Revert "checkpoint model on first step callback (#2906)"

This reverts commit 10ba1622f7.

.bandit

@@ -1,3 +1,3 @@
 [bandit]
 exclude = tests
-skips = B101
+skips = B101,B615

.github/workflows/base.yml

@@ -5,11 +5,13 @@ on:
     branches:
       - "main"
     paths:
-      - 'Dockerfile-base'
+      - 'docker/Dockerfile-base'
+      - 'docker/Dockerfile-uv-base'
       - '.github/workflows/base.yml'
   pull_request:
     paths:
-      - 'Dockerfile-base'
+      - 'docker/Dockerfile-base'
+      - 'docker/Dockerfile-uv-base'
       - '.github/workflows/base.yml'
   workflow_dispatch:
 
@@ -27,11 +29,11 @@ jobs:
             cuda_version: 12.4.1
             cudnn_version: ""
             python_version: "3.11"
-            pytorch: 2.5.1
+            pytorch: 2.6.0
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
             dockerfile: "Dockerfile-base"
-          - cuda: "124"
-            cuda_version: 12.4.1
+          - cuda: "126"
+            cuda_version: 12.6.3
             cudnn_version: ""
             python_version: "3.11"
             pytorch: 2.6.0
@@ -41,7 +43,7 @@ jobs:
             cuda_version: 12.6.3
             cudnn_version: ""
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.7.0
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
             dockerfile: "Dockerfile-base"
           - cuda: "126"

.github/workflows/main.yml

@@ -15,17 +15,16 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.1
-            axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.6.0
+            axolotl_extras:
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.0
             axolotl_extras: vllm
-            is_latest: true
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
@@ -83,17 +82,17 @@ jobs:
     strategy:
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.1
-            axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.6.0
             axolotl_extras:
             is_latest: true
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.0
+            axolotl_extras:
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
@@ -146,8 +145,8 @@ jobs:
     strategy:
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.6.0
             axolotl_extras:

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

@@ -26,18 +26,18 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.6.0
-            axolotl_extras: vllm
+            axolotl_extras:
             num_gpus: 2
             nightly_build: "true"
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
-            pytorch: 2.5.1
-            axolotl_extras:
+            pytorch: 2.7.0
+            axolotl_extras: vllm
             num_gpus: 2
             nightly_build: "true"
           - cuda: 126

.github/workflows/nightlies.yml

@@ -12,16 +12,16 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.1
-            axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.6.0
             axolotl_extras:
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.1
+            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -65,16 +65,16 @@ jobs:
     strategy:
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.1
-            axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.6.0
             axolotl_extras:
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.1
+            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout

.github/workflows/preview-docs.yml

@@ -28,6 +28,8 @@ jobs:
     steps:
       - name: Check out repository
         uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event.pull_request.head.sha }}
 
       - name: Set up Quarto
         uses: quarto-dev/quarto-actions/setup@v2
@@ -50,10 +52,11 @@ jobs:
 
       - name: Netlify Publish
         uses: nwtgck/actions-netlify@v3.0
+        id: netlify
         with:
           publish-dir: './_site'
-          enable-pull-request-comment: true
-          enable-github-deployment: true
+          enable-pull-request-comment: false
+          enable-github-deployment: false
           github-token: ${{ secrets.GITHUB_TOKEN }}
           deploy-message: "Deployed On Netlify"
           github-deployment-environment: 'preview'
@@ -61,3 +64,13 @@ jobs:
         env:
           NETLIFY_AUTH_TOKEN: ${{ secrets.NETLIFY_AUTH_TOKEN }}
           NETLIFY_SITE_ID: ${{ secrets.NETLIFY_SITE_ID }}
+
+      - name: Update PR with preview link
+        if: ${{ steps.netlify.outcome == 'success' }}
+        uses: marocchino/sticky-pull-request-comment@v2
+        with:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          message: |
+            📖 **Documentation Preview**: ${{ steps.netlify.outputs.deploy-url }}
+
+            Deployed on Netlify from commit ${{ github.event.pull_request.head.sha }}

.github/workflows/tests-nightly.yml

@@ -18,116 +18,26 @@ jobs:
         env:
           SKIP: no-commit-to-branch
 
-  preload-cache:
-    name: Preload HF cache
-    runs-on: ubuntu-latest
-    strategy:
-      fail-fast: false
-      matrix:
-        python_version: ["3.11"]
-        pytorch_version: ["2.6.0"]
-    timeout-minutes: 20
-
-    env:
-      AXOLOTL_IS_CI_CACHE_PRELOAD: "1"
-
-    steps:
-      - name: Check out repository code
-        uses: actions/checkout@v4
-
-      - name: Restore HF cache
-        id: hf-cache-restore
-        uses: actions/cache/restore@v4
-        with:
-          path: |
-            /home/runner/.cache/huggingface/hub/datasets--*
-            /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ runner.os }}-hf-hub-cache-v2
-
-      - name: Setup Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: ${{ matrix.python_version }}
-          cache: 'pip' # caching pip dependencies
-
-      - name: upgrade pip
-        run: |
-          pip3 install --upgrade pip
-          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
-
-      - name: Install PyTorch
-        run: |
-          pip3 install torch==${{ matrix.pytorch_version }}
-
-      - name: Install dependencies
-        run: |
-          pip3 show torch
-          pip3 install --no-build-isolation -U -e .
-          python scripts/unsloth_install.py | sh
-          python scripts/cutcrossentropy_install.py | sh
-          pip3 install -r requirements-dev.txt -r requirements-tests.txt
-
-      - name: Make sure PyTorch version wasn't clobbered
-        run: |
-          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
-
-      - name: Ensure axolotl CLI was installed
-        run: |
-          axolotl --help
-
-      - name: Pre-Download dataset fixture
-        run: |
-          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
-
-      - name: Run tests
-        run: |
-          pytest -v tests/conftest.py
-
-      - name: Upload coverage to Codecov
-        uses: codecov/codecov-action@v5
-        with:
-          token: ${{ secrets.CODECOV_TOKEN }}
-          files: ./coverage.xml
-          flags: unittests,pytorch-${{ matrix.pytorch_version }}
-          fail_ci_if_error: false
-
-      - name: cleanup pip cache
-        run: |
-          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
-
-      - name: Save HF cache
-        id: hf-cache
-        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]
     strategy:
       fail-fast: false
       max-parallel: 2
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.5.1", "2.6.0", "2.7.0"]
+        pytorch_version: ["2.6.0", "2.7.0"]
     timeout-minutes: 20
 
     steps:
       - name: Check out repository code
         uses: actions/checkout@v4
 
-      - name: Restore HF cache
-        id: hf-cache-restore
-        uses: actions/cache/restore@v4
-        with:
-          path: |
-            /home/runner/.cache/huggingface/hub/datasets--*
-            /home/runner/.cache/huggingface/hub/models--*
-          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
@@ -168,15 +78,11 @@ jobs:
         run: |
           axolotl --help
 
-      - name: Pre-Download dataset fixture
-        run: |
-          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
-
       - name: Run tests
         run: |
-          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/
-          pytest -v tests/patched/
-          pytest -v tests/cli/
+          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/
+          pytest -v --durations=10 tests/patched/
+          pytest -v --durations=10 tests/cli/
 
       - name: cleanup pip cache
         run: |
@@ -193,15 +99,8 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.1
-            num_gpus: 1
-            axolotl_extras:
-            nightly_build: "true"
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.6.0
             num_gpus: 1

.github/workflows/tests.yml

@@ -52,7 +52,7 @@ jobs:
       fail-fast: false
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.5.1", "2.6.0", "2.7.1"]
+        pytorch_version: ["2.6.0", "2.7.0", "2.7.1"]
     timeout-minutes: 20
 
     steps:
@@ -102,9 +102,9 @@ jobs:
 
