refactor build() into smaller fns

.github/workflows/main.yml

@@ -31,11 +31,6 @@ jobs:
             python_version: "3.11"
             pytorch: 2.7.0
             axolotl_extras:
-          - cuda: 128
-            cuda_version: 12.8.1
-            python_version: "3.11"
-            pytorch: 2.7.0
-            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -99,11 +94,6 @@ jobs:
             python_version: "3.11"
             pytorch: 2.7.0
             axolotl_extras:
-          - cuda: 128
-            cuda_version: 12.8.1
-            python_version: "3.11"
-            pytorch: 2.7.0
-            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout

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

@@ -3,7 +3,7 @@ name: docker-multigpu-tests-biweekly
 on:
   pull_request:
     paths:
-      - 'tests/e2e/multigpu/**.py'
+      - 'tests/e2e/multigpu/*.py'
       - 'requirements.txt'
       - 'setup.py'
       - 'pyproject.toml'

.github/workflows/tests.yml

@@ -44,102 +44,96 @@ jobs:
         env:
           SKIP: no-commit-to-branch
 
-#  preload-cache:
+  preload-cache:
-#    name: Preload HF cache
+    name: Preload HF cache
-#    runs-on: ubuntu-latest
+    runs-on: ubuntu-latest
-#    strategy:
+    strategy:
-#      fail-fast: false
+      fail-fast: false
-#      matrix:
+      matrix:
-#        python_version: ["3.11"]
+        python_version: ["3.11"]
-#        pytorch_version: ["2.6.0"]
+        pytorch_version: ["2.6.0"]
-#    timeout-minutes: 20
+    timeout-minutes: 20
-#
+
-#    env:
+    env:
-#      AXOLOTL_IS_CI_CACHE_PRELOAD: "1"
+      AXOLOTL_IS_CI_CACHE_PRELOAD: "1"
-#
+
-#    steps:
+    steps:
-#      - name: Check out repository code
+      - name: Check out repository code
-#        uses: actions/checkout@v4
+        uses: actions/checkout@v4
-#
+
-#      - name: Restore HF cache
+      - name: Restore HF cache
-#        id: hf-cache-restore
+        id: hf-cache-restore
-#        uses: actions/cache/restore@v4
+        uses: actions/cache/restore@v4
-#        with:
+        with:
-#          path: |
+          path: |
-#            /home/runner/.cache/huggingface/hub/datasets--*
+            /home/runner/.cache/huggingface/hub/datasets--*
-#            /home/runner/.cache/huggingface/hub/models--*
+            /home/runner/.cache/huggingface/hub/models--*
-#          key: ${{ runner.os }}-hf-hub-cache-v2
+          key: ${{ runner.os }}-hf-hub-cache-v2
-#
+
-#      - name: Restore Cache from S3
+      - name: Setup Python
-#        id: hf-cache-restore-s3
+        uses: actions/setup-python@v5
-#        run: |
+        with:
-#          mkdir -p /home/runner/.cache/huggingface/hub
+          python-version: ${{ matrix.python_version }}
-#          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C /home/runner/.cache/huggingface/hub/  --use-compress-program unzstd
+          cache: 'pip' # caching pip dependencies
-#
+
-#      - name: Setup Python
+      - name: upgrade pip
-#        uses: actions/setup-python@v5
+        run: |
-#        with:
+          pip3 install --upgrade pip
-#          python-version: ${{ matrix.python_version }}
+          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
-#          cache: 'pip' # caching pip dependencies
+
-#
+      - name: Install PyTorch
-#      - name: upgrade pip
+        run: |
-#        run: |
+          pip3 install torch==${{ matrix.pytorch_version }}
-#          pip3 install --upgrade pip
+
-#          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
+      - name: Install dependencies
-#
+        run: |
-#      - name: Install PyTorch
+          pip3 show torch
-#        run: |
+          pip3 install --no-build-isolation -U -e .
-#          pip3 install torch==${{ matrix.pytorch_version }}
+          python scripts/unsloth_install.py | sh
-#
+          python scripts/cutcrossentropy_install.py | sh
-#      - name: Install dependencies
+          pip3 install -r requirements-dev.txt -r requirements-tests.txt
-#        run: |
+
-#          pip3 show torch
+      - name: Make sure PyTorch version wasn't clobbered
-#          pip3 install --no-build-isolation -U -e .
+        run: |
-#          python scripts/unsloth_install.py | sh
+          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
-#          python scripts/cutcrossentropy_install.py | sh
+
-#          pip3 install -r requirements-dev.txt -r requirements-tests.txt
+      - name: Ensure axolotl CLI was installed
-#
+        run: |
-#      - name: Make sure PyTorch version wasn't clobbered
+          axolotl --help
-#        run: |
+
-#          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
+      - name: Pre-Download dataset fixture
-#
+        run: |
-#      - name: Ensure axolotl CLI was installed
+          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
-#        run: |
+
-#          axolotl --help
+      - name: Run tests
-#
+        run: |
-#      - name: Pre-Download dataset fixture
+          pytest -v tests/conftest.py
-#        run: |
+
-#          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
+      - name: Upload coverage to Codecov
-#
+        uses: codecov/codecov-action@v5
-#      - name: Run tests
+        with:
-#        run: |
+          token: ${{ secrets.CODECOV_TOKEN }}
-#          pytest -v tests/conftest.py
+          files: ./coverage.xml
-#
+          flags: unittests,pytorch-${{ matrix.pytorch_version }}
-#      - name: Upload coverage to Codecov
+          fail_ci_if_error: false
-#        uses: codecov/codecov-action@v5
+
-#        with:
+      - name: cleanup pip cache
-#          token: ${{ secrets.CODECOV_TOKEN }}
+        run: |
-#          files: ./coverage.xml
+          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
-#          flags: unittests,pytorch-${{ matrix.pytorch_version }}
+
-#          fail_ci_if_error: false
+      - name: Save HF cache
-#
+        id: hf-cache
-#      - name: cleanup pip cache
+        uses: actions/cache/save@v4
-#        run: |
+        with:
-#          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
+          path: |
-#
+            /home/runner/.cache/huggingface/hub/datasets--*
-#      - name: Save HF cache
+            /home/runner/.cache/huggingface/hub/models--*
-#        id: hf-cache
+          key: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
-#        uses: actions/cache/save@v4
-#        with:
-#          path: |
-#            /home/runner/.cache/huggingface/hub/datasets--*
-#            /home/runner/.cache/huggingface/hub/models--*
-#          key: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
 
   pytest:
     name: PyTest
     runs-on: ubuntu-latest
-#    needs: [preload-cache]
+    needs: [preload-cache]
     strategy:
       fail-fast: false
       matrix:
@@ -151,20 +145,14 @@ jobs:
       - name: Check out repository code
         uses: actions/checkout@v4
 
-#      - name: Restore HF cache
+      - name: Restore HF cache
-#        id: hf-cache-restore
+        id: hf-cache-restore
-#        uses: actions/cache/restore@v4
+        uses: actions/cache/restore@v4
-#        with:
+        with:
-#          path: |
+          path: |
-#            /home/runner/.cache/huggingface/hub/datasets--*
+            /home/runner/.cache/huggingface/hub/datasets--*
-#            /home/runner/.cache/huggingface/hub/models--*
+            /home/runner/.cache/huggingface/hub/models--*
-#          key: ${{ runner.os }}-hf-hub-cache-v2
+          key: ${{ runner.os }}-hf-hub-cache-v2
-
-      - name: Restore Cache from S3
-        id: hf-cache-restore-s3
-        run: |
-          mkdir -p /home/runner/.cache/huggingface/hub
-          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C /home/runner/.cache/huggingface/hub/  --use-compress-program unzstd
 
       - name: Setup Python
         uses: actions/setup-python@v5
@@ -222,7 +210,7 @@ jobs:
   pytest-sdist:
     name: PyTest from Source Dist
     runs-on: ubuntu-latest
-#    needs: [preload-cache]
+    needs: [preload-cache]
     strategy:
       fail-fast: false
       matrix:
@@ -234,20 +222,14 @@ jobs:
       - name: Check out repository code
         uses: actions/checkout@v4
 
-#      - name: Restore HF cache
+      - name: Restore HF cache
-#        id: hf-cache-restore
+        id: hf-cache-restore
-#        uses: actions/cache/restore@v4
+        uses: actions/cache/restore@v4
-#        with:
+        with:
-#          path: |
+          path: |
-#            /home/runner/.cache/huggingface/hub/datasets--*
+            /home/runner/.cache/huggingface/hub/datasets--*
-#            /home/runner/.cache/huggingface/hub/models--*
+            /home/runner/.cache/huggingface/hub/models--*
-#          key: ${{ runner.os }}-hf-hub-cache-v2
+          key: ${{ runner.os }}-hf-hub-cache-v2
-
-      - name: Restore Cache from S3
-        id: hf-cache-restore-s3
-        run: |
-          mkdir -p /home/runner/.cache/huggingface/hub
-          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C /home/runner/.cache/huggingface/hub/  --use-compress-program unzstd
 
       - name: Setup Python
         uses: actions/setup-python@v5
@@ -295,7 +277,6 @@ jobs:
           find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
 
   docker-e2e-tests-1st:
-    # Run this job first as a gate for running the remainder of the test matrix
     if: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' }}
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
@@ -342,8 +323,6 @@ jobs:
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
     timeout-minutes: 90
-    # Only run the remainder of the matrix if the first e2e check passed;
-    # this is to save on wasted compute costs for known failures that get caught in the first run
     needs: [pre-commit, pytest, docker-e2e-tests-1st]
 
     strategy:
@@ -368,12 +347,6 @@ jobs:
             pytorch: 2.7.0
             num_gpus: 1
             axolotl_extras:
-          - cuda: 128
-            cuda_version: 12.8.1
-            python_version: "3.11"
-            pytorch: 2.7.0
-            num_gpus: 1
-            axolotl_extras:
     steps:
       - name: Checkout
         uses: actions/checkout@v4

.runpod/src/handler.py

@@ -57,9 +57,7 @@ async def handler(job):
     logger.info("Training Complete.")
 
     # Cleanup
-    if "WANDB_API_KEY" in os.environ:
     del os.environ["WANDB_API_KEY"]
-    if "HF_TOKEN" in os.environ:
     del os.environ["HF_TOKEN"]
 
 

_quarto.yml

@@ -48,23 +48,8 @@ quartodoc:
       contents:
         - core.trainers.base
         - core.trainers.trl
-        - core.trainers.mamba
-        - core.trainers.relora
         - core.trainers.dpo.trainer
         - core.trainers.grpo.trainer
-        - core.trainers.grpo.sampler
-        - core.trainers.utils
-    - title: Mixins
-      desc: Mixin classes for augmenting trainers
-      contents:
-        - core.trainers.mixins.optimizer
-        - core.trainers.mixins.rng_state_loader
-        - core.trainers.mixins.scheduler
-        - core.trainers.mixins.sequence_parallel
-    - title: Context Managers
-      desc: Context managers for altering trainer behaviors
-      contents:
-        - utils.ctx_managers.sequence_parallel
     - title: Prompt Strategies
       desc: Prompt formatting strategies
       contents:
@@ -101,7 +86,7 @@ quartodoc:
         - kernels.swiglu
         - kernels.quantize
         - kernels.utils
-    - title: Monkey Patches
+    - title: MonkeyPatches
       desc: Runtime patches for model optimizations
       contents:
         - monkeypatch.llama_attn_hijack_flash
@@ -139,8 +124,7 @@ quartodoc:
         - utils.optimizers.adopt
         - utils.data.pretraining
         - utils.data.sft
-        - utils.gradient_checkpointing.offload_cpu
+        - utils.gradient_checkpointing.unsloth
-        - utils.gradient_checkpointing.offload_disk
     - title: Schemas
       desc: Pydantic data models for Axolotl config
       contents:

cicd/e2e_tests.py

@@ -6,7 +6,7 @@ from .single_gpu import GPU_CONFIG, VOLUME_CONFIG, app, cicd_image, run_cmd
 @app.function(
     image=cicd_image,
     gpu=GPU_CONFIG,
-    timeout=90 * 60,  # 90 min
+    timeout=60 * 60,
     cpu=8.0,
     memory=131072,
     volumes=VOLUME_CONFIG,

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=16.0,
+    cpu=8.0,
     memory=131072 * N_GPUS,
     volumes=VOLUME_CONFIG,
 )

codecov.yml

@@ -19,7 +19,7 @@ coverage:
         if_no_uploads: error
         if_not_found: success
         if_ci_failed: error
-        only_pulls: true
+        only_pulls: false
         flags: null
         paths: null
     patch:

docs/config.qmd

@@ -505,7 +505,6 @@ save_strategy: # Set to `"no"` to skip checkpoint saves, `"epoch"` at end of eac
 save_steps: # Leave empty to save at each epoch, integer for every N steps. float for fraction of total steps
 saves_per_epoch: # number of times per epoch to save a checkpoint, mutually exclusive with save_steps
 save_total_limit: # Checkpoints saved at a time
-save_only_model: # Save only the model weights, skipping the optimizer. Using this means you can't resume from checkpoints.
 # Maximum number of iterations to train for. It precedes num_epochs which means that
 # if both are set, num_epochs will not be guaranteed.
 # e.g., when 1 epoch is 1000 steps => `num_epochs: 2` and `max_steps: 100` will train for 100 steps
@@ -539,7 +538,7 @@ train_on_inputs: false
 # Note that training loss may have an oscillating pattern with this enabled.
 group_by_length: false
 
-# Whether to use gradient checkpointing. Available options are: true, false, "offload", "offload_disk".
+# Whether to use gradient checkpointing. Available options are: true, false, "offload".
 # https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing
 gradient_checkpointing: false
 # additional kwargs to pass to the trainer for gradient checkpointing
@@ -633,9 +632,7 @@ 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:
 

docs/getting-started.qmd

@@ -104,7 +104,7 @@ the `alpaca` dataset format, which has the following format:
 Please see our [Dataset Formats](dataset-formats) for more dataset formats and how to
 format them.
 
-2. Prepare your JSONL data in the specified format (in this case, the expected `alpaca`
+2. Prepare your JSONL data in the specified format (in this case, the expected `alpaca
 format):
 
 ```json
@@ -120,12 +120,6 @@ axolotl train my_training.yml
 
 ## Common Tasks {#sec-common-tasks}
 
-::: {.callout-tip}
-
-The same yaml file is used for training, inference, and merging.
-
-:::
-
 ### Testing Your Model {#sec-testing}
 
 After training, test your model:
@@ -134,16 +128,6 @@ After training, test your model:
 axolotl inference my_training.yml --lora-model-dir="./outputs/lora-out"
 ```
 
-More details can be found in [Inference](inference.qmd).
-
-### Using a UI {#sec-ui}
-
-Launch a Gradio interface:
-
-```bash
-axolotl inference my_training.yml --lora-model-dir="./outputs/lora-out" --gradio
-```
-
 ### Preprocessing Data {#sec-preprocessing}
 
 For large datasets, preprocess first:
@@ -152,22 +136,14 @@ For large datasets, preprocess first:
 axolotl preprocess my_training.yml
 ```
 
-Please make sure to set `dataset_prepared_path: ` in your config to set the path to save the prepared dataset.
+### Using a UI {#sec-ui}
 
-More details can be found in [Dataset Preprocessing](dataset_preprocessing.qmd).
+Launch a Gradio interface:
-
-### Merging LoRA weights {#sec-merging-lora}
-
-To merge the LoRA weights back into the base model, run:
 
 ```bash
-axolotl merge-lora my_training.yml --lora-model-dir="./outputs/lora-out"
+axolotl inference my_training.yml --lora-model-dir="./outputs/lora-out" --gradio
 ```
 
-The merged model will be saved in the `{output_dir}/merged` directory.
-
-More details can be found in [Merging LoRA weights](inference.qmd#sec-merging).
-
 ## Next Steps {#sec-next-steps}
 
 Now that you have the basics, you might want to:
@@ -180,7 +156,6 @@ Now that you have the basics, you might want to:
 Check our other guides for details on these topics:
 
 - [Configuration Guide](config.qmd) - Full configuration options
-- [Dataset Loading](dataset-loading.qmd) - Loading datasets from various sources
 - [Dataset Formats](dataset-formats) - Working with different data formats
 - [Multi-GPU Training](multi-gpu.qmd)
 - [Multi-Node Training](multi-node.qmd)

docs/sequence_parallelism.qmd

@@ -3,6 +3,8 @@ title: Sequence Parallelism
 description: Train with long sequences split across multiple GPUs.
 ---
 
+# Sequence Parallelism
+
 Sequence parallelism is a technique that splits sequences across multiple GPUs,
 allowing you to train with very long sequences that wouldn't fit on a single GPU. Each
 GPU processes a different portion of the sequence, and the results are aggregated
@@ -25,7 +27,7 @@ To enable sequence parallelism, add the following to your configuration file:
 sequence_parallel_degree: 4  # Split sequences across 4 GPUs
 # Optional; strides across the key dimension. Larger values use more memory but should make training faster.
 heads_k_stride: 1
-# Optional; one of "varlen_llama3" or "batch_ring". Defaults to
+# Optional; one of "varlen_llama3", "batch_ring", "batch_zigzag", "batch_stripe". Defaults to
 # "varlen_llama3" when `sample_packing: true`, and "batch_ring" otherwise.
 ring_attn_func:
 ```

src/axolotl/cli/args.py

@@ -82,12 +82,6 @@ class VllmServeCliArgs:
             "hardware support this feature."
         },
     )
-    serve_module: Optional[str] = field(
-        default=None,
-        metadata={
-            "help": "Module to serve. If not set, the default module will be used."
-        },
-    )
 
 
 @dataclass

src/axolotl/cli/vllm_serve.py

@@ -6,6 +6,7 @@ from pathlib import Path
 from typing import Union
 
 from trl.scripts.vllm_serve import ScriptArguments
+from trl.scripts.vllm_serve import main as vllm_serve_main
 
 from axolotl.cli.config import load_cfg
 
@@ -27,9 +28,6 @@ def do_vllm_serve(
     cfg = load_cfg(config)
     model = cfg.base_model
 
-    serve_module = cli_args.get("serve_module", "trl.scripts.vllm_serve")
-    vllm_serve_main = getattr(__import__(serve_module, fromlist=["main"]), "main")
-
     tensor_parallel_size = (
         cli_args.get("tensor_parallel_size") or cfg.vllm.tensor_parallel_size
     )

src/axolotl/common/datasets.py

@@ -14,7 +14,6 @@ from axolotl.utils.data import prepare_dataset
 from axolotl.utils.data.rl import load_prepare_preference_datasets
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.models import load_processor, load_tokenizer
-from axolotl.utils.schemas.enums import RLType
 from axolotl.utils.tokenization import check_dataset_labels
 
 LOG = logging.getLogger(__name__)
@@ -134,7 +133,7 @@ def load_preference_datasets(
     total_num_steps: Optional[int] = int(
         math.ceil(len(train_dataset) * cfg.num_epochs / cfg.batch_size)
     )
-    if cfg.rl is RLType.GRPO:
+    if cfg.rl == "grpo":
         total_num_steps = None
 
     if cli_args.debug or cfg.debug:

src/axolotl/core/trainers/__init__.py

@@ -5,7 +5,7 @@
 
 from .base import AxolotlTrainer
 from .dpo.trainer import AxolotlDPOTrainer
-from .grpo.trainer import AxolotlGRPOSequenceParallelTrainer, AxolotlGRPOTrainer
+from .grpo.trainer import AxolotlGRPOTrainer
 from .mamba import AxolotlMambaTrainer
 from .relora import ReLoRATrainer
 from .trl import (

src/axolotl/core/trainers/base.py

@@ -373,13 +373,15 @@ class AxolotlTrainer(
                 num_items_in_batch=num_items_in_batch,
             )
 
-        return super().compute_loss(
+        loss = super().compute_loss(
             model,
             inputs,
             return_outputs=return_outputs,
             num_items_in_batch=num_items_in_batch,
         )
 
+        return loss
+
     @staticmethod
     def orpo_concatenate_inputs(inputs, label_pad_token=-100, pad_token=0, device=None):
         concatenated_batch = {}

src/axolotl/core/trainers/builders/__init__.py

@@ -0,0 +1,21 @@
+# Copyright 2024 Axolotl AI. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Init for axolotl.core.trainers.builders"""
+
+# pylint: disable=unused-import
+# flake8: noqa
+
+from .causal import HFCausalTrainerBuilder
+from .rl import HFRLTrainerBuilder

