fix compile

.coveragerc

@@ -2,6 +2,7 @@
 source = axolotl
 omit =
     */tests/*
+    setup.py
 
 [report]
 exclude_lines =

.github/CONTRIBUTING.md

@@ -29,18 +29,13 @@ PRs are **greatly welcome**!
 2. Set up the development environment by following the instructions in the [README.md](https://github.com/axolotl-ai-cloud/axolotl/tree/main/README.md) file.
 3. Explore the codebase, run tests, and verify that everything works as expected.
 
-Please run the below to setup:
-
+Please run below to setup env
 ```bash
-git clone https://github.com/axolotl-ai-cloud/axolotl.git
-cd axolotl
+pip3 install -r requirements-dev.txt -r requirements-tests.txt
+pre-commit install
 
-uv sync --dev && uv pip install flash-attn --no-build-isolation
-source .venv/bin/activate
-
-pre-commit install  # install pre-commit hooks
-
-pytest tests/  # optional; run test suite
+# test
+pytest tests/
 ```
 
 ## How to Contribute

.github/workflows/base.yml

@@ -39,6 +39,13 @@ jobs:
             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: "126"
+            cuda_version: 12.6.3
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.7.0
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-base"
           - cuda: "126"
             cuda_version: 12.6.3
             cudnn_version: ""
@@ -98,9 +105,7 @@ jobs:
           context: .
           file: ./docker/${{ matrix.dockerfile }}
           push: ${{ github.event_name != 'pull_request' }}
-          tags: |
-            ${{ steps.metadata.outputs.tags }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
-            ${{ steps.metadata.outputs.tags }}-base-uv-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
+          tags: ${{ steps.metadata.outputs.tags }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
           labels: ${{ steps.metadata.outputs.labels }}
           build-args: |
             CUDA_VERSION=${{ matrix.cuda_version }}

.github/workflows/docs.yml

@@ -20,14 +20,10 @@ jobs:
           uses: actions/setup-python@v5
           with:
             python-version: '3.11'
-        - name: Install uv
-          uses: astral-sh/setup-uv@v4
-          with:
-            version: "latest"
         - name: Install dependencies
           run: |
-            uv pip install --system jupyter quartodoc
-            uv pip install --system -e .
+            python3 -m pip install jupyter quartodoc
+            python3 -m pip install -e .
         - name: Build autodoc
           run: quartodoc build
         - name: Publish to GitHub Pages (and render)

.github/workflows/lint.yml

@@ -6,7 +6,7 @@ on:
       types: [opened, synchronize, reopened, ready_for_review]
       paths:
        - '**.py'
-       - 'pyproject.toml'
+       - 'requirements.txt'
        - '.github/workflows/*.yml'
        - "*.[q]md"
        - "examples/**/*.y[a]?ml"
@@ -23,4 +23,5 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: "3.11"
+          cache: 'pip' # caching pip dependencies
       - uses: pre-commit/action@v3.0.1

.github/workflows/main.yml

@@ -20,6 +20,11 @@ jobs:
             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:
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
@@ -68,8 +73,6 @@ jobs:
             PYTORCH_VERSION=${{ matrix.pytorch }}
             AXOLOTL_ARGS=${{ matrix.axolotl_args }}
             AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}
-            GIT_REF=${{ github.ref }}
-            GIT_SHA=${{ github.sha }}
           file: ./docker/Dockerfile
           push: ${{ github.event_name != 'pull_request' }}
           tags: |
@@ -90,6 +93,11 @@ jobs:
             python_version: "3.11"
             pytorch: 2.6.0
             axolotl_extras:
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.0
+            axolotl_extras:
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
@@ -140,8 +148,6 @@ jobs:
           build-args: |
             BASE_TAG=${{ github.ref_type == 'tag' && 'main' || github.ref_name }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
             CUDA=${{ matrix.cuda }}
-            GIT_REF=${{ github.ref }}
-            GIT_SHA=${{ github.sha }}
           file: ./docker/Dockerfile-cloud
           push: ${{ github.event_name != 'pull_request' }}
           tags: |
@@ -207,8 +213,6 @@ jobs:
           build-args: |
             BASE_TAG=${{ github.ref_type == 'tag' && 'main' || github.ref_name }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
             CUDA=${{ matrix.cuda }}
-            GIT_REF=${{ github.ref }}
-            GIT_SHA=${{ github.sha }}
           file: ./docker/Dockerfile-cloud-no-tmux
           push: ${{ github.event_name != 'pull_request' }}
           tags: |

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

@@ -4,6 +4,8 @@ on:
   pull_request:
     paths:
       - 'tests/e2e/multigpu/**.py'
+      - 'requirements.txt'
+      - 'setup.py'
       - 'pyproject.toml'
       - '.github/workflows/multi-gpu-e2e.yml'
       - 'src/axolotl/core/trainers/mixins/sequence_parallel.py'
@@ -54,17 +56,13 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: "3.11"
-      - name: Install uv
-        uses: astral-sh/setup-uv@v4
-        with:
-          version: "latest"
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2 protobuf
+          pip install modal==1.0.2 jinja2
       - name: Update env vars
         run: |
-          echo "BASE_TAG=${{ github.ref_name }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
+          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
           echo "PYTORCH_VERSION=${{ matrix.pytorch}}" >> $GITHUB_ENV
           echo "AXOLOTL_ARGS=${{ matrix.axolotl_args}}" >> $GITHUB_ENV
           echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV
@@ -74,4 +72,4 @@ jobs:
           echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         run: |
-          modal run -m cicd.multigpu
+          modal run cicd.multigpu

.github/workflows/nightlies.yml

@@ -52,8 +52,6 @@ jobs:
             CUDA=${{ matrix.cuda }}
             PYTORCH_VERSION=${{ matrix.pytorch }}
             AXOLOTL_ARGS=${{ matrix.axolotl_args }}
-            GIT_REF=${{ github.ref }}
-            GIT_SHA=${{ github.sha }}
           file: ./docker/Dockerfile
           push: ${{ github.event_name != 'pull_request' }}
           tags: |
@@ -104,8 +102,6 @@ jobs:
           build-args: |
             BASE_TAG=${{ github.ref_name }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
             CUDA=${{ matrix.cuda }}
-            GIT_REF=${{ github.ref }}
-            GIT_SHA=${{ github.sha }}
           file: ./docker/Dockerfile-cloud
           push: ${{ github.event_name != 'pull_request' }}
           tags: |

.github/workflows/precommit-autoupdate.yml

@@ -18,15 +18,10 @@ jobs:
         with:
           python-version: '3.11'
 
-      - name: Install uv
-        uses: astral-sh/setup-uv@v4
-        with:
-          version: "latest"
-
       - name: Update pre-commit hooks
         id: update
         run: |
-          uv pip install --system pre-commit
+          pip install pre-commit
           pre-commit autoupdate
           if [[ -n $(git status --porcelain) ]]; then
             echo "changes=true" >> $GITHUB_OUTPUT

.github/workflows/preview-docs.yml

@@ -40,15 +40,10 @@ jobs:
         with:
           python-version: '3.11'
 
-      - name: Install uv
-        uses: astral-sh/setup-uv@v4
-        with:
-          version: "latest"
-
       - name: Install dependencies
         run: |
-          uv pip install --system jupyter quartodoc
-          uv pip install --system -e .
+          python3 -m pip install jupyter quartodoc
+          python3 -m pip install -e .
 
       - name: Build autodoc
         run: quartodoc build

.github/workflows/pypi.yml

@@ -38,24 +38,23 @@ jobs:
         with:
           python-version: "3.11"
 
-      - name: Install uv
-        uses: astral-sh/setup-uv@v4
-        with:
-          version: "latest"
-
       - name: Install dependencies
         run: |
-          uv pip install --system wheel packaging==23.2
-          uv pip install --system --no-build-isolation -e ".[dev]"
+          pip3 install wheel packaging==23.2
+          pip3 install --no-build-isolation -e .
+          pip3 install -r requirements-dev.txt -r requirements-tests.txt
 
       - name: Extract tag name
         id: tag
-        run: echo "TAG_NAME=$(echo "$GITHUB_REF" | cut -d / -f 3)" >> "$GITHUB_OUTPUT"
+        run: echo ::set-output name=TAG_NAME::$(echo $GITHUB_REF | cut -d / -f 3)
 
-      - name: Build package
+      - name: Update version in setup.py
         run: |
-          uv pip install --system build
-          python -m build
+          sed -i -E 's/version="([0-9.]+)",/version="${{ steps.tag.outputs.TAG_NAME }}",/g' setup.py
+
+      - name: Build a source dist
+        run: |
+          python setup.py sdist
 
       - name: Publish package distributions to PyPI
         uses: pypa/gh-action-pypi-publish@release/v1

.github/workflows/tests-nightly.yml

@@ -13,6 +13,7 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: "3.11"
+          cache: 'pip' # caching pip dependencies
       - uses: pre-commit/action@v3.0.1
         env:
           SKIP: no-commit-to-branch
@@ -42,30 +43,32 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.python_version }}
+          cache: 'pip' # caching pip dependencies
 
-      - name: Install uv
-        uses: astral-sh/setup-uv@v4
-        with:
-          version: "latest"
+      - name: upgrade pip
+        run: |
+          pip3 install --upgrade pip
+          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
 
       - name: Install PyTorch
         run: |
-          uv pip install --system torch==${{ matrix.pytorch_version }} torchvision
+          pip3 install torch==${{ matrix.pytorch_version }} torchvision
 
-      - name: Update pyproject.toml for nightly builds
+      - name: Update requirements.txt
         run: |
-          sed -i 's#"transformers==.*"#"transformers @ git+https://github.com/huggingface/transformers.git@main"#' pyproject.toml
-          sed -i 's#"peft==.*"#"peft @ git+https://github.com/huggingface/peft.git@main"#' pyproject.toml
-          sed -i 's#"accelerate==.*"#"accelerate @ git+https://github.com/huggingface/accelerate.git@main"#' pyproject.toml
-          sed -i 's#"trl==.*"#"trl @ git+https://github.com/huggingface/trl.git@main"#' pyproject.toml
-          sed -i 's#"datasets==.*"#"datasets @ git+https://github.com/huggingface/datasets.git@main"#' pyproject.toml
+          sed -i 's#^transformers.*#transformers @ git+https://github.com/huggingface/transformers.git@main#' requirements.txt
+          sed -i 's#^peft.*#peft @ git+https://github.com/huggingface/peft.git@main#' requirements.txt
+          sed -i 's#^accelerate.*#accelerate @ git+https://github.com/huggingface/accelerate.git@main#' requirements.txt
+          sed -i 's#^trl.*#trl @ git+https://github.com/huggingface/trl.git@main#' requirements.txt
+          sed -i 's#^datasets.*#datasets @ git+https://github.com/huggingface/datasets.git@main#' requirements.txt
 
       - name: Install dependencies
         run: |
-          uv pip show --system torch
-          uv pip install --system --no-build-isolation -e ".[dev]"
+          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: |
@@ -81,6 +84,9 @@ jobs:
           pytest -v --durations=10 tests/patched/
           pytest -v --durations=10 tests/cli/
 
+      - name: cleanup pip cache
+        run: |
+          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
 
   docker-e2e-tests:
     if: github.repository_owner == 'axolotl-ai-cloud'
@@ -114,16 +120,13 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: "3.11"
-      - name: Install uv
-        uses: astral-sh/setup-uv@v4
-        with:
-          version: "latest"
       - name: Install Modal
         run: |
-          uv pip install --system modal==1.0.2 jinja2
+          python -m pip install --upgrade pip
+          pip install modal==1.0.2 jinja2
       - name: Update env vars
         run: |
-          echo "BASE_TAG=main-base-uv-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
+          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
           echo "PYTORCH_VERSION=${{ matrix.pytorch}}" >> $GITHUB_ENV
           echo "AXOLOTL_ARGS=${{ matrix.axolotl_args}}" >> $GITHUB_ENV
           echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV
@@ -133,7 +136,7 @@ jobs:
           echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         run: |
-          modal run -m cicd.e2e_tests
+          modal run cicd.e2e_tests
   docker-e2e-multigpu-tests:
     if: github.repository_owner == 'axolotl-ai-cloud'
     # this job needs to be run on self-hosted GPU runners...
@@ -159,16 +162,13 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: "3.11"
-      - name: Install uv
-        uses: astral-sh/setup-uv@v4
-        with:
-          version: "latest"
       - name: Install Modal
         run: |
-          uv pip install --system modal==1.0.2 jinja2
+          python -m pip install --upgrade pip
+          pip install modal==1.0.2 jinja2
       - name: Update env vars
         run: |
-          echo "BASE_TAG=main-base-uv-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
+          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
           echo "PYTORCH_VERSION=${{ matrix.pytorch}}" >> $GITHUB_ENV
           echo "AXOLOTL_ARGS=${{ matrix.axolotl_args}}" >> $GITHUB_ENV
           echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV

.github/workflows/tests.yml

@@ -7,16 +7,18 @@ on:
       - "main"
     paths:
       - '**.py'
-      - 'pyproject.toml'
+      - 'requirements.txt'
       - '.github/workflows/*.yml'
+      - 'requirements-tests.txt'
       - 'cicd/cicd.sh'
       - 'cicd/Dockerfile.jinja'
   pull_request:
       types: [opened, synchronize, reopened, ready_for_review]
       paths:
        - '**.py'
-       - 'pyproject.toml'
+       - 'requirements.txt'
        - '.github/workflows/*.yml'
+       - 'requirements-tests.txt'
        - 'cicd/cicd.sh'
        - 'cicd/Dockerfile.jinja'
   workflow_dispatch:
@@ -39,6 +41,7 @@ jobs:
       - uses: actions/setup-python@v5
         with:
           python-version: "3.11"
+          cache: 'pip' # caching pip dependencies
       - uses: pre-commit/action@v3.0.1
         env:
           SKIP: no-commit-to-branch
@@ -69,25 +72,24 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.python_version }}
+          cache: 'pip' # caching pip dependencies
 
-      - name: Install uv
-        uses: astral-sh/setup-uv@v4
-        with:
-          version: "latest"
+      - name: upgrade pip
+        run: |
+          pip3 install --upgrade pip
+          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
 
       - name: Install PyTorch
         run: |
-          uv pip install --system torch==${{ matrix.pytorch_version }} torchvision
+          pip3 install torch==${{ matrix.pytorch_version }} torchvision
 
       - name: Install dependencies
         run: |
-          uv pip show --system torch
-          uv pip install --system wheel
-          printf "torch==${{ matrix.pytorch_version }}\n" > torch-constraints.txt
-          uv pip install --system --no-cache-dir --no-build-isolation -e ".[dev]" --constraints torch-constraints.txt
-          set -o pipefail
-          python scripts/unsloth_install.py | bash
-          python scripts/cutcrossentropy_install.py | bash
+          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: |
@@ -103,10 +105,10 @@ jobs:
 
       - name: Run tests
         run: |
-          python -m pytest -v --durations=10 -n 8 --dist loadfile --cov=axolotl --cov-report=xml --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ --ignore=tests/monkeypatch/ tests/
-          python -m pytest -v --durations=10 -n 8 --cov=axolotl --cov-append --cov-report=xml tests/monkeypatch/
-          python -m pytest -v --durations=10 -n 8 --cov=axolotl --cov-append --cov-report=xml tests/patched/
-          python -m pytest -v --durations=10 -n 8 --cov=axolotl --cov-append --cov-report=xml tests/cli/
+          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ --ignore=tests/monkeypatch/ tests/ --cov=axolotl --cov-report=xml
+          pytest -v --durations=10 tests/monkeypatch/ --cov=axolotl --cov-append --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
@@ -116,6 +118,9 @@ jobs:
           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 {} \;
 
   pytest-sdist:
     name: PyTest from Source Dist
@@ -142,26 +147,25 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.python_version }}
+          cache: 'pip' # caching pip dependencies
 
-      - name: Install uv
-        uses: astral-sh/setup-uv@v4
-        with:
-          version: "latest"
+      - name: upgrade pip
+        run: |
+          pip3 install --upgrade pip
+          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 setuptools_scm build wheel
 
       - name: Install PyTorch
         run: |
-          uv pip install --system torch==${{ matrix.pytorch_version }} torchvision
+          pip3 install torch==${{ matrix.pytorch_version }} torchvision
 
       - name: Install dependencies
         run: |
-          uv pip show --system torch
-          uv pip install --system wheel build setuptools_scm
-          python -m build --sdist
-          printf "torch==${{ matrix.pytorch_version }}\n" > torch-constraints.txt
-          tarball_path=$(echo dist/axolotl*.tar.gz)
-          uv pip install --no-cache-dir --no-build-isolation --system "${tarball_path}[dev]" --constraints torch-constraints.txt
+          pip3 show torch
+          python -m build --no-isolation --sdist
+          pip3 install --no-build-isolation dist/axolotl*.tar.gz
           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: |
@@ -176,9 +180,13 @@ jobs:
 
       - name: Run tests
         run: |
-          python -m pytest -v --durations=10 -n 8 --dist loadfile --cov=axolotl --cov-report=xml --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ --ignore=tests/monkeypatch/ tests/
-          python -m pytest -v --durations=10 -n 8 --cov=axolotl --cov-append --cov-report=xml tests/monkeypatch/
-          python -m pytest -v --durations=10 -n 8 tests/cli/
+          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ --ignore=tests/monkeypatch/ tests/ --cov=axolotl --cov-report=xml
+          pytest -v --durations=10 tests/monkeypatch/ --cov=axolotl --cov-append --cov-report=xml
+          pytest -v --durations=10 tests/cli/
+
+      - name: cleanup pip cache
+        run: |
+          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
 
   gate-skip-e2e:
     needs: [pre-commit, pytest, pytest-sdist]
@@ -235,7 +243,7 @@ jobs:
             pytorch: 2.7.1
             num_gpus: 1
             axolotl_extras:
-            dockerfile: "Dockerfile.jinja"
+            dockerfile: "Dockerfile-uv.jinja"
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -243,17 +251,13 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: "3.11"
-      - name: Install uv
-        uses: astral-sh/setup-uv@v4
-        with:
-          version: "latest"
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2 protobuf
+          pip install modal==1.0.2 jinja2
       - name: Update env vars
         run: |
-          echo "BASE_TAG=${{ github.ref_name }}-base-uv-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
+          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
           echo "PYTORCH_VERSION=${{ matrix.pytorch}}" >> $GITHUB_ENV
           echo "AXOLOTL_ARGS=${{ matrix.axolotl_args}}" >> $GITHUB_ENV
           echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV
@@ -308,17 +312,13 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: "3.11"
-      - name: Install uv
-        uses: astral-sh/setup-uv@v4
-        with:
-          version: "latest"
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2 protobuf
+          pip install modal==1.0.2 jinja2
       - name: Update env vars
         run: |
-          echo "BASE_TAG=${{ github.ref_name }}-base-uv-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
+          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
           echo "PYTORCH_VERSION=${{ matrix.pytorch}}" >> $GITHUB_ENV
           echo "AXOLOTL_ARGS=${{ matrix.axolotl_args}}" >> $GITHUB_ENV
           echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV
@@ -355,17 +355,13 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: "3.11"
-      - name: Install uv
-        uses: astral-sh/setup-uv@v4
-        with:
-          version: "latest"
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2 protobuf
+          pip install modal==1.0.2 jinja2
       - name: Update env vars
         run: |
-          echo "BASE_TAG=${{ github.ref_name }}-base-uv-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
+          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
           echo "PYTORCH_VERSION=${{ matrix.pytorch}}" >> $GITHUB_ENV
           echo "AXOLOTL_ARGS=${{ matrix.axolotl_args}}" >> $GITHUB_ENV
           echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV

.gitignore

@@ -191,5 +191,5 @@ out/
 # vim
 *.swp
 
-# setuptools-scm generated version file
+# scm auto-versioning
 src/axolotl/_version.py

.runpod/Dockerfile

@@ -1,8 +1,9 @@
 FROM axolotlai/axolotl-cloud:main-py3.11-cu124-2.6.0
 
 COPY .runpod/requirements.txt /requirements.txt
-RUN curl -LsSf https://astral.sh/uv/install.sh | sh && \
-    /root/.local/bin/uv pip install --system -r /requirements.txt
+RUN --mount=type=cache,target=/root/.cache/pip \
+    python3 -m pip install --upgrade pip && \
+    python3 -m pip install --upgrade -r /requirements.txt
 
 # Environment settings
 ARG BASE_VOLUME="/runpod-volume"

MANIFEST.in

@@ -1,5 +1,6 @@
-include pyproject.toml
+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 src/axolotl *.py
+recursive-include axolotl *.py

README.md

@@ -65,9 +65,15 @@ Features:
 - **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.
 