       - name: Run tests
         run: |
-          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/ --cov=axolotl --cov-report=xml
-          pytest -v tests/patched/ --cov=axolotl --cov-append --cov-report=xml
-          pytest -v tests/cli/ --cov=axolotl --cov-append --cov-report=xml
+          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/ --cov=axolotl --cov-report=xml
+          pytest -v --durations=10 tests/patched/ --cov=axolotl --cov-append --cov-report=xml
+          pytest -v --durations=10 tests/cli/ --cov=axolotl --cov-append --cov-report=xml
 
       - name: Upload coverage to Codecov
         uses: codecov/codecov-action@v5
@@ -125,7 +125,7 @@ jobs:
       fail-fast: false
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.5.1", "2.6.0", "2.7.1"]
+        pytorch_version: ["2.6.0", "2.7.0", "2.7.1"]
     timeout-minutes: 20
 
     steps:
@@ -175,9 +175,9 @@ jobs:
 
       - name: Run tests
         run: |
-          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/
-          pytest -v tests/patched/
-          pytest -v tests/cli/
+          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/
+          pytest -v --durations=10 tests/patched/
+          pytest -v --durations=10 tests/cli/
 
       - name: cleanup pip cache
         run: |
@@ -195,12 +195,12 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.7.1
             num_gpus: 1
-            axolotl_extras: vllm
+            axolotl_extras:
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
@@ -247,22 +247,10 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.6.0
-            num_gpus: 1
-            axolotl_extras: llmcompressor
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.1
-            num_gpus: 1
-            axolotl_extras:
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
-            pytorch: 2.7.1
+            pytorch: 2.6.0
             num_gpus: 1
             axolotl_extras:
           - cuda: 128
@@ -311,7 +299,7 @@ jobs:
             python_version: "3.11"
             pytorch: 2.6.0
             num_gpus: 1
-            axolotl_extras: vllm
+            axolotl_extras:
     steps:
       - name: Checkout
         uses: actions/checkout@v4

.pre-commit-config.yaml

@@ -19,7 +19,7 @@ repos:
     hooks:
       - id: isort
 -   repo: https://github.com/PyCQA/flake8
-    rev: 7.2.0
+    rev: 7.3.0
     hooks:
     - id: flake8
 -   repo: https://github.com/pylint-dev/pylint
@@ -27,7 +27,7 @@ repos:
     hooks:
     - id: pylint
 -   repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.16.0
+    rev: v1.16.1
     hooks:
     - id: mypy
       additional_dependencies:
@@ -36,7 +36,7 @@ repos:
             'pydantic>=2.5.3',
         ]
 -   repo: https://github.com/PyCQA/bandit
-    rev: 1.8.3
+    rev: 1.8.6
     hooks:
     -   id: bandit
         args: [

.runpod/src/config/config.yaml

@@ -97,7 +97,7 @@
 #       # 'no_input_format' cannot include {input}
 #       no_input_format: "{instruction} "
 
-#       # For `completion` datsets only, uses the provided field instead of `text` column
+#       # For `completion` datasets only, uses the provided field instead of `text` column
 #       field:
 
 # # Axolotl attempts to save the dataset as an arrow after packing the data together so

MANIFEST.in

@@ -2,4 +2,5 @@ include requirements.txt
 include README.md
 include LICENSE
 include src/setuptools_axolotl_dynamic_dependencies.py
+include src/axolotl/utils/chat_templates/templates/*.jinja
 recursive-include axolotl *.py

README.md

@@ -43,7 +43,7 @@ Features:
 - **Multiple Model Support**: Train various models like LLaMA, Mistral, Mixtral, Pythia, and more. We are compatible with HuggingFace transformers causal language models.
 - **Training Methods**: Full fine-tuning, LoRA, QLoRA, GPTQ, QAT, Preference Tuning (DPO, IPO, KTO, ORPO), RL (GRPO), Multimodal, and Reward Modelling (RM) / Process Reward Modelling (PRM).
 - **Easy Configuration**: Re-use a single YAML file between dataset preprocess, training, evaluation, quantization, and inference.
-- **Performance Optimizations**: [Multipacking](https://docs.axolotl.ai/docs/multipack.html), [Flash Attention](https://github.com/Dao-AILab/flash-attention), [Xformers](https://github.com/facebookresearch/xformers), [Flex Attention](https://pytorch.org/blog/flexattention/), [Liger Kernel](https://github.com/linkedin/Liger-Kernel), [Cut Cross Entropy](https://github.com/apple/ml-cross-entropy/tree/main), Sequence Parallelism (SP), LoRA optimizations, Multi-GPU training (FSDP1, FSDP2, DeepSpeed), Multi-node training (Torchrun, Ray), and many more!
+- **Performance Optimizations**: [Multipacking](https://docs.axolotl.ai/docs/multipack.html), [Flash Attention](https://github.com/Dao-AILab/flash-attention), [Xformers](https://github.com/facebookresearch/xformers), [Flex Attention](https://pytorch.org/blog/flexattention/), [Liger Kernel](https://github.com/linkedin/Liger-Kernel), [Cut Cross Entropy](https://github.com/apple/ml-cross-entropy/tree/main), [Sequence Parallelism (SP)](https://docs.axolotl.ai/docs/sequence_parallelism.html), [LoRA optimizations](https://docs.axolotl.ai/docs/lora_optims.html), [Multi-GPU training (FSDP1, FSDP2, DeepSpeed)](https://docs.axolotl.ai/docs/multi-gpu.html), [Multi-node training (Torchrun, Ray)](https://docs.axolotl.ai/docs/multi-node.html), and many more!
 - **Flexible Dataset Handling**: Load from local, HuggingFace, and cloud (S3, Azure, GCP, OCI) datasets.
 - **Cloud Ready**: We ship [Docker images](https://hub.docker.com/u/axolotlai) and also [PyPI packages](https://pypi.org/project/axolotl/) for use on cloud platforms and local hardware.
 
@@ -55,10 +55,12 @@ Features:
 
 - NVIDIA GPU (Ampere or newer for `bf16` and Flash Attention) or AMD GPU
 - Python 3.11
-- PyTorch ≥2.5.1
+- PyTorch ≥2.6.0
 
 ### Installation
 
+#### Using pip
+
 ```bash
 pip3 install -U packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation axolotl[flash-attn,deepspeed]
@@ -68,6 +70,13 @@ axolotl fetch examples
 axolotl fetch deepspeed_configs  # OPTIONAL
 ```
 
+#### Using Docker
+
+Installing with Docker can be less error prone than installing in your own environment.
+```bash
+docker run --gpus '"all"' --rm -it axolotlai/axolotl:main-latest
+```
+
 Other installation approaches are described [here](https://docs.axolotl.ai/docs/installation.html).
 
