_len).tolist() + + def next_batch(self, global_tokens, grad_accum_steps): + device_tokens = global_tokens // (self.world_size * grad_accum_steps) + device_batch_size = device_tokens // self.seq_len + remaining = np.array([len(s) for s in self.start_inds], dtype=np.float64) + x = torch.empty((device_batch_size, self.seq_len), dtype=torch.int64) + y = torch.empty((device_batch_size, self.seq_len), dtype=torch.int64) + for bi in range(device_batch_size): + total = remaining.sum() + if total <= 0: + for si in range(len(self.files)): + self._reset_shard(si) + remaining = np.array( + [len(s) for s in self.start_inds], dtype=np.float64 + ) + total = remaining.sum() + probs = remaining / total + si = int(self.rng.choice(len(self.files), p=probs)) + start_ind = self.start_inds[si].pop() + remaining[si] -= 1 + mm = _get_shard_memmap(self.files[si]) + window = torch.as_tensor( + np.array(mm[start_ind : start_ind + self.seq_len + 1], dtype=np.int64) + ) + x[bi] = window[:-1] + y[bi] = window[1:] + return x.to(self.device, non_blocking=True), y.to( + self.device, non_blocking=True + ) + class RMSNorm(nn.Module): - def __init__(self, eps: float | None = None): + def __init__(self, eps=None): super().__init__() self.eps = eps - def forward(self, x: Tensor) -> Tensor: + def forward(self, x): return F.rms_norm(x, (x.size(-1),), eps=self.eps) class CastedLinear(nn.Linear): - # Keep weights in fp32 for optimizer/state quality, cast at matmul time for bf16 compute. - def forward(self, x: Tensor) -> Tensor: + def forward(self, x): + w = self.weight.to(x.dtype) bias = self.bias.to(x.dtype) if self.bias is not None else None - return F.linear(x, self.weight.to(x.dtype), bias) - - -def restore_low_dim_params_to_fp32(module: nn.Module) -> None: - # Keep small/control parameters in fp32 even when the model body runs in bf16. - 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() + return F.linear(x, w, bias) + + +@triton.jit +def linear_xielu_kernel( + a_desc, + b_desc, + c_desc, + aux_desc, + M, + N, + K, + ap, an, bp, bn, + BLOCK_SIZE_M: tl.constexpr, + BLOCK_SIZE_N: tl.constexpr, + BLOCK_SIZE_K: tl.constexpr, + NUM_SMS: tl.constexpr, + FORWARD: tl.constexpr, +): + dtype = tl.bfloat16 + start_pid = tl.program_id(axis=0) + num_pid_m = tl.cdiv(M, BLOCK_SIZE_M) + num_pid_n = tl.cdiv(N, BLOCK_SIZE_N) + k_tiles = tl.cdiv(K, BLOCK_SIZE_K) + num_tiles = num_pid_m * num_pid_n + tile_id_c = start_pid - NUM_SMS + for tile_id in tl.range(start_pid, num_tiles, NUM_SMS, flatten=True): + pid_m = tile_id // num_pid_n + pid_n = tile_id % num_pid_n + offs_am = pid_m * BLOCK_SIZE_M + offs_bn = pid_n * BLOCK_SIZE_N + accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32) + for ki in range(k_tiles): + offs_k = ki * BLOCK_SIZE_K + a = a_desc.load([offs_am, offs_k]) + b = b_desc.load([offs_bn, offs_k]) + accumulator = tl.dot(a, b.T, accumulator) + tile_id_c += NUM_SMS + offs_am_c = offs_am + offs_bn_c = offs_bn + acc = tl.reshape(accumulator, (BLOCK_SIZE_M, 2, BLOCK_SIZE_N // 2)) + acc = tl.permute(acc, (0, 2, 1)) + acc0, acc1 = tl.split(acc) + c0 = acc0.to(dtype) + c1 = acc1.to(dtype) + if not FORWARD: + # backward: multiply grad by d/dh[xIELU(h)] + pre0 = aux_desc.load([offs_am_c, offs_bn_c]) + pre1 = aux_desc.load([offs_am_c, offs_bn_c + BLOCK_SIZE_N // 2]) + c0 = c0 * tl.where(pre0 > 0, 2.0 * ap * pre0 + bp, 2.0 * an * pre0 + bn) + c1 = c1 * tl.where(pre1 > 0, 2.0 * ap * pre1 + bp, 2.0 * an * pre1 + bn) + # forward: c_desc = raw matmul (pre); backward: c_desc = d_activation * grad + c_desc.store([offs_am_c, offs_bn_c], c0) + c_desc.store([offs_am_c, offs_bn_c + BLOCK_SIZE_N // 2], c1) + if FORWARD: + # xIELU: h * where(h > 0, ap*h + bp, an*h + bn) + aux0 = c0 * tl.where(c0 > 0, ap * c0 + bp, an * c0 + bn) + aux1 = c1 * tl.where(c1 > 0, ap * c1 + bp, an * c1 + bn) + aux_desc.store([offs_am_c, offs_bn_c], aux0) + aux_desc.store([offs_am_c, offs_bn_c + BLOCK_SIZE_N // 2], aux1) + + +def linear_xielu(a, b, ap, an, bp, bn, aux=None): + M, K = a.shape + N, K2 = b.shape + assert K == K2 + c = torch.empty((M, N), device=a.device, dtype=a.dtype) + forward = aux is None + if aux is None: + aux = torch.empty((M, N), device=a.device, dtype=a.dtype) + num_sms = torch.cuda.get_device_properties(a.device).multi_processor_count + BLOCK_SIZE_M, BLOCK_SIZE_N, BLOCK_SIZE_K = 128, 256, 64 + num_stages = 4 if forward else 3 + a_desc = TensorDescriptor.from_tensor(a, [BLOCK_SIZE_M, BLOCK_SIZE_K]) + b_desc = TensorDescriptor.from_tensor(b, [BLOCK_SIZE_N, BLOCK_SIZE_K]) + c_desc = TensorDescriptor.from_tensor(c, [BLOCK_SIZE_M, BLOCK_SIZE_N // 2]) + aux_desc = TensorDescriptor.from_tensor(aux, [BLOCK_SIZE_M, BLOCK_SIZE_N // 2]) + grid = lambda _meta: ( + min(num_sms, triton.cdiv(M, BLOCK_SIZE_M) * triton.cdiv(N, BLOCK_SIZE_N)), + ) + linear_xielu_kernel[grid]( + a_desc, + b_desc, + c_desc, + aux_desc, + M, + N, + K, + ap, an, bp, bn, + BLOCK_SIZE_M=BLOCK_SIZE_M, + BLOCK_SIZE_N=BLOCK_SIZE_N, + BLOCK_SIZE_K=BLOCK_SIZE_K, + NUM_SMS=num_sms, + FORWARD=forward, + num_stages=num_stages, + num_warps=8, + ) + if forward: + return c, aux + return c + + +class FusedXieluMLPFunction(torch.autograd.Function): + @staticmethod + def forward(ctx, x, w1, w2, ap, an, bp, bn): + x_flat = x.reshape(-1, x.shape[-1]) + pre, post = linear_xielu(x_flat, w1, ap, an, bp, bn) + out = F.linear(post, w2) + ctx.save_for_backward(x, w1, w2, pre, post) + ctx.ap = ap + ctx.an = an + ctx.bp = bp + ctx.bn = bn + return out.view(*x.shape[:-1], out.shape[-1]) + + @staticmethod + def backward(ctx, grad_output): + x, w1, w2, pre, post = ctx.saved_tensors + x_flat = x.reshape(-1, x.shape[-1]) + grad_output_flat = grad_output.reshape(-1, grad_output.shape[-1]) + dw2 = grad_output_flat.T @ post + dpre = linear_xielu(grad_output_flat, w2.T.contiguous(), ctx.ap, ctx.an, ctx.bp, ctx.bn, aux=pre) + dw1 = dpre.T @ x_flat + dx = dpre @ w1 + return dx.view_as(x), dw1, dw2, None, None, None, None + + +FusedXieluMLP = FusedXieluMLPFunction.apply class Rotary(nn.Module): - # Caches cos/sin tables per sequence length on the current device. - def __init__(self, dim: int, base: float = 10000.0): + def __init__(self, dim, base=1e4, train_seq_len=1024, rope_dims=0, yarn=True): super().__init__() - inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) + self.dim = dim + self.base = base + self.train_seq_len = train_seq_len + self.yarn = yarn + 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 + self._cos_cached = None + self._sin_cached = None - def forward(self, seq_len: int, device: torch.device, dtype: torch.dtype) -> tuple[Tensor, Tensor]: + def forward(self, seq_len, device, dtype): if ( self._cos_cached is None or self._sin_cached is None - or self._seq_len_cached != seq_len + or self._seq_len_cached < seq_len or self._cos_cached.device != device ): - t = torch.arange(seq_len, device=device, dtype=self.inv_freq.dtype) - freqs = torch.outer(t, self.inv_freq.to(device)) - self._cos_cached = freqs.cos()[None, None, :, :] - self._sin_cached = freqs.sin()[None, None, :, :] + rd = self.rope_dims + if self.yarn and 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.float().to(device) + t = torch.arange(seq_len, device=device, dtype=torch.float32) + 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) + return self._cos_cached[:, :seq_len].to(dtype=dtype), self._sin_cached[:, :seq_len].to(dtype=dtype) -def apply_rotary_emb(x: Tensor, cos: Tensor, sin: Tensor) -> Tensor: +def apply_rotary_emb(x, cos, sin, rope_dims=0): + 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) + 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, + self, dim, num_heads, num_kv_heads, rope_base, qk_gain_init, train_seq_len, yarn=True ): super().__init__() if dim % num_heads != 0: @@ -571,553 +634,2255 @@ def __init__( self.head_dim = dim // num_heads if self.head_dim % 2 != 0: raise ValueError("head_dim must be even for RoPE") - kv_dim = self.num_kv_heads * self.head_dim - self.c_q = CastedLinear(dim, dim, bias=False) - self.c_k = CastedLinear(dim, kv_dim, bias=False) - self.c_v = CastedLinear(dim, kv_dim, bias=False) - self.proj = CastedLinear(dim, dim, bias=False) - self.proj._zero_init = True - self.q_gain = nn.Parameter(torch.full((num_heads,), qk_gain_init, dtype=torch.float32)) - self.rotary = Rotary(self.head_dim, base=rope_base) - - def forward(self, x: Tensor) -> Tensor: + self.q_gain = nn.Parameter( + torch.full((num_heads,), qk_gain_init, dtype=torch.float32) + ) + self.rope_dims = 0 + self.rotary = Rotary(self.head_dim, base=rope_base, train_seq_len=train_seq_len, yarn=yarn) + self.use_xsa = False + + def _xsa_efficient(self, y, v): + B, T, H, D = y.shape + Hkv = v.size(-2) + group = H // Hkv + y_g = y.reshape(B, T, Hkv, group, D) + vn = F.normalize(v, dim=-1).unsqueeze(-2) + proj = (y_g * vn).sum(dim=-1, keepdim=True) * vn + return (y_g - proj).reshape(B, T, H, D) + + def forward(self, x, q_w, k_w, v_w, out_w, cu_seqlens=None, max_seqlen=0): bsz, seqlen, dim = x.shape - q = self.c_q(x).reshape(bsz, seqlen, self.num_heads, self.head_dim).transpose(1, 2) - k = self.c_k(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) - v = self.c_v(x).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim).transpose(1, 2) + 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)).reshape(bsz, seqlen, self.num_kv_heads, self.head_dim) 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) - k = apply_rotary_emb(k, cos, sin) - q = q * self.q_gain.to(dtype=q.dtype)[None, :, None, None] - y = F.scaled_dot_product_attention( - q, - k, - v, - attn_mask=None, - is_causal=True, - enable_gqa=(self.num_kv_heads != self.num_heads), - ) - y = y.transpose(1, 2).contiguous().reshape(bsz, seqlen, dim) - return self.proj(y) + 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 cu_seqlens is not None: + y = flash_attn_varlen_func( + q[0], + k[0], + v[0], + cu_seqlens_q=cu_seqlens, + cu_seqlens_k=cu_seqlens, + max_seqlen_q=max_seqlen, + max_seqlen_k=max_seqlen, + causal=True, + window_size=(-1, -1), + )[None] + else: + y = flash_attn_3_func(q, k, v, causal=True) + if self.use_xsa: + y = self._xsa_efficient(y, v) + y = y.reshape(bsz, seqlen, dim) + self._last_proj_input = y.detach() if getattr(self, "_calib", False) else None + return F.linear(y, out_w.to(x.dtype)) + + +# Per-layer xIELU coefficients discovered by convergence loop (Run 2, 8xH100, seed 1337). +# Activation: torch.where(x > 0, ap*x² + bp*x, an*x² + bn*x) +QK_GAIN_INIT_PER_LAYER = [2.3495, 2.8818, 2.7627, 2.8148, 2.7893, 2.8762, 2.5657, 2.7206, 2.6426, 2.2737, 1.9741] + +XIELU_AP = [0.103, 0.196, 1.415, 1.196, 1.485, 1.546, 1.337, 1.727, 1.495, 0.988, 0.917] +XIELU_AN = [0.39, 0.578, 0.363, 0.491, 0.536, 0.548, 0.579, 0.983, 1.058, 0.935, 0.845] +XIELU_BP = [0.126, 0.07, 0.0, 0.0, 0.0, 0.002, 0.017, 0.067, 0.005, 0.058, 0.568] +XIELU_BN = [0.785, 0.638, 0.405, 0.377, 0.314, 0.289, 0.313, 0.571, 0.42, 0.286, 0.52] class MLP(nn.Module): - # relu^2 MLP from the original modded-nanogpt setup - def __init__(self, dim: int, mlp_mult: int): + def __init__(self, dim, mlp_mult, layer_idx=0): super().