src/axolotl/core/trainers/builders/base.py

@@ -0,0 +1,331 @@
+"""Base class trainer / training args builder implementation"""
+
+import abc
+from typing import Any
+
+from torch import Type
+from transformers import TrainerCallback
+from transformers.training_args import TrainingArguments
+
+from axolotl.integrations.base import PluginManager
+from axolotl.monkeypatch.trainer.lr import patch_trainer_get_lr
+from axolotl.utils import is_comet_available, is_mlflow_available
+from axolotl.utils.callbacks import GCCallback, SaveAxolotlConfigtoWandBCallback
+from axolotl.utils.callbacks.profiler import PytorchProfilerCallback
+
+PLUGIN_MANAGER = PluginManager.get_instance()
+
+
+class TrainerBuilderBase(abc.ABC):
+    """Base class for trainer builder."""
+
+    _train_dataset = None
+    _eval_dataset = None
+    _model_ref = None
+    _peft_config = None
+
+    def __init__(self, cfg, model, tokenizer, processor=None):
+        self.cfg = cfg
+        self.model = model
+        self.tokenizer = tokenizer
+        self.processor = processor
+
+        # If the model supports tagging, add the axolotl tag.
+        # This makes sure the tag is correctly pushed even if a user calls
+        # model.push_to_hub instead of trainer.push_to_hub.
+        if hasattr(model, "add_model_tags"):
+            model.add_model_tags(["axolotl"])
+
+        patch_trainer_get_lr()
+
+    @property
+    def model_ref(self):
+        return self._model_ref
+
+    @model_ref.setter
+    def model_ref(self, model):
+        self._model_ref = model
+
+    @property
+    def train_dataset(self):
+        return self._train_dataset
+
+    @train_dataset.setter
+    def train_dataset(self, dataset):
+        self._train_dataset = dataset
+
+    @property
+    def eval_dataset(self):
+        return self._eval_dataset
+
+    @eval_dataset.setter
+    def eval_dataset(self, dataset):
+        self._eval_dataset = dataset
+
+    @property
+    def peft_config(self):
+        return self._peft_config
+
+    @peft_config.setter
+    def peft_config(self, peft_config):
+        self._peft_config = peft_config
+
+    @abc.abstractmethod
+    def build(self, total_num_steps):
+        pass
+
+    def get_common_training_args_kwargs(
+        self, total_num_steps: int | None = None
+    ) -> dict[str, Any]:
+        """Get common training arguments kwargs used across different trainer types."""
+        training_args_kwargs = {}
+
+        # Common parameters
+        for arg in [
+            "adam_beta1",
+            "adam_beta2",
+            "adam_epsilon",
+            "max_grad_norm",
+            "dataloader_num_workers",
+            "dataloader_pin_memory",
+            "dataloader_prefetch_factor",
+            "dataloader_drop_last",
+            "remove_unused_columns",
+        ]:
+            if hasattr(self.cfg, arg) and getattr(self.cfg, arg) is not None:
+                training_args_kwargs[arg] = getattr(self.cfg, arg)
+
+        # Add Hub integration arguments if needed
+        if self.cfg.hub_model_id:
+            training_args_kwargs["hub_model_id"] = self.cfg.hub_model_id
+            training_args_kwargs["push_to_hub"] = True
+            training_args_kwargs["hub_private_repo"] = True
+            training_args_kwargs["hub_always_push"] = True
+
+            if self.cfg.hub_strategy:
+                training_args_kwargs["hub_strategy"] = self.cfg.hub_strategy
+
+        # BF16/FP16 settings
+        if hasattr(self.cfg, "bf16") and self.cfg.bf16:
+            if self.cfg.bf16 == "full":
+                training_args_kwargs["bf16_full_eval"] = True
+            else:
+                training_args_kwargs["bf16"] = self.cfg.bf16
+        elif hasattr(self.cfg, "bfloat16") and self.cfg.bfloat16:
+            training_args_kwargs["bf16"] = True
+
+        if hasattr(self.cfg, "fp16"):
+            training_args_kwargs["fp16"] = (
+                getattr(self.cfg, "fp16", False)
+                and not getattr(self.cfg, "bf16", False)
+            ) or False
+
+        # Set save_strategy and save_steps
+        if self.cfg.save_steps:
+            training_args_kwargs["save_strategy"] = "steps"
+            training_args_kwargs["save_steps"] = self.cfg.save_steps
+        elif self.cfg.save_strategy:
+            training_args_kwargs["save_strategy"] = self.cfg.save_strategy
+        else:
+            # default to saving each epoch if not defined
+            training_args_kwargs["save_strategy"] = "epoch"
+
+        # Handle safetensors
+        if self.cfg.save_safetensors is not None:
+            training_args_kwargs["save_safetensors"] = self.cfg.save_safetensors
+
+        # Handle gradient checkpointing
+        if self.cfg.gradient_checkpointing:
+            training_args_kwargs["gradient_checkpointing"] = (
+                self.cfg.gradient_checkpointing
+            )
+            if self.cfg.gradient_checkpointing_kwargs is not None:
+                training_args_kwargs["gradient_checkpointing_kwargs"] = (
+                    self.cfg.gradient_checkpointing_kwargs
+                )
+
+        # Common optimizer and LR scheduler settings
+        training_args_kwargs["optim"] = self.cfg.optimizer
+        if hasattr(self.cfg, "lr_scheduler") and self.cfg.lr_scheduler:
+            training_args_kwargs["lr_scheduler_type"] = self.cfg.lr_scheduler
+        else:
+            training_args_kwargs["lr_scheduler_type"] = "cosine"
+
+        if hasattr(self.cfg, "lr_scheduler_kwargs") and self.cfg.lr_scheduler_kwargs:
+            training_args_kwargs["lr_scheduler_kwargs"] = self.cfg.lr_scheduler_kwargs
+        else:
+            training_args_kwargs["lr_scheduler_kwargs"] = {}
+
+        # LoRA+ specific settings
+        if hasattr(self.cfg, "loraplus_lr_ratio"):
+            training_args_kwargs["loraplus_lr_ratio"] = self.cfg.loraplus_lr_ratio
+        if hasattr(self.cfg, "loraplus_lr_embedding"):
+            training_args_kwargs["loraplus_lr_embedding"] = (
+                self.cfg.loraplus_lr_embedding
+            )
+
+        # Reporting tools
+        report_to = []
+        if self.cfg.use_wandb:
+            report_to.append("wandb")
+            if self.cfg.wandb_name:
+                training_args_kwargs["run_name"] = self.cfg.wandb_name
+        if self.cfg.use_mlflow:
+            report_to.append("mlflow")
+        if self.cfg.use_tensorboard:
+            report_to.append("tensorboard")
+        if self.cfg.use_comet:
+            report_to.append("comet_ml")
+
+        if report_to:
+            training_args_kwargs["report_to"] = report_to
+
+        # Basic training settings
+        if hasattr(self.cfg, "sequence_len"):
+            training_args_kwargs["max_length"] = self.cfg.sequence_len
+
+        training_args_kwargs["save_only_model"] = getattr(
+            self.cfg, "save_only_model", False
+        )
+        training_args_kwargs["save_total_limit"] = getattr(
+            self.cfg, "save_total_limit", 5
+        )
+
+        # Compute warmup steps
+        if hasattr(self.cfg, "warmup_steps") and self.cfg.warmup_steps is not None:
+            training_args_kwargs["warmup_steps"] = self.cfg.warmup_steps
+        elif (
+            total_num_steps
+            and hasattr(self.cfg, "warmup_ratio")
+            and self.cfg.warmup_ratio is not None
+        ):
+            training_args_kwargs["warmup_steps"] = max(
+                int(self.cfg.warmup_ratio * total_num_steps), 0
+            )
+        elif total_num_steps:
+            training_args_kwargs["warmup_steps"] = min(int(0.03 * total_num_steps), 100)
+
+        return training_args_kwargs
+
+    def create_training_args(
+        self,
+        args_cls: Type[TrainingArguments],
+        total_num_steps: int | None = None,
+        **additional_kwargs,
+    ) -> TrainingArguments:
+        """Create training arguments with common logic."""
+        # Get common trainings args and update with trainer-specific args
+        training_args_kwargs = self.get_common_training_args_kwargs(total_num_steps)
+        training_args_kwargs.update(additional_kwargs)
+
+        # Create training args with pre- and post-creation hooks
+        training_args_kwargs = self.hook_pre_create_training_args(training_args_kwargs)
+        training_args = args_cls(**training_args_kwargs)
+        training_args = self.hook_post_create_training_args(training_args)
+
+        # Unset run_name so wandb sets up experiment names properly
+        if self.cfg.use_wandb and training_args.run_name == training_args.output_dir:
+            training_args.run_name = None
+
+        return training_args
+
+    def create_trainer(
+        self, trainer_cls, training_args, trainer_args=None, trainer_kwargs=None
+    ):
+        """Create trainer with common logic."""
+        if trainer_args is None:
+            trainer_args = []
+        if trainer_kwargs is None:
+            trainer_kwargs = {}
+
+        # Create trainer with pre- and post- creation hooks
+        trainer_kwargs, trainer_cls = self.hook_pre_create_trainer(
+            trainer_kwargs, trainer_cls
+        )
+        trainer = trainer_cls(
+            *trainer_args,
+            args=training_args,
+            train_dataset=self.train_dataset,
+            eval_dataset=self.eval_dataset,
+            callbacks=self.get_callbacks(),
+            **trainer_kwargs,
+        )
+        trainer = self.hook_post_create_trainer(trainer)
+
+        # Add post-creation callbacks
+        for callback in self.get_post_trainer_create_callbacks(trainer):
+            trainer.add_callback(callback)
+
+        return trainer
+
+    def get_callbacks(self) -> list[TrainerCallback]:
+        callbacks = []
+        callbacks.extend(
+            PLUGIN_MANAGER.add_callbacks_pre_trainer(cfg=self.cfg, model=self.model)
+        )
+
+        if self.cfg.profiler_steps:
+            callbacks.append(
+                PytorchProfilerCallback(
+                    steps_to_profile=self.cfg.profiler_steps,
+                )
+            )
+
+        if self.cfg.gc_steps:
+            callbacks.append(GCCallback(gc_steps=self.cfg.gc_steps))
+
+        if self.cfg.use_wandb:
+            callbacks.append(
+                SaveAxolotlConfigtoWandBCallback(self.cfg.axolotl_config_path)
+            )
+        if self.cfg.use_mlflow and is_mlflow_available():
+            from axolotl.utils.callbacks.mlflow_ import (
+                SaveAxolotlConfigtoMlflowCallback,
+            )
+
+            callbacks.extend(
+                [
+                    SaveAxolotlConfigtoMlflowCallback(self.cfg.axolotl_config_path),
+                ]
+            )
+        if self.cfg.use_comet and is_comet_available():
+            from axolotl.utils.callbacks.comet_ import SaveAxolotlConfigtoCometCallback
+
+            callbacks.append(
+                SaveAxolotlConfigtoCometCallback(self.cfg.axolotl_config_path)
+            )
+
+        return callbacks
+
+    def get_post_trainer_create_callbacks(self, trainer):
+        """Callbacks added after the trainer is created, usually because these need
+        access to the trainer.
+        """
+        callbacks = []
+        if self.cfg.plugins:
+            callbacks.extend(
+                [
+                    cb
+                    for cb in PLUGIN_MANAGER.add_callbacks_post_trainer(
+                        self.cfg, trainer
+                    )
+                    if cb
+                ]
+            )
+        return callbacks
+
+    def hook_pre_create_training_args(self, training_arguments_kwargs):
+        # TODO
+        return training_arguments_kwargs
+
+    def hook_post_create_training_args(self, training_arguments):
+        # TODO
+        return training_arguments
+
+    def hook_pre_create_trainer(self, trainer_kwargs, trainer_cls):
+        # TODO
+        return trainer_kwargs, trainer_cls
+
+    def hook_post_create_trainer(self, trainer):
+        # TODO
+        return trainer

src/axolotl/core/trainers/builders/causal.py

@@ -0,0 +1,619 @@
+"""Causal trainer / training args builder implementation"""
+
+import importlib
+import inspect
+import logging
+import math
+import os
+import sys
+from pathlib import Path
+from typing import Type
+
+import transformers
+from transformers import (
+    DataCollatorWithFlattening,
+    EarlyStoppingCallback,
+)
+from transformers.training_args import OptimizerNames
+from trl.trainer.utils import RewardDataCollatorWithPadding
+
+from axolotl.core.trainers.base import AxolotlTrainer
+from axolotl.core.trainers.builders.base import TrainerBuilderBase
+from axolotl.core.trainers.mamba import AxolotlMambaTrainer
+from axolotl.core.trainers.relora import ReLoRATrainer
+from axolotl.core.trainers.trl import AxolotlPRMTrainer, AxolotlRewardTrainer
+from axolotl.core.training_args import (
+    AxolotlPRMConfig,
+    AxolotlRewardConfig,
+    AxolotlTrainingArguments,
+)
+from axolotl.integrations.base import PluginManager
+from axolotl.monkeypatch.multipack import SUPPORTED_MULTIPACK_MODEL_TYPES
+from axolotl.monkeypatch.relora import ReLoRACallback
+from axolotl.processing_strategies import get_processing_strategy
+from axolotl.utils import is_comet_available, is_mlflow_available
+from axolotl.utils.callbacks import (
+    EvalFirstStepCallback,
+    GPUStatsCallback,
+    LossWatchDogCallback,
+    SaveBetterTransformerModelCallback,
+    bench_eval_callback_factory,
+    causal_lm_bench_eval_callback_factory,
+    colab_inference_post_train_callback,
+    log_prediction_callback_factory,
+)
+from axolotl.utils.callbacks.lisa import lisa_callback_factory
+from axolotl.utils.chat_templates import get_chat_template_from_config
+from axolotl.utils.collators.batching import (
+    BatchSamplerDataCollatorForSeq2Seq,
+    DataCollatorForSeq2Seq,
+    V2BatchSamplerDataCollatorForSeq2Seq,
+)
+from axolotl.utils.collators.mamba import MambaDataCollator
+from axolotl.utils.collators.mm_chat import MultiModalChatDataCollator
+from axolotl.utils.schemas.enums import CustomSupportedOptimizers
+
+LOG = logging.getLogger(__name__)
+PLUGIN_MANAGER = PluginManager.get_instance()
+
+
+class HFCausalTrainerBuilder(TrainerBuilderBase):
+    """Build the HuggingFace training args / trainer for causal models and reward
+    modeling using TRL.
+    """
+
+    def get_callbacks(self):
+        callbacks = super().get_callbacks()
+        callbacks.append(GPUStatsCallback(self.cfg))
+        callbacks.append(EvalFirstStepCallback())
+
+        if self.cfg.relora_steps:
+            callbacks.append(ReLoRACallback(self.cfg))
+
+        if (
+            hasattr(self.model, "use_bettertransformer")
+            and self.model.use_bettertransformer is True
+        ):
+            callbacks.append(SaveBetterTransformerModelCallback())
+
+        if self.cfg.loss_watchdog_threshold is not None:
+            callbacks.append(LossWatchDogCallback(self.cfg))
+
+        return callbacks
+
+    def get_post_trainer_create_callbacks(self, trainer):
+        callbacks = []
+        if self.cfg.use_wandb and self.cfg.eval_table_size > 0:
+            LogPredictionCallback = log_prediction_callback_factory(
+                trainer, self.tokenizer, "wandb"
+            )
+            callbacks.append(LogPredictionCallback(self.cfg))
+        if (
+            self.cfg.use_mlflow
+            and is_mlflow_available()
+            and self.cfg.eval_table_size > 0
+        ):
+            LogPredictionCallback = log_prediction_callback_factory(
+                trainer, self.tokenizer, "mlflow"
+            )
+            callbacks.append(LogPredictionCallback(self.cfg))
+        if self.cfg.use_comet and is_comet_available() and self.cfg.eval_table_size > 0:
+            LogPredictionCallback = log_prediction_callback_factory(
+                trainer, self.tokenizer, "comet_ml"
+            )
+            callbacks.append(LogPredictionCallback(self.cfg))
+
+        if self.cfg.do_bench_eval:
+            callbacks.append(bench_eval_callback_factory(trainer, self.tokenizer))
+        if self.cfg.do_causal_lm_eval:
+            CausalLMBenchEvalCallback = causal_lm_bench_eval_callback_factory(
+                trainer, self.tokenizer
+            )
+            callbacks.append(CausalLMBenchEvalCallback(self.cfg))
+
+        if self.cfg.early_stopping_patience:
+            early_stop_cb = EarlyStoppingCallback(
+                self.cfg.early_stopping_patience,
+            )
+            callbacks.append(early_stop_cb)
+
+        if self.cfg.lisa_step_interval and self.cfg.lisa_n_layers:
+            callbacks.append(lisa_callback_factory(trainer))
+
+        if any("COLAB_" in key for key in os.environ):
+            ColabCallback = colab_inference_post_train_callback(trainer)
+            callbacks.append(ColabCallback(self.cfg))
+
+        callbacks.extend(super().get_post_trainer_create_callbacks(trainer=trainer))
+        return callbacks
+
+    def _get_trainer_cls(self):
+        if self.cfg.plugins:
+            trainer_cls = PLUGIN_MANAGER.get_trainer_cls(self.cfg)
+            if trainer_cls:
+                return trainer_cls
+        if self.cfg.relora_steps:
+            return ReLoRATrainer
+        if self.cfg.model_config_type == "mamba":
+            return AxolotlMambaTrainer
+        if self.cfg.reward_model:
+            return AxolotlRewardTrainer
+        if self.cfg.process_reward_model:
+            return AxolotlPRMTrainer
+
+        return AxolotlTrainer
+
+    def build(self, total_num_steps):
+        """Build and return a causal trainer instance using the refactored base class."""
+        # Get trainer class
+        trainer_cls = self._get_trainer_cls()
+
+        # Prepare training arguments
+        training_args = self._prepare_training_args(total_num_steps)
+
+        # Prepare data collators
+        data_collator_kwargs = self._prepare_data_collator_kwargs()
+
+        # Prepare trainer kwargs
+        trainer_kwargs = self._prepare_trainer_kwargs(
+            trainer_cls=trainer_cls,
+            data_collator_kwargs=data_collator_kwargs,
+            training_args=training_args,
+        )
+
+        # Create the trainer
+        trainer = self.create_trainer(
+            trainer_cls=trainer_cls,
+            training_args=training_args,
+            trainer_kwargs={
+                "model": self.model,
+                "data_collator": self.build_collator(
+                    training_args, **data_collator_kwargs
+                ),
+                **trainer_kwargs,
+            },
+        )
+
+        # Handle DeepSpeed config for sample packing if needed
+        if self.cfg.deepspeed and self.cfg.sample_packing:
+            trainer.accelerator.state.deepspeed_plugin.deepspeed_config[
+                "train_micro_batch_size_per_gpu"
+            ] = self.cfg.micro_batch_size
+
+        return trainer
+
+    def _prepare_training_args(self, total_num_steps):
+        """Prepare and return training arguments."""
+        # Base training arguments
+        training_args_kwargs = self._get_base_training_args()
+
+        # Add feature configurations
+        self._add_feature_configs(training_args_kwargs)
+
+        # Handle optimizer configuration
+        self._configure_optimizer(training_args_kwargs)
+
+        # Create training args using the base class method
+        training_args_cls = self._get_training_args_cls()
+
+        return self.create_training_args(
+            args_cls=training_args_cls,
+            total_num_steps=total_num_steps,
+            **training_args_kwargs,
+        )
+
+    def _get_base_training_args(self):
+        """Return the base training arguments."""
+        return {
+            "max_steps": self.cfg.max_steps if self.cfg.max_steps else -1,
+            "max_seq_length": self.cfg.sequence_len,
+            "per_device_train_batch_size": self.cfg.micro_batch_size,
+            "gradient_accumulation_steps": self.cfg.gradient_accumulation_steps,
+            "eval_accumulation_steps": self.cfg.gradient_accumulation_steps,
+            "num_train_epochs": self.cfg.num_epochs,
+            "learning_rate": self.cfg.learning_rate,
+            "output_dir": self.cfg.output_dir,
+            "weight_decay": (
+                self.cfg.weight_decay if self.cfg.weight_decay is not None else 0.0
+            ),
+            "model_type": self.cfg.model_config_type,
+            "pretraining": bool(self.cfg.pretraining_dataset),
+            "sequence_parallel_degree": self.cfg.sequence_parallel_degree,
+            "ring_attn_func": self.cfg.ring_attn_func,
+            "embedding_lr": self.cfg.embedding_lr,
+            "embedding_lr_scale": self.cfg.embedding_lr_scale,
+            "loraplus_lr_ratio": self.cfg.loraplus_lr_ratio,
+            "loraplus_lr_embedding": self.cfg.loraplus_lr_embedding,
+            "lr_groups": self.cfg.lr_groups,
+        }
+
+    def _add_feature_configs(self, training_args_kwargs):
+        """Add various feature configurations."""
+        # Sample packing configurations
+        self._add_sample_packing_configs(training_args_kwargs)
+
+        # Batch size configurations
+        if self.cfg.eval_batch_size:
+            training_args_kwargs["per_device_eval_batch_size"] = (
+                self.cfg.eval_batch_size
+            )
+        if self.cfg.auto_find_batch_size is not None:
+            training_args_kwargs["auto_find_batch_size"] = self.cfg.auto_find_batch_size
+
+        # Advanced training techniques (ReLoRA & Lisa)
+        self._add_advanced_training_configs(training_args_kwargs)
+
+        # Model-specific configurations
+        self._add_model_specific_configs(training_args_kwargs)
+
+    def _add_sample_packing_configs(self, training_args_kwargs):
+        """Add sample packing configurations if applicable."""
+        if hasattr(self.cfg, "sample_packing") and self.cfg.sample_packing:
+            training_args_kwargs.update(
+                {
+                    "sample_packing": bool(self.cfg.sample_packing),
+                    "multipack_real_batches": not self.cfg.flash_attention
+                    or self.cfg.multipack_real_batches,
+                    "eval_sample_packing": bool(self.cfg.eval_sample_packing),
+                }
+            )
+
+            if self.cfg.sample_packing_bin_size is not None:
+                training_args_kwargs["sample_packing_bin_size"] = (
+                    self.cfg.sample_packing_bin_size
+                )
+
+            if self.cfg.sample_packing_group_size is not None:
+                training_args_kwargs["sample_packing_group_size"] = (
+                    self.cfg.sample_packing_group_size
+                )
+
+            if self.cfg.sample_packing_eff_est:
+                training_args_kwargs["sample_packing_efficiency"] = (
+                    self.cfg.sample_packing_eff_est
+                )
+
+    def _add_advanced_training_configs(self, training_args_kwargs):
+        """Add advanced training techniques configurations (ReLoRA & Lisa)."""
+        # ReLoRA configurations
+        if self.cfg.relora_steps:
+            training_args_kwargs.update(
+                {
+                    "relora_steps": self.cfg.relora_steps,
+                    "relora_warmup_steps": self.cfg.relora_warmup_steps,
+                }
+            )
+            if self.cfg.relora_anneal_steps:
+                training_args_kwargs["relora_anneal_steps"] = (
+                    self.cfg.relora_anneal_steps
+                )
+            if self.cfg.relora_prune_ratio:
+                training_args_kwargs["relora_prune_ratio"] = self.cfg.relora_prune_ratio
+
+        # Lisa configurations
+        if self.cfg.lisa_step_interval and self.cfg.lisa_n_layers:
+            training_args_kwargs.update(
+                {
+                    "lisa_n_layers": self.cfg.lisa_n_layers,
+                    "lisa_step_interval": self.cfg.lisa_step_interval,
+                    "lisa_layers_attribute": self.cfg.lisa_layers_attribute,
+                }
+            )
+
+    def _add_model_specific_configs(self, training_args_kwargs):
+        """Add model-specific configurations."""
+        # Chat template
+        if self.cfg.chat_template:
+            training_args_kwargs["chat_template"] = get_chat_template_from_config(
+                cfg=self.cfg,
+                tokenizer=self.tokenizer,
+            )
+
+        # NEFTune
+        if self.cfg.neftune_noise_alpha is not None:
+            training_args_kwargs["neftune_noise_alpha"] = self.cfg.neftune_noise_alpha
+
+        # Knowledge distillation configurations
+        if self.cfg.kd_ce_alpha is not None:
+            training_args_kwargs["kd_ce_alpha"] = self.cfg.kd_ce_alpha
+        if self.cfg.kd_alpha is not None:
+            training_args_kwargs["kd_alpha"] = self.cfg.kd_alpha
+        if self.cfg.kd_temperature is not None:
+            training_args_kwargs["kd_temperature"] = self.cfg.kd_temperature
+        if self.cfg.kd_zscore_base_temp is not None:
+            training_args_kwargs["kd_zscore_base_temp"] = self.cfg.kd_zscore_base_temp
+        if self.cfg.kd_top_k_before_softmax is not None:
+            training_args_kwargs["kd_top_k_before_softmax"] = (
+                self.cfg.kd_top_k_before_softmax
+            )
+
+        # Image configurations
+        if self.cfg.image_size:
+            training_args_kwargs["image_size"] = self.cfg.image_size
+        if self.cfg.image_resize_algorithm:
+            training_args_kwargs["image_resize_algorithm"] = (
+                self.cfg.image_resize_algorithm
+            )
+
+        # Accelerator configuration
+        if self.cfg.accelerator_config:
+            training_args_kwargs["accelerator_config"] = self.cfg.accelerator_config
+
+    def _configure_optimizer(self, training_args_kwargs):
+        """Configure optimizer settings."""
+        custom_supported_optimizers = [opt.value for opt in CustomSupportedOptimizers]
+
+        if self.cfg.optimizer in custom_supported_optimizers:
+            # Use custom optimizer implementation
+            self._configure_custom_optimizer(training_args_kwargs)
+        else:
+            # Use transformers' optimizer
+            training_args_kwargs["optim"] = self.cfg.optimizer
+            self._add_optimizer_args(training_args_kwargs)
+
+        # Handle optimizer targeting specific modules
+        if self.cfg.optim_target_modules:
+            training_args_kwargs["optim_target_modules"] = self.cfg.optim_target_modules
+
+        # Special case for anyprecision optimizer
+        if self.cfg.optimizer == "adamw_anyprecision":
+            if Path(self.cfg.torchdistx_path).exists():
+                sys.path.append(self.cfg.torchdistx_path)
+                importlib.import_module("torchdistx")
+
+    def _configure_custom_optimizer(self, training_args_kwargs):
+        """Configure custom optimizer settings."""
+        # Common optimizer kwargs
+        optimizer_kwargs = {
+            "lr": training_args_kwargs.get("learning_rate"),
+            "weight_decay": training_args_kwargs.get("weight_decay"),
+        }
+
+        # Add Adam-specific kwargs if available
+        adam_kwargs = self._get_adam_kwargs(training_args_kwargs)
+
+        # Get optimizer class and update kwargs based on optimizer type
+        optimizer_cls = self._get_optimizer_class(
+            training_args_kwargs, optimizer_kwargs, adam_kwargs
+        )
+
+        # Add any additional optimizer args from config
+        self._update_optimizer_kwargs_from_config(optimizer_kwargs)
+
+        training_args_kwargs["optimizer_cls_and_kwargs"] = (
+            optimizer_cls,
+            optimizer_kwargs,
+        )
+
+    def _get_adam_kwargs(self, training_args_kwargs):
+        """Get Adam-specific kwargs if available."""
+        adam_kwargs = {}
+        if training_args_kwargs.get("adam_beta1") and training_args_kwargs.get(
+            "adam_beta2"
+        ):
+            adam_kwargs["betas"] = (
+                training_args_kwargs.get("adam_beta1"),
+                training_args_kwargs.get("adam_beta2"),
+            )
+        if training_args_kwargs.get("adam_epsilon"):
+            adam_kwargs["eps"] = training_args_kwargs.get("adam_epsilon")
+        return adam_kwargs
+
+    def _get_optimizer_class(self, training_args_kwargs, optimizer_kwargs, adam_kwargs):
+        """Get optimizer class based on configuration."""
+        if self.cfg.optimizer == "muon":
+            from axolotl.contribs.mit.muon import MuonOptimizerFactory   # pylint: disable=no-name-in-module
+
+            optimizer_cls = MuonOptimizerFactory
+            optimizer_kwargs.update(adam_kwargs)
+        elif self.cfg.optimizer == "optimi_adamw":
+            from optimi import AdamW
+
+            optimizer_kwargs["foreach"] = False
+            optimizer_cls = AdamW
+            optimizer_kwargs.update(adam_kwargs)
+        elif self.cfg.optimizer == "ao_adamw_4bit":
+            from torchao.prototype.low_bit_optim import AdamW4bit
+
+            optimizer_cls = AdamW4bit
+            optimizer_kwargs.update(adam_kwargs)
+            LOG.warning(
+                f"`ao_adamw_4bit` will be deprecated soon. Please use `{OptimizerNames.ADAMW_TORCH_4BIT}` instead."
+            )
+        elif self.cfg.optimizer == "ao_adamw_8bit":
+            from torchao.prototype.low_bit_optim import AdamW8bit
+
+            optimizer_cls = AdamW8bit
+            optimizer_kwargs.update(adam_kwargs)
+        elif self.cfg.optimizer == "ao_adamw_fp8":
+            from torchao.prototype.low_bit_optim import AdamWFp8
+
+            optimizer_cls = AdamWFp8
+            optimizer_kwargs.update(adam_kwargs)
+        elif self.cfg.optimizer == "adopt_adamw":
+            from axolotl.utils.optimizers.adopt import ADOPT
+
+            optimizer_cls = ADOPT
+            adam_kwargs["decouple"] = True
+            optimizer_kwargs.update(adam_kwargs)
+        elif self.cfg.optimizer == "came_pytorch":
+            from came_pytorch import CAME
+
+            optimizer_cls = CAME
+
+            beta1 = training_args_kwargs.get("adam_beta1", 0.9)
+            beta2 = training_args_kwargs.get("adam_beta2", 0.999)
+            beta3 = training_args_kwargs.get("adam_beta2", 0.9999)
+            eps1 = training_args_kwargs.get("adam_epsilon", 1e-30)
+            eps2 = training_args_kwargs.get("adam_epsilon2", 1e-16)
+
+            adam_kwargs["betas"] = (beta1, beta2, beta3)
+            adam_kwargs["eps"] = (eps1, eps2)
+            optimizer_kwargs.update(adam_kwargs)
+        else:
+            # Default case or unsupported optimizer
+            optimizer_cls = None
+
+        return optimizer_cls
+
+    def _update_optimizer_kwargs_from_config(self, optimizer_kwargs):
+        """Update optimizer kwargs from config."""
+        if self.cfg.optim_args:
+            if isinstance(self.cfg.optim_args, dict):
+                optimizer_kwargs.update(self.cfg.optim_args)
+            else:
+                # Parse string format "key1=value1,key2=value2"
+                for mapping in self.cfg.optim_args.replace(" ", "").split(","):
+                    key, value = mapping.split("=")
+                    optimizer_kwargs[key] = value
+
+    def _add_optimizer_args(self, training_args_kwargs):
+        """Add optimizer arguments if available."""
+        if self.cfg.optim_args:
+            if isinstance(self.cfg.optim_args, dict):
+                optim_args = ",".join(
+                    [f"{key}={value}" for key, value in self.cfg.optim_args.items()]
+                )
+            else:
+                optim_args = self.cfg.optim_args
+            training_args_kwargs["optim_args"] = optim_args
+
+    def _get_training_args_cls(self):
+        """Get the appropriate training arguments class."""
+        if self.cfg.reward_model:
+            return AxolotlRewardConfig
+        if self.cfg.process_reward_model:
+            return AxolotlPRMConfig
+        return AxolotlTrainingArguments
+
+    def _prepare_data_collator_kwargs(self):
+        """Prepare data collator kwargs."""
+        data_collator_kwargs = {"padding": True}  # True/"longest" is the default
+
+        if self.cfg.pad_to_sequence_len:
+            data_collator_kwargs["pad_to_multiple_of"] = 64 * math.ceil(
+                self.cfg.sequence_len / 64
+            )
+        else:
+            data_collator_kwargs["pad_to_multiple_of"] = 64
+
+        if self.cfg.reward_model:
+            data_collator_kwargs["max_length"] = self.cfg.sequence_len
+
+        return data_collator_kwargs
+
+    def _prepare_trainer_kwargs(self, trainer_cls, data_collator_kwargs, training_args):
+        """Prepare trainer kwargs."""
+        trainer_kwargs = {}
+
+        # Handle special data collators for evaluation
+        if eval_data_collator := self.build_collator(
+            training_args, is_eval=True, **data_collator_kwargs
+        ):
+            if not (self.cfg.reward_model or self.cfg.process_reward_model):
+                trainer_kwargs["eval_data_collator"] = eval_data_collator
+
+        # Add bench data collator if needed
+        if not (self.cfg.reward_model or self.cfg.process_reward_model):
+            trainer_kwargs["bench_data_collator"] = transformers.DataCollatorForSeq2Seq(
+                self.tokenizer,
+                return_tensors="pt",
+                **data_collator_kwargs,
+            )
+
+        # Add tokenizer or processing class
+        sig = inspect.signature(trainer_cls)
+        if "processing_class" in sig.parameters.keys():
+            trainer_kwargs["processing_class"] = self.tokenizer
+        else:
+            trainer_kwargs["tokenizer"] = self.tokenizer
+
+        # Add dataset tags if available
+        if (
+            not (trainer_cls in [AxolotlRewardTrainer, AxolotlPRMTrainer])
+            and self.cfg.datasets is not None
+        ):
+            trainer_kwargs["dataset_tags"] = [
+                d["path"] for d in self.cfg.datasets if not Path(d["path"]).is_dir()
+            ]
+
+        return trainer_kwargs
+
+    def build_collator(
+        self, training_args: AxolotlTrainingArguments, is_eval=False, **kwargs
+    ):
+        if training_args.pretraining:
+            if (
+                self.cfg.pretraining_sample_concatenation is False
+                or self.cfg.micro_batch_size > 1
+            ):
+                return DataCollatorForSeq2Seq(self.tokenizer, **kwargs)
+            return None
+
+        if self.cfg.model_config_type == "mamba":
+            return MambaDataCollator(tokenizer=self.tokenizer)
+
+        use_batch_sampler_collator = False
+        if is_eval is False and training_args.sample_packing:
+            use_batch_sampler_collator = True
+        if is_eval and training_args.eval_sample_packing:
+            use_batch_sampler_collator = True
+
+        collator: Type[
+            V2BatchSamplerDataCollatorForSeq2Seq
+            | BatchSamplerDataCollatorForSeq2Seq
+            | DataCollatorForSeq2Seq
+            | DataCollatorWithFlattening
+            | RewardDataCollatorWithPadding
+        ]
+        collator_args = [self.tokenizer]
+        if self.cfg.reward_model:
+            collator = RewardDataCollatorWithPadding
+            if "max_length" in kwargs:
+                kwargs.pop("max_length")
+        elif use_batch_sampler_collator:
+            if self.cfg.flex_attention:
+                collator = V2BatchSamplerDataCollatorForSeq2Seq
+            elif self.cfg.model_config_type in SUPPORTED_MULTIPACK_MODEL_TYPES:
+                collator = V2BatchSamplerDataCollatorForSeq2Seq
+            elif (
+                self.cfg.model_config_type in ["llama"]
+                and self.cfg.flash_attention is not True
+            ):
+                collator = V2BatchSamplerDataCollatorForSeq2Seq
+            else:
+                collator = BatchSamplerDataCollatorForSeq2Seq
+        else:
+            if self.cfg.processor_type and self.processor:
+                collator = MultiModalChatDataCollator
+                kwargs["processing_strategy"] = get_processing_strategy(
+                    self.processor,
+                    training_args.chat_template,
+                    self.cfg.chat_template,
+                    image_size=training_args.image_size,
+                    image_resize_algorithm=training_args.image_resize_algorithm,
+                )
+            elif self.cfg.batch_flattening:
+                collator = DataCollatorWithFlattening
+                collator_args.pop(0)
+                kwargs.pop("pad_to_multiple_of", None)
+                kwargs.pop("padding", None)
+            elif self.cfg.kd_trainer:
+                from axolotl.integrations.kd.collator import (
+                    DataCollatorForKD,
+                    KDBatchSamplerDataCollatorForSeq2Seq,
+                )
+
+                if self.cfg.sample_packing:
+                    collator = KDBatchSamplerDataCollatorForSeq2Seq
+                else:
+                    collator = DataCollatorForKD
+            else:
+                collator = DataCollatorForSeq2Seq
+
+        kwargs["return_tensors"] = "pt"
+
+        return collator(
+            *collator_args,
+            **kwargs,
+        )