 ### Your First Fine-tune

_quarto.yml

@@ -276,6 +276,7 @@ website:
             - docs/torchao.qmd
             - docs/custom_integrations.qmd
             - docs/sequence_parallelism.qmd
+            - docs/gradient_checkpointing.qmd
 
         - section: "Troubleshooting"
           contents:

cicd/Dockerfile.jinja

@@ -9,6 +9,7 @@ ENV GITHUB_REF="{{ GITHUB_REF }}"
 ENV GITHUB_SHA="{{ GITHUB_SHA }}"
 ENV NIGHTLY_BUILD="{{ NIGHTLY_BUILD }}"
 ENV HF_HOME="{{ HF_HOME }}"
+ENV AXOLOTL_DATASET_PROCESSES="8"
 
 RUN apt-get update && \
     apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev

cicd/multigpu.py

@@ -24,9 +24,9 @@ df_template = template_env.get_template("Dockerfile.jinja")
 df_args = {
     "AXOLOTL_EXTRAS": os.environ.get("AXOLOTL_EXTRAS", ""),
     "AXOLOTL_ARGS": os.environ.get("AXOLOTL_ARGS", ""),
-    "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.5.1"),
-    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-py3.11-cu124-2.5.1"),
-    "CUDA": os.environ.get("CUDA", "124"),
+    "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.6.0"),
+    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-py3.11-cu126-2.6.0"),
+    "CUDA": os.environ.get("CUDA", "126"),
     "GITHUB_REF": os.environ.get("GITHUB_REF", "refs/heads/main"),
     "GITHUB_SHA": os.environ.get("GITHUB_SHA", ""),
     "CODECOV_TOKEN": os.environ.get("CODECOV_TOKEN", ""),

cicd/single_gpu.py

@@ -24,14 +24,16 @@ df_template = template_env.get_template(dockerfile)
 df_args = {
     "AXOLOTL_EXTRAS": os.environ.get("AXOLOTL_EXTRAS", ""),
     "AXOLOTL_ARGS": os.environ.get("AXOLOTL_ARGS", ""),
-    "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.5.1"),
-    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-py3.11-cu124-2.5.1"),
-    "CUDA": os.environ.get("CUDA", "124"),
+    "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.6.0"),
+    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-py3.11-cu126-2.6.0"),
+    "CUDA": os.environ.get("CUDA", "126"),
     "GITHUB_REF": os.environ.get("GITHUB_REF", "refs/heads/main"),
     "GITHUB_SHA": os.environ.get("GITHUB_SHA", ""),
     "NIGHTLY_BUILD": os.environ.get("NIGHTLY_BUILD", ""),
     "CODECOV_TOKEN": os.environ.get("CODECOV_TOKEN", ""),
     "HF_HOME": "/workspace/data/huggingface-cache/hub",
+    "PYTHONUNBUFFERED": os.environ.get("PYTHONUNBUFFERED", "1"),
+    "DEEPSPEED_LOG_LEVEL": os.environ.get("DEEPSPEED_LOG_LEVEL", "WARNING"),
 }
 
 dockerfile_contents = df_template.render(**df_args)

docker/Dockerfile-base

@@ -38,6 +38,6 @@ RUN git lfs install --skip-repo && \
     # The base image ships with `pydantic==1.8.2` which is not working
     pip3 install -U --no-cache-dir pydantic==1.10.10
 
-RUN if [ "$PYTORCH_VERSION" = "2.7.1" ] ; then \
-        pip3 install flash-attn==2.7.4.post1; \
+RUN if [ "$PYTORCH_VERSION" = "2.6.0" ] && [ "$CUDA" = "124" ] ; then \
+        FLASH_ATTENTION_FORCE_BUILD="TRUE" pip3 install --no-build-isolation flash-attn==2.8.0.post2; \
     fi

docker/Dockerfile-uv-base

@@ -34,7 +34,3 @@ RUN uv pip install packaging setuptools wheel psutil \
     && uv pip install --no-build-isolation "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main" \
     && uv pip install "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main" \
     && uv pip install awscli pydantic
-
-RUN if [ "$PYTORCH_VERSION" = "2.7.1" ] ; then \
-        uv pip install --no-build-isolation flash-attn==2.7.4.post1; \
-    fi

docs/custom_integrations.qmd

@@ -7,6 +7,7 @@ toc-depth: 3
 ```{python}
 #| echo: false
 
+import os
 import re
 
 def process_readme(integration_name):
@@ -53,6 +54,24 @@ sections = [
     ("LLMCompressor", "llm_compressor")
 ]
 
+for folder_name in os.listdir("../src/axolotl/integrations/"):
+    if folder_name in [path for name, path in sections]:
+        # skip if already in sections
+        continue
+    if os.path.exists(f"../src/axolotl/integrations/{folder_name}/README.md"):
+        # grab the first heading in README.md as the section name
+        with open(f"../src/axolotl/integrations/{folder_name}/README.md", "r") as f:
+            txt = f.read()
+            matches = re.search(r'^# (.*)\n?', txt, flags=re.MULTILINE)
+            if matches:
+                name = matches.group(1)
+            else:
+                continue
+            sections.append((name, folder_name))
+
+# sort sections by name
+sections = sorted(sections, key=lambda x: x[0])
+
 for section_name, folder_name in sections:
     print(print_section(section_name, folder_name))
 ```

docs/dataset-formats/conversation.qmd

@@ -9,7 +9,7 @@ order: 3
 Chat Template strategy uses a jinja2 template that converts a list of messages into a prompt. Support using tokenizer's template, a supported template, or custom jinja2.
 
 ```{.json filename="data.jsonl"}
-{"conversations": [{"role": "...", "content": "..."}]}
+{"messages": [{"role": "...", "content": "..."}, {"role": "...", "content": "..."}, ...]}
 ```
 
 See [configs](../config-reference.qmd) for full configs and supported templates.
@@ -187,6 +187,7 @@ Instead of passing `tools` via the system prompt, an alternative method would be
             "role": "assistant", // call the function via assistant
             "tool_calls": [
                 {
+                    "id": "...",  // required only for mistral
                     "type": "function",
                     "function": {
                         "name": "...",
@@ -199,6 +200,7 @@ Instead of passing `tools` via the system prompt, an alternative method would be
         },
         {
             "role": "tool",
+            "tool_call_id": "...",  // required only for mistral
             "name": "...",
             "content": "..."
         },

docs/docker.qmd

@@ -9,7 +9,7 @@ format:
 This section describes the different Docker images that are released by AxolotlAI at [Docker Hub](https://hub.docker.com/u/axolotlai).
 
 ::: {.callout-important}
-For Blackwell GPUs, please use the tags with Pytorch 2.7.1 and CUDA 12.8.
+For Blackwell GPUs, please use the tags with PyTorch 2.7.1 and CUDA 12.8.
 :::
 
 ## Base
@@ -34,8 +34,9 @@ Tags examples:
 
 - `main-base-py3.11-cu128-2.7.1`
 - `main-base-py3.11-cu126-2.7.1`
+- `main-base-py3.11-cu126-2.7.0`
+- `main-base-py3.11-cu126-2.6.0`
 - `main-base-py3.11-cu124-2.6.0`
-- `main-base-py3.11-cu124-2.5.1`
 
 ## Main
 
@@ -73,13 +74,15 @@ There may be some extra tags appended to the image, like `-vllm` which installs
 
 Tags examples:
 
+- `main-py3.11-cu128-2.7.1`
+- `main-py3.11-cu126-2.7.1`
 - `main-py3.11-cu126-2.7.0`
+- `main-py3.11-cu126-2.6.0`
 - `main-py3.11-cu124-2.6.0`
-- `main-py3.11-cu124-2.5.1`
 - `main-latest`
 - `main-20250303-py3.11-cu124-2.6.0`
-- `main-20250303-py3.11-cu124-2.5.1`
-- `0.9.2`
+- `main-20250303-py3.11-cu126-2.6.0`
+- `0.10.1`
 
 ## Cloud
 

docs/faq.qmd

@@ -51,6 +51,18 @@ description: Frequently asked questions
 >   pad_token: "..."
 > ```
 
+**Q: `IterableDataset error` or `KeyError: 'input_ids'` when using `preprocess` CLI**
+
+> A: This is because you may be using `preprocess` CLI with `pretraining_dataset:` or `skip_prepare_dataset: true` respectively. Please use `axolotl train` CLI directly instead as these datasets are prepared on demand.
+
+**Q: vLLM is not working with Axolotl**
+
+> A: We currently recommend torch 2.6.0 for use with `vllm`. Please ensure you use the right version. For Docker, please use the `main-py3.11-cu124-2.6.0` tag.
+
+**Q: FA2 2.8.0 `undefined symbol` runtime error on CUDA 12.4**
+
+> A: There seems to be a wheel issue with FA2 2.8.0 on CUDA 12.4. Try CUDA 12.6 instead or downgrade to FA2 2.7.4. Please refer to the upstream issue: https://github.com/Dao-AILab/flash-attention/issues/1717.
+
 ### Chat templates
 
 **Q: `jinja2.exceptions.UndefinedError: 'dict object' has no attribute 'content' / 'role' / ____`**

docs/fsdp_qlora.qmd

@@ -20,7 +20,7 @@ To enable `QLoRA` with `FSDP`, you need to perform the following steps:
 > See the [example config](#example-config) file in addition to reading these instructions.
 