__init__() - hidden = mlp_mult * dim - self.fc = CastedLinear(dim, hidden, bias=False) - self.proj = CastedLinear(hidden, dim, bias=False) - self.proj._zero_init = True - - def forward(self, x: Tensor) -> Tensor: - x = torch.relu(self.fc(x)) - return self.proj(x.square()) + self.use_fused = True + idx = min(layer_idx, len(XIELU_AP) - 1) + self.ap = XIELU_AP[idx] + self.an = XIELU_AN[idx] + self.bp = XIELU_BP[idx] + self.bn = XIELU_BN[idx] + + def forward(self, x, up_w, down_w): + if self.training and self.use_fused: + return FusedXieluMLP(x, up_w.to(x.dtype), down_w.to(x.dtype), + self.ap, self.an, self.bp, self.bn) + h = F.linear(x, up_w.to(x.dtype)) + hidden = h * torch.where(h > 0, self.ap * h + self.bp, + self.an * h + self.bn) + self._last_down_input = hidden.detach() if getattr(self, "_calib", False) else None + return F.linear(hidden, down_w.to(x.dtype)) + + +def log_qk_gain_converged(log0, model): + """Print per-layer q_gain mean values for convergence tracking.""" + qk_vals = [] + for i, block in enumerate(model.blocks): + v = block.attn.q_gain.detach().cpu().mean().item() + log0(f"qk_gain:layer {i}: mean={v:.4f}") + qk_vals.append(round(v, 4)) + qk_str = ", ".join(f"{v}" for v in qk_vals) + log0(f"QK_GAIN_INIT_PER_LAYER = [{qk_str}]") 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, + dim, + num_heads, + num_kv_heads, + mlp_mult, + rope_base, + qk_gain_init, + train_seq_len, + layer_idx=0, + ln_scale=False, + yarn=True, ): super().__init__() self.attn_norm = RMSNorm() self.mlp_norm = RMSNorm() - self.attn = CausalSelfAttention(dim, num_heads, num_kv_heads, rope_base, qk_gain_init) - self.mlp = MLP(dim, mlp_mult) + self.attn = CausalSelfAttention( + dim, num_heads, num_kv_heads, rope_base, qk_gain_init, train_seq_len, yarn=yarn + ) + self.mlp = MLP(dim, mlp_mult, layer_idx=layer_idx) self.attn_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) self.mlp_scale = nn.Parameter(torch.ones(dim, dtype=torch.float32)) - self.resid_mix = nn.Parameter(torch.stack((torch.ones(dim), torch.zeros(dim))).float()) + self.resid_mix = nn.Parameter( + torch.stack((torch.ones(dim), torch.zeros(dim))).float() + ) + self.ln_scale_factor = 1.0 / math.sqrt(layer_idx + 1) if ln_scale else 1.0 - def forward(self, x: Tensor, x0: Tensor) -> Tensor: + def forward(self, x, x0, q_w, k_w, v_w, out_w, up_w, down_w, cu_seqlens=None, max_seqlen=0): mix = self.resid_mix.to(dtype=x.dtype) - x = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 - attn_out = self.attn(self.attn_norm(x)) - x = x + self.attn_scale.to(dtype=x.dtype)[None, None, :] * attn_out - x = x + self.mlp_scale.to(dtype=x.dtype)[None, None, :] * self.mlp(self.mlp_norm(x)) - return x - + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + attn_out = self.attn( + self.attn_norm(x_in) * self.ln_scale_factor, + q_w, k_w, v_w, out_w, + cu_seqlens=cu_seqlens, + max_seqlen=max_seqlen, + ) + 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) + return x_out 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, - ): + def __init__(self, h): super().__init__() - 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.tok_emb = nn.Embedding(vocab_size, 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)) + if h.logit_softcap <= 0.0: + raise ValueError(f"logit_softcap must be positive, got {h.logit_softcap}") + self.tie_embeddings = h.tie_embeddings + self.tied_embed_init_std = h.tied_embed_init_std + self.logit_softcap = h.logit_softcap + self.tok_emb = nn.Embedding(h.vocab_size, h.embedding_dim) + if h.embedding_dim != h.model_dim: + self.embed_proj = CastedLinear(h.embedding_dim, h.model_dim, bias=False) + self.head_proj = CastedLinear(h.model_dim, h.embedding_dim, bias=False) + else: + self.embed_proj = None + self.head_proj = None + self.num_layers = h.num_layers + head_dim = h.model_dim // h.num_heads + kv_dim = h.num_kv_heads * head_dim + hidden_dim = int(h.mlp_mult * h.model_dim) + self.qo_bank = nn.Parameter(torch.empty(2 * h.num_layers, h.model_dim, h.model_dim)) + self.kv_bank = nn.Parameter(torch.empty(2 * h.num_layers, kv_dim, h.model_dim)) + self.mlp_up_bank = nn.Parameter(torch.empty(h.num_layers, hidden_dim, h.model_dim)) + self.mlp_down_bank = nn.Parameter(torch.empty(h.num_layers, h.model_dim, hidden_dim)) + self.num_encoder_layers = h.num_layers // 2 + self.num_decoder_layers = h.num_layers - self.num_encoder_layers self.blocks = nn.ModuleList( [ Block( - model_dim, - num_heads, - num_kv_heads, - mlp_mult, - rope_base, - qk_gain_init, + h.model_dim, + h.num_heads, + h.num_kv_heads, + h.mlp_mult, + h.rope_base, + QK_GAIN_INIT_PER_LAYER[i] if QK_GAIN_INIT_PER_LAYER is not None else h.qk_gain_init, + h.train_seq_len, + layer_idx=i, + ln_scale=h.ln_scale, + yarn=h.rope_yarn, ) - for i in range(num_layers) + for i in range(h.num_layers) ] ) + if h.rope_dims > 0: + head_dim = h.model_dim // h.num_heads + for block in self.blocks: + block.attn.rope_dims = h.rope_dims + block.attn.rotary = Rotary( + head_dim, + base=h.rope_base, + train_seq_len=h.train_seq_len, + rope_dims=h.rope_dims, + yarn=h.rope_yarn, + ) self.final_norm = RMSNorm() - self.lm_head = None if tie_embeddings else CastedLinear(model_dim, vocab_size, bias=False) + self.lm_head = ( + None + if h.tie_embeddings + else CastedLinear(h.embedding_dim, h.vocab_size, bias=False) + ) if self.lm_head is not None: self.lm_head._zero_init = True + if h.xsa_last_n > 0: + for i in range(max(0, h.num_layers - h.xsa_last_n), h.num_layers): + self.blocks[i].attn.use_xsa = True + self.looping_active = False + if h.num_loops > 0: + loop_seg = list(range(h.loop_start, h.loop_end + 1)) + all_indices = list(range(h.loop_start)) + for _ in range(h.num_loops + 1): + all_indices.extend(loop_seg) + all_indices.extend(range(h.loop_end + 1, h.num_layers)) + num_enc = len(all_indices) // 2 + self.encoder_indices = all_indices[:num_enc] + self.decoder_indices = all_indices[num_enc:] + else: + self.encoder_indices = list(range(self.num_encoder_layers)) + self.decoder_indices = list(range(self.num_encoder_layers, h.num_layers)) + self.num_skip_weights = min( + len(self.encoder_indices), len(self.decoder_indices) + ) + self.skip_weights = nn.Parameter( + torch.ones(self.num_skip_weights, h.model_dim, dtype=torch.float32) + ) + self.skip_gates = ( + nn.Parameter( + torch.zeros(self.num_skip_weights, h.model_dim, dtype=torch.float32) + ) + if h.skip_gates_enabled + else None + ) + self.parallel_start_layer = h.parallel_start_layer + self.parallel_final_lane = h.parallel_final_lane.lower() + self.parallel_post_lambdas = nn.Parameter( + torch.ones(h.num_layers, 2, 2, dtype=torch.float32) + ) + self.parallel_resid_lambdas = nn.Parameter( + torch.full((h.num_layers, 2), 1.1, dtype=torch.float32) + ) self._init_weights() - def _init_weights(self) -> None: + def _init_weights(self): if self.tie_embeddings: nn.init.normal_(self.tok_emb.weight, mean=0.0, std=self.tied_embed_init_std) - for module in self.modules(): - if isinstance(module, nn.Linear) and getattr(module, "_zero_init", False): - nn.init.zeros_(module.weight) + n = self.num_layers + proj_scale = 1.0 / math.sqrt(2 * n) + for i in range(n): + nn.init.orthogonal_(self.qo_bank.data[i], gain=1.0) + nn.init.zeros_(self.qo_bank.data[n + i]) + self.qo_bank.data[n + i].mul_(proj_scale) + nn.init.orthogonal_(self.kv_bank.data[i], gain=1.0) + nn.init.orthogonal_(self.kv_bank.data[n + i], gain=1.0) + nn.init.orthogonal_(self.mlp_up_bank.data[i], gain=1.0) + nn.init.zeros_(self.mlp_down_bank.data[i]) + self.mlp_down_bank.data[i].mul_(proj_scale) + 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 _bank_weights(self, i): + n = self.num_layers + return ( + 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], + ) - def forward(self, input_ids: Tensor, target_ids: Tensor) -> Tensor: + def _parallel_block( + self, block_idx, lane0, lane1, x0, + q_w, k_w, v_w, out_w, up_w, down_w, + cu_seqlens=None, max_seqlen=0, + ): + block = self.blocks[block_idx] + mix = block.resid_mix.to(dtype=lane0.dtype) + attn_read = mix[0][None, None, :] * lane0 + mix[1][None, None, :] * x0 + attn_out = block.attn( + block.attn_norm(attn_read) * block.ln_scale_factor, + q_w, k_w, v_w, out_w, + cu_seqlens=cu_seqlens, max_seqlen=max_seqlen, + ) + attn_out = block.attn_scale.to(dtype=attn_out.dtype)[None, None, :] * attn_out + mlp_read = mix[0][None, None, :] * lane1 + mix[1][None, None, :] * x0 + mlp_out = block.mlp_scale.to(dtype=lane1.dtype)[None, None, :] * block.mlp( + block.mlp_norm(mlp_read) * block.ln_scale_factor, up_w, down_w + ) + attn_resid = self.parallel_resid_lambdas[block_idx, 0].to(dtype=lane0.dtype) + attn_post = self.parallel_post_lambdas[block_idx, 0].to(dtype=lane0.dtype) + mlp_resid = self.parallel_resid_lambdas[block_idx, 1].to(dtype=lane0.dtype) + mlp_post = self.parallel_post_lambdas[block_idx, 1].to(dtype=lane0.dtype) + lane0 = attn_resid * lane0 + attn_post[0] * attn_out + mlp_post[0] * mlp_out + lane1 = mlp_resid * lane1 + attn_post[1] * attn_out + mlp_post[1] * mlp_out + return lane0, lane1 + + def _final_parallel_hidden(self, lane0, lane1): + if self.parallel_final_lane == "mlp": + return lane1 + if self.parallel_final_lane == "attn": + return lane0 + return 0.5 * (lane0 + lane1) + + def forward_logits(self, input_ids, cu_seqlens=None, max_seqlen=0): x = self.tok_emb(input_ids) x = F.rms_norm(x, (x.size(-1),)) + if self.embed_proj is not None: + x = self.embed_proj(x) x0 = x - skips: list[Tensor] = [] - - # First half stores skips; second half reuses them in reverse order. - for i in range(self.num_encoder_layers): - x = self.blocks[i](x, x0) + skips = [] + enc_iter = ( + self.encoder_indices + if self.looping_active + else range(self.num_encoder_layers) + ) + dec_iter = ( + self.decoder_indices + if self.looping_active + else range( + self.num_encoder_layers, + self.num_encoder_layers + self.num_decoder_layers, + ) + ) + for i in enc_iter: + q_w, k_w, v_w, out_w, up_w, down_w = self._bank_weights(i) + x = self.blocks[i](x, x0, q_w, k_w, v_w, out_w, up_w, down_w, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen) skips.append(x) - for i in range(self.num_decoder_layers): - if skips: - x = x + self.skip_weights[i].to(dtype=x.dtype)[None, None, :] * skips.pop() - x = self.blocks[self.num_encoder_layers + i](x, x0) - - x = self.final_norm(x).reshape(-1, x.size(-1)) - targets = target_ids.reshape(-1) + psl = self.parallel_start_layer + lane0 = None + lane1 = None + for skip_idx, i in enumerate(dec_iter): + q_w, k_w, v_w, out_w, up_w, down_w = self._bank_weights(i) + if