+
+
 ## 🚀 Quick Start - LLM Fine-tuning in Minutes
 
-**Requirements**: NVIDIA GPU (Ampere+) or AMD GPU, Python 3.11+
+**Requirements**:
+
+- NVIDIA GPU (Ampere or newer for `bf16` and Flash Attention) or AMD GPU
+- Python 3.11
+- PyTorch ≥2.6.0
 
 ### Google Colab
 
@@ -75,35 +81,15 @@ Features:
 
 ### Installation
 
-#### Project setup (uv add)
+#### Using pip
 
 ```bash
-# Install uv
-curl -LsSf https://astral.sh/uv/install.sh | sh
-
-# Initialize or enter your project
-uv init my-project && cd my-project
-uv add axolotl
-uv pip install flash-attn --no-build-isolation
-source .venv/bin/activate
+pip3 install -U packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install --no-build-isolation axolotl[flash-attn,deepspeed]
 
 # Download example axolotl configs, deepspeed configs
 axolotl fetch examples
-axolotl fetch deepspeed_configs  # optional
-```
-
-#### Quick try (uv pip)
-
-```bash
-# Install uv if needed
-curl -LsSf https://astral.sh/uv/install.sh | sh
-
-uv pip install axolotl
-uv pip install flash-attn --no-build-isolation
-
-# Download example axolotl configs, deepspeed configs
-axolotl fetch examples
-axolotl fetch deepspeed_configs  # optional
+axolotl fetch deepspeed_configs  # OPTIONAL
 ```
 
 #### Using Docker

_quarto.yml

@@ -267,7 +267,6 @@ website:
             - docs/dataset_loading.qmd
             - docs/qat.qmd
             - docs/quantize.qmd
-            - docs/optimizations.qmd
 
         - section: "Core Concepts"
           contents:
@@ -286,6 +285,7 @@ website:
             - docs/custom_integrations.qmd
             - docs/sequence_parallelism.qmd
             - docs/gradient_checkpointing.qmd
+            - docs/moe_backends.md
             - docs/nd_parallelism.qmd
 
         - section: "Troubleshooting"

cicd/Dockerfile-uv.jinja

@@ -0,0 +1,52 @@
+FROM axolotlai/axolotl-base-uv:{{ BASE_TAG }}
+
+ENV TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 9.0+PTX"
+ENV AXOLOTL_EXTRAS="{{ AXOLOTL_EXTRAS }}"
+ENV AXOLOTL_ARGS="{{ AXOLOTL_ARGS }}"
+ENV CUDA="{{ CUDA }}"
+ENV PYTORCH_VERSION="{{ PYTORCH_VERSION }}"
+ENV GITHUB_REF="{{ GITHUB_REF }}"
+ENV GITHUB_SHA="{{ GITHUB_SHA }}"
+ENV NIGHTLY_BUILD="{{ NIGHTLY_BUILD }}"
+ENV HF_HOME="{{ HF_HOME }}"
+
+RUN apt-get update && \
+    apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev ibverbs-providers ibverbs-utils infiniband-diags librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm
+
+WORKDIR /workspace
+
+RUN git clone --depth=1 https://github.com/axolotl-ai-cloud/axolotl.git
+
+WORKDIR /workspace/axolotl
+
+RUN git fetch origin +$GITHUB_REF && \
+    git checkout FETCH_HEAD
+
+# If AXOLOTL_EXTRAS is set, append it in brackets
+RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
+        sed -i 's#^transformers.*#transformers @ git+https://github.com/huggingface/transformers.git@main#' requirements.txt; \
+        sed -i 's#^peft.*#peft @ git+https://github.com/huggingface/peft.git@main#' requirements.txt; \
+        sed -i 's#^accelerate.*#accelerate @ git+https://github.com/huggingface/accelerate.git@main#' requirements.txt; \
+        sed -i 's#^trl.*#trl @ git+https://github.com/huggingface/trl.git@main#' requirements.txt; \
+        sed -i 's#^datasets.*#datasets @ git+https://github.com/huggingface/datasets.git@main#' requirements.txt; \
+    fi
+
+RUN uv pip install packaging==23.2 setuptools==75.8.0
+RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
+        uv pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
+    else \
+        uv pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray] $AXOLOTL_ARGS; \
+    fi
+
+RUN python scripts/unsloth_install.py --uv | sh
+RUN python scripts/cutcrossentropy_install.py --uv | sh
+
+# So we can test the Docker image
+RUN uv pip install -r requirements-dev.txt -r requirements-tests.txt
+
+# fix so that git fetch/pull from remote works
+RUN git config remote.origin.fetch "+refs/heads/*:refs/remotes/origin/*" && \
+    git config --get remote.origin.fetch
+
+# helper for huggingface-login cli
+RUN git config --global credential.helper store

cicd/Dockerfile.jinja

@@ -1,10 +1,6 @@
-FROM axolotlai/axolotl-base-uv:{{ BASE_TAG }}
+FROM axolotlai/axolotl-base:{{ BASE_TAG }}
 
-SHELL ["/bin/bash", "-euxo", "pipefail", "-c"]
-
-ARG VENV_PYTHON="/workspace/axolotl-venv/bin/python"
-
-ENV TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 9.0+PTX"
+ENV TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX"
 ENV AXOLOTL_EXTRAS="{{ AXOLOTL_EXTRAS }}"
 ENV AXOLOTL_ARGS="{{ AXOLOTL_ARGS }}"
 ENV CUDA="{{ CUDA }}"
@@ -13,7 +9,7 @@ ENV GITHUB_REF="{{ GITHUB_REF }}"
 ENV GITHUB_SHA="{{ GITHUB_SHA }}"
 ENV NIGHTLY_BUILD="{{ NIGHTLY_BUILD }}"
 ENV HF_HOME="{{ HF_HOME }}"
-ENV VENV_PYTHON=$VENV_PYTHON
+ENV AXOLOTL_DATASET_PROCESSES="8"
 
 RUN apt-get update && \
     apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev ibverbs-providers ibverbs-utils infiniband-diags librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm
@@ -29,27 +25,25 @@ RUN git fetch origin +$GITHUB_REF && \
 
 # If AXOLOTL_EXTRAS is set, append it in brackets
 RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
-        sed -i 's#"transformers[^"]*"#"transformers @ git+https://github.com/huggingface/transformers.git@main"#' pyproject.toml; \
-        sed -i 's#"peft[^"]*"#"peft @ git+https://github.com/huggingface/peft.git@main"#' pyproject.toml; \
-        sed -i 's#"accelerate[^"]*"#"accelerate @ git+https://github.com/huggingface/accelerate.git@main"#' pyproject.toml; \
-        sed -i 's#"trl[^"]*"#"trl @ git+https://github.com/huggingface/trl.git@main"#' pyproject.toml; \
-        sed -i 's#"datasets[^"]*"#"datasets @ git+https://github.com/huggingface/datasets.git@main"#' pyproject.toml; \
+        sed -i 's#^transformers.*#transformers @ git+https://github.com/huggingface/transformers.git@main#' requirements.txt; \
+        sed -i 's#^peft.*#peft @ git+https://github.com/huggingface/peft.git@main#' requirements.txt; \
+        sed -i 's#^accelerate.*#accelerate @ git+https://github.com/huggingface/accelerate.git@main#' requirements.txt; \
+        sed -i 's#^trl.*#trl @ git+https://github.com/huggingface/trl.git@main#' requirements.txt; \
+        sed -i 's#^datasets.*#datasets @ git+https://github.com/huggingface/datasets.git@main#' requirements.txt; \
     fi
 
-RUN uv pip install --python "$VENV_PYTHON" packaging==23.2 setuptools==75.8.0 pip
+RUN pip install packaging==23.2 setuptools==75.8.0
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        uv pip install --python "$VENV_PYTHON" --no-build-isolation -e .[ring-flash-attn,optimizers,ray,${AXOLOTL_EXTRAS}] $AXOLOTL_ARGS; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \
-        uv pip install --python "$VENV_PYTHON" --no-build-isolation -e .[ring-flash-attn,optimizers,ray] $AXOLOTL_ARGS; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray] $AXOLOTL_ARGS; \
     fi
 
-RUN uv pip install --python "$VENV_PYTHON" --no-build-isolation flash-attn $AXOLOTL_ARGS
-
-RUN "$VENV_PYTHON" scripts/unsloth_install.py | sh
-RUN "$VENV_PYTHON" scripts/cutcrossentropy_install.py | sh
+RUN python scripts/unsloth_install.py | sh
+RUN python scripts/cutcrossentropy_install.py | sh
 
 # So we can test the Docker image
-RUN uv pip install --python "$VENV_PYTHON" -e ".[dev]"
+RUN pip install -r requirements-dev.txt -r requirements-tests.txt
 
 # fix so that git fetch/pull from remote works
 RUN git config remote.origin.fetch "+refs/heads/*:refs/remotes/origin/*" && \

cicd/cicd.sh

@@ -4,7 +4,7 @@ set -e
 python -c "import torch; assert '$PYTORCH_VERSION' in torch.__version__"
 
 # Run unit tests with initial coverage report
-uv run pytest -v --durations=10 -n8 \
+pytest -v --durations=10 -n8 \
   --ignore=tests/e2e/ \
   --ignore=tests/patched/ \
   --ignore=tests/cli \
@@ -12,36 +12,36 @@ uv run pytest -v --durations=10 -n8 \
   --cov=axolotl
 
 # Run lora kernels tests with coverage append
-uv run pytest -v --durations=10 \
+pytest -v --durations=10 \
   /workspace/axolotl/tests/e2e/patched/lora_kernels \
   --cov=axolotl \
   --cov-append
 
 # Run patched tests excluding lora kernels with coverage append
-uv run pytest --full-trace -vvv --durations=10 \
+pytest --full-trace -vvv --durations=10 \
   --ignore=tests/e2e/patched/lora_kernels \
   /workspace/axolotl/tests/e2e/patched \
   --cov=axolotl \
   --cov-append
 
 # Run solo tests with coverage append
-uv run pytest -v --durations=10 -n1 \
+pytest -v --durations=10 -n1 \
   /workspace/axolotl/tests/e2e/solo/ \
   --cov=axolotl \
   --cov-append
 
 # Run integration tests with coverage append
-uv run pytest -v --durations=10 \
+pytest -v --durations=10 \
   /workspace/axolotl/tests/e2e/integrations/ \
   --cov=axolotl \
   --cov-append
 
-uv run pytest -v --durations=10 /workspace/axolotl/tests/cli \
+pytest -v --durations=10 /workspace/axolotl/tests/cli \
   --cov=axolotl \
   --cov-append
 
 # Run remaining e2e tests with coverage append and final report
-uv run pytest -v --durations=10 \
+pytest -v --durations=10 \
   --ignore=tests/e2e/solo/ \
   --ignore=tests/e2e/patched/ \
   --ignore=tests/e2e/multigpu/ \
@@ -52,4 +52,4 @@ uv run pytest -v --durations=10 \
   --cov-append \
   --cov-report=xml:e2e-coverage.xml
 
-uv run codecov upload-process -t $CODECOV_TOKEN -f e2e-coverage.xml -F e2e,pytorch-${PYTORCH_VERSION} || true
+codecov upload-process -t $CODECOV_TOKEN -f e2e-coverage.xml -F e2e,pytorch-${PYTORCH_VERSION} || true

cicd/multigpu.py

@@ -23,7 +23,7 @@ df_args = {
     "AXOLOTL_EXTRAS": os.environ.get("AXOLOTL_EXTRAS", ""),
     "AXOLOTL_ARGS": os.environ.get("AXOLOTL_ARGS", ""),
     "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.6.0"),
-    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-uv-py3.11-cu126-2.6.0"),
+    "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", ""),

cicd/single_gpu.py

@@ -23,7 +23,7 @@ df_args = {
     "AXOLOTL_EXTRAS": os.environ.get("AXOLOTL_EXTRAS", ""),
     "AXOLOTL_ARGS": os.environ.get("AXOLOTL_ARGS", ""),
     "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.6.0"),
-    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-uv-py3.11-cu126-2.6.0"),
+    "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", ""),

docker/Dockerfile

@@ -1,19 +1,13 @@
-ARG BASE_TAG=main-base-uv
-FROM axolotlai/axolotl-base-uv:$BASE_TAG
+ARG BASE_TAG=main-base
+FROM axolotlai/axolotl-base:$BASE_TAG
 
 ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX"
 ARG AXOLOTL_EXTRAS=""
 ARG AXOLOTL_ARGS=""
 ARG CUDA="118"
 ARG PYTORCH_VERSION="2.1.2"
-ARG GIT_REF="refs/heads/main"
-ARG GIT_SHA="HEAD"
-ARG VENV_PYTHON="/workspace/axolotl-venv/bin/python"
 
 ENV PYTORCH_VERSION=$PYTORCH_VERSION
-ENV GIT_REF=$GIT_REF
-ENV GIT_SHA=$GIT_SHA
-ENV VENV_PYTHON=$VENV_PYTHON
 
 RUN apt-get update && \
     apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev rsync s3fs && \
@@ -26,19 +20,16 @@ RUN git clone --depth=1 https://github.com/axolotl-ai-cloud/axolotl.git
 
 WORKDIR /workspace/axolotl
 
-# Ensure we are on the expected commit and break Docker cache between revisions
-RUN git fetch origin "$GIT_REF" && git checkout "$GIT_SHA"
-
 # If AXOLOTL_EXTRAS is set, append it in brackets
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        uv pip install --python "$VENV_PYTHON" --no-build-isolation -e .[ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \
-        uv pip install --python "$VENV_PYTHON" --no-build-isolation -e .[ring-flash-attn,optimizers,ray] $AXOLOTL_ARGS; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray] $AXOLOTL_ARGS; \
     fi && \
-    uv pip install --python "$VENV_PYTHON" --no-build-isolation flash-attn $AXOLOTL_ARGS && \
-    "$VENV_PYTHON" scripts/unsloth_install.py | sh && \
-    "$VENV_PYTHON" scripts/cutcrossentropy_install.py | sh && \
-    uv pip install --python "$VENV_PYTHON" pytest
+    python scripts/unsloth_install.py | sh && \
+    python scripts/cutcrossentropy_install.py | sh && \
+    pip install pytest && \
+    pip cache purge
 
 # fix so that git fetch/pull from remote works with shallow clone
 RUN git config remote.origin.fetch "+refs/heads/*:refs/remotes/origin/*" && \

docker/Dockerfile-base

@@ -48,5 +48,5 @@ RUN git lfs install --skip-repo && \
     pip3 cache purge
 
 RUN if [ "$PYTORCH_VERSION" = "2.6.0" ] && [ "$CUDA" = "124" ] ; then \
-        FLASH_ATTENTION_FORCE_BUILD="TRUE" uv pip install --no-build-isolation flash-attn==2.8.0.post2; \
+        FLASH_ATTENTION_FORCE_BUILD="TRUE" pip3 install --no-build-isolation flash-attn==2.8.0.post2; \
     fi

docker/Dockerfile-cloud

@@ -12,8 +12,8 @@ EXPOSE 22
 COPY scripts/cloud-entrypoint.sh /root/cloud-entrypoint.sh
 COPY scripts/motd /etc/motd
 
-RUN uv pip install --python "$VENV_PYTHON" jupyterlab notebook ipywidgets && \
-    "$VENV_PYTHON" -m jupyter lab clean
+RUN pip install jupyterlab notebook ipywidgets && \
+    jupyter lab clean
 RUN apt update && \
     apt install --yes --no-install-recommends openssh-server tmux iproute2 nvtop && \
     rm -rf /var/cache/apt/archives && \

docker/Dockerfile-cloud-no-tmux

@@ -12,8 +12,8 @@ EXPOSE 22
 COPY scripts/cloud-entrypoint.sh /root/cloud-entrypoint.sh
 COPY scripts/motd /etc/motd
 
-RUN uv pip install --python "$VENV_PYTHON" jupyterlab notebook ipywidgets && \
-    "$VENV_PYTHON" -m jupyter lab clean
+RUN pip install jupyterlab notebook ipywidgets && \
+    jupyter lab clean
 RUN apt update && \
     apt install --yes --no-install-recommends openssh-server tmux iproute2 nvtop ibverbs-providers ibverbs-utils infiniband-diags librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm && \
     rm -rf /var/cache/apt/archives && \

docker/Dockerfile-tests

@@ -24,14 +24,13 @@ RUN git fetch origin +$GITHUB_REF && \
 
 # If AXOLOTL_EXTRAS is set, append it in brackets
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        uv pip install --no-build-isolation -e .[deepspeed,mamba-ssm,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,mamba-ssm,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \
-        uv pip install --no-build-isolation -e .[deepspeed,mamba-ssm] $AXOLOTL_ARGS; \
-    fi && \
-    uv pip install --no-build-isolation flash-attn $AXOLOTL_ARGS
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,mamba-ssm] $AXOLOTL_ARGS; \
+    fi
 
 # So we can test the Docker image
-RUN uv pip install pytest
+RUN pip install pytest
 
 # fix so that git fetch/pull from remote works
 RUN git config remote.origin.fetch "+refs/heads/*:refs/remotes/origin/*" && \

docker/Dockerfile-uv-base

@@ -13,7 +13,6 @@ ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 9.0+PTX"
 ENV PYTHON_VERSION=$PYTHON_VERSION
 ENV TORCH_CUDA_ARCH_LIST=$TORCH_CUDA_ARCH_LIST
 ENV UV_TORCH_BACKEND="cu${CUDA}"
-ENV VENV_PYTHON=/workspace/axolotl-venv/bin/python
 
 RUN apt-get update \
     && apt-get install -y wget git build-essential ninja-build git-lfs libaio-dev pkg-config curl && rm -rf /var/lib/apt/lists/* \
@@ -30,8 +29,8 @@ RUN uv venv --no-project --relocatable axolotl-venv
 
 ENV PATH="/workspace/axolotl-venv/bin:${PATH}"
 
-RUN uv pip install --python "$VENV_PYTHON" packaging setuptools wheel psutil protobuf grpclib \
-    && uv pip install --python "$VENV_PYTHON" torch==${PYTORCH_VERSION} \
-    && uv pip install --python "$VENV_PYTHON" --no-build-isolation "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main" \
-    && uv pip install --python "$VENV_PYTHON" "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main" \
-    && uv pip install --python "$VENV_PYTHON" awscli pydantic
+RUN uv pip install packaging setuptools wheel psutil \
+    && uv pip install torch==${PYTORCH_VERSION} \
+    && 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

docs/dataset-formats/conversation.qmd

@@ -212,14 +212,6 @@ Instead of passing `tools` via the system prompt, an alternative method would be
 Tools need to follow [JSON schema](https://json-schema.org/learn/getting-started-step-by-step).
 :::
 
-::: {.callout-warning}
-If you have tool arguments with same name but different dtypes (like `"time": string` and `"time": number`), please save `arguments: ` as JSON string to prevent `datasets` from having casting issues.
-
-```
-"arguments": "{\"...\": \"...\"}"
-```
-:::
-
 Example config for Llama4:
 ```yaml
 chat_template: llama4

docs/dataset-formats/index.qmd

@@ -61,7 +61,7 @@ While we recommend `.jsonl`, you can also use the other formats (`csv`, `parquet
 
 ### Pre-training without streaming
 
-In the case that the dataset is small and can be loaded entirely into memory, another approach to running pre-training is to use the `completion` format. This would mean that the entire dataset is pre-tokenized instead of on-demand in streaming.
+On the rare case that the dataset is small and can be loaded entirely into memory, another approach to running pre-training is to use the `completion` format. This would mean that the entire dataset is pre-tokenized instead of on-demand in streaming.
 
 One benefit of this is that the tokenization can be performed separately on a CPU-only machine, and then transferred to a GPU machine for training to save costs.
 

docs/debugging.qmd

@@ -72,8 +72,8 @@ datasets:
 Make sure you have an [editable install](https://setuptools.pypa.io/en/latest/userguide/development_mode.html) of Axolotl, which ensures that changes you make to the code are reflected at runtime.  Run the following commands from the root of this project:
 
 ```bash
-uv sync --extra deepspeed
-uv pip install flash-attn --no-build-isolation
+pip3 install packaging
+pip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'
 ```
 
 #### Remote Hosts
@@ -213,8 +213,8 @@ docker run --privileged --gpus '"all"' --shm-size 10g --rm -it --name axolotl --
 You will now be in the container.  Next, perform an editable install of Axolotl:
 
 ```bash
-uv sync --extra deepspeed
-uv pip install flash-attn --no-build-isolation
+pip3 install packaging
+pip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'
 ```
 
 ### Attach To Container

docs/faq.qmd

@@ -140,7 +140,3 @@ description: Frequently asked questions
 **Q: `ValueError("Backward pass should have cleared tracker of all tensors")`
 
 > A: This may happen due to edge cases in using the modern OffloadActivations context manager for CUDA streams. If you encounter this error, you may have success using the naive implementation with `offload_activations: legacy` in your YAML.
-
-**Q: `Error parsing tool_calls arguments as JSON.`
-
-> A: There is an error parsing string arguments to a dict. Please check your dataset and the error message for more details.

docs/fsdp_qlora.qmd

@@ -1,5 +1,5 @@
 ---
-title: "FSDP + QLoRA"
+title: "FDSP + QLoRA"
 description: Use FSDP with QLoRA to fine-tune large LLMs on consumer GPUs.
 format:
   html:
@@ -23,12 +23,6 @@ To enable `QLoRA` with `FSDP`, you need to perform the following steps:
 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`.
 
-## Enabling Swap for FSDP2
-
-If available memory is insufficient even after FSDP's CPU offloading, you can enable swap memory usage by setting `cpu_offload_pin_memory: false` alongside `offload_params: true` in FSDP config.
-
-This disables memory pinning, allowing FSDP to use disk swap space as fallback. Disabling memory pinning itself incurs performance overhead, and actually having to use swap adds more, but it may enable training larger models that would otherwise cause OOM errors on resource constrained systems.
-
 ## Example Config
 
 [examples/llama-2/qlora-fsdp.yml](../examples/llama-2/qlora-fsdp.yml) contains an example of how to enable QLoRA + FSDP in axolotl.

docs/installation.qmd

@@ -29,40 +29,19 @@ Follow the instructions at: [https://pytorch.org/get-started/locally/](https://p
 For Blackwell GPUs, please use Pytorch 2.7.0 and CUDA 12.8.
 :::
 
-### uv Installation (Recommended) {#sec-uv-quick}
+### PyPI Installation (Recommended) {#sec-pypi}
 