 1. Set `adapter: qlora` in your axolotl config file.
-2. Enable FSDP in your axolotl config, as [described here](https://github.com/axolotl-ai-cloud/axolotl?tab=readme-ov-file#fsdp).
+2. Enable FSDP in your axolotl config, as [described here](multi-gpu.qmd#sec-fsdp).
 3. Use one of the supported model types: `llama`, `mistral` or `mixtral`.
 
 ## Example Config

docs/gradient_checkpointing.qmd

@@ -0,0 +1,29 @@
+---
+title: Gradient Checkpointing and Activation Offloading
+---
+
+Gradient checkpointing and activation offloading are techniques used to optimize the performance of deep learning
+models by reducing the memory footprint and improving computational efficiency.
+
+### Enabling Gradient Checkpointing
+
+```yaml
+gradient_checkpointing: true
+```
+
+### Enabling Activation Offloading
+
+```yaml
+gradient_checkpointing: true  # required for activation offloading
+activation_offloading: true
+```
+
+Activation offloading variants:
+
+The default `activation_offloading: true` offloads activations to CPU and uses CUDA streams
+to overlap the communications and computations when offloading.
+
+The `activation_offloading: legacy` naively offloads activations to CPU and without additional optimizations.
+
+For resource constrained environments with limited CPU memory, `activation_offloading: disk` offloads
+activations to disk instead of CPU RAM so that much larger context lengths can be trained with minimal memory.

docs/installation.qmd

@@ -15,7 +15,7 @@ This guide covers all the ways you can install and set up Axolotl for your envir
 
 - NVIDIA GPU (Ampere architecture or newer for `bf16` and Flash Attention) or AMD GPU
 - Python ≥3.11
-- PyTorch ≥2.5.1
+- PyTorch ≥2.6.0
 
 ## Installation Methods {#sec-installation-methods}
 

docs/multi-gpu.qmd

@@ -23,8 +23,6 @@ Axolotl supports several methods for multi-GPU training:
 
 ## DeepSpeed {#sec-deepspeed}
 
-DeepSpeed is the recommended approach for multi-GPU training due to its stability and performance. It provides various optimization levels through ZeRO stages.
-
 ### Configuration {#sec-deepspeed-config}
 
 Add to your YAML config:
@@ -32,7 +30,6 @@ Add to your YAML config:
 ```{.yaml}
 deepspeed: deepspeed_configs/zero1.json
 ```
-
 ### Usage {#sec-deepspeed-usage}
 
 ```{.bash}
@@ -66,9 +63,75 @@ Start from Stage 1 -> Stage 2 -> Stage 3.
 
 :::
 
-## FSDP {#sec-fsdp}
+::: {.callout-tip}
 
-### Basic FSDP Configuration {#sec-fsdp-config}
+Using ZeRO Stage 3 with Single-GPU training
+
+ZeRO Stage 3 can be used for training on a single GPU by manually setting the environment variables:
+`WORLD_SIZE=1 LOCAL_RANK=0 MASTER_ADDR=0.0.0.0 MASTER_PORT=29500`
+
+:::
+
+## Fully Sharded Data Parallel (FSDP) {#sec-fsdp}
+
+::: {.callout-note}
+
+FSDP2 is recommended for new users. FSDP1 is deprecated and will be removed in an upcoming release of Axolotl.
+
+:::
+
+### Migrating from FSDP1 to FSDP2 {#sec-migrate-fsdp1-fsdp2}
+
+To migrate your config from FSDP1 to FSDP2, you must use the `fsdp_version` top-level config field to specify the FSDP version, and
+also follow the config field mapping below to update field names.
+
+#### Config mapping
+
+FSDP1 | FSDP2
+-------- | --------
+fsdp_sharding_strategy | reshard_after_forward
+fsdp_backward_prefetch_policy | **REMOVED**
+fsdp_backward_prefetch | **REMOVED**
+fsdp_forward_prefetch | **REMOVED**
+fsdp_sync_module_states | **REMOVED**
+fsdp_cpu_ram_efficient_loading | cpu_ram_efficient_loading
+fsdp_state_dict_type | state_dict_type
+fsdp_use_orig_params | **REMOVED**
+
+
+For example, if you were using the following FSDP1 config:
+
+```{.yaml}
+fsdp_version: 1
+fsdp_config:
+  fsdp_offload_params: false
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: Qwen3DecoderLayer
+  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_sharding_strategy: FULL_SHARD
+```
+
+You can migrate to the following FSDP2 config:
+
+```{.yaml}
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Qwen3DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  reshard_after_forward: true
+```
+
+### FSDP1 (deprecated) {#sec-fsdp-config}
+
+::: {.callout-note}
+
+Using `fsdp` to configure FSDP is deprecated and will be removed in an upcoming release of Axolotl. Please use `fsdp_config` as above instead.
+
+:::
 
 ```{.yaml}
 fsdp:
@@ -80,6 +143,7 @@ fsdp_config:
   fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
 ```
 
+
 ## Sequence parallelism {#sec-sequence-parallelism}
 
 We support sequence parallelism (SP) via the

docs/multi-node.qmd

@@ -40,13 +40,13 @@ use_cpu: false
 
 Configure your model to use FSDP in the Axolotl yaml. For example:
 ```yaml
-fsdp:
-  - full_shard
-  - auto_wrap
+fsdp_version: 2
 fsdp_config:
-  fsdp_offload_params: true
-  fsdp_state_dict_type: FULL_STATE_DICT
-  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  offload_params: true
+  state_dict_type: FULL_STATE_DICT
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  reshard_after_forward: true
 ```
 
 All you have to do now is launch using accelerate as you would usually do on each machine and voila, the processes will start once you have launched accelerate on every machine.

docs/rlhf.qmd

@@ -17,7 +17,6 @@ feedback. Various methods include, but not limited to:
 - [Kahneman-Tversky Optimization (KTO)](#kto)
 - [Odds Ratio Preference Optimization (ORPO)](#orpo)
 - [Group Relative Policy Optimization (GRPO)](#grpo)
-- Proximal Policy Optimization (PPO) (not yet supported in axolotl, if you're interested in contributing, please reach out!)
 
 
 ## RLHF using Axolotl
@@ -275,15 +274,14 @@ rl: dpo
 datasets:
   - path: ...
     split: train
-    type: user_defined.default
-
-    field_prompt: "prompt"
-    field_system: "system"
-    field_chosen: "chosen"
-    field_rejected: "rejected"
-    prompt_format: "{prompt}"
-    chosen_format: "{chosen}"
-    rejected_format: "{rejected}"
+    type:
+      field_prompt: "prompt"
+      field_system: "system"
+      field_chosen: "chosen"
+      field_rejected: "rejected"
+      prompt_format: "{prompt}"
+      chosen_format: "{chosen}"
+      rejected_format: "{rejected}"
 ```
 