src/axolotl/core/trainers/builders/rl.py

@@ -0,0 +1,367 @@
+"""RL trainer / training args builder implementation"""
+
+import inspect
+from pathlib import Path
+
+from axolotl.core.trainers.builders.base import TrainerBuilderBase
+from axolotl.core.trainers.dpo import DPOStrategy
+from axolotl.core.trainers.dpo.args import AxolotlDPOConfig
+from axolotl.core.trainers.grpo import GRPOStrategy
+from axolotl.core.trainers.trl import (
+    AxolotlCPOTrainer,
+    AxolotlKTOTrainer,
+    AxolotlORPOTrainer,
+)
+from axolotl.core.training_args import (
+    AxolotlCPOConfig,
+    AxolotlKTOConfig,
+    AxolotlORPOConfig,
+)
+from axolotl.utils.models import ensure_dtype
+
+
+class HFRLTrainerBuilder(TrainerBuilderBase):
+    """Trainer factory class for TRL-based RLHF trainers (e.g. DPO)"""
+
+    def get_callbacks(self):
+        callbacks = super().get_callbacks()
+
+        return callbacks
+
+    def get_post_trainer_create_callbacks(self, trainer):
+        callbacks = super().get_post_trainer_create_callbacks(trainer=trainer)
+        return callbacks
+
+    def build_training_arguments(self, total_num_steps):
+        training_args_kwargs = {}
+        for arg in [
+            "adam_beta1",
+            "adam_beta2",
+            "adam_epsilon",
+            "dataloader_num_workers",
+            "dataloader_pin_memory",
+        ]:
+            if hasattr(self.cfg, arg) and getattr(self.cfg, arg) is not None:
+                training_args_kwargs[arg] = getattr(self.cfg, arg)
+
+        if self.cfg.hub_model_id:
+            training_args_kwargs["hub_model_id"] = self.cfg.hub_model_id
+            training_args_kwargs["push_to_hub"] = True
+            training_args_kwargs["hub_private_repo"] = True
+            training_args_kwargs["hub_always_push"] = True
+
+            if self.cfg.hub_strategy:
+                training_args_kwargs["hub_strategy"] = self.cfg.hub_strategy
+
+        if self.cfg.save_safetensors is not None:
+            training_args_kwargs["save_safetensors"] = self.cfg.save_safetensors
+
+        if self.eval_dataset:
+            training_args_kwargs["eval_strategy"] = "steps"
+            training_args_kwargs["eval_steps"] = self.cfg.eval_steps
+        else:
+            training_args_kwargs["eval_strategy"] = "no"
+
+        if self.cfg.bf16 or self.cfg.bfloat16:
+            training_args_kwargs["bf16"] = True
+
+        training_args_kwargs["loraplus_lr_ratio"] = self.cfg.loraplus_lr_ratio
+        training_args_kwargs["loraplus_lr_embedding"] = self.cfg.loraplus_lr_embedding
+        training_args_kwargs["lr_scheduler_type"] = (
+            self.cfg.lr_scheduler if self.cfg.lr_scheduler else "cosine"
+        )
+        training_args_kwargs["lr_scheduler_kwargs"] = (
+            self.cfg.lr_scheduler_kwargs if self.cfg.lr_scheduler_kwargs else {}
+        )
+        if self.cfg.remove_unused_columns is not None:
+            training_args_kwargs["remove_unused_columns"] = (
+                self.cfg.remove_unused_columns
+            )
+        else:
+            training_args_kwargs["remove_unused_columns"] = False
+
+        if self.cfg.dataloader_pin_memory is not None:
+            training_args_kwargs["dataloader_pin_memory"] = (
+                self.cfg.dataloader_pin_memory
+            )
+        if self.cfg.dataloader_num_workers is not None:
+            training_args_kwargs["dataloader_num_workers"] = (
+                self.cfg.dataloader_num_workers
+            )
+        if self.cfg.dataloader_prefetch_factor is not None:
+            training_args_kwargs["dataloader_prefetch_factor"] = (
+                self.cfg.dataloader_prefetch_factor
+            )
+        if self.cfg.gradient_checkpointing:
+            training_args_kwargs["gradient_checkpointing"] = (
+                self.cfg.gradient_checkpointing
+            )
+            if self.cfg.gradient_checkpointing_kwargs is not None:
+                training_args_kwargs["gradient_checkpointing_kwargs"] = (
+                    self.cfg.gradient_checkpointing_kwargs
+                )
+            else:
+                training_args_kwargs["gradient_checkpointing_kwargs"] = {
+                    "use_reentrant": False
+                }
+
+        # set save_strategy and save_steps
+        if self.cfg.save_steps:
+            training_args_kwargs["save_strategy"] = "steps"
+            training_args_kwargs["save_steps"] = self.cfg.save_steps
+        elif self.cfg.save_strategy:
+            training_args_kwargs["save_strategy"] = self.cfg.save_strategy
+        else:
+            # default to saving each epoch if not defined
+            training_args_kwargs["save_strategy"] = "epoch"
+
+        training_args_kwargs["save_only_model"] = self.cfg.save_only_model
+
+        if self.cfg.dataset_processes:
+            training_args_kwargs["dataset_num_proc"] = self.cfg.dataset_processes
+
+        if self.cfg.trl and self.cfg.trl.beta is not None:
+            training_args_kwargs["beta"] = self.cfg.trl.beta
+        elif self.cfg.rl_beta is not None:
+            training_args_kwargs["beta"] = self.cfg.rl_beta
+        elif self.cfg.orpo_alpha is not None:
+            # trl does some odd mapping of alpha to beta to reuse the beta parameter ???
+            training_args_kwargs["beta"] = self.cfg.orpo_alpha
+
+        if self.cfg.rpo_alpha is not None:
+            training_args_kwargs["rpo_alpha"] = self.cfg.rpo_alpha
+
+        if self.cfg.use_wandb:
+            training_args_kwargs["run_name"] = self.cfg.wandb_name
+
+        training_args_cls = None
+        blocklist_args_kwargs = []
+        if self.cfg.rl == "simpo":
+            training_args_cls = AxolotlCPOConfig
+            training_args_kwargs["loss_type"] = "simpo"
+            training_args_kwargs["max_length"] = self.cfg.sequence_len
+            training_args_kwargs["simpo_gamma"] = self.cfg.simpo_gamma
+            if self.cfg.cpo_alpha is not None:
+                training_args_kwargs["cpo_alpha"] = self.cfg.cpo_alpha
+
+        elif self.cfg.rl == "orpo":
+            training_args_cls = AxolotlORPOConfig
+            training_args_kwargs["max_length"] = self.cfg.sequence_len
+            if self.cfg.max_prompt_len:
+                training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
+
+        elif self.cfg.rl == "kto":
+            training_args_cls = AxolotlKTOConfig
+
+            training_args_kwargs["desirable_weight"] = (
+                self.cfg.kto_desirable_weight or 1.0
+            )
+            training_args_kwargs["undesirable_weight"] = (
+                self.cfg.kto_undesirable_weight or 1.0
+            )
+
+            training_args_kwargs["max_length"] = self.cfg.sequence_len
+            if self.cfg.max_prompt_len:
+                training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
+
+        elif self.cfg.rl == "grpo":
+            training_args_cls = GRPOStrategy.get_training_args_class()
+            training_args_kwargs.update(GRPOStrategy.set_training_args_kwargs(self.cfg))
+            blocklist_args_kwargs = GRPOStrategy.get_blocklist_args_kwargs()
+
+        else:
+            training_args_cls = AxolotlDPOConfig
+            if self.cfg.rl == "ipo":
+                training_args_kwargs["loss_type"] = "ipo"
+            training_args_kwargs["max_length"] = self.cfg.sequence_len
+            training_args_kwargs["max_completion_length"] = None
+            training_args_kwargs["max_prompt_length"] = self.cfg.sequence_len
+            training_args_kwargs["generate_during_eval"] = self.cfg.use_wandb
+            if self.cfg.dpo_use_weighting is not None:
+                training_args_kwargs["use_weighting"] = self.cfg.dpo_use_weighting
+            if self.cfg.dpo_use_logits_to_keep is not None:
+                training_args_kwargs["use_logits_to_keep"] = (
+                    self.cfg.dpo_use_logits_to_keep
+                )
+
+        for blocklist_key in blocklist_args_kwargs:
+            if blocklist_key in training_args_kwargs:
+                del training_args_kwargs[blocklist_key]
+
+        max_steps = self.cfg.max_steps or total_num_steps or -1
+        training_args_kwargs["num_train_epochs"] = self.cfg.num_epochs
+        training_args = training_args_cls(  # pylint: disable=unexpected-keyword-arg
+            self.cfg.output_dir,
+            per_device_train_batch_size=self.cfg.micro_batch_size,
+            max_steps=max_steps,
+            gradient_accumulation_steps=self.cfg.gradient_accumulation_steps,
+            learning_rate=self.cfg.learning_rate,
+            warmup_steps=self.cfg.warmup_steps,
+            logging_first_step=True,
+            logging_steps=1,
+            optim=self.cfg.optimizer,
+            save_total_limit=self.cfg.save_total_limit or 5,
+            **training_args_kwargs,
+        )
+
+        # unset run_name so wandb sets up experiment names
+        if self.cfg.use_wandb and training_args.run_name == training_args.output_dir:
+            training_args.run_name = (  # pylint: disable=attribute-defined-outside-init
+                None
+            )
+
+        return training_args
+
+    def build(self, total_num_steps):
+        """Build and return an RL trainer instance"""
+        # Prepare RL-specific training args kwargs
+        training_args_kwargs = {
+            "per_device_train_batch_size": self.cfg.micro_batch_size,
+            "max_steps": self.cfg.max_steps or total_num_steps or -1,
+            "gradient_accumulation_steps": self.cfg.gradient_accumulation_steps,
+            "learning_rate": self.cfg.learning_rate,
+            "warmup_steps": self.cfg.warmup_steps,
+            "logging_first_step": True,
+            "logging_steps": 1,
+            "output_dir": self.cfg.output_dir,
+            "num_train_epochs": self.cfg.num_epochs,
+        }
+
+        # Handle dataset processes
+        if self.cfg.dataset_processes:
+            training_args_kwargs["dataset_num_proc"] = self.cfg.dataset_processes
+
+        # Handle beta/alpha parameters for different RL algorithms
+        if self.cfg.trl and self.cfg.trl.beta is not None:
+            training_args_kwargs["beta"] = self.cfg.trl.beta
+        elif self.cfg.rl_beta is not None:
+            training_args_kwargs["beta"] = self.cfg.rl_beta
+        elif self.cfg.orpo_alpha is not None:
+            # trl does some odd mapping of alpha to beta to reuse the beta parameter
+            training_args_kwargs["beta"] = self.cfg.orpo_alpha
+
+        if self.cfg.rpo_alpha is not None:
+            training_args_kwargs["rpo_alpha"] = self.cfg.rpo_alpha
+
+        # Determine training args class and add RL-specific parameters
+        training_args_cls = None
+        blocklist_args_kwargs = []
+
+        if self.cfg.rl == "simpo":
+            training_args_cls = AxolotlCPOConfig
+            training_args_kwargs["loss_type"] = "simpo"
+            training_args_kwargs["simpo_gamma"] = self.cfg.simpo_gamma
+            if self.cfg.cpo_alpha is not None:
+                training_args_kwargs["cpo_alpha"] = self.cfg.cpo_alpha
+        elif self.cfg.rl == "orpo":
+            training_args_cls = AxolotlORPOConfig
+            if self.cfg.max_prompt_len:
+                training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
+        elif self.cfg.rl == "kto":
+            training_args_cls = AxolotlKTOConfig
+            training_args_kwargs["desirable_weight"] = (
+                self.cfg.kto_desirable_weight or 1.0
+            )
+            training_args_kwargs["undesirable_weight"] = (
+                self.cfg.kto_undesirable_weight or 1.0
+            )
+            if self.cfg.max_prompt_len:
+                training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
+        elif self.cfg.rl == "grpo":
+            training_args_cls = GRPOStrategy.get_training_args_class()
+            training_args_kwargs.update(GRPOStrategy.set_training_args_kwargs(self.cfg))
+            blocklist_args_kwargs = GRPOStrategy.get_blocklist_args_kwargs()
+        else:  # Default to DPO
+            training_args_cls = AxolotlDPOConfig
+            if self.cfg.rl == "ipo":
+                training_args_kwargs["loss_type"] = "ipo"
+            training_args_kwargs["max_prompt_length"] = self.cfg.sequence_len
+            training_args_kwargs["max_completion_length"] = None
+            training_args_kwargs["generate_during_eval"] = self.cfg.use_wandb
+            if self.cfg.dpo_use_weighting is not None:
+                training_args_kwargs["use_weighting"] = self.cfg.dpo_use_weighting
+            if self.cfg.dpo_use_logits_to_keep is not None:
+                training_args_kwargs["use_logits_to_keep"] = (
+                    self.cfg.dpo_use_logits_to_keep
+                )
+
+        # Remove any blocklisted arguments
+        for blocklist_key in blocklist_args_kwargs:
+            if blocklist_key in training_args_kwargs:
+                del training_args_kwargs[blocklist_key]
+
+        # Create training args using the base class method
+        training_args = self.create_training_args(
+            args_cls=training_args_cls,
+            total_num_steps=total_num_steps,
+            **training_args_kwargs,
+        )
+
+        # Prepare trainer kwargs
+        trainer_kwargs = {}
+        if self.cfg.rl == "ipo" and self.cfg.dpo_label_smoothing:
+            trainer_kwargs["label_smoothing"] = self.cfg.dpo_label_smoothing
+        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.precompute_ref_log_probs is not None:
+            trainer_kwargs["precompute_ref_log_probs"] = (
+                self.cfg.precompute_ref_log_probs
+            )
+
+        # Determine trainer class and arguments
+        if self.cfg.rl == "grpo":
+            trainer_cls = GRPOStrategy.get_trainer_class()
+            trainer_args = [self.model]
+            trainer_args.extend(GRPOStrategy.set_trainer_args(self.cfg))
+            trainer_kwargs.update(GRPOStrategy.set_trainer_kwargs(self.cfg))
+        elif self.cfg.rl in ["dpo", "ipo"]:
+            trainer_cls = DPOStrategy.get_trainer_class()
+            trainer_args = [self.model, self.model_ref]
+        elif self.cfg.rl == "orpo":
+            trainer_cls = AxolotlORPOTrainer
+            trainer_args = [self.model]
+        elif self.cfg.rl in ["kto"]:
+            trainer_cls = AxolotlKTOTrainer
+            trainer_args = [self.model]
+        elif self.cfg.rl in ["simpo"]:
+            trainer_cls = AxolotlCPOTrainer
+            trainer_args = [self.model]
+        else:
+            raise ValueError(f"Unsupported RL: {self.cfg.rl}")
+
+        # Add tokenizer or processing class
+        sig = inspect.signature(trainer_cls)
+        if "tokenizer" in sig.parameters.keys():
+            trainer_kwargs["tokenizer"] = self.tokenizer
+        else:
+            trainer_kwargs["processing_class"] = self.tokenizer
+
+        # Add dataset tags if available
+        if self.cfg.datasets is not None and (
+            trainer_cls is DPOStrategy.get_trainer_class()
+        ):
+            trainer_kwargs["dataset_tags"] = [
+                d["path"] for d in self.cfg.datasets if not Path(d["path"]).is_dir()
+            ]
+
+        # Create the trainer
+        trainer = self.create_trainer(
+            trainer_cls=trainer_cls,
+            training_args=training_args,
+            trainer_args=trainer_args,
+            trainer_kwargs=trainer_kwargs,
+        )
+
+        # Handle FSDP specific settings
+        if self.cfg.fsdp:
+            ensure_dtype(trainer.model, dtype=self.cfg.torch_dtype)
+            if (
+                self.cfg.rl in ["dpo", "ipo"]
+                and hasattr(trainer, "ref_model")
+                and trainer.ref_model
+            ):
+                ensure_dtype(trainer.ref_model, dtype=self.cfg.torch_dtype)
+
+        return trainer

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

@@ -1,11 +1,14 @@
-"""DPO Specific Strategy for training"""
+"""
+DPO Specific Strategy for training
+"""
 
 from axolotl.core.trainers.dpo.trainer import AxolotlDPOTrainer
-from axolotl.utils.schemas.enums import RLType
 
 
 class DPOStrategy:
-    """Strategy for DPO training"""
+    """
+    Strategy for DPO training
+    """
 
     @classmethod
     def get_trainer_class(cls):
@@ -20,7 +23,7 @@ class DPOStrategy:
     @classmethod
     def set_training_args_kwargs(cls, cfg):
         training_args_kwargs = {}
-        if cfg.rl is RLType.IPO:
+        if cfg.rl == "ipo":
             training_args_kwargs["loss_type"] = "ipo"
         training_args_kwargs["max_length"] = cfg.sequence_len
         training_args_kwargs["max_completion_length"] = None

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

@@ -1,41 +1,37 @@
-"""GRPO Specific Strategy for training"""
+"""
+GRPO Specific Strategy for training
+"""
 
 import importlib
 import inspect
 import logging
-from typing import Any
 
 from trl.trainer.grpo_trainer import RewardFunc
 
-from axolotl.core.trainers.grpo.args import AxolotlGRPOConfig
+from axolotl.core.trainers.grpo.trainer import AxolotlGRPOTrainer
-from axolotl.core.trainers.grpo.trainer import (
-    AxolotlGRPOSequenceParallelTrainer,
-    AxolotlGRPOTrainer,
-)
-from axolotl.utils.dict import DictDefault
 from axolotl.utils.schemas.trl import TRLConfig
 
-LOG = logging.getLogger(__name__)
+LOG = logging.getLogger("axolotl")
 
 
 class GRPOStrategy:
-    """Strategy for GRPO training"""
+    """
+    Strategy for GRPO training
+    """
 
     @classmethod
-    def get_trainer_class(
+    def get_trainer_class(cls):
-        cls, sequence_parallel: bool
-    ) -> type[AxolotlGRPOTrainer] | type[AxolotlGRPOSequenceParallelTrainer]:
-        if sequence_parallel:
-            return AxolotlGRPOSequenceParallelTrainer
         return AxolotlGRPOTrainer
 
     @classmethod
-    def get_training_args_class(cls) -> type[AxolotlGRPOConfig]:
+    def get_training_args_class(cls):
+        from axolotl.core.trainers.grpo.args import AxolotlGRPOConfig
+
         return AxolotlGRPOConfig
 
     @classmethod
-    def set_training_args_kwargs(cls, cfg: DictDefault) -> dict[str, Any]:
+    def set_training_args_kwargs(cls, cfg):
-        grpo_args_kwargs: dict[str, Any] = {}
+        grpo_args_kwargs = {}
 
         if not hasattr(cfg, "trl") or not cfg.trl:
             return grpo_args_kwargs
@@ -44,8 +40,8 @@ class GRPOStrategy:
 
         if trl.use_vllm:
             grpo_args_kwargs["use_vllm"] = trl.use_vllm
-            grpo_args_kwargs["vllm_server_host"] = trl.vllm_server_host or trl.vllm.host  # type: ignore[attr-defined]
+            grpo_args_kwargs["vllm_server_host"] = trl.vllm_server_host or trl.vllm.host
-            grpo_args_kwargs["vllm_server_port"] = trl.vllm_server_port or trl.vllm.port  # type: ignore[attr-defined]
+            grpo_args_kwargs["vllm_server_port"] = trl.vllm_server_port or trl.vllm.port
             if trl.vllm_server_timeout:
                 grpo_args_kwargs["vllm_server_timeout"] = trl.vllm_server_timeout
             if trl.vllm_guided_decoding_regex:
@@ -106,18 +102,17 @@ class GRPOStrategy:
         return grpo_args_kwargs
 
     @classmethod
-    def set_trainer_args(cls, cfg: DictDefault) -> list[Any]:
+    def set_trainer_args(cls, cfg):
         trainer_args = []
         if cfg.trl and cfg.trl.reward_funcs:
             reward_funcs = []
             for reward_func_fqn in cfg.trl.reward_funcs:
                 reward_funcs.append(cls.get_reward_func(reward_func_fqn))
             trainer_args.append(reward_funcs)
-
         return trainer_args
 
     @classmethod
-    def set_trainer_kwargs(cls, cfg: DictDefault) -> dict[str, Any]:
+    def set_trainer_kwargs(cls, cfg):
         trainer_kwargs = {}
         if cfg.trl and cfg.trl.reward_processing_classes:
             trainer_kwargs["reward_processing_classes"] = (
@@ -131,7 +126,7 @@ class GRPOStrategy:
         return None
 
     @classmethod
-    def get_blocklist_args_kwargs(cls) -> list[str]:
+    def get_blocklist_args_kwargs(cls):
         return ["dataset_num_proc"]
 
     @classmethod
@@ -142,13 +137,13 @@ class GRPOStrategy:
         Args:
             reward_func_fqn (str): Fully qualified name of the reward function (e.g. r1_grpo.gsm8k_transform),
                 or a HF hub path to the reward model.
+        Raises:
+            ValueError: If the reward function does not accept at least two arguments.
 
         Returns:
             RewardFunc: A callable that accepts prompts and completions and returns rewards,
                 or a path to a reward model.
 
-        Raises:
-            ValueError: If the reward function does not accept at least two arguments.
         """
         try:
             # use importlib to dynamically load the reward function from the module

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

@@ -11,4 +11,6 @@ from axolotl.core.training_args import AxolotlTrainingMixins
 
 @dataclass
 class AxolotlGRPOConfig(AxolotlTrainingMixins, GRPOConfig):
-    """Axolotl GRPO Config for GRPO training"""
+    """
+    Axolotl GRPO Config for GRPO training
+    """

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

@@ -1,172 +0,0 @@
-"""Repeat random sampler (similar to the one implemented in
-https://github.com/huggingface/trl/blob/main/trl/trainer/grpo_trainer.py) that adds
-sequence parallelism functionality; i.e., duplicating data across ranks in the same
-sequence parallel group.
-"""
-
-from typing import Iterator, Sized
-
-import torch
-from torch.utils.data import Sampler
-
-
-class SequenceParallelRepeatRandomSampler(Sampler):
-    """Sampler for GRPO training with sequence parallelism.
-
-    This sampler ensures:
-    - Ranks in the same sequence parallel (SP) group receive identical data.
-    - Each index is repeated multiple times for sampling different completions.
-    - Entire batches are repeated for reuse in multiple updates.
-    - Data is properly distributed across SP groups.
-
-    In the table below, the values represent dataset indices. Each SP group has
-    `sequence_parallel_degree = 2` GPUs working together on the same data. There are 2
-    SP groups (SP0 and SP1), with `world_size = 4` total GPUs.
-
-                                               Sequence Parallel Groups
-                                        |       SP0        |       SP1        |
-                                        |  GPU 0  |  GPU 1 |  GPU 2  |  GPU 3 |
-                    global_step  step    <---> mini_repeat_count=3
-                                            <----------> batch_size=2 per SP group
-    grad_accum=2   ▲  ▲  0       0         [0 0 0  1 1 1]     [2 2 2  3 3 3]   <- SP groups get different data
-                   ▼  |  0       1         [0 0 0  1 1 1]     [2 2 2  3 3 3]   <- Same data for each SP group GPU
-                      |
-                      |  1       2         [0 0 0  1 1 1]     [2 2 2  3 3 3]   <- Repeat same indices for iterations
-    num_iterations=2  ▼  1       3         [0 0 0  1 1 1]     [2 2 2  3 3 3]   <- When using gradient accumulation
-
-                         2       4         [4 4 4  5 5 5]     [6 6 6  7 7 7]   <- New batch of data indices
-                         2       5         [4 4 4  5 5 5]     [6 6 6  7 7 7]
-                                            ...
-
-    Args:
-        dataset: Dataset to sample from.
-        mini_repeat_count: How many times to repeat each sample immediately.
-        world_size: Total number of processes.
-        rank: Rank of current process.
-        batch_size: Number of samples per batch.
-        repeat_count: How many times to repeat the full sampling process.
-        sequence_parallel_degree: Number of ranks in a sequence parallel group.
-        shuffle: Whether to shuffle the dataset.
-        seed: Random seed for shuffling.
-        drop_last: Whether to drop the last incomplete batch.
-    """
-
-    def __init__(
-        self,
-        dataset: Sized,
-        mini_repeat_count: int,
-        world_size: int,
-        rank: int,
-        batch_size: int = 1,
-        repeat_count: int = 1,
-        sequence_parallel_degree: int = 1,
-        shuffle: bool = True,
-        seed: int = 0,
-        drop_last: bool = False,
-    ):
-        self.dataset = dataset
-        self.mini_repeat_count = mini_repeat_count
-        self.batch_size = batch_size
-        self.repeat_count = repeat_count
-        self.shuffle = shuffle
-        self.seed = seed
-        self.drop_last = drop_last
-        self.epoch = 0
-
-        self.world_size = world_size
-        self.rank = rank
-
-        # Sequence parallelism parameters
-        self.sequence_parallel_degree = sequence_parallel_degree
-        self.num_sp_groups = world_size // sequence_parallel_degree
-        self.sp_group_id = rank // sequence_parallel_degree
-
-        # Adjust dataset size for distributed sampling
-        self.num_samples = len(self.dataset)
-        self.total_size = self.num_samples
-
-        # Calculate effective number of samples per SP group
-        if (
-            self.drop_last
-            and self.total_size % (self.num_sp_groups * self.batch_size) != 0
-        ):
-            # Drop last incomplete batch if drop_last is True
-            self.num_samples_per_sp_group = (
-                self.total_size // self.batch_size // self.num_sp_groups
-            ) * self.batch_size
-        else:
-            # Round up to include last batch if drop_last is False
-            self.num_samples_per_sp_group = (
-                (self.total_size + self.batch_size * self.num_sp_groups - 1)
-                // (self.batch_size * self.num_sp_groups)
-                * self.batch_size
-            )
-
-        if shuffle:
-            self.generator = torch.Generator()
-            self.generator.manual_seed(seed)
-
-    def __iter__(self) -> Iterator[int]:
-        """Creates iterator over dataset indices.
-
-        Returns:
-            Iterator that yields indices into the dataset.
-        """
-        # Deterministically shuffle based on epoch and seed
-        if self.shuffle:
-            indices = torch.randperm(
-                self.num_samples, generator=self.generator
-            ).tolist()
-        else:
-            indices = list(range(self.num_samples))
-
-        # Add extra samples to make it evenly divisible by batch_size
-        if len(indices) % self.batch_size != 0:
-            padding = indices[: self.batch_size - len(indices) % self.batch_size]
-            indices += padding
-
-        # Subsample based on SP group ID
-        # Each SP group gets distinct batches of data
-        batch_indices = []
-        for i in range(0, len(indices), self.batch_size * self.num_sp_groups):
-            start_idx = i + self.sp_group_id * self.batch_size
-            end_idx = min(start_idx + self.batch_size, len(indices))
-            if start_idx < len(indices):
-                for j in range(self.batch_size):
-                    if start_idx + j < end_idx:
-                        batch_indices.append(indices[start_idx + j])
-
-        # Make sure batch_indices is exactly batch_size * num_batches_per_sp_group
-        if self.drop_last:
-            num_batches_per_sp_group = self.num_samples_per_sp_group // self.batch_size
-            target_len = self.batch_size * num_batches_per_sp_group
-            if len(batch_indices) > target_len:
-                batch_indices = batch_indices[:target_len]
-
-        # Apply the GRPO repeat pattern
-        final_indices = []
-        for _ in range(self.repeat_count):
-            for idx in batch_indices:
-                for _ in range(self.mini_repeat_count):
-                    final_indices.append(idx)
-
-        return iter(final_indices)
-
-    def __len__(self) -> int:
-        """Returns the total length of the iterable including repetitions.
-
-        Returns:
-            Total number of samples.
-        """
-        # Total length including all repetitions
-        return (
-            self.num_samples_per_sp_group * self.mini_repeat_count * self.repeat_count
-        )
-
-    def set_epoch(self, epoch: int) -> None:
-        """Sets the epoch for this sampler.
-
-        Args:
-            epoch: Epoch number to use for shuffling.
-        """
-        self.epoch = epoch

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

@@ -1,653 +1,69 @@
-"""Axolotl GRPO trainers (with and without sequence parallelism handling)"""
+"""
+Axolotl GRPO trainer
+"""
 
-# pylint: disable=too-many-lines,duplicate-code,protected-access,no-member
+from contextlib import nullcontext
 
-import warnings
+from accelerate.utils import is_deepspeed_available, is_peft_model
-from typing import Any
-
-import datasets
-import torch
-import torch.distributed as dist
-import torch.utils.data
-from accelerate.utils import (
-    broadcast_object_list,
-    gather,
-    gather_object,
-    is_peft_available,
-)
-from datasets import Dataset, IterableDataset
-from torch import nn
-from torch.utils.data import (
-    BatchSampler,
-    DataLoader,
-    Sampler,
-)
-from transformers import (
-    PreTrainedModel,
-    PreTrainedTokenizerBase,
-    Trainer,
-    TrainerCallback,
-)
-from transformers.trainer_utils import seed_worker
 from trl import GRPOTrainer
-from trl.data_utils import (
+from trl.extras.profiling import profiling_decorator
-    apply_chat_template,
-    is_conversational,
-    maybe_apply_chat_template,
-)
-from trl.extras.profiling import profiling_context
-from trl.models import unwrap_model_for_generation
-from trl.trainer.grpo_config import GRPOConfig
-from trl.trainer.grpo_trainer import RewardFunc, nanstd
-from trl.trainer.utils import pad
 
-from axolotl.core.trainers.grpo.sampler import SequenceParallelRepeatRandomSampler
 from axolotl.core.trainers.mixins import RngLoaderMixin, SchedulerMixin
-from axolotl.monkeypatch.attention.ring_attn.patch import get_ring_attn_group
 
-if is_peft_available():
+if is_deepspeed_available():
-    # pylint: disable=unused-import
+    import deepspeed
-    from peft import PeftConfig
 
 
 class AxolotlGRPOTrainer(RngLoaderMixin, SchedulerMixin, GRPOTrainer):
-    """Extend the base GRPOTrainer for axolotl helpers"""
+    """
+    Extend the base GRPOTrainer for axolotl helpers
+    """
 
     _tag_names = ["trl", "grpo", "axolotl"]
 
-
+    @profiling_decorator
-class AxolotlGRPOSequenceParallelTrainer(AxolotlGRPOTrainer):
+    def _move_model_to_vllm(self):
-    """Extend the base GRPOTrainer for sequence parallelism handling"""
+        # For DeepSpeed ZeRO-3, we need to gather all parameters before operations
-
+        deepspeed_plugin = self.accelerator.state.deepspeed_plugin
-    def __init__(
+        zero_stage_3 = deepspeed_plugin is not None and deepspeed_plugin.zero_stage == 3
-        self,
+        gather_if_zero3 = (
-        model: str | PreTrainedModel,
+            deepspeed.zero.GatheredParameters if zero_stage_3 else nullcontext
-        reward_funcs: RewardFunc | list[RewardFunc],
-        args: GRPOConfig | None = None,
-        train_dataset: Dataset | IterableDataset | None = None,
-        eval_dataset: (
-            Dataset | IterableDataset | dict[str, Dataset | IterableDataset] | None
-        ) = None,
-        processing_class: PreTrainedTokenizerBase | None = None,
-        reward_processing_classes: (
-            PreTrainedTokenizerBase | list[PreTrainedTokenizerBase] | None
-        ) = None,
-        callbacks: list[TrainerCallback] | None = None,
-        optimizers: tuple[
-            torch.optim.Optimizer | None, torch.optim.lr_scheduler.LambdaLR | None
-        ] = (None, None),
-        peft_config: "PeftConfig | None" = None,
-    ):
-        # First call the superclass constructor with all arguments
-        super().__init__(
-            model=model,
-            reward_funcs=reward_funcs,
-            args=args,
-            train_dataset=train_dataset,
-            eval_dataset=eval_dataset,
-            processing_class=processing_class,
-            reward_processing_classes=reward_processing_classes,
-            callbacks=callbacks,
-            optimizers=optimizers,
-            peft_config=peft_config,
         )
 
-        # Get number of SP groups (number of processes divided by SP degree)
+        if is_peft_model(self.model):
-        num_processes = self.accelerator.num_processes
+            # With PEFT and DeepSpeed ZeRO Stage 3, we must gather the full model at once before merging, as merging
-        num_sp_groups = num_processes // self.args.sequence_parallel_degree
+            # adapters in a sharded manner is not supported.
+            with gather_if_zero3(list(self.model.parameters())):
+                self.model.merge_adapter()
 
-        # Calculate batch size per SP group (not per process)
+                # Update vLLM weights while parameters are gathered
-        sp_group_batch_size = self.args.per_device_train_batch_size * num_sp_groups
+                for name, param in self.model.named_parameters():
-        possible_values = [
+                    # When using PEFT, we need to recover the original parameter name and discard some parameters
-            n_gen
+                    name = (
-            for n_gen in range(2, sp_group_batch_size + 1)
+                        name.removeprefix("base_model.model.")
-            if (sp_group_batch_size) % n_gen == 0
+                        .removeprefix("base_model.model.")
-        ]
+                        .replace(".base_layer", "")
-
-        if self.num_generations not in possible_values:
-            raise ValueError(
-                f"The batch size per SP group ({num_sp_groups} x "
-                f"{self.args.per_device_train_batch_size}) must be evenly divisible by "
-                f"the number of generations per prompt ({self.num_generations}). Given "
-                "the current configuration, the valid values for the number of "
-                f"generations are: {possible_values}."
                     )
+                    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.args.eval_strategy != "no":
-            # If sequence parallelism is enabled, calculate batch size per SP group
-            sp_group_eval_batch_size = args.per_device_eval_batch_size * num_sp_groups  # type: ignore[union-attr]
-            possible_values = [
-                n_gen
-                for n_gen in range(2, sp_group_eval_batch_size + 1)
-                if (sp_group_eval_batch_size) % n_gen == 0
-            ]
-
-            if self.num_generations not in possible_values:
-                raise ValueError(
-                    f"With sequence parallelism (degree {self.args.sequence_parallel_degree}), "
-                    f"the eval batch size per SP group ({num_sp_groups} x {self.args.per_device_eval_batch_size}) "
-                    f"must be evenly divisible by the number of generations per prompt "
-                    f"({self.num_generations}). Given the current eval batch size, "
-                    f"the valid values for the number of generations are: {possible_values}."
-                )
-
-        # Initialize the SP group
-        self.sp_group = get_ring_attn_group()
-        self.rank = dist.get_rank()
-        self.world_size = dist.get_world_size()
-        self.local_rank = dist.get_rank(group=self.sp_group)
-        self.local_world_size = dist.get_world_size(group=self.sp_group)
-
-    def _get_train_sampler(self) -> Sampler:
-        effective_batch_size = (
-            self.args.per_device_train_batch_size
-            * self.world_size
-            * self.args.gradient_accumulation_steps
-        )
-
-        return SequenceParallelRepeatRandomSampler(
-            dataset=self.train_dataset,
-            mini_repeat_count=self.num_generations,
-            world_size=self.world_size,
-            rank=self.rank,
-            batch_size=effective_batch_size
-            // self.num_generations
-            // self.args.sequence_parallel_degree,
-            repeat_count=self.num_iterations * self.args.gradient_accumulation_steps,
-            sequence_parallel_degree=self.args.sequence_parallel_degree,
-            shuffle=True,
-            seed=self.args.seed,
-            drop_last=True,
-        )
-
-    def _create_dataloader_params(self, is_eval=False, custom_batch_size=None):
-        """Create common dataloader parameters for train or eval."""
-        batch_size = custom_batch_size or (
-            self.args.eval_batch_size if is_eval else self._train_batch_size
-        )
-
-        params = {
-            "batch_size": batch_size,
-            "collate_fn": self.data_collator,
-            "num_workers": self.args.dataloader_num_workers,
-            "pin_memory": self.args.dataloader_pin_memory,
-        }
-
-        # Add persistent workers only for training
-        if not is_eval and hasattr(self.args, "dataloader_persistent_workers"):
-            params["persistent_workers"] = self.args.dataloader_persistent_workers
-
-        # Add prefetch factor if specified
-        if self.args.dataloader_prefetch_factor:
-            params["prefetch_factor"] = self.args.dataloader_prefetch_factor
-
-        return params
-
-    def _prepare_dataloader(
-        self, dataset, sampler, is_eval=False, custom_batch_size=None
-    ):
-        """Prepare a dataloader with the given dataset and sampler."""
-        # Get base parameters
-        dataloader_params = self._create_dataloader_params(is_eval, custom_batch_size)
-
-        # Add sampler configuration
-        if not isinstance(dataset, torch.utils.data.IterableDataset):
-            if isinstance(sampler, BatchSampler):
-                # batch_size and batch_sampler are mutually exclusive
-                dataloader_params["batch_sampler"] = sampler
-                del dataloader_params["batch_size"]
-            else:
-                dataloader_params["sampler"] = sampler
-                dataloader_params["drop_last"] = self.args.dataloader_drop_last
-
-            if not is_eval:
-                dataloader_params["worker_init_fn"] = seed_worker
-
-        # Create the dataloader
-        dataloader = DataLoader(dataset, **dataloader_params)
-
-        if self.args.sample_packing and (
-            (not is_eval and not self.args.pretraining)
-            or (is_eval and self.args.eval_sample_packing is not False)
-        ):
-            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
-
-        # Otherwise prepare with accelerator
-        return self.accelerator.prepare_data_loader(dataloader)
-
-    def get_train_dataloader(self) -> DataLoader:
-        """Get dataloader for training"""
-        train_dataset = self.train_dataset
-        # pylint: disable=access-member-before-definition
-        data_collator = self.data_collator  # type: ignore
-
-        # Handle dataset preprocessing
-        if isinstance(train_dataset, datasets.Dataset):
-            # Add debug print before any modifications
-            if self.args.sample_packing and not self.args.pretraining:
-                train_dataset = train_dataset.remove_columns(["length"])
-            if not self.args.sample_packing or self.args.pretraining:
-                train_dataset = self._remove_unused_columns(
-                    train_dataset, description="training"
-                )
-        else:
-            self.data_collator = self._get_collator_with_removed_columns(  # pylint: disable=attribute-defined-outside-init
-                data_collator,
-                description="training",
-            )
-
-        # Get sampler and create dataloader
-        sampler = self._get_train_sampler()
-        dataloader = self._prepare_dataloader(train_dataset, sampler, is_eval=False)
-
-        return dataloader
-
-    def _generate_and_score_completions(
-        self, inputs: list[dict[str, torch.Tensor | Any]]
-    ) -> dict[str, torch.Tensor | Any]:
-        device = self.accelerator.device
-        mode = "eval" if self.control.should_evaluate else "train"
-
-        prompts = [x["prompt"] for x in inputs]
-        prompts_text = [
-            maybe_apply_chat_template(example, self.processing_class)["prompt"]
-            for example in inputs
-        ]
-        prompt_inputs = self.processing_class(
-            text=prompts_text,
-            return_tensors="pt",
-            padding=True,
-            padding_side="left",
-            add_special_tokens=False,
-        )
-        prompt_inputs = Trainer._prepare_inputs(self, prompt_inputs)
-        prompt_ids, prompt_mask = (
-            prompt_inputs["input_ids"],
-            prompt_inputs["attention_mask"],
-        )
-
-        if self.max_prompt_length is not None:
-            prompt_ids = prompt_ids[:, -self.max_prompt_length :]
-            prompt_mask = prompt_mask[:, -self.max_prompt_length :]
-
-        # Generate completions using either vLLM or regular generation
-        if self.args.use_vllm:
-            # First, have main process load weights if needed
-            # pylint: disable=access-member-before-definition
-            if self.state.global_step != self._last_loaded_step:  # type: ignore[has-type]
-                self._move_model_to_vllm()
-                # pylint: disable=attribute-defined-outside-init
-                self._last_loaded_step = self.state.global_step
-
-            # Generate completions using vLLM: gather all prompts and use them in a single call in the main process
-            all_prompts_text = gather_object(prompts_text)
                     if self.accelerator.is_main_process:
-                if self.args.sequence_parallel_degree > 1:
+                        self.vllm_client.update_named_param(name, param.data)
-                    # Calculate sequence parallel group information
-                    world_size = self.accelerator.num_processes
-                    sequence_parallel_degree = self.args.sequence_parallel_degree
-                    num_sp_groups = world_size // sequence_parallel_degree
 
-                    # Since processes in the same SP group have the same prompts, we need to ensure
+                # Unmerge adapters while parameters are still gathered
-                    # we only take one copy of each prompt from each SP group
+                self.model.unmerge_adapter()
-                    ordered_set_of_prompts = []
+                # Parameters will automatically be repartitioned when exiting the context
-                    for sp_group_id in range(num_sp_groups):
-                        # Get the first process from each SP group (typically the group leader)
-                        group_leader_rank = sp_group_id * sequence_parallel_degree
-
-                        # Extract prompts from this SP group, accounting for num_generations duplicates
-                        # We only need prompts from one rank in each SP group
-                        group_prompts = all_prompts_text[
-                            group_leader_rank
-                            * len(prompts_text) : (group_leader_rank + 1)
-                            * len(prompts_text) : self.num_generations
-                        ]
-
-                        ordered_set_of_prompts.extend(group_prompts)
         else:
-                    # Since 'prompts' contains 'num_generations' duplicates, we first take unique prompts, and generate
+            # For non-PEFT models, simply gather and update each parameter individually.
-                    # num_generations outputs for each one. This is faster than generating outputs for each duplicate
+            for name, param in self.model.named_parameters():
-                    # prompt individually.
+                with gather_if_zero3([param]):
-                    ordered_set_of_prompts = all_prompts_text[
+                    if self.accelerator.is_main_process:
-                        :: self.num_generations * self.args.sequence_parallel_degree
+                        self.vllm_client.update_named_param(name, param.data)
-                    ]
 