 ```{.bash}
-# Install uv if not already installed
-curl -LsSf https://astral.sh/uv/install.sh | sh
-
-# Add Axolotl to a project (recommended)
-uv init my-project && cd my-project
-uv add axolotl
-uv pip install flash-attn --no-build-isolation
-source .venv/bin/activate
-```
-
-For a quick one-off install without creating a project:
-
-```{.bash}
-uv pip install axolotl
-uv pip install flash-attn --no-build-isolation
-```
-
-### pip Installation {#sec-pypi}
-
-```{.bash}
-pip install --no-build-isolation axolotl[deepspeed]
-pip install --no-build-isolation flash-attn
+pip3 install -U packaging setuptools wheel ninja
+pip3 install --no-build-isolation axolotl[flash-attn,deepspeed]
 ```
 
 We use `--no-build-isolation` in order to detect the installed PyTorch version (if
 installed) in order not to clobber it, and so that we set the correct version of
 dependencies that are specific to the PyTorch version or other installed
-co-dependencies. Flash Attention is resolved separately so it can be built against
-the environment configured by the previous step.
+co-dependencies.
 
-### Advanced uv Installation {#sec-uv}
+### uv Installation {#sec-uv}
 
 uv is a fast, reliable Python package installer and resolver built in Rust. It offers significant performance improvements over pip and provides better dependency resolution, making it an excellent choice for complex environments.
 
@@ -83,38 +62,28 @@ source .venv/bin/activate
 Install PyTorch
 - PyTorch 2.6.0 recommended
 ```{.bash}
+uv pip install packaging setuptools wheel
 uv pip install torch==2.6.0
 uv pip install awscli pydantic
 ```
 
 Install axolotl from PyPi
 ```{.bash}
-uv pip install --no-build-isolation axolotl[deepspeed]
-# optionally install with vLLM if you're using torch==2.6.0 and want to train w/ GRPO
-# uv pip install --no-build-isolation axolotl[deepspeed,vllm]
+uv pip install --no-build-isolation axolotl[deepspeed,flash-attn]
 
-uv pip install flash-attn --no-build-isolation
+# optionally install with vLLM if you're using torch==2.6.0 and want to train w/ GRPO
+uv pip install --no-build-isolation axolotl[deepspeed,flash-attn,vllm]
 ```
 
 ### Edge/Development Build {#sec-edge-build}
 
 For the latest features between releases:
 
-#### Using uv (recommended)
 ```{.bash}
 git clone https://github.com/axolotl-ai-cloud/axolotl.git
 cd axolotl
-curl -LsSf https://astral.sh/uv/install.sh | sh  # If not already installed
-uv sync
-uv pip install flash-attn --no-build-isolation
-```
-
-#### Using pip
-```{.bash}
-git clone https://github.com/axolotl-ai-cloud/axolotl.git
-cd axolotl
-pip install --no-build-isolation -e '.[deepspeed]'
-pip install --no-build-isolation flash-attn
+pip3 install -U packaging setuptools wheel ninja
+pip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'
 ```
 
 ### Docker {#sec-docker}
@@ -172,7 +141,7 @@ For providers supporting Docker:
 ### macOS {#sec-macos}
 
 ```{.bash}
-uv pip install --no-build-isolation -e '.'
+pip3 install --no-build-isolation -e '.'
 ```
 
 See @sec-troubleshooting for Mac-specific issues.
@@ -190,15 +159,10 @@ We recommend using WSL2 (Windows Subsystem for Linux) or Docker.
 1. Install Python ≥3.11
 2. Install PyTorch: https://pytorch.org/get-started/locally/
 3. Install Axolotl:
-```{.bash}
-# Option A: add Axolotl to the environment
-uv add axolotl
-uv pip install flash-attn --no-build-isolation
-
-# Option B: quick install
-uv pip install axolotl
-uv pip install flash-attn --no-build-isolation
-```
+   ```{.bash}
+   pip3 install -U packaging setuptools wheel ninja
+   pip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'
+   ```
 4. (Optional) Login to Hugging Face:
    ```{.bash}
    huggingface-cli login

docs/lora_optims.qmd

@@ -5,11 +5,10 @@ description: "Custom autograd functions and Triton kernels in Axolotl for optimi
 
 Inspired by [Unsloth](https://github.com/unslothai/unsloth), we've implemented two
 optimizations for LoRA and QLoRA fine-tuning, supporting both single GPU and multi-GPU
-(including the DDP, DeepSpeed, and FSDP2 settings) training. These include (1) SwiGLU
-and GEGLU activation function Triton kernels, and (2) LoRA MLP and attention custom
-autograd functions. Our goal was to leverage operator fusion and tensor re-use in order
-to improve speed and reduce memory usage during the forward and backward passes of
-these calculations.
+(in the DDP and DeepSpeed settings) training. These include (1) SwiGLU and GEGLU activation function
+Triton kernels, and (2) LoRA MLP and attention custom autograd functions. Our goal was
+to leverage operator fusion and tensor re-use in order to improve speed and reduce
+memory usage during the forward and backward passes of these calculations.
 
 We currently support several common model architectures, including (but not limited to):
 
@@ -132,5 +131,6 @@ computation path.
 ## Future Work
 
 - Support for additional model architectures
+- Support for the FSDP setting
 - Support for dropout and bias
 - Additional operator fusions

docs/moe_backends.md

@@ -0,0 +1,18 @@
+MoE Backends in Axolotl
+
+Axolotl supports selecting a Mixture-of-Experts (MoE) compute backend via the training config (YAML):
+
+- Set `moe_backend: auto|torch_grouped|naive`
+
+Behavior
+- auto (default): prefers PyTorch 2.8+ grouped GEMM; otherwise naive.
+- torch_grouped: targets PyTorch 2.8+ grouped GEMM (H100/SM90+ recommended).
+- naive: keeps the reference per-expert loop.
+
+Notes
+- Current implementation wires the backend selector and routes Mixtral MoE through it. Torch grouped uses cuBLASLt grouped GEMM when available; otherwise, the code falls back to the naive per-expert loop.
+- No changes to training scripts are required; selection happens inside the model forward.
+
+Example
+moe_backend: torch_grouped
+accelerate launch -m axolotl.cli.train path/to/config.yaml

docs/multimodal.qmd

@@ -13,7 +13,6 @@ format:
 - [Pixtral](#sec-pixtral)
 - [Llava-1.5](#sec-llava-15)
 - [Mistral-Small-3.1](#sec-mistral-small-31)
-- [Magistral-Small-2509](#sec-magistral-small-2509)
 - [Voxtral](#sec-voxtral)
 - [Gemma-3](#sec-gemma-3)
 - [Gemma-3n](#sec-gemma-3n)
@@ -42,6 +41,7 @@ datasets:
   - path: HuggingFaceH4/llava-instruct-mix-vsft
     type: chat_template
     split: train[:1%]
+    field_messages: messages
 
 # (optional) if doing lora, only finetune the Language model,
 # leave the vision model and vision tower frozen
@@ -94,28 +94,16 @@ chat_template: llava
 
 ### Mistral-Small-3.1 {#sec-mistral-small-31}
 
-::: {.callout-tip}
-Please make sure to install vision lib via `uv pip install 'mistral-common[opencv]==1.8.5'`
-:::
-
 ```yaml
 base_model: mistralai/Mistral-Small-3.1-24B-Instruct-2503
-```
 
-### Magistral-Small-2509 {#sec-magistral-small-2509}
-
-::: {.callout-tip}
-Please make sure to install vision lib via `uv pip install 'mistral-common[opencv]==1.8.5'`
-:::
-
-```yaml
-base_model: mistralai/Magistral-Small-2509
+chat_template: mistral_v7_tekken
 ```
 
 ### Voxtral {#sec-voxtral}
 
 ::: {.callout-tip}
-Please make sure to install audio lib via `uv pip install librosa==0.11.0 'mistral_common[audio]==1.8.3'`
+Please make sure to install audio lib via `pip3 install librosa==0.11.0 'mistral_common[audio]==1.8.3'`
 :::
 
 ```yaml
@@ -143,7 +131,7 @@ The model's initial loss and grad norm will be very high. We suspect this to be
 :::
 
 ::: {.callout-tip}
-Please make sure to install `timm` via `uv pip install timm==1.0.17`
+Please make sure to install `timm` via `pip3 install timm==1.0.17`
 :::
 
 ```yaml
@@ -171,7 +159,7 @@ chat_template: qwen2_vl  # same as qwen2-vl
 ### SmolVLM2 {#sec-smolvlm2}
 
 ::: {.callout-tip}
-Please make sure to install `num2words` via `uv pip install num2words==0.5.14`
+Please make sure to install `num2words` via `pip3 install num2words==0.5.14`
 :::
 
 ```yaml
@@ -181,7 +169,7 @@ base_model: HuggingFaceTB/SmolVLM2-500M-Video-Instruct
 ### LFM2-VL {#sec-lfm2-vl}
 
 ::: {.callout-warning}
-Please uninstall `causal-conv1d` via `uv pip uninstall -y causal-conv1d`
+Please uninstall `causal-conv1d` via `pip3 uninstall -y causal-conv1d`
 :::
 
 ```yaml
@@ -222,7 +210,7 @@ For audio loading, you can use the following keys within `content` alongside `"t
 
 ::: {.callout-tip}
 
-You may need to install `librosa` via `uv pip install librosa==0.11.0`.
+You may need to install `librosa` via `pip3 install librosa==0.11.0`.
 
 :::
 

docs/optimizations.qmd

@@ -1,133 +0,0 @@
----
-title: Optimizations Guide
-description: A guide to the performance and memory optimizations available in Axolotl.
----
-
-Axolotl includes numerous optimizations to speed up training, reduce memory usage, and handle large models.
-
-This guide provides a high-level overview and directs you to the detailed documentation for each feature.
-
-## Speed Optimizations
-
-These optimizations focus on increasing training throughput and reducing total training time.
-
-### Sample Packing
-
-Improves GPU utilization by combining multiple short sequences into a single packed sequence for training. This requires enabling one of the [attention](#attention-implementations) implementations below.
-
-- **Config:** `sample_packing: true`
-- **Learn more:** [Sample Packing](multipack.qmd)
-
-### Attention Implementations
-
-Using an optimized attention implementation is critical for training speed.
-
-- **[Flash Attention 2](https://github.com/Dao-AILab/flash-attention)**: `flash_attention: true`. **(Recommended)** The industry standard for fast attention on modern GPUs. Requires Ampere or higher. For AMD, check [AMD Support](https://github.com/Dao-AILab/flash-attention?tab=readme-ov-file#amd-rocm-support).
-- **[Flex Attention](https://pytorch.org/blog/flexattention/)**: `flex_attention: true`.
-- **[SDP Attention](https://docs.pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html)**: `sdp_attention: true`. PyTorch's native implementation.
-- **[Xformers](https://github.com/facebookresearch/xformers)**: `xformers_attention: true`. Works with FP16.
-
-*Note: You should only enable one attention backend.*
-
-### LoRA Optimizations
-
-Leverages optimized kernels to accelerate LoRA training and reduce memory usage.
-
-- **Learn more:** [LoRA Optimizations Documentation](lora_optims.qmd)
-
-## Memory Optimizations
-
-These techniques help you fit larger models or use bigger batch sizes on your existing hardware.
-
-### Parameter Efficient Finetuning (LoRA & QLoRA)
-
-Drastically reduces memory by training a small set of "adapter" parameters instead of the full model. This is the most common and effective memory-saving technique.
-
-- Examples: Find configs with `lora` or `qlora` in the [examples directory](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/llama-3).
-- Config Reference: See `adapter`, `load_in_4bit`, and `load_in_8bit` in the [Configuration Reference](config-reference.qmd).
-
-### Gradient Checkpointing & Activation Offloading
-
-These techniques save VRAM by changing how activations are handled.
-
-- Gradient Checkpointing: re-computes activations during the backward pass, trading compute time for VRAM.
-- Activation Offloading: moves activations to CPU RAM or disk, trading I/O overhead for VRAM.
-- Learn more: [Gradient Checkpointing and Offloading Docs](gradient_checkpointing.qmd)
-
-### Cut Cross Entropy (CCE)
-
-Reduces VRAM usage by using an optimized cross-entropy loss calculation.
-
-- **Learn more:** [Custom Integrations - CCE](custom_integrations.qmd#cut-cross-entropy)
-
-### Liger Kernels
-
-Provides efficient Triton kernels to improve training speed and reduce memory usage.
-
-- **Learn more:** [Custom Integrations - Liger Kernels](custom_integrations.qmd#liger-kernels)
-
-## Long Context Models
-
-Techniques to train models on sequences longer than their original context window.
-
-### RoPE Scaling
-
-Extends a model's context window by interpolating its Rotary Position Embeddings.
-
-- **Config:** Pass the `rope_scaling` config under the `overrides_of_model_config: `. To learn how to set RoPE, check the respective model config.
-
-### Sequence Parallelism
-
-Splits long sequences across multiple GPUs, enabling training with sequence lengths that would not fit on a single device.
-
-- **Learn more:** [Sequence Parallelism Documentation](sequence_parallelism.qmd)
-
-### Artic Long Sequence Training (ALST)
-
-ALST is a recipe that combines several techniques to train long-context models efficiently. It typically involves:
-
-- TiledMLP to reduce memory usage in MLP layers.
-- Tiled Loss functions (like [CCE](#cut-cross-entropy-(cce) or [Liger](#liger-kernels)).
-- Activation Offloading to CPU.
-
-- Example: [ALST Example Configuration](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/alst)
-
-## Large Models (Distributed Training)
-
-To train models that don't fit on a single GPU, you'll need to use a distributed training strategy like FSDP or DeepSpeed. These frameworks shard the model weights, gradients, and optimizer states across multiple GPUs and nodes.
-
-- **Learn more:** [Multi-GPU Guide](multi-gpu.qmd)
-- **Learn more:** [Multi-Node Guide](multi-node.qmd)
-
-### N-D Parallelism (Beta)
-
-For advanced scaling, Axolotl allows you to compose different parallelism techniques (e.g., Data, Tensor, Sequence Parallelism). This is a powerful approach to train an extremely large model by overcoming multiple bottlenecks at once.
-
-- **Learn more:** [N-D Parallelism Guide](nd_parallelism.qmd)
-
-
-## Quantization
-
-Techniques to reduce the precision of model weights for memory savings.
-
-### 4-bit Training (QLoRA)
-
-The recommended approach for quantization-based training. It loads the base model in 4-bit using `bitsandbytes` and then trains QLoRA adapters. See [Adapter Finetuning](#adapter-finetuning-lora-qlora) for details.
-
-### FP8 Training
-
-Enables training with 8-bit floating point precision on supported hardware (e.g., NVIDIA Hopper series GPUs) for significant speed and memory gains.
-
-- **Example:** [Llama 3 FP8 FSDP Example](https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-3/3b-fp8-fsdp2.yaml)
-
-### Quantization Aware Training (QAT)
-
-Simulates quantization effects during training, helping the model adapt and potentially improving the final accuracy of the quantized model.
-
-- **Learn more:** [QAT Documentation](qat.qmd)
-
-### GPTQ
-
-Allows you to finetune LoRA adapters on top of a model that has already been quantized using the GPTQ method.
-
-- **Example:** [GPTQ LoRA Example](https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-2/gptq-lora.yml)

docs/qat.qmd

@@ -30,7 +30,6 @@ qat:
 ```
 
 We support the following quantization schemas:
-
 - `Int4WeightOnly` (requires the `fbgemm-gpu` extra when installing Axolotl)
 - `Int8DynamicActivationInt4Weight`
 - `Float8DynamicActivationFloat8Weight`

docs/sequence_parallelism.qmd

@@ -49,9 +49,9 @@ When sequence parallelism is enabled:
 To use sequence parallelism, you need:
 
 - Multiple GPUs (at least 2)
-- The `ring-flash-attn` package. Install with either `uv sync --extra ring-flash-attn`
-  (from a cloned repository) or `uv pip install ring-flash-attn>=0.1.4`.
-- Flash Attention installed separately with `uv pip install flash-attn --no-build-isolation`.
+- The `ring-flash-attn` package. Install with:
+  - `pip install axolotl[ring-flash-attn]` (preferred)
+  - `pip install ring-flash-attn>=0.1.4`
 
 ## Limitations
 

examples/LiquidAI/README.md

@@ -12,14 +12,9 @@ This guide shows how to fine-tune both the LFM2 and LFM2-VL models with Axolotl.
 
     Here is an example of how to install from pip:
     ```bash
-    # Ensure you have a compatible version of PyTorch installed
-    # Option A: manage dependencies in your project
-    uv add 'axolotl>=0.12.0'
-    uv pip install flash-attn --no-build-isolation
-
-    # Option B: quick install
-    uv pip install 'axolotl>=0.12.0'
-    uv pip install flash-attn --no-build-isolation
+    # Ensure you have a compatible version of Pytorch installed
+    pip3 install packaging setuptools wheel ninja
+    pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
     ```
 
 2.  Run one of the finetuning examples below.
@@ -40,7 +35,7 @@ This guide shows how to fine-tune both the LFM2 and LFM2-VL models with Axolotl.
 
 - **Installation Error**: If you encounter `ImportError: ... undefined symbol ...` or `ModuleNotFoundError: No module named 'causal_conv1d_cuda'`, the `causal-conv1d` package may have been installed incorrectly. Try uninstalling it:
   ```bash
-  uv pip uninstall -y causal-conv1d
+  pip uninstall -y causal-conv1d
   ```
 
 - **Dataset Loading**: Read more on how to load your own dataset in our [documentation](https://docs.axolotl.ai/docs/dataset_loading.html).

examples/alst/README.md

@@ -7,24 +7,3 @@ techniques. It is a combination of:
 - Activation Offloading: Offload activations to CPU RAM to reduce memory usage
 
 For more information, you can check out the ALST paper [here](https://www.arxiv.org/abs/2506.13996).
-
-## Usage
-
-```yaml
-tiled_mlp: true
-
-# See Sequence Parallelism docs
-# https://docs.axolotl.ai/docs/sequence_parallelism.html
-context_parallel_size: int
-
-plugins:
-# See Cut Cross Entropy docs
-# https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
-  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
-
-# or Liger Kernel docs
-# https://docs.axolotl.ai/docs/custom_integrations.html#liger-kernels
-  - axolotl.integrations.liger.LigerPlugin
-# ...
-
-```

examples/apertus/README.md

@@ -1,110 +0,0 @@
-# Finetune Swiss-AI's Apertus with Axolotl
-
-[Apertus](https://huggingface.co/collections/swiss-ai/apertus-llm-68b699e65415c231ace3b059) is a family of opensource models trained by Swiss-ai.
-
-This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
-
-## Getting started
-
-1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html). You need to install from main as Apertus 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 min)
-git clone https://github.com/axolotl-ai-cloud/axolotl.git
-cd axolotl
-
-uv sync
-uv pip install flash-attn --no-build-isolation
-
-# Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
-python scripts/cutcrossentropy_install.py | sh
-```
-
-2. (Optional, highly recommended) Install XIELU CUDA
-
-```bash
-## Recommended for reduced VRAM and faster speeds
-
-# Point to CUDA toolkit directory
-# For those using our Docker image, use the below path.
-export CUDA_HOME=/usr/local/cuda
-
-uv pip install git+https://github.com/nickjbrowning/XIELU@59d6031 --no-build-isolation --no-deps
-```
-
-For any installation errors, see [XIELU Installation Issues](#xielu-installation-issues)
-
-3. Run the finetuning example:
-
-```bash
-axolotl train examples/apertus/apertus-8b-qlora.yaml
-```
-
-This config uses about 8.7 GiB VRAM.
-
-Let us know how it goes. Happy finetuning! 🚀
-
-### Tips
-
-- For inference, the official Apertus team recommends `top_p=0.9` and `temperature=0.8`.
-- You can instead use full paremter fine-tuning 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).
-
-### XIELU Installation Issues
-
-#### `ModuleNotFoundError: No module named 'torch'`
-
-Please check these one by one:
-- Running in correct environment
-- Env has PyTorch installed
-- CUDA toolkit is at `CUDA_HOME`
-
-If those didn't help, please try the below solutions:
-
-1. Pass env for CMAKE and try install again:
-
-    ```bash
-    Python_EXECUTABLE=$(which python) uv pip install git+https://github.com/nickjbrowning/XIELU@59d6031 --no-build-isolation --no-deps
-    ```
-
-2. Git clone the repo and manually hardcode python path:
-
-    ```bash
-    git clone https://github.com/nickjbrowning/XIELU
-    cd xielu
-    git checkout 59d6031
-
-    cd xielu
-    nano CMakeLists.txt  # or vi depending on your preference
-    ```
-
-    ```diff
-    execute_process(
-    -    COMMAND ${Python_EXECUTABLE} -c "import torch.utils; print(torch.utils.cmake_prefix_path)"
-    +    COMMAND /root/miniconda3/envs/py3.11/bin/python -c "import torch.utils; print(torch.utils.cmake_prefix_path)"
-        RESULT_VARIABLE TORCH_CMAKE_PATH_RESULT
-        OUTPUT_VARIABLE TORCH_CMAKE_PATH_OUTPUT
-        ERROR_VARIABLE TORCH_CMAKE_PATH_ERROR
-    )
-    ```
-
-    ```bash
-    uv pip install . --no-build-isolation --no-deps
-    ```
-
-## 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)
-
-## Related Resources
-
-- [Apertus Tech Report](https://github.com/swiss-ai/apertus-tech-report/blob/main/Apertus_Tech_Report.pdf)
-- [Axolotl Docs](https://docs.axolotl.ai)
-- [Axolotl Website](https://axolotl.ai)
-- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
-- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/apertus/apertus-8b-qlora.yaml

@@ -1,64 +0,0 @@
-base_model: swiss-ai/Apertus-8B-Instruct-2509
-
-# Automatically upload checkpoint and final model to HF
-# hub_model_id: username/custom_model_name
-
-plugins:
-  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
-
-load_in_8bit: false
-load_in_4bit: true
-
-datasets:
-  - path: fozziethebeat/alpaca_messages_2k_test
-    type: chat_template
-
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.1
-output_dir: ./outputs/lora-out
-
-adapter: qlora
-lora_model_dir:
-
-sequence_len: 2048
-sample_packing: true
-
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_linear: true
-lora_target_modules:
-  - gate_proj
-  - down_proj
-  - up_proj
-  - q_proj
-  - v_proj
-  - k_proj
-  - o_proj
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 4
-micro_batch_size: 2
-num_epochs: 1
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-bf16: auto
-tf32: false
-
-gradient_checkpointing: true
-resume_from_checkpoint:
-logging_steps: 1
-flash_attention: true
-
-warmup_ratio: 0.1
-evals_per_epoch: 1
-saves_per_epoch: 1
-
-# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/arcee/README.md