 The input format is a simple JSON input with customizable fields based on the above config.
@@ -476,14 +474,13 @@ rl: kto
 datasets:
   - path: ...
     split: train
-    type: user_defined.default
-
-    field_prompt: "prompt"
-    field_system: "system"
-    field_completion: "completion"
-    field_label: "label"
-    prompt_format: "{prompt}"
-    completion_format: "{completion}"
+    type:
+      field_prompt: "prompt"
+      field_system: "system"
+      field_completion: "completion"
+      field_label: "label"
+      prompt_format: "{prompt}"
+      completion_format: "{completion}"
 ```
 
 The input format is a simple JSON input with customizable fields based on the above config.

examples/archived/README.md

@@ -0,0 +1,5 @@
+# Archived Examples
+
+This directory contains examples that are no longer maintained and may no longer be functional.
+
+We keep them around for archival purposes in case they are useful to others.

examples/devstral/README.md

@@ -0,0 +1,70 @@
+# Finetune Devstral with Axolotl
+
+Devstral Small is a 24B parameter opensource model from MistralAI found on HuggingFace [Devstral-Small-2505](https://huggingface.co/mistralai/Devstral-Small-2505) and [Devstral-Small-2507](https://huggingface.co/mistralai/Devstral-Small-2507). `Devstral-Small-2507` is the latest version of the model and has [function calling](https://mistralai.github.io/mistral-common/usage/tools/) support.
+
+This guide shows how to fine-tune it with Axolotl with multi-turn conversations with proper masking.
+
+The model was fine-tuned ontop of [Mistral-Small-3.1](https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Base-2503) without the vision layer and has a context of up to 128k tokens.
+
+Thanks to the team at MistralAI for giving us early access to prepare for this release.
+
+## Getting started
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html). You need to install from main as Devstral is only on nightly or use our latest [Docker images](https://docs.axolotl.ai/docs/docker.html).
+
+    Here is an example of how to install from main for pip:
+
+```bash
+# Ensure you have Pytorch installed (Pytorch 2.6.0+)
+git clone https://github.com/axolotl-ai-cloud/axolotl.git
+cd axolotl
+
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install --no-build-isolation -e '.[flash-attn]'
+```
+
+2. Run the finetuning example:
+
+```bash
+axolotl train examples/devstral/devstral-small-qlora.yml
+```
+
+This config uses about 21GB VRAM.
+
+Let us know how it goes. Happy finetuning! 🚀
+
+### TIPS
+
+- You can run a full finetuning by removing the `adapter: qlora` and `load_in_4bit: true` from the config.
+- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+- The dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+- Learn how to use function calling with Axolotl at [docs](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#using-tool-use).
+
+## Optimization Guides
+
+- [Multi-GPU Training](https://docs.axolotl.ai/docs/multi-gpu.html)
+- [Multi-Node Training](https://docs.axolotl.ai/docs/multi-node.html)
+- [LoRA Optimizations](https://docs.axolotl.ai/docs/lora_optims.html)
+- [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy)
+- [Liger Kernel](https://docs.axolotl.ai/docs/custom_integrations.html#liger-kernels)
+
+## Limitations
+
+We only support the `mistral-common` tokenizer for Supervised Fine-tuning at the moment and for `type: chat_template` only.
+
+In addition, we do not support overriding tokens yet.
+
+## Related Resources
+
+- [MistralAI Devstral Blog](https://mistral.ai/news/devstral)
+- [MistralAI Devstral 1.1 Blog](https://mistral.ai/news/devstral-2507)
+- [Axolotl Docs](https://docs.axolotl.ai)
+- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
+- [Axolotl Website](https://axolotl.ai)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)
+
+
+## Future Work
+
+- Add parity to Preference Tuning, RL, Multi-modal, etc.
+- Add parity to other tokenizer configs like overriding tokens.

examples/devstral/devstral-small-qlora.yml

@@ -0,0 +1,64 @@
+base_model: mistralai/Devstral-Small-2507
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+# Enable to use mistral-common tokenizer
+tokenizer_use_mistral_common: true
+
+load_in_8bit: false
+load_in_4bit: true
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/qlora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+pad_to_sequence_len: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_torch
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+loss_watchdog_threshold: 5.0
+loss_watchdog_patience: 3
+
+warmup_ratio: 0.05
+evals_per_epoch: 4
+saves_per_epoch: 1
+
+weight_decay: 0.0
+special_tokens:

examples/lfm2/README.md

@@ -0,0 +1,7 @@
+# Liquid Foundation Models 2
+
+LFM2 support in transformers exists in the main branch, but is not yet included in the transformers release.
+
+```bash
+pip install --upgrade --no-deps --force-reinstall git+https://github.com/huggingface/transformers.git
+```

examples/lfm2/lfm2-350m-fft.yaml

@@ -0,0 +1,48 @@
+base_model: LiquidAI/LFM2-350M
+
+chunked_cross_entropy: true
+
+chat_template: tokenizer_default
+eot_tokens:
+  - "<|im_end|>"
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_field_role: from
+    message_field_content: value
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.05
+output_dir: ./outputs/out
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 4
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 5e-5
+
+bf16: true
+tf32: true
+
+gradient_checkpointing: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 2
+saves_per_epoch: 1
+
+weight_decay: 0.0

examples/magistral/README.md

@@ -18,16 +18,10 @@ git clone https://github.com/axolotl-ai-cloud/axolotl.git
 cd axolotl
 
 pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
-pip3 install --no-build-isolation -e '.[flash-attn,mistral]'
+pip3 install --no-build-isolation -e '.[flash-attn]'
 ```
 
