Torchao FP8 not working with LoRA

[kailashg26]

Hello,

I followed full weight finetuning and implemented it to LoRA. But I see some errors with and without torch.compile.
This is done with llama-3.1 8B BS=16 and Seq len = 8192 with compile and packed = True

Note: The same code works for 70B - lora but I don't see any performance improvement.
lora_finetune_distributed.py

import sys
import time

from functools import partial
from typing import Any, Optional, Union
from warnings import warn

import torch
from omegaconf import DictConfig, ListConfig

from torch import nn
from torch.distributed import destroy_process_group, init_process_group
from torch.distributed.tensor import DTensor

from torch.optim import Optimizer
from torchao.float8 import precompute_float8_dynamic_scale_for_fsdp
from torchdata.stateful_dataloader import StatefulDataLoader
from torchdata.stateful_dataloader.sampler import StatefulDistributedSampler
from torchtune import config, modules, training, utils
from torchtune.config._utils import _get_component_from_path
from torchtune.data import padded_collate_packed
from torchtune.datasets import ConcatDataset
from torchtune.modules.loss import SFTLoss
from torchtune.modules.peft import (
    AdapterModule,
    get_adapter_params,
    get_lora_module_names,
    set_trainable_params,
    validate_missing_and_unexpected_for_lora,
)
from torchtune.recipe_interfaces import FTRecipeInterface
from torchtune.training import (
    DummyProfiler,
    PROFILER_KEY,
    VALID_BACKENDS_FOR_MEMORY_STATS,
)
from torchtune.training.checkpointing._checkpoint_client import (
    CheckpointClient,
    TrainingProgress,
)
from torchtune.training.quantization import (
    convert_to_float8_training,
    is_fp8_tensorwise_scaling,
)

from tqdm import tqdm


class LoRAFinetuneRecipeDistributed(FTRecipeInterface):
    """
    Distributed LoRA finetuning recipe for dense transformer-based LLMs such as Llama2. This recipe supports
    distributed training and can be run on a single node (1 to 8 GPUs).

    Features:
        - FSDP. Supported using PyTorch's FSDP APIs. CPU offload of parameters, gradients, and optimizer states
            is supported via ``fsdp_cpu_offload``. Resharding of parameters after the forward pass is
            done by default (corresponding to FULL_SHARD sharding strategy), but can be disabled by setting the config
            ``fsdp_reshard_after_forward`` to False (this corresponds to SHARD_GRAD_OP sharding strategy).
            DDP is currently not supported. Training on CPU is not supported.

        - Activation Checkpointing. This can be controlled using the ``enable_activation_checkpointing``
            flag. Activation checkpointing helps reduce the memory footprint since we no longer keep
            activations in memory and instead recompute them during the backward pass. This is especially
            helpful for larger batch sizes when you're memory constrained. But these savings in memory
            come at the cost of training performance. In most cases training can slow-down quite a bit as
            a result of this activation recomputation.

        - Activation Offloading. This can be controlled using the ``enable_activation_offloading``
            flag. Activation offloading is a technique similar to activations checkpointing that helps
            reduce the memory footprint to prevent OOMs on CUDA and enable bigger batches. Where activations
            checkpointing drops the activation in the forward to recompute it later in the backward,
            activations offloading will drop the activation in the forward to the CPU and bring it
            back during the backward pass. As always, there is a tradeoff--these savings in memory can
            come at the cost of training performance and CPU resources. To recover some runtime cost,
            we've added an option to enable offloading on a different stream to permit overlapping with
            the computation. This option is currently only available on PyTorch 2.5.0 or later and will be
            enabled by default if an acceptable torch version is found. Activation offloading can be used in
            conjunction with activation checkpointing.

        - Precision. Full fp32 and bf16 training are supported. Precision is controlled using the ``dtype``
            flag. When ``dtype=bf16``, all activations, gradients and optimizer states are in bfloat16. In
            most cases this should halve the memory footprint of full precision (fp32) training, without
            loss in model quality (will depend on the model, training data and other settings). For
            GPUs which do not support bfloat16, we fall back to fp32. Mixed precision training and fp16
            precision are currently not supported.

        - Gradient Accumulation. You can simulate larger batch sizes by accumulating gradients. This is
            controlled using the ``gradient_accumulation_steps`` flag.

                Total Batch Size = batch_size * number of GPUs * gradient accumulation steps.

            For example: with batch_size=1, nproc_per_node=2 and gradient_accumulation_steps=32 we get a
            total batch size of 64.

            Gradient accumulation is especially useful when you are memory constrained. In this case,
            accumulating gradients might give you better training speed than enabling activation
            checkpointing.

        - Checkpointing. Model weights are checkpointed both at the end of each epoch and at the end of
            training. Currently we checkpoint both the adapter weights (trainable params only) and the
            complete merged weights (adapter weights added back to the base model). For more details
            please take a look at our LoRA tutorial
            (https://pytorch.org/torchtune/main/tutorials/lora_finetune.html).

            Optimizer State and recipe state (seed, total_epochs, number of epochs run etc) are
            only saved at the end of a given epoch and used in case of resuming training. Resuming
            training is controlled by the ``resume_from_checkpoint`` flag. Mid-epoch checkpointing is
            currently not supported.

            For more details on the checkpointer, please take a look at
            our checkpointer deepdive (https://pytorch.org/torchtune/main/tutorials/checkpointer.html).

        - Logging. Terminal, Disk, WandB and TensorBoard are all supported.

        - Gradient Clipping. Gradient clipping is supported using the ``clip_grad_norm`` flag. By default,
            ``clip_grad_norm`` is set to ``None``. If you only want to log the grad norm, you can set
            ``clip_grad_norm='inf'``.

    For a full list of example configs for this recipe, run ``tune ls`` on the command line. Each config
    has example commands for how to kick-off training.

    Args:
        cfg (DictConfig): OmegaConf object parsed from yaml file

    Raises:
        ValueError: If ``dtype`` is set to fp16.
        ValueError: If world_size is 1
        RuntimeError: If ``dtype`` is set to bf16 and the hardware does not support bf16.
        RuntimeError: If ``left_pad_sequence`` is set as the data collator.
        RuntimeError: If ``enable_activation_offloading`` is True and device is not CUDA.
        RuntimeError: If ``enable_activation_offloading`` is True and ``enable_activation_checkpointing`` is False.
    """

    def __init__(self, cfg: DictConfig) -> None:
        self._device = utils.get_device(device=cfg.device)
        self._dtype = training.get_dtype(cfg.dtype, device=self._device)

        if self._dtype == torch.float16:
            raise ValueError(
                "full fp16 training is not supported with this recipe. Please use bf16 or fp32 instead."
            )