@@ -17,11 +17,8 @@ Thanks to the team at Arcee.ai for using Axolotl in supervised fine-tuning the A
 git clone https://github.com/axolotl-ai-cloud/axolotl.git
 cd axolotl
 
-uv sync
-uv pip install flash-attn --no-build-isolation
-
-# Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
-python scripts/cutcrossentropy_install.py | sh
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install --no-build-isolation -e '.[flash-attn]'
 ```
 
 2. Run the finetuning example:

examples/archived/deepcoder/deepcoder-14B-preview-lora.yml

@@ -9,6 +9,10 @@ strict: false
 datasets:
   - path: fozziethebeat/alpaca_messages_2k_test
     type: chat_template
+    field_messages: messages
+    message_property_mappings:
+      role: role
+      content: content
 
 dataset_prepared_path:
 val_set_size: 0.05

examples/colab-notebooks/colab-axolotl-example.ipynb

@@ -12,10 +12,10 @@
     "\n",
     "Axolotl is the most performant LLM post-training framework available, delivering faster training with efficient, consistent and stable performance. Train your workload and ship your product 30% faster; saving you both time and money.\n",
     "\n",
-    "- \u2b50 us on [GitHub](https://github.com/axolotl-ai-cloud/axolotl)\n",
-    "- \ud83d\udcdc Read the [Docs](http://docs.axolotl.ai/)\n",
-    "- \ud83d\udcac Chat with us on [Discord](https://discord.gg/mnpEYgRUmD)\n",
-    "- \ud83d\udcf0 Get updates on [X/Twitter](https://x.com/axolotl_ai)\n"
+    "- ⭐ us on [GitHub](https://github.com/axolotl-ai-cloud/axolotl)\n",
+    "- 📜 Read the [Docs](http://docs.axolotl.ai/)\n",
+    "- 💬 Chat with us on [Discord](https://discord.gg/mnpEYgRUmD)\n",
+    "- 📰 Get updates on [X/Twitter](https://x.com/axolotl_ai)\n"
    ]
   },
   {
@@ -39,8 +39,8 @@
    "source": [
     "%%capture\n",
     "# This step can take ~5-10 minutes to install dependencies\n",
-    "!uv pip install --no-build-isolation axolotl>=0.9.1\n!uv pip install flash-attn --no-build-isolation\n",
-    "!uv pip install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@147ea28\""
+    "!pip install --no-build-isolation axolotl[flash-attn]>=0.9.1\n",
+    "!pip install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@c6a32c5\""
    ]
   },
   {
@@ -1371,7 +1371,7 @@
        "version_minor": 0
       },
       "text/plain": [
-       "VBox(children=(HTML(value='<center> <img\\nsrc=https://huggingface.co/front/assets/huggingface_logo-noborder.sv\u2026"
+       "VBox(children=(HTML(value='<center> <img\\nsrc=https://huggingface.co/front/assets/huggingface_logo-noborder.sv…"
       ]
      },
      "metadata": {},
@@ -1729,9 +1729,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_12815f401eba44658caa7b2e490137a8",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_30e02aa2d0d241979369e598287f2639",
-      "value": "Drop\u2007Samples\u2007with\u2007Zero\u2007Trainable\u2007Tokens\u2007(num_proc=2):\u2007100%"
+      "value": "Drop Samples with Zero Trainable Tokens (num_proc=2): 100%"
      }
     },
     "083f9cda8d754c168beee10d2f8955a2": {
@@ -1774,9 +1774,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_b195f160ca20442fadd8b5aed0ee41af",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_ca65e32eb52f48c09a84b33cb18f22cd",
-      "value": "\u200711.4M/11.4M\u2007[00:00&lt;00:00,\u200721.8MB/s]"
+      "value": " 11.4M/11.4M [00:00&lt;00:00, 21.8MB/s]"
      }
     },
     "0a46ad75c198463d843fb35e813642cb": {
@@ -1917,7 +1917,7 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_b1bea589efa14258a9982071b87938bf",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_590eef89881545aa8bbef9a8bbe7fb00",
       "value": "\n<b>Pro Tip:</b> If you don't already have one, you can create a dedicated\n'notebooks' token with 'write' access, that you can then easily reuse for all\nnotebooks. </center>"
      }
@@ -1938,9 +1938,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_bfcdbba993b74972a9e3e575f86908ff",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_6ebb2ec171414e47a14765505f64bb3c",
-      "value": "\u20073.84G/3.84G\u2007[00:09&lt;00:00,\u2007664MB/s]"
+      "value": " 3.84G/3.84G [00:09&lt;00:00, 664MB/s]"
      }
     },
     "0e936d9dbf9c4fdd86bbfe9730dedc47": {
@@ -2296,9 +2296,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_349eee9f56d64f0cba6fc24ff2c50c9b",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_7e5d3774060e4589aa65982da5ea4ef4",
-      "value": "\u20079985/9985\u2007[00:04&lt;00:00,\u20072604.11\u2007examples/s]"
+      "value": " 9985/9985 [00:04&lt;00:00, 2604.11 examples/s]"
      }
     },
     "16d1283741404b7bb319094c992fce01": {
@@ -2317,9 +2317,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_a4e5789584564049b83df7c6c54a3e08",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_ff3a94b146a948b6907f5d80c7157f99",
-      "value": "\u20079985/0\u2007[00:00&lt;00:00,\u200750763.46\u2007examples/s]"
+      "value": " 9985/0 [00:00&lt;00:00, 50763.46 examples/s]"
      }
     },
     "1811cda0644e4190a9469d1774435d82": {
@@ -2390,9 +2390,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_e366ae3fceec4566b9ed303d6c5f90af",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_5dd7d150dbe04f08b165ce7f2c27cd11",
-      "value": "model-00008-of-00008.safetensors:\u2007100%"
+      "value": "model-00008-of-00008.safetensors: 100%"
      }
     },
     "19127c7bb1554ccbac877059f9a82db0": {
@@ -2561,9 +2561,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_0dea5caa27384f5689e3cab51f558727",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_a6f48410b9964fefba0c3009a77dc838",
-      "value": "\u20079.68k/9.68k\u2007[00:00&lt;00:00,\u2007812kB/s]"
+      "value": " 9.68k/9.68k [00:00&lt;00:00, 812kB/s]"
      }
     },
     "1f7d30f71bbd4547a9150d21da071055": {
@@ -2634,9 +2634,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_f4a1795dc7514a718f478245f521f0ba",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_5e746eb25bbe416fb585fa24e79f5177",
-      "value": "model-00002-of-00008.safetensors:\u2007100%"
+      "value": "model-00002-of-00008.safetensors: 100%"
      }
     },
     "20352e5f58d24bb8b1f3940efd14fe4a": {
@@ -2707,9 +2707,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_1c6f1f10667545aaab958016ba7e2c94",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_e6e969610738449887259063967f82b0",
-      "value": "\u20072.78M/2.78M\u2007[00:00&lt;00:00,\u200717.8MB/s]"
+      "value": " 2.78M/2.78M [00:00&lt;00:00, 17.8MB/s]"
      }
     },
     "258b7c635c1045329d4669e48c46ccd5": {
@@ -3056,9 +3056,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_be724f04b03942b2a033a7e8898bb4fd",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_fcbab4d8dced41a18dfccce81e3a45a0",
-      "value": "model-00005-of-00008.safetensors:\u2007100%"
+      "value": "model-00005-of-00008.safetensors: 100%"
      }
     },
     "3036608c71904ce9ae4bb2a9fa8802d9": {
@@ -3077,9 +3077,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_5ca6be24acb548cea130bd58e9954c7c",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_5cfb02ee044b4011a378efa8b54a370f",
-      "value": "\u20073.96G/3.96G\u2007[00:10&lt;00:00,\u2007531MB/s]"
+      "value": " 3.96G/3.96G [00:10&lt;00:00, 531MB/s]"
      }
     },
     "30a81da86f8043eca301e86a8651201a": {
@@ -3629,9 +3629,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_8f5bd719974e41c3a8dd9a5b0d3d71e6",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_b87c84de30e84b3abf4871461fb9cbd3",
-      "value": "Loading\u2007checkpoint\u2007shards:\u2007100%"
+      "value": "Loading checkpoint shards: 100%"
      }
     },
     "41f3b32c2f6b4034ae7a3b9124e28bc7": {
@@ -3791,7 +3791,7 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_39789237703c4a418134243055c9cbf5",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_a3a945817f684328b34651fe052393ec",
       "value": "Connecting..."
      }
@@ -4077,9 +4077,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_4d468f96ec924681ad65eb671674b93e",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_ad7599de524549c48bf2d3124ad4b299",
-      "value": "Dropping\u2007Long\u2007Sequences\u2007(num_proc=2):\u2007100%"
+      "value": "Dropping Long Sequences (num_proc=2): 100%"
      }
     },
     "5ca240f31e6b44e3882c5eb37cd5a309": {
@@ -4471,9 +4471,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_5e18768f7ad6434ba8b8b8a2e853e204",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_bb33aec33a6447078c31bfd728942994",
-      "value": "\u2007728/728\u2007[00:00&lt;00:00,\u200720.3kB/s]"
+      "value": " 728/728 [00:00&lt;00:00, 20.3kB/s]"
      }
     },
     "62e302ebdad64aada0ffe64ae1c873f3": {
@@ -4636,9 +4636,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_81c3db71ac704280ad030072655f1537",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_042e091f75694c47aee761e760e76773",
-      "value": "\u20079985/9985\u2007[00:02&lt;00:00,\u20073977.47\u2007examples/s]"
+      "value": " 9985/9985 [00:02&lt;00:00, 3977.47 examples/s]"
      }
     },
     "67da6c4260574869aa24c3cbc1bc1654": {
@@ -4778,7 +4778,7 @@
       "description_tooltip": null,
       "disabled": false,
       "layout": "IPY_MODEL_2e257c8be2da40b4bb67a9e4ab6811f3",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_56e3768bef5a4b9db4168c5c17f509c2",
       "value": ""
      }
@@ -4823,9 +4823,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_41f3b32c2f6b4034ae7a3b9124e28bc7",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_a10d0a76010f4e508c65a9b69ebc5156",
-      "value": "Tokenizing\u2007Prompts\u2007(num_proc=2):\u2007100%"
+      "value": "Tokenizing Prompts (num_proc=2): 100%"
      }
     },
     "704f2f5a9b1c49d5a75a0025a5dda11b": {
@@ -5071,9 +5071,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_93a44a11aa4846fa8efc6c1413ef1627",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_a55060adc3564407ac81ad7297d34aaa",
-      "value": "train.jsonl:\u2007100%"
+      "value": "train.jsonl: 100%"
      }
     },
     "7be6f04c284e4326bb4ff3d301e7b3c6": {
@@ -5138,9 +5138,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_7fd44cf9ca6e4726bfd7ac21846d6a14",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_366a343b62fa47d8985a3bd464d99f9e",
-      "value": "config.json:\u2007100%"
+      "value": "config.json: 100%"
      }
     },
     "7cd0b85ebd204b7aba908417811ce4e0": {
@@ -5339,9 +5339,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_67da6c4260574869aa24c3cbc1bc1654",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_94b9088614464f60a203de39dbcae853",
-      "value": "\u20078/8\u2007[01:47&lt;00:00,\u200711.64s/it]"
+      "value": " 8/8 [01:47&lt;00:00, 11.64s/it]"
      }
     },
     "823f1c78f15043e38bbd4dca3932a86a": {
@@ -5488,7 +5488,7 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_8640ac440fbc4644b9a3af7ba3ae7183",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_5cea7996f02040b187ece0bb2d6a8d1f",
       "value": "<center> <img\nsrc=https://huggingface.co/front/assets/huggingface_logo-noborder.svg\nalt='Hugging Face'> <br> Copy a token from <a\nhref=\"https://huggingface.co/settings/tokens\" target=\"_blank\">your Hugging Face\ntokens page</a> and paste it below. <br> Immediately click login after copying\nyour token or it might be stored in plain text in this notebook file. </center>"
      }
@@ -5509,9 +5509,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_ef223e8504b64e3592589880326aaf41",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_598da69727bd4fb8b1caf465ac736d7a",
-      "value": "\u20071.67M/1.67M\u2007[00:00&lt;00:00,\u200719.0MB/s]"
+      "value": " 1.67M/1.67M [00:00&lt;00:00, 19.0MB/s]"
      }
     },
     "897b77a56c09479bb11d7f2a30997e55": {
@@ -5717,9 +5717,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_37de928300e34184881039378bd75e7f",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_0e936d9dbf9c4fdd86bbfe9730dedc47",
-      "value": "\u20073.96G/3.96G\u2007[00:13&lt;00:00,\u2007273MB/s]"
+      "value": " 3.96G/3.96G [00:13&lt;00:00, 273MB/s]"
      }
     },
     "936d04b5fe1b4c63bf0b080e423d051b": {
@@ -6050,9 +6050,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_d955dcaa0e944e719f3a06139dd54a03",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_d3de2662c7964f1ba96e58da382af720",
-      "value": "merges.txt:\u2007100%"
+      "value": "merges.txt: 100%"
      }
     },
     "9cd5211b5d8b457aa0002f1d17b80028": {
@@ -6071,9 +6071,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_6932489232ec4ab18a160b1e7fbcdfe1",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_4540927d98f54466b434ba4c0edf045d",
-      "value": "model-00007-of-00008.safetensors:\u2007100%"
+      "value": "model-00007-of-00008.safetensors: 100%"
      }
     },
     "9d4897eefb5f48259ffb2d23e332f752": {
@@ -6303,9 +6303,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_3aaecbf540f54a2db9ab0931e3b1fe57",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_9e333ed3b5014069ac1dd969255dd591",
-      "value": "\u2007239/239\u2007[00:00&lt;00:00,\u200730.9kB/s]"
+      "value": " 239/239 [00:00&lt;00:00, 30.9kB/s]"
      }
     },
     "a20927bf5f2c41f58c1e31ac858ab36c": {
@@ -6324,9 +6324,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_1811cda0644e4190a9469d1774435d82",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_35c811d2ae8e43f3b5cecbdd3cfa857f",
-      "value": "tokenizer.json:\u2007100%"
+      "value": "tokenizer.json: 100%"
      }
     },
     "a3a945817f684328b34651fe052393ec": {
@@ -6360,9 +6360,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_ed5ca967ad5342929e578ac6aa4dc4c0",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_af401d117d5047629d3a6e2361757b62",
-      "value": "model-00001-of-00008.safetensors:\u2007100%"
+      "value": "model-00001-of-00008.safetensors: 100%"
      }
     },
     "a4e5789584564049b83df7c6c54a3e08": {
@@ -6494,9 +6494,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_fa1282ccc7544e4f818e2f03ccffe4a5",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_bbbf575d2a4b4c6ea8389be79b2a6039",
-      "value": "model.safetensors.index.json:\u2007100%"
+      "value": "model.safetensors.index.json: 100%"
      }
     },
     "ab93eabd7cea4b94b4b7a387f101e8a1": {
@@ -6582,9 +6582,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_62e302ebdad64aada0ffe64ae1c873f3",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_bd1b0dfed6d34d16af33a4a58330f5ec",
-      "value": "Saving\u2007the\u2007dataset\u2007(1/1\u2007shards):\u2007100%"
+      "value": "Saving the dataset (1/1 shards): 100%"
      }
     },
     "ad7599de524549c48bf2d3124ad4b299": {
@@ -6967,9 +6967,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_2b3a2659b12244bd8548320320016dbf",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_0cd7efffbb3c4c4b972e63749f61ab97",
-      "value": "Generating\u2007train\u2007split:\u2007"
+      "value": "Generating train split: "
      }
     },
     "b87c84de30e84b3abf4871461fb9cbd3": {
@@ -7085,9 +7085,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_0f480e3a0b0a45d2a2d2dec3cad923f3",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_fcb30372e7404c5d8a1ad4df91e6c7b2",
-      "value": "\u20071.91G/1.91G\u2007[00:05&lt;00:00,\u2007444MB/s]"
+      "value": " 1.91G/1.91G [00:05&lt;00:00, 444MB/s]"
      }
     },
     "bd1b0dfed6d34d16af33a4a58330f5ec": {
@@ -7325,9 +7325,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_158c8b85dbf34de6a94b4e35e2fc7d5a",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_0b4c9753a7cb4354b8e5f187e6e1ad7c",
-      "value": "\u20073.96G/3.96G\u2007[00:15&lt;00:00,\u2007564MB/s]"
+      "value": " 3.96G/3.96G [00:15&lt;00:00, 564MB/s]"
      }
     },
     "c0991cf63ee6458b96e9a75e7a88b61a": {
@@ -7346,9 +7346,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_ed28e2e0410d4e0b855467e798e53d66",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_d93f134f802b4b69b575bdaf07dbd27c",
-      "value": "tokenizer_config.json:\u2007100%"
+      "value": "tokenizer_config.json: 100%"
      }
     },
     "c12ea43372ac4d57bb9605f1a429b397": {
@@ -7581,9 +7581,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_8bc9d8ba866c442b9118d9630009939c",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_9f56a2d9979c4bd8928c644c22c3ecdf",
-      "value": "model-00003-of-00008.safetensors:\u2007100%"
+      "value": "model-00003-of-00008.safetensors: 100%"
      }
     },
     "c6164e05a1914ae48083db9ad7f4ef7c": {
@@ -7694,9 +7694,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_e40d1c1ac9494b3bade9858324e7ffdf",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_d65b6b060d9845779299491ac5599c31",
-      "value": "\u20079985/9985\u2007[01:04&lt;00:00,\u2007189.08\u2007examples/s]"
+      "value": " 9985/9985 [01:04&lt;00:00, 189.08 examples/s]"
      }
     },
     "c7433acd3c4841e6958ae8f7e87b1808": {
@@ -7737,9 +7737,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_0077aedc3d174560bce924ee89e9c006",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_00321cce58884f6f9b3855a21fcd9187",
-      "value": "Add\u2007position_id\u2007column\u2007(Sample\u2007Packing)\u2007(num_proc=2):\u2007100%"
+      "value": "Add position_id column (Sample Packing) (num_proc=2): 100%"
      }
     },
     "ca65e32eb52f48c09a84b33cb18f22cd": {
@@ -8162,9 +8162,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_63580b6fb30642479fe3000915bf551a",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_8f726dbfb45d4528afa33e36a6313267",
-      "value": "\u200727.3M/27.3M\u2007[00:00&lt;00:00,\u200731.0MB/s]"
+      "value": " 27.3M/27.3M [00:00&lt;00:00, 31.0MB/s]"
      }
     },
     "d43c6df07ddb466587807d6dbe1ff614": {
@@ -8183,9 +8183,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_8c4d4fc5a30f4e7cb3be53fe2adda33d",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_e90658f4bcb642baa78426012f863152",
-      "value": "model-00004-of-00008.safetensors:\u2007100%"
+      "value": "model-00004-of-00008.safetensors: 100%"
      }
     },
     "d65b6b060d9845779299491ac5599c31": {
@@ -8474,9 +8474,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_34cf3df51fbc41cabfdbba153c007f0e",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_ac764024cf1c4e08ba7749afd2cd20ac",
-      "value": "vocab.json:\u2007100%"
+      "value": "vocab.json: 100%"
      }
     },
     "dfd2a2649b8341ef913207526708aff1": {
@@ -8669,9 +8669,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_c6164e05a1914ae48083db9ad7f4ef7c",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_813621384dc748b0ad06775e22761c0b",
-      "value": "\u20079985/9985\u2007[00:03&lt;00:00,\u20073622.89\u2007examples/s]"
+      "value": " 9985/9985 [00:03&lt;00:00, 3622.89 examples/s]"
      }
     },
     "e400cbf14bcc446a9d33b210cd93550b": {
@@ -9065,9 +9065,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_fba7aa824b38467ab3061b226114cdec",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_f3075dccbd2747b4a7913b66f44f2596",
-      "value": "\u20073.96G/3.96G\u2007[00:13&lt;00:00,\u2007398MB/s]"
+      "value": " 3.96G/3.96G [00:13&lt;00:00, 398MB/s]"
      }
     },
     "ec030fc3c346426f9abc3a89892258d3": {
@@ -9110,9 +9110,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_936d04b5fe1b4c63bf0b080e423d051b",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_f1cef8e8dc2646fb9fd09f3b09081074",
-      "value": "\u200736.5k/36.5k\u2007[00:00&lt;00:00,\u20074.32MB/s]"
+      "value": " 36.5k/36.5k [00:00&lt;00:00, 4.32MB/s]"
      }
     },
     "ed28e2e0410d4e0b855467e798e53d66": {
@@ -9422,9 +9422,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_735d4f225b24414294fc1b213c61223c",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_5e5e15b0569b474c9620083b3ec6af55",
-      "value": "generation_config.json:\u2007100%"
+      "value": "generation_config.json: 100%"
      }
     },
     "f4667818b9d34a09891cd727a429a610": {
@@ -9443,9 +9443,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_4b27c267393640f28f6eae0875bd2ed9",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_9858cb74a09748a39e8149baac96702c",
-      "value": "\u20073.96G/3.96G\u2007[00:11&lt;00:00,\u2007457MB/s]"
+      "value": " 3.96G/3.96G [00:11&lt;00:00, 457MB/s]"
      }
     },
     "f4a1795dc7514a718f478245f521f0ba": {
@@ -9830,9 +9830,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_d1f9b10c130542f094c8fd3d1e23b5e9",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_e575d87a7efe4ec7b1efde489839d4a6",
-      "value": "model-00006-of-00008.safetensors:\u2007100%"
+      "value": "model-00006-of-00008.safetensors: 100%"
      }
     },
     "fe18bba7f3fb4c31bf840541f36b3425": {
@@ -9873,9 +9873,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_e5a82df528bb4e408797a3b6c2758f4a",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_f113ebd8c1c34806bea4dd7ed3035173",
-      "value": "\u20079985/9985\u2007[00:00&lt;00:00,\u200744264.88\u2007examples/s]"
+      "value": " 9985/9985 [00:00&lt;00:00, 44264.88 examples/s]"
      }
     },
     "fea1b70fb46745feb5111b3929175b5d": {
@@ -9931,9 +9931,9 @@
       "description": "",
       "description_tooltip": null,
       "layout": "IPY_MODEL_ab93eabd7cea4b94b4b7a387f101e8a1",
-      "placeholder": "\u200b",
+      "placeholder": "​",
       "style": "IPY_MODEL_704f2f5a9b1c49d5a75a0025a5dda11b",
-      "value": "\u20073.96G/3.96G\u2007[00:12&lt;00:00,\u2007656MB/s]"
+      "value": " 3.96G/3.96G [00:12&lt;00:00, 656MB/s]"
      }
     }
    }

examples/deepcogito/cogito-v1-preview-llama-3B-lora.yml

@@ -9,6 +9,10 @@ strict: false
 datasets:
   - path: fozziethebeat/alpaca_messages_2k_test
     type: chat_template
+    field_messages: messages
+    message_property_mappings:
+      role: role
+      content: content
 
 dataset_prepared_path:
 val_set_size: 0.05

examples/deepcogito/cogito-v1-preview-qwen-14B-lora.yml

@@ -9,6 +9,10 @@ strict: false
 datasets:
   - path: fozziethebeat/alpaca_messages_2k_test
     type: chat_template
+    field_messages: messages
+    message_property_mappings:
+      role: role
+      content: content
 
 dataset_prepared_path:
 val_set_size: 0.05

examples/devstral/README.md

@@ -16,13 +16,8 @@ Thanks to the team at MistralAI for giving us early access to prepare for this r
 
