improve handling and error if fa3 requested but not installeD

* don't set peft_config on grpo to prevent double peft wrap

* remove overrides needed to support bug

* fix grpo tests

* require more CPU for multigpu to help with torch compile for vllm

.github/workflows/base.yml

@@ -47,11 +47,18 @@ jobs:
             pytorch: 2.7.0
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
           - cuda: "128"
-            cuda_version: 12.6.3
+            cuda_version: 12.8.1
             cudnn_version: ""
             python_version: "3.11"
             pytorch: 2.7.0
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+          - cuda: "126"
+            cuda_version: 12.6.3
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.6.0
+            suffix: "-hopper"
+            torch_cuda_arch_list: "9.0+PTX"
           - cuda: "128"
             cuda_version: 12.8.1
             cudnn_version: ""
@@ -87,7 +94,7 @@ jobs:
           context: .
           file: ${{ matrix.pytorch == 'nightly' && './docker/Dockerfile-base-nightly' || matrix.pytorch == 'next' && './docker/Dockerfile-base-next' || './docker/Dockerfile-base' }}
           push: ${{ github.event_name != 'pull_request' }}
-          tags: ${{ steps.metadata.outputs.tags }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
+          tags: ${{ steps.metadata.outputs.tags }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}${{ matrix.suffix || '' }}
           labels: ${{ steps.metadata.outputs.labels }}
           build-args: |
             CUDA_VERSION=${{ matrix.cuda_version }}

.github/workflows/multi-gpu-e2e.yml

@@ -32,21 +32,25 @@ jobs:
             pytorch: 2.6.0
             axolotl_extras: vllm
             num_gpus: 2
-            nightly_build: "true"
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.6.0
+            axolotl_extras:
+            suffix: "-hopper"
+            num_gpus: 2
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
             pytorch: 2.5.1
             axolotl_extras:
             num_gpus: 2
-            nightly_build: "true"
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.7.0
             axolotl_extras:
             num_gpus: 2
-            nightly_build: "true"
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
     steps:
@@ -68,7 +72,6 @@ jobs:
           echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
-          echo "NIGHTLY_BUILD=${{ matrix.nightly_build }}" >> $GITHUB_ENV
           echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         run: |

cicd/Dockerfile.jinja

@@ -32,6 +32,11 @@ RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
     fi
 
 RUN pip install packaging==23.2 setuptools==75.8.0
+RUN if [ "$PYTORCH_VERSION" = "2.6.0" ] && [ "$CUDA" = "126" ] ; then \
+        curl -L -O https://d1dttdx32dkk5p.cloudfront.net/fa3/cu${CUDA}/torch-${PYTORCH_VERSION}/flash_attn_3-3.0.0b1-cp311-cp311-linux_x86_64.whl; \
+        pip3 install --no-cache-dir flash_attn_3-3.0.0b1-cp311-cp311-linux_x86_64.whl; \
+        rm flash_attn_3-3.0.0b1-cp311-cp311-linux_x86_64.whl; \
+    fi
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
         pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \

cicd/multigpu.py

@@ -70,7 +70,7 @@ def run_cmd(cmd: str, run_folder: str):
     image=cicd_image,
     gpu=GPU_CONFIG,
     timeout=90 * 60,
-    cpu=8.0,
+    cpu=16.0,
     memory=131072 * N_GPUS,
     volumes=VOLUME_CONFIG,
 )

docker/Dockerfile-base

@@ -1,5 +1,5 @@
-ARG CUDA_VERSION="11.8.0"
+ARG CUDA_VERSION="12.4.1"
-ARG CUDNN_VERSION="8"
+ARG CUDNN_VERSION=""
 ARG UBUNTU_VERSION="22.04"
 ARG MAX_JOBS=4
 
@@ -7,16 +7,16 @@ FROM nvidia/cuda:$CUDA_VERSION-cudnn$CUDNN_VERSION-devel-ubuntu$UBUNTU_VERSION A
 
 ENV PATH="/root/miniconda3/bin:${PATH}"
 
-ARG PYTHON_VERSION="3.10"
+ARG PYTHON_VERSION="3.11"
-ARG PYTORCH_VERSION="2.1.2"
+ARG PYTORCH_VERSION="2.5.1"
-ARG CUDA="118"
+ARG CUDA="124"
 ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 9.0+PTX"
 
 ENV PYTHON_VERSION=$PYTHON_VERSION
 ENV TORCH_CUDA_ARCH_LIST=$TORCH_CUDA_ARCH_LIST
 
 RUN apt-get update \
-    && apt-get install -y wget git build-essential ninja-build git-lfs libaio-dev pkg-config && rm -rf /var/lib/apt/lists/* \
+    && apt-get install -y wget git build-essential ninja-build git-lfs libaio-dev pkg-config curl && rm -rf /var/lib/apt/lists/* \
     && wget \
     https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh \
     && mkdir /root/.conda \
@@ -38,6 +38,10 @@ RUN git lfs install --skip-repo && \
     # The base image ships with `pydantic==1.8.2` which is not working
     pip3 install -U --no-cache-dir pydantic==1.10.10
 
-RUN if [ "$PYTORCH_VERSION" = "2.7.0" ] ; then \
+RUN if [ "$TORCH_CUDA_ARCH_LIST" = "9.0+PTX" ] ; then \
+        curl -L -O https://d1dttdx32dkk5p.cloudfront.net/fa3/cu${CUDA}/torch-${PYTORCH_VERSION}/flash_attn_3-3.0.0b1-cp311-cp311-linux_x86_64.whl; \
+        pip3 install --no-cache-dir flash_attn_3-3.0.0b1-cp311-cp311-linux_x86_64.whl; \
+        rm flash_attn_3-3.0.0b1-cp311-cp311-linux_x86_64.whl; \
+    elif [ "$PYTORCH_VERSION" = "2.7.0" ] ; then \
         pip3 install flash-attn==2.7.4.post1; \
     fi

docs/config.qmd

@@ -633,7 +633,9 @@ weight_decay:
 # adamw hyperparams
 adam_beta1:
 adam_beta2:
+adam_beta3:  # only used for CAME Optimizer
 adam_epsilon:
+adam_epsilon2:  # only used for CAME Optimizer
 # Gradient clipping max norm
 max_grad_norm:
 

src/axolotl/core/trainer_builder.py

@@ -387,8 +387,12 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             training_arguments_kwargs["adam_beta1"] = self.cfg.adam_beta1
         if self.cfg.adam_beta2:
             training_arguments_kwargs["adam_beta2"] = self.cfg.adam_beta2
+        if self.cfg.adam_beta3:
+            training_arguments_kwargs["adam_beta3"] = self.cfg.adam_beta3
         if self.cfg.adam_epsilon:
             training_arguments_kwargs["adam_epsilon"] = self.cfg.adam_epsilon
+        if self.cfg.adam_epsilon2:
+            training_arguments_kwargs["adam_epsilon2"] = self.cfg.adam_epsilon2
         if self.cfg.max_grad_norm:
             training_arguments_kwargs["max_grad_norm"] = self.cfg.max_grad_norm
 
@@ -713,7 +717,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
 
                 beta1 = training_arguments_kwargs.get("adam_beta1", 0.9)
                 beta2 = training_arguments_kwargs.get("adam_beta2", 0.999)
-                beta3 = training_arguments_kwargs.get("adam_beta2", 0.9999)
+                beta3 = training_arguments_kwargs.get("adam_beta3", 0.9999)
                 eps1 = training_arguments_kwargs.get("adam_epsilon", 1e-30)
                 eps2 = training_arguments_kwargs.get("adam_epsilon2", 1e-16)
                 adam_kwargs["betas"] = (beta1, beta2, beta3)
@@ -1170,7 +1174,8 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         if self.eval_dataset:
             trainer_kwargs["eval_dataset"] = self.eval_dataset
         if self.cfg.adapter and self.peft_config:
-            trainer_kwargs["peft_config"] = self.peft_config
+            if self.cfg.rl is not RLType.GRPO:
+                trainer_kwargs["peft_config"] = self.peft_config
         if self.cfg.precompute_ref_log_probs is not None:
             trainer_kwargs["precompute_ref_log_probs"] = (
                 self.cfg.precompute_ref_log_probs

src/axolotl/core/trainers/base.py

@@ -156,9 +156,6 @@ class AxolotlTrainer(
         Helper method to get the sampler for evaluation. Handles sequence parallelism
         and sample packing cases.
 
-        Args:
-            eval_dataset: Evaluation dataset.
-
         Returns:
             If the dataset is non-empty, a sampler is returned, the type of which
                 depends on the passed training args.
@@ -240,6 +237,9 @@ class AxolotlTrainer(
             self.accelerator.even_batches = False
 
         # Return unprepared dataloader if using sequence parallelism
+        # TODO(djsaunde): We might be able to use `accelerate`'s dataloader preparation
+        # if we use `dispatch_batches` and `slice_fn_for_dispatch` properly (i.e.,
+        # slice each batch along the sequence dimension).
         if self.args.sequence_parallel_degree > 1:
             return dataloader
 

src/axolotl/core/trainers/dpo/trainer.py

@@ -1,25 +1,33 @@
-"""DPO trainer for Axolotl"""
+"""
+DPO trainer for axolotl
+"""
 
 import gc
+import random
 from functools import wraps
-from typing import Any, Dict, Union
+from typing import Any, Dict, Optional, Union
 
+import pandas as pd
 import torch
-from datasets import Dataset
+import wandb
+from accelerate import PartialState
+from datasets import Dataset, IterableDataset
 from peft.optimizers import create_loraplus_optimizer
 from torch import nn
-from torch.utils.data import Sampler
+from torch.utils.data import DataLoader
 from transformers import (
+    BaseImageProcessor,
+    FeatureExtractionMixin,
+    PreTrainedTokenizerBase,
+    ProcessorMixin,
     Trainer,
 )
+from transformers.trainer_utils import EvalLoopOutput
 from transformers.utils import is_sagemaker_mp_enabled
-from trl import DPOTrainer
+from trl import DPOConfig, DPOTrainer, maybe_apply_chat_template, maybe_extract_prompt
+from trl.trainer.utils import log_table_to_comet_experiment
 
-from axolotl.core.trainers.mixins import (
+from axolotl.core.trainers.mixins import RngLoaderMixin, SchedulerMixin
-    RngLoaderMixin,
-    SchedulerMixin,
-    SequenceParallelMixin,
-)
 from axolotl.core.trainers.utils import (
     sanitize_kwargs_for_ds_tagging,
     sanitize_kwargs_for_tagging,
@@ -29,10 +37,10 @@ if is_sagemaker_mp_enabled():
     import smdistributed.modelparallel.torch as smp
 
 
-class AxolotlDPOTrainer(
+class AxolotlDPOTrainer(RngLoaderMixin, SchedulerMixin, DPOTrainer):
-    RngLoaderMixin, SchedulerMixin, SequenceParallelMixin, DPOTrainer
+    """
-):
+    Extend the base DPOTrainer for axolotl helpers
-    """Extend the base DPOTrainer for axolotl helpers"""
+    """
 
     tag_names = ["axolotl", "dpo"]
 
@@ -87,6 +95,64 @@ class AxolotlDPOTrainer(
 
         return super().push_to_hub(*args, **kwargs)
 
+    # TODO: remove this once https://github.com/huggingface/trl/pull/3377 is in a release
+    def _prepare_dataset(
+        self,
+        dataset: Union[Dataset, IterableDataset],
+        processing_class: Union[
+            PreTrainedTokenizerBase,
+            BaseImageProcessor,
+            FeatureExtractionMixin,
+            ProcessorMixin,
+        ],
+        args: DPOConfig,
+        dataset_name: str,
+    ) -> Union[Dataset, IterableDataset]:
+        # Build the kwargs for the `map` function
+        map_kwargs: Dict[str, Any] = {"writer_batch_size": 10}
+        if isinstance(dataset, Dataset):  # IterableDataset does not support num_proc
+            map_kwargs["num_proc"] = args.dataset_num_proc
+
+        with PartialState().main_process_first():
+            # Extract prompt if needed
+            if isinstance(
+                dataset, Dataset
+            ):  # `IterableDataset.map` does not support `desc`
+                map_kwargs["desc"] = f"Extracting prompt in {dataset_name} dataset"
+            dataset = dataset.map(maybe_extract_prompt, **map_kwargs)
+
+            # Apply the chat template if needed
+            if isinstance(
+                dataset, Dataset
+            ):  # `IterableDataset.map` does not support `desc`
+                map_kwargs["desc"] = f"Applying chat template to {dataset_name} dataset"
+            dataset = dataset.map(
+                maybe_apply_chat_template,
+                fn_kwargs={"tokenizer": processing_class, "tools": args.tools},
+                **map_kwargs,
+            )
+
+            # Tokenize the dataset
+            if isinstance(
+                dataset, Dataset
+            ):  # `IterableDataset.map` does not support `desc`
+                map_kwargs["desc"] = f"Tokenizing {dataset_name} dataset"
+
+            dataset = dataset.map(
+                self.tokenize_row if not self.is_vision_model else self.process_row,
+                remove_columns=["chosen", "rejected"],
+                fn_kwargs={
+                    "processing_class": processing_class,
+                    "max_prompt_length": args.max_prompt_length,
+                    "max_completion_length": args.max_completion_length,
+                    # for enc-dec, we add the special tokens ([bos_token] + prompt + [eos_token]; completion + [eos_token])
+                    "add_special_tokens": False,
+                },
+                **map_kwargs,
+            )
+
+        return dataset
+
     @staticmethod
     def tokenize_row(
         features,
@@ -127,48 +193,68 @@ class AxolotlDPOTrainer(
         torch.cuda.empty_cache()
         return loss
 
-    def _get_train_sampler(self) -> Sampler | None:
+    # TODO: remove this once https://github.com/huggingface/trl/pull/3377 is in a release
+    def evaluation_loop(
+        self,
+        dataloader: DataLoader,
+        description: str,
+        prediction_loss_only: Optional[bool] = None,
+        ignore_keys: Optional[list[str]] = None,
+        metric_key_prefix: str = "eval",
+    ) -> EvalLoopOutput:
         """
-        Helper method to get the sampler for training. Handles cases for sequence
+        Overriding built-in evaluation loop to store metrics for each batch.
-        parallelism, sample packing, and curriculum sampling (sequential).
+        Prediction/evaluation loop, shared by `Trainer.evaluate()` and `Trainer.predict()`.
 
-        Returns:
+        Works both with or without labels.
-            If the dataset is non-empty, a sampler is returned, the type of which
-                depends on the passed training args.
         """
-        import torch.distributed as dist
 
-        if dist.get_rank() == 0:
+        # Sample and save to game log if requested (for one batch to save time)
-            import ipdb
+        if self.generate_during_eval:
+            # Generate random indices within the range of the total number of samples
+            num_samples = len(dataloader.dataset)
+            random_indices = random.sample(
+                range(num_samples), k=self.args.eval_batch_size
+            )
 
-            ipdb.set_trace()
+            # Use dataloader.dataset.select to get the random batch without iterating over the DataLoader
-        dist.barrier()
+            random_batch_dataset = dataloader.dataset.select(random_indices)
-        if dist.get_rank() == 1:
+            random_batch = self.data_collator(random_batch_dataset)
-            import ipdb
+            random_batch = self._prepare_inputs(random_batch)
 
-            ipdb.set_trace()
+            policy_output_decoded, ref_output_decoded = (
-        dist.barrier()
+                self.generate_from_model_and_ref(self.model, random_batch)
+            )
 
-        if self.args.sequence_parallel_degree > 1:
+            table = pd.DataFrame(
-            return self._sp_get_train_sampler(self.train_dataset)
+                columns=["Prompt", "Policy", "Ref Model"],
+                data=[
+                    [prompt, pol[len(prompt) :], ref[len(prompt) :]]
+                    for prompt, pol, ref in zip(
+                        random_batch_dataset["prompt"],
+                        policy_output_decoded,
+                        ref_output_decoded,
+                    )
+                ],
+            )
+            if "wandb" in self.args.report_to and self.accelerator.is_main_process:
+                wandb.log({"game_log": wandb.Table(data=table)})
 
-        return super()._get_train_sampler()
+            if "comet_ml" in self.args.report_to:
+                log_table_to_comet_experiment(
+                    name="game_log.csv",
+                    table=table,
+                )
 
-    def _get_eval_sampler(self, eval_dataset: Dataset | None = None) -> Sampler | None:
+        # Base evaluation
-        """
+        initial_output = super(  # pylint: disable=bad-super-call
-        Helper method to get the sampler for evaluation. Handles sequence parallelism
+            DPOTrainer, self
-        and sample packing cases.
+        ).evaluation_loop(
+            dataloader,
+            description,
+            prediction_loss_only,
+            ignore_keys,
+            metric_key_prefix,
+        )
 
-        Args:
+        return initial_output
-            eval_dataset: Evaluation dataset.
-
-        Returns:
-            If the dataset is non-empty, a sampler is returned, the type of which
-                depends on the passed training args.
-        """
-        eval_dataset = eval_dataset if eval_dataset is not None else self.eval_dataset
-
-        if self.args.sequence_parallel_degree > 1:
-            return self._sp_get_eval_sampler(eval_dataset)
-
-        return super()._get_eval_sampler(eval_dataset)

src/axolotl/core/trainers/grpo/trainer.py

@@ -3,7 +3,6 @@
 # pylint: disable=too-many-lines,duplicate-code,protected-access,no-member
 
 import warnings
-from contextlib import nullcontext
 from typing import Any
 
 import datasets
@@ -14,7 +13,7 @@ from accelerate.utils import (
     broadcast_object_list,
     gather,
     gather_object,
-    is_peft_model,
+    is_peft_available,
 )
 from datasets import Dataset, IterableDataset
 from torch import nn
@@ -30,15 +29,13 @@ from transformers import (
     TrainerCallback,
 )
 from transformers.trainer_utils import seed_worker
-from transformers.utils import is_peft_available
 from trl import GRPOTrainer
 from trl.data_utils import (
     apply_chat_template,
     is_conversational,
     maybe_apply_chat_template,
 )
-from trl.extras.profiling import profiling_context, profiling_decorator
+from trl.extras.profiling import profiling_context
-from trl.import_utils import is_deepspeed_available
 from trl.models import unwrap_model_for_generation
 from trl.trainer.grpo_config import GRPOConfig
 from trl.trainer.grpo_trainer import RewardFunc, nanstd
@@ -52,62 +49,12 @@ if is_peft_available():
     # pylint: disable=unused-import
     from peft import PeftConfig
 
-if is_deepspeed_available():
-    import deepspeed
-
 
 class AxolotlGRPOTrainer(RngLoaderMixin, SchedulerMixin, GRPOTrainer):
     """Extend the base GRPOTrainer for axolotl helpers"""
 
     _tag_names = ["trl", "grpo", "axolotl"]
 
-    @profiling_decorator
-    def _move_model_to_vllm(self):
-        # For DeepSpeed ZeRO-3, we need to gather all parameters before operations
-        deepspeed_plugin = self.accelerator.state.deepspeed_plugin
-        zero_stage_3 = deepspeed_plugin is not None and deepspeed_plugin.zero_stage == 3
-        gather_if_zero3 = (
-            deepspeed.zero.GatheredParameters if zero_stage_3 else nullcontext
-        )
-
-        if is_peft_model(self.model):
-            # With PEFT and DeepSpeed ZeRO Stage 3, we must gather the full model at once before merging, as merging
-            # adapters in a sharded manner is not supported.
-            with gather_if_zero3(list(self.model.parameters())):
-                self.model.merge_adapter()
-
-                # Update vLLM weights while parameters are gathered
-                for name, param in self.model.named_parameters():
-                    # When using PEFT, we need to recover the original parameter name and discard some parameters
-                    name = (
-                        name.removeprefix("base_model.model.")
-                        .removeprefix("base_model.model.")
-                        .replace(".base_layer", "")
-                    )
-                    if self.model.prefix in name:
-                        continue
-                    # When module to save, remove its prefix and discard the original module
-                    if "original_module" in name:
-                        continue
-                    name = name.replace("modules_to_save.default.", "")
-
-                    if self.accelerator.is_main_process:
-                        self.vllm_client.update_named_param(name, param.data)
-
-                # Unmerge adapters while parameters are still gathered
-                self.model.unmerge_adapter()
-                # Parameters will automatically be repartitioned when exiting the context
-        else:
-            # For non-PEFT models, simply gather and update each parameter individually.
-            for name, param in self.model.named_parameters():
-                with gather_if_zero3([param]):
-                    if self.accelerator.is_main_process:
-                        self.vllm_client.update_named_param(name, param.data)
-
-        # Reset cache on main process
-        if self.accelerator.is_main_process:
-            self.vllm_client.reset_prefix_cache()
-
 
 class AxolotlGRPOSequenceParallelTrainer(AxolotlGRPOTrainer):
     """Extend the base GRPOTrainer for sequence parallelism handling"""
@@ -266,6 +213,9 @@ class AxolotlGRPOSequenceParallelTrainer(AxolotlGRPOTrainer):
             self.accelerator.even_batches = False
 
         # Return unprepared dataloader if using sequence parallelism
+        # TODO(djsaunde): We might be able to use `accelerate`'s dataloader preparation
+        # if we use `dispatch_batches` and `slice_fn_for_dispatch` properly (i.e.,
+        # slice each batch along the sequence dimension).
         if self.args.sequence_parallel_degree > 1:
             return dataloader
 

src/axolotl/core/training_args.py

@@ -227,6 +227,19 @@ class AxolotlTrainingMixins:
         },
     )
 
+    adam_beta3: Optional[float] = field(
+        default=None,
+        metadata={
+            "help": "The beta3 hyperparameter used in some optimizers such as CAME"
+        },
+    )
+    adam_epsilon2: Optional[float] = field(
+        default=None,
+        metadata={
+            "help": "The epsilon2 hyperparameter used in some optimizers such as CAME"
+        },
+    )
+
     # multi-modal section
 
     image_size: int | tuple[int, int] | None = field(

src/axolotl/utils/callbacks/mlflow_.py

@@ -1,6 +1,7 @@
 """MLFlow module for trainer callbacks"""
 
 import logging
+import os
 from shutil import copyfile
 from tempfile import NamedTemporaryFile
 from typing import TYPE_CHECKING
@@ -16,6 +17,11 @@ if TYPE_CHECKING:
 LOG = logging.getLogger("axolotl.callbacks")
 
 
+def should_log_artifacts() -> bool:
+    truths = ["TRUE", "1", "YES"]
+    return os.getenv("HF_MLFLOW_LOG_ARTIFACTS", "FALSE").upper() in truths
+
+
 class SaveAxolotlConfigtoMlflowCallback(TrainerCallback):
     # pylint: disable=duplicate-code
     """Callback to save axolotl config to mlflow"""
@@ -32,13 +38,18 @@ class SaveAxolotlConfigtoMlflowCallback(TrainerCallback):
     ):
         if is_main_process():
             try:
-                with NamedTemporaryFile(
+                if should_log_artifacts():
-                    mode="w", delete=False, suffix=".yml", prefix="axolotl_config_"
+                    with NamedTemporaryFile(
-                ) as temp_file:
+                        mode="w", delete=False, suffix=".yml", prefix="axolotl_config_"
-                    copyfile(self.axolotl_config_path, temp_file.name)
+                    ) as temp_file:
-                    mlflow.log_artifact(temp_file.name, artifact_path="")
+                        copyfile(self.axolotl_config_path, temp_file.name)
+                        mlflow.log_artifact(temp_file.name, artifact_path="")
+                        LOG.info(
+                            "The Axolotl config has been saved to the MLflow artifacts."
+                        )
+                else:
                     LOG.info(
-                        "The Axolotl config has been saved to the MLflow artifacts."
+                        "Skipping logging artifacts to MLflow (hf_mlflow_log_artifacts is false)"
                     )
             except (FileNotFoundError, ConnectionError) as err:
                 LOG.warning(f"Error while saving Axolotl config to MLflow: {err}")

src/axolotl/utils/ctx_managers/sequence_parallel.py

@@ -1,7 +1,6 @@
 """Module for Axolotl trainer sequence parallelism manager and utilities"""
 
 import functools
-import inspect
 
 import torch
 import torch.distributed as dist
@@ -33,7 +32,7 @@ def apply_sequence_parallelism(
     to only keep the last N tokens in the sequence during generation.
 
     Args:
-        batch: Dictionary of model arguments (e.g., input_ids, attention_mask, etc.).
+        batch: Batch dictionary (e.g., input_ids, attention_mask, etc.).
         local_rank: Local rank in the sequence parallel group.
         local_world_size: World size of the sequence parallel group.
         gradient_accumulation_steps: Number of steps to accumulate gradients over.
@@ -207,26 +206,12 @@ class SequenceParallelContextManager:
     def __enter__(self):
         # Forward pre-hook to apply sequence parallelism
         def sequence_parallel_pre_hook(_, args, kwargs):
-            # Convert all args to kwargs using the model's forward function signature
+            # Apply sequence parallelism to kwargs and get original sequence length and padding info
-            updated_kwargs = kwargs.copy()
+            kwargs, self.original_seq_len, self.pad_len = (
-
+                self.apply_sequence_parallelism(batch=kwargs)
-            # Get parameter names from the model's forward function
-            forward_params = list(
-                inspect.signature(self.models[0].forward).parameters.keys()
             )
 
-            # Map args to their parameter names
+            return args, kwargs
-            for i, arg in enumerate(args):
-                if i < len(forward_params):
-                    param_name = forward_params[i]
-                    updated_kwargs[param_name] = arg
-
-            # Apply sequence parallelism to empty args and updated kwargs
-            updated_kwargs, self.original_seq_len, self.pad_len = (
-                self.apply_sequence_parallelism(updated_kwargs)
-            )
-
-            return (), updated_kwargs
 
         # Forward post-hook to gather outputs
         def sequence_parallel_post_hook(_, __, output: ModelOutput) -> ModelOutput:

src/axolotl/utils/models.py

@@ -629,6 +629,49 @@ class ModelLoader:
             )
 
         if self.cfg.flash_attention:
+            use_fa3 = False
+            if self.cfg.use_flash_attention_3 is True:
+                use_fa3 = True
+            elif self.cfg.use_flash_attention_3 == "auto":
+                if torch.cuda.get_device_capability() >= (9, 0):
+                    # FA3 is only available on Hopper GPUs and newer
+                    use_fa3 = True
+                if not importlib.util.find_spec("flash_attn_interface"):
+                    use_fa3 = False
+            if use_fa3 and not importlib.util.find_spec("flash_attn_interface"):
+                # this can happen when use_flash_attention_3 is explicity set to True
+                # and flash_attn_interface is not installed
+                raise ModuleNotFoundError(
+                    "Please install the flash_attn_interface library to use Flash Attention 3.x"
+                )
+            if use_fa3 and importlib.util.find_spec("flash_attn_interface") is not None:
+                from flash_attn_interface import flash_attn_func as flash_attn_func_v3
+                from flash_attn_interface import (
+                    flash_attn_varlen_func as flash_attn_varlen_func_v3,
+                )
+
+                def flash_attn_func_v3_wrapper(*args, **kwargs):
+                    kwargs.pop("dropout_p", None)
+                    if "softmax_scale" in kwargs and len(args) >= 4:
+                        # if softmax_scale is provided, then the 3rd position is dropout_p that we need to drop
+                        args = (*args[:3],) + args[4:]
+                    return flash_attn_func_v3(*args, **kwargs)[0]
+
+                def flash_attn_varlen_func_v3_wrapper(*args, **kwargs):
+                    kwargs.pop("dropout_p", None)
+                    if "softmax_scale" in kwargs and len(args) >= 4:
+                        # if softmax_scale is provided, then the 3rd position is dropout_p that we need to drop
+                        args = (*args[:3],) + args[4:]
+                    return flash_attn_varlen_func_v3(*args, **kwargs)[0]
+
+                transformers.modeling_flash_attention_utils.flash_attn_func = (
+                    flash_attn_func_v3_wrapper
+                )
+                transformers.modeling_flash_attention_utils.flash_attn_varlen_func = (
+                    flash_attn_varlen_func_v3_wrapper
+                )
+                LOG.info("Switched to Flash Attention v3")
+
             self.patch_attention()
 
         if self.cfg.sample_packing and self.cfg.s2_attention:
@@ -699,6 +742,7 @@ class ModelLoader:
 
                 patch_mllama()
 
+            # TODO deprecate soon
             if self.model_config.model_type == "btlm":
                 from axolotl.monkeypatch.btlm_attn_hijack_flash import (
                     replace_btlm_attn_with_flash_attn,
@@ -706,6 +750,7 @@ class ModelLoader:
 
                 replace_btlm_attn_with_flash_attn(self.cfg.base_model)
 
+            # TODO deprecate soon
             if (
                 self.model_config.model_type == "stablelm_epoch"
                 and self.cfg.sample_packing

src/axolotl/utils/schemas/config.py

@@ -233,6 +233,7 @@ class AxolotlInputConfig(
     flash_attn_fuse_qkv: bool | None = None
     flash_attn_fuse_mlp: bool | None = None
     flash_optimum: bool | None = None
+    use_flash_attention_3: Literal["auto"] | bool | None = None
 
     eager_attention: bool | None = None
 

tests/conftest.py

@@ -421,6 +421,7 @@ def temp_dir():
 
 @pytest.fixture(scope="function", autouse=True)
 def cleanup_monkeypatches():
+    import transformers.modeling_flash_attention_utils
     from transformers import Trainer
     from transformers.models.llama.modeling_llama import (  # LlamaFlashAttention2,
         LlamaAttention,
@@ -434,6 +435,19 @@ def cleanup_monkeypatches():
         Trainer._inner_training_loop  # pylint: disable=protected-access
     )
     original_trainer_training_step = Trainer.training_step
+    original_fa_func = None
+    original_fa_varlen_func = None
+    if (
+        importlib.util.find_spec("flash_attn")
+        and hasattr(transformers.modeling_flash_attention_utils, "flash_attn_func")
+        and hasattr(
+            transformers.modeling_flash_attention_utils, "flash_attn_varlen_func"
+        )
+    ):
+        original_fa_func = transformers.modeling_flash_attention_utils.flash_attn_func
+        original_fa_varlen_func = (
+            transformers.modeling_flash_attention_utils.flash_attn_varlen_func
+        )
     # monkey patches can happen inside the tests
     yield
     # Reset LlamaFlashAttention2 forward
@@ -444,6 +458,11 @@ def cleanup_monkeypatches():
         original_trainer_inner_training_loop
     )
     Trainer.training_step = original_trainer_training_step
+    if original_fa_func:
+        transformers.modeling_flash_attention_utils.flash_attn_func = original_fa_func
+        transformers.modeling_flash_attention_utils.flash_attn_varlen_func = (
+            original_fa_varlen_func
+        )
 
     # Reset other known monkeypatches
     modules_to_reset: list[tuple[str, list[str]]] = [
@@ -458,6 +477,7 @@ def cleanup_monkeypatches():
         ("transformers.trainer",),
         ("transformers", ["Trainer"]),
         ("transformers.loss.loss_utils",),
+        ("transformers.modeling_flash_attention_utils",),
     ]
     for module_name_tuple in modules_to_reset:
         module_name = module_name_tuple[0]

tests/e2e/multigpu/solo/test_grpo.py

@@ -166,7 +166,6 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
 """
             )
 
-    @pytest.mark.skip(reason="flaky test")
     @pytest.mark.parametrize(
         "num_gpus",
         [1, 2],
@@ -231,8 +230,6 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
             "NCCL_P2P_LEVEL": "LOC",
             **current_env,
             "CUDA_VISIBLE_DEVICES": "1",
-            "VLLM_DISABLE_COMPILE_CACHE": "1",
-            # "VLLM_USE_V1": "0",
         }
         vllm_process = start_vllm(
             cfg.base_model,
@@ -266,7 +263,6 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
         finally:
             recursive_kill(vllm_process)
 
-    @pytest.mark.skip(reason="flaky test")
     @pytest.mark.parametrize(
         "num_gpus",
         [1, 2],
@@ -325,8 +321,6 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
             "NCCL_P2P_LEVEL": "LOC",  # nccl can be brittle, assume P2P isn't reliable
             **current_env,
             "CUDA_VISIBLE_DEVICES": "1",
-            "VLLM_DISABLE_COMPILE_CACHE": "1",
-            # "VLLM_USE_V1": "0",
         }
         vllm_process = start_vllm(
             cfg.base_model,

tests/e2e/multigpu/test_llama.py

@@ -101,7 +101,13 @@ class TestMultiGPULlama:
         "gradient_accumulation_steps",
         [1, 2],
     )
-    def test_lora_ddp_packed(self, temp_dir, gradient_accumulation_steps):
+    @pytest.mark.parametrize(
+        "use_flash_attention_3",
+        [False, "auto"],
+    )
+    def test_lora_ddp_packed(
+        self, temp_dir, gradient_accumulation_steps, use_flash_attention_3
+    ):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
@@ -138,6 +144,7 @@ class TestMultiGPULlama:
                 "flash_attention": True,
                 "use_tensorboard": True,
                 "bf16": True,
+                "use_flash_attention_3": use_flash_attention_3,
             }
         )
 

tests/e2e/test_packing_loss.py

@@ -4,7 +4,6 @@ E2E tests for packed training
 
 import logging
 import os
-import unittest
 
 from transformers.utils import is_torch_bf16_gpu_available
 
@@ -14,18 +13,17 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
-from .utils import check_tensorboard, with_temp_dir
+from .utils import check_tensorboard
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
 
 
-class TestPackedLlama(unittest.TestCase):
+class TestPackedLlama:
     """
     Test case for Packed training of llama models
     """
 
-    @with_temp_dir
     def test_loss_packed(self, temp_dir):
         # pylint: disable=duplicate-code
         cfg = DictDefault(