-2. Download the example config:
-
-```bash
-axolotl fetch examples
-```
-
-3. Run the finetuning example:
+2. Run the finetuning example:
 
 ```bash
 axolotl train examples/magistral/magistral-small-qlora.yaml
@@ -42,7 +36,7 @@ Let us know how it goes. Happy finetuning! 🚀
 - For inference, the official MistralAI team recommends `top_p: 0.95` and `temperature: 0.7` with `max_tokens: 40960`.
 - You can run a full finetuning by removing the `adapter: qlora` and `load_in_4bit: true` from the config.
 - Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
-- The dataset format is the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+- The dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
 
 ## Optimization Guides
 
@@ -54,7 +48,7 @@ Let us know how it goes. Happy finetuning! 🚀
 
 We only support the `mistral-common` tokenizer for Supervised Fine-tuning at the moment and for `type: chat_template` only.
 
-The tokenizer does not work with `dataset.map` with multiprocessing, so we had to disable it. In addition, we do not support overriding tokens yet.
+In addition, we do not support overriding tokens yet.
 
 ## Related Resources
 

requirements.txt

@@ -1,24 +1,24 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
 
 # START section of dependencies that don't install on Darwin/MacOS
-bitsandbytes==0.45.4
+bitsandbytes==0.46.0
 triton>=3.0.0
 mamba-ssm==1.2.0.post1
 xformers>=0.0.23.post1
 autoawq==0.2.7.post3
-liger-kernel==0.5.10
+liger-kernel==0.6.0
 # END section
 
 packaging==23.2
 
-huggingface_hub==0.32.2
-peft==0.15.2
-transformers==4.52.4
+huggingface_hub>=0.33.0
+peft==0.16.0
+transformers==4.53.2
 tokenizers>=0.21.1
-accelerate==1.7.0
-datasets==3.6.0
+accelerate==1.8.1
+datasets==4.0.0
 deepspeed>=0.17.0
-trl==0.18.2
+trl==0.19.1
 hf_xet==1.1.2
 
 optimum==1.16.2
@@ -68,4 +68,4 @@ schedulefree==1.4.1
 axolotl-contribs-lgpl==0.0.6
 axolotl-contribs-mit==0.0.3
 
-mistral-common==1.6.0
+mistral-common==1.7.0

scripts/cutcrossentropy_install.py

@@ -29,5 +29,5 @@ UV_PREFIX = "uv " if USE_UV else ""
 
 print(
     UNINSTALL_PREFIX
-    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@78b2a45713a54c9bedf8b33f5e31cf07a1a57154"'
+    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@865b899"'
 )

setup.py

@@ -66,13 +66,16 @@ def parse_requirements(extras_require_map):
 
             if (major, minor) >= (2, 7):
                 _install_requires.pop(_install_requires.index(xformers_version))
-                # _install_requires.append("xformers==0.0.29.post3")  # xformers seems to be hard pinned to 2.6.0
-                extras_require_map["vllm"] = ["vllm==0.8.5.post1"]
+                if patch == 0:
+                    _install_requires.append("xformers==0.0.30")
+                else:
+                    _install_requires.append("xformers==0.0.31.post1")
+                extras_require_map["vllm"] = ["vllm>=0.9.0"]
             elif (major, minor) >= (2, 6):
                 _install_requires.pop(_install_requires.index(xformers_version))
-                _install_requires.append(
-                    "xformers==0.0.29.post2"
-                )  # vllm needs post2 w torch 2.6
+                _install_requires.append("xformers==0.0.29.post3")
+                # since we only support 2.6.0+cu126
+                _dependency_links.append("https://download.pytorch.org/whl/cu126")
                 extras_require_map["vllm"] = ["vllm==0.8.5.post1"]
             elif (major, minor) >= (2, 5):
                 _install_requires.pop(_install_requires.index(xformers_version))
@@ -111,14 +114,14 @@ def get_package_version():
 
 
 extras_require = {
-    "flash-attn": ["flash-attn==2.7.4.post1"],
+    "flash-attn": ["flash-attn==2.8.0.post2"],
     "ring-flash-attn": [
-        "flash-attn==2.7.4.post1",
-        "ring-flash-attn>=0.1.4",
+        "flash-attn==2.8.0.post2",
+        "ring-flash-attn>=0.1.5",
         "yunchang==0.6.0",
     ],
     "deepspeed": [
-        "deepspeed==0.17.1",
+        "deepspeed==0.17.2",
         "deepspeed-kernels",
     ],
     "mamba-ssm": [

src/axolotl/__init__.py

@@ -4,4 +4,4 @@ import pkgutil
 
 __path__ = pkgutil.extend_path(__path__, __name__)  # Make this a namespace package
 
-__version__ = "0.11.0.dev"
+__version__ = "0.12.0.dev"

src/axolotl/cli/preprocess.py

@@ -1,5 +1,6 @@
 """CLI to run preprocessing of a dataset."""
 
+import os
 import warnings
 from pathlib import Path
 from typing import Union
@@ -35,6 +36,12 @@ def do_preprocess(cfg: DictDefault, cli_args: PreprocessCliArgs) -> None:
     check_accelerate_default_config()
     check_user_token()
 
+    for key in ["skip_prepare_dataset", "pretraining_dataset"]:
+        if cfg.get("key"):
+            raise ValueError(
+                f"You have set `{key}:`. `preprocess` is not needed. Run the `axolotl train` CLI directly instead."
+            )
+
     if not cfg.dataset_prepared_path:
         msg = (
             Fore.RED
@@ -89,6 +96,7 @@ def do_cli(
         kwargs: Additional keyword arguments to override config file values.
     """
     # pylint: disable=duplicate-code
+    os.environ["AXOLOTL_IS_PREPROCESS"] = "1"
     parsed_cfg = load_cfg(config, **kwargs)
     parsed_cfg.is_preprocess = True
     parser = transformers.HfArgumentParser(PreprocessCliArgs)

src/axolotl/cli/train.py

@@ -109,6 +109,13 @@ def ray_train_func(kwargs: dict):
     # initialize accelerator before model instantiation
     Accelerator(gradient_accumulation_steps=cfg.gradient_accumulation_steps)
 
+    # Register plugins in Ray workers
+    if cfg.get("plugins"):
+        from axolotl.cli.config import plugin_set_cfg, prepare_plugins
+
+        prepare_plugins(cfg)
+        plugin_set_cfg(cfg)
+
     kwargs["cfg"] = cfg
 
     do_train(**kwargs)

src/axolotl/cli/vllm_serve.py

@@ -37,7 +37,6 @@ def do_vllm_serve(
     Returns:
         process_id: the process id of the started VLLM server
     """
-    patch_vllm_worker()
     cfg = load_cfg(config)
     model = cfg.base_model
 
@@ -47,6 +46,9 @@ def do_vllm_serve(
     tensor_parallel_size = (
         cli_args.get("tensor_parallel_size") or cfg.vllm.tensor_parallel_size
     )
+    data_parallel_size = (
+        cli_args.get("data_parallel_size") or cfg.vllm.data_parallel_size
+    )
     host = cli_args.get("host") or cfg.vllm.host
     port = cli_args.get("port") or cfg.vllm.port
     gpu_memory_utilization = (
@@ -68,6 +70,7 @@ def do_vllm_serve(
     vllm_script_args = AxolotlScriptArguments(
         model=model,
         tensor_parallel_size=tensor_parallel_size,
+        data_parallel_size=data_parallel_size,
         host=host,
         port=port,
         gpu_memory_utilization=gpu_memory_utilization,

src/axolotl/common/datasets.py

@@ -75,13 +75,17 @@ def load_datasets(
 
         num_examples = cli_args.debug_num_examples if cli_args else 1
         text_only = cli_args.debug_text_only if cli_args else False
-        train_samples = sample_dataset(train_dataset, num_examples)
-        check_dataset_labels(
-            train_samples,
-            tokenizer,
-            num_examples=num_examples,
-            text_only=text_only,
-        )
+        try:
+            train_samples = sample_dataset(train_dataset, num_examples)
+            check_dataset_labels(
+                train_samples,
+                tokenizer,
+                num_examples=num_examples,
+                text_only=text_only,
+            )
+        except AttributeError:
+            # can't sample iterable datasets
+            pass
 
         LOG.info("printing prompters...")
         for prompter in prompters:

src/axolotl/core/attention/flex_block_mask.py

@@ -0,0 +1,162 @@
+"""
+monkeypatch for flex + packing
+"""
+
+import sys
+from typing import Callable, Optional, Union
+
+import torch
+from torch.nn.attention.flex_attention import BlockMask
+from transformers import Cache, PretrainedConfig
+from transformers.masking_utils import (
+    ALL_MASK_ATTENTION_FUNCTIONS,
+    _preprocess_mask_arguments,
+    and_masks,
+    causal_mask_function,
+    or_masks,
+)
+from transformers.utils import is_torch_greater_or_equal
+
+_is_torch_greater_or_equal_than_2_6 = is_torch_greater_or_equal("2.6", accept_dev=True)
+
+
+def create_causal_mask(
+    config: PretrainedConfig,