 ```bash
 # Ensure you have Pytorch installed (Pytorch 2.6.0 min)
-# Option A: manage dependencies in your project
-uv add 'axolotl>=0.12.0'
-uv pip install flash-attn --no-build-isolation
-
-# Option B: quick install
-uv pip install 'axolotl>=0.12.0'
-uv pip install flash-attn --no-build-isolation
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
 2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage

examples/gemma3/gemma-3-4b-vision-qlora.yml

@@ -18,7 +18,7 @@ datasets:
   - path: HuggingFaceH4/llava-instruct-mix-vsft
     type: chat_template
     split: train[:1%]
-
+    field_messages: messages
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.01
 output_dir: ./outputs/out

examples/gemma3n/README.md

@@ -10,33 +10,20 @@ Gemma-3n is a family of multimodal models from Google found on [HuggingFace](htt
 
 ```bash
 # Ensure you have Pytorch installed (Pytorch 2.6.0 min)
-# Option A: manage dependencies in your project
-uv add 'axolotl>=0.12.0'
-uv pip install flash-attn --no-build-isolation
-
-# Option B: quick install
-uv pip install 'axolotl>=0.12.0'
-uv pip install flash-attn --no-build-isolation
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
 2. In addition to Axolotl's requirements, Gemma-3n requires:
 
 ```bash
-uv pip install timm==1.0.17
+pip3 install timm==1.0.17
 
 # for loading audio data
-uv pip install librosa==0.11.0
+pip3 install librosa==0.11.0
 ```
 
-3. Download sample dataset files
-
-```bash
-# for text + vision + audio only
-wget https://huggingface.co/datasets/Nanobit/text-vision-audio-2k-test/resolve/main/African_elephant.jpg
-wget https://huggingface.co/datasets/Nanobit/text-vision-audio-2k-test/resolve/main/En-us-African_elephant.oga
-```
-
-4. Run the finetuning example:
+3. Run the finetuning example:
 
 ```bash
 # text only

examples/gpt-oss/README.md

@@ -12,13 +12,8 @@ This guide shows how to fine-tune it with Axolotl with multi-turn conversations
 
 ```bash
 # Ensure you have Pytorch installed (Pytorch 2.6.0 min)
-# Option A: manage dependencies in your project
-uv add 'axolotl>=0.12.0'
-uv pip install flash-attn --no-build-isolation
-
-# Option B: quick install
-uv pip install 'axolotl>=0.12.0'
-uv pip install flash-attn --no-build-isolation
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
 2. Choose one of the following configs below for training the 20B model. (for 120B, see [below](#training-120b))
@@ -80,7 +75,7 @@ for more information about using a special vllm-openai docker image for inferenc
 Optionally, vLLM can be installed from nightly:
 
 ```bash
-uv pip install --no-build-isolation --pre -U vllm --extra-index-url https://wheels.vllm.ai/nightly
+pip install --no-build-isolation --pre -U vllm --extra-index-url https://wheels.vllm.ai/nightly
 ```
 and the vLLM server can be started with the following command (modify `--tensor-parallel-size 8` to match your environment):
 ```bash

examples/hunyuan/README.md

@@ -13,8 +13,8 @@ Tencent released a family of opensource models called HunYuan with varying param
 git clone https://github.com/axolotl-ai-cloud/axolotl.git
 cd axolotl
 
-uv sync
-uv pip install flash-attn --no-build-isolation
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install --no-build-isolation -e '.[flash-attn]'
 
 # Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
 python scripts/cutcrossentropy_install.py | sh

examples/llama-2/qlora-fsdp.yml

@@ -66,7 +66,6 @@ fsdp_config:
   fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
   fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
   fsdp_state_dict_type: FULL_STATE_DICT
-  # fsdp_cpu_offload_pin_memory: false  # uncomment to enable swap memory usage when RAM is insufficient
 special_tokens:
 
 # save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/llama-3/instruct-lora-8b.yml

@@ -12,6 +12,15 @@ chat_template: llama3
 datasets:
   - path: fozziethebeat/alpaca_messages_2k_test
     type: chat_template
+    field_messages: messages
+    message_property_mappings:
+      role: role
+      content: content
+    roles:
+      user:
+        - user
+      assistant:
+        - assistant
 
 dataset_prepared_path:
 val_set_size: 0.05

examples/llama-4/do-no-use-fa2/scout-vision-qlora-fsdp.yaml

@@ -46,6 +46,7 @@ datasets:
   - path: HuggingFaceH4/llava-instruct-mix-vsft
     type: chat_template
     split: train[:1%]
+    field_messages: messages
 
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0

examples/llama-4/scout-vision-qlora-fsdp2-flex.yaml

@@ -45,6 +45,7 @@ datasets:
   - path: HuggingFaceH4/llava-instruct-mix-vsft
     type: chat_template
     split: train[:1%]
+    field_messages: messages
 
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0

examples/magistral/README.md

@@ -1,10 +1,10 @@
 # Finetune Magistral Small with Axolotl
 
-Magistral Small is a 24B parameter opensource model from MistralAI found on HuggingFace at [2506](https://huggingface.co/mistralai/Magistral-Small-2506), [2507](https://huggingface.co/mistralai/Magistral-Small-2507) (see [Thinking](#thinking)), and [2509](https://huggingface.co/mistralai/Magistral-Small-2509) (see [Vision](#vision)). This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
+Magistral Small is a 24B parameter opensource model from MistralAI found on HuggingFace at [2506](https://huggingface.co/mistralai/Magistral-Small-2506) and [2507](https://huggingface.co/mistralai/Magistral-Small-2507) (see [Thinking](#thinking)). This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
 
 MistralAI has also released a proprietary medium-sized version called Magistral Medium.
 
-Thanks to the team at MistralAI for giving us early access to prepare for these releases.
+Thanks to the team at MistralAI for giving us early access to prepare for this release.
 
 ## Getting started
 
@@ -13,14 +13,9 @@ Thanks to the team at MistralAI for giving us early access to prepare for these
     Here is an example of how to install from pip:
 
 ```bash
-# Ensure you have PyTorch installed (PyTorch 2.6.0 min)
-# Option A: manage dependencies in your project
-uv add 'axolotl>=0.12.0'
-uv pip install flash-attn --no-build-isolation
-
-# Option B: quick install
-uv pip install 'axolotl>=0.12.0'
-uv pip install flash-attn --no-build-isolation
+# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
 2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage
@@ -41,17 +36,29 @@ Let us know how it goes. Happy finetuning! 🚀
 
 ### Thinking
 
-MistralAI has released their [2507](https://huggingface.co/mistralai/Magistral-Small-2507) model with thinking capabilities, enabling Chain-of-Thought reasoning with explicit thinking steps.
+MistralAI has released their [2507](https://huggingface.co/mistralai/Magistral-Small-2507) model with thinking capabilities. The model requires the multi-content dataset format with support for an extra `role: thinking` within system and assistant messages.
 
-📚 **[See the Thinking fine-tuning guide →](./think/README.md)**
+Example format:
 
-### Vision
+```json
+{
+    "messages": [
+        {"role": "system", "content": [{ "type": "text", "text": "{SYSTEM_PROMPT}"}]},
+        {"role": "user", "content": [{ "type": "text", "text": "..."}]},
+        {"role": "assistant", "content": [{ "type": "thinking", "thinking": "..."}, { "type": "text", "text": "..." }]},
+    ],
+}
+```
 
-MistralAI has released their [2509](https://huggingface.co/mistralai/Magistral-Small-2509) model with vision capabilities.
+Example config: `./magistral-small-think-qlora.yaml`.
 
-📚 **[See the Vision fine-tuning guide →](./vision/README.md)**
+The `thinking` section also supports an optional arg `closed: bool` (`True` default) which controls adding the closing `[/THINK]` tag.
 
-### Tips
+Limitations:
+- You cannot mix `content: str` with `content: list[dict]` as the `dataset.load_dataset` may complain about different types for `content` key.
+- This mode does not work with custom `train_detail` and `training` at the moment.
+
+### TIPS
 
 - We recommend adding the same/similar SystemPrompt that the model is tuned for. You can find this within the repo's files titled `SYSTEM_PROMPT.txt`.
 - For inference, the official MistralAI team recommends `top_p: 0.95` and `temperature: 0.7` with `max_tokens: 40960`.
@@ -82,5 +89,5 @@ In addition, we do not support overriding tokens yet.
 
 ## Future Work
 
-- Add parity to Preference Tuning, RL, etc.
+- Add parity to Preference Tuning, RL, Multi-modal, etc.
 - Add parity to other tokenizer configs like overriding tokens.

examples/magistral/think/README.md

@@ -1,73 +0,0 @@
-# Magistral Small Thinking Fine-tuning
-
-This guide covers fine-tuning [Magistral Small 2507](https://huggingface.co/mistralai/Magistral-Small-2507) with thinking capabilities using Axolotl. The thinking model enables explicit Chain-of-Thought reasoning with separate thinking and response sections.
-
-## Prerequisites
-
-Before starting, ensure you have:
-- Installed Axolotl (see [main README](../README.md))
-
-## Getting Started
-
-Run the thinking model fine-tuning:
-
-```bash
-axolotl train magistral-small-think-qlora.yaml
-```
-
-This config uses about 19.1 GiB VRAM.
-
-### Tips
-
-- Dataset uses multi-content format with `type: thinking` support. See [Dataset Format](#dataset-format) below.
-- You cannot mix `content: str` and `content: list[dict]`, otherwise, dataset loading will fail. Keep it consistent.
-
-## Dataset Format
-
-The thinking model requires the multi-content dataset format with support for an extra `role: thinking` within system and assistant messages.
-
-Example format:
-
-```json
-{
-    "messages": [
-        {
-            "role": "system",
-            "content": [
-                { "type": "text", "text": "{SYSTEM_PROMPT}"}
-            ]
-        },
-        {
-            "role": "user",
-            "content": [
-                { "type": "text", "text": "Solve this step by step: What is 15% of 240?"}
-            ]
-        },
-        {
-            "role": "assistant",
-            "content": [
-                {
-                    "type": "thinking",
-                    "thinking": "I need to calculate 15% of 240. First, I'll convert 15% to decimal: 0.15. Then multiply: 0.15 × 240 = 36."
-                },
-                {
-                    "type": "text",
-                    "text": "To find 15% of 240, I'll multiply 240 by 0.15:\n\n240 × 0.15 = 36\n\nTherefore, 15% of 240 is 36."
-                }
-            ]
-        }
-    ]
-}
-```
-
-### Advanced Options
-
-The `thinking` section supports an optional `closed` parameter:
-
-```json
-{
-    "type": "thinking",
-    "thinking": "Internal reasoning here...",
-    "closed": true  // Default: true, controls adding the closing [/THINK] tag
-}
-```

examples/magistral/vision/README.md

@@ -1,60 +0,0 @@
-# Magistral Small Vision Fine-tuning
-
-This guide covers fine-tuning [Magistral Small 2509](https://huggingface.co/mistralai/Magistral-Small-2509) with vision capabilities using Axolotl.
-
-## Prerequisites
-
-Before starting, ensure you have:
-- Installed Axolotl from source (see [main README](../README.md#getting-started))
-
-## Getting started
-
-1. Install the required vision lib:
-    ```bash
-    pip install 'mistral-common[opencv]==1.8.5'
-    ```
-
-2. Download the example dataset image:
-   ```bash
-   wget https://huggingface.co/datasets/Nanobit/text-vision-2k-test/resolve/main/African_elephant.jpg
-   ```
-
-3. Run the fine-tuning:
-   ```bash
-   axolotl train magistral-small-vision-24B-qlora.yml
-   ```
-
-This config uses about 17GiB VRAM.
-
-WARNING: The loss and grad norm will be much higher than normal at first. We suspect this to be inherent to the model as of the moment. If anyone would like to submit a fix for this, we are happy to take a look.
-
-### Tips
-
-Key differences from text-only model:
-- `max_tokens: 131072` for inference
-- Multi-modal dataset format required
-- Sample packing not supported
-
-## Dataset Format
-
-The vision model requires multi-modal dataset format as documented [here](https://docs.axolotl.ai/docs/multimodal.html#dataset-format).
-
-One exception is that, passing `"image": PIL.Image` is not supported. MistralTokenizer only supports `path`, `url`, and `base64` for now.
-
-Example:
-```json
-{
-    "messages": [
-        {"role": "system", "content": [{ "type": "text", "text": "{SYSTEM_PROMPT}"}]},
-        {"role": "user", "content": [
-            { "type": "text", "text": "What's in this image?"},
-            {"type": "image", "path": "path/to/image.jpg" }
-        ]},
-        {"role": "assistant", "content": [{ "type": "text", "text": "..." }]},
-    ],
-}
-```
-
-## Limitations
-
-- Sample Packing is not supported for multi-modality training currently.

examples/magistral/vision/magistral-small-vision-24B-qlora.yml

@@ -1,64 +0,0 @@
-base_model: mistralai/Magistral-Small-2509
-processor_type: AutoProcessor
-
-# Enable to use mistral-common tokenizer
-tokenizer_use_mistral_common: true
-
-plugins:
-  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
-
-load_in_4bit: true
-
-# these 3 lines are needed for now to handle vision chat templates w images
-skip_prepare_dataset: true
-remove_unused_columns: false
-sample_packing: false
-
-# sample dataset below requires downloading image in advance
-# wget https://huggingface.co/datasets/Nanobit/text-vision-2k-test/resolve/main/African_elephant.jpg
-datasets:
-  - path: Nanobit/text-vision-2k-test
-    type: chat_template
-
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
-output_dir: ./outputs/out
-
-adapter: qlora
-lora_model_dir:
-
-sequence_len: 2048
-
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 1
-micro_batch_size: 1
-num_epochs: 1
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-bf16: true
-fp16:
-tf32: true
-
-gradient_checkpointing: true
-logging_steps: 1
-flash_attention: true
-
-warmup_ratio: 0.1
-evals_per_epoch: 1
-saves_per_epoch: 1
-weight_decay: 0.0
-special_tokens:
-
-# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/mistral/mistral-small-3.1-24B-lora.yml

@@ -1,9 +1,6 @@
 base_model: mistralai/Mistral-Small-3.1-24B-Instruct-2503
 processor_type: AutoProcessor
 
-# Enable to use mistral-common tokenizer
-tokenizer_use_mistral_common: true
-
 load_in_8bit: true
 
 # these 3 lines are needed for now to handle vision chat templates w images
@@ -11,12 +8,12 @@ skip_prepare_dataset: true
 remove_unused_columns: false
 sample_packing: false
 
-# sample dataset below requires downloading image in advance
-# wget https://huggingface.co/datasets/Nanobit/text-vision-2k-test/resolve/main/African_elephant.jpg
+chat_template: mistral_v7_tekken
 datasets:
-  - path: Nanobit/text-vision-2k-test
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
     type: chat_template
-
+    split: train[:1%]
+    field_messages: messages
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.01
 output_dir: ./outputs/out
@@ -51,7 +48,8 @@ tf32: true
 
 gradient_checkpointing: true
 logging_steps: 1
-flash_attention: true
+# flash_attention: false # PixtralVisionModel does not support Flash Attention 2.0 yet.
+sdp_attention: true
 
 warmup_ratio: 0.1
 evals_per_epoch: 1

examples/moe/qwen2-moe-qlora-10gb.yaml

@@ -0,0 +1,53 @@
+base_model: Qwen/Qwen1.5-MoE-A2.7B
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+trust_remote_code: true
+
+# Keep VRAM low
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.05
+output_dir: ./outputs/qwen2-moe-qlora-10gb
+
+# Train small to fit 10GB
+sequence_len: 512
+sample_packing: false
+pad_to_sequence_len: false
+
+adapter: qlora
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+gradient_accumulation_steps: 8
+micro_batch_size: 1
+num_epochs: 1
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 5
+flash_attention: true
+
+warmup_ratio: 0.03
+evals_per_epoch: 2
+saves_per_epoch: 1
+weight_decay: 0.0
+
+model_config:
+  output_router_logits: true
+
+special_tokens:

examples/phi/lora-3.5.yaml

@@ -12,6 +12,15 @@ chat_template: phi_3
 datasets:
   - path: fozziethebeat/alpaca_messages_2k_test
     type: chat_template
+    field_messages: messages
+    message_property_mappings:
+      role: role
+      content: content
+    roles:
+      user:
+        - user
+      assistant:
+        - assistant
 
 dataset_prepared_path:
 val_set_size: 0.05

examples/pixtral/lora-12b.yml

@@ -45,7 +45,8 @@ tf32: true
 
 gradient_checkpointing: true
 logging_steps: 1
-flash_attention: true
+# flash_attention:  # PixtralVisionModel does not support Flash Attention 2.0 yet
+sdp_attention: true
 
 warmup_ratio: 0.1
 evals_per_epoch: 1

examples/qwen2-vl/lora-7b.yaml

@@ -11,7 +11,7 @@ datasets:
   - path: HuggingFaceH4/llava-instruct-mix-vsft
     type: chat_template
     split: train[:1%]
-
+    field_messages: messages
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0
 output_dir: ./outputs/out

examples/qwen2_5-vl/lora-7b.yaml

@@ -11,7 +11,7 @@ datasets:
   - path: HuggingFaceH4/llava-instruct-mix-vsft
     type: chat_template
     split: train[:1%]
-
+    field_messages: messages
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0
 output_dir: ./outputs/out

examples/qwen3-next/README.md

@@ -1,64 +0,0 @@
-# Finetune Qwen3-Next with Axolotl
-
-[Qwen3-Next](https://huggingface.co/collections/Qwen/qwen3-next-68c25fd6838e585db8eeea9d) represents the next-generation foundation models optimized for extreme context length and large-scale parameter efficiency. The series introduces architectural innovations including Hybrid Attention (Gated DeltaNet + Gated Attention), High-Sparsity MoE with 1:50 activation ratio, and Multi-Token Prediction for enhanced performance and inference acceleration.
-
-This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
-
-## Getting started
-
-1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html). You need to install from main as Qwen3-Next 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 min)
-git clone https://github.com/axolotl-ai-cloud/axolotl.git
-cd axolotl
-
-uv sync
-uv pip install flash-attn --no-build-isolation
-
-# Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
-python scripts/cutcrossentropy_install.py | sh
-```
-
-2. Install Qwen3-Next transformers commit
-```bash
-uv pip uninstall -y transformers && uv pip install "git+https://github.com/huggingface/transformers.git@b9282355bea846b54ed850a066901496b19da654"
-```
-
-3. Install FLA for improved performance
-```bash
-uv pip uninstall -y causal-conv1d && uv pip install flash-linear-attention==0.3.2
-```
-
-4. Run the finetuning example:
-
-```bash
-axolotl train examples/qwen3-next/qwen3-next-80b-a3b-qlora.yaml
-```
-
-This config uses about 45.62 GiB VRAM.
-
-Let us know how it goes. Happy finetuning! 🚀
-
-### TIPS
-
-- For inference, you can experiment with `temperature: 0.7`, `top_p: 0.8`, `top_k: 20`, and `min_p: 0`.
-- You can run a full finetuning by removing the `adapter: qlora` and `load_in_4bit: true` from the config. See [Multi-GPU](#optimization-guides) section below.
-- 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).
-
-## 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)
-
-## Related Resources
-
-- [Qwen3-Next Blog](https://qwenlm.github.io/blog/qwen3_next/)
-- [Axolotl Docs](https://docs.axolotl.ai)
-- [Axolotl Website](https://axolotl.ai)
-- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
-- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/qwen3-next/qwen3-next-80b-a3b-qlora.yaml

@@ -1,68 +0,0 @@
-base_model: Qwen/Qwen3-Next-80B-A3B-Instruct
-
-# Automatically upload checkpoint and final model to HF
-# hub_model_id: username/custom_model_name
-
-plugins:
-  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
-
-load_in_8bit: false
-load_in_4bit: true
-
-datasets:
-  - path: fozziethebeat/alpaca_messages_2k_test
-    type: chat_template
-
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.1
-output_dir: ./outputs/lora-out
-
-adapter: qlora
-lora_model_dir:
-
-sequence_len: 2048
-sample_packing: true
-
-lora_r: 16
-lora_alpha: 8
-lora_dropout: 0.05
-lora_target_modules:
-  - linear_attn.in_proj_ba
-  - linear_attn.in_proj_qkvz
-  - linear_attn.out_proj
-  - shared_expert.up_proj
-  - shared_expert.down_proj
-  - shared_expert.gate_proj
-  - shared_expert_gate
-  - mlp.gate
-  - q_proj
-  - v_proj
-  - k_proj
-  - o_proj
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 2
-micro_batch_size: 2
-num_epochs: 1
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-bf16: auto
-tf32: false
-
-gradient_checkpointing: true
-resume_from_checkpoint:
-logging_steps: 1
-flash_attention: true
-
-warmup_ratio: 0.1
-evals_per_epoch: 1
-saves_per_epoch: 1
-
-# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/seed-oss/README.md

@@ -15,8 +15,8 @@ This guide shows how to fine-tune it with Axolotl with multi-turn conversations
 git clone https://github.com/axolotl-ai-cloud/axolotl.git
 cd axolotl
 
-uv sync --extra deepspeed
-uv pip install flash-attn --no-build-isolation
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install --no-build-isolation -e '.[flash-attn]'
 
 # Install Cut Cross Entropy
 python scripts/cutcrossentropy_install.py | sh

examples/smolvlm2/README.md

@@ -13,19 +13,14 @@ This guide shows how to fine-tune SmolVLM2 models with Axolotl.
     Here is an example of how to install from pip:
     ```bash
     # Ensure you have a compatible version of Pytorch installed
-    # Option A: manage dependencies in your project
-    uv add 'axolotl>=0.12.0'
-    uv pip install flash-attn --no-build-isolation
-
-    # Option B: quick install
-    uv pip install 'axolotl>=0.12.0'
-    uv pip install flash-attn --no-build-isolation
+    pip3 install packaging setuptools wheel ninja
+    pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
     ```
 
 2. Install an extra dependency:
 
     ```bash
-    uv pip install num2words==0.5.14
+    pip3 install num2words==0.5.14
     ```
 
 3.  Run the finetuning example:

examples/voxtral/README.md

@@ -12,34 +12,22 @@ Thanks to the team at MistralAI for giving us early access to prepare for this r
 
 ```bash
 # Ensure you have Pytorch installed (Pytorch 2.6.0 min)
-# Option A: manage dependencies in your project
-uv add 'axolotl>=0.12.0'
-uv pip install flash-attn --no-build-isolation
-
-# Option B: quick install
-uv pip install 'axolotl>=0.12.0'
-uv pip install flash-attn --no-build-isolation
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
 2. Please install the below.
 
 ```bash
 # audio
-uv pip install librosa==0.11.0
-uv pip install 'mistral_common[audio]==1.8.3'
+pip3 install librosa==0.11.0
+pip3 install 'mistral_common[audio]==1.8.3'
 
 # Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
 python scripts/cutcrossentropy_install.py | sh
 ```
 
-3. Download sample dataset files
-
-```bash
-# for text + audio only
-wget https://huggingface.co/datasets/Nanobit/text-audio-2k-test/resolve/main/En-us-African_elephant.oga
-```
-
-4. Run the finetuning example:
+3. Run the finetuning example:
 
 ```bash
 # text only

pyproject.toml

@@ -1,131 +1,14 @@
 [build-system]
-requires = ["setuptools>=64", "wheel", "setuptools_scm>=8"]
+requires = ["setuptools>=64", "wheel", "setuptools_scm>=8", "packaging==23.2"]
 build-backend = "setuptools.build_meta"
 
 [project]
 name = "axolotl"
-dynamic = ["version"]
+dynamic = ["version", "dependencies", "optional-dependencies"]
 description = "LLM Trainer"
 readme = "README.md"
-requires-python = ">=3.10,<3.13"
-license = {text = "Apache-2.0"}
-authors = [
-    {name = "Axolotl AI"},
-]
-maintainers = [
-    {name = "Axolotl AI"},
-]
-classifiers = [
-    "Development Status :: 4 - Beta",
-    "License :: OSI Approved :: Apache Software License",
-    "Programming Language :: Python :: 3",
-    "Programming Language :: Python :: 3.10",
-    "Programming Language :: Python :: 3.11",
-    "Programming Language :: Python :: 3.12",
-]
-
-dependencies = [
-    "torch>=2.6.0",
-    "packaging>=23.2",
-    "huggingface_hub>=0.33.0",
-    "peft==0.17.0",
-    "transformers==4.56.1",
-    "tokenizers>=0.21.1",
-    "accelerate==1.10.1",
-    "datasets==4.0.0",
-    "trl==0.23.0",
-    "hf_xet==1.1.5",
-    "kernels==0.9.0",
-    "trackio",
-    "optimum==1.16.2",
-    "hf_transfer",
-    "sentencepiece",
-    "gradio==5.41.1",
-    "modal==1.0.2",
-    "pydantic>=2.10.6",
-    "addict",
-    "fire",
-    "PyYAML>=6.0",
-    "requests",
-    "wandb",
-    "einops",
-    "colorama",
-    "numba",
-    "numpy>=1.24.4,<3.0",
-    "evaluate==0.4.1",
-    "scipy",
-    "scikit-learn>=1.7.0",
-    "nvidia-ml-py==12.560.30",
-    "art",
-    "tensorboard",
-    "python-dotenv==1.0.1",
-    "s3fs>=2024.5.0",
-    "gcsfs>=2024.5.0",
-    "adlfs>=2024.5.0",
-    "ocifs==1.3.2",
-    "zstandard>=0.23.0",
-    "fastcore",
-    "lm_eval==0.4.7",
-    "langdetect==1.0.9",
-    "immutabledict==4.2.0",
-    "antlr4-python3-runtime==4.13.2",
-    "schedulefree==1.4.1",
-    "mistral-common==1.8.5",
-
-    # Axolotl contribs
-    "axolotl-contribs-lgpl @ git+https://github.com/axolotl-ai-cloud/axolotl-contribs-lgpl.git@numpy",
-    "axolotl-contribs-mit==0.0.5",
-
-    # Platform-specific dependencies (Linux by default, excluded on macOS)
-    "triton>=3.0.0 ; sys_platform != 'darwin'",
-    "xformers>=0.0.28 ; sys_platform != 'darwin'",
-    "autoawq==0.2.7.post3 ; sys_platform != 'darwin'",
-    "liger-kernel==0.6.1 ; sys_platform != 'darwin'",
-    "torchao==0.13.0 ; sys_platform != 'darwin'",
-    "bitsandbytes==0.47.0 ; sys_platform != 'darwin'",
-    "deepspeed>=0.17.5 ; sys_platform != 'darwin'",
-    "deepspeed-kernels ; sys_platform != 'darwin'",
-]
-
-[project.optional-dependencies]
-ring-flash-attn = [
-    "ring-flash-attn>=0.1.7",
-    "yunchang==0.6.0",
-]
-mamba-ssm = ["mamba-ssm>=2.2.0", "causal_conv1d>=1.4.0",]
-gptqmodel = ["gptqmodel>=4.0.0"]
-mlflow = ["mlflow"]
-galore = ["galore_torch"]
-apollo = ["apollo-torch"]
-optimizers = [
-    "galore_torch",
-    "apollo-torch",
-    "lomo-optim==0.1.1",
-    "torch-optimi==0.2.1",
-    "came_pytorch==0.1.3",
-]
-ray = ["ray[train]"]
-vllm = ["vllm>=0.10.0"]
-llmcompressor = ["llmcompressor>=0.5.1"]
-fbgemm-gpu = ["fbgemm-gpu-genai>=1.2.0"]
-dev = [
-    "pytest",
-    "pytest-cov",
-    "pytest-retry",
-    "pytest-sugar",
-    "pytest-xdist",
-    "codecov",
-    "codecov-cli",
-    "tbparse",
-    "ruff",
-    "mypy",
-    "pre-commit",
-    "types-requests",
-    "quartodoc",
-    "jupyter",
-    "blobfile",
-    "tiktoken",
-]
+requires-python = ">=3.10"
+# license = "Apache-2.0"
 
 [project.scripts]
 axolotl = "axolotl.cli.main:main"
@@ -134,20 +17,15 @@ axolotl = "axolotl.cli.main:main"
 Homepage = "https://axolotl.ai/"
 Documentation = "https://docs.axolotl.ai/"
 Repository = "https://github.com/axolotl-ai-cloud/axolotl.git"
-Issues = "https://github.com/axolotl-ai-cloud/axolotl/issues"
-
-[tool.setuptools]
-package-dir = {"" = "src"}
-include-package-data = true
-
-[tool.setuptools.packages.find]
-where = ["src"]
-
-[tool.setuptools.package-data]
-"*" = ["*.yaml", "*.yml", "*.json"]
 
 [tool.setuptools_scm]
-write_to = "src/axolotl/_version.py"
+
+[tool.setuptools]
+py-modules = ["setuptools_axolotl_dynamic_dependencies"]
+include-package-data = true
+
+[tool.setuptools.cmdclass]
+build_py = "setuptools_axolotl_dynamic_dependencies.BuildPyCommand"
 
 [tool.ruff]
 line-length = 88
@@ -179,60 +57,3 @@ indent-style = "space"
 skip-magic-trailing-comma = false
 line-ending = "auto"
 docstring-code-format = false
-
-[tool.mypy]
-python_version = "3.11"
-warn_return_any = true
-warn_unused_configs = true
-ignore_missing_imports = true
-
-[tool.pytest.ini_options]
-testpaths = ["tests"]
-python_files = ["test_*.py", "*_test.py"]
-addopts = "-v --tb=short"
-
-# UV specific configuration
-[tool.uv]
-prerelease = "allow"
-default-groups = ["default"]
-conflicts = [
-    [
-        { group = "default" },
-        { extra = "vllm" },
-    ],
-]
-
-[dependency-groups]
-default = ["torch>=2.6.0"]
-dev = [
-    "pytest",
-    "pytest-cov",
-    "pytest-retry",
-    "pytest-sugar",
-    "pytest-xdist",
-    "codecov",
-    "codecov-cli",
-    "tbparse",
-    "ruff",
-    "mypy",
-    "pre-commit",
-    "types-requests",
-    "quartodoc",
-    "jupyter",
-    "blobfile",
-    "tiktoken",
-]
-
-[[tool.uv.index]]
-name = "autogptq"
-url = "https://huggingface.github.io/autogptq-index/whl/"
-
-[tool.uv.extra-build-dependencies]
-mamba-ssm = ["torch", "causal_conv1d"]
-gptqmodel = [
-    { requirement = "torch", match-runtime = true },
-]
-autoawq = ["torch"]
-triton = ["torch"]
-bitsandbytes = ["torch"]
-grpclib = ["wheel"]

requirements-dev.txt

@@ -0,0 +1,8 @@
+black
+mypy
+pre-commit
+types-requests
+quartodoc
+jupyter
+blobfile
+tiktoken

requirements-tests.txt

@@ -0,0 +1,8 @@
+codecov
+codecov-cli
+pytest
+pytest-cov
+pytest-retry
+pytest-sugar
+pytest-xdist
+tbparse

requirements.txt

@@ -0,0 +1,73 @@
+--extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
+
+# START section of dependencies that don't install on Darwin/MacOS
+bitsandbytes==0.47.0
+triton>=3.0.0
+mamba-ssm==1.2.0.post1
+xformers>=0.0.23.post1
+autoawq==0.2.7.post3
+liger-kernel==0.6.1
+# END section
+
+packaging==23.2
+
+huggingface_hub>=0.33.0
+peft>=0.17.0
+transformers==4.56.1
+tokenizers>=0.21.1
+accelerate==1.10.1
+datasets==4.0.0
+deepspeed>=0.17.0
+trl==0.23.0
+hf_xet==1.1.5
+kernels==0.9.0
+trackio
+
+optimum==1.16.2
+hf_transfer
+sentencepiece
+gradio==5.41.1
+
+modal==1.0.2
+pydantic==2.10.6
+addict
+fire
+PyYAML>=6.0
+requests
+wandb
+einops
+colorama
+numba
+numpy>=1.24.4,<=2.0.1
+
+# qlora things
+evaluate==0.4.1
+scipy
+scikit-learn==1.4.2
+nvidia-ml-py==12.560.30
+art
+tensorboard
+python-dotenv==1.0.1
+
+# remote filesystems
+s3fs>=2024.5.0
+gcsfs>=2024.5.0
+adlfs>=2024.5.0
+ocifs==1.3.2
+
+zstandard==0.22.0
+fastcore
+
+# lm eval harness
+lm_eval==0.4.7
+langdetect==1.0.9
+immutabledict==4.2.0
+antlr4-python3-runtime==4.13.2
+
+torchao==0.13.0
+schedulefree==1.4.1
+
+axolotl-contribs-lgpl==0.0.6
+axolotl-contribs-mit==0.0.5
+
+mistral-common==1.8.3

scripts/bench_moe.py

@@ -0,0 +1,209 @@
+#!/usr/bin/env python
+"""Benchmark Hugging Face Qwen2 MoE block with and without grouped_mm."""
+
+from __future__ import annotations
+
+import argparse
+import sys
+import time
+import weakref
+from pathlib import Path
+
+import torch
+import torch._dynamo as dynamo
+
+try:
+    from axolotl.kernels.moe import torch_grouped as tg
+except Exception:  # pragma: no cover
+    tg = None
+
+
+def bench(run, *, iters: int, warmup: int, sync: bool = True) -> float:
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    for _ in range(warmup):
+        run()
+        if sync and device.type == "cuda":
+            torch.cuda.synchronize()
+    times = []
+    for _ in range(iters):
+        if sync and device.type == "cuda":
+            torch.cuda.synchronize()
+        start = time.perf_counter()
+        run()
+        if sync and device.type == "cuda":
+            torch.cuda.synchronize()
+        times.append((time.perf_counter() - start) * 1000.0)
+    return sum(times) / len(times)
+
+
+def estimate_moe_flops(tokens: int, hidden: int, inter: int, top_k: int) -> float:
+    return 6.0 * tokens * top_k * hidden * inter
+
+
+def load_hf_block(
+    hidden: int,
+    inter: int,
+    experts: int,
+    top_k: int,
+    *,
+    device: torch.device,
+    dtype: torch.dtype,
+):
+    project_root = Path(__file__).resolve().parents[2]
+    transformers_src = project_root / "transformers" / "src"
+    if transformers_src.exists() and str(transformers_src) not in sys.path:
+        sys.path.append(str(transformers_src))
+
+    from transformers.models.qwen2_moe.configuration_qwen2_moe import Qwen2MoeConfig
+    from transformers.models.qwen2_moe.modeling_qwen2_moe import Qwen2MoeSparseMoeBlock
+
+    cfg = Qwen2MoeConfig(
+        hidden_size=hidden,
+        moe_intermediate_size=inter,
+        shared_expert_intermediate_size=inter,
+        num_experts=experts,
+        num_experts_per_tok=top_k,
+        norm_topk_prob=True,
+        qkv_bias=True,
+    )
+
+    block = Qwen2MoeSparseMoeBlock(cfg).to(device=device, dtype=dtype)
+    block_grouped = Qwen2MoeSparseMoeBlock(cfg).to(device=device, dtype=dtype)
+    block_grouped.load_state_dict(block.state_dict())
+    return block, block_grouped
+
+
+def main() -> None:
+    p = argparse.ArgumentParser(description="Qwen2 MoE grouped_mm benchmark")
+    p.add_argument("--bsz", type=int, default=8)
+    p.add_argument("--seq", type=int, default=1024)
+    p.add_argument("--hidden", type=int, default=4096)
+    p.add_argument("--inter", type=int, default=14336)
+    p.add_argument("--experts", type=int, default=32)
+    p.add_argument("--top_k", type=int, default=4)
+    p.add_argument("--dtype", choices=["bf16", "fp16", "fp32"], default="bf16")
+    p.add_argument("--iters", type=int, default=50)
+    p.add_argument("--warmup", type=int, default=10)
+    p.add_argument("--profile", action="store_true")
+    p.add_argument(
+        "--compile",
+        action="store_true",
+        help="Torch.compile both paths before benchmarking",
+    )
+    args = p.parse_args()
+
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    dtype = {
+        "bf16": torch.bfloat16,
+        "fp16": torch.float16,
+        "fp32": torch.float32,
+    }[args.dtype]
+
+    torch.manual_seed(0)
+    if device.type == "cuda":
+        torch.cuda.manual_seed(0)
+
+    block_naive, block_grouped = load_hf_block(
+        args.hidden,
+        args.inter,
+        args.experts,
+        args.top_k,
+        device=device,
+        dtype=dtype,
+    )
+
+    tokens = args.bsz * args.seq
+    flops_total = estimate_moe_flops(tokens, args.hidden, args.inter, args.top_k)
+    print(
+        f"Device={device} dtype={dtype} tokens={tokens} hidden={args.hidden} inter={args.inter} "
+        f"experts={args.experts} top_k={args.top_k}"
+    )
+
+    x = torch.randn(args.bsz, args.seq, args.hidden, device=device, dtype=dtype)
+
+    # Optional torch.compile
+    run_grouped_impl = None
+    if args.compile:
+        dynamo.config.capture_scalar_outputs = True
+        dynamo.config.allow_unspec_int_on_nn_module = True
+        try:
+            block_naive = torch.compile(block_naive)  # type: ignore[arg-type]
+        except Exception as exc:  # pragma: no cover
+            print(f"torch.compile naive failed ({exc}); using eager")
+        else:
+
+            def grouped_forward(inp, *, block=block_grouped):
+                block.experts._ax_parent_block_ref = weakref.ref(block)  # type: ignore[attr-defined]
+                y, _ = tg.moe_ffn_forward_grouped(
+                    inp, block.gate, block.experts, block.top_k
+                )
+                return y
+
+            try:
+                run_grouped_impl = torch.compile(grouped_forward)  # type: ignore[arg-type]
+            except Exception as exc:  # pragma: no cover
+                print(f"torch.compile grouped failed ({exc}); using eager")
+                run_grouped_impl = None
+
+    def run_naive(block=block_naive, data=x):
+        y, _ = block(data)
+        return y
+
+    def run_grouped(block=block_grouped, data=x, impl=run_grouped_impl):
+        if impl is not None:
+            return impl(data)
+        if tg is None or not tg.available():
+            return torch.empty(0)
+        block.experts._ax_parent_block_ref = weakref.ref(block)  # type: ignore[attr-defined]
+        y, _ = tg.moe_ffn_forward_grouped(data, block.gate, block.experts, block.top_k)
+        return y if y is not None else torch.empty(0)
+
+    t_naive = bench(run_naive, iters=args.iters, warmup=args.warmup)
+    tflops_naive = flops_total / ((t_naive / 1000.0) * 1e12)
+    print(
+        f"naive\t{t_naive:.2f} ms\t{tokens / (t_naive / 1000.0):.1f} tok/s\t{tflops_naive:.2f} TFLOP/s"
+    )
+
+    with torch.no_grad():
+        y_ref = run_naive()
+
+    if tg is None or not tg.available():
+        print("torch_grouped\tN/A (unavailable)")
+        return
+
+    y_grouped = run_grouped()
+    if y_grouped.numel() == 0:
+        print("torch_grouped\tN/A (op not callable)")
+        return
+
+    t_grouped = bench(run_grouped, iters=args.iters, warmup=args.warmup)
+    tflops_grouped = flops_total / ((t_grouped / 1000.0) * 1e12)
+    speedup = t_naive / t_grouped
+    print(
+        f"torch_grouped\t{t_grouped:.2f} ms\t{tokens / (t_grouped / 1000.0):.1f} tok/s\t"
+        f"{tflops_grouped:.2f} TFLOP/s\t{speedup:.2f}×"
+    )
+
+    diff = (y_ref.float() - y_grouped.float()).abs()
+    print(
+        "torch_grouped_check: "
+        f"max_abs={diff.max().item():.3e} mean_abs={diff.mean().item():.3e} "
+        f"rel_l2={(diff.pow(2).sum() / (y_ref.float().pow(2).sum() + 1e-12)).sqrt().item():.3e}"
+    )
+
+    if args.profile:
+        with torch.profiler.profile(
+            activities=[torch.profiler.ProfilerActivity.CUDA], record_shapes=True
+        ) as prof:
+            run_naive()
+        print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=20))
+
+        with torch.profiler.profile(
+            activities=[torch.profiler.ProfilerActivity.CUDA], record_shapes=True
+        ) as prof:
+            run_grouped()
+        print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=20))
+
+
+if __name__ == "__main__":
+    main()

scripts/bench_moe_sweep.py

@@ -0,0 +1,311 @@
+#!/usr/bin/env python
+"""Sweep grouped_mm vs naive performance for Qwen2 MoE block."""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import sys
+import time
+import weakref
+from dataclasses import dataclass
+from pathlib import Path
+from typing import List
+
+import torch
+import torch._dynamo as dynamo
+
+try:
+    from axolotl.kernels.moe import torch_grouped as tg
+except Exception:  # pragma: no cover
+    tg = None
+
+
+def _parse_list(arg: str) -> List[int]:
+    return [int(v) for v in arg.split(",") if v]
+
+
+def _bench(run, *, iters: int, warmup: int, device: torch.device) -> float:
+    for _ in range(warmup):
+        run()
+        if device.type == "cuda":
+            torch.cuda.synchronize()
+    times: List[float] = []
+    for _ in range(iters):
+        if device.type == "cuda":
+            torch.cuda.synchronize()
+        start = time.perf_counter()
+        run()
+        if device.type == "cuda":
+            torch.cuda.synchronize()
+        times.append((time.perf_counter() - start) * 1000.0)
+    return sum(times) / len(times)
+
+
+def _estimate_flops(tokens: int, hidden: int, inter: int, top_k: int) -> float:
+    return 6.0 * tokens * top_k * hidden * inter
+
+
+def _load_block(
+    hidden: int,
+    inter: int,
+    experts: int,
+    top_k: int,
+    *,
+    device: torch.device,
+    dtype: torch.dtype,
+):
+    project_root = Path(__file__).resolve().parents[2]
+    transformers_src = project_root / "transformers" / "src"
+    if transformers_src.exists() and str(transformers_src) not in sys.path:
+        sys.path.append(str(transformers_src))
+
+    from transformers.models.qwen2_moe.configuration_qwen2_moe import Qwen2MoeConfig
+    from transformers.models.qwen2_moe.modeling_qwen2_moe import Qwen2MoeSparseMoeBlock
+
+    cfg = Qwen2MoeConfig(
+        hidden_size=hidden,
+        moe_intermediate_size=inter,
+        shared_expert_intermediate_size=inter,
+        num_experts=experts,
+        num_experts_per_tok=top_k,
+        norm_topk_prob=True,
+        qkv_bias=True,
+    )
+
+    block = Qwen2MoeSparseMoeBlock(cfg).to(device=device, dtype=dtype)
+    block_grouped = Qwen2MoeSparseMoeBlock(cfg).to(device=device, dtype=dtype)
+    block_grouped.load_state_dict(block.state_dict())
+    return block, block_grouped
+
+
+@dataclass
+class Result:
+    bsz: int
+    seq: int
+    hidden: int
+    inter: int
+    experts: int
+    top_k: int
+    dtype: str
+    naive_ms: float
+    grouped_ms: float
+    speedup: float
+    naive_tflops: float
+    grouped_tflops: float
+    max_abs: float
+    mean_abs: float
+    rel_l2: float
+
+
+def main() -> None:
+    p = argparse.ArgumentParser(description="Grouped MoE sweep")
+    p.add_argument("--batch-sizes", default="4,8,16")
+    p.add_argument("--seq-lens", default="512,1024,2048")
+    p.add_argument("--hidden", default="2048,4096")
+    p.add_argument("--inter", default="5632,8192,14336")
+    p.add_argument("--experts", default="8,16,32")
+    p.add_argument("--top-k", default="1,2,4")
+    p.add_argument("--dtype", choices=["bf16", "fp16", "fp32"], default="bf16")
+    p.add_argument("--iters", type=int, default=25)
+    p.add_argument("--warmup", type=int, default=5)
+    p.add_argument("--csv", type=Path, default=None)
+    p.add_argument("--compile", action="store_true")
+    args = p.parse_args()
+
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    dtype = {
+        "bf16": torch.bfloat16,
+        "fp16": torch.float16,
+        "fp32": torch.float32,
+    }[args.dtype]
+
+    if tg is None or not tg.available():
+        print("torch_grouped unavailable; sweep aborted")
+        return
+
+    bs_list = _parse_list(args.batch_sizes)
+    seq_list = _parse_list(args.seq_lens)
+    hidden_list = _parse_list(args.hidden)
+    inter_list = _parse_list(args.inter)
+    expert_list = _parse_list(args.experts)
+    topk_list = _parse_list(args.top_k)
+
+    results: List[Result] = []
+
+    print(
+        "bsz\tseq\thidden\tinter\texperts\ttop_k\tnaive(ms)\tgrouped(ms)\tspeedup\t"
+        "naive TF/s\tgrouped TF/s\tmax_abs\tmean_abs\trel_l2"
+    )
+
+    for bsz in bs_list:
+        for seq in seq_list:
+            tokens = bsz * seq
+            for hidden in hidden_list:
+                for inter in inter_list:
+                    for experts in expert_list:
+                        for top_k in topk_list:
+                            torch.manual_seed(0)