i >= psl and psl > 0: + if lane0 is None: + lane0 = x + lane1 = x + if skip_idx < self.num_skip_weights and skips: + skip = skips.pop() + w = self.skip_weights[skip_idx].to(dtype=lane0.dtype)[None, None, :] + if self.skip_gates is not None: + g = torch.sigmoid(self.skip_gates[skip_idx].to(dtype=lane0.dtype))[None, None, :] + lane0 = torch.lerp(w * skip, lane0, g) + else: + lane0 = lane0 + w * skip + lane0, lane1 = self._parallel_block( + i, lane0, lane1, x0, q_w, k_w, v_w, out_w, up_w, down_w, + cu_seqlens=cu_seqlens, max_seqlen=max_seqlen, + ) + else: + if skip_idx < self.num_skip_weights and skips: + scaled_skip = ( + self.skip_weights[skip_idx].to(dtype=x.dtype)[None, None, :] + * skips.pop() + ) + if self.skip_gates is not None: + g = torch.sigmoid(self.skip_gates[skip_idx].to(dtype=x.dtype))[None, None, :] + x = torch.lerp(scaled_skip, x, g) + else: + x = x + scaled_skip + x = self.blocks[i](x, x0, q_w, k_w, v_w, out_w, up_w, down_w, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen) + if lane0 is not None: + x = self._final_parallel_hidden(lane0, lane1) + x = self.final_norm(x) + if self.head_proj is not None: + x = self.head_proj(x) if self.tie_embeddings: logits_proj = F.linear(x, 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) - logits = self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) - return F.cross_entropy(logits.float(), targets, reduction="mean") - + return self.logit_softcap * torch.tanh(logits_proj / self.logit_softcap) -# ----------------------------- -# TRAINING -# ----------------------------- + def forward(self, input_ids, target_ids, cu_seqlens=None, max_seqlen=0): + logits = self.forward_logits( + input_ids, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen + ) + return F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + target_ids.reshape(-1), + reduction="mean", + ) -def main() -> None: - global zeropower_via_newtonschulz5 + def forward_ttt(self, input_ids, target_ids, lora): + x = self.tok_emb(input_ids) + x = F.rms_norm(x, (x.size(-1),)) + if self.embed_proj is not None: + x = self.embed_proj(x) + x0 = x + skips = [] + enc_iter = ( + self.encoder_indices + if self.looping_active + else list(range(self.num_encoder_layers)) + ) + dec_iter = ( + self.decoder_indices + if self.looping_active + else list( + range( + self.num_encoder_layers, + self.num_encoder_layers + self.num_decoder_layers, + ) + ) + ) + slot = 0 + for i in enc_iter: + q_w, k_w, v_w, out_w, up_w, down_w = self._bank_weights(i) + x = self._block_with_lora(self.blocks[i], x, x0, lora, slot, q_w, k_w, v_w, out_w, up_w, down_w) + slot += 1 + skips.append(x) + psl = self.parallel_start_layer + lane0 = None + lane1 = None + for skip_idx, i in enumerate(dec_iter): + q_w, k_w, v_w, out_w, up_w, down_w = self._bank_weights(i) + if i >= psl and psl > 0: + if lane0 is None: + lane0 = x + lane1 = x + if skip_idx < self.num_skip_weights and skips: + skip = skips.pop() + w = self.skip_weights[skip_idx].to(dtype=lane0.dtype)[None, None, :] + if self.skip_gates is not None: + g = torch.sigmoid(self.skip_gates[skip_idx].to(dtype=lane0.dtype))[None, None, :] + lane0 = torch.lerp(w * skip, lane0, g) + else: + lane0 = lane0 + w * skip + lane0, lane1 = self._parallel_block_with_lora( + i, lane0, lane1, x0, lora, slot, + q_w, k_w, v_w, out_w, up_w, down_w, + ) + else: + if skip_idx < self.num_skip_weights and skips: + scaled_skip = ( + self.skip_weights[skip_idx].to(dtype=x.dtype)[None, None, :] + * skips.pop() + ) + if self.skip_gates is not None: + g = torch.sigmoid(self.skip_gates[skip_idx].to(dtype=x.dtype))[None, None, :] + x = torch.lerp(scaled_skip, x, g) + else: + x = x + scaled_skip + x = self._block_with_lora(self.blocks[i], x, x0, lora, slot, q_w, k_w, v_w, out_w, up_w, down_w) + slot += 1 + if lane0 is not None: + x = self._final_parallel_hidden(lane0, lane1) + x = self.final_norm(x) + if self.head_proj is not None: + x = self.head_proj(x) + if self.tie_embeddings: + logits = F.linear(x, self.tok_emb.weight) + else: + logits = self.lm_head(x) + logits = logits + lora.lm_head_lora(x) + logits = self.logit_softcap * torch.tanh(logits / self.logit_softcap) + bsz, sl, V = logits.shape + return F.cross_entropy( + logits.float().reshape(-1, V), target_ids.reshape(-1), reduction="none" + ).reshape(bsz, sl) + + def _block_with_lora(self, block, x, x0, lora, slot, q_w, k_w, v_w, out_w, up_w, down_w): + mix = block.resid_mix.to(dtype=x.dtype) + x_in = mix[0][None, None, :] * x + mix[1][None, None, :] * x0 + n = block.attn_norm(x_in) * block.ln_scale_factor + attn = block.attn + bsz, seqlen, dim = n.shape + q = (F.linear(n, q_w.to(n.dtype)) + lora.q_loras[slot](n)).reshape( + bsz, seqlen, attn.num_heads, attn.head_dim + ) + k = F.linear(n, k_w.to(n.dtype)) + if lora.k_loras is not None: + k = k + lora.k_loras[slot](n) + k = k.reshape(bsz, seqlen, attn.num_kv_heads, attn.head_dim) + v = (F.linear(n, v_w.to(n.dtype)) + lora.v_loras[slot](n)).reshape( + bsz, seqlen, attn.num_kv_heads, attn.head_dim + ) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = attn.rotary(seqlen, n.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, attn.rope_dims) + k = apply_rotary_emb(k, cos, sin, attn.rope_dims) + q = q * attn.q_gain.to(dtype=q.dtype)[None, None, :, None] + y = flash_attn_3_func(q, k, v, causal=True) + if attn.use_xsa: + y = attn._xsa_efficient(y, v) + y = y.reshape(bsz, seqlen, dim) + attn_out = F.linear(y, out_w.to(n.dtype)) + if lora.o_loras is not None: + attn_out = attn_out + lora.o_loras[slot](n) + x_out = x_in + block.attn_scale.to(dtype=x_in.dtype)[None, None, :] * attn_out + mlp_n = block.mlp_norm(x_out) * block.ln_scale_factor + mlp_out = block.mlp(mlp_n, up_w, down_w) + if lora.mlp_loras is not None: + mlp_out = mlp_out + lora.mlp_loras[slot](mlp_n) + x_out = x_out + block.mlp_scale.to(dtype=x_out.dtype)[None, None, :] * mlp_out + return x_out + + def _parallel_block_with_lora( + self, block_idx, lane0, lane1, x0, lora, slot, + q_w, k_w, v_w, out_w, up_w, down_w, + ): + block = self.blocks[block_idx] + mix = block.resid_mix.to(dtype=lane0.dtype) + attn_read = mix[0][None, None, :] * lane0 + mix[1][None, None, :] * x0 + n = block.attn_norm(attn_read) * block.ln_scale_factor + attn = block.attn + bsz, seqlen, dim = n.shape + q = (F.linear(n, q_w.to(n.dtype)) + lora.q_loras[slot](n)).reshape( + bsz, seqlen, attn.num_heads, attn.head_dim + ) + k = F.linear(n, k_w.to(n.dtype)) + if lora.k_loras is not None: + k = k + lora.k_loras[slot](n) + k = k.reshape(bsz, seqlen, attn.num_kv_heads, attn.head_dim) + v = (F.linear(n, v_w.to(n.dtype)) + lora.v_loras[slot](n)).reshape( + bsz, seqlen, attn.num_kv_heads, attn.head_dim + ) + q = F.rms_norm(q, (q.size(-1),)) + k = F.rms_norm(k, (k.size(-1),)) + cos, sin = attn.rotary(seqlen, n.device, q.dtype) + q = apply_rotary_emb(q, cos, sin, attn.rope_dims) + k = apply_rotary_emb(k, cos, sin, attn.rope_dims) + q = q * attn.q_gain.to(dtype=q.dtype)[None, None, :, None] + y = flash_attn_3_func(q, k, v, causal=True) + if attn.use_xsa: + y = attn._xsa_efficient(y, v) + y = y.reshape(bsz, seqlen, dim) + attn_out = F.linear(y, out_w.to(n.dtype)) + if lora.o_loras is not None: + attn_out = attn_out + lora.o_loras[slot](n) + attn_out = block.attn_scale.to(dtype=attn_out.dtype)[None, None, :] * attn_out + mlp_read = mix[0][None, None, :] * lane1 + mix[1][None, None, :] * x0 + mlp_n = block.mlp_norm(mlp_read) * block.ln_scale_factor + mlp_out = block.mlp(mlp_n, up_w, down_w) + if lora.mlp_loras is not None: + mlp_out = mlp_out + lora.mlp_loras[slot](mlp_n) + mlp_out = block.mlp_scale.to(dtype=lane1.dtype)[None, None, :] * mlp_out + attn_resid = self.parallel_resid_lambdas[block_idx, 0].to(dtype=lane0.dtype) + attn_post = self.parallel_post_lambdas[block_idx, 0].to(dtype=lane0.dtype) + mlp_resid = self.parallel_resid_lambdas[block_idx, 1].to(dtype=lane0.dtype) + mlp_post = self.parallel_post_lambdas[block_idx, 1].to(dtype=lane0.dtype) + lane0 = attn_resid * lane0 + attn_post[0] * attn_out + mlp_post[0] * mlp_out + lane1 = mlp_resid * lane1 + attn_post[1] * attn_out + mlp_post[1] * mlp_out + return lane0, lane1 + + +class BatchedLinearLoRA(nn.Module): + def __init__(self, bsz, in_features, out_features, rank): + super().__init__() + self._bound = 1.0 / math.sqrt(in_features) + self.A = nn.Parameter( + torch.empty(bsz, rank, in_features).uniform_(-self._bound, self._bound) + ) + self.B = nn.Parameter(torch.zeros(bsz, out_features, rank)) - code = Path(__file__).read_text(encoding="utf-8") - args = Hyperparameters() - zeropower_via_newtonschulz5 = torch.compile(zeropower_via_newtonschulz5) + def reset(self): + with torch.no_grad(): + self.A.uniform_(-self._bound, self._bound) + self.B.zero_() - # ----------------------------- - # DISTRIBUTED + CUDA SETUP - # ----------------------------- + def forward(self, x): + return (x @ self.A.transpose(1, 2)) @ self.B.transpose(1, 2) - 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 <= 0: - raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") - if 8 % world_size != 0: - raise ValueError(f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral") - 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 - # Fast math knobs - 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 +class BatchedTTTLoRA(nn.Module): + def __init__(self, bsz, model, rank, k_lora=True, mlp_lora=True, o_lora=True): + super().__init__() + self.bsz = bsz + dim = model.qo_bank.shape[-1] + vocab = model.tok_emb.num_embeddings + if getattr(model, "looping_active", False): + num_slots = len(model.encoder_indices) + len(model.decoder_indices) + else: + num_slots = len(model.blocks) + kv_dim = model.blocks[0].attn.num_kv_heads * ( + dim // model.blocks[0].attn.num_heads + ) + embed_dim = model.tok_emb.embedding_dim + self.lm_head_lora = BatchedLinearLoRA(bsz, embed_dim, vocab, rank) + self.q_loras = nn.ModuleList( + [BatchedLinearLoRA(bsz, dim, dim, rank) for _ in range(num_slots)] + ) + self.v_loras = nn.ModuleList( + [BatchedLinearLoRA(bsz, dim, kv_dim, rank) for _ in range(num_slots)] + ) + self.k_loras = ( + nn.ModuleList( + [BatchedLinearLoRA(bsz, dim, kv_dim, rank) for _ in range(num_slots)] + ) + if k_lora + else None + ) + self.mlp_loras = ( + nn.ModuleList( + [BatchedLinearLoRA(bsz, dim, dim, rank) for _ in range(num_slots)] + ) + if mlp_lora + else None + ) + self.o_loras = ( + nn.ModuleList( + [BatchedLinearLoRA(bsz, dim, dim, rank) for _ in range(num_slots)] + ) + if o_lora + else None + ) - enable_cudnn_sdp(False) - enable_flash_sdp(True) - enable_mem_efficient_sdp(False) - enable_math_sdp(False) + def reset(self): + with torch.no_grad(): + self.lm_head_lora.reset() + for loras in [self.q_loras, self.v_loras, self.k_loras, + self.mlp_loras, self.o_loras]: + if loras is not None: + for lora in loras: + lora.reset() + + +def classify_param(name): + if "tok_emb" in name or "lm_head" in name: + return "embed" + if ".mlp." in name: + return "mlp" + if ".attn." in name or (".proj." in name and ".mlp." not in name): + return "attn" + return "other" + + +@torch.compile +def zeropower_via_newtonschulz5(G, steps=10, eps=1e-07): + a, b, c = 3.4445, -4.775, 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 - 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: +class Muon(torch.optim.Optimizer): + def __init__( + self, + params, + lr, + momentum, + backend_steps, + nesterov=True, + weight_decay=0.0, + row_normalize=False, + ): + super().