        # Set up the backend for distributed training (NCCL, GLOO, etc.)
        self._enable_async_checkpointing = cfg.get("enable_async_checkpointing", False)
        self.fsdp_cpu_offload = cfg.get("fsdp_cpu_offload", False)
        self.distributed_backend = training.get_distributed_backend(
            cfg.device,
            offload_ops_to_cpu=self.fsdp_cpu_offload
            or self._enable_async_checkpointing,
        )
        init_process_group(self.distributed_backend)

        self.world_size, self.rank = utils.get_world_size_and_rank()

        self._is_rank_zero = self.rank == 0

        # logging attributes
        self._output_dir = cfg.output_dir
        self._log_every_n_steps = cfg.get("log_every_n_steps", 1)
        self._log_peak_memory_stats = cfg.get("log_peak_memory_stats", False)
        self._logger = utils.get_logger(cfg.log_level)

        if (
            self._log_peak_memory_stats
            and self._device.type not in VALID_BACKENDS_FOR_MEMORY_STATS
        ):
            self._logger.info(
                f"log_peak_memory_stats was set to True; however, training device is not in {VALID_BACKENDS_FOR_MEMORY_STATS}."
                "Setting log_peak_memory_stats=False."
            )
            self._log_peak_memory_stats = False

        self._enable_async_checkpointing = cfg.get("enable_async_checkpointing", False)
        self._checkpoint_client = CheckpointClient(cfg)
        self._enable_fp8_training = cfg.get("enable_fp8_training", False)
        self._fp8_recipe_name = cfg.get("fp8_recipe_name", None)

        # These attributes constitute the recipe state and are updated by ``load_checkpoint``
        # when ``resume_from_checkpoint`` is ``True``
        self.seed = training.set_seed(
            seed=cfg.seed, debug_mode=cfg.get("cudnn_deterministic_mode", None)
        )
        self.epochs_run = 0
        self.total_epochs = cfg.epochs
        self.max_steps_per_epoch = cfg.max_steps_per_epoch
        self.global_step = 0
        self._clip_grad_norm = cfg.get("clip_grad_norm", None)

        self._save_adapter_weights_only = cfg.get("save_adapter_weights_only", False)
        self._resume_from_checkpoint = cfg.resume_from_checkpoint
        self._gradient_accumulation_steps = cfg.gradient_accumulation_steps

        self._run_val_every_n_steps = cfg.get("run_val_every_n_steps", None)
        if self._run_val_every_n_steps is not None:
            assert (
                cfg.get("dataset_val") is not None
            ), "run_val_every_n_steps is set but dataset_val is not provided"

        # activation checkpointing/offloading
        self._enable_activation_checkpointing = cfg.get(
            "enable_activation_checkpointing", False
        )
        self._enable_activation_offloading = cfg.get(
            "enable_activation_offloading", False
        )
        if self._enable_activation_offloading:
            if self._device.type != "cuda":
                raise RuntimeError(
                    "enable_activation_offloading should only be True when training on CUDA"
                )
            if not self._enable_activation_checkpointing:
                raise RuntimeError(
                    "enable_activation_offloading should only be True when enable_activation_checkpointing is True"
                )
        elif (
            self._enable_activation_checkpointing
            and cfg.checkpointer.model_type != "LLAMA3_VISION"
        ):
            utils.log_rank_zero(
                self._logger,
                "Hint: enable_activation_checkpointing is True, but enable_activation_offloading isn't. "
                "Enabling activation offloading should reduce memory further.",
            )

    def _update_recipe_state(self, ckpt_dict: dict[str, Any]) -> None:
        """
        Updates the recipe state from checkpoint.
        """
        try:
            self.epochs_run = ckpt_dict[training.EPOCHS_KEY]

            # on mismatch, warn the user and prevent the override
            if self.seed != ckpt_dict[training.SEED_KEY]:
                warn(
                    message=(
                        "Config value for seed does not match the checkpoint value, "
                        f"using the checkpoint value: {ckpt_dict[training.SEED_KEY]}"
                    )
                )
                self.seed = ckpt_dict[training.SEED_KEY]
            if self.max_steps_per_epoch != ckpt_dict[training.MAX_STEPS_KEY]:
                warn(
                    message=(
                        "Config value for max_steps_per_epoch does not match the checkpoint value, "
                        f"using the checkpoint value: {ckpt_dict[training.MAX_STEPS_KEY]}"
                    )
                )
                self.max_steps_per_epoch = ckpt_dict[training.MAX_STEPS_KEY]

            # on mismatch, warn the user but allow the override
            if self.total_epochs != ckpt_dict[training.TOTAL_EPOCHS_KEY]:
                warn(
                    message=(
                        "Config value for total_epochs does not match the checkpoint value, "
                        f"using the config value: {self.total_epochs}"
                    )
                )

        except KeyError as e:
            raise KeyError(
                "Checkpoint does not contain the required keys needed for updating recipe state. "
                "Are you sure you passed in the right recipe checkpoint?"
            ) from e

    def setup(self, cfg: DictConfig) -> None:
        """
        Setup the recipe state. This includes recipe state (if resume_from_checkpoint is True),
        model, tokenizer, loss, optimizer, learning rate scheduler, sampler, and dataloader.
        """
        if self._is_rank_zero:
            self._metric_logger = config.instantiate(cfg.metric_logger)

            # log config with parameter override
            self._metric_logger.log_config(cfg)

        if (
            cfg.checkpointer.model_type == "LLAMA4"
            and self._save_adapter_weights_only is False
        ):
            raise ValueError(
                "For Llama4 training, you should set save_adapter_weights_only to True."
            )

        checkpoint_dict = self._checkpoint_client.load_base_checkpoint()

        self._compile = cfg.get("compile", False)

        self._model = self._setup_model(
            cfg_model=cfg.model,
            enable_activation_checkpointing=self._enable_activation_checkpointing,
            enable_activation_offloading=self._enable_activation_offloading,
            custom_sharded_layers=cfg.get("custom_sharded_layers", None),
            fsdp_cpu_offload=cfg.get("fsdp_cpu_offload", False),
            reshard_after_forward=cfg.get("fsdp_reshard_after_forward", True),
            base_model_state_dict=checkpoint_dict[training.MODEL_KEY],
            lora_weights_state_dict=(
                checkpoint_dict[training.ADAPTER_KEY]
                if training.ADAPTER_KEY in checkpoint_dict
                else None
            ),
        )
        self._tokenizer = config.instantiate(cfg.tokenizer)

        self._optimizer = self._setup_optimizer(
            cfg_optimizer=cfg.optimizer,
            opt_state_dict=(
                checkpoint_dict[training.OPT_KEY]
                if training.OPT_KEY in checkpoint_dict
                else None
            ),
        )

        if self._resume_from_checkpoint:
            # If async checkpointing is enabled, intermediate checkpoints are saved asynchronously
            # using the DistributedCheckpointer.
            # Therefore the recipe needs to load the distributed checkpoint to restore the training
            # progress.
            if self._enable_async_checkpointing:
                try:
                    checkpoint_dict = (
                        self._checkpoint_client.load_distributed_checkpoint(
                            self._model,
                            self._optimizer,
                            self._adapter_config,
                        )
                    )
                except Exception as e:
                    self._logger.warning(
                        f"Failed to load distributed checkpoint: {e}. Training will start from the base checkpoint."
                    )

            if training.ADAPTER_KEY not in checkpoint_dict:
                raise ValueError(
                    "Adapter weights not found. Please ensure a valid adapter checkpoint is provided."
                )

            # Update the recipe state from the checkpoint state dict.
            self._update_recipe_state(checkpoint_dict)

        # initialize loss
        self._loss_fn = config.instantiate(cfg.loss)
        if isinstance(self._loss_fn, SFTLoss):
            self._loss_fn.set_model_output(self._model)

        if self._compile:
            training.compile_loss(self._loss_fn, verbose=self._is_rank_zero)

        utils.log_rank_zero(self._logger, "Loss is initialized.")

        # sampler and dataloader depend on the tokenizer and loss_fn and should be
        # setup after all of these are setup
        collate_name = cfg.get("collate_fn", "torchtune.data.padded_collate_sft")
        self._dataloader = self._setup_data(
            cfg_dataset=cfg.dataset,
            shuffle=cfg.shuffle,
            batch_size=cfg.batch_size,
            collate_fn=collate_name,
        )

        # Setup validation dataloader if validation dataset is provided
        self._val_dataloader = None
        if cfg.get("dataset_val") is not None:
            batch_size_val = cfg.get("batch_size_val", cfg.batch_size)
            self._val_dataloader = self._setup_data(
                cfg_dataset=cfg.dataset_val,
                batch_size=batch_size_val,
                collate_fn=collate_name,
                shuffle=False,
            )

        # Finally update the recipe state which can only be correctly set after all of the
        # other components have been initialized and updated.

        # Number of training steps in each epoch depends on the number of batches produced
        # by the dataloader and the max_steps_per_epoch param set by the user and is used
        # for logging and tracking training state. This should be computed after the dataloader
        # has been setup
        self._steps_per_epoch = (
            len(self._dataloader) // self._gradient_accumulation_steps
        )
        if (
            self.max_steps_per_epoch is not None
            and self.max_steps_per_epoch < self._steps_per_epoch
        ):
            self._steps_per_epoch = self.max_steps_per_epoch
        self.global_step = self.epochs_run * self._steps_per_epoch

        # Learning rate scheduler can only be set up after number of steps
        # has been computed
        self._lr_scheduler = self._setup_lr_scheduler(
            cfg_lr_scheduler=cfg.lr_scheduler,
            num_training_steps=self.total_epochs * self._steps_per_epoch,
            last_epoch=self.global_step - 1,
        )

        # Set up profiler, returns DummyProfiler (nullcontext object with no-op `step` method)
        # if cfg is missing profiler key or if `cfg.profiler.enabled = False`
        self._profiler = self._setup_profiler(cfg.get(PROFILER_KEY, None))

    def _setup_profiler(
        self, cfg_profiler: Optional[DictConfig] = None
    ) -> Union[torch.profiler.profile, DummyProfiler]:
        """
        Parses the `profiler` section of top-level `cfg` and sets up profiler
        """
        # Missing profiler section in config, assume disabled
        if cfg_profiler is None:
            cfg_profiler = DictConfig({"enabled": False})

        # Check that component is included and set correctly
        if cfg_profiler.get("_component_", None) is None:
            cfg_profiler["_component_"] = "torchtune.training.setup_torch_profiler"
        else:
            assert (
                cfg_profiler.get("_component_")
                == "torchtune.training.setup_torch_profiler"
            ), "Only torch profiler supported currently: component must be `torchtune.training.setup_torch_profiler`"

        profiler, profiler_cfg = config.instantiate(cfg_profiler)

        utils.log_rank_zero(
            self._logger, f" Profiler config after instantiation: {profiler_cfg}"
        )
        if self._is_rank_zero:
            self.profiler_profile_memory = profiler_cfg.get("profile_memory", False)
            if profiler_cfg["enabled"]:
                self.profiler_wait_steps = profiler_cfg["wait_steps"]
                self.profiler_warmup_steps = profiler_cfg["warmup_steps"]
                self.profiler_active_steps = profiler_cfg["active_steps"]

        return profiler

    def _setup_model(
        self,
        cfg_model: DictConfig,
        enable_activation_checkpointing: bool,
        enable_activation_offloading: bool,
        fsdp_cpu_offload: bool,
        reshard_after_forward: bool,
        base_model_state_dict: dict[str, Any],
        custom_sharded_layers: Optional[list[str]] = None,
        lora_weights_state_dict: Optional[dict[str, Any]] = None,
    ) -> nn.Module:
        """
        Model initialization has some important considerations:
           a. To minimize GPU peak memory, we initialize the model on meta device with
              the right dtype
           b. All ranks calls ``load_state_dict`` without peaking CPU RAMs since
              full state dicts are loaded with ``torch.load(mmap=True)``
           c. We register (pre-)forward hooks with ``fully_shard`` instead of wrapping `nn.Module`
        """

        self._lora_rank = cfg_model.lora_rank
        self._lora_alpha = cfg_model.lora_alpha
        self._lora_attn_modules = list(cfg_model.lora_attn_modules)
        self._apply_lora_to_mlp = cfg_model.apply_lora_to_mlp
        self._apply_lora_to_output = getattr(cfg_model, "apply_lora_to_output", False)
        self._adapter_config = {
            "r": self._lora_rank,
            "lora_alpha": self._lora_alpha,
            "target_modules": get_lora_module_names(
                self._lora_attn_modules,
                self._apply_lora_to_mlp,
                self._apply_lora_to_output,
            ),
            "peft_type": "LORA",
        }

        utils.log_rank_zero(
            self._logger,
            "FSDP is enabled. Instantiating model and loading checkpoint on Rank 0 ...",
        )
        init_start = time.perf_counter()

        with training.set_default_dtype(self._dtype), torch.device("meta"):
            model = config.instantiate(cfg_model)

        set_trainable_params(model, get_adapter_params(model))

        if self._compile:
            training.compile_model(model, verbose=self._is_rank_zero)

        if self._enable_fp8_training:
            # Requires https://github.com/pytorch/pytorch/pull/148922
            if torch.__version__ < "2.8.0.dev20250318":
                raise RuntimeError(
                    "Float8 fine-tuning requires PyTorch 2.8.0.dev20250318 or later."
                )
            '''    
            if self.tp_plan is not None:
                raise ValueError(
                    "FP8 training does not support tensor parallelism yet. "
                    "This will be enabled in the near future."
                )
            '''
            model = convert_to_float8_training(model, self._fp8_recipe_name)

        if enable_activation_checkpointing:
            training.set_activation_checkpointing(
                model, auto_wrap_policy={modules.TransformerSelfAttentionLayer}
            )

        # For FSDP sharding
        fsdp_shard_conditions = [
            partial(
                training.get_shard_conditions,
                names_to_match=custom_sharded_layers,
            )
        ]
        training.shard_model(
            model=model,
            shard_conditions=fsdp_shard_conditions,
            cpu_offload=fsdp_cpu_offload,
            reshard_after_forward=reshard_after_forward,
        )

        if lora_weights_state_dict:
            lora_missing, lora_unexpected = training.load_from_full_model_state_dict(
                model,
                lora_weights_state_dict,
                self._device,
                cpu_offload=fsdp_cpu_offload,
            )
        else:
            lora_missing, lora_unexpected = None, None

        # Initialize LoRA params and RoPE buffers
        with training.set_default_dtype(self._dtype), self._device:
            lora_device = "cpu" if fsdp_cpu_offload else self._device
            for m in model.modules():
                if (isinstance(m, AdapterModule)) and not lora_weights_state_dict:
                    # lora may not be covered in state dict
                    # if finetune for the 1st time
                    m.to_empty(device=lora_device)
                    m.initialize_parameters()

                if hasattr(m, "rope_init"):
                    m.rope_init()

        base_missing, base_unexpected = training.load_from_full_model_state_dict(
            model,
            base_model_state_dict,
            self._device,
            cpu_offload=fsdp_cpu_offload,
        )
        for m in model.modules():
            if hasattr(m, "initialize_dora_magnitude"):
                m.initialize_dora_magnitude()

        validate_missing_and_unexpected_for_lora(
            lora_attn_modules=self._lora_attn_modules,
            apply_lora_to_mlp=self._apply_lora_to_mlp,
            apply_lora_to_output=self._apply_lora_to_output,
            state_dict_keys=model.state_dict().keys(),
            base_missing=base_missing,
            base_unexpected=base_unexpected,
            lora_missing=lora_missing,
            lora_unexpected=lora_unexpected,
        )
        # Ensure no params and buffers are on meta device
        training.validate_no_params_on_meta_device(model)

        # activation offloading
        self.activations_handling_ctx = training.get_act_offloading_ctx_manager(
            model, enable_activation_offloading
        )

        # log
        utils.log_rank_zero(
            self._logger,
            f"Instantiating model and loading checkpoint took {time.perf_counter() - init_start:.2f} secs",
        )
        if self._is_rank_zero:
            memory_stats = training.get_memory_stats(device=self._device)
            training.log_memory_stats(memory_stats)

        # synchronize before training begins
        torch.distributed.barrier()

        return model

    def _setup_optimizer(
        self, cfg_optimizer: DictConfig, opt_state_dict: Optional[dict[str, Any]] = None
    ) -> Optimizer:
        optimizer = config.instantiate(cfg_optimizer, self._model.parameters())
        if opt_state_dict:
            training.load_from_full_optimizer_state_dict(
                self._model,
                optimizer,
                opt_state_dict,
                self._device,
            )

        utils.log_rank_zero(self._logger, "Optimizer is initialized.")
        return optimizer

    def _setup_lr_scheduler(
        self,
        cfg_lr_scheduler: DictConfig,
        num_training_steps: int,
        last_epoch: int,
    ) -> Optimizer:
        lr_scheduler = config.instantiate(
            cfg_lr_scheduler,
            self._optimizer,
            num_training_steps=num_training_steps,
            last_epoch=last_epoch,
        )
        utils.log_rank_zero(self._logger, "Learning rate scheduler is initialized.")
        return lr_scheduler

    def _setup_data(
        self,
        cfg_dataset: DictConfig,
        shuffle: bool,
        batch_size: int,
        collate_fn: str,
    ) -> StatefulDataLoader:
        """
        All data related setup happens here. This recipe currently supports only
        map-style datasets. If a state_dict is provided (meaning we are resuming a training run),
        it is loaded into the dataloader.
        """
        if isinstance(cfg_dataset, ListConfig):
            datasets = [
                config.instantiate(single_cfg_dataset, self._tokenizer)
                for single_cfg_dataset in cfg_dataset
            ]
            ds = ConcatDataset(datasets=datasets)
            packed = getattr(ds, "packed", False)
        else:
            ds = config.instantiate(cfg_dataset, self._tokenizer)
            packed = cfg_dataset.get("packed", False)

        # Instantiate collate_fn
        if "left_pad_sequence" in collate_fn:
            raise RuntimeError("left_pad_sequence collator is only for inference.")
        collate_fn = _get_component_from_path(collate_fn)

        sampler = StatefulDistributedSampler(
            ds,
            num_replicas=self.world_size,
            rank=self.rank,
            shuffle=shuffle,
        )
        dataloader = StatefulDataLoader(
            dataset=ds,
            batch_size=batch_size,
            sampler=sampler,
            collate_fn=(
                partial(
                    collate_fn,
                    padding_idx=self._tokenizer.pad_id,
                    ignore_idx=self._loss_fn.ignore_index,
                )
                if not packed
                else padded_collate_packed
            ),
            # dropping last avoids shape issues with compile + flex attention
            drop_last=True,
        )

        return dataloader

    def save_checkpoint(
        self,
        epoch: int,
    ) -> None:
        self._checkpoint_client.save_checkpoint(
            model=self._model,
            optimizer=self._optimizer,
            training_progress=TrainingProgress(
                seed=self.seed,
                epochs_run=self.epochs_run,
                total_epochs=self.total_epochs,
                max_steps_per_epoch=self.max_steps_per_epoch,
                dataloader_state_dict=self._dataloader.state_dict(),
            ),
            epoch=epoch,
            adapter_config=self._adapter_config.copy(),
            adapter_only=self._save_adapter_weights_only,
        )

    def train(self) -> None:
        """
        The core training loop.
        """
        # clean up before training begins
        training.cleanup_before_training()

        # zero out the gradients before starting training
        self._optimizer.zero_grad()

        # Initialize tokens count and running loss (for grad accumulation)
        t0 = time.perf_counter()
        running_loss = 0
        num_tokens = 0

        self._profiler.start()
        # self.epochs_run should be non-zero when we're resuming from a checkpoint
        for curr_epoch in range(self.epochs_run, self.total_epochs):
            pbar = tqdm(total=self._steps_per_epoch, disable=not (self.rank == 0))
            self._dataloader.sampler.set_epoch(curr_epoch)
            for idx, batch in enumerate(self._dataloader):
                # Start tracking CUDA memory for active steps for just the first epoch
                if (
                    self._is_rank_zero
                    and curr_epoch == 0
                    and self.profiler_profile_memory
                    and idx == self.profiler_wait_steps + self.profiler_warmup_steps
                    and self._device.type == "cuda"
                ):
                    torch.cuda.memory._record_memory_history()

                utils.batch_to_device(batch, self._device)

                # Calculate the number of unmasked tokens in the current batch
                # and increment the total number of tokens seen in the step
                current_num_tokens = (
                    batch["labels"] != self._loss_fn.ignore_index
                ).sum()
                num_tokens += current_num_tokens

                # Loss is normalized by default so we multiply by the number of tokens
                # This way we can normalize by the total number of tokens if we're accumulating gradients
                current_loss = self._loss_step(batch) * current_num_tokens
                running_loss += current_loss
                current_loss.backward()

                # Step with optimizer
                if (idx + 1) % self._gradient_accumulation_steps == 0:
                    # Get total number of tokens across all ranks to normalize gradients
                    torch.distributed.all_reduce(num_tokens)
                    # This will ensure that the logged loss matches what we're optimizing
                    torch.distributed.all_reduce(running_loss)
                    # Manually scale the gradients from unnormalized loss by total # of tokens
                    # We multiply by world_size to undo FSDP2 gradient normalization.
                    training.scale_grads(self._model, self.world_size / num_tokens)
                    if self._clip_grad_norm is not None:
                        grad_norm = torch.nn.utils.clip_grad_norm_(
                            self._model.parameters(),
                            max_norm=float(self._clip_grad_norm),
                        ).full_tensor()
                    self._optimizer.step()
                    self._optimizer.zero_grad(set_to_none=True)
                    self._lr_scheduler.step()

                    # Update the number of steps when the weights are updated
                    self.global_step += 1

                    # If float8 training is enabled, perform a single all-reduce to compute the
                    # scale for all float8 parameters efficiently instead of doing many small
                    # all-reduces for each parameter
                    if (
                        self._enable_fp8_training
                        and is_fp8_tensorwise_scaling(self._fp8_recipe_name)
                    ):
                        precompute_float8_dynamic_scale_for_fsdp(self._model)

                    loss_to_log = running_loss.detach().item() / num_tokens
                    pbar.update(1)
                    pbar.set_description(
                        f"{curr_epoch + 1}|{self.global_step}|Loss: {loss_to_log}"
                    )

                    # Log per-step metrics
                    if (
                        self.global_step % self._log_every_n_steps == 0
                        and self._is_rank_zero
                    ):
                        time_per_step = time.perf_counter() - t0
                        log_dict = {
                            "loss": loss_to_log,
                            "lr": self._optimizer.param_groups[0]["lr"],
                            "tokens_per_second_per_gpu": num_tokens
                            / (time_per_step * self.world_size),
                        }
                        if self._log_peak_memory_stats:
                            log_dict.update(
                                training.get_memory_stats(device=self._device)
                            )

                        if self._clip_grad_norm is not None:
                            log_dict.update({"grad_norm": grad_norm})
                        self._metric_logger.log_dict(
                            log_dict,
                            step=self.global_step,
                        )

                    # Reset running stats for the next step
                    running_loss = 0
                    num_tokens = 0
                    t0 = time.perf_counter()

                    # Stop tracking CUDA memory now that active steps are complete
                    if (
                        self._is_rank_zero
                        and curr_epoch == 0
                        and self.profiler_profile_memory
                        and idx
                        == self.profiler_wait_steps
                        + self.profiler_warmup_steps
                        + self.profiler_active_steps
                        and self._device.type == "cuda"
                    ):
                        torch.cuda.memory._record_memory_history(enabled=None)

                    # Step profiler
                    # Note that this is called within gradient accumulation block, hence
                    # will include multiple forward / backward passes if gradient accumulation > 1
                    self._profiler.step()

                    # Run validation after gradient update
                    if (
                        self._run_val_every_n_steps is not None
                        and self.global_step % self._run_val_every_n_steps == 0
                    ):
                        pbar.refresh()
                        self.validate()

                if (
                    (idx + 1) // self._gradient_accumulation_steps
                ) == self.max_steps_per_epoch:
                    break

            self.epochs_run += 1
            self.save_checkpoint(epoch=curr_epoch)

        self._profiler.stop()

    def _loss_step(self, batch: dict[str, torch.Tensor]) -> torch.Tensor:
        # Shape [b, s], needed for the loss not the model
        labels = batch.pop("labels")

        with self.activations_handling_ctx:
            outputs = self._model(**batch)

        # post process for third party loss functions
        if not isinstance(self._loss_fn, SFTLoss):
            labels = labels.reshape(-1)
            outputs = outputs.reshape(-1, outputs.size(-1))
            if isinstance(outputs, DTensor):
                outputs = outputs.full_tensor()

        # Compute loss
        loss = self._loss_fn(outputs, labels)

        # free logits otherwise it peaks backward memory
        del outputs

        return loss

    def validate(self) -> dict[str, float]:
        """
        Run validation loop and return average validation loss.
        """

        self._model.eval()
        total_val_loss = torch.tensor(0.0, device=self._device)
        total_val_tokens = torch.tensor(0.0, device=self._device)

        with torch.no_grad():
            for batch_idx, batch in enumerate(self._val_dataloader):
                utils.batch_to_device(batch, self._device)

                # Count tokens excluding padding
                current_num_tokens = (
                    batch["labels"] != self._loss_fn.ignore_index
                ).sum()

                # Compute loss
                val_loss = self._loss_step(batch) * current_num_tokens

                total_val_loss += val_loss
                total_val_tokens += current_num_tokens

        # Aggregate validation metrics across all ranks
        torch.distributed.all_reduce(total_val_loss)
        torch.distributed.all_reduce(total_val_tokens)

        avg_val_loss = (
            (total_val_loss / total_val_tokens).item()
            if total_val_tokens > 0
            else float("inf")
        )
        log_dict = {"val_loss": avg_val_loss}

        if self._is_rank_zero:
            self._logger.info(f"Validation loss: {avg_val_loss:.4f}")
            self._metric_logger.log_dict(
                log_dict,
                step=self.global_step,
            )

        self._model.train()
        return log_dict

    def cleanup(self) -> None:
        if self._is_rank_zero:
            self._metric_logger.close()
        destroy_process_group()


@config.parse
def recipe_main(cfg: DictConfig) -> None:
    """
    Entry point for the recipe.

    Configurable parameters are read in the following order:
        - Parameters specified in config (see available configs through ``tune ls``)
        - Overwritten by arguments from the command-line
    """
    if not training.is_distributed():
        raise RuntimeError(
            "Distributed finetune recipe should be run via a distributed launcher."
            "If using tune CLI, please specify --nnodes 1 and --nproc_per_node [num_gpus]"
        )
    if cfg.get("fsdp_cpu_offload", False):
        # Utilize all available CPU cores for intra-op parallelism. This provides ~2x
        # speed up when benchmarking fused AdamW on CPU
        training.set_torch_num_threads()

    config.log_config(recipe_name="LoRAFinetuneRecipeDistributed", cfg=cfg)

    recipe = LoRAFinetuneRecipeDistributed(cfg=cfg)
    recipe.setup(cfg=cfg)
    recipe.train()
    recipe.cleanup()


if __name__ == "__main__":
    sys.exit(recipe_main())

`

Error:

[rank6]: from user code:
[rank6]:    File "/workspace/torchtune/torchtune/modules/transformer.py", line 132, in forward
[rank6]:     attn_out = self.attn(h, h, mask=mask, input_pos=input_pos)
[rank6]:   File "/workspace/torchtune/torchtune/modules/attention.py", line 234, in forward
[rank6]:     q = self.q_proj(x)
[rank6]:   File "/workspace/torchtune/torchtune/modules/peft/lora.py", line 144, in forward
[rank6]:     lora_out = self.lora_a(self.dropout(x))
[rank6]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torchao/float8/float8_linear.py", line 352, in forward
[rank6]:     output = matmul_with_hp_or_float8_args.apply(

[rank6]: Set TORCHDYNAMO_VERBOSE=1 for the internal stack trace (please do this especially if you're reporting a bug to PyTorch). For even more developer context, set TORCH_LOGS="+dynamo"

[rank0]:E0617 21:55:42.880000 53586 site-packages/torch/_subclasses/fake_tensor.py:2558] [0/0] failed while attempting to run meta for aten._scaled_mm.default
[rank0]:E0617 21:55:42.880000 53586 site-packages/torch/_subclasses/fake_tensor.py:2558] [0/0] Traceback (most recent call last):
[rank0]:E0617 21:55:42.880000 53586 site-packages/torch/_subclasses/fake_tensor.py:2558] [0/0]   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py", line 2554, in _dispatch_impl
[rank0]:E0617 21:55:42.880000 53586 site-packages/torch/_subclasses/fake_tensor.py:2558] [0/0]     r = func(*args, **kwargs)
[rank0]:E0617 21:55:42.880000 53586 site-packages/torch/_subclasses/fake_tensor.py:2558] [0/0]   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_ops.py", line 776, in __call__
[rank0]:E0617 21:55:42.880000 53586 site-packages/torch/_subclasses/fake_tensor.py:2558] [0/0]     return self._op(*args, **kwargs)
[rank0]:E0617 21:55:42.880000 53586 site-packages/torch/_subclasses/fake_tensor.py:2558] [0/0]   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_meta_registrations.py", line 6244, in meta_scaled_mm
[rank0]:E0617 21:55:42.880000 53586 site-packages/torch/_subclasses/fake_tensor.py:2558] [0/0]     torch._check(
[rank0]:E0617 21:55:42.880000 53586 site-packages/torch/_subclasses/fake_tensor.py:2558] [0/0]   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/__init__.py", line 1669, in _check
[rank0]:E0617 21:55:42.880000 53586 site-packages/torch/_subclasses/fake_tensor.py:2558] [0/0]     _check_with(RuntimeError, cond, message)
[rank0]:E0617 21:55:42.880000 53586 site-packages/torch/_subclasses/fake_tensor.py:2558] [0/0]   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/__init__.py", line 1651, in _check_with
[rank0]:E0617 21:55:42.880000 53586 site-packages/torch/_subclasses/fake_tensor.py:2558] [0/0]     raise error_type(message_evaluated)
[rank0]:E0617 21:55:42.880000 53586 site-packages/torch/_subclasses/fake_tensor.py:2558] [0/0] RuntimeError: Expected both dimensions of mat2 to be divisble by 16 but got torch.Size([4096, 8])
[rank0]: Traceback (most recent call last):
[rank0]:   File "/workspace/torchtune/recipes/lora_finetune_distributed.py", line 939, in <module>
[rank0]:     sys.exit(recipe_main())
[rank0]:   File "/workspace/torchtune/torchtune/config/_parse.py", line 99, in wrapper
[rank0]:     sys.exit(recipe_main(conf))
[rank0]:   File "/workspace/torchtune/recipes/lora_finetune_distributed.py", line 934, in recipe_main
[rank0]:     recipe.train()
[rank0]:   File "/workspace/torchtune/recipes/lora_finetune_distributed.py", line 732, in train
[rank0]:     current_loss = self._loss_step(batch) * current_num_tokens
[rank0]:   File "/workspace/torchtune/recipes/lora_finetune_distributed.py", line 843, in _loss_step
[rank0]:     outputs = self._model(**batch)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1751, in _wrapped_call_impl
[rank0]:     return self._call_impl(*args, **kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1857, in _call_impl
[rank0]:     return inner()
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1805, in inner
[rank0]:     result = forward_call(*args, **kwargs)
[rank0]:   File "/workspace/torchtune/torchtune/modules/transformer.py", line 659, in forward
[rank0]:     h = layer(
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1751, in _wrapped_call_impl
[rank0]:     return self._call_impl(*args, **kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1857, in _call_impl
[rank0]:     return inner()
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1805, in inner
[rank0]:     result = forward_call(*args, **kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/distributed/algorithms/_checkpoint/checkpoint_wrapper.py", line 171, in forward
[rank0]:     return self.checkpoint_fn(  # type: ignore[misc]
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_compile.py", line 51, in inner
[rank0]:     return disable_fn(*args, **kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py", line 850, in _fn
[rank0]:     return fn(*args, **kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/utils/checkpoint.py", line 495, in checkpoint
[rank0]:     ret = function(*args, **kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1749, in _wrapped_call_impl
[rank0]:     return self._compiled_call_impl(*args, **kwargs)  # type: ignore[misc]
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py", line 658, in _fn
[rank0]:     return fn(*args, **kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1762, in _call_impl
[rank0]:     return forward_call(*args, **kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 1453, in __call__
[rank0]:     return self._torchdynamo_orig_callable(
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 1234, in __call__
[rank0]:     result = self._inner_convert(
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 619, in __call__
[rank0]:     return _compile(
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 1080, in _compile
[rank0]:     guarded_code = compile_inner(code, one_graph, hooks, transform)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_utils_internal.py", line 97, in wrapper_function
[rank0]:     return function(*args, **kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 779, in compile_inner
[rank0]:     return _compile_inner(code, one_graph, hooks, transform)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 815, in _compile_inner
[rank0]:     out_code = transform_code_object(code, transform)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/bytecode_transformation.py", line 1422, in transform_code_object
[rank0]:     transformations(instructions, code_options)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 264, in _fn
[rank0]:     return fn(*args, **kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 736, in transform
[rank0]:     tracer.run()
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 3491, in run
[rank0]:     super().run()
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 1345, in run
[rank0]:     while self.step():
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 1254, in step
[rank0]:     self.dispatch_table[inst.opcode](self, inst)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 827, in wrapper
[rank0]:     return inner_fn(self, inst)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 2256, in CALL_FUNCTION_KW
[rank0]:     self.call_function(fn, args, kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 1178, in call_function
[rank0]:     self.push(fn.call_function(self, args, kwargs))  # type: ignore[arg-type]
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/variables/lazy.py", line 201, in realize_and_forward
[rank0]:     return getattr(self.realize(), name)(*args, **kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/variables/nn_module.py", line 952, in call_function
[rank0]:     return variables.UserFunctionVariable(fn, source=source).call_function(
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/variables/functions.py", line 414, in call_function
[rank0]:     return super().call_function(tx, args, kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/variables/functions.py", line 186, in call_function
[rank0]:     return tx.inline_user_function_return(self, [*self.self_args(), *args], kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 1195, in inline_user_function_return
[rank0]:     return InliningInstructionTranslator.inline_call(self, fn, args, kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 3717, in inline_call
[rank0]:     return tracer.inline_call_()
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 3900, in inline_call_
[rank0]:     self.run()
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 1345, in run
[rank0]:     while self.step():
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 1254, in step
[rank0]:     self.dispatch_table[inst.opcode](self, inst)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 827, in wrapper
[rank0]:     return inner_fn(self, inst)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 2146, in CALL_FUNCTION
[rank0]:     self.call_function(fn, args, {})
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 1178, in call_function
[rank0]:     self.push(fn.call_function(self, args, kwargs))  # type: ignore[arg-type]
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/variables/lazy.py", line 201, in realize_and_forward
[rank0]:     return getattr(self.realize(), name)(*args, **kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/variables/nn_module.py", line 952, in call_function
[rank0]:     return variables.UserFunctionVariable(fn, source=source).call_function(
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/variables/functions.py", line 414, in call_function
[rank0]:     return super().call_function(tx, args, kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/variables/functions.py", line 186, in call_function
[rank0]:     return tx.inline_user_function_return(self, [*self.self_args(), *args], kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 1195, in inline_user_function_return
[rank0]:     return InliningInstructionTranslator.inline_call(self, fn, args, kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 3717, in inline_call
[rank0]:     return tracer.inline_call_()
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 3900, in inline_call_
[rank0]:     self.run()
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 1345, in run
[rank0]:     while self.step():
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 1254, in step
[rank0]:     self.dispatch_table[inst.opcode](self, inst)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 827, in wrapper
[rank0]:     return inner_fn(self, inst)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 2146, in CALL_FUNCTION
[rank0]:     self.call_function(fn, args, {})
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 1178, in call_function
[rank0]:     self.push(fn.call_function(self, args, kwargs))  # type: ignore[arg-type]
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/variables/lazy.py", line 201, in realize_and_forward
[rank0]:     return getattr(self.realize(), name)(*args, **kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/variables/nn_module.py", line 952, in call_function
[rank0]:     return variables.UserFunctionVariable(fn, source=source).call_function(
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/variables/functions.py", line 414, in call_function
[rank0]:     return super().call_function(tx, args, kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/variables/functions.py", line 186, in call_function
[rank0]:     return tx.inline_user_function_return(self, [*self.self_args(), *args], kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 1195, in inline_user_function_return
[rank0]:     return InliningInstructionTranslator.inline_call(self, fn, args, kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 3717, in inline_call
[rank0]:     return tracer.inline_call_()
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 3900, in inline_call_
[rank0]:     self.run()
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 1345, in run
[rank0]:     while self.step():
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 1254, in step
[rank0]:     self.dispatch_table[inst.opcode](self, inst)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 827, in wrapper
[rank0]:     return inner_fn(self, inst)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 2146, in CALL_FUNCTION
[rank0]:     self.call_function(fn, args, {})
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 1178, in call_function
[rank0]:     self.push(fn.call_function(self, args, kwargs))  # type: ignore[arg-type]
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/variables/misc.py", line 968, in call_function
[rank0]:     return self.obj.call_method(tx, self.name, args, kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/variables/misc.py", line 698, in call_method
[rank0]:     return wrap_fx_proxy(
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/variables/builder.py", line 2421, in wrap_fx_proxy
[rank0]:     return wrap_fx_proxy_cls(target_cls=TensorVariable, **kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/variables/builder.py", line 2487, in wrap_fx_proxy_cls
[rank0]:     return _wrap_fx_proxy(
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/variables/builder.py", line 2585, in _wrap_fx_proxy
[rank0]:     example_value = get_fake_value(proxy.node, tx, allow_non_graph_fake=True)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/utils.py", line 3269, in get_fake_value
[rank0]:     raise TorchRuntimeError(str(e)).with_traceback(e.__traceback__) from None
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/utils.py", line 3167, in get_fake_value
[rank0]:     ret_val = wrap_fake_exception(
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/utils.py", line 2681, in wrap_fake_exception
[rank0]:     return fn()
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/utils.py", line 3168, in <lambda>
[rank0]:     lambda: run_node(tx.output, node, args, kwargs, nnmodule)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/utils.py", line 3365, in run_node
[rank0]:     raise RuntimeError(make_error_message(e)).with_traceback(
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/utils.py", line 3324, in run_node
[rank0]:     return node.target(*args, **kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_dynamo/variables/misc.py", line 556, in trampoline_autograd_apply
[rank0]:     return fn_cls.apply(*args, **kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/autograd/function.py", line 575, in apply
[rank0]:     return super().apply(*args, **kwargs)  # type: ignore[misc]
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torchao/float8/float8_linear.py", line 121, in forward
[rank0]:     res_bits = torch.mm(input_maybe_fp8_reshaped, weight_maybe_fp8_t)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torchao/float8/float8_tensor.py", line 374, in __torch_dispatch__
[rank0]:     return FLOAT8_OPS_TABLE[func](func, args, kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torchao/float8/float8_ops.py", line 373, in float8_mm
[rank0]:     tensor_out = addmm_float8_unwrapped(
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torchao/float8/float8_ops.py", line 63, in addmm_float8_unwrapped
[rank0]:     output = torch._scaled_mm(
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/utils/_stats.py", line 27, in wrapper
[rank0]:     return fn(*args, **kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py", line 1311, in __torch_dispatch__
[rank0]:     return self.dispatch(func, types, args, kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py", line 1932, in dispatch
[rank0]:     return self._cached_dispatch_impl(func, types, args, kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py", line 1414, in _cached_dispatch_impl
[rank0]:     output = self._dispatch_impl(func, types, args, kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_subclasses/fake_tensor.py", line 2554, in _dispatch_impl
[rank0]:     r = func(*args, **kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_ops.py", line 776, in __call__
[rank0]:     return self._op(*args, **kwargs)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_meta_registrations.py", line 6244, in meta_scaled_mm
[rank0]:     torch._check(
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/__init__.py", line 1669, in _check
[rank0]:     _check_with(RuntimeError, cond, message)
[rank0]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/__init__.py", line 1651, in _check_with
[rank0]:     raise error_type(message_evaluated)
[rank0]: torch._dynamo.exc.TorchRuntimeError: Dynamo failed to run FX node with fake tensors: call_function <function produce_trampoline_autograd_apply.<locals>.trampoline_autograd_apply at 0x7f50eb366ef0>(*(FakeTensor(..., device='cuda:0', size=(16, 8192, 4096), dtype=torch.bfloat16), Float8Tensor(dtype=torch.float8_e4m3fnuz, scale=FakeTensor(..., device='cuda:0', size=()), linear_mm_config=LinearMMConfig(output=ScaledMMConfig(emulate=False, use_fast_accum=True, fp8_output=False, pad_inner_dim=False), grad_input=ScaledMMConfig(emulate=False, use_fast_accum=False, fp8_output=False, pad_inner_dim=False), grad_weight=ScaledMMConfig(emulate=False, use_fast_accum=False, fp8_output=False, pad_inner_dim=False)), axiswise_dim=None
[rank0]: gemm_input_role=GemmInputRole.WEIGHT
[rank0]: as_orig_prec=FakeTensor(..., device='cuda:0', size=(4096, 8), dtype=torch.bfloat16,
[rank0]:            grad_fn=<_FromFloat8ConstrFuncBackward>), LinearMMConfig(output=ScaledMMConfig(emulate=False, use_fast_accum=True, fp8_output=False, pad_inner_dim=False), grad_input=ScaledMMConfig(emulate=False, use_fast_accum=False, fp8_output=False, pad_inner_dim=False), grad_weight=ScaledMMConfig(emulate=False, use_fast_accum=False, fp8_output=False, pad_inner_dim=False)), Float8LinearConfig(cast_config_input=CastConfig(scaling_type=<ScalingType.DYNAMIC: 'dynamic'>, scaling_granularity=<ScalingGranularity.TENSORWISE: 'tensorwise'>, target_dtype=torch.float8_e4m3fnuz), cast_config_input_for_grad_weight=CastConfig(scaling_type=<ScalingType.DYNAMIC: 'dynamic'>, scaling_granularity=<ScalingGranularity.TENSORWISE: 'tensorwise'>, target_dtype=torch.float8_e4m3fnuz), cast_config_weight=CastConfig(scaling_type=<ScalingType.DYNAMIC: 'dynamic'>, scaling_granularity=<ScalingGranularity.TENSORWISE: 'tensorwise'>, target_dtype=torch.float8_e4m3fnuz), cast_config_weight_for_grad_input=CastConfig(scaling_type=<ScalingType.DYNAMIC: 'dynamic'>, scaling_granularity=<ScalingGranularity.TENSORWISE: 'tensorwise'>, target_dtype=torch.float8_e4m3fnuz), cast_config_grad_output=CastConfig(scaling_type=<ScalingType.DYNAMIC: 'dynamic'>, scaling_granularity=<ScalingGranularity.TENSORWISE: 'tensorwise'>, target_dtype=torch.float8_e5m2fnuz), cast_config_grad_output_for_grad_weight=CastConfig(scaling_type=<ScalingType.DYNAMIC: 'dynamic'>, scaling_granularity=<ScalingGranularity.TENSORWISE: 'tensorwise'>, target_dtype=torch.float8_e5m2fnuz), gemm_config_output=Float8GemmConfig(use_fast_accum=True), gemm_config_grad_input=Float8GemmConfig(use_fast_accum=False), gemm_config_grad_weight=Float8GemmConfig(use_fast_accum=False), enable_fsdp_float8_all_gather=False, pad_inner_dim=False, emulate=False, force_recompute_fp8_weight_in_bwd=False, round_scales_to_power_of_2=False)), **{}): got RuntimeError('Expected both dimensions of mat2 to be divisble by 16 but got torch.Size([4096, 8])')`

Can someone help me with this?

[felipemello1]

hey @kailashg26 , I dont think that it works well with LoRA. My understanding is that you wouldnt do FP8 on the quantized frozen weights, which are NF4. You could do FP8 on the LoRA adapters, but these are already so tiny that i cant think that you would see improvements anyway.

We dont have bandwidth to look at this at the moment, but if you decide to keep looking, I asked @ebsmothers about it and he shared this:

I suspect it's because of how torchao does module swap, so it won't pick up our LoRALinear layers: https://github.com/pytorch/ao/blob/8b12ddf7735ab82026e8fcd99e5232f44e5e779f/torchao/float8/float8_linear_utils.py#L84

[kailashg26]

Yeah, for 70B lora it worked, but indeed there is a bit of performance loss (tokens processed).

I tried llama-4 17B_16E on single node and run into same issue even with full weight finetuning. Do you know why this happens?

[rank3]:   File "/workspace/torchtune/recipes/full_finetune_distributed.py", line 1088, in recipe_main
[rank3]:     recipe.train()
[rank3]:   File "/workspace/torchtune/recipes/full_finetune_distributed.py", line 933, in train
[rank3]:     current_loss.backward()
[rank3]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_tensor.py", line 648, in backward
[rank3]:     torch.autograd.backward(
[rank3]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/autograd/__init__.py", line 353, in backward
[rank3]:     _engine_run_backward(
[rank3]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/autograd/graph.py", line 824, in _engine_run_backward
[rank3]:     return Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
[rank3]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/autograd/function.py", line 307, in apply
[rank3]:     return user_fn(self, *args)
[rank3]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torchao/float8/float8_linear.py", line 244, in backward
[rank3]:     grad_weight = torch.mm(
[rank3]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torchao/float8/float8_tensor.py", line 374, in __torch_dispatch__
[rank3]:     return FLOAT8_OPS_TABLE[func](func, args, kwargs)
[rank3]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torchao/float8/float8_ops.py", line 373, in float8_mm
[rank3]:     tensor_out = addmm_float8_unwrapped(
[rank3]:   File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torchao/float8/float8_ops.py", line 63, in addmm_float8_unwrapped
[rank3]:     output = torch._scaled_mm(
[rank3]: RuntimeError: Expected trailing dimension of mat1 to be divisible by 16 but got mat1 shape: (202048x4072).

[felipemello1]

@ebsmothers , mind taking a look?

[vkuzo]

[rank3]: RuntimeError: Expected trailing dimension of mat1 to be divisible by 16 but got mat1 shape: (202048x4072).

The float8 matrix multiply requires shapes with very specific divisibility requirements of operands, and it looks like one of your shapes does not match those requirements. You have some options:

1. exclude whichever linear this appears on from float8 conversion
2. enable padding (https://github.com/pytorch/ao/blob/420f782b5769256cb25575927566a39387bb3467/torchao/float8/config.py#L190) which may help in some cases

[kailashg26]

Thanks @vkuzo

I was trying to integrate fp8 for lora. I did not get the speedup I did MLP-> apply lora->FP8, but rather I changed the lora implementation to MLP->FP8 -> apply lora. I tested this on llama 3.1 70b and was able to improve the throughput from 1080.72 (t/sec/gpu) to 1409.82 (t/sec/gpu). I have done this on single node with 8 GPUs.
I was using this commit of torchutne (b4c98ac). However, one issue, I'm still trying to figure out is the loss, FP8_tweaked is my modification and FP8_lora_orig is basically I followed how FP8 is implemented on full weight finetuning and transplanted it to LoRA. However, the FP8_lora_orig was not giving any speedups, so I had to make some changes.
Can anyone give me insights/help on this one?

lora_finetune_distributed.py

I also shared by lora_finetune_distributed.py file in the chat! Appreciate the help!