+                            if device.type == "cuda":
+                                torch.cuda.manual_seed(0)
+
+                            block_naive, block_grouped = _load_block(
+                                hidden,
+                                inter,
+                                experts,
+                                top_k,
+                                device=device,
+                                dtype=dtype,
+                            )
+
+                            x = torch.randn(
+                                bsz, seq, hidden, device=device, dtype=dtype
+                            )
+
+                            compiled_impl = None
+                            if args.compile:
+                                dynamo.config.capture_scalar_outputs = True
+                                dynamo.config.allow_unspec_int_on_nn_module = True
+                                try:
+                                    block_naive = torch.compile(block_naive)  # type: ignore[arg-type]
+                                except Exception as exc:
+                                    print(
+                                        f"torch.compile naive failed ({exc}); using eager"
+                                    )
+                                else:
+
+                                    def grouped_forward(inp, *, block=block_grouped):
+                                        block.experts._ax_parent_block_ref = (
+                                            weakref.ref(block)
+                                        )  # type: ignore[attr-defined]
+                                        y, _ = tg.moe_ffn_forward_grouped(
+                                            inp,
+                                            block.gate,
+                                            block.experts,
+                                            block.top_k,
+                                        )
+                                        return y
+
+                                    try:
+                                        compiled_impl = torch.compile(grouped_forward)  # type: ignore[arg-type]
+                                    except Exception as exc:
+                                        print(
+                                            f"torch.compile grouped failed ({exc}); using eager"
+                                        )
+                                        compiled_impl = None
+
+                            def run_naive(block=block_naive, data=x):
+                                y, _ = block(data)
+                                return y
+
+                            def run_grouped(
+                                block=block_grouped, data=x, impl=compiled_impl
+                            ):
+                                if impl is not None:
+                                    return impl(data)
+                                block.experts._ax_parent_block_ref = weakref.ref(block)  # type: ignore[attr-defined]
+                                y, _ = tg.moe_ffn_forward_grouped(
+                                    data,
+                                    block.gate,
+                                    block.experts,
+                                    block.top_k,
+                                )
+                                return y
+
+                            naive_ms = _bench(
+                                run_naive,
+                                iters=args.iters,
+                                warmup=args.warmup,
+                                device=device,
+                            )
+                            y_naive = run_naive()
+
+                            grouped_ms = _bench(
+                                run_grouped,
+                                iters=args.iters,
+                                warmup=args.warmup,
+                                device=device,
+                            )
+                            y_grouped = run_grouped()
+
+                            diff = (y_naive.float() - y_grouped.float()).abs()
+                            res = Result(
+                                bsz,
+                                seq,
+                                hidden,
+                                inter,
+                                experts,
+                                top_k,
+                                args.dtype,
+                                naive_ms,
+                                grouped_ms,
+                                naive_ms / grouped_ms,
+                                _estimate_flops(tokens, hidden, inter, top_k)
+                                / ((naive_ms / 1000.0) * 1e12),
+                                _estimate_flops(tokens, hidden, inter, top_k)
+                                / ((grouped_ms / 1000.0) * 1e12),
+                                diff.max().item(),
+                                diff.mean().item(),
+                                (
+                                    (
+                                        diff.pow(2).sum()
+                                        / (y_naive.float().pow(2).sum() + 1e-12)
+                                    )
+                                    .sqrt()
+                                    .item()
+                                ),
+                            )
+                            results.append(res)
+                            print(
+                                f"{bsz}\t{seq}\t{hidden}\t{inter}\t{experts}\t{top_k}\t{res.naive_ms:.2f}\t"
+                                f"{res.grouped_ms:.2f}\t{res.speedup:.2f}\t{res.naive_tflops:.2f}\t"
+                                f"{res.grouped_tflops:.2f}\t{res.max_abs:.2e}\t{res.mean_abs:.2e}\t{res.rel_l2:.2e}"
+                            )
+
+    if args.csv:
+        fieldnames = [
+            "bsz",
+            "seq",
+            "hidden",
+            "inter",
+            "experts",
+            "top_k",
+            "dtype",
+            "naive_ms",
+            "grouped_ms",
+            "speedup",
+            "naive_tflops",
+            "grouped_tflops",
+            "max_abs",
+            "mean_abs",
+            "rel_l2",
+        ]
+        with args.csv.open("w", newline="") as f:
+            writer = csv.DictWriter(f, fieldnames=fieldnames)
+            writer.writeheader()
+            for r in results:
+                writer.writerow(
+                    {
+                        "bsz": r.bsz,
+                        "seq": r.seq,
+                        "hidden": r.hidden,
+                        "inter": r.inter,
+                        "experts": r.experts,
+                        "top_k": r.top_k,
+                        "dtype": r.dtype,
+                        "naive_ms": f"{r.naive_ms:.4f}",
+                        "grouped_ms": f"{r.grouped_ms:.4f}",
+                        "speedup": f"{r.speedup:.4f}",
+                        "naive_tflops": f"{r.naive_tflops:.4f}",
+                        "grouped_tflops": f"{r.grouped_tflops:.4f}",
+                        "max_abs": f"{r.max_abs:.6e}",
+                        "mean_abs": f"{r.mean_abs:.6e}",
+                        "rel_l2": f"{r.rel_l2:.6e}",
+                    }
+                )
+
+
+if __name__ == "__main__":
+    import weakref
+
+    main()

scripts/bench_torchtitan_moe.py

@@ -0,0 +1,205 @@
+#!/usr/bin/env python
+"""Benchmark Torchtitan MoE grouped vs naive expert execution."""
+
+from __future__ import annotations
+
+import argparse
+import sys
+import time
+from pathlib import Path
+
+import torch
+
+# Ensure torchtitan is importable when running from the axolotl tree
+_PROJECT_ROOT = Path(__file__).resolve().parents[2]
+_TITAN_PATH = _PROJECT_ROOT / "torchtitan"
+if str(_TITAN_PATH) not in sys.path:
+    sys.path.append(str(_TITAN_PATH))
+
+from torchtitan.models.moe import MoE, MoEArgs
+
+
+def _parse_args() -> argparse.Namespace:
+    p = argparse.ArgumentParser(description="Torchtitan MoE microbenchmark")
+    p.add_argument("--bsz", type=int, default=8)
+    p.add_argument("--seq", type=int, default=1024)
+    p.add_argument("--hidden", type=int, default=4096)
+    p.add_argument("--inter", type=int, default=14336)
+    p.add_argument("--experts", type=int, default=8)
+    p.add_argument("--top_k", type=int, default=2)
+    p.add_argument("--dtype", choices=["bf16", "fp16", "fp32"], default="bf16")
+    p.add_argument("--iters", type=int, default=50)
+    p.add_argument("--warmup", type=int, default=10)
+    p.add_argument("--init-std", type=float, default=0.02)
+    p.add_argument(
+        "--score-before",
+        action="store_true",
+        help="Apply routing scores before expert computation (default: after)",
+    )
+    p.add_argument(
+        "--score-func",
+        choices=["softmax", "sigmoid"],
+        default="softmax",
+    )
+    p.add_argument(
+        "--route-norm",
+        action="store_true",
+        help="Enable Torchtitan router normalization when using sigmoid scores.",
+    )
+    return p.parse_args()
+
+
+def _map_dtype(arg: str) -> torch.dtype:
+    return {
+        "bf16": torch.bfloat16,
+        "fp16": torch.float16,
+        "fp32": torch.float32,
+    }[arg]
+
+
+def _estimate_moe_flops(tokens: int, hidden: int, inter: int, top_k: int) -> float:
+    # Two up projections + one down projection per expert/token combination.
+    return 6.0 * tokens * top_k * hidden * inter
+
+
+def _prepare_module(
+    moe: MoE,
+    *,
+    device: torch.device,
+    dtype: torch.dtype,
+) -> MoE:
+    moe = moe.to(device=device)
+    for param in moe.parameters():
+        param.data = param.data.to(dtype)
+        if param.grad is not None:
+            param.grad = None
+
+    buffers = dict(moe.named_buffers())
+    for name, buf in buffers.items():
+        if name == "tokens_per_expert":
+            moe._buffers[name] = torch.zeros_like(
+                buf, dtype=torch.float32, device=device
+            )
+        elif name == "expert_bias" and buf is not None:
+            moe._buffers[name] = torch.zeros_like(
+                buf, dtype=torch.float32, device=device
+            )
+        else:
+            moe._buffers[name] = buf.to(device=device, dtype=dtype)
+    moe.eval()
+    return moe
+
+
+@torch.inference_mode()
+def _forward_fn(module: MoE, x: torch.Tensor) -> torch.Tensor:
+    return module(x)
+
+
+def _bench(fn, *, iters: int, warmup: int, sync: bool = True) -> float:
+    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+    for _ in range(warmup):
+        fn()
+        if sync and device.type == "cuda":
+            torch.cuda.synchronize()
+    times = []
+    for _ in range(iters):
+        if sync and device.type == "cuda":
+            torch.cuda.synchronize()
+        start = time.perf_counter()
+        fn()
+        if sync and device.type == "cuda":
+            torch.cuda.synchronize()
+        times.append((time.perf_counter() - start) * 1000.0)
+    return sum(times) / len(times)
+
+
+def main() -> None:
+    args = _parse_args()
+    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+    dtype = _map_dtype(args.dtype)
+
+    torch.manual_seed(0)
+    if device.type == "cuda":
+        torch.cuda.manual_seed(0)
+
+    moe_args_grouped = MoEArgs(
+        num_experts=args.experts,
+        num_shared_experts=0,
+        score_func=args.score_func,
+        route_norm=args.route_norm,
+        top_k=args.top_k,
+        use_grouped_mm=True,
+        score_before_experts=args.score_before,
+        load_balance_coeff=None,
+    )
+    moe_grouped = MoE(moe_args_grouped, dim=args.hidden, hidden_dim=args.inter)
+    moe_grouped.init_weights(args.init_std, buffer_device=device)
+
+    moe_args_naive = MoEArgs(
+        num_experts=args.experts,
+        num_shared_experts=0,
+        score_func=args.score_func,
+        route_norm=args.route_norm,
+        top_k=args.top_k,
+        use_grouped_mm=False,
+        score_before_experts=args.score_before,
+        load_balance_coeff=None,
+    )
+    moe_naive = MoE(moe_args_naive, dim=args.hidden, hidden_dim=args.inter)
+    moe_naive.load_state_dict(moe_grouped.state_dict(), strict=True)
+
+    moe_grouped = _prepare_module(moe_grouped, device=device, dtype=dtype)
+    moe_naive = _prepare_module(moe_naive, device=device, dtype=dtype)
+
+    x = torch.randn(args.bsz, args.seq, args.hidden, device=device, dtype=dtype)
+
+    tokens = args.bsz * args.seq
+    print(
+        f"Device={device} dtype={dtype} tokens={tokens} hidden={args.hidden} "
+        f"inter={args.inter} experts={args.experts} top_k={args.top_k}"
+    )
+
+    def run_naive():
+        return _forward_fn(moe_naive, x)
+
+    def run_grouped():
+        return _forward_fn(moe_grouped, x)
+
+    if hasattr(moe_naive, "tokens_per_expert"):
+        moe_naive.tokens_per_expert.zero_()
+    if hasattr(moe_grouped, "tokens_per_expert"):
+        moe_grouped.tokens_per_expert.zero_()
+
+    t_naive = _bench(run_naive, iters=args.iters, warmup=args.warmup)
+    flops = _estimate_moe_flops(tokens, args.hidden, args.inter, args.top_k)
+    tflops_naive = flops / ((t_naive / 1000.0) * 1e12)
+    print(
+        f"naive\t{t_naive:.2f} ms\t{tokens / (t_naive / 1000.0):.1f} tok/s\t"
+        f"{tflops_naive:.2f} TFLOP/s"
+    )
+
+    y_naive = run_naive()
+
+    if hasattr(moe_grouped, "tokens_per_expert"):
+        moe_grouped.tokens_per_expert.zero_()
+
+    t_grouped = _bench(run_grouped, iters=args.iters, warmup=args.warmup)
+    tflops_grouped = flops / ((t_grouped / 1000.0) * 1e12)
+    speedup = t_naive / t_grouped if t_grouped > 0 else float("nan")
+    print(
+        f"grouped\t{t_grouped:.2f} ms\t{tokens / (t_grouped / 1000.0):.1f} tok/s\t"
+        f"{tflops_grouped:.2f} TFLOP/s\t{speedup:.2f}×"
+    )
+
+    y_grouped = run_grouped()
+    diff = (y_naive.float() - y_grouped.float()).abs()
+    max_abs = diff.max().item()
+    mean_abs = diff.mean().item()
+    rel_l2 = (diff.pow(2).sum() / (y_naive.float().pow(2).sum() + 1e-12)).sqrt().item()
+    print(
+        f"grouped_check: max_abs={max_abs:.3e} mean_abs={mean_abs:.3e} rel_l2={rel_l2:.3e}"
+    )
+
+
+if __name__ == "__main__":
+    main()

scripts/bench_torchtitan_moe_sweep.py

@@ -0,0 +1,328 @@
+#!/usr/bin/env python
+"""Sweep Torchtitan MoE grouped vs naive configurations and report performance."""
+
+from __future__ import annotations
+
+import argparse
+import csv
+import sys
+import time
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Iterable, List
+
+import torch
+
+_PROJECT_ROOT = Path(__file__).resolve().parents[2]
+_TITAN_PATH = _PROJECT_ROOT / "torchtitan"
+if str(_TITAN_PATH) not in sys.path:
+    sys.path.append(str(_TITAN_PATH))
+
+from torchtitan.models.moe import MoE, MoEArgs
+
+
+def _parse_int_list(value: str) -> List[int]:
+    return [int(v) for v in value.split(",") if v]
+
+
+def _parse_args() -> argparse.Namespace:
+    p = argparse.ArgumentParser(description="Torchtitan MoE grouped vs naive sweep")
+    p.add_argument(
+        "--batch-sizes", default="4,8,16", help="Comma separated batch sizes"
+    )
+    p.add_argument(
+        "--seq-lens", default="1024,2048", help="Comma separated sequence lengths"
+    )
+    p.add_argument(
+        "--experts", default="8,16,32,64", help="Comma separated expert counts"
+    )
+    p.add_argument("--top-ks", default="1,2,4", help="Comma separated top_k choices")
+    p.add_argument("--hidden", type=int, default=4096)
+    p.add_argument("--inter", type=int, default=14336)
+    p.add_argument("--dtype", choices=["bf16", "fp16", "fp32"], default="bf16")
+    p.add_argument("--iters", type=int, default=25)
+    p.add_argument("--warmup", type=int, default=5)
+    p.add_argument("--init-std", type=float, default=0.02)
+    p.add_argument("--score-before", action="store_true")
+    p.add_argument("--score-func", choices=["softmax", "sigmoid"], default="softmax")
+    p.add_argument("--route-norm", action="store_true")
+    p.add_argument("--csv", type=Path, default=None, help="Optional CSV output path")
+    return p.parse_args()
+
+
+def _map_dtype(arg: str) -> torch.dtype:
+    return {
+        "bf16": torch.bfloat16,
+        "fp16": torch.float16,
+        "fp32": torch.float32,
+    }[arg]
+
+
+def _estimate_flops(tokens: int, hidden: int, inter: int, top_k: int) -> float:
+    return 6.0 * tokens * top_k * hidden * inter
+
+
+def _prepare_module(module: MoE, *, device: torch.device, dtype: torch.dtype) -> MoE:
+    module = module.to(device=device)
+    for param in module.parameters():
+        param.data = param.data.to(dtype)
+        if param.grad is not None:
+            param.grad = None
+    for name, buf in module.named_buffers():
+        if name == "tokens_per_expert":
+            module._buffers[name] = torch.zeros_like(
+                buf, dtype=torch.float32, device=device
+            )
+        elif name == "expert_bias" and buf is not None:
+            module._buffers[name] = torch.zeros_like(
+                buf, dtype=torch.float32, device=device
+            )
+        else:
+            module._buffers[name] = buf.to(device=device, dtype=dtype)
+    module.eval()
+    return module
+
+
+@torch.inference_mode()
+def _forward(module: MoE, x: torch.Tensor) -> torch.Tensor:
+    return module(x)
+
+
+def _bench(callable_, *, iters: int, warmup: int, device: torch.device) -> float:
+    for _ in range(warmup):
+        callable_()
+        if device.type == "cuda":
+            torch.cuda.synchronize()
+    timings: List[float] = []
+    for _ in range(iters):
+        if device.type == "cuda":
+            torch.cuda.synchronize()
+        start = time.perf_counter()
+        callable_()
+        if device.type == "cuda":
+            torch.cuda.synchronize()
+        timings.append((time.perf_counter() - start) * 1000.0)
+    return sum(timings) / len(timings)
+
+
+@dataclass
+class SweepResult:
+    bsz: int
+    seq: int
+    experts: int
+    top_k: int
+    dtype: str
+    naive_ms: float
+    grouped_ms: float
+    speedup: float
+    naive_tflops: float
+    grouped_tflops: float
+    max_abs: float
+    mean_abs: float
+    rel_l2: float
+
+
+def _run_case(
+    *,
+    bsz: int,
+    seq: int,
+    experts: int,
+    top_k: int,
+    hidden: int,
+    inter: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    iters: int,
+    warmup: int,
+    init_std: float,
+    score_before: bool,
+    score_func: str,
+    route_norm: bool,
+) -> SweepResult:
+    torch.manual_seed(0)
+    if device.type == "cuda":
+        torch.cuda.manual_seed(0)
+
+    moe_args_grouped = MoEArgs(
+        num_experts=experts,
+        num_shared_experts=0,
+        score_func=score_func,
+        route_norm=route_norm,
+        top_k=top_k,
+        use_grouped_mm=True,
+        score_before_experts=score_before,
+        load_balance_coeff=None,
+    )
+    moe_grouped = MoE(moe_args_grouped, dim=hidden, hidden_dim=inter)
+    moe_grouped.init_weights(init_std, buffer_device=device)
+
+    moe_args_naive = MoEArgs(
+        num_experts=experts,
+        num_shared_experts=0,
+        score_func=score_func,
+        route_norm=route_norm,
+        top_k=top_k,
+        use_grouped_mm=False,
+        score_before_experts=score_before,
+        load_balance_coeff=None,
+    )
+    moe_naive = MoE(moe_args_naive, dim=hidden, hidden_dim=inter)
+    moe_naive.load_state_dict(moe_grouped.state_dict(), strict=True)
+
+    moe_grouped = _prepare_module(moe_grouped, device=device, dtype=dtype)
+    moe_naive = _prepare_module(moe_naive, device=device, dtype=dtype)
+
+    x = torch.randn(bsz, seq, hidden, device=device, dtype=dtype)
+
+    def run_naive():
+        if hasattr(moe_naive, "tokens_per_expert"):
+            moe_naive.tokens_per_expert.zero_()
+        return _forward(moe_naive, x)
+
+    def run_grouped():
+        if hasattr(moe_grouped, "tokens_per_expert"):
+            moe_grouped.tokens_per_expert.zero_()
+        return _forward(moe_grouped, x)
+
+    naive_ms = _bench(run_naive, iters=iters, warmup=warmup, device=device)
+    y_naive = run_naive()
+
+    grouped_ms = _bench(run_grouped, iters=iters, warmup=warmup, device=device)
+    y_grouped = run_grouped()
+
+    diff = (y_naive.float() - y_grouped.float()).abs()
+    max_abs = diff.max().item()
+    mean_abs = diff.mean().item()
+    rel_l2 = (diff.pow(2).sum() / (y_naive.float().pow(2).sum() + 1e-12)).sqrt().item()
+
+    tokens = bsz * seq
+    flops = _estimate_flops(tokens, hidden, inter, top_k)
+    naive_tflops = flops / ((naive_ms / 1000.0) * 1e12)
+    grouped_tflops = flops / ((grouped_ms / 1000.0) * 1e12)
+    speedup = naive_ms / grouped_ms if grouped_ms > 0 else float("nan")
+
+    return SweepResult(
+        bsz=bsz,
+        seq=seq,
+        experts=experts,
+        top_k=top_k,
+        dtype=str(dtype),
+        naive_ms=naive_ms,
+        grouped_ms=grouped_ms,
+        speedup=speedup,
+        naive_tflops=naive_tflops,
+        grouped_tflops=grouped_tflops,
+        max_abs=max_abs,
+        mean_abs=mean_abs,
+        rel_l2=rel_l2,
+    )
+
+
+def _print_header(
+    hidden: int, inter: int, dtype: torch.dtype, device: torch.device
+) -> None:
+    print(f"Device={device} dtype={dtype} hidden={hidden} inter={inter}")
+    print(
+        "bsz\tseq\texperts\ttop_k\tnaive(ms)\tgrouped(ms)\tspeedup\t"
+        "naive TF/s\tgrouped TF/s\tmax_abs\tmean_abs\trel_l2"
+    )
+
+
+def _print_result(res: SweepResult) -> None:
+    print(
+        f"{res.bsz}\t{res.seq}\t{res.experts}\t{res.top_k}\t"
+        f"{res.naive_ms:.2f}\t{res.grouped_ms:.2f}\t{res.speedup:.2f}\t"
+        f"{res.naive_tflops:.2f}\t{res.grouped_tflops:.2f}\t"
+        f"{res.max_abs:.2e}\t{res.mean_abs:.2e}\t{res.rel_l2:.2e}"
+    )
+
+
+def _write_csv(path: Path, results: Iterable[SweepResult]) -> None:
+    fieldnames = [
+        "batch_size",
+        "seq_len",
+        "experts",
+        "top_k",
+        "dtype",
+        "naive_ms",
+        "grouped_ms",
+        "speedup",
+        "naive_tflops",
+        "grouped_tflops",
+        "max_abs",
+        "mean_abs",
+        "rel_l2",
+    ]
+    with path.open("w", newline="") as f:
+        writer = csv.DictWriter(f, fieldnames=fieldnames)
+        writer.writeheader()
+        for r in results:
+            writer.writerow(
+                {
+                    "batch_size": r.bsz,
+                    "seq_len": r.seq,
+                    "experts": r.experts,
+                    "top_k": r.top_k,
+                    "dtype": r.dtype,
+                    "naive_ms": f"{r.naive_ms:.4f}",
+                    "grouped_ms": f"{r.grouped_ms:.4f}",
+                    "speedup": f"{r.speedup:.4f}",
+                    "naive_tflops": f"{r.naive_tflops:.4f}",
+                    "grouped_tflops": f"{r.grouped_tflops:.4f}",
+                    "max_abs": f"{r.max_abs:.6e}",
+                    "mean_abs": f"{r.mean_abs:.6e}",
+                    "rel_l2": f"{r.rel_l2:.6e}",
+                }
+            )
+
+
+def main() -> None:
+    args = _parse_args()
+    dtype = _map_dtype(args.dtype)
+    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
+
+    batch_sizes = _parse_int_list(args.batch_sizes)
+    seq_lens = _parse_int_list(args.seq_lens)
+    experts_list = _parse_int_list(args.experts)
+    top_ks = _parse_int_list(args.top_ks)
+
+    results: List[SweepResult] = []
+    _print_header(args.hidden, args.inter, dtype, device)
+
+    for bsz in batch_sizes:
+        for seq in seq_lens:
+            for experts in experts_list:
+                for top_k in top_ks:
+                    try:
+                        res = _run_case(
+                            bsz=bsz,
+                            seq=seq,
+                            experts=experts,
+                            top_k=top_k,
+                            hidden=args.hidden,
+                            inter=args.inter,
+                            dtype=dtype,
+                            device=device,
+                            iters=args.iters,
+                            warmup=args.warmup,
+                            init_std=args.init_std,
+                            score_before=args.score_before,
+                            score_func=args.score_func,
+                            route_norm=args.route_norm,
+                        )
+                    except RuntimeError as err:
+                        print(
+                            f"{bsz}\t{seq}\t{experts}\t{top_k}\tERROR: {err}",
+                            file=sys.stderr,
+                        )
+                        continue
+                    results.append(res)
+                    _print_result(res)
+
+    if args.csv and results:
+        _write_csv(args.csv, results)
+        print(f"Wrote {len(results)} rows to {args.csv}")
+
+
+if __name__ == "__main__":
+    main()

scripts/cutcrossentropy_install.py

@@ -1,24 +1,33 @@
-"""Print the pip command to install Axolotl's cut_cross_entropy fork."""
-
-from __future__ import annotations
+"""Script to output the correct installation command for cut-cross-entropy."""
 
+import importlib.util
 import sys
-from shlex import quote
 
 try:
     import torch
-except ImportError as exc:  # pragma: no cover
+except ImportError as exc:
     raise ImportError("Install torch via `pip install torch`") from exc
-
 from packaging.version import Version as V
 
-if V(torch.__version__.split("+")[0]) < V("2.6.0"):
+USE_UV = "--uv" in sys.argv[1:]
+
+v = V(torch.__version__)
+
+# no cut-cross-entropy support for torch < 2.4.0
+if v < V("2.4.0"):
     print("")
     sys.exit(0)
 
-python_exe = quote(sys.executable)
+cce_spec = importlib.util.find_spec("cut_cross_entropy")
+
+UNINSTALL_PREFIX = ""
+if cce_spec:
+    if not importlib.util.find_spec("cut_cross_entropy.transformers"):
+        UNINSTALL_PREFIX = "pip uninstall -y cut-cross-entropy && "
+
+UV_PREFIX = "uv " if USE_UV else ""
+
 print(
-    f"{python_exe} -m pip install "
-    '"cut-cross-entropy[transformers] '
-    '@ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@147ea28"'
+    UNINSTALL_PREFIX
+    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@c6a32c5"'
 )

scripts/debug_qwen2_experts.py

@@ -0,0 +1,53 @@
+#!/usr/bin/env python
+"""Inspect Qwen2 MoE expert implementations for grouped-mm debugging."""
+
+from __future__ import annotations
+
+import sys
+from pathlib import Path
+
+import torch
+
+ROOT = Path(__file__).resolve().parents[2]
+sys.path.extend(
+    [
+        str(ROOT / "transformers" / "src"),
+        str(ROOT / "src"),
+    ]
+)
+
+from transformers.models.qwen2_moe.configuration_qwen2_moe import Qwen2MoeConfig
+from transformers.models.qwen2_moe.modeling_qwen2_moe import Qwen2MoeSparseMoeBlock
+
+from axolotl.kernels.moe.torch_grouped import _iter_expert_impls
+
+
+def main() -> None:
+    cfg = Qwen2MoeConfig(
+        hidden_size=4096,
+        moe_intermediate_size=14336,
+        shared_expert_intermediate_size=14336,
+        num_experts=32,
+        num_experts_per_tok=4,
+    )
+
+    block = Qwen2MoeSparseMoeBlock(cfg).to("cuda", dtype=torch.bfloat16)
+    experts = block.experts
+    experts._ax_parent_block = block
+
+    impls = _iter_expert_impls(experts)
+    print(f"impl count: {len(impls)}")
+    for idx, impl in enumerate(impls[:8]):
+        has_gate = hasattr(impl, "gate_proj")
+        has_up = hasattr(impl, "up_proj")
+        print(
+            f"impl[{idx}] type={impl.__class__.__name__} has_gate={has_gate} has_up={has_up}"
+        )
+        if has_gate:
+            print(f"  gate shape {tuple(impl.gate_proj.weight.shape)}")
+            print(f"  up shape   {tuple(impl.up_proj.weight.shape)}")
+            print(f"  down shape {tuple(impl.down_proj.weight.shape)}")
+
+
+if __name__ == "__main__":
+    main()

scripts/probe_torch_grouped_ops.py

@@ -0,0 +1,47 @@
+#!/usr/bin/env python
+"""
+Probe PyTorch for grouped GEMM operator names and namespaces.
+Run: python scripts/probe_torch_grouped_ops.py
+"""
+
+import sys
+
+
+def main():
+    try:
+        import torch
+    except Exception as e:
+        print("Failed to import torch:", e)
+        sys.exit(1)
+
+    print("torch version:", torch.__version__)
+    namespaces = [n for n in dir(torch.ops) if not n.startswith("_")]
+    print("ops namespaces:", namespaces)
+
+    found_any = False
+    for ns in namespaces:
+        obj = getattr(torch.ops, ns, None)
+        ops = []
+        if obj is not None:
+            try:
+                ops = dir(obj)
+            except Exception as e:
+                print(f"warning: failed to list ops for namespace {ns}: {e}")
+        cands = [
+            o
+            for o in ops
+            if ("group" in o.lower())
+            or ("mm_grouped" in o.lower())
+            or ("matmul_grouped" in o.lower())
+            or ("grouped" in o.lower())
+        ]
+        if cands:
+            found_any = True
+            print(f"namespace {ns} candidates:", cands)
+
+    if not found_any:
+        print("No grouped GEMM candidates found. PyTorch >= 2.8 is recommended.")
+
+
+if __name__ == "__main__":
+    main()

scripts/unsloth_install.py

@@ -1,48 +1,40 @@
-"""Emit the install commands for Unsloth without altering torch."""
-
-from __future__ import annotations
-
-import shutil
+# noqa
 import sys
-from shlex import quote
 
 try:
     import torch
-except ImportError as exc:  # pragma: no cover
-    raise ImportError("Install torch via `pip install torch`") from exc
-
+except ImportError as error:
+    raise ImportError("Install torch via `pip install torch`") from error
 from packaging.version import Version as V
 
-MIN_TORCH = V("2.6.0")
+use_uv = "--uv" in sys.argv[1:]
 
-if V(torch.__version__.split("+")[0]) < MIN_TORCH:
-    raise RuntimeError(
-        f"Torch {torch.__version__} detected, but Unsloth requires >= {MIN_TORCH}."
-    )
-
-USE_UV_FLAG = "--uv" in sys.argv[1:]
-USE_PIP_FLAG = "--pip" in sys.argv[1:]
-
-if USE_UV_FLAG and USE_PIP_FLAG:
-    raise SystemExit("Specify only one of --uv or --pip")
-
-if USE_PIP_FLAG:
-    use_uv = False
-elif USE_UV_FLAG:
-    use_uv = True
+v = V(torch.__version__)
+cuda = str(torch.version.cuda)
+try:
+    is_ampere = torch.cuda.get_device_capability()[0] >= 8
+except RuntimeError:
+    is_ampere = False
+if cuda != "12.1" and cuda != "11.8" and cuda != "12.4":
+    raise RuntimeError(f"CUDA = {cuda} not supported!")
+if v <= V("2.1.0"):
+    raise RuntimeError(f"Torch = {v} too old!")
+elif v <= V("2.1.1"):
+    x = "cu{}{}-torch211"
+elif v <= V("2.1.2"):
+    x = "cu{}{}-torch212"
+elif v < V("2.3.0"):
+    x = "cu{}{}-torch220"
+elif v < V("2.4.0"):
+    x = "cu{}{}-torch230"
+elif v < V("2.5.0"):
+    x = "cu{}{}-torch240"
+elif v < V("2.6.0"):
+    x = "cu{}{}-torch250"
 else:
-    use_uv = shutil.which("uv") is not None
-
-python_exe = quote(sys.executable or shutil.which("python3") or "python")
-
-if use_uv:
-    installer = "uv pip install --system --no-deps"
-else:
-    installer = f"{python_exe} -m pip install --no-deps"
-
-commands = [
-    f"{installer} unsloth-zoo==2025.9.12",
-    f'{installer} "unsloth[huggingface]==2025.9.9"',
-]
-
-print(" && ".join(commands))
+    raise RuntimeError(f"Torch = {v} too new!")
+x = x.format(cuda.replace(".", ""), "-ampere" if is_ampere else "")
+uv_prefix = "uv " if use_uv else ""
+print(
+    f'{uv_prefix}pip install unsloth-zoo==2024.12.1 && {uv_prefix}pip install --no-deps "unsloth[{x}]==2024.12.4"'
+)

setup.py

@@ -0,0 +1,183 @@
+"""setup.py for axolotl"""
+
+import ast
+import os
+import platform
+import re
+from importlib.metadata import PackageNotFoundError, version
+from pathlib import Path
+
+from setuptools import find_packages, setup
+
+
+def parse_requirements(extras_require_map):
+    _install_requires = []
+    _dependency_links = []
+    with open("./requirements.txt", encoding="utf-8") as requirements_file:
+        lines = [r.strip() for r in requirements_file.readlines()]
+        for line in lines:
+            is_extras = "deepspeed" in line or "mamba-ssm" in line
+            if line.startswith("--extra-index-url"):
+                # Handle custom index URLs
+                _, url = line.split()
+                _dependency_links.append(url)
+            elif not is_extras and line and line[0] != "#":
+                # Handle standard packages
+                _install_requires.append(line)
+    try:
+        xformers_version = [req for req in _install_requires if "xformers" in req][0]
+        autoawq_version = [req for req in _install_requires if "autoawq" in req][0]
+        if "Darwin" in platform.system():
+            # skip packages not compatible with OSX
+            skip_packages = [
+                "bitsandbytes",
+                "triton",
+                "mamba-ssm",
+                "xformers",
+                "autoawq",
+                "liger-kernel",
+            ]
+            _install_requires = [
+                req
+                for req in _install_requires
+                if re.split(r"[>=<]", req)[0].strip() not in skip_packages
+            ]
+            print(
+                _install_requires, [req in skip_packages for req in _install_requires]
+            )
+        else:
+            # detect the version of torch already installed
+            # and set it so dependencies don't clobber the torch version
+            try:
+                torch_version = version("torch")
+            except PackageNotFoundError:
+                torch_version = "2.6.0"  # default to torch 2.6
+            _install_requires.append(f"torch=={torch_version}")
+
+            version_match = re.match(r"^(\d+)\.(\d+)(?:\.(\d+))?", torch_version)
+            if version_match:
+                major, minor, patch = version_match.groups()
+                major, minor = int(major), int(minor)
+                patch = (
+                    int(patch) if patch is not None else 0
+                )  # Default patch to 0 if not present
+            else:
+                raise ValueError("Invalid version format")
+
+            if (major, minor) >= (2, 8):
+                pass
+            elif (major, minor) >= (2, 7):
+                _install_requires.pop(_install_requires.index(xformers_version))
+                if patch == 0:
+                    _install_requires.append("xformers==0.0.30")
+                    # vllm 0.9.x is incompatible with latest transformers
+                    extras_require_map.pop("vllm")
+                else:
+                    _install_requires.append("xformers==0.0.31")
+                    extras_require_map["vllm"] = ["vllm>=0.10.0"]
+            elif (major, minor) >= (2, 6):
+                _install_requires.pop(_install_requires.index(xformers_version))
+                _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.pop("vllm")
+            elif (major, minor) >= (2, 5):
+                _install_requires.pop(_install_requires.index(xformers_version))
+                if patch == 0:
+                    _install_requires.append("xformers==0.0.28.post2")
+                else:
+                    _install_requires.append("xformers>=0.0.28.post3")
+                _install_requires.pop(_install_requires.index(autoawq_version))
+                extras_require_map.pop("vllm")
+            elif (major, minor) >= (2, 4):
+                extras_require_map.pop("vllm")
+                if patch == 0:
+                    _install_requires.pop(_install_requires.index(xformers_version))
+                    _install_requires.append("xformers>=0.0.27")
+                else:
+                    _install_requires.pop(_install_requires.index(xformers_version))
+                    _install_requires.append("xformers==0.0.28.post1")
+            else:
+                raise ValueError("axolotl requires torch>=2.4")
+
+    except PackageNotFoundError:
+        pass
+    return _install_requires, _dependency_links, extras_require_map
+
+
+def get_package_version():
+    with open(
+        Path(os.path.dirname(os.path.abspath(__file__)))
+        / "src"
+        / "axolotl"
+        / "__init__.py",
+        "r",
+        encoding="utf-8",
+    ) as fin:
+        version_match = re.search(r"^__version__\s*=\s*(.*)$", fin.read(), re.MULTILINE)
+    version_ = ast.literal_eval(version_match.group(1))
+    return version_
+
+
+extras_require = {
+    "flash-attn": ["flash-attn==2.8.3"],
+    "ring-flash-attn": [
+        "flash-attn==2.8.3",
+        "ring-flash-attn>=0.1.7",
+        "yunchang==0.6.0",
+    ],
+    "deepspeed": [
+        "deepspeed==0.17.5",
+        "deepspeed-kernels",
+    ],
+    "mamba-ssm": [
+        "mamba-ssm==1.2.0.post1",
+        "causal_conv1d",
+    ],
+    "auto-gptq": [
+        "auto-gptq==0.5.1",
+    ],
+    "mlflow": [
+        "mlflow",
+    ],
+    "galore": [
+        "galore_torch",
+    ],
+    "apollo": [
+        "apollo-torch",
+    ],
+    "optimizers": [
+        "galore_torch",
+        "apollo-torch",
+        "lomo-optim==0.1.1",
+        "torch-optimi==0.2.1",
+        "came_pytorch==0.1.3",
+    ],
+    "ray": [
+        "ray[train]",
+    ],
+    "vllm": [
+        "vllm==0.10.0",
+    ],
+    "llmcompressor": [
+        "llmcompressor==0.5.1",
+    ],
+    "fbgemm-gpu": ["fbgemm-gpu-genai>=1.2.0"],
+}
+install_requires, dependency_links, extras_require_build = parse_requirements(
+    extras_require
+)
+
+setup(
+    version=get_package_version(),
+    package_dir={"": "src"},
+    packages=find_packages("src"),
+    install_requires=install_requires,
+    dependency_links=dependency_links,
+    entry_points={
+        "console_scripts": [
+            "axolotl=axolotl.cli.main:main",
+        ],
+    },
+    extras_require=extras_require_build,
+)

src/axolotl/__init__.py

@@ -1,17 +1,7 @@
-"""Axolotl - Train and fine-tune large language models."""
-
-from __future__ import annotations
+"""Axolotl - Train and fine-tune large language models"""
 
 import pkgutil
-from importlib import metadata
 
-try:
-    from ._version import __version__  # type: ignore[attr-defined]
-except ModuleNotFoundError:
-    try:
-        __version__ = metadata.version("axolotl")
-    except metadata.PackageNotFoundError:  # pragma: no cover
-        __version__ = "0+unknown"
+__path__ = pkgutil.extend_path(__path__, __name__)  # Make this a namespace package
 
-__path__ = pkgutil.extend_path(__path__, __name__)
-__all__ = ["__version__"]
+__version__ = "0.13.0.dev"

src/axolotl/cli/delinearize_llama4.py

@@ -85,7 +85,9 @@ def do_cli(model: Union[Path, str], output: Union[Path, str]) -> None:
     unpatch_llama4 = patch_llama4_linearized_modeling()
     from transformers import Llama4ForConditionalGeneration
 
-    model_ = Llama4ForConditionalGeneration.from_pretrained(model, dtype=torch.bfloat16)
+    model_ = Llama4ForConditionalGeneration.from_pretrained(
+        model, torch_dtype=torch.bfloat16
+    )
     processor = AutoProcessor.from_pretrained(model)
     processor.save_pretrained(output)
 

src/axolotl/cli/quantize.py

@@ -69,7 +69,7 @@ def do_quantize(
     config = AutoConfig.from_pretrained(model_path)
     torch_dtype = config.torch_dtype if hasattr(config, "torch_dtype") else None
     model = AutoModelForCausalLM.from_pretrained(
-        model_path, device_map="auto", dtype=torch_dtype
+        model_path, device_map="auto", torch_dtype=torch_dtype
     )
 
     LOG.info(