+    input_embeds: torch.Tensor,
+    attention_mask: torch.Tensor,
+    cache_position: torch.Tensor,
+    past_key_values: Optional[Cache],
+    or_mask_function: Optional[Callable] = None,
+    and_mask_function: Optional[Callable] = None,
+) -> Optional[Union[torch.Tensor, BlockMask]]:
+    """
+    Create a standard causal mask based on the attention implementation used (stored in the config). If `past_key_values`
+    has an HybridCache structure, this function will return the mask corresponding to one of the "full_attention" layers (to align
+    to what is needed in the `modeling_xxx.py` files).
+
+    Args:
+        config (`PretrainedConfig`):
+            The model config.
+        input_embeds (`torch.Tensor`):
+            The input embeddings of shape (batch_size, query_length, hidden_dim). This is used only to infer the
+            batch size, query length and dtype.
+        attention_mask (`torch.Tensor`, optional):
+            The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length).
+            It can also be an already prepared 4D mask, in which case it is returned as-is.
+        cache_position (`torch.Tensor`):
+            A tensor of shape (query_length,) indicating the current indices of the input sequence elements.
+        past_key_values (`Cache`, optional):
+            The past key values, if we use a cache.
+        or_mask_function (`Callable`, optional):
+            An optional mask function to combine with the causal mask function (by doing the union of both). This is
+            useful to easily overlay another mask on top of the causal one, for example for image tokens handling.
+        and_mask_function (`Callable`, optional):
+            An optional mask function to combine with the causal mask function (by doing the intersection of both). This is
+            useful to easily overlay another mask on top of the causal one, for example for image tokens handling.
+    """
+    # If we have an HybridCache structure, here we want to create the mask for the full layers
+    if (
+        past_key_values
+        and hasattr(past_key_values, "is_sliding")
+        and False in past_key_values.is_sliding
+    ):
+        layer_idx = past_key_values.is_sliding.index(False)
+    else:
+        layer_idx = 0
+
+    original_attention_mask = (
+        None
+        if attention_mask is None
+        else attention_mask.clone().to(cache_position.device)
+    )
+    early_exit, attention_mask, kv_length, kv_offset = _preprocess_mask_arguments(
+        config, input_embeds, attention_mask, cache_position, past_key_values, layer_idx
+    )
+    if early_exit:
+        return attention_mask
+
+    batch_size, total_seq_len = cache_position.shape
+    key_length = total_seq_len
+    document_ids = torch.nn.functional.pad(
+        original_attention_mask, value=0, pad=(0, key_length)
+    )
+
+    batch_size, dtype = input_embeds.shape[0], input_embeds.dtype
+    if attention_mask is not None:
+
+        def causal_doc_mask_mod(
+            batch_idx, head_idx, q_idx, kv_idx
+        ):  # pylint: disable=unused-argument
+            """
+            Defines the logic of a block causal mask by combining both a standard causal mask
+            and a block diagonal document mask.
+            See :func:`~torchtune.modules.attention_utils.create_block_causal_mask`
+            for an illustration.
+            """
+            causal_mask_ = q_idx >= kv_idx  # not valid when decoding
+            document_mask = (
+                document_ids[batch_idx, q_idx] == document_ids[batch_idx, kv_idx]
+            )
+            final_mask = causal_mask_ & document_mask
+            return final_mask
+
+        mask_factory_function = causal_doc_mask_mod
+    else:
+        mask_factory_function = causal_mask_function
+    mask_interface = ALL_MASK_ATTENTION_FUNCTIONS[
+        config._attn_implementation  # pylint: disable=protected-access
+    ]
+
+    # Do not allow skip if we are compiling (this is to match BC)
+    allow_is_causal_skip = (
+        not past_key_values.is_compileable if past_key_values is not None else True
+    )
+
+    # Allow slight deviations from causal mask
+    if or_mask_function is not None:
+        if not _is_torch_greater_or_equal_than_2_6:
+            raise ValueError(
+                "Using `or_mask_function` or `and_mask_function` arguments require torch>=2.6"
+            )
+        mask_factory_function = or_masks(mask_factory_function, or_mask_function)
+        allow_is_causal_skip = False
+    if and_mask_function is not None:
+        if not _is_torch_greater_or_equal_than_2_6:
+            raise ValueError(
+                "Using `or_mask_function` or `and_mask_function` arguments require torch>=2.6"
+            )
+        mask_factory_function = and_masks(mask_factory_function, and_mask_function)
+        allow_is_causal_skip = False
+
+    # We now create the mask
+    causal_mask = mask_interface(
+        batch_size=batch_size,
+        cache_position=cache_position,
+        kv_length=kv_length,
+        kv_offset=kv_offset,
+        mask_function=mask_factory_function,
+        attention_mask=attention_mask,
+        allow_is_causal_skip=allow_is_causal_skip,  # additional kwarg for sdpa
+        dtype=dtype,  # Additional kwarg for eager
+        config=config,  # Pass the config as well, in case someone wants to easily have their own mask_interface
+    )
+    return causal_mask
+
+
+def patch_create_causal_mask(model_type):
+    import transformers.masking_utils
+
+    transformers.masking_utils.create_causal_mask = create_causal_mask
+
+    if model_type:
+        try:
+            # Dynamically import the module and attention class
+            module_path = f"transformers.models.{model_type}.modeling_{model_type}"
+            module = __import__(module_path)
+            module.create_causal_mask = create_causal_mask
+            del sys.modules[module_path]
+        except (ImportError, AttributeError) as e:
+            raise ValueError(
+                f"Could not import attention class for model_type: {model_type}. "
+                f"Error: {str(e)}"
+            ) from e

src/axolotl/core/builders/base.py

@@ -112,13 +112,6 @@ class TrainerBuilderBase(abc.ABC):
             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))
 
@@ -145,6 +138,14 @@ class TrainerBuilderBase(abc.ABC):
 
         callbacks.append(GPUStatsCallback(cfg=self.cfg))
 
+        if self.cfg.profiler_steps:
+            callbacks.append(
+                PytorchProfilerCallback(
+                    steps_to_profile=self.cfg.profiler_steps,
+                    profiler_steps_start=self.cfg.profiler_steps_start,
+                )
+            )
+
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
@@ -219,7 +220,9 @@ class TrainerBuilderBase(abc.ABC):
         if self.cfg.bf16 == "full":
             training_args_kwargs["bf16_full_eval"] = True
         else:
-            training_args_kwargs["bf16"] = self.cfg.bf16 or self.cfg.bfloat16
+            bf16 = self.cfg.bf16 or self.cfg.bfloat16
+            bf16 = bf16 if bf16 is not None else False
+            training_args_kwargs["bf16"] = bf16
 
     def _configure_scheduler(self, training_args_kwargs: dict):
         if self.cfg.lr_scheduler in ["one_cycle", "rex"]:
@@ -416,6 +419,9 @@ class TrainerBuilderBase(abc.ABC):
             torch._dynamo.config.suppress_errors = (  # pylint: disable=protected-access
                 True
             )
+            torch._dynamo.config.accumulated_cache_size_limit = (  # pylint: disable=protected-access
+                256
+            )
             training_args_kwargs["torch_compile"] = self.cfg.torch_compile
             if self.cfg.torch_compile_backend:
                 training_args_kwargs["torch_compile_backend"] = (
@@ -424,8 +430,16 @@ class TrainerBuilderBase(abc.ABC):
             if self.cfg.torch_compile_mode:
                 training_args_kwargs["torch_compile_mode"] = self.cfg.torch_compile_mode
 
+    def _configure_accelerator_config(self, training_args_kwargs: dict):
+        if self.cfg.accelerator_config:
+            training_args_kwargs["accelerator_config"] = self.cfg.accelerator_config
+