__init__( + params, + dict( + lr=lr, + momentum=momentum, + backend_steps=backend_steps, + nesterov=nesterov, + weight_decay=weight_decay, + row_normalize=row_normalize, + ), + ) + 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, + }) + self._bank_meta.sort(key=lambda m: -m["p"].numel()) + self._built = True + + def launch_reduce_scatters(self): + if not self._built: + self._build() + if not self._distributed: return - if console: - print(msg) - if logfile is not None: - with open(logfile, "a", encoding="utf-8") as f: - print(msg, file=f) + 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): + 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) + row_normalize = group.get("row_normalize", False) + prev_ag_handle = None + prev_m = None + sharded = self._distributed and hasattr(self, "_rs_futures") + for idx, 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[idx] is not None: + self._rs_futures[idx].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 + if row_normalize: + rn = update.float().norm(dim=-1, keepdim=True).clamp_min(1e-07) + update = update / rn.to(update.dtype) + 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 - log0(code, console=False) - log0("=" * 100, console=False) - 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) - # ----------------------------- - # TOKENIZER + VALIDATION METRIC SETUP - # ----------------------------- +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,skip_gates,parallel_post_lambdas,parallel_resid_lambdas", + ).split(",") + if pattern +) - 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())}" +PACKED_REPLICATED_GRAD_MAX_NUMEL = 1 << 15 + + +class Optimizers: + def __init__(self, h, base_model): + 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) + if base_model.skip_gates is not None and base_model.skip_gates.numel() > 0: + scalar_params.append(base_model.skip_gates) + if base_model.parallel_post_lambdas is not None: + scalar_params.append(base_model.parallel_post_lambdas) + if base_model.parallel_resid_lambdas is not None: + scalar_params.append(base_model.parallel_resid_lambdas) + token_lr = h.tied_embed_lr if h.tie_embeddings else h.embed_lr + tok_params = [ + {"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr} + ] + self.optimizer_tok = torch.optim.AdamW( + tok_params, + betas=(h.beta1, h.beta2), + eps=h.adam_eps, + weight_decay=h.embed_wd, + fused=True, ) - dataset_dir = Path(args.data_path).resolve() - actual_train_files = len(list(dataset_dir.glob("fineweb_train_*.bin"))) - val_tokens = load_validation_tokens(args.val_files, args.train_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}") - 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}") - - # ----------------------------- - # MODEL + OPTIMIZER SETUP - # ----------------------------- - - 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, - ).to(device).bfloat16() - for module in base_model.modules(): + self.optimizer_muon = Muon( + matrix_params, + lr=h.matrix_lr, + momentum=h.muon_momentum, + backend_steps=h.muon_backend_steps, + weight_decay=h.muon_wd, + row_normalize=h.muon_row_normalize, + ) + for group in self.optimizer_muon.param_groups: + group["base_lr"] = h.matrix_lr + self.optimizer_scalar = torch.optim.AdamW( + [{"params": scalar_params, "lr": h.scalar_lr, "base_lr": h.scalar_lr}], + betas=(h.beta1, h.beta2), + eps=h.adam_eps, + weight_decay=h.adam_wd, + fused=True, + ) + self.optimizers = [ + self.optimizer_tok, + self.optimizer_muon, + self.optimizer_scalar, + ] + if base_model.lm_head is not None: + self.optimizer_head = torch.optim.Adam( + [ + { + "params": [base_model.lm_head.weight], + "lr": h.head_lr, + "base_lr": h.head_lr, + } + ], + betas=(h.beta1, h.beta2), + eps=h.adam_eps, + fused=True, + ) + self.optimizers.insert(1, self.optimizer_head) + else: + self.optimizer_head = None + self.replicated_params = list(tok_params[0]["params"]) + self.replicated_params.extend(scalar_params) + if base_model.lm_head is not None: + self.replicated_params.append(base_model.lm_head.weight) + self.replicated_large_params = [] + self.replicated_packed_params = [] + for p in self.replicated_params: + if p.numel() <= PACKED_REPLICATED_GRAD_MAX_NUMEL: + self.replicated_packed_params.append(p) + else: + self.replicated_large_params.append(p) + + def __iter__(self): + return iter(self.optimizers) + + def zero_grad_all(self): + for opt in self.optimizers: + opt.zero_grad(set_to_none=True) + + def _all_reduce_packed_grads(self): + grads_by_key = collections.defaultdict(list) + for p in self.replicated_packed_params: + if p.grad is not None: + grads_by_key[(p.grad.device, p.grad.dtype)].append(p.grad) + for grads in grads_by_key.values(): + flat = torch.empty( + sum(g.numel() for g in grads), + device=grads[0].device, + dtype=grads[0].dtype, + ) + offset = 0 + for g in grads: + n = g.numel() + flat[offset : offset + n].copy_(g.contiguous().view(-1)) + offset += n + dist.all_reduce(flat, op=dist.ReduceOp.AVG) + offset = 0 + for g in grads: + n = g.numel() + g.copy_(flat[offset : offset + n].view_as(g)) + offset += n + + def step(self, distributed=False): + self.optimizer_muon.launch_reduce_scatters() + if distributed: + reduce_handles = [ + dist.all_reduce(p.grad, op=dist.ReduceOp.AVG, async_op=True) + for p in self.replicated_large_params + if p.grad is not None + ] + self._all_reduce_packed_grads() + for handle in reduce_handles: + handle.wait() + self.optimizer_tok.step() + self.optimizer_scalar.step() + if self.optimizer_head is not None: + self.optimizer_head.step() + self.optimizer_muon.step() + self.zero_grad_all() + + +def restore_fp32_params(model): + for module in model.modules(): if isinstance(module, CastedLinear): module.float() - restore_low_dim_params_to_fp32(base_model) - compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) - model: nn.Module = DDP(compiled_model, device_ids=[local_rank], broadcast_buffers=False) if distributed else compiled_model - - # Optimizer split: - # - token embedding (Adam) uses EMBED_LR - # - untied lm_head (Adam) uses HEAD_LR - # - matrix params in transformer blocks use MATRIX_LR via Muon - # - vectors/scalars use SCALAR_LR via Adam - block_named_params = list(base_model.blocks.named_parameters()) - matrix_params = [ - p - for name, p in block_named_params - if p.ndim == 2 and not any(pattern in name for pattern in CONTROL_TENSOR_NAME_PATTERNS) - ] - 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) - token_lr = args.tied_embed_lr if args.tie_embeddings else args.embed_lr - optimizer_tok = torch.optim.Adam( - [{"params": [base_model.tok_emb.weight], "lr": token_lr, "base_lr": token_lr}], - betas=(args.beta1, args.beta2), - eps=args.adam_eps, - fused=True, + for name, param in model.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() + if hasattr(model, "qo_bank"): + model.qo_bank.data = model.qo_bank.data.float() + model.kv_bank.data = model.kv_bank.data.float() + model.mlp_up_bank.data = model.mlp_up_bank.data.float() + model.mlp_down_bank.data = model.mlp_down_bank.data.float() + + +def collect_hessians(model, train_loader, h, device, n_calibration_batches=64): + hessians = {} + hooks = [] + for i, block in enumerate(model.blocks): + block.attn._calib = True + block.mlp._calib = True + block.mlp.use_fused = False + + def make_attn_hook(layer_idx): + def hook_fn(module, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + for suffix in ["c_q", "c_k", "c_v"]: + name = f"blocks.{layer_idx}.attn.{suffix}.weight" + if name not in hessians: + hessians[name] = torch.zeros( + x.shape[1], x.shape[1], dtype=torch.float32, device=device + ) + hessians[name].addmm_(x.T, x) + y = module._last_proj_input + if y is not None: + y = y.float() + if y.ndim == 3: + y = y.reshape(-1, y.shape[-1]) + name = f"blocks.{layer_idx}.attn.proj.weight" + if name not in hessians: + hessians[name] = torch.zeros( + y.shape[1], y.shape[1], dtype=torch.float32, device=device + ) + hessians[name].addmm_(y.T, y) + return hook_fn + + def make_mlp_hook(layer_idx): + def hook_fn(module, inp, out): + x = inp[0].detach().float() + if x.ndim == 3: + x = x.reshape(-1, x.shape[-1]) + name = f"blocks.{layer_idx}.mlp.fc.weight" + if name not in hessians: + hessians[name] = torch.zeros( + x.shape[1], x.shape[1], dtype=torch.float32, device=device + ) + hessians[name].addmm_(x.T, x) + h_act = module._last_down_input + if h_act is not None: + h_act = h_act.float() + if h_act.ndim == 3: + h_act = h_act.reshape(-1, h_act.shape[-1]) + name = f"blocks.{layer_idx}.mlp.proj.weight" + if name not in hessians: + hessians[name] = torch.zeros( + h_act.shape[1], h_act.shape[1], dtype=torch.float32, device=device + ) + hessians[name].addmm_(h_act.T, h_act) + return hook_fn + + for i, block in enumerate(model.blocks): + hooks.append(block.attn.register_forward_hook(make_attn_hook(i))) + hooks.append(block.mlp.register_forward_hook(make_mlp_hook(i))) + if model.tie_embeddings: + hook_module = ( + model.head_proj if model.head_proj is not None else model.final_norm + ) + + def make_output_hook(name): + def hook_fn(module, inp, out): + x = out.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], dtype=torch.float32, device=device + ) + hessians[name].addmm_(x.T, x) + return hook_fn + + hooks.append( + hook_module.register_forward_hook(make_output_hook("tok_emb.weight")) + ) + model.eval() + with torch.no_grad(): + for _ in range(n_calibration_batches): + x, _ = train_loader.next_batch(h.train_batch_tokens, h.grad_accum_steps) + model.forward_logits(x) + for hook in hooks: + hook.remove() + for i, block in enumerate(model.blocks): + block.attn._calib = False + block.mlp._calib = False + block.mlp.use_fused = True + for name in hessians: + hessians[name] = hessians[name].cpu() / n_calibration_batches + return hessians + + +def gptq_quantize_weight(w, H, clip_sigmas=3.0, clip_range=63, block_size=128): + W_orig = w.float().clone() + rows, cols = W_orig.shape + H = H.float().clone() + dead = torch.diag(H) == 0 + H[dead, dead] = 1 + damp = 0.01 * H.diag().mean() + H.diagonal().add_(damp) + perm = torch.argsort(H.diag(), descending=True) + invperm = torch.argsort(perm) + W_perm = W_orig[:, perm].clone() + W_perm[:, dead[perm]] = 0 + H = H[perm][:, perm] + Hinv = torch.cholesky_inverse(torch.linalg.cholesky(H)) + Hinv = torch.linalg.cholesky(Hinv, upper=True) + row_std = W_orig.std(dim=1) + s = (clip_sigmas * row_std / clip_range).clamp_min(1e-10).to(torch.float16) + sf = s.float() + Q = torch.zeros(rows, cols, dtype=torch.int8) + W_work = W_perm.clone() + for i1 in range(0, cols, block_size): + i2 = min(i1 + block_size, cols) + W_block = W_work[:, i1:i2].clone() + Hinv_block = Hinv[i1:i2, i1:i2] + Err = torch.zeros(rows, i2 - i1) + for j in range(i2 - i1): + w_col = W_block[:, j] + d = Hinv_block[j, j] + q_col = torch.clamp(torch.round(w_col / sf), -clip_range, clip_range) + Q[:, i1 + j] = q_col.to(torch.int8) + err = (w_col - q_col.float() * sf) / d + Err[:, j] = err + W_block[:, j:] -= err.unsqueeze(1) * Hinv_block[j, j:].unsqueeze(0) + if i2 < cols: + W_work[:, i2:] -= Err @ Hinv[i1:i2, i2:] + return Q[:, invperm], s + + +def gptq_mixed_quantize(state_dict, hessians, h): + result = {} + meta = {} + for (name, tensor) in state_dict.items(): + t = tensor.detach().cpu().contiguous() + 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 (float16)" + continue + if "tok_emb" in name: + cs = h.embed_clip_sigmas + elif ".mlp." in name: + cs = h.mlp_clip_sigmas + elif ".attn." in name: + cs = h.attn_clip_sigmas + else: + cs = h.matrix_clip_sigmas + bits = h.embed_bits if "tok_emb" in name else h.matrix_bits + q, s = gptq_quantize_weight( + t, hessians[name], clip_sigmas=cs, clip_range=2 ** (bits - 1) - 1 + ) + result[name + ".q"] = q + result[name + ".scale"] = s + meta[name] = f"gptq (int{bits})" + categories = collections.defaultdict(set) + for (name, cat) in meta.items(): + short = re.sub("\\.\\d+$", "", re.sub("blocks\\.\\d+", "blocks", name)) + categories[cat].add(short) + log("Quantized weights:") + for cat in sorted(categories): + log(f" {cat}: {', '.join(sorted(categories[cat]))}") + return result, meta + + +def dequantize_mixed(result, meta, template_sd): + out = {} + for (name, orig) in template_sd.items(): + info = meta.get(name) + if info is None: + continue + orig_dtype = orig.dtype + if "passthrough" in info: + 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: + out[name] = ( + q.float() * s.float().view(q.shape[0], *[1] * (q.ndim - 1)) + ).to(orig_dtype) + else: + out[name] = (q.float() * float(s.item())).to(orig_dtype) + return out + + +_BSHF_MAGIC = b"BSHF" + + +def _byte_shuffle(data, stride=2): + 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): + 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(data, compressor): + data = _byte_shuffle(data) + if compressor == "lzma": + return lzma.compress(data, preset=6) + elif compressor == "brotli": + import brotli + + return brotli.compress(data, quality=11) + raise ValueError(f"Unknown compressor: {compressor!r}") + + +def _decompress(data, compressor): + if compressor == "lzma": + raw = lzma.decompress(data) + elif compressor == "brotli": + import brotli + + raw = brotli.decompress(data) + else: + raise ValueError(f"Unknown compressor: {compressor!r}") + raw = _byte_unshuffle(raw) + return raw + + +def _unbank_state_dict(state_dict, num_layers): + sd = {} + n = num_layers + for k, v in state_dict.items(): + t = v.detach().cpu() + if k == "qo_bank": + for i in range(n): + sd[f"blocks.{i}.attn.c_q.weight"] = t[i] + sd[f"blocks.{i}.attn.proj.weight"] = t[n + i] + elif k == "kv_bank": + for i in range(n): + sd[f"blocks.{i}.attn.c_k.weight"] = t[i] + sd[f"blocks.{i}.attn.c_v.weight"] = t[n + i] + elif k == "mlp_up_bank": + for i in range(n): + sd[f"blocks.{i}.mlp.fc.weight"] = t[i] + elif k == "mlp_down_bank": + for i in range(n): + sd[f"blocks.{i}.mlp.proj.weight"] = t[i] + else: + sd[k] = t + return sd + + +def _rebank_state_dict(flat_sd, num_layers, model_dim, kv_dim, hidden_dim): + sd = {} + n = num_layers + sd["qo_bank"] = torch.zeros(2 * n, model_dim, model_dim) + sd["kv_bank"] = torch.zeros(2 * n, kv_dim, model_dim) + sd["mlp_up_bank"] = torch.zeros(n, hidden_dim, model_dim) + sd["mlp_down_bank"] = torch.zeros(n, model_dim, hidden_dim) + for i in range(n): + sd["qo_bank"][i] = flat_sd[f"blocks.{i}.attn.c_q.weight"] + sd["qo_bank"][n + i] = flat_sd[f"blocks.{i}.attn.proj.weight"] + sd["kv_bank"][i] = flat_sd[f"blocks.{i}.attn.c_k.weight"] + sd["kv_bank"][n + i] = flat_sd[f"blocks.{i}.attn.c_v.weight"] + sd["mlp_up_bank"][i] = flat_sd[f"blocks.{i}.mlp.fc.weight"] + sd["mlp_down_bank"][i] = flat_sd[f"blocks.{i}.mlp.proj.weight"] + for k, v in flat_sd.items(): + if not ( + k.startswith("blocks.") + and any( + p in k + for p in [ + ".attn.c_q.", ".attn.c_k.", ".attn.c_v.", + ".attn.proj.", ".mlp.fc.", ".mlp.proj.", + ] + ) + ): + sd[k] = v + return sd + + +def _compressed_code_size(code): + code_raw = code.encode("utf-8") + minified = subprocess.run( + ["pyminify", "--no-rename-locals", "--no-hoist-literals", "--remove-literal-statements", "-"], + input=code_raw, capture_output=True, check=True, + ).stdout + compressed = lzma.compress(minified) + encoded = base64.b85encode(compressed) + wrapper = b'import lzma as L,base64 as B\nexec(L.decompress(B.b85decode("' + encoded + b'")))\n' + return len(code_raw), len(wrapper) + + +def serialize(h, base_model, code): + code_bytes_uncompressed, code_bytes = _compressed_code_size(code) + if h.is_main_process: + torch.save(base_model.state_dict(), h.model_path) + model_bytes = os.path.getsize(h.model_path) + log(f"Serialized model: {model_bytes} bytes") + log(f"Code size (uncompressed): {code_bytes_uncompressed} bytes") + log(f"Code size (compressed): {code_bytes} bytes") + sd_cpu = _unbank_state_dict(base_model.state_dict(), h.num_layers) + device = torch.device("cuda", h.local_rank) + log("GPTQ:collecting Hessians from calibration data...") + t0 = time.perf_counter() + calib_loader = ShuffledSequenceLoader(h, device) + hessians = collect_hessians( + base_model, + calib_loader, + h, + device, + n_calibration_batches=h.gptq_calibration_batches, ) - optimizer_muon = Muon( - matrix_params, - lr=args.matrix_lr, - momentum=args.muon_momentum, - backend_steps=args.muon_backend_steps, + log(f"GPTQ:collected {len(hessians)} Hessians in {time.perf_counter()-t0:.1f}s") + quant_result, quant_meta = gptq_mixed_quantize(sd_cpu, hessians, h) + quant_buf = io.BytesIO() + torch.save({"w": quant_result, "m": quant_meta}, quant_buf) + quant_raw = quant_buf.getvalue() + quant_blob = _compress(quant_raw, h.compressor) + quant_file_bytes = len(quant_blob) + bytes_total = quant_file_bytes + code_bytes + if h.is_main_process: + with open(h.quantized_model_path, "wb") as f: + f.write(quant_blob) + log(f"Serialized model quantized+{h.compressor}: {quant_file_bytes} bytes") + log(f"Total submission size quantized+{h.compressor}: {bytes_total} bytes") + return bytes_total, quant_file_bytes + + +def deserialize(h, device): + eval_model = GPT(h).to(device).bfloat16() + restore_fp32_params(eval_model) + flat_template = _unbank_state_dict(eval_model.state_dict(), h.num_layers) + with open(h.quantized_model_path, "rb") as f: + quant_blob_disk = f.read() + quant_state = torch.load( + io.BytesIO(_decompress(quant_blob_disk, h.compressor)), map_location="cpu" ) - for group in optimizer_muon.param_groups: - group["base_lr"] = args.matrix_lr - optimizer_scalar = torch.optim.Adam( - [{"params": scalar_params, "lr": args.scalar_lr, "base_lr": args.scalar_lr}], - betas=(args.beta1, args.beta2), - eps=args.adam_eps, - fused=True, + deq_flat = dequantize_mixed(quant_state["w"], quant_state["m"], flat_template) + head_dim = h.model_dim // h.num_heads + kv_dim = h.num_kv_heads * head_dim + hidden_dim = int(h.mlp_mult * h.model_dim) + deq_state = _rebank_state_dict(deq_flat, h.num_layers, h.model_dim, kv_dim, hidden_dim) + eval_model.load_state_dict(deq_state, strict=True) + return eval_model + + +def _loss_bpb(loss_sum, token_count, byte_count): + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_count.item()) + return val_loss, val_bpb + + +def eval_val(h, device, val_data, model, forward_logits_fn=None): + seq_len = h.eval_seq_len + local_batch_tokens = h.val_batch_tokens // (h.world_size * h.grad_accum_steps) + if local_batch_tokens < seq_len: + raise ValueError( + f"VAL_BATCH_SIZE must provide at least one sequence per rank; got VAL_BATCH_SIZE={h.val_batch_tokens}, WORLD_SIZE={h.world_size}, GRAD_ACCUM_STEPS={h.grad_accum_steps}, seq_len={seq_len}" + ) + local_batch_seqs = local_batch_tokens // seq_len + total_seqs = (val_data.val_tokens.numel() - 1) // seq_len + seq_start = total_seqs * h.rank // h.world_size + seq_end = total_seqs * (h.rank + 1) // h.world_size + + # TODO: Don't truncate this. + seq_end = seq_start + ((seq_end - seq_start) // local_batch_seqs) * local_batch_seqs + + 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) + run_forward_logits = ( + (model.module.forward_logits if hasattr(model, "module") else model.forward_logits) + if forward_logits_fn is None + else forward_logits_fn ) - optimizers: list[torch.optim.Optimizer] = [optimizer_tok, optimizer_muon, optimizer_scalar] - 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, + model.eval() + global BOS_ID + if BOS_ID is None: + BOS_ID = 1 + with torch.no_grad(): + 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_data.val_tokens[raw_start:raw_end].to( + device=device, dtype=torch.int64, non_blocking=True + ) + x = local[:-1] + y = local[1:] + bos_pos = (x == BOS_ID).nonzero(as_tuple=True)[0].tolist() + cu_seqlens, max_seqlen = _build_cu_seqlens( + bos_pos, x.numel(), x.device, h.eval_seq_len, 64 + ) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + logits = run_forward_logits( + x[None], cu_seqlens=cu_seqlens, max_seqlen=max_seqlen + ).detach() + per_token_loss = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y.reshape(-1), + reduction="none", + ) + val_loss_sum += per_token_loss.to(torch.float64).sum() + val_token_count += float(y.numel()) + prev_ids = x + tgt_ids = y + token_bytes = val_data.base_bytes_lut[tgt_ids].to(dtype=torch.int16) + token_bytes += ( + val_data.has_leading_space_lut[tgt_ids] + & ~val_data.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) + model.train() + return _loss_bpb(val_loss_sum, val_token_count, val_byte_count) + + +def eval_val_sliding(h, device, val_data, base_model, forward_logits_fn=None, batch_seqs=32): + global BOS_ID + if BOS_ID is None: + BOS_ID = 1 + base_model.eval() + run_forward_logits = base_model.forward_logits if forward_logits_fn is None else forward_logits_fn + seq_len = h.eval_seq_len + stride = h.eval_stride + total_tokens = val_data.val_tokens.numel() - 1 + context_size = seq_len - stride + window_starts = [ws for ws in range(0, total_tokens, stride) + if ws + context_size < total_tokens] + total_windows = len(window_starts) + my_s = (total_windows * h.rank) // h.world_size + my_e = (total_windows * (h.rank + 1)) // h.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) + total_batches = (len(my_windows) + batch_seqs - 1) // batch_seqs + is_master = h.rank == 0 + cu_bucket = 64 + t_sw_start = time.perf_counter() + with torch.no_grad(): + for bi in range(0, len(my_windows), batch_seqs): + batch_idx = bi // batch_seqs + if is_master and (batch_idx % 50 == 0 or batch_idx == total_batches - 1): + elapsed = time.perf_counter() - t_sw_start + rl = float(loss_sum.item() / token_count.item()) if token_count.item() > 0 else 0.0 + rb = float((rl / math.log(2.0)) * token_count.item() / byte_count.item()) if byte_count.item() > 0 else 0.0 + log(f"sliding_progress: batch {batch_idx+1}/{total_batches} " + f"tokens:{int(token_count.item())} running_loss:{rl:.4f} running_bpb:{rb:.4f} " + f"elapsed:{elapsed:.1f}s") + batch_ws = my_windows[bi:bi + batch_seqs] + x_parts = [] + y_parts = [] + cu_starts = [] + score_ranges = [] + offset = 0 + for ws in batch_ws: + end = min(ws + seq_len, total_tokens) + wlen = end - ws + chunk_cpu = val_data.val_tokens[ws:end + 1] + bos_pos = (chunk_cpu[:-1] == BOS_ID).nonzero(as_tuple=True)[0].tolist() + if not bos_pos or bos_pos[0] != 0: + bos_pos = [0] + bos_pos + cu_starts.extend(offset + pos for pos in bos_pos) + chunk = chunk_cpu.to(dtype=torch.int64, device=device) + x_parts.append(chunk[:-1]) + y_parts.append(chunk[1:]) + score_ranges.append((offset, wlen, ws)) + offset += wlen + x_cat = torch.cat(x_parts, dim=0)[None] + y_cat = torch.cat(y_parts, dim=0) + boundaries = cu_starts + [offset] + padded_len = get_next_multiple_of_n(len(boundaries), cu_bucket) + cu_seqlens = torch.full((padded_len,), offset, dtype=torch.int32, device=device) + cu_seqlens[:len(boundaries)] = torch.tensor(boundaries, dtype=torch.int32, device=device) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits = run_forward_logits(x_cat, cu_seqlens=cu_seqlens, max_seqlen=seq_len) + flat_nll = F.cross_entropy( + logits.reshape(-1, logits.size(-1)).float(), + y_cat, + reduction="none", + ) + flat_x = x_cat.reshape(-1) + for off, wlen, ws in score_ranges: + s = 0 if ws == 0 else context_size + lo = off + s + hi = off + wlen + scored_nll = flat_nll[lo:hi].to(torch.float64) + loss_sum += scored_nll.sum() + token_count += float(hi - lo) + tgt = y_cat[lo:hi] + prev = flat_x[lo:hi] + tb = val_data.base_bytes_lut[tgt].to(torch.float64) + tb += (val_data.has_leading_space_lut[tgt] & ~val_data.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) + base_model.train() + return _loss_bpb(loss_sum, token_count, byte_count) + + +def _find_docs(all_tokens): + bos_positions = (all_tokens == BOS_ID).nonzero(as_tuple=True)[0].numpy() + docs = [] + for i in range(len(bos_positions)): + start = int(bos_positions[i]) + end = ( + int(bos_positions[i + 1]) + if i + 1 < len(bos_positions) + else all_tokens.numel() ) - optimizers.insert(1, optimizer_head) - - n_params = sum(p.numel() for p in base_model.parameters()) - log0(f"model_params:{n_params}") - 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}" + if i + 1 < len(bos_positions): + end += 1 + assert end - start >= 2 + docs.append((start, end - start)) + return docs + + +def _build_ttt_global_batches(doc_entries, h, ascending=False): + batch_size = h.ttt_batch_size + global_doc_entries = sorted(doc_entries, key=lambda x: x[1][1]) + global_batches = [ + global_doc_entries[i : i + batch_size] + for i in range(0, len(global_doc_entries), batch_size) + ] + indexed = list(enumerate(global_batches)) + if not ascending: + indexed.sort(key=lambda ib: -max(dl for _, (_, dl) in ib[1])) + return indexed + + +def _init_batch_counter(path): + with open(path, "wb") as f: + f.write((0).to_bytes(4, "little")) + + +def _claim_next_batch(counter_path, queue_len): + try: + with open(counter_path, "r+b") as f: + fcntl.flock(f, fcntl.LOCK_EX) + idx = int.from_bytes(f.read(4), "little") + f.seek(0) + f.write((idx + 1).to_bytes(4, "little")) + f.flush() + except FileNotFoundError: + return queue_len + return idx + + +def _compute_chunk_window(ci, pred_len, num_chunks, chunk_size, eval_seq_len): + chunk_end = pred_len if ci == num_chunks - 1 else (ci + 1) * chunk_size + win_start = max(0, chunk_end - eval_seq_len) + win_len = chunk_end - win_start + chunk_start = ci * chunk_size + chunk_offset = chunk_start - win_start + chunk_len = chunk_end - chunk_start + return win_start, win_len, chunk_offset, chunk_len + + +def _accumulate_bpb( + ptl, + x, + y, + chunk_offsets, + chunk_lens, + pos_idx, + base_bytes_lut, + has_leading_space_lut, + is_boundary_token_lut, + loss_sum, + byte_sum, + token_count, +): + pos = pos_idx[: x.size(1)].unsqueeze(0) + mask = ( + (chunk_lens.unsqueeze(1) > 0) + & (pos >= chunk_offsets.unsqueeze(1)) + & (pos < (chunk_offsets + chunk_lens).unsqueeze(1)) ) - 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}" + mask_f64 = mask.to(torch.float64) + tok_bytes = base_bytes_lut[y].to(torch.float64) + tok_bytes += (has_leading_space_lut[y] & ~is_boundary_token_lut[x]).to( + torch.float64 ) - log0(f"seed:{args.seed}") + loss_sum += (ptl.to(torch.float64) * mask_f64).sum() + byte_sum += (tok_bytes * mask_f64).sum() + token_count += chunk_lens.to(torch.float64).sum() + +def eval_val_ttt_lora(h, base_model, device, val_data, forward_ttt_train): + global BOS_ID + if BOS_ID is None: + BOS_ID = 1 + base_model.eval() + for p in base_model.parameters(): + p.requires_grad_(False) + all_tokens = val_data.val_tokens + all_tokens_idx = all_tokens.to(torch.int32) + docs = _find_docs(all_tokens) + doc_entries = list(enumerate(docs)) + if h.val_doc_fraction < 1.0: + sample_n = max(1, int(round(len(docs) * h.val_doc_fraction))) + sampled_indices = sorted( + random.Random(h.seed).sample(range(len(docs)), sample_n) + ) + doc_entries = [(i, docs[i]) for i in sampled_indices] + log( + f"ttt_lora:docs:{len(doc_entries)} rank:{h.ttt_lora_rank} lr:{h.ttt_lora_lr} chunk:{h.ttt_chunk_size}" + ) + if os.environ.get("TTT_DEBUG_BYPASS") and h.rank == 0: + test_doc = doc_entries[0][1] + ds, dl = test_doc + log(f"DEBUG: test doc start={ds} len={dl}") + toks = all_tokens_idx[ds : ds + dl].to(device=device, dtype=torch.int64) + x_d = toks[:-1].unsqueeze(0) + y_d = toks[1:].unsqueeze(0) + with torch.no_grad(), torch.autocast(device_type="cuda", dtype=torch.bfloat16): + logits_d = base_model.forward_logits(x_d) + ptl_d = F.cross_entropy( + logits_d.float().reshape(-1, logits_d.size(-1)), + y_d.reshape(-1), reduction="none", + ) + direct_loss = ptl_d.mean().item() + direct_bpb = direct_loss / math.log(2.0) + log(f"DEBUG: direct forward_logits loss={direct_loss:.6f} bpb={direct_bpb:.6f} ntokens={y_d.numel()}") + toks_first5 = toks[:5].tolist() + ptl_first5 = ptl_d[:5].tolist() + log(f"DEBUG: first 5 tokens={toks_first5} ptl={[f'{v:.4f}' for v in ptl_first5]}") + chunk_size, eval_seq_len = h.ttt_chunk_size, h.ttt_eval_seq_len + eval_batch_set = None + if h.ttt_eval_batches: + eval_batch_set = set(int(x) for x in h.ttt_eval_batches.split(",") if x.strip()) + use_ascending = eval_batch_set is not None + global_batches_sorted = _build_ttt_global_batches(doc_entries, h, ascending=use_ascending) + queue_len = len(global_batches_sorted) + counter_path = f"/tmp/ttt_counter_{h.run_id}" + if h.rank == 0: + _init_batch_counter(counter_path) + if dist.is_available() and dist.is_initialized(): + path_list = [counter_path] + dist.broadcast_object_list(path_list, src=0) + counter_path = path_list[0] + dist.barrier() + loss_sum = torch.zeros((), device=device, dtype=torch.float64) + byte_sum = torch.zeros((), device=device, dtype=torch.float64) + token_count = torch.zeros((), device=device, dtype=torch.float64) + t_start = time.perf_counter() + reusable_lora = BatchedTTTLoRA( + h.ttt_batch_size, base_model, h.ttt_lora_rank, + k_lora=h.ttt_k_lora, mlp_lora=h.ttt_mlp_lora, o_lora=h.ttt_o_lora, + ).to(device) + + def _build_opt(lora): + if h.ttt_optimizer == "sgd": + return torch.optim.SGD( + lora.parameters(), lr=h.ttt_lora_lr, + momentum=h.ttt_beta1, weight_decay=h.ttt_weight_decay, + ) + return torch.optim.AdamW( + lora.parameters(), lr=h.ttt_lora_lr, + betas=(h.ttt_beta1, h.ttt_beta2), + eps=1e-10, weight_decay=h.ttt_weight_decay, fused=True, + ) - # ----------------------------- - # DATA LOADER & MODEL WARMUP - # ----------------------------- + reusable_opt = _build_opt(reusable_lora) + progress_f = None + if h.ttt_output_dir and h.rank == 0: + os.makedirs(h.ttt_output_dir, exist_ok=True) + progress_f = open(os.path.join(h.ttt_output_dir, "progress.jsonl"), "w") + try: + while True: + queue_idx = _claim_next_batch(counter_path, queue_len) + if queue_idx >= queue_len: + break + orig_batch_idx, batch_entries = global_batches_sorted[queue_idx] + batch = [doc for _, doc in batch_entries] + bsz = len(batch) + prev_loss = loss_sum.item() + prev_bytes = byte_sum.item() + prev_tokens = token_count.item() + if bsz == reusable_lora.bsz: + reusable_lora.reset() + for s in reusable_opt.state.values(): + for k, v in s.items(): + if isinstance(v, torch.Tensor): + v.zero_() + elif k == "step": + s[k] = 0 + cur_lora = reusable_lora + cur_opt = reusable_opt + else: + cur_lora = BatchedTTTLoRA( + bsz, base_model, h.ttt_lora_rank, + k_lora=h.ttt_k_lora, mlp_lora=h.ttt_mlp_lora, o_lora=h.ttt_o_lora, + ).to(device) + cur_opt = _build_opt(cur_lora) + pred_lens = [doc_len - 1 for _, doc_len in batch] + num_chunks = [(pl + chunk_size - 1) // chunk_size for pl in pred_lens] + max_nc = max(num_chunks) + num_chunks_t = torch.tensor(num_chunks, dtype=torch.int64, device=device) + for ci in range(max_nc): + active = [ci < nc for nc in num_chunks] + needs_train = any(ci < nc - 1 for nc in num_chunks) + tok_starts = torch.zeros(bsz, dtype=torch.int64) + tok_wls = torch.zeros(bsz, dtype=torch.int64) + chunk_offsets_cpu = torch.zeros(bsz, dtype=torch.int64) + chunk_lens_cpu = torch.zeros(bsz, dtype=torch.int64) + for b in range(bsz): + if not active[b]: + continue + doc_start, doc_len = batch[b] + win_start, win_len, chunk_offset, chunk_len = _compute_chunk_window( + ci, pred_lens[b], num_chunks[b], chunk_size, eval_seq_len + ) + tok_starts[b] = doc_start + win_start + tok_wls[b] = win_len + chunk_offsets_cpu[b] = chunk_offset + chunk_lens_cpu[b] = chunk_len + _, context_size, chunk_offset, _ = _compute_chunk_window( + ci, (ci + 1) * chunk_size, ci + 1, chunk_size, eval_seq_len + ) + col_idx = torch.arange(context_size + 1) + idx = tok_starts.unsqueeze(1) + col_idx.unsqueeze(0) + idx.clamp_(max=all_tokens.numel() - 1) + gathered_gpu = all_tokens_idx[idx].to( + device=device, dtype=torch.int64, non_blocking=True + ) + valid = (col_idx[:context_size].unsqueeze(0) < tok_wls.unsqueeze(1)).to( + device, non_blocking=True + ) + chunk_offsets = chunk_offsets_cpu.to(device, non_blocking=True) + chunk_lens = chunk_lens_cpu.to(device, non_blocking=True) + x = torch.where(valid, gathered_gpu[:, :context_size], 0) + y = torch.where(valid, gathered_gpu[:, 1 : context_size + 1], 0) + ctx_pos = torch.arange(context_size, device=device, dtype=torch.int64) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + per_tok_loss = forward_ttt_train(x, y, lora=cur_lora) + with torch.no_grad(): + _accumulate_bpb( + per_tok_loss, + x, + y, + chunk_offsets, + chunk_lens, + ctx_pos, + val_data.base_bytes_lut, + val_data.has_leading_space_lut, + val_data.is_boundary_token_lut, + loss_sum, + byte_sum, + token_count, + ) + if needs_train: + activate_chunk_mask = (num_chunks_t - 1 > ci).float() + for gi in range(h.ttt_grad_steps): + if gi > 0: + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + per_tok_loss = forward_ttt_train(x, y, lora=cur_lora) + per_doc = per_tok_loss[ + :, chunk_offset : chunk_offset + chunk_size + ].mean(dim=-1) + cur_opt.zero_grad(set_to_none=True) + (per_doc * activate_chunk_mask).sum().backward() + cur_opt.step() + else: + del per_tok_loss + batch_num = orig_batch_idx + 1 + doc_lens = [dl for _, dl in batch] + should_report = False + if eval_batch_set is not None: + should_report = batch_num in eval_batch_set + else: + # should_report = local_batch_count % 10 == 0 + should_report = True + if should_report: + cur_tokens = token_count.item() + cur_loss_val = loss_sum.item() + cur_bytes_val = byte_sum.item() + dt = cur_tokens - prev_tokens + if dt > 0: + b_loss = (cur_loss_val - prev_loss) / dt + b_bpb = b_loss / math.log(2.0) * (dt / (cur_bytes_val - prev_bytes)) + else: + b_loss = b_bpb = 0.0 + r_loss = cur_loss_val / max(cur_tokens, 1) + r_bpb = r_loss / math.log(2.0) * (cur_tokens / max(cur_bytes_val, 1)) + elapsed = time.perf_counter() - t_start + log( + f"ttt_progress: batch {batch_num}/{queue_len} batch_loss:{b_loss:.4f} " + f"batch_bpb:{b_bpb:.4f} running_loss:{r_loss:.4f} running_bpb:{r_bpb:.4f} " + f"doc_len:{min(doc_lens)}-{max(doc_lens)}" + ) + if progress_f is not None: + progress_f.write( + json.dumps({ + "batch": batch_num, "total_batches": queue_len, + "batch_loss": round(b_loss, 8), "batch_bpb": round(b_bpb, 8), + "running_loss": round(r_loss, 8), "running_bpb": round(r_bpb, 8), + "doc_len_min": min(doc_lens), "doc_len_max": max(doc_lens), + "chunk_size": chunk_size, + "elapsed_s": round(elapsed, 3), + "batch_t_s": round(elapsed, 3), + }) + "\n" + ) + progress_f.flush() + del cur_lora, cur_opt + finally: + if progress_f is not None: + progress_f.close() + if dist.is_available() and dist.is_initialized(): + dist.all_reduce(loss_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(byte_sum, op=dist.ReduceOp.SUM) + dist.all_reduce(token_count, op=dist.ReduceOp.SUM) + for p in base_model.parameters(): + p.requires_grad_(True) + base_model.train() + val_loss = (loss_sum / token_count).item() + val_bpb = val_loss / math.log(2.0) * (token_count.item() / byte_sum.item()) + return val_loss, val_bpb + + +def timed_eval(label, fn, *args, **kwargs): + torch.cuda.synchronize() + t0 = time.perf_counter() + val_loss, val_bpb = fn(*args, **kwargs) + torch.cuda.synchronize() + elapsed_ms = 1e3 * (time.perf_counter() - t0) + log( + f"{label} val_loss:{val_loss:.8f} val_bpb:{val_bpb:.8f} eval_time:{elapsed_ms:.0f}ms" + ) + return val_loss, val_bpb - train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) - def zero_grad_all() -> None: - for opt in optimizers: - opt.zero_grad(set_to_none=True) +def train_model(h, device, val_data): + base_model = GPT(h).to(device).bfloat16() + restore_fp32_params(base_model) + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + compiled_forward_logits = torch.compile( + base_model.forward_logits, dynamic=False, fullgraph=True + ) + model = compiled_model + log(f"model_params:{sum(p.numel()for p in base_model.parameters())}") + optimizers = Optimizers(h, base_model) + train_loader = DocumentPackingLoader(h, device) + max_wallclock_ms = ( + 1e3 * h.max_wallclock_seconds if h.max_wallclock_seconds > 0 else None + ) + if max_wallclock_ms is not None: + max_wallclock_ms -= h.gptq_reserve_seconds * 1e3 + log( + f"gptq:reserving {h.gptq_reserve_seconds:.0f}s, effective={max_wallclock_ms:.0f}ms" + ) - max_wallclock_ms = 1000.0 * args.max_wallclock_seconds if args.max_wallclock_seconds > 0 else None + def training_frac(step, elapsed_ms): + if max_wallclock_ms is None: + return step / max(h.iterations, 1) + return elapsed_ms / max(max_wallclock_ms, 1e-09) - def lr_mul(step: int, elapsed_ms: float) -> float: - if args.warmdown_iters <= 0: + def lr_mul(frac): + if h.warmdown_frac <= 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 - - # Warmup primes the compiled forward/backward/optimizer paths, then we restore the - # initial weights/optimizer state so measured training starts from the true init. - 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] + if frac >= 1.0 - h.warmdown_frac: + return max((1.0 - frac) / h.warmdown_frac, h.min_lr) + return 1.0 + + def step_fn(step, lr_scale): + optimizers.zero_grad_all() + train_loss = torch.zeros((), device=device) + for micro_step in range(h.grad_accum_steps): + x, y, cu_seqlens, _max_seqlen = train_loader.next_batch( + h.train_batch_tokens, h.grad_accum_steps + ) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + loss = model(x, y, cu_seqlens=cu_seqlens, max_seqlen=h.train_seq_len) + train_loss += loss.detach() + (loss / h.grad_accum_steps).backward() + train_loss /= h.grad_accum_steps + frac = ( + min(step / h.muon_momentum_warmup_steps, 1.0) + if h.muon_momentum_warmup_steps > 0 + else 1.0 + ) + muon_momentum = ( + 1 - frac + ) * h.muon_momentum_warmup_start + frac * h.muon_momentum + for group in optimizers.optimizer_muon.param_groups: + group["momentum"] = muon_momentum + for opt in optimizers: + for group in opt.param_groups: + group["lr"] = group["base_lr"] * lr_scale + if h.grad_clip_norm > 0: + torch.nn.utils.clip_grad_norm_(base_model.parameters(), h.grad_clip_norm) + optimizers.step(distributed=h.distributed) + return train_loss + + if h.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): - if distributed: - model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 - 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() - 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}") + num_tokens_local = h.train_batch_tokens // h.world_size + for blk in base_model.blocks: + blk.attn.rotary(num_tokens_local, device, torch.bfloat16) + cu_bucket_size = train_loader.cu_bucket_size + warmup_cu_buckets = tuple(cu_bucket_size * i for i in range(1, 5)) + warmup_cu_iters = 3 + x, y, cu_seqlens, _ = train_loader.next_batch( + h.train_batch_tokens, h.grad_accum_steps + ) + log(f"warmup_cu_buckets:{','.join(str(b) for b in warmup_cu_buckets)} iters_each:{warmup_cu_iters}") + def _run_cu_bucket_warmup(): + for bucket_len in warmup_cu_buckets: + boundaries = list(range(0, x.size(1), max(h.train_seq_len, 1))) + if boundaries[-1] != x.size(1): + boundaries.append(x.size(1)) + cu = torch.full((bucket_len,), x.size(1), dtype=torch.int32, device=device) + cu[: len(boundaries)] = torch.tensor(boundaries, dtype=torch.int32, device=device) + for _ in range(warmup_cu_iters): + optimizers.zero_grad_all() + with torch.autocast(device_type="cuda", dtype=torch.bfloat16, enabled=True): + wloss = model(x, y, cu_seqlens=cu, max_seqlen=h.train_seq_len) + (wloss / h.grad_accum_steps).backward() + optimizers.zero_grad_all() + _run_cu_bucket_warmup() + if h.num_loops > 0: + base_model.looping_active = True + _run_cu_bucket_warmup() + base_model.looping_active = False + for warmup_step in range(h.warmup_steps): + step_fn(warmup_step, 1.0) + if ( + warmup_step <= 5 + or (warmup_step + 1) % 10 == 0 + or warmup_step + 1 == h.warmup_steps + ): + log(f"warmup_step: {warmup_step+1}/{h.warmup_steps}") + if h.num_loops > 0: + base_model.looping_active = True + log( + f"loop_warmup:enabled encoder:{base_model.encoder_indices} decoder:{base_model.decoder_indices}" + ) + for warmup_step in range(h.warmup_steps): + step_fn(warmup_step, 1.0) + if ( + warmup_step <= 5 + or (warmup_step + 1) % 10 == 0 + or warmup_step + 1 == h.warmup_steps + ): + log(f"loop_warmup_step: {warmup_step+1}/{h.warmup_steps}") + base_model.looping_active = False base_model.load_state_dict(initial_model_state, strict=True) - for opt, state in zip(optimizers, initial_optimizer_states, strict=True): + for (opt, state) in zip(optimizers, initial_optimizer_states, strict=True): opt.load_state_dict(state) - zero_grad_all() - if distributed: - model.require_backward_grad_sync = True - train_loader = DistributedTokenLoader(args.train_files, rank, world_size, device) - - # ----------------------------- - # MAIN TRAINING LOOP - # ----------------------------- - + optimizers.zero_grad_all() + train_loader = DocumentPackingLoader(h, device) + ema_state = { + name: t.detach().float().clone() + for (name, t) in base_model.state_dict().items() + } + ema_decay = h.ema_decay training_time_ms = 0.0 - stop_after_step: int | None = None + stop_after_step = 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) + last_step = ( + step == h.iterations + or stop_after_step is not None + and step >= stop_after_step + ) + should_validate = ( + last_step or h.val_loss_every > 0 and step % h.val_loss_every == 0 + ) if should_validate: torch.cuda.synchronize() - training_time_ms += 1000.0 * (time.perf_counter() - t0) + training_time_ms += 1e3 * (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, + h, device, val_data, model, compiled_forward_logits ) - 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" + log( + f"{step}/{h.iterations} val_loss: {val_loss:.4f} val_bpb: {val_bpb:.4f}" ) 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}" + if stop_after_step is not None and step < h.iterations: + log( + f"stopping_early: wallclock_cap train_time: {training_time_ms:.0f}ms step: {step}/{h.iterations}" ) break - - elapsed_ms = training_time_ms + 1000.0 * (time.perf_counter() - t0) - scale = lr_mul(step, elapsed_ms) - zero_grad_all() - train_loss = torch.zeros((), device=device) - for micro_step in range(grad_accum_steps): - if distributed: - model.require_backward_grad_sync = micro_step == grad_accum_steps - 1 - 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() - 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) - for opt in optimizers: - opt.step() - zero_grad_all() - + elapsed_ms = training_time_ms + 1e3 * (time.perf_counter() - t0) + frac = training_frac(step, elapsed_ms) + scale = lr_mul(frac) + if ( + h.num_loops > 0 + and not base_model.looping_active + and frac >= h.enable_looping_at + ): + base_model.looping_active = True + log( + f"layer_loop:enabled step:{step} frac:{frac:.3f} encoder:{base_model.encoder_indices} decoder:{base_model.decoder_indices}" + ) + train_loss = step_fn(step, scale) + 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) - 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) + approx_training_time_ms = training_time_ms + 1e3 * (time.perf_counter() - t0) + should_log_train = h.train_log_every > 0 and ( + step <= 5 or step % h.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_per_sec = step * h.train_batch_tokens / (approx_training_time_ms / 1e3) + log( + f"{step}/{h.iterations} train_loss: {train_loss.item():.4f} train_time: {approx_training_time_ms/60000:.1f}m tok/s: {tok_per_sec:.0f}" ) - - # Needed to sync whether we've reached the wallclock cap. - reached_cap = max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms - if distributed and max_wallclock_ms is not None: + reached_cap = ( + max_wallclock_ms is not None and approx_training_time_ms >= max_wallclock_ms + ) + if h.distributed and 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" + log( + f"peak memory allocated: {torch.cuda.max_memory_allocated()//1024//1024} MiB reserved: {torch.cuda.max_memory_reserved()//1024//1024} MiB" ) - - # ----------------------------- - # SERIALIZATION + ROUNDTRIP VALIDATION - # ----------------------------- - # Save the raw state (useful for debugging/loading in PyTorch directly), then always produce - # the compressed int8+zlib artifact and validate the round-tripped weights. - - if master_process: - torch.save(base_model.state_dict(), "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") - log0(f"Total submission size: {model_bytes + code_bytes} bytes") - - quant_obj, quant_stats = quantize_state_dict_int8(base_model.state_dict()) - quant_buf = io.BytesIO() - torch.save(quant_obj, quant_buf) - quant_raw = quant_buf.getvalue() - quant_blob = zlib.compress(quant_raw, level=9) - quant_raw_bytes = len(quant_raw) - if master_process: - with open("final_model.int8.ptz", "wb") as f: - f.write(quant_blob) - quant_file_bytes = os.path.getsize("final_model.int8.ptz") - code_bytes = len(code.encode("utf-8")) - ratio = quant_stats["baseline_tensor_bytes"] / max(quant_stats["int8_payload_bytes"], 1) - log0( - f"Serialized model int8+zlib: {quant_file_bytes} bytes " - f"(payload:{quant_stats['int8_payload_bytes']} raw_torch:{quant_raw_bytes} payload_ratio:{ratio:.2f}x)" + log("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) + log_qk_gain_converged(log, base_model) + return base_model, compiled_model, compiled_forward_logits + + +def train_and_eval(h, device): + random.seed(h.seed) + np.random.seed(h.seed) + torch.manual_seed(h.seed) + torch.cuda.manual_seed_all(h.seed) + if h.artifact_dir and h.is_main_process: + os.makedirs(h.artifact_dir, exist_ok=True) + val_data = ValidationData(h, device) + if h.eval_only_path: + log(f"eval_only:loading checkpoint from {h.eval_only_path}") + base_model = GPT(h).to(device).bfloat16() + restore_fp32_params(base_model) + base_model.load_state_dict(torch.load(h.eval_only_path, map_location=device)) + if h.num_loops > 0: + base_model.looping_active = True + compiled_model = torch.compile(base_model, dynamic=False, fullgraph=True) + compiled_forward_logits = torch.compile( + base_model.forward_logits, dynamic=False, fullgraph=True ) - log0(f"Total submission size int8+zlib: {quant_file_bytes + code_bytes} bytes") - - if distributed: + else: + log( + f"train_shards: {len(list(Path(h.datasets_dir).resolve().glob('fineweb_train_*.bin')))}" + ) + log(f"val_tokens: {val_data.val_tokens.numel()-1}") + base_model, compiled_model, compiled_forward_logits = train_model( + h, device, val_data + ) + _skip_training = bool(h.eval_only_path) + torch._dynamo.reset() + timed_eval( + "diagnostic pre-quantization post-ema", + eval_val, + h, + device, + val_data, + compiled_model, + compiled_forward_logits, + ) + if not _skip_training: + serialize(h, base_model, Path(__file__).read_text(encoding="utf-8")) + else: + log("eval_only: skipping serialize (already have quantized model)") + if not os.path.exists(h.quantized_model_path): + log("eval_only: no quantized model found, running serialize anyway") + serialize(h, base_model, Path(__file__).read_text(encoding="utf-8")) + if h.distributed: dist.barrier() - with open("final_model.int8.ptz", "rb") as f: - quant_blob_disk = f.read() - quant_state = torch.load(io.BytesIO(zlib.decompress(quant_blob_disk)), map_location="cpu") - base_model.load_state_dict(dequantize_state_dict_int8(quant_state), strict=True) - torch.cuda.synchronize() - t_qeval = time.perf_counter() - q_val_loss, q_val_bpb = eval_val( - args, - model, - rank, - world_size, + eval_model = deserialize(h, device) + if h.num_loops > 0: + eval_model.looping_active = True + compiled_model = torch.compile(eval_model, dynamic=False, fullgraph=True) + compiled_forward_logits = torch.compile( + eval_model.forward_logits, dynamic=False, fullgraph=True + ) + timed_eval( + "diagnostic quantized", + eval_val, + h, device, - grad_accum_steps, - val_tokens, - base_bytes_lut, - has_leading_space_lut, - is_boundary_token_lut, + val_data, + compiled_model, + compiled_forward_logits, ) - torch.cuda.synchronize() - log0( - f"final_int8_zlib_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" + if h.sliding_window_enabled: + timed_eval( + "diagnostic quantized_sliding_window", + eval_val_sliding, + h, + device, + val_data, + eval_model, + forward_logits_fn=compiled_forward_logits, + ) + if h.ttt_enabled: + del eval_model, compiled_model + torch._dynamo.reset() + torch.cuda.empty_cache() + ttt_model = deserialize(h, device) + if h.num_loops > 0: + ttt_model.looping_active = True + for p in ttt_model.parameters(): + p.requires_grad_(False) + + if h.rope_yarn: + _yarn_seqlen = h.train_batch_tokens // h.grad_accum_steps + for block in ttt_model.blocks: + block.attn.rotary(_yarn_seqlen, device, torch.bfloat16) + else: + for block in ttt_model.blocks: + block.attn.rotary._cos_cached = None + block.attn.rotary._sin_cached = None + block.attn.rotary._seq_len_cached = 0 + block.attn.rotary(h.ttt_eval_seq_len, device, torch.bfloat16) + + def _fwd_ttt_inner(input_ids, target_ids, lora): + return ttt_model.forward_ttt(input_ids, target_ids, lora=lora) + + _fwd_ttt_compiled_inner = None + + def _fwd_ttt(input_ids, target_ids, lora): + nonlocal _fwd_ttt_compiled_inner + if _fwd_ttt_compiled_inner is None: + _fwd_ttt_compiled_inner = torch.compile(_fwd_ttt_inner, dynamic=True) + return _fwd_ttt_compiled_inner(input_ids, target_ids, lora=lora) + + _ttt_debug_bypass = bool(os.environ.get("TTT_DEBUG_BYPASS")) + if _ttt_debug_bypass: + def _fwd_ttt_bypass(input_ids, target_ids, lora): + logits = ttt_model.forward_logits(input_ids) + dummy = lora.q_loras[0].B.sum() * 0 + logits = logits + dummy + bsz, sl, V = logits.shape + return F.cross_entropy( + logits.float().reshape(-1, V), target_ids.reshape(-1), reduction="none" + ).reshape(bsz, sl) + fwd_ttt_compiled = _fwd_ttt_bypass + log("ttt_lora:DEBUG BYPASS active - using forward_logits directly (no compile warmup)") + else: + fwd_ttt_compiled = _fwd_ttt + log(f"ttt_lora:warming up compile (random tokens, no val data)") + global BOS_ID + if BOS_ID is None: + BOS_ID = 1 + t_warmup = time.perf_counter() + warmup_bszes = [h.ttt_batch_size] + for bsz in warmup_bszes: + wl = BatchedTTTLoRA( + bsz, ttt_model, h.ttt_lora_rank, + k_lora=h.ttt_k_lora, mlp_lora=h.ttt_mlp_lora, o_lora=h.ttt_o_lora, + ).to(device) + wo = torch.optim.AdamW( + wl.parameters(), + lr=h.ttt_lora_lr, + betas=(h.ttt_beta1, h.ttt_beta2), + eps=1e-10, + weight_decay=h.ttt_weight_decay, + fused=True, + ) + for ctx_len in (h.ttt_chunk_size, h.ttt_eval_seq_len): + xw = torch.randint(0, h.vocab_size, (bsz, ctx_len), device=device, dtype=torch.int64) + yw = torch.randint(0, h.vocab_size, (bsz, ctx_len), device=device, dtype=torch.int64) + with torch.autocast(device_type="cuda", dtype=torch.bfloat16): + ptl = fwd_ttt_compiled(xw, yw, lora=wl) + ptl[:, : min(h.ttt_chunk_size, ctx_len)].mean(dim=-1).sum().backward() + wo.step() + wo.zero_grad(set_to_none=True) + del wl, wo + torch.cuda.empty_cache() + compile_elapsed = time.perf_counter() - t_warmup + log(f"ttt_lora:compile warmup done ({compile_elapsed:.1f}s)") + log("\nbeginning TTT eval timer") + torch.cuda.synchronize() + t_ttt = time.perf_counter() + ttt_val_loss, ttt_val_bpb = eval_val_ttt_lora( + h, ttt_model, device, val_data, forward_ttt_train=fwd_ttt_compiled + ) + torch.cuda.synchronize() + ttt_eval_elapsed = time.perf_counter() - t_ttt + log( + f"quantized_ttt_lora val_loss:{ttt_val_loss:.8f} val_bpb:{ttt_val_bpb:.8f} eval_time:{1e3*ttt_eval_elapsed:.0f}ms" + ) + log(f"total_eval_time:{ttt_eval_elapsed:.1f}s") + del ttt_model + + +def main(): + world_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + distributed = "RANK" in os.environ and "WORLD_SIZE" in os.environ + if not torch.cuda.is_available(): + raise RuntimeError("CUDA is required") + if world_size <= 0: + raise ValueError(f"WORLD_SIZE must be positive, got {world_size}") + if 8 % world_size != 0: + raise ValueError( + f"WORLD_SIZE={world_size} must divide 8 so grad_accum_steps stays integral" + ) + device = torch.device("cuda", local_rank) + torch.cuda.set_device(device) + if distributed: + dist.init_process_group(backend="nccl", device_id=device) + dist.barrier() + torch.backends.cuda.matmul.allow_tf32 = True + torch.backends.cudnn.allow_tf32 = True + torch.set_float32_matmul_precision("high") + from torch.backends.cuda import ( + enable_cudnn_sdp, + enable_flash_sdp, + enable_math_sdp, + enable_mem_efficient_sdp, ) - log0(f"final_int8_zlib_roundtrip_exact val_loss:{q_val_loss:.8f} val_bpb:{q_val_bpb:.8f}") + enable_cudnn_sdp(False) + enable_flash_sdp(True) + enable_mem_efficient_sdp(False) + enable_math_sdp(False) + torch._dynamo.config.optimize_ddp = False + torch._dynamo.config.cache_size_limit = 16 + h = Hyperparameters() + set_logging_hparams(h) + if h.is_main_process: + os.makedirs(h.artifact_dir if h.artifact_dir else "logs", exist_ok=True) + log(100 * "=", console=False) + log("Hyperparameters:", console=True) + for (k, v) in sorted(vars(type(h)).items()): + if not k.startswith("_"): + log(f" {k}: {v}", console=True) + log("=" * 100, console=False) + log("Source code:", console=False) + log("=" * 100, console=False) + with open(__file__, "r", encoding="utf-8") as _src: + log(_src.read(), console=False) + log("=" * 100, console=False) + log(f"Running Python {sys.version}", console=False) + log(f"Running PyTorch {torch.__version__}", console=False) + log( + subprocess.run( + ["nvidia-smi"], + stdout=subprocess.PIPE, + stderr=subprocess.PIPE, + text=True, + check=False, + ).stdout, + console=False, + ) + log("=" * 100, console=False) + train_and_eval(h, device) if distributed: dist.destroy_process_group()