-                with profiling_context(self, "vLLM.generate"):
+        # Reset cache on main process
-                    completion_ids = self.vllm_client.generate(
+        if self.accelerator.is_main_process:
-                        prompts=ordered_set_of_prompts,
+            self.vllm_client.reset_prefix_cache()
-                        n=self.num_generations,
-                        repetition_penalty=self.repetition_penalty,
-                        temperature=self.temperature,
-                        top_p=self.top_p,
-                        top_k=-1 if self.top_k is None else self.top_k,
-                        min_p=0.0 if self.min_p is None else self.min_p,
-                        max_tokens=self.max_completion_length,
-                        guided_decoding_regex=self.guided_decoding_regex,
-                    )
-            else:
-                completion_ids = [None] * (
-                    len(all_prompts_text) // self.args.sequence_parallel_degree
-                )
-
-            # Broadcast the completions from the main process to all processes
-            completion_ids = broadcast_object_list(completion_ids, from_process=0)
-
-            # Determine the appropriate slice based on sequence parallelism
-            if self.args.sequence_parallel_degree > 1:
-                # Calculate SP group ID (which group of ranks this rank belongs to)
-                sp_group_id = self.accelerator.process_index // self.local_world_size
-
-                # Calculate the start index for this SP group
-                sp_group_start = sp_group_id * len(prompts) * self.local_world_size
-
-                # All ranks in the same SP group get the same data slice
-                process_slice = slice(
-                    sp_group_start,
-                    sp_group_start + len(prompts),
-                )
-                completion_ids = completion_ids[process_slice]
-            else:
-                # Original behavior for non-sequence parallel case
-                process_slice = slice(
-                    self.accelerator.process_index * len(prompts),
-                    (self.accelerator.process_index + 1) * len(prompts),
-                )
-                completion_ids = completion_ids[process_slice]
-
-            # Pad the completions, and concatenate them with the prompts
-            completion_ids = [
-                torch.tensor(ids, device=device) for ids in completion_ids
-            ]
-            completion_ids = pad(
-                completion_ids, padding_value=self.processing_class.pad_token_id
-            )
-            prompt_completion_ids = torch.cat([prompt_ids, completion_ids], dim=1)
-        else:
-            # Regular generation path
-            with unwrap_model_for_generation(
-                self.model_wrapped,
-                self.accelerator,
-                gather_deepspeed3_params=self.args.ds3_gather_for_generation,
-            ) as unwrapped_model:
-                prompt_completion_ids = unwrapped_model.generate(
-                    prompt_ids,
-                    attention_mask=prompt_mask,
-                    generation_config=self.generation_config,
-                )
-
-            # Compute prompt length and extract completion ids
-            prompt_length = prompt_ids.size(1)
-            prompt_ids = prompt_completion_ids[:, :prompt_length]
-            completion_ids = prompt_completion_ids[:, prompt_length:]
-
-        # Mask everything after the first EOS token
-        is_eos = completion_ids == self.processing_class.eos_token_id
-        eos_idx = torch.full(
-            (is_eos.size(0),), is_eos.size(1), dtype=torch.long, device=device
-        )
-        eos_idx[is_eos.any(dim=1)] = is_eos.int().argmax(dim=1)[is_eos.any(dim=1)]
-        sequence_indices = torch.arange(is_eos.size(1), device=device).expand(
-            is_eos.size(0), -1
-        )
-        completion_mask = (sequence_indices <= eos_idx.unsqueeze(1)).int()
-
-        # If mask_truncated_completions is enabled, zero out truncated completions in completion_mask
-        if self.args.mask_truncated_completions:
-            truncated_completions = ~is_eos.any(dim=1)
-            completion_mask = (
-                completion_mask * (~truncated_completions).unsqueeze(1).int()
-            )
-
-        # Concatenate prompt_mask with completion_mask for logit computation
-        attention_mask = torch.cat([prompt_mask, completion_mask], dim=1)  # (B, P+C)
-
-        logits_to_keep = completion_ids.size(
-            1
-        )  # we only need to compute the logits for the completion tokens
-        batch_size = (
-            self.args.per_device_train_batch_size
-            if mode == "train"
-            else self.args.per_device_eval_batch_size
-        )
-
-        with torch.no_grad():
-            # When using num_iterations == 1, old_per_token_logps == per_token_logps, so we can skip it's
-            # computation here, and use per_token_logps.detach() instead.
-            if self.num_iterations > 1:
-                old_per_token_logps = self._get_per_token_logps(
-                    self.model,
-                    prompt_completion_ids,
-                    attention_mask,
-                    logits_to_keep,
-                    batch_size,
-                )
-            else:
-                old_per_token_logps = None
-
-            if self.beta == 0.0:
-                ref_per_token_logps = None
-            elif self.ref_model is not None:
-                ref_per_token_logps = self._get_per_token_logps(
-                    self.ref_model,
-                    prompt_completion_ids,
-                    attention_mask,
-                    logits_to_keep,
-                    batch_size,
-                )
-            else:
-                with self.accelerator.unwrap_model(self.model).disable_adapter():
-                    ref_per_token_logps = self._get_per_token_logps(
-                        self.model,
-                        prompt_completion_ids,
-                        attention_mask,
-                        logits_to_keep,
-                        batch_size,
-                    )
-
-        # Decode the generated completions
-        completions_text = self.processing_class.batch_decode(
-            completion_ids, skip_special_tokens=True
-        )
-        if is_conversational(inputs[0]):
-            completions = []
-            for prompt, completion in zip(prompts, completions_text):
-                bootstrap = (
-                    prompt.pop()["content"] if prompt[-1]["role"] == "assistant" else ""
-                )
-                completions.append(
-                    [{"role": "assistant", "content": bootstrap + completion}]
-                )
-        else:
-            completions = completions_text
-
-        rewards_per_func = torch.zeros(
-            len(prompts), len(self.reward_funcs), device=device
-        )
-        for i, (reward_func, reward_processing_class, reward_func_name) in enumerate(
-            zip(
-                self.reward_funcs,
-                self.reward_processing_classes,
-                self.reward_func_names,
-            )
-        ):
-            with profiling_context(self, reward_func_name):
-                if isinstance(
-                    reward_func, nn.Module
-                ):  # Module instead of PretrainedModel for compat with compiled models
-                    if is_conversational(inputs[0]):
-                        messages = [
-                            {"messages": p + c} for p, c in zip(prompts, completions)
-                        ]
-                        texts = [
-                            apply_chat_template(x, reward_processing_class)["text"]
-                            for x in messages
-                        ]
-                    else:
-                        texts = [p + c for p, c in zip(prompts, completions)]
-                    reward_inputs = reward_processing_class(
-                        text=texts,
-                        return_tensors="pt",
-                        padding=True,
-                        padding_side="right",
-                        add_special_tokens=False,
-                    )
-                    reward_inputs = Trainer._prepare_inputs(self, reward_inputs)
-                    with torch.inference_mode():
-                        rewards_per_func[:, i] = reward_func(**reward_inputs).logits[
-                            :, 0
-                        ]  # Shape (B*G,)
-                else:
-                    # Repeat all input columns (but "prompt" and "completion") to match the number of generations
-                    keys = [
-                        key for key in inputs[0] if key not in ["prompt", "completion"]
-                    ]
-                    reward_kwargs = {
-                        key: [example[key] for example in inputs] for key in keys
-                    }
-                    output_reward_func = reward_func(
-                        prompts=prompts, completions=completions, **reward_kwargs
-                    )
-                    # Convert None values to NaN
-                    output_reward_func = [
-                        reward if reward is not None else torch.nan
-                        for reward in output_reward_func
-                    ]
-
-                    rewards_per_func[:, i] = torch.tensor(
-                        output_reward_func, dtype=torch.float32, device=device
-                    )
-
-        # If all reward functions return None for a given row, issue a detailed warning
-        if torch.isnan(rewards_per_func).all(dim=1).any():
-            nan_row_idx = (
-                torch.isnan(rewards_per_func).all(dim=1).nonzero(as_tuple=True)[0][0]
-            )
-            row_reward_kwargs = {
-                key: value[nan_row_idx] for key, value in reward_kwargs.items()
-            }
-            row_reward_kwargs["prompt"] = prompts[nan_row_idx]
-            row_reward_kwargs["completion"] = completions[nan_row_idx]
-            warnings.warn(
-                f"All reward functions returned None for the following kwargs: {row_reward_kwargs}. "
-                "Please ensure that at least one reward function returns a valid reward."
-            )
-
-        # Gather the reward per function: this part is crucial, because the rewards are normalized per group and the
-        # completions may be distributed across processes
-        rewards_per_func = gather(rewards_per_func)
-
-        # Apply weights to each reward function's output and sum
-        rewards = (
-            rewards_per_func * self.reward_weights.to(device).unsqueeze(0)
-        ).nansum(dim=1)
-
-        # Compute grouped-wise rewards
-        mean_grouped_rewards = rewards.view(-1, self.num_generations).mean(dim=1)
-        std_grouped_rewards = rewards.view(-1, self.num_generations).std(dim=1)
-
-        # Normalize the rewards to compute the advantages
-        mean_grouped_rewards = mean_grouped_rewards.repeat_interleave(
-            self.num_generations, dim=0
-        )
-        std_grouped_rewards = std_grouped_rewards.repeat_interleave(
-            self.num_generations, dim=0
-        )
-        advantages = rewards - mean_grouped_rewards
-        if self.args.scale_rewards:
-            advantages = advantages / (std_grouped_rewards + 1e-4)
-
-        # Slice to keep only the local part of the data
-        if self.args.sequence_parallel_degree > 1:
-            # Calculate SP group ID (which group of ranks this rank belongs to)
-            sp_group_id = self.accelerator.process_index // self.local_world_size
-
-            # Calculate the start index for this SP group
-            sp_group_start = sp_group_id * len(prompts) * self.local_world_size
-
-            # All ranks in the same SP group get the same data slice
-            process_slice = slice(
-                sp_group_start,
-                sp_group_start + len(prompts),
-            )
-        else:
-            # Original behavior for non-sequence parallel case
-            process_slice = slice(
-                self.accelerator.process_index * len(prompts),
-                (self.accelerator.process_index + 1) * len(prompts),
-            )
-        advantages = advantages[process_slice]
-
-        # Log the metrics
-        if mode == "train":
-            self._total_train_tokens += (
-                self.accelerator.gather_for_metrics(attention_mask.sum()).sum().item()
-            )
-        self._metrics[mode]["num_tokens"] = [self._total_train_tokens]
-
-        # log completion lengths, mean, min, max
-        agg_completion_mask = self.accelerator.gather_for_metrics(
-            completion_mask.sum(1)
-        )
-        self._metrics[mode]["completions/mean_length"].append(
-            agg_completion_mask.float().mean().item()
-        )
-        self._metrics[mode]["completions/min_length"].append(
-            agg_completion_mask.float().min().item()
-        )
-        self._metrics[mode]["completions/max_length"].append(
-            agg_completion_mask.float().max().item()
-        )
-
-        # identify sequences that terminated with EOS and log their lengths
-        agg_terminated_with_eos = self.accelerator.gather_for_metrics(is_eos.any(dim=1))
-        term_completion_mask = agg_completion_mask[agg_terminated_with_eos]
-        clipped_completions_ratio = 1 - len(term_completion_mask) / len(
-            agg_completion_mask
-        )
-        self._metrics[mode]["completions/clipped_ratio"].append(
-            clipped_completions_ratio
-        )
-        if len(term_completion_mask) == 0:
-            # edge case where no completed sequences are found
-            term_completion_mask = torch.zeros(1, device=device)
-        self._metrics[mode]["completions/mean_terminated_length"].append(
-            term_completion_mask.float().mean().item()
-        )
-        self._metrics[mode]["completions/min_terminated_length"].append(
-            term_completion_mask.float().min().item()
-        )
-        self._metrics[mode]["completions/max_terminated_length"].append(
-            term_completion_mask.float().max().item()
-        )
-
-        # Calculate mean reward per function, but only for samples where the function was applied (non-NaN values)
-        for i, reward_func_name in enumerate(self.reward_func_names):
-            mean_rewards = torch.nanmean(rewards_per_func[:, i]).item()
-            self._metrics[mode][f"rewards/{reward_func_name}/mean"].append(mean_rewards)
-            std_rewards = nanstd(rewards_per_func[:, i]).item()
-            self._metrics[mode][f"rewards/{reward_func_name}/std"].append(std_rewards)
-        self._metrics[mode]["reward"].append(mean_grouped_rewards.mean().item())
-        self._metrics[mode]["reward_std"].append(std_grouped_rewards.mean().item())
-
-        # Log prompt and completion texts
-        self._textual_logs["prompt"].extend(gather_object(prompts_text))
-        self._textual_logs["completion"].extend(gather_object(completions_text))
-        for i, name in enumerate(self.reward_func_names):
-            self._textual_logs["rewards"][name].extend(rewards_per_func[:, i].tolist())
-
-        return {
-            "prompt_ids": prompt_ids,
-            "prompt_mask": prompt_mask,
-            "completion_ids": completion_ids,
-            "completion_mask": completion_mask,
-            "advantages": advantages,
-            "old_per_token_logps": old_per_token_logps,
-            "ref_per_token_logps": ref_per_token_logps,
-        }

src/axolotl/core/trainers/mixins/__init__.py

@@ -6,4 +6,4 @@
 from .optimizer import OptimizerMixin
 from .rng_state_loader import RngLoaderMixin
 from .scheduler import SchedulerMixin
-from .sequence_parallel import SequenceParallelMixin
+from .sequence_parallel import SequenceParallelContextManager, SequenceParallelMixin

src/axolotl/core/trainers/mixins/sequence_parallel.py

@@ -1,13 +1,85 @@
-"""Module for Axolotl trainer sequence parallelism mixin"""
+"""
+Module for Axolotl trainer sequence parallelism mixin and training context manager
+"""
 
+import functools
+import logging
+
+import torch
 import torch.distributed as dist
 from datasets import Dataset
+from torch import nn
 from torch.utils.data import DistributedSampler, Sampler
+from torch.utils.hooks import RemovableHandle
 
 from axolotl.monkeypatch.attention.ring_attn import (
+    RingAttnFunc,
     get_ring_attn_group,
+    update_ring_attn_params,
 )
 
+LOG = logging.getLogger(__name__)
+
+
+def apply_sequence_parallelism(
+    batch: dict[str, torch.Tensor],
+    local_rank: int,
+    local_world_size: int,
+    ring_attn_func: RingAttnFunc,
+) -> dict[str, torch.Tensor]:
+    """
+    Apply sequence parallelism slicing to a batch.
+
+    Args:
+        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
+        ring_attn_func: The ring attention function to use
+
+    Returns:
+        Sliced batch dictionary.
+    """
+    # Update ring attention params if needed
+    if batch.get("position_ids") is not None:
+        update_ring_attn_params(position_ids=batch["position_ids"])
+
+    # Slice batch for sequence parallel processing
+    total_seq_len = batch["input_ids"].size(1)
+    for key in batch:
+        if (
+            key in batch
+            and isinstance(batch[key], torch.Tensor)
+            and batch[key].dim() > 1
+            and batch[key].size(1) == total_seq_len
+        ):
+
+            if ring_attn_func in [
+                RingAttnFunc.VARLEN_LLAMA3,
+                RingAttnFunc.BATCH_RING,
+            ]:
+                # Split in sequential fashion and grab this rank's chunk
+                batch[key] = (
+                    batch[key].chunk(local_world_size, dim=1)[local_rank].contiguous()
+                )
+            elif ring_attn_func is RingAttnFunc.BATCH_ZIGZAG:
+                chunks = batch[key].chunk(2 * local_world_size, dim=1)
+
+                # Take rank's chunk and opposing chunk for zigzag pattern
+                selected_chunks = [
+                    chunks[local_rank],
+                    chunks[2 * local_world_size - local_rank - 1],
+                ]
+                batch[key] = torch.cat(selected_chunks, dim=1).contiguous()
+            elif ring_attn_func is RingAttnFunc.BATCH_STRIPE:
+                # Split into striped data and stack
+                tensor = torch.stack(
+                    batch[key].split(local_world_size, dim=1),
+                    dim=1,
+                ).transpose(1, 2)
+                batch[key] = tensor[:, local_rank].contiguous()
+
+    return batch
+
 
 class SequenceParallelMixin:
     """
@@ -85,3 +157,157 @@ class SequenceParallelMixin:
         return self._create_sequence_parallel_sampler(
             eval_dataset, shuffle=False, is_eval=True
         )
+
+
+class SequenceParallelContextManager:
+    """
+    Context manager for sequence parallelism operations.
+
+    This class provides a context that will automatically apply sequence parallelism
+    during model forward passes using a pre-forward hook, and gather outputs from
+    across the sequence parallelism group using a post-forward hook.
+    """
+
+    def __init__(
+        self,
+        model: nn.Module,
+        sequence_parallel_degree: int,
+        ring_attn_func: RingAttnFunc,
+    ):
+        self.model = model
+        self.sequence_parallel_degree = sequence_parallel_degree
+        self.ring_attn_func = ring_attn_func
+        self.process_group = get_ring_attn_group()
+
+        # Initialize sequence parallel group details
+        self.local_rank = dist.get_rank(self.process_group)
+        self.local_world_size = dist.get_world_size(self.process_group)
+
+        # Will store hook handles for removal
+        self.hook_handles: list[RemovableHandle] = []
+
+        # Create a partially applied version of the apply_sequence_parallelism function
+        # with pre-configured params
+        self.apply_sequence_parallelism = functools.partial(
+            apply_sequence_parallelism,
+            local_rank=self.local_rank,
+            local_world_size=self.local_world_size,
+            ring_attn_func=self.ring_attn_func,
+        )
+
+    def __enter__(self):
+        # Forward pre-hook to apply sequence parallelism
+        def sequence_parallel_pre_hook(_, args, kwargs):
+            # Apply sequence parallelism to kwargs
+            kwargs = self.apply_sequence_parallelism(batch=kwargs)
+            return args, kwargs
+
+        # Forward post-hook to gather outputs
+        def sequence_parallel_post_hook(_, __, output):
+            # Gather the sharded outputs
+            return self.gather_outputs(output)
+
+        # Register both hooks
+        self.hook_handles.append(
+            self.model.register_forward_pre_hook(
+                sequence_parallel_pre_hook, with_kwargs=True
+            )
+        )
+        self.hook_handles.append(
+            self.model.register_forward_hook(sequence_parallel_post_hook)
+        )
+
+        return self
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        # Remove all hooks
+        for handle in self.hook_handles:
+            handle.remove()
+        self.hook_handles = []
+
+    def gather_outputs(self, output):
+        """Gather sharded outputs from all ranks and reconstruct the full tensor."""
+        # Handle different output formats (dict, tensor, etc.)
+        if isinstance(output, dict):
+            gathered_output = {}
+            for key, value in output.items():
+                if isinstance(value, torch.Tensor) and value.dim() > 1:
+                    # Gather logits or other sequence-sharded tensors
+                    gathered_value = self.gather_tensor(value)
+                    gathered_output[key] = gathered_value
+                else:
+                    gathered_value = value.clone()
+                    dist.all_reduce(
+                        gathered_value, op=dist.ReduceOp.SUM, group=self.process_group
+                    )
+                    gathered_output[key] = gathered_value
+            return gathered_output
+        if isinstance(output, torch.Tensor):
+            return self.gather_tensor(output)
+
+        return output
+
+    def gather_tensor(self, tensor):
+        """Gather a sharded tensor from all ranks."""
+        # Prepare tensors for all_gather
+        world_size = self.local_world_size
+
+        # Create list to store tensors from all ranks
+        gathered_tensors = [torch.zeros_like(tensor) for _ in range(world_size)]
+
+        # All-gather operation
+        dist.all_gather(gathered_tensors, tensor, group=self.process_group)
+
+        # Concatenate along sequence dimension (typically dim=1)
+        if self.ring_attn_func in [RingAttnFunc.VARLEN_LLAMA3, RingAttnFunc.BATCH_RING]:
+            # Simple concatenation for standard sharding
+            return torch.cat(gathered_tensors, dim=1)
+
+        if self.ring_attn_func is RingAttnFunc.BATCH_ZIGZAG:
+            # Each rank has a pattern of (rank, world_size*2-rank-1)
+            reconstituted_tensors = [None] * (world_size * 2)
+
+            # First, split each gathered tensor into its two chunks
+            for rank, gathered_tensor in enumerate(gathered_tensors):
+                # Each tensor contains two chunks in the sequence dimension
+                chunk_size = gathered_tensor.size(1) // 2
+                chunk1, chunk2 = gathered_tensor.split(chunk_size, dim=1)
+
+                # Place chunks in their original positions
+                reconstituted_tensors[rank] = chunk1
+                reconstituted_tensors[world_size * 2 - rank - 1] = chunk2
+
+            # Concatenate the reconstituted tensors in the correct order
+            return torch.cat(reconstituted_tensors, dim=1)
+
+        # Otherwise, RingAttnFunc.BATCH_STRIPE
+        # In striping, each rank has every world_size-th slice
+        batch_size = tensor.size(0)
+        hidden_dim = tensor.size(-1)
+
+        # First, determine the full sequence length
+        total_seq_len = 0
+        for t in gathered_tensors:
+            total_seq_len += t.size(1)
+
+        # Create a tensor to hold the unstriped result
+        result = torch.zeros(
+            batch_size,
+            total_seq_len,
+            hidden_dim,
+            dtype=tensor.dtype,
+            device=tensor.device,
+        )
+
+        # For each rank's tensor, distribute its slices to the correct positions
+        for rank, gathered_tensor in enumerate(gathered_tensors):
+            # The rank's tensor contains every world_size-th slice
+            # starting from its rank position
+            seq_len = gathered_tensor.size(1)
+            for i in range(seq_len):
+                # Calculate the position in the full tensor
+                pos = i * world_size + rank
+                if pos < total_seq_len:
+                    result[:, pos] = gathered_tensor[:, i]
+
+        return result

src/axolotl/core/training_args.py

@@ -9,7 +9,7 @@ from PIL.Image import Resampling
 from transformers import TrainingArguments
 from trl import CPOConfig, KTOConfig, ORPOConfig, PRMConfig, RewardConfig
 
-from axolotl.utils.schemas.enums import RingAttnFunc
+from axolotl.monkeypatch.attention.ring_attn.patch import RingAttnFunc
 
 
 @dataclass
@@ -227,19 +227,6 @@ 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/integrations/cut_cross_entropy/monkeypatch/cohere.py

@@ -20,15 +20,25 @@ from cut_cross_entropy.transformers.utils import (
 from transformers.cache_utils import Cache
 from transformers.modeling_outputs import CausalLMOutputWithPast
 from transformers.models.cohere.modeling_cohere import (
+    _CONFIG_FOR_DOC,
+    COHERE_INPUTS_DOCSTRING,
     KwargsForCausalLM,
 )
 from transformers.processing_utils import Unpack
+from transformers.utils import (
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
 from transformers.utils.deprecation import deprecate_kwarg
 
 _PATCH_OPTS: PatchOptions | None = None
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+@add_start_docstrings_to_model_forward(COHERE_INPUTS_DOCSTRING)
+@replace_return_docstrings(
+    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
+)
 def cce_forward(
     self,
     input_ids: torch.LongTensor | None = None,

src/axolotl/integrations/cut_cross_entropy/monkeypatch/gemma.py

@@ -17,15 +17,25 @@ from cut_cross_entropy.transformers.utils import (
 from transformers.cache_utils import Cache
 from transformers.modeling_outputs import CausalLMOutputWithPast
 from transformers.models.gemma.modeling_gemma import (
+    _CONFIG_FOR_DOC,
+    GEMMA_INPUTS_DOCSTRING,
     KwargsForCausalLM,
 )
 from transformers.processing_utils import Unpack
+from transformers.utils import (
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
 from transformers.utils.deprecation import deprecate_kwarg
 
 _PATCH_OPTS: PatchOptions | None = None
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+@add_start_docstrings_to_model_forward(GEMMA_INPUTS_DOCSTRING)
+@replace_return_docstrings(
+    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
+)
 def cce_forward(
     self,
     input_ids: torch.LongTensor | None = None,

src/axolotl/integrations/cut_cross_entropy/monkeypatch/gemma3.py

@@ -20,11 +20,15 @@ from torch import nn
 from transformers.cache_utils import Cache, HybridCache
 from transformers.modeling_outputs import CausalLMOutputWithPast
 from transformers.models.gemma3.modeling_gemma3 import (
+    _CONFIG_FOR_DOC,
+    GEMMA3_INPUTS_DOCSTRING,
     Gemma3CausalLMOutputWithPast,
     logger,
 )
 from transformers.utils import (
+    add_start_docstrings_to_model_forward,
     is_torchdynamo_compiling,
+    replace_return_docstrings,
 )
 from transformers.utils.deprecation import deprecate_kwarg
 
@@ -34,6 +38,10 @@ _PATCH_OPTS: PatchOptions | None = None
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+@add_start_docstrings_to_model_forward(GEMMA3_INPUTS_DOCSTRING)
+@replace_return_docstrings(
+    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
+)
 def cce_forward(
     self,
     input_ids: torch.LongTensor | None = None,
@@ -162,6 +170,10 @@ def cce_forward(
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+@add_start_docstrings_to_model_forward(GEMMA3_INPUTS_DOCSTRING)
+@replace_return_docstrings(
+    output_type=Gemma3CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
+)
 def cce_forward_multimodal(
     self,
     input_ids: torch.LongTensor | None = None,

src/axolotl/integrations/cut_cross_entropy/monkeypatch/llama.py

@@ -19,9 +19,15 @@ from transformers.modeling_outputs import (
     CausalLMOutputWithPast,
 )
 from transformers.models.llama.modeling_llama import (
+    _CONFIG_FOR_DOC,
+    LLAMA_INPUTS_DOCSTRING,
     KwargsForCausalLM,
 )
 from transformers.processing_utils import Unpack
+from transformers.utils import (
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
 from transformers.utils.deprecation import deprecate_kwarg
 from transformers.utils.generic import can_return_tuple
 
@@ -30,6 +36,10 @@ _PATCH_OPTS: PatchOptions | None = None
 
 @can_return_tuple
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+@add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
+@replace_return_docstrings(
+    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
+)
 def cce_forward(
     self,
     input_ids: Optional[torch.LongTensor] = None,

src/axolotl/integrations/cut_cross_entropy/monkeypatch/llama4.py

@@ -16,12 +16,22 @@ from torch import nn
 from transformers.cache_utils import Cache
 from transformers.modeling_outputs import CausalLMOutputWithPast
 from transformers.models.llama4.modeling_llama4 import (
+    _CONFIG_FOR_DOC,
+    LLAMA4_INPUTS_DOCSTRING,
     Llama4CausalLMOutputWithPast,
 )
+from transformers.utils import (
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
 
 _PATCH_OPTS: PatchOptions | None = None
 
 
+@add_start_docstrings_to_model_forward(LLAMA4_INPUTS_DOCSTRING)
+@replace_return_docstrings(
+    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
+)
 def cce_forward(
     self,
     input_ids: torch.LongTensor | None = None,
@@ -150,6 +160,9 @@ def cce_forward(
     )
 
 
+@replace_return_docstrings(
+    output_type=Llama4CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
+)
 def cce_forward_multimodal(
     self,
     input_ids: torch.LongTensor | None = None,  # type: ignore

src/axolotl/integrations/cut_cross_entropy/monkeypatch/mistral3.py

@@ -19,11 +19,15 @@ from transformers.models.mistral3.modeling_mistral3 import (
     Mistral3CausalLMOutputWithPast,
 )
 from transformers.models.mistral.modeling_mistral import (
+    _CONFIG_FOR_DOC,
+    MISTRAL_INPUTS_DOCSTRING,
     KwargsForCausalLM,
 )
 from transformers.processing_utils import Unpack
 from transformers.utils import (
+    add_start_docstrings_to_model_forward,
     is_torchdynamo_compiling,
+    replace_return_docstrings,
 )
 from transformers.utils.deprecation import deprecate_kwarg
 
@@ -31,6 +35,10 @@ _PATCH_OPTS: PatchOptions | None = None
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+@add_start_docstrings_to_model_forward(MISTRAL_INPUTS_DOCSTRING)
+@replace_return_docstrings(
+    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
+)
 def cce_forward(
     self,
     input_ids: torch.LongTensor | None = None,

src/axolotl/integrations/cut_cross_entropy/monkeypatch/qwen2_moe.py

@@ -13,10 +13,16 @@ from cut_cross_entropy.transformers.utils import (
     apply_lce,
 )
 from transformers.models.qwen2_moe.modeling_qwen2_moe import (
+    _CONFIG_FOR_DOC,
+    QWEN2MOE_INPUTS_DOCSTRING,
     MoeCausalLMOutputWithPast,
     MoeModelOutputWithPast,
     load_balancing_loss_func,
 )
+from transformers.utils import (
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
 from transformers.utils.deprecation import deprecate_kwarg
 from transformers.utils.generic import can_return_tuple
 
@@ -25,6 +31,10 @@ _PATCH_OPTS: PatchOptions | None = None
 
 @can_return_tuple
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+@add_start_docstrings_to_model_forward(QWEN2MOE_INPUTS_DOCSTRING)
+@replace_return_docstrings(
+    output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
+)
 def forward(
     self,
     input_ids: Optional[torch.LongTensor] = None,

src/axolotl/integrations/cut_cross_entropy/monkeypatch/qwen2_vl.py

@@ -14,12 +14,22 @@ from cut_cross_entropy.transformers.utils import (
 )
 from torch.nn import CrossEntropyLoss
 from transformers.models.qwen2_vl.modeling_qwen2_vl import (
+    _CONFIG_FOR_DOC,
+    QWEN2_VL_INPUTS_DOCSTRING,
     Qwen2VLCausalLMOutputWithPast,
 )
+from transformers.utils import (
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
 
 _PATCH_OPTS: PatchOptions | None = None
 
 
+@add_start_docstrings_to_model_forward(QWEN2_VL_INPUTS_DOCSTRING)
+@replace_return_docstrings(
+    output_type=Qwen2VLCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
+)
 def cce_forward_multimodal(
     self,
     input_ids: Optional[torch.LongTensor] = None,

src/axolotl/integrations/cut_cross_entropy/monkeypatch/qwen3_moe.py

@@ -12,13 +12,20 @@ from cut_cross_entropy.transformers.utils import (
     TransformersModelT,
     apply_lce,
 )
+from transformers.modeling_outputs import CausalLMOutputWithPast
 from transformers.models.qwen3_moe.modeling_qwen3_moe import (
+    _CONFIG_FOR_DOC,
+    QWEN3_MOE_INPUTS_DOCSTRING,
     KwargsForCausalLM,
     MoeCausalLMOutputWithPast,
     MoeModelOutputWithPast,
     load_balancing_loss_func,
 )
 from transformers.processing_utils import Unpack
+from transformers.utils import (
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
 from transformers.utils.deprecation import deprecate_kwarg
 from transformers.utils.generic import can_return_tuple
 
@@ -27,6 +34,10 @@ _PATCH_OPTS: PatchOptions | None = None
 
 @can_return_tuple
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+@add_start_docstrings_to_model_forward(QWEN3_MOE_INPUTS_DOCSTRING)
+@replace_return_docstrings(
+    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
+)
 def forward(
     self,
     input_ids: Optional[torch.LongTensor] = None,

src/axolotl/integrations/kd/trainer.py

@@ -74,9 +74,6 @@ class AxolotlKDTrainer(AxolotlTrainer):
         target_token_ids_for_loss = target_token_ids[..., 1:, :].contiguous()
         target_mask_for_loss = target_mask[..., 1:, :].contiguous()
 
-        if num_items_in_batch is None:
-            num_items_in_batch = -1
-
         if self.args.kd_zscore_base_temp:
             loss_kd = topk_kd_loss_with_zscore(
                 shift_logits,

src/axolotl/integrations/liger/models/deepseekv2.py

@@ -14,6 +14,10 @@ from torch.nn import CrossEntropyLoss
 from transformers.modeling_outputs import CausalLMOutputWithPast
 
 
+# @add_start_docstrings_to_model_forward(DeepseekV2_INPUTS_DOCSTRING)
+# @replace_return_docstrings(
+#    output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
+# )
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,

src/axolotl/integrations/liger/models/jamba.py

@@ -13,11 +13,21 @@ from liger_kernel.transformers.fused_linear_cross_entropy import (
 from torch.nn import CrossEntropyLoss
 from transformers.modeling_outputs import MoeCausalLMOutputWithPast
 from transformers.models.jamba.modeling_jamba import (
+    _CONFIG_FOR_DOC,
+    JAMBA_INPUTS_DOCSTRING,
     HybridMambaAttentionDynamicCache,
     load_balancing_loss_func,
 )
+from transformers.utils import (
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
 
 
+@add_start_docstrings_to_model_forward(JAMBA_INPUTS_DOCSTRING)
+@replace_return_docstrings(
+    output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
+)
 def lce_forward(
     self,
     input_ids: torch.LongTensor = None,

src/axolotl/monkeypatch/attention/ring_attn/__init__.py

@@ -4,6 +4,7 @@
 # flake8: noqa
 
 from .patch import (
+    RingAttnFunc,
     get_ring_attn_group,
     register_ring_attn,
     set_ring_attn_group,

src/axolotl/monkeypatch/attention/ring_attn/adapters/batch.py

@@ -16,7 +16,11 @@ import torch
 import torch.distributed as dist
 import transformers
 import transformers.modeling_flash_attention_utils
-from ring_flash_attn import ring_flash_attn_func
+from ring_flash_attn import (
+    ring_flash_attn_func,
+    stripe_flash_attn_func,
+    zigzag_ring_flash_attn_func,
+)
 from ring_flash_attn.adapters.hf_adapter import check_params
 from transformers.modeling_flash_attention_utils import (
     _flash_supports_window_size,
@@ -24,12 +28,12 @@ from transformers.modeling_flash_attention_utils import (
 )
 from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
 
-from axolotl.utils.schemas.enums import RingAttnFunc
+from axolotl.monkeypatch.attention.ring_attn.patch import RingAttnFunc
 
 RING_ATTN_FUNC_MAPPING = {
-    RingAttnFunc.BATCH_RING: torch.compile(ring_flash_attn_func),
+    RingAttnFunc.BATCH_RING: ring_flash_attn_func,
-    # RingAttnFunc.BATCH_ZIGZAG: torch.compile(zigzag_ring_flash_attn_func),
+    RingAttnFunc.BATCH_ZIGZAG: zigzag_ring_flash_attn_func,
-    # RingAttnFunc.BATCH_STRIPE: torch.compile(stripe_flash_attn_func),
+    RingAttnFunc.BATCH_STRIPE: stripe_flash_attn_func,
 }
 
 

src/axolotl/monkeypatch/attention/ring_attn/patch.py

@@ -6,12 +6,13 @@ package, specifically the `hf_adapter.substitute_hf_flash_attn` function to patc
 their sequence parallel version of Flash Attention 2.
 """
 
+from enum import Enum
+
 import torch
 import torch.distributed as dist
 from accelerate.logging import get_logger
 
 from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
-from axolotl.utils.schemas.enums import RingAttnFunc
 
 LOG = get_logger(__name__)
 
@@ -40,6 +41,17 @@ def set_ring_attn_group(ring_attn_group: dist.ProcessGroup | None):
     RING_ATTN_GROUP = ring_attn_group
 
 
+class RingAttnFunc(str, Enum):
+    """Enum class for supported `ring-flash-attn` implementations"""
+
+    # VARLEN_RING = "varlen_ring"
+    # VARLEN_ZIGZAG = "varlen_zigzag"
+    VARLEN_LLAMA3 = "varlen_llama3"
+    BATCH_RING = "batch_ring"
+    BATCH_ZIGZAG = "batch_zigzag"
+    BATCH_STRIPE = "batch_stripe"
+
+
 def register_ring_attn(
     sequence_parallel_degree: int,
     heads_k_stride: int | None,
@@ -105,7 +117,11 @@ def register_ring_attn(
         substitute_hf_flash_attn(
             process_group=get_ring_attn_group(), heads_k_stride=heads_k_stride or 1
         )
-    elif ring_attn_func is RingAttnFunc.BATCH_RING:
+    elif ring_attn_func in [
+        RingAttnFunc.BATCH_RING,
+        RingAttnFunc.BATCH_ZIGZAG,
+        RingAttnFunc.BATCH_STRIPE,
+    ]:
         from axolotl.monkeypatch.attention.ring_attn.adapters.batch import (
             substitute_hf_flash_attn,
         )

src/axolotl/monkeypatch/gemma3.py

@@ -7,16 +7,24 @@ from typing import Optional, Tuple, Union
 import torch
 from transformers.cache_utils import Cache
 from transformers.models.gemma3.modeling_gemma3 import (
+    _CONFIG_FOR_DOC,
+    GEMMA3_INPUTS_DOCSTRING,
     Gemma3CausalLMOutputWithPast,
     logger,
 )
 from transformers.utils import (
+    add_start_docstrings_to_model_forward,
     is_torchdynamo_compiling,
+    replace_return_docstrings,
 )
 from transformers.utils.deprecation import deprecate_kwarg
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+@add_start_docstrings_to_model_forward(GEMMA3_INPUTS_DOCSTRING)
+@replace_return_docstrings(
+    output_type=Gemma3CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
+)
 def new_forward(
     self,
     input_ids: torch.LongTensor = None,

src/axolotl/train.py

@@ -7,7 +7,7 @@ import os
 import signal
 import sys
 import weakref
-from contextlib import ExitStack
+from contextlib import nullcontext
 from pathlib import Path
 from typing import Any, Dict
 
@@ -26,14 +26,15 @@ from axolotl.common.datasets import TrainDatasetMeta
 from axolotl.contribs.lgpl import (  # pylint: disable = no-name-in-module
     fix_untrained_tokens,
 )
-from axolotl.core.trainer_builder import HFCausalTrainerBuilder, HFRLTrainerBuilder
+from axolotl.core.trainers.builders import HFCausalTrainerBuilder, HFRLTrainerBuilder
+from axolotl.core.trainers.mixins.sequence_parallel import (
+    SequenceParallelContextManager,
+)
 from axolotl.integrations.base import PluginManager
-from axolotl.utils.ctx_managers.sequence_parallel import SequenceParallelContextManager
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import cleanup_distributed
 from axolotl.utils.freeze import freeze_layers_except
 from axolotl.utils.models import load_model, load_processor, load_tokenizer
-from axolotl.utils.schemas.enums import RLType
 from axolotl.utils.trainer import setup_trainer
 
 try:
@@ -106,7 +107,7 @@ def setup_reference_model(
         Reference model if needed for RL training, `None` otherwise.
     """
     model_ref = None
-    if cfg.rl and cfg.rl != RLType.ORPO:
+    if cfg.rl and cfg.rl != "orpo":
         if cfg.adapter and not cfg.rl_adapter_ref_model:
             # use built-in trl autounwrap
             LOG.debug("Passing model_ref: None to RL trainer")
@@ -187,32 +188,28 @@ def execute_training(
         trainer: The configured trainer object.
         resume_from_checkpoint: Path to checkpoint to resume from, if applicable.
     """
-    with ExitStack() as stack:
     # Define the context managers to use
-        if cfg.flash_optimum:
+    flash_context = (
-            stack.enter_context(
         torch.backends.cuda.sdp_kernel(
             enable_flash=True,
             enable_math=True,
             enable_mem_efficient=True,
         )
+        if cfg.flash_optimum
+        else nullcontext()
     )
-
+    sequence_parallel_context = (
-        if cfg.sequence_parallel_degree > 1:
-            models = [trainer.model]
-            if hasattr(trainer, "ref_model"):
-                models.append(trainer.ref_model)
-
-            stack.enter_context(
         SequenceParallelContextManager(
-                    models=models,
+            model=trainer.model,
             sequence_parallel_degree=cfg.sequence_parallel_degree,
-                    gradient_accumulation_steps=cfg.gradient_accumulation_steps,
             ring_attn_func=cfg.ring_attn_func,
         )
+        if cfg.sequence_parallel_degree > 1
+        else nullcontext()
     )
 
     LOG.info("Starting trainer...")
+    with flash_context, sequence_parallel_context:
         trainer.train(resume_from_checkpoint=resume_from_checkpoint)
 
 

src/axolotl/utils/callbacks/__init__.py

@@ -46,11 +46,11 @@ from axolotl.utils.distributed import (
 from axolotl.utils.schemas.config import AxolotlInputConfig
 
 if TYPE_CHECKING:
-    from axolotl.core.trainer_builder import AxolotlTrainingArguments
+    from axolotl.core.training_args import AxolotlTrainingArguments
 
 
 IGNORE_INDEX = -100
-LOG = logging.getLogger("axolotl.callbacks")
+LOG = logging.getLogger(__name__)
 
 
 class EvalFirstStepCallback(

src/axolotl/utils/callbacks/mlflow_.py

@@ -1,7 +1,6 @@
 """MLFlow module for trainer callbacks"""
 
 import logging
-import os
 from shutil import copyfile
 from tempfile import NamedTemporaryFile
 from typing import TYPE_CHECKING
@@ -17,11 +16,6 @@ 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"""
@@ -38,7 +32,6 @@ class SaveAxolotlConfigtoMlflowCallback(TrainerCallback):
     ):
         if is_main_process():
             try:
-                if should_log_artifacts():
                 with NamedTemporaryFile(
                     mode="w", delete=False, suffix=".yml", prefix="axolotl_config_"
                 ) as temp_file:
@@ -47,10 +40,6 @@ class SaveAxolotlConfigtoMlflowCallback(TrainerCallback):
                     LOG.info(
                         "The Axolotl config has been saved to the MLflow artifacts."
                     )
-                else:
-                    LOG.info(
-                        "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}")
         return control

src/axolotl/utils/ctx_managers/__init__.py

@@ -1,6 +0,0 @@
-"""Init for context manager submodule"""
-
-# pylint: disable=unused-import
-# flake8: noqa
-
-from .sequence_parallel import SequenceParallelContextManager

src/axolotl/utils/ctx_managers/sequence_parallel.py

@@ -1,335 +0,0 @@
-"""Module for Axolotl trainer sequence parallelism manager and utilities"""
-
-import functools
-
-import torch
-import torch.distributed as dist
-from torch import nn
-from torch.utils.hooks import RemovableHandle
-from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.utils import ModelOutput
-
-from axolotl.monkeypatch.attention.ring_attn.patch import (
-    get_ring_attn_group,
-    update_ring_attn_params,
-)
-from axolotl.utils.schemas.enums import RingAttnFunc
-
-
-# TODO(djsaunde): implement zigzag, stripe patterns here (and elsewhere) in this
-# module. Currently, we just focus on batch ring and varlen llama3 for simplicity.
-def apply_sequence_parallelism(
-    batch: dict[str, torch.Tensor],
-    local_rank: int,
-    local_world_size: int,
-    gradient_accumulation_steps: int,
-    ring_attn_func: RingAttnFunc,  # pylint: disable=unused-argument
-) -> tuple[dict[str, torch.Tensor], int, int]:
-    """
-    Apply sequence parallelism slicing to a batch.
-
-    Special handling is implemented for integer logits_to_keep, which indicates
-    to only keep the last N tokens in the sequence during generation.
-
-    Args:
-        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.
-        ring_attn_func: Which ring attention function to use. Currently unused, but
-            related to above TODO.
-
-    Returns:
-        tuple of:
-            - Batch dictionary with sliced tensors.
-            - The original sequence length before padding.
-            - The number of padding tokens added.
-    """
-    original_seq_len = batch["input_ids"].size(1)
-
-    # Update ring attention params if needed
-    if batch.get("position_ids") is not None:
-        update_ring_attn_params(position_ids=batch["position_ids"])
-    else:
-        # If position_ids aren't already in the batch, create them
-        batch["position_ids"] = torch.arange(
-            0,
-            original_seq_len,
-            dtype=torch.long,
-            device=batch["input_ids"].device,
-        ).expand(batch["input_ids"].size(0), -1)
-
-    if "logits_to_keep" in batch and isinstance(batch["logits_to_keep"], int):
-        logits_to_keep = batch["logits_to_keep"]
-
-        # Calculate which positions in the full sequence contain the last N tokens
-        start_position = max(0, original_seq_len - logits_to_keep)
-        chunk_size = original_seq_len // local_world_size
-        rank_start = local_rank * chunk_size
-        rank_end = rank_start + chunk_size
-
-        # Create a boolean mask tensor for this rank's chunk
-        mask = torch.zeros(
-            chunk_size,
-            dtype=torch.bool,
-            device=batch["input_ids"].device,
-        )
-
-        if rank_end > start_position:
-            # Calculate how many of the last N tokens fall within this rank's range
-            tokens_in_rank = min(rank_end, original_seq_len) - max(
-                rank_start, start_position
-            )
-
-            # Calculate where these tokens start in the local chunk
-            local_start_idx = max(0, start_position - rank_start)
-
-            # Set the appropriate positions in the mask to True
-            mask[local_start_idx : local_start_idx + tokens_in_rank] = True
-
-        # Replace the integer with the boolean mask
-        batch["logits_to_keep"] = mask
-
-    # Add padding to make sequence length divisible by local_world_size
-    total_seq_len = original_seq_len
-    pad_len = 0
-    divisor = min(local_world_size, 64)
-    if total_seq_len % divisor != 0:
-        pad_len = divisor - (total_seq_len % divisor)
-
-        # Apply padding to all relevant tensors
-        for key in batch:
-            if (
-                isinstance(batch[key], torch.Tensor)
-                and batch[key].dim() > 1
-                and batch[key].size(1) == total_seq_len
-            ):
-                # Create padding tensor
-                pad_value = -100 if key == "labels" else 0
-                padding = torch.full(
-                    (batch[key].size(0), pad_len, *batch[key].shape[2:]),
-                    pad_value,
-                    dtype=batch[key].dtype,
-                    device=batch[key].device,
-                )
-
-                # Concatenate padding to the right side of the tensor
-                batch[key] = torch.cat([batch[key], padding], dim=1)
-            if key == "logits_to_keep":
-                # Create padding tensor
-                padding = torch.ones(
-                    1,
-                    dtype=batch[key].dtype,
-                    device=batch[key].device,
-                )
-
-                # Concatenate padding to the right side of the tensor
-                batch[key] = torch.cat([batch[key], padding], dim=0)
-
-        # Update the total sequence length after padding
-        total_seq_len = batch["input_ids"].size(1)
-
-    # Slice batch for sequence parallel
-    for key in batch:
-        if not isinstance(batch[key], torch.Tensor) or batch[key].dim() <= 1:
-            continue
-
-        # Split in sequential fashion and grab this rank's chunk
-        if batch[key].size(1) == total_seq_len:
-            batch[key] = (
-                batch[key].chunk(local_world_size, dim=1)[local_rank].contiguous()
-            )
-        elif key == "logits_to_keep":
-            batch[key] = (
-                batch[key].chunk(local_world_size, dim=0)[local_rank].contiguous()
-            )
-
-        # Handle num_items_in_batch
-        if "num_items_in_batch" in batch:
-            # Approximation; this needed since num_items_in_batch may be counted across
-            # all samples in a gradient accumulated batch, not on a per-step basis.
-            batch["num_items_in_batch"] = (
-                batch["labels"] != -100
-            ).sum() * gradient_accumulation_steps
-
-    return batch, original_seq_len, pad_len
-
-
-class SequenceParallelContextManager:
-    """Context manager for sequence parallelism operations.
-
-    This class provides a context that will automatically apply sequence parallelism
-    during model forward passes using a pre-forward hook, and gather outputs from
-    across the sequence parallelism group using a post-forward hook.
-
-    Args:
-        models: List of models to apply sequence parallelism to pre- and post- forward
-            hooks.
-        sequence_parallel_degree: Number of processes to split sequences over.
-        gradient_accumulation_steps: Number of steps to accumulate gradients over.
-        ring_attn_func: Which ring attention function to use. Currently unused.
-    """
-
-    def __init__(
-        self,
-        models: list[nn.Module],
-        sequence_parallel_degree: int,
-        gradient_accumulation_steps: int,
-        ring_attn_func: RingAttnFunc,
-    ):
-        self.models = models
-        self.sequence_parallel_degree = sequence_parallel_degree
-        self.gradient_accumulation_steps = gradient_accumulation_steps
-        self.ring_attn_func = ring_attn_func
-        self.process_group = get_ring_attn_group()
-
-        # Initialize sequence parallel group details
-        self.local_rank = dist.get_rank(self.process_group)
-        self.local_world_size = dist.get_world_size(self.process_group)
-
-        # Will store hook handles for removal
-        self.hook_handles: list[RemovableHandle] = []
-
-        # Store original sequence length and padding information
-        self.original_seq_len = 0
-        self.pad_len = 0
-
-        # Create a partially applied version of the apply_sequence_parallelism function
-        self.apply_sequence_parallelism = functools.partial(
-            apply_sequence_parallelism,
-            local_rank=self.local_rank,
-            local_world_size=self.local_world_size,
-            gradient_accumulation_steps=self.gradient_accumulation_steps,
-            ring_attn_func=self.ring_attn_func,
-        )
-
-    def __enter__(self):
-        # Forward pre-hook to apply sequence parallelism
-        def sequence_parallel_pre_hook(_, args, kwargs):
-            # Apply sequence parallelism to kwargs and get original sequence length and padding info
-            kwargs, self.original_seq_len, self.pad_len = (
-                self.apply_sequence_parallelism(batch=kwargs)
-            )
-
-            return args, kwargs
-
-        # Forward post-hook to gather outputs
-        def sequence_parallel_post_hook(_, __, output: ModelOutput) -> ModelOutput:
-            # Gather the sharded outputs
-            output = self.gather_outputs(output)
-
-            # Remove padding if it was added
-            if self.pad_len > 0:
-                for key, value in output.items():
-                    if isinstance(value, torch.Tensor) and value.dim() > 1:
-                        if value.size(1) == self.original_seq_len + self.pad_len:
-                            # Slice to remove padding
-                            output[key] = value[:, : self.original_seq_len].contiguous()
-
-            return output
-
-        # Register both hooks
-        for model in self.models:
-            self.hook_handles.append(
-                model.register_forward_pre_hook(
-                    sequence_parallel_pre_hook, with_kwargs=True
-                )
-            )
-            self.hook_handles.append(
-                model.register_forward_hook(sequence_parallel_post_hook)
-            )
-
-        return self
-
-    def __exit__(self, exc_type, exc_val, exc_tb):
-        # Remove all hooks
-        for handle in self.hook_handles:
-            handle.remove()
-        self.hook_handles = []
-
-    def gather_outputs(self, output: CausalLMOutputWithPast) -> CausalLMOutputWithPast:
-        """Gather sharded outputs from all ranks and reconstruct the full tensor."""
-        for key, value in output.items():
-            if isinstance(value, torch.Tensor) and value.dim() > 1:
-                output[key] = AllGatherWithGrad.apply(value, self.process_group)
-
-        return output
-
-
-class AllGatherWithGrad(torch.autograd.Function):
-    """Custom autograd function for all-gather to preserve gradients."""
-
-    @staticmethod
-    def forward(
-        ctx: torch.autograd.function.FunctionCtx,
-        input_tensor: torch.Tensor,
-        group: dist.ProcessGroup,
-    ) -> torch.Tensor:
-        """
-        Forward pass of all-gather of data with sequence dimension.
-
-        Args:
-            ctx: `torch.autograd` function context.
-            input_tensor: Tensor from model output with sequence dimension.
-            group: `torch.distributed` process group.
-
-        Returns:
-            Tensor from gathering the `input_tensor` from across the process group and
-                concatenating along the sequence dimension.
-        """
-        ctx.group = group
-        ctx.rank = dist.get_rank(group)
-        world_size = dist.get_world_size(group)
-
-        # Gather shape metadata
-        local_shape = torch.tensor(list(input_tensor.shape), device=input_tensor.device)
-        all_shapes = [torch.zeros_like(local_shape) for _ in range(world_size)]
-        dist.all_gather(all_shapes, local_shape, group=group)
-
-        # Store sequence lengths for backward pass
-        seq_lens = [int(shape[1].item()) for shape in all_shapes]
-        ctx.seq_lens = seq_lens
-
-        # Perform all_gather operation
-        gathered = [
-            torch.zeros(
-                tuple(shape.tolist()),
-                dtype=input_tensor.dtype,
-                device=input_tensor.device,
-            )
-            for shape in all_shapes
-        ]
-        dist.all_gather(gathered, input_tensor, group=group)
-
-        # Concatenate tensors along sequence dimension
-        result = torch.cat(gathered, dim=1)
-
-        return result
-
-    @staticmethod
-    def backward(
-        ctx: torch.autograd.function.FunctionCtx, grad_output: torch.Tensor
-    ) -> tuple[torch.Tensor, None]:
-        """
-        Backward pass for all-gather operation.
-
-        Extracts the gradient slice corresponding to this rank's original input
-        from the full gradient tensor.
-
-        Args:
-            ctx: `torch.autograd` function context.
-            grad_output: Gradient from subsequent layers with respect to the
-                concatenated output tensor.
-
-        Returns:
-            Tuple containing the gradient slice for this rank's input tensor and `None`
-                for the process group parameter which doesn't require gradients.
-        """
-        rank = ctx.rank
-        seq_lens = ctx.seq_lens
-
-        # Extract gradient for this rank's chunk
-        offset = sum(seq_lens[:rank])
-        grad_slice = grad_output[:, offset : offset + seq_lens[rank]].contiguous()
-
-        return grad_slice, None

src/axolotl/utils/data/rl.py

@@ -18,9 +18,8 @@ from axolotl.utils.data.utils import deduplicate_and_log_datasets, md5
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import is_main_process, zero_first
 from axolotl.utils.models import load_tokenizer
-from axolotl.utils.schemas.enums import RLType
 
-LOG = logging.getLogger(__name__)
+LOG = logging.getLogger("axolotl")
 
 
 def _get_path(ds_hash, cfg):
@@ -72,7 +71,6 @@ def map_dataset(cfg, data_set, ds_transform_fn, tokenizer, **map_kwargs):
     data_set = data_set.map(
         ds_transform_fn,
         desc="Mapping RL Dataset",
-        num_proc=cfg.dataset_processes,
         **map_kwargs,
     )
 
@@ -82,7 +80,7 @@ def map_dataset(cfg, data_set, ds_transform_fn, tokenizer, **map_kwargs):
 def drop_long_rl_seq(
     sample, rl, tokenizer, sequence_len  # pylint: disable=invalid-name
 ):
-    if rl in (RLType.DPO, RLType.IPO, RLType.ORPO, RLType.SIMPO):
+    if rl in ("dpo", "ipo", "orpo", "simpo"):
         if not (
             sample.get("prompt") and sample.get("chosen") and sample.get("rejected")
         ):
@@ -102,7 +100,7 @@ def drop_long_rl_seq(
             len_prompt + len_rejected
         ) <= sequence_len
 
-    if rl is RLType.KTO:
+    if rl == "kto":
         if not (sample.get("prompt") and sample.get("completion")):
             raise ValueError("Prompt and completion keys are required for KTO datasets")
 
@@ -116,7 +114,7 @@ def drop_long_rl_seq(
 
         return (len_prompt + len_completion) <= sequence_len
 
-    if rl is RLType.GRPO:
+    if rl == "grpo":
         return True
 
     raise ValueError("Unknown RL type")
@@ -139,9 +137,9 @@ def load_prepare_preference_datasets(cfg):
             if _type:
                 if isinstance(_type, DictDefault):
                     _type = "user_defined.default"
-                if _cfg.rl is RLType.ORPO:
+                if _cfg.rl == "orpo":
                     ds_transform_fn = load_orpo(_type, _cfg, dataset_idx=i)
-                elif _cfg.rl is RLType.KTO:
+                elif _cfg.rl == "kto":
                     ds_transform_fn = load_kto(_type, _cfg, dataset_idx=i)
                 else:
                     ds_transform_fn = load_dpo(_type, _cfg, dataset_idx=i)
@@ -152,7 +150,7 @@ def load_prepare_preference_datasets(cfg):
                 split_datasets[i] = map_dataset(
                     cfg, data_set, ds_transform_fn, tokenizer, **map_kwargs
                 )
-            elif _cfg.rl is RLType.KTO:
+            elif _cfg.rl == "kto":
                 ds_transform_fn = load_kto(_type, _cfg, dataset_idx=i)
                 map_kwargs = {}
                 if isinstance(ds_transform_fn, tuple):
@@ -187,7 +185,7 @@ def load_prepare_preference_datasets(cfg):
                     )
 
         combined_datasets = concatenate_datasets(split_datasets)
-        combined_datasets = combined_datasets.shuffle(seed=cfg.seed or 42)
+        combined_datasets = combined_datasets.shuffle(seed=cfg.seed)
 
         return combined_datasets
 
@@ -207,8 +205,6 @@ def load_prepare_preference_datasets(cfg):
                 eval_dataset = load_split(cfg.test_datasets, cfg)
         if not eval_dataset:
             if cfg.val_set_size:
-                seed = cfg.seed if cfg.seed is not None else 42
-
                 # ensure we end up with the same fingerprint by doing rank0 first and being able to cache
                 to_hash_train = (
                     train_dataset._fingerprint  # pylint: disable=protected-access
@@ -217,7 +213,7 @@ def load_prepare_preference_datasets(cfg):
                     + "|"
                     + "train"
                     + "|"
-                    + str(seed)
+                    + str(cfg.seed or 42)
                 )
                 to_hash_test = (
                     train_dataset._fingerprint  # pylint: disable=protected-access
@@ -226,13 +222,13 @@ def load_prepare_preference_datasets(cfg):
                     + "|"
                     + "test"
                     + "|"
-                    + str(seed)
+                    + str(cfg.seed or 42)
                 )
                 train_fingerprint = md5(to_hash_train)
                 test_fingerprint = md5(to_hash_test)
                 ds_w_test_split = train_dataset.train_test_split(
                     test_size=cfg.val_set_size,
-                    seed=seed,
+                    seed=cfg.seed,
                     shuffle=False,
                     train_new_fingerprint=train_fingerprint,
                     test_new_fingerprint=test_fingerprint,

src/axolotl/utils/data/sft.py

@@ -148,7 +148,7 @@ def prepare_dataset(cfg, tokenizer, processor=None, preprocess_iterable=None):
             ds_wrapper_partial,
             max_tokens=cfg.sequence_len,
             batch_size=cfg.micro_batch_size,
-            seed=cfg.seed if cfg.seed is not None else 42,
+            seed=cfg.seed or 42,
             buffer_size=cfg.pretrain_multipack_buffer_size or 10_000,
         )
         # https://discuss.huggingface.co/t/how-to-use-huggingface-trainer-streaming-datasets-without-wrapping-it-with-torchdatas-iterablewrapper/25230
@@ -416,8 +416,6 @@ def load_prepare_datasets(
     )
 
     if split == "train" and val_set_size:
-        seed = cfg.seed if cfg.seed is not None else 42
-
         # ensure we end up with the same fingerprint by doing rank0 first and being able to cache
         to_hash_train = (
             dataset._fingerprint  # pylint: disable=protected-access
@@ -426,7 +424,7 @@ def load_prepare_datasets(
             + "|"
             + "train"
             + "|"
-            + str(seed)
+            + str(cfg.seed or 42)
         )
         to_hash_test = (
             dataset._fingerprint  # pylint: disable=protected-access
@@ -435,7 +433,7 @@ def load_prepare_datasets(
             + "|"
             + "test"
             + "|"
-            + str(seed)
+            + str(cfg.seed or 42)
         )
         train_fingerprint = md5(to_hash_train)
         test_fingerprint = md5(to_hash_test)
@@ -444,7 +442,7 @@ def load_prepare_datasets(
         dataset = dataset.train_test_split(
             test_size=val_set_size,
             shuffle=False,
-            seed=seed,
+            seed=cfg.seed or 42,
             train_new_fingerprint=train_fingerprint,
             test_new_fingerprint=test_fingerprint,
         )
@@ -484,7 +482,7 @@ def get_dataset_wrapper(
     }
 
     LOG.info(
-        f"Loading dataset: {config_dataset['path']} with base_type: {d_base_type} and prompt_style: {d_prompt_style}"
+        f"Loading dataset with base_type: {d_base_type} and prompt_style: {d_prompt_style}"
     )
 
     if (

src/axolotl/utils/gradient_checkpointing/__init__.py

@@ -5,11 +5,8 @@ from functools import partial
 
 from packaging import version
 
-from axolotl.utils.gradient_checkpointing.offload_cpu import (
+from axolotl.utils.gradient_checkpointing.unsloth import (
-    CPU_Offloaded_Gradient_Checkpointer,
+    Unsloth_Offloaded_Gradient_Checkpointer,
-)
-from axolotl.utils.gradient_checkpointing.offload_disk import (
-    Disco,
 )
 
 transformers_version = version.parse(importlib.metadata.version("transformers"))
@@ -29,31 +26,12 @@ def hf_grad_checkpoint_offload_wrapper(
     decoder_layer, *args, use_reentrant=None
 ):  # pylint: disable=unused-argument
     if uses_gc_layers(decoder_layer):
-        return CPU_Offloaded_Gradient_Checkpointer.apply(
+        return Unsloth_Offloaded_Gradient_Checkpointer.apply(
             decoder_layer,
             *args,
         )
 
-    return CPU_Offloaded_Gradient_Checkpointer.apply(
+    return Unsloth_Offloaded_Gradient_Checkpointer.apply(
-        (
-            decoder_layer.func.__self__
-            if isinstance(decoder_layer, partial)
-            else decoder_layer.__self__
-        ),
-        *args,
-    )
-
-
-def hf_grad_checkpoint_disk_offload_wrapper(
-    decoder_layer, *args, use_reentrant=None
-):  # pylint: disable=unused-argument
-    if uses_gc_layers(decoder_layer):
-        return Disco.apply(
-            decoder_layer,
-            *args,
-        )
-
-    return Disco.apply(
         (
             decoder_layer.func.__self__
             if isinstance(decoder_layer, partial)

src/axolotl/utils/gradient_checkpointing/offload_disk.py

@@ -1,531 +0,0 @@
-"""
-DISCO - DIsk-based Storage and Checkpointing with Optimized prefetching
-"""
-
-# Copyright 2025 Axolotl AI. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-import atexit
-import concurrent.futures
-import logging
-import os
-import queue
-import shutil
-import tempfile
-import threading
-import time
-import uuid
-from collections import deque
-from concurrent.futures import Future
-from typing import Dict
-
-import torch
-
-torch_cuda_amp_custom_fwd = torch.amp.custom_fwd(device_type="cuda")
-torch_cuda_amp_custom_bwd = torch.amp.custom_bwd(device_type="cuda")
-
-# Setup logger
-logger = logging.getLogger(__name__)
-
-
-class DiskOffloadManager:
-    """
-    Manages offloaded tensors and handles prefetching in a separate thread.
-    Includes synchronization to prevent race conditions.
-    """
-
-    def __init__(
-        self,
-        prefetch_size: int = 3,
-        prefetch_to_gpu: bool = True,
-        save_workers: int = 4,
-    ):
-        """
-        Args:
-            prefetch_size: Maximum number of tensors to prefetch in the background.
-            prefetch_to_gpu: Whether to prefetch tensors directly to GPU memory.
-            save_workers: Maximum number of concurrent save operations.
-        """
-        self.temp_dir = tempfile.mkdtemp(prefix="disco_")
-
-        # Track tensor paths and their status
-        self.tensor_paths: deque = deque()  # Ordered history of tensor paths (LIFO)
-        self.file_locks: Dict[str, threading.Lock] = (
-            {}
-        )  # Maps file_path -> threading.Lock()
-        # Maps file_path -> status ("saving", "ready", "prefetching", "loaded", "deleted")
-        self.file_status: Dict[str, str] = {}
-
-        self.max_prefetch = prefetch_size
-        self.prefetch_to_gpu = prefetch_to_gpu
-
-        # Thread synchronization
-        self.manager_lock = threading.RLock()  # Used for thread-safe operations
-
-        # Prefetch queue and cache
-        self.prefetch_queue: queue.Queue = queue.Queue()
-        self.prefetch_cache: Dict[str, torch.Tensor] = {}  # Maps file_path -> tensor
-
-        # Save queue and thread pool
-        self.save_queue: queue.Queue = queue.Queue()
-        self.save_pool = concurrent.futures.ThreadPoolExecutor(max_workers=save_workers)
-        self.save_futures: Dict[str, Future] = {}
-        self.save_semaphore = threading.Semaphore(
-            save_workers * 2
-        )  # Limit concurrent save operations
-
-        # Start prefetch worker thread
-        self.stop_event = threading.Event()
-        # start multiple threads for prefetching
-        self.prefetch_worker_count = 2
-        self.prefetch_workers = []
-        for _ in range(self.prefetch_worker_count):
-            worker = threading.Thread(target=self._prefetch_worker, daemon=True)
-            worker.start()
-            self.prefetch_workers.append(worker)
-
-        # Start save worker thread
-        self.save_worker = threading.Thread(target=self._save_worker, daemon=True)
-        self.save_worker.start()
-        self.idx = 0
-
-        atexit.register(self.cleanup)
-
-    def _save_worker(self):
-        """Background thread that processes the save queue"""
-        while not self.stop_event.is_set():
-            try:
-                save_item = self.save_queue.get(timeout=0.5)
-                if save_item is None:
-                    continue
-
-                tensor, file_path = save_item
-
-                # Submit the save task to the thread pool
-                future = self.save_pool.submit(
-                    self._save_tensor_to_disk, tensor, file_path
-                )
-                with self.manager_lock:
-                    self.save_futures[file_path] = future
-
-                self.save_queue.task_done()
-
-            except queue.Empty:
-                time.sleep(0.01)  # Small sleep to prevent CPU spinning
-                continue
-
-    def _save_tensor_to_disk(self, tensor: torch.Tensor, file_path: str):
-        """Actually save the tensor to disk"""
-        try:
-            # Save tensor to disk
-            cpu_tensor = tensor.detach().cpu()
-            torch.save(cpu_tensor, file_path)
-            del cpu_tensor
-
-            with self.manager_lock:
-                # Mark file as ready
-                self.file_status[file_path] = "ready"
-
-            # Release semaphore
-            self.save_semaphore.release()
-
-            return True
-        except FileNotFoundError as e:
-            logger.error(f"Error saving tensor to {file_path}: {e}")
-            with self.manager_lock:
-                self.file_status[file_path] = "error"
-
-            # Release semaphore
-            self.save_semaphore.release()
-
-            return False
-
-    def _prefetch_worker(self):
-        """Background thread that loads tensors from disk ahead of time"""
-        while not self.stop_event.is_set():
-            try:
-                file_path = self.prefetch_queue.get(timeout=0.5)
-                if file_path is None:
-                    continue
-
-                # Check if file is available and not already in cache
-                with self.manager_lock:
-                    if (
-                        file_path not in self.file_status
-                        or self.file_status[file_path] == "deleted"
-                    ):
-                        self.prefetch_queue.task_done()
-                    if file_path in self.prefetch_cache:
-                        self.prefetch_queue.task_done()
-                        continue
-
-                    # If file is still being saved, wait for it
-                    if (
-                        self.file_status[file_path] == "saving"
-                        and file_path in self.save_futures
-                    ):
-                        # Re-queue this prefetch request with a little delay
-                        self.prefetch_queue.task_done()
-                        time.sleep(0.1)
-                        self.prefetch_queue.put(file_path)
-                        continue
-
-                    # Mark file as being prefetched
-                    self.file_status[file_path] = "prefetching"
-
-                # Load tensor from disk and store in cache
-                try:
-                    if os.path.exists(file_path):
-                        if self.prefetch_to_gpu:
-                            tensor = torch.load(
-                                file_path,
-                                map_location=torch.device("cuda"),
-                                weights_only=True,
-                            )
-                        else:
-                            tensor = torch.load(file_path, weights_only=True)
-
-                        with self.manager_lock:
-                            self.prefetch_cache[file_path] = tensor
-                            self.file_status[file_path] = "ready"
-                    else:
-                        with self.manager_lock:
-                            if self.file_status.get(file_path) != "deleted":
-                                logger.warning(
-                                    f"Prefetch error: File not found {file_path}"
-                                )
-                                self.file_status[file_path] = "missing"
-
-                except FileNotFoundError as e:
-                    with self.manager_lock:
-                        if self.file_status.get(file_path) != "deleted":
-                            logger.warning(f"Prefetch error for {file_path}: {e}")
-                            self.file_status[file_path] = "error"
-
-                self.prefetch_queue.task_done()
-
-            except queue.Empty:
-                time.sleep(0.01)  # Small sleep to prevent CPU spinning
-                continue
-
-    def save_tensor(self, tensor: torch.Tensor):
-        """Save tensor to disk asynchronously and return file path with thread-safe operations"""
-        # Generate unique file path
-        self.idx += 1
-        file_path: str = os.path.join(
-            self.temp_dir, f"{self.idx:06d}-{uuid.uuid4()}.pt"
-        )
-
-        with self.manager_lock:
-            # Mark file as being saved
-            self.file_locks[file_path] = threading.Lock()
-            self.file_status[file_path] = "saving"
-            # Add to history
-            self.tensor_paths.append(file_path)
-
-        # Acquire semaphore to limit concurrent save operations
-        self.save_semaphore.acquire()  # pylint: disable=consider-using-with
-        # Queue tensor for saving in background
-        self.save_queue.put((tensor.detach(), file_path))
-
-        return file_path
-
-    def wait_for_save(self, file_path, timeout=None) -> None:
-        """Wait for a tensor to be saved to disk"""
-        start_time = time.time()
-        while timeout is None or time.time() - start_time < timeout:
-            with self.manager_lock:
-                if self.file_status.get(file_path) == "ready":
-                    return
-                if self.file_status.get(file_path) in ["error", "missing", "deleted"]:
-                    return
-
-                if file_path in self.save_futures:
-                    future = self.save_futures[file_path]
-                    if future.done():
-                        return
-
-            # Small sleep to prevent CPU spinning
-            time.sleep(0.01)
-
-        # Timeout
-        logger.warning(f"Timeout waiting for tensor to be saved: {file_path}")
-        return
-
-    def load_tensor(self, file_path, target_device="cuda"):
-        """Load tensor from disk or prefetch cache with proper synchronization"""
-        # Wait for tensor to be saved if it's still in progress
-        self.wait_for_save(file_path)
-
-        tensor = None
-
-        # Try to get from cache first
-        with self.manager_lock:
-            # Check if tensor is already in cache
-            if file_path in self.prefetch_cache:
-                tensor = self.prefetch_cache[file_path]
-                del self.prefetch_cache[file_path]
-                self.file_status[file_path] = "loaded"
-
-        if tensor is not None:
-            # Ensure tensor is on correct device
-            if target_device != "cpu" and tensor.device.type == "cpu":
-                tensor = tensor.to(target_device, non_blocking=True)
-            return tensor
-
-        # If not in cache, load directly from disk
-        try:
-            if not os.path.exists(file_path):
-                logger.error(f"File not found for loading: {file_path}")
-                raise FileNotFoundError(f"File not found: {file_path}")
-
-            tensor = torch.load(file_path, weights_only=True)
-
-            with self.manager_lock:
-                self.file_status[file_path] = "loaded"
-
-            if target_device != "cpu":
-                tensor = tensor.to(target_device, non_blocking=True)
-
-            return tensor
-
-        except Exception as e:
-            logger.error(f"Error loading tensor from {file_path}: {e}")
-            raise
-
-    def _safe_delete_file(self, file_path):
-        """Safely delete a file with proper synchronization"""
-        with self.manager_lock:
-            # Make sure any save operation is completed
-            if file_path in self.save_futures:
-                future = self.save_futures[file_path]
-                try:
-                    if not future.done():
-                        future.cancel()
-                    del self.save_futures[file_path]
-                except FileNotFoundError as e:
-                    logger.warning(
-                        f"Error canceling save operation for {file_path}: {e}"
-                    )
-
-            # Only delete if file exists and is not being prefetched
-            status = self.file_status.get(file_path)
-            if status in ["ready", "loaded", "error", "missing"]:
-                try:
-                    if os.path.exists(file_path):
-                        os.remove(file_path)
-                    self.file_status[file_path] = "deleted"
-                    return True
-                except FileNotFoundError as e:
-                    logger.warning(f"Error deleting file {file_path}: {e}")
-            return False
-
-    def trigger_prefetch(self, n=None):
-        """Trigger prefetching of the next N tensors with proper synchronization"""
-        if n is None:
-            n = self.max_prefetch
-
-        prefetch_paths = []
-        with self.manager_lock:
-            # Find files that are ready to be prefetched (not already in cache or being prefetched)
-            for path in reversed(self.tensor_paths):
-                if (
-                    path not in self.prefetch_cache
-                    and self.file_status.get(path) == "ready"
-                ):
-                    prefetch_paths.append(path)
-                    if len(prefetch_paths) >= n:
-                        break
-
-        # Queue files for prefetching
-        for path in prefetch_paths:
-            self.prefetch_queue.put(path)
-
-    def cleanup_tensor(self, file_path: str):
-        """Clean up a specific tensor file after it's been used"""
-        with self.manager_lock:
-            if file_path in self.tensor_paths:
-                self.tensor_paths.remove(file_path)
-
-            # Remove from prefetch cache if present
-            if file_path in self.prefetch_cache:
-                del self.prefetch_cache[file_path]
-
-            # Remove from save futures if present
-            if file_path in self.save_futures:
-                future = self.save_futures[file_path]
-                if not future.done():
-                    future.cancel()
-                del self.save_futures[file_path]
-
-        # Try to delete the file
-        self._safe_delete_file(file_path)
-
-    def cleanup(self):
-        """Clean up all temp files and stop prefetch thread with proper synchronization"""
-        self.stop_event.set()
-
-        # Cancel all pending save operations
-        with self.manager_lock:
-            for _, future in self.save_futures.items():
-                if not future.done():
-                    future.cancel()
-            self.save_futures.clear()
-
-        # Drain the save queue
-        while not self.save_queue.empty():
-            try:
-                self.save_queue.get_nowait()
-                self.save_queue.task_done()
-            except queue.Empty:
-                break
-
-        # Shutdown the save pool
-        self.save_pool.shutdown(wait=False)
-
-        # Join the save worker thread
-        if self.save_worker.is_alive():
-            self.save_worker.join(timeout=2.0)
-
-        # Join the prefetch worker threads
-        for thread in self.prefetch_workers:
-            if thread.is_alive():
-                thread.join(timeout=2.0)
-
-        # Clear cache and remove all temporary files
-        with self.manager_lock:
-            self.prefetch_cache.clear()
-            paths_to_delete = list(self.tensor_paths)
-            self.tensor_paths.clear()
-
-        # Delete all temporary files
-        for path in paths_to_delete:
-            self._safe_delete_file(path)
-
-        # Remove temp directory
-        try:
-            if os.path.exists(self.temp_dir):
-                shutil.rmtree(self.temp_dir, ignore_errors=True)
-        except FileNotFoundError as e:
-            logger.warning(f"Error removing temporary directory {self.temp_dir}: {e}")
-
-
-class Disco(torch.autograd.Function):
-    """
-    Disco: DIsk-based Storage and Checkpointing with Optimized prefetching
-    Advanced disk-based gradient checkpointer with prefetching.
-    """
-
-    # Shared manager instance across all checkpointing operations
-    _manager = None
-
-    @staticmethod
-    def get_instance(prefetch_size=1, prefetch_to_gpu=True, save_workers=4):
-        """Get or create the offload manager"""
-        if Disco._manager is None:
-            Disco._manager = DiskOffloadManager(
-                prefetch_size=prefetch_size,
-                prefetch_to_gpu=prefetch_to_gpu,
-                save_workers=save_workers,
-            )
-        return Disco._manager
-
-    @staticmethod
-    @torch_cuda_amp_custom_fwd
-    def forward(
-        ctx,
-        forward_function,
-        hidden_states,
-        *args,
-        prefetch_size=1,
-        prefetch_to_gpu=True,
-        save_workers=4,
-    ):
-        """Forward pass that offloads activations to disk asynchronously"""
-        # Get or create the manager
-        manager = Disco.get_instance(
-            prefetch_size=prefetch_size,
-            prefetch_to_gpu=prefetch_to_gpu,
-            save_workers=save_workers,
-        )
-
-        # Save tensor to disk asynchronously
-        file_path = manager.save_tensor(hidden_states)
-
-        # Run forward pass immediately without waiting for save to complete
-        with torch.no_grad():
-            output = forward_function(hidden_states, *args)
-
-        # Store what we need for backward
-        ctx.save_for_backward(torch.tensor([0]))  # Dummy tensor
-        ctx.file_path = file_path
-        ctx.forward_function = forward_function
-        ctx.args = args
-
-        return output
-
-    @staticmethod
-    @torch_cuda_amp_custom_bwd
-    def backward(ctx, *grad_outputs):
-        """Backward pass that loads activations from disk with prefetching"""
-        # Get the manager
-        manager = Disco._manager
-
-        # Trigger prefetching for future tensors
-        # This happens at the start of backward, so should have time to complete
-        manager.trigger_prefetch()
-
-        # Load hidden states from disk or prefetch cache
-        file_path = ctx.file_path
-        try:
-            # Ensure the file is saved before we try to load it
-            manager.wait_for_save(file_path)
-
-            hidden_states = manager.load_tensor(file_path)
-            hidden_states.requires_grad = True
-
-            # Compute gradients
-            with torch.enable_grad():
-                output = ctx.forward_function(hidden_states, *ctx.args)
-
-                # Handle tuple outputs properly
-                if isinstance(output, tuple):
-                    if len(grad_outputs) == len(output):
-                        torch.autograd.backward(output, grad_outputs)
-                    else:
-                        torch.autograd.backward(output, grad_outputs[0])
-                else:
-                    torch.autograd.backward(output, grad_outputs[0])
-
-            # Clean up the file after we're done with it
-            manager.cleanup_tensor(file_path)
-
-            return (
-                (
-                    None,  # forward_function
-                    hidden_states.grad,  # hidden_states grad
-                )
-                + (None,) * len(ctx.args)  # for each arg
-                + (
-                    None,  # prefetch_size
-                    None,  # prefetch_to_gpu
-                    None,  # save_workers
-                )
-            )
-
-        except Exception as e:
-            logger.error(f"Error in backward pass: {e}")
-            # Clean up the file even on error
-            manager.cleanup_tensor(file_path)
-            raise

src/axolotl/utils/gradient_checkpointing/unsloth.py

@@ -1,4 +1,4 @@
-"""CPU offloaded checkpointing"""
+"""Unsloth checkpointing"""
 
 # Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
 #
@@ -26,7 +26,7 @@ else:
     torch_cuda_amp_custom_bwd = torch.amp.custom_bwd(device_type="cuda")
 
 
-class CPU_Offloaded_Gradient_Checkpointer(  # pylint: disable=invalid-name
+class Unsloth_Offloaded_Gradient_Checkpointer(  # pylint: disable=invalid-name
     torch.autograd.Function
 ):
     """

src/axolotl/utils/models.py

@@ -70,13 +70,9 @@ from axolotl.utils.distributed import (
     is_local_main_process,
     is_main_process,
 )
-from axolotl.utils.gradient_checkpointing import (
+from axolotl.utils.gradient_checkpointing import hf_grad_checkpoint_offload_wrapper
-    hf_grad_checkpoint_disk_offload_wrapper,
-    hf_grad_checkpoint_offload_wrapper,
-)
 from axolotl.utils.lora_embeddings import get_linear_embedding_layers
 from axolotl.utils.model_shard_quant import load_sharded_model, load_sharded_model_quant
-from axolotl.utils.schemas.enums import RLType
 
 LOG = logging.getLogger(__name__)
 PLUGIN_MANAGER = PluginManager.get_instance()
@@ -623,10 +619,6 @@ class ModelLoader:
 
         if self.cfg.gradient_checkpointing in ["unsloth", "offload"]:
             transformers.modeling_utils.checkpoint = hf_grad_checkpoint_offload_wrapper
-        if self.cfg.gradient_checkpointing == "offload_disk":
-            transformers.modeling_utils.checkpoint = (
-                hf_grad_checkpoint_disk_offload_wrapper
-            )
 
         if self.cfg.flash_attention:
             self.patch_attention()
@@ -1380,7 +1372,7 @@ class ModelLoader:
             # then the dpo trainer doesn't want the peft model loaded over it, it just wants the lora/peft config
             if (
                 self.cfg.adapter
-                and self.cfg.rl in [RLType.DPO, RLType.IPO, RLType.KTO]
+                and self.cfg.rl in ["dpo", "ipo", "kto"]
                 and not self.cfg.merge_lora
             ):
                 _, lora_config = load_lora(

src/axolotl/utils/schemas/config.py

@@ -27,7 +27,7 @@ from axolotl.utils.schemas.datasets import (
     StepwiseSupervisedDataset,
 )
 from axolotl.utils.schemas.deprecated import DeprecatedParameters, RemappedParameters
-from axolotl.utils.schemas.enums import ChatTemplate, RingAttnFunc, RLType
+from axolotl.utils.schemas.enums import ChatTemplate, RLType
 from axolotl.utils.schemas.integrations import (
     CometConfig,
     GradioConfig,
@@ -178,7 +178,7 @@ class AxolotlInputConfig(
 
     # torch_dtype: torch.dtype | None
 
-    gradient_checkpointing: Literal["offload", "offload_disk"] | bool | None = Field(
+    gradient_checkpointing: Literal["unsloth", "offload"] | bool | None = Field(
         default=False
     )
     gradient_checkpointing_kwargs: dict[str, Any] | None = None
@@ -260,7 +260,7 @@ class AxolotlInputConfig(
 
     sequence_parallel_degree: int | None = None
     heads_k_stride: int | None = None
-    ring_attn_func: RingAttnFunc | None = None
+    ring_attn_func: str | None = None
 
     special_tokens: SpecialTokensConfig | None = None
     tokens: list[str] | None = None
@@ -782,7 +782,7 @@ class AxolotlInputConfig(
 
     @model_validator(mode="after")
     def check_simpo_warmup(self):
-        if self.rl is RLType.SIMPO and self.warmup_ratio:
+        if self.rl == "simpo" and self.warmup_ratio:
             raise ValueError(
                 "warmup_ratio is not supported with the simpo trainer. Please use `warmup_steps` instead"
             )
@@ -1149,28 +1149,16 @@ class AxolotlInputConfig(
 
         return data
 
-    # @model_validator(mode="before")
-    # @classmethod
-    # def check_grpo_peft_liger(cls, data):
-    #     if (
-    #         data.get("rl") == "grpo"
-    #         and data.get("trl", {})
-    #         and data.get("trl").get("use_liger_loss")
-    #         and data.get("adapter")
-    #     ):
-    #         raise ValueError("PEFT + GRPO + Liger is not yet supported")
-    #     return data
-    #
     @model_validator(mode="before")
     @classmethod
-    def check_grpo_liger_sequence_parallel(cls, data):
+    def check_grpo_peft_liger(cls, data):
         if (
             data.get("rl") == "grpo"
             and data.get("trl", {})
             and data.get("trl").get("use_liger_loss")
-            and data.get("sequence_parallel_degree", 1) > 1
+            and data.get("adapter")
         ):
-            raise ValueError("GRPO + SP + Liger not currently supported")
+            raise ValueError("PEFT + GRPO + Liger is not yet supported")
         return data
 
     @model_validator(mode="after")
@@ -1185,7 +1173,7 @@ class AxolotlInputConfig(
 
             if self.sample_packing and self.micro_batch_size > 1:
                 raise ValueError(
-                    "micro_batch_size must be set to 1 when sample_packing is enabled "
+                    "micro_batch_size must be set to 1 when sample_packing is enabled"
                     "due to a `ring-flash-attn` requirement"
                 )
 
@@ -1217,8 +1205,16 @@ class AxolotlInputConfig(
         if getattr(self, "sequence_parallel_degree", 1) == 1:
             return self
 
+        from axolotl.monkeypatch.attention.ring_attn.patch import RingAttnFunc
+
         if self.ring_attn_func is not None:
+            valid_funcs = list(RingAttnFunc)
+            if self.ring_attn_func in valid_funcs:
                 self.ring_attn_func = RingAttnFunc(self.ring_attn_func)
+            else:
+                raise ValueError(
+                    f"ring_attn_func: {self.ring_attn_func} must be in {valid_funcs}"
+                )
         else:
             # Default ring attention function selection
             sample_packing = getattr(self, "sample_packing", False)
@@ -1349,10 +1345,6 @@ class AxolotlConfigWCapabilities(AxolotlInputConfig):
             ):
                 return data
 
-            # Skip if dropout is not 0, as auto enabling it would just disable it during runtime patch checks
-            if data.get("lora_dropout") != 0:
-                return data
-
             # Check multi-GPU compatibility
             capabilities = data.get("capabilities")
             is_multi_gpu = capabilities and capabilities.get("n_gpu", 0) > 1

src/axolotl/utils/schemas/enums.py

@@ -6,12 +6,12 @@ from enum import Enum
 class RLType(str, Enum):
     """RL trainer type configuration subset"""
 
-    DPO = "dpo"  # pylint: disable=invalid-name
+    dpo = "dpo"  # pylint: disable=invalid-name
-    GRPO = "grpo"  # pylint: disable=invalid-name
+    grpo = "grpo"  # pylint: disable=invalid-name
-    IPO = "ipo"  # pylint: disable=invalid-name
+    ipo = "ipo"  # pylint: disable=invalid-name
-    ORPO = "orpo"  # pylint: disable=invalid-name
+    orpo = "orpo"  # pylint: disable=invalid-name
-    KTO = "kto"  # pylint: disable=invalid-name
+    kto = "kto"  # pylint: disable=invalid-name
-    SIMPO = "simpo"  # pylint: disable=invalid-name
+    simpo = "simpo"  # pylint: disable=invalid-name
 
 
 class ChatTemplate(str, Enum):
@@ -55,14 +55,3 @@ class CustomSupportedOptimizers(str, Enum):
     adopt_adamw = "adopt_adamw"  # pylint: disable=invalid-name
     came_pytorch = "came_pytorch"  # pylint: disable=invalid-name
     muon = "muon"  # pylint: disable=invalid-name
-
-
-class RingAttnFunc(str, Enum):
-    """Enum class for supported `ring-flash-attn` implementations"""
-
-    # VARLEN_RING = "varlen_ring"
-    # VARLEN_ZIGZAG = "varlen_zigzag"
-    VARLEN_LLAMA3 = "varlen_llama3"
-    BATCH_RING = "batch_ring"
-    # BATCH_ZIGZAG = "batch_zigzag"
-    # BATCH_STRIPE = "batch_stripe"

src/axolotl/utils/trainer.py

@@ -16,7 +16,7 @@ from datasets import IterableDataset, disable_caching, enable_caching
 from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
 from transformers.utils import is_torch_bf16_gpu_available
 
-from axolotl.core.trainer_builder import HFCausalTrainerBuilder, HFRLTrainerBuilder
+from axolotl.core.trainers.builders import HFCausalTrainerBuilder, HFRLTrainerBuilder
 from axolotl.monkeypatch.trainer_eval_guard import patch_evaluation_loop_for_fsdp2
 from axolotl.utils.distributed import reduce_and_broadcast
 from axolotl.utils.environment import check_cuda_p2p_ib_support
@@ -633,8 +633,7 @@ def setup_trainer(
         peft_config: Optional PEFT (Parameter-Efficient Fine-Tuning) configuration. Default is None.
 
     Returns:
-        A trainer instance (either `HFRLTrainer` or `HFCausalTrainer`) configured based
+        A trainer instance configured based on the provided parameters.
-            on the provided parameters.
     """
     if (
         cfg.torch_compile

tests/core/test_trainer_builder.py

@@ -1,10 +1,8 @@
-"""
+"""Unit tests for axolotl.core.trainers.builders"""
-unit tests for axolotl.core.trainer_builder
-"""
 
 import pytest
 
-from axolotl.core.trainer_builder import HFRLTrainerBuilder
+from axolotl.core.trainers.builders import HFRLTrainerBuilder
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.models import load_model, load_tokenizer
@@ -53,9 +51,7 @@ def fixture_model(cfg, tokenizer):
 
 
 class TestHFRLTrainerBuilder:
-    """
+    """Test case class for RL trainer builder"""
-    TestCase class for DPO trainer builder
-    """
 
     def test_build_training_arguments(self, cfg, model, tokenizer):
         builder = HFRLTrainerBuilder(cfg, model, tokenizer)

tests/e2e/multigpu/patched/test_sp.py

@@ -25,7 +25,6 @@ class TestSequenceParallelism:
         micro_batch_size=1,
         pad_to_sequence_len=True,
         ring_attn_func=None,
-        threshold=2.0,
     ):
         """Helper method to run sequence parallel tests with different configurations"""
         cfg = DictDefault(
@@ -94,22 +93,22 @@ class TestSequenceParallelism:
         )
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", threshold, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 2.6, "Train Loss is too high"
         )
 
     @pytest.mark.parametrize(
-        "sample_packing, micro_batch_size, pad_to_sequence_len, ring_attn_func, threshold",
+        "sample_packing, micro_batch_size, pad_to_sequence_len, ring_attn_func",
         [
-            (True, 1, True, None, 2.5),  # defaults to varlen_llama3 ring_attn_func
+            (True, 1, True, None),  # defaults to varlen_llama3 ring_attn_func
-            (False, 2, True, None, 2.5),  # defaults to batch_ring ring_attn_func
+            (False, 2, True, None),  # defaults to batch_ring ring_attn_func
-            # (False, 2, True, "batch_zigzag", 2.5),
+            (False, 2, True, "batch_zigzag"),
-            (False, 2, False, None, 2.5),  # defaults to batch_ring ring_attn_func
+            # (False, 2, False),  # not yet working
         ],
         ids=[
             "sample_packing, varlen_llama3 ring_attn_func",
-            "no sample_packing, pad_to_sequence_len, batch_ring ring_attn_func",
-            # "no sample_packing, no pad_to_sequence_len, batch_zigzag ring_attn_func",
             "no sample_packing, no pad_to_sequence_len, batch_ring ring_attn_func",
+            "no sample_packing, no pad_to_sequence_len, batch_zigzag ring_attn_func",
+            # "no sample_packing, pad_to_sequence_len",  # not yet working
         ],
     )
     def test_sequence_parallel_training(
@@ -119,7 +118,6 @@ class TestSequenceParallelism:
         micro_batch_size,
         pad_to_sequence_len,
         ring_attn_func,
-        threshold,
     ):
         """Test sequence parallel training with different configurations"""
         self._run_sequence_parallel_test(
@@ -128,5 +126,4 @@ class TestSequenceParallelism:
             micro_batch_size=micro_batch_size,
             pad_to_sequence_len=pad_to_sequence_len,
             ring_attn_func=ring_attn_func,
-            threshold=threshold,
         )

tests/e2e/multigpu/solo/test_grpo.py

@@ -227,9 +227,11 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
 
         current_env = os.environ.copy()
         env = {
-            "NCCL_P2P_LEVEL": "LOC",
+            "NCCL_P2P_LEVEL": "NVL",
             **current_env,
             "CUDA_VISIBLE_DEVICES": "1",
+            "VLLM_DISABLE_COMPILE_CACHE": "1",
+            # "VLLM_USE_V1": "0",
         }
         vllm_process = start_vllm(
             cfg.base_model,
@@ -255,7 +257,7 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
                     f"{get_torch_dist_unique_port()}",
                 ],
                 env={
-                    "NCCL_P2P_LEVEL": "LOC",
+                    "NCCL_P2P_LEVEL": "NVL",
                     "NCCL_DEBUG": "INFO",
                     **current_env,
                 },
@@ -318,9 +320,11 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
 
         current_env = os.environ.copy()
         env = {
-            "NCCL_P2P_LEVEL": "LOC",  # nccl can be brittle, assume P2P isn't reliable
+            "NCCL_P2P_LEVEL": "NVL",  # 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,
@@ -346,7 +350,7 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
                     f"{get_torch_dist_unique_port()}",
                 ],
                 env={
-                    "NCCL_P2P_LEVEL": "LOC",
+                    "NCCL_P2P_LEVEL": "NVL",
                     "NCCL_DEBUG": "INFO",
                     **current_env,
                 },

tests/e2e/patched/test_activation_checkpointing.py

@@ -26,15 +26,10 @@ class TestActivationCheckpointing:
     E2E tests for activation checkpointing
     """
 
-    @pytest.mark.parametrize(
-        "gradient_checkpointing",
-        ["offload", "offload_disk"],
-    )
     def test_activation_checkpointing_offload(
         self,
         temp_dir,
         fix_checkpoint_after_test,  # pylint: disable=unused-argument,redefined-outer-name
-        gradient_checkpointing,
     ):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
@@ -69,7 +64,7 @@ class TestActivationCheckpointing:
                 "sample_packing": True,
                 "bf16": True,
                 "save_safetensors": True,
-                "gradient_checkpointing": gradient_checkpointing,
+                "gradient_checkpointing": "offload",
             }
         )
 

tests/e2e/patched/test_sp.py

@@ -10,15 +10,14 @@ import pytest
 import torch
 from accelerate.state import PartialState
 
+from axolotl.core.trainers.mixins.sequence_parallel import apply_sequence_parallelism
 from axolotl.monkeypatch.attention.ring_attn import (
+    RingAttnFunc,
     get_ring_attn_group,
     register_ring_attn,
     set_ring_attn_group,
 )
-from axolotl.utils.ctx_managers.sequence_parallel import apply_sequence_parallelism
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.schemas.enums import RingAttnFunc
-from axolotl.utils.schemas.trl import TRLConfig
 
 
 @pytest.fixture
@@ -63,14 +62,12 @@ def sequence_parallel_batch():
     input_ids = torch.arange(batch_size * seq_len).reshape(batch_size, seq_len)
     attention_mask = torch.ones(batch_size, seq_len)
     position_ids = torch.arange(seq_len).expand(batch_size, seq_len)
-    labels = input_ids.clone()
 
     # Create test batch
     batch = {
         "input_ids": input_ids,
         "attention_mask": attention_mask,
         "position_ids": position_ids,
-        "labels": labels,
     }
 
     return batch
@@ -182,44 +179,12 @@ class TestConfigValidation:
                 False,
                 "micro_batch_size must be set to 1",
             ),
-            # Valid: Basic GRPO config
-            (
-                {
-                    "sequence_parallel_degree": 2,
-                    "flash_attention": True,
-                    "micro_batch_size": 2,
-                    "trl": {"use_liger_loss": True},
-                },
-                {
-                    "sequence_parallel_degree": 2,
-                    "flash_attention": True,
-                    "micro_batch_size": 2,
-                    "trl": TRLConfig(use_liger_loss=True),
-                },
-                True,
-                "GRPO + SP + Liger not currently supported",
-            ),
-            # Invalid: GRPO config with Liger loss
-            (
-                {
-                    "rl": "grpo",
-                    "sequence_parallel_degree": 2,
-                    "flash_attention": True,
-                    "micro_batch_size": 2,
-                    "trl": {"use_liger_loss": True},
-                },
-                None,
-                False,
-                "GRPO + SP + Liger not currently supported",
-            ),
         ],
         ids=[
             "valid_config",
             "default_sp_degree",
             "without_flash_attention",
             "sample_packing_with_large_batch",
-            "valid_grpo",
-            "grpo_with_liger_loss",
         ],
     )
     def test_sequence_parallel_config_validation(
@@ -291,7 +256,7 @@ class TestConfigValidation:
             AxolotlInputConfig(**cfg)
 
         # Verify error message
-        assert "Input should be 'varlen_llama3' or 'batch_ring'" in str(excinfo.value)
+        assert "ring_attn_func: INVALID_FUNC must be in" in str(excinfo.value)
 
 
 class TestApplySequenceParallelism:
@@ -325,11 +290,10 @@ class TestApplySequenceParallelism:
 
     def test_world_size_one(self, sequence_parallel_batch):
         """Test that function returns original batch when world size is 1."""
-        result, _, _ = apply_sequence_parallelism(
+        result = apply_sequence_parallelism(
             batch=sequence_parallel_batch,
             local_rank=0,
             local_world_size=1,
-            gradient_accumulation_steps=1,
             ring_attn_func=RingAttnFunc.BATCH_RING,
         )
 
@@ -341,11 +305,10 @@ class TestApplySequenceParallelism:
         batch = sequence_parallel_batch
         seq_len = batch["input_ids"].size(1)
 
-        result, _, _ = apply_sequence_parallelism(
+        result = apply_sequence_parallelism(
             batch=batch,
             local_rank=0,
             local_world_size=2,
-            gradient_accumulation_steps=1,
             ring_attn_func=RingAttnFunc.BATCH_RING,
         )
 
@@ -365,59 +328,57 @@ class TestApplySequenceParallelism:
         seq_len = batch["input_ids"].size(1)
         original_input_ids = batch["input_ids"].clone()
 
-        result, _, _ = apply_sequence_parallelism(
+        result = apply_sequence_parallelism(
             batch=batch,
             local_rank=1,
             local_world_size=2,
-            gradient_accumulation_steps=1,
             ring_attn_func=RingAttnFunc.BATCH_RING,
         )
 
         # Verify content: rank 1 should get the second half of the sequence
         assert torch.equal(result["input_ids"], original_input_ids[:, seq_len // 2 :])
 
-    # TODO(djsaunde): add back once implemented.
+    def test_batch_zigzag(self, sequence_parallel_batch):
-    # def test_batch_zigzag(self, sequence_parallel_batch):
+        """Test BATCH_ZIGZAG sharding pattern."""
-    #     """Test BATCH_ZIGZAG sharding pattern."""
+        batch = sequence_parallel_batch
-    #     batch = sequence_parallel_batch
+        original_input_ids = batch["input_ids"].clone()
-    #     original_input_ids = batch["input_ids"].clone()
+        seq_len = batch["input_ids"].size(1)
-    #     seq_len = batch["input_ids"].size(1)
 
-    #     # Test rank 0
+        # Test rank 0
-    #     result_rank0 = apply_sequence_parallelism(
+        result_rank0 = apply_sequence_parallelism(
-    #         batch={k: v.clone() for k, v in batch.items()},
+            batch={k: v.clone() for k, v in batch.items()},
-    #         local_rank=0,
+            local_rank=0,
-    #         local_world_size=2,
+            local_world_size=2,
-    #         ring_attn_func=RingAttnFunc.BATCH_ZIGZAG,
+            ring_attn_func=RingAttnFunc.BATCH_ZIGZAG,
-    #     )
+        )
 
-    #     # Test rank 1
+        # Test rank 1
-    #     result_rank1 = apply_sequence_parallelism(
+        result_rank1 = apply_sequence_parallelism(
-    #         batch={k: v.clone() for k, v in batch.items()},
+            batch={k: v.clone() for k, v in batch.items()},
-    #         local_rank=1,
+            local_rank=1,
-    #         local_world_size=2,
+            local_world_size=2,
-    #         ring_attn_func=RingAttnFunc.BATCH_ZIGZAG,
+            ring_attn_func=RingAttnFunc.BATCH_ZIGZAG,
-    #     )
+        )
 
-    #     # Checks for both ranks
+        # Checks for both ranks
-    #     assert result_rank0["input_ids"].shape[1] == seq_len // 2
+        assert result_rank0["input_ids"].shape[1] == seq_len // 2
-    #     assert result_rank1["input_ids"].shape[1] == seq_len // 2
+        assert result_rank1["input_ids"].shape[1] == seq_len // 2
 
-    #     # For a 2-rank system with 8 tokens, check specific zigzag pattern
+        # For a 2-rank system with 8 tokens, check specific zigzag pattern
-    #     # Rank 0 should get chunks [0, 1] and [6, 7]
+        # Rank 0 should get chunks [0, 1] and [6, 7]
-    #     # Rank 1 should get chunks [2, 3] and [4, 5]
+        # Rank 1 should get chunks [2, 3] and [4, 5]
-    #     if seq_len == 8:
+        if seq_len == 8:
-    #         # Create expected tensors for comparison
+            # Create expected tensors for comparison
-    #         rank0_expected = torch.cat(
+            rank0_expected = torch.cat(
-    #             [original_input_ids[:, :2], original_input_ids[:, 6:8]], dim=1
+                [original_input_ids[:, :2], original_input_ids[:, 6:8]], dim=1
-    #         )
+            )
 
-    #         rank1_expected = torch.cat(
+            rank1_expected = torch.cat(
-    #             [original_input_ids[:, 2:4], original_input_ids[:, 4:6]], dim=1
+                [original_input_ids[:, 2:4], original_input_ids[:, 4:6]], dim=1
-    #         )
+            )
 
-    #         assert torch.equal(result_rank0["input_ids"], rank0_expected)
+            assert torch.equal(result_rank0["input_ids"], rank0_expected)
-    #         assert torch.equal(result_rank1["input_ids"], rank1_expected)
+            assert torch.equal(result_rank1["input_ids"], rank1_expected)
 
     def test_partial_application(self, sequence_parallel_batch):
         """Test that we can create a partially applied version of the function."""
@@ -429,12 +390,11 @@ class TestApplySequenceParallelism:
             apply_sequence_parallelism,
             local_rank=0,
             local_world_size=2,
-            gradient_accumulation_steps=1,
             ring_attn_func=RingAttnFunc.BATCH_RING,
         )
 
         # Use the partially applied function
-        result, _, _ = rank0_ring_parallel(batch=batch)
+        result = rank0_ring_parallel(batch=batch)
 
         # Verify it works as expected
         assert result["input_ids"].shape[1] == original_input_ids.shape[1] // 2
@@ -452,15 +412,13 @@ class TestApplySequenceParallelism:
         original_input_ids = batch["input_ids"].clone()
 
         # This should run without error even though position_ids is missing
-        result, _, _ = apply_sequence_parallelism(
+        result = apply_sequence_parallelism(
             batch=batch,
             local_rank=0,
             local_world_size=2,
-            gradient_accumulation_steps=1,
             ring_attn_func=RingAttnFunc.BATCH_RING,
         )
 
         # Verification should pass
-        assert "position_ids" in result
+        assert "position_ids" not in result
-        assert result["input_ids"].shape[1] == result["position_ids"].shape[1]
         assert result["input_ids"].shape[1] == original_input_ids.shape[1] // 2

tests/e2e/test_imports.py

@@ -1,21 +1,21 @@
-"""
+"""Test module to import various submodules that have historically broken due to
-test module to import various submodules that have historically broken due to dependency issues
+dependency issues.
 """
 
 import unittest
 
 
 class TestImports(unittest.TestCase):
-    """
+    """Test class to import various submodules that have historically broken due to
-    Test class to import various submodules that have historically broken due to dependency issues
+    dependency issues.
     """
 
     def test_import_causal_trainer(self):
-        from axolotl.core.trainer_builder import (  # pylint: disable=unused-import  # noqa: F401
+        from axolotl.core.trainers.builders import (  # pylint: disable=unused-import  # noqa: F401
             HFCausalTrainerBuilder,
         )
 
     def test_import_rl_trainer(self):
-        from axolotl.core.trainer_builder import (  # pylint: disable=unused-import  # noqa: F401
+        from axolotl.core.trainers.builders import (  # pylint: disable=unused-import  # noqa: F401
             HFRLTrainerBuilder,
         )