     def _configure_gradient_checkpointing(self, training_args_kwargs: dict):
-        if self.cfg.gradient_checkpointing:
+        if self.cfg.activation_offloading is True:
+            # don't use the HF gradient checkpointing, manually wrap
+            training_args_kwargs["gradient_checkpointing"] = False
+            training_args_kwargs["activation_offloading"] = True
+        elif self.cfg.gradient_checkpointing:
             training_args_kwargs["gradient_checkpointing"] = (
                 self.cfg.gradient_checkpointing
             )
@@ -499,10 +513,15 @@ class TrainerBuilderBase(abc.ABC):
         if self.cfg.reward_model or self.cfg.rl:
             training_args_kwargs["max_length"] = self.cfg.sequence_len
 
+        if self.cfg.fsdp_config or self.cfg.fsdp:
+            training_args_kwargs["fsdp_config"] = self.cfg.fsdp_config
+            training_args_kwargs["fsdp"] = self.cfg.fsdp if self.cfg.fsdp else True
+
         self._configure_reporting(training_args_kwargs)
         self._configure_hub_parameters(training_args_kwargs)
         self._configure_scheduler(training_args_kwargs)
         self._configure_optimizer(training_args_kwargs, trainer_kwargs)
         self._configure_torch_compile(training_args_kwargs)
+        self._configure_accelerator_config(training_args_kwargs)
 
         return training_args_kwargs, trainer_kwargs

src/axolotl/core/builders/causal.py

@@ -151,14 +151,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs, trainer_kwargs = self._set_base_training_args(
             total_num_steps
         )
-
-        if self.cfg.fsdp:
-            training_arguments_kwargs["fsdp"] = self.cfg.fsdp
-            if self.cfg.fsdp_config:
-                training_arguments_kwargs["fsdp_config"] = {
-                    k.lstrip("fsdp_"): v for k, v in dict(self.cfg.fsdp_config).items()
-                }
-
         if self.cfg.adapter == "qlora":
             training_arguments_kwargs["qlora"] = True
 
@@ -245,14 +237,27 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs["curriculum_sampling"] = self.cfg.curriculum_sampling
 
         training_arguments_kwargs["sample_packing"] = bool(self.cfg.sample_packing)
+        training_arguments_kwargs["sample_packing_drop_attention_mask"] = bool(
+            self.cfg.flash_attention
+            or self.cfg.xformers_attention
+            or self.cfg.flex_attention
+        )
         training_arguments_kwargs["multipack_real_batches"] = (
             self.cfg.multipack_real_batches
             if self.cfg.multipack_real_batches is not None
-            else not self.cfg.flash_attention
+            else not (
+                self.cfg.flash_attention
+                or self.cfg.flex_attention
+                or self.cfg.xformers_attention
+            )
         )
         training_arguments_kwargs["eval_sample_packing"] = bool(
             self.cfg.eval_sample_packing
         )
+        if self.cfg.sample_packing_sequentially is not None:
+            training_arguments_kwargs["sample_packing_sequentially"] = (
+                self.cfg.sample_packing_sequentially
+            )
         if self.cfg.sample_packing_bin_size is not None:
             training_arguments_kwargs["sample_packing_bin_size"] = (
                 self.cfg.sample_packing_bin_size
@@ -305,11 +310,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 self.cfg.neftune_noise_alpha
             )
 
-        if self.cfg.accelerator_config:
-            training_arguments_kwargs["accelerator_config"] = (
-                self.cfg.accelerator_config
-            )
-
         if self.cfg.image_size:
             training_arguments_kwargs["image_size"] = self.cfg.image_size
         if self.cfg.image_resize_algorithm:
@@ -413,7 +413,8 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 or self.cfg.micro_batch_size > 1
             ):
                 return DataCollatorForSeq2Seq(self.tokenizer, **kwargs)
-            return None
+            if not (self.cfg.sample_packing and self.cfg.pretrain_multipack_attn):
+                return None
 
         if self.cfg.model_config_type == "mamba":
             return MambaDataCollator(tokenizer=self.tokenizer)

src/axolotl/core/builders/rl.py

@@ -208,7 +208,7 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             callbacks=self.get_callbacks(),
             **trainer_kwargs,
         )
-        if self.cfg.fsdp:
+        if self.cfg.fsdp_config or self.cfg.fsdp:
             ensure_dtype(trainer.model, dtype=self.cfg.torch_dtype)
             if self.cfg.rl in [RLType.DPO, RLType.IPO] and trainer.ref_model:
                 ensure_dtype(trainer.ref_model, dtype=self.cfg.torch_dtype)
@@ -218,21 +218,3 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             trainer.add_callback(callback)
 
         return trainer
-
-
-class HFPPOTrainerBuilder(TrainerBuilderBase):
-    """
-    HF Factory class for PPO Trainer
-    """
-
-    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(self, total_num_steps):
-        # TODO: build PPOConfig
-        raise NotImplementedError("PPO trainer builder is not implemented yet.")

src/axolotl/core/trainers/__init__.py

@@ -14,5 +14,4 @@ from .trl import (
     AxolotlORPOTrainer,
     AxolotlPRMTrainer,
     AxolotlRewardTrainer,
-    TRLPPOTrainer,
 )

src/axolotl/core/trainers/base.py

@@ -25,8 +25,10 @@ from trl.trainer.utils import pad_to_length
 from typing_extensions import override
 
 from axolotl.core.trainers.mixins import (
+    ActivationOffloadingMixin,
     CheckpointSaveMixin,
     OptimizerMixin,
+    PackingMixin,
     RngLoaderMixin,
     SchedulerMixin,
 )
@@ -42,7 +44,13 @@ LOG = get_logger(__name__)
 
 
 class AxolotlTrainer(
-    SchedulerMixin, OptimizerMixin, RngLoaderMixin, CheckpointSaveMixin, Trainer
+    PackingMixin,
+    SchedulerMixin,
+    OptimizerMixin,
+    RngLoaderMixin,
+    CheckpointSaveMixin,
+    ActivationOffloadingMixin,
+    Trainer,
 ):
     """Extend the base Trainer for axolotl helpers"""
 
@@ -69,18 +77,6 @@ class AxolotlTrainer(
         if self.args.orpo_alpha:
             self.loss_fct = torch.nn.CrossEntropyLoss(reduction="none")
 
-    def _wrap_model(self, model, training=True, dataloader=None):
-        if self.args.torch_compile:
-            torch._dynamo.config.accumulated_cache_size_limit = (  # pylint: disable=protected-access
-                256
-            )
-            model = torch.compile(
-                model,
-                backend=self.args.torch_compile_backend,
-                mode=self.args.torch_compile_mode,
-            )
-        return super()._wrap_model(model, training=training, dataloader=dataloader)
-
     def _create_multipack_sampler(
         self, base_sampler: Sampler, dataset: Dataset
     ) -> MultipackBatchSampler:
@@ -116,6 +112,7 @@ class AxolotlTrainer(
             sequential=self.args.sample_packing_sequentially,
             drop_last=True,
             num_processes=self.args.dataset_num_proc,
+            mp_start_method=self.args.sample_packing_mp_start_method or "fork",
         )
 
         len(sampler)
@@ -205,6 +202,14 @@ class AxolotlTrainer(
 
         if dataset.column_names and "length" in dataset.column_names:
             dataset = dataset.remove_columns(["length"])
+        if (
+            dataset.column_names
+            and "position_ids" in dataset.column_names
+            and "attention_mask" in dataset.column_names
+            and self.args.sample_packing
+            and self.args.sample_packing_drop_attention_mask
+        ):
+            dataset = dataset.remove_columns(["attention_mask"])
 
         if isinstance(dataset, datasets.Dataset):
             if is_training:

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

@@ -28,7 +28,7 @@ class DPOStrategy:
         training_args_kwargs["max_completion_length"] = None
         training_args_kwargs["max_length"] = cfg.sequence_len
         training_args_kwargs["max_prompt_length"] = cfg.sequence_len
-        training_args_kwargs["generate_during_eval"] = cfg.use_wandb
+        training_args_kwargs["generate_during_eval"] = cfg.dpo_generate_during_eval
         if cfg.dpo_use_weighting is not None:
             training_args_kwargs["use_weighting"] = cfg.dpo_use_weighting
         if cfg.dpo_padding_free is not None: