draft config for devstral

* models.py -> loaders/ module refactor

* refactor ModelLoader class

* plugin manager changes

* circular import fix

* pytest

* pytest

* minor improvements

* fix

* minor changes

* fix test

* remove dead code

* coderabbit comments

* lint

* fix

* coderabbit suggestion I liked

* more coderabbit

* review comments, yak shaving

* lint

* updating in light of SP ctx manager changes

* review comment

* review comment 2

.github/workflows/base.yml

@@ -22,12 +22,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: "124"
-            cuda_version: 12.4.1
-            cudnn_version: ""
-            python_version: "3.11"
-            pytorch: 2.4.1
-            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
           - cuda: "124"
             cuda_version: 12.4.1
             cudnn_version: ""

.github/workflows/main.yml

@@ -15,16 +15,11 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.4.1
-            axolotl_extras:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
             pytorch: 2.5.1
-            axolotl_extras: vllm
+            axolotl_extras:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
@@ -35,7 +30,12 @@ jobs:
             cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.7.0
-            axolotl_extras: vllm
+            axolotl_extras:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.7.0
+            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -67,6 +67,7 @@ jobs:
             CUDA=${{ matrix.cuda }}
             PYTORCH_VERSION=${{ matrix.pytorch }}
             AXOLOTL_ARGS=${{ matrix.axolotl_args }}
+            AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}
           file: ./docker/Dockerfile
           push: ${{ github.event_name != 'pull_request' }}
           tags: |
@@ -82,11 +83,6 @@ jobs:
     strategy:
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.4.1
-            axolotl_extras:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
@@ -103,6 +99,11 @@ jobs:
             python_version: "3.11"
             pytorch: 2.7.0
             axolotl_extras:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.7.0
+            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout

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

@@ -3,11 +3,13 @@ name: docker-multigpu-tests-biweekly
 on:
   pull_request:
     paths:
-      - 'tests/e2e/multigpu/*.py'
+      - 'tests/e2e/multigpu/**.py'
       - 'requirements.txt'
       - 'setup.py'
       - 'pyproject.toml'
       - '.github/workflows/multi-gpu-e2e.yml'
+      - 'src/axolotl/core/trainers/mixins/sequence_parallel.py'
+      - 'src/axolotl/utils/distributed.py'
   workflow_dispatch:
   schedule:
     - cron: '0 0 * * 1,4'  # Runs at 00:00 UTC every monday & thursday
@@ -31,18 +33,11 @@ jobs:
             axolotl_extras: vllm
             num_gpus: 2
             nightly_build: "true"
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.4.1
-            axolotl_extras:  # no vllm support for 2.4.1
-            num_gpus: 2
-            nightly_build: "true"
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
             pytorch: 2.5.1
-            axolotl_extras: vllm
+            axolotl_extras:
             num_gpus: 2
             nightly_build: "true"
           - cuda: 126

.github/workflows/nightlies.yml

@@ -12,11 +12,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.4.1
-            axolotl_extras:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
@@ -70,11 +65,6 @@ jobs:
     strategy:
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.4.1
-            axolotl_extras:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"

.github/workflows/preview-docs.yml

@@ -0,0 +1,61 @@
+name: Preview
+on:
+  workflow_dispatch:
+  pull_request:
+    types: [opened, synchronize, reopened]
+
+    # Run the workflow only when one of these files changes
+    paths:
+      - '**/*.md'      # any Markdown file
+      - '**/*.qmd'     # any Quarto file
+      - '_quarto.yaml'
+
+permissions:
+  checks: write
+  contents: write
+  deployments: write
+  issues: write
+  discussions: write
+  pages: write
+  pull-requests: write
+  statuses: write
+
+jobs:
+  preview:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Check out repository
+        uses: actions/checkout@v4
+
+      - name: Set up Quarto
+        uses: quarto-dev/quarto-actions/setup@v2
+
+      - name: Setup Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+
+      - name: Install dependencies
+        run: |
+          python3 -m pip install jupyter quartodoc
+          python3 -m pip install -e . --no-deps
+
+      - name: Build autodoc
+        run: quartodoc build
+
+      - name: Quarto render
+        run: quarto render
+
+      - name: Netlify Publish
+        uses: nwtgck/actions-netlify@v3.0
+        with:
+          publish-dir: './_site'
+          enable-pull-request-comment: true
+          enable-github-deployment: true
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          deploy-message: "Deployed On Netlify"
+          github-deployment-environment: 'preview'
+          github-deployment-description: 'Preview Deployment'
+        env:
+          NETLIFY_AUTH_TOKEN: ${{ secrets.NETLIFY_AUTH_TOKEN }}
+          NETLIFY_SITE_ID: ${{ secrets.NETLIFY_SITE_ID }}

.github/workflows/tests-nightly.yml

@@ -18,15 +18,102 @@ jobs:
         env:
           SKIP: no-commit-to-branch
 
+  preload-cache:
+    name: Preload HF cache
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        python_version: ["3.11"]
+        pytorch_version: ["2.6.0"]
+    timeout-minutes: 20
+
+    env:
+      AXOLOTL_IS_CI_CACHE_PRELOAD: "1"
+
+    steps:
+      - name: Check out repository code
+        uses: actions/checkout@v4
+
+      - name: Restore HF cache
+        id: hf-cache-restore
+        uses: actions/cache/restore@v4
+        with:
+          path: |
+            /home/runner/.cache/huggingface/hub/datasets--*
+            /home/runner/.cache/huggingface/hub/models--*
+          key: ${{ runner.os }}-hf-hub-cache-v2
+
+      - name: Setup Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python_version }}
+          cache: 'pip' # caching pip dependencies
+
+      - name: upgrade pip
+        run: |
+          pip3 install --upgrade pip
+          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
+
+      - name: Install PyTorch
+        run: |
+          pip3 install torch==${{ matrix.pytorch_version }}
+
+      - name: Install dependencies
+        run: |
+          pip3 show torch
+          pip3 install --no-build-isolation -U -e .
+          python scripts/unsloth_install.py | sh
+          python scripts/cutcrossentropy_install.py | sh
+          pip3 install -r requirements-dev.txt -r requirements-tests.txt
+
+      - name: Make sure PyTorch version wasn't clobbered
+        run: |
+          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
+
+      - name: Ensure axolotl CLI was installed
+        run: |
+          axolotl --help
+
+      - name: Pre-Download dataset fixture
+        run: |
+          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
+
+      - name: Run tests
+        run: |
+          pytest -v tests/conftest.py
+
+      - name: Upload coverage to Codecov
+        uses: codecov/codecov-action@v5
+        with:
+          token: ${{ secrets.CODECOV_TOKEN }}
+          files: ./coverage.xml
+          flags: unittests,pytorch-${{ matrix.pytorch_version }}
+          fail_ci_if_error: false
+
+      - name: cleanup pip cache
+        run: |
+          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
+
+      - name: Save HF cache
+        id: hf-cache
+        uses: actions/cache/save@v4
+        with:
+          path: |
+            /home/runner/.cache/huggingface/hub/datasets--*
+            /home/runner/.cache/huggingface/hub/models--*
+          key: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
+
   pytest:
     name: PyTest
     runs-on: ubuntu-latest
+    needs: [preload-cache]
     strategy:
       fail-fast: false
       max-parallel: 2
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.4.1", "2.5.1", "2.6.0"]
+        pytorch_version: ["2.5.1", "2.6.0", "2.7.0"]
     timeout-minutes: 20
 
     steps:
@@ -106,13 +193,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.4.1
-            num_gpus: 1
-            axolotl_extras:
-            nightly_build: "true"
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"

.github/workflows/tests.yml

@@ -27,6 +27,9 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
   cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
 
+env:
+  TRANSFORMERS_IS_CI: "yes"
+
 jobs:
   pre-commit:
     name: pre-commit
@@ -41,29 +44,127 @@ jobs:
         env:
           SKIP: no-commit-to-branch
 
+#  preload-cache:
+#    name: Preload HF cache
+#    runs-on: ubuntu-latest
+#    strategy:
+#      fail-fast: false
+#      matrix:
+#        python_version: ["3.11"]
+#        pytorch_version: ["2.6.0"]
+#    timeout-minutes: 20
+#
+#    env:
+#      AXOLOTL_IS_CI_CACHE_PRELOAD: "1"
+#
+#    steps:
+#      - name: Check out repository code
+#        uses: actions/checkout@v4
+#
+#      - name: Restore HF cache
+#        id: hf-cache-restore
+#        uses: actions/cache/restore@v4
+#        with:
+#          path: |
+#            /home/runner/.cache/huggingface/hub/datasets--*
+#            /home/runner/.cache/huggingface/hub/models--*
+#          key: ${{ runner.os }}-hf-hub-cache-v2
+#
+#      - name: Restore Cache from S3
+#        id: hf-cache-restore-s3
+#        run: |
+#          mkdir -p /home/runner/.cache/huggingface/hub
+#          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C /home/runner/.cache/huggingface/hub/  --use-compress-program unzstd
+#
+#      - name: Setup Python
+#        uses: actions/setup-python@v5
+#        with:
+#          python-version: ${{ matrix.python_version }}
+#          cache: 'pip' # caching pip dependencies
+#
+#      - name: upgrade pip
+#        run: |
+#          pip3 install --upgrade pip
+#          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
+#
+#      - name: Install PyTorch
+#        run: |
+#          pip3 install torch==${{ matrix.pytorch_version }}
+#
+#      - name: Install dependencies
+#        run: |
+#          pip3 show torch
+#          pip3 install --no-build-isolation -U -e .
+#          python scripts/unsloth_install.py | sh
+#          python scripts/cutcrossentropy_install.py | sh
+#          pip3 install -r requirements-dev.txt -r requirements-tests.txt
+#
+#      - name: Make sure PyTorch version wasn't clobbered
+#        run: |
+#          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
+#
+#      - name: Ensure axolotl CLI was installed
+#        run: |
+#          axolotl --help
+#
+#      - name: Pre-Download dataset fixture
+#        run: |
+#          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
+#
+#      - name: Run tests
+#        run: |
+#          pytest -v tests/conftest.py
+#
+#      - name: Upload coverage to Codecov
+#        uses: codecov/codecov-action@v5
+#        with:
+#          token: ${{ secrets.CODECOV_TOKEN }}
+#          files: ./coverage.xml
+#          flags: unittests,pytorch-${{ matrix.pytorch_version }}
+#          fail_ci_if_error: false
+#
+#      - name: cleanup pip cache
+#        run: |
+#          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
+#
+#      - name: Save HF cache
+#        id: hf-cache
+#        uses: actions/cache/save@v4
+#        with:
+#          path: |
+#            /home/runner/.cache/huggingface/hub/datasets--*
+#            /home/runner/.cache/huggingface/hub/models--*
+#          key: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
+
   pytest:
     name: PyTest
     runs-on: ubuntu-latest
+#    needs: [preload-cache]
     strategy:
       fail-fast: false
-      max-parallel: 2
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.4.1", "2.5.1", "2.6.0", "2.7.0"]
+        pytorch_version: ["2.5.1", "2.6.0", "2.7.0"]
     timeout-minutes: 20
 
     steps:
       - name: Check out repository code
         uses: actions/checkout@v4
 
-      - name: Restore HF cache
-        id: hf-cache-restore
-        uses: actions/cache/restore@v4
-        with:
-          path: |
-            /home/runner/.cache/huggingface/hub/datasets--*
-            /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ runner.os }}-hf-hub-cache-v2
+#      - name: Restore HF cache
+#        id: hf-cache-restore
+#        uses: actions/cache/restore@v4
+#        with:
+#          path: |
+#            /home/runner/.cache/huggingface/hub/datasets--*
+#            /home/runner/.cache/huggingface/hub/models--*
+#          key: ${{ runner.os }}-hf-hub-cache-v2
+
+      - name: Restore Cache from S3
+        id: hf-cache-restore-s3
+        run: |
+          mkdir -p /home/runner/.cache/huggingface/hub
+          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C /home/runner/.cache/huggingface/hub/  --use-compress-program unzstd
 
       - name: Setup Python
         uses: actions/setup-python@v5
@@ -118,38 +219,35 @@ jobs:
         run: |
           find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
 
-      - name: Save HF cache
-        id: hf-cache
-        uses: actions/cache/save@v4
-        with:
-          path: |
-            /home/runner/.cache/huggingface/hub/datasets--*
-            /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
-
   pytest-sdist:
     name: PyTest from Source Dist
     runs-on: ubuntu-latest
+#    needs: [preload-cache]
     strategy:
       fail-fast: false
-      max-parallel: 1
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.4.1", "2.5.1", "2.6.0"]
+        pytorch_version: ["2.5.1", "2.6.0", "2.7.0"]
     timeout-minutes: 20
 
     steps:
       - name: Check out repository code
         uses: actions/checkout@v4
 
-      - name: Restore HF cache
-        id: hf-cache-restore
-        uses: actions/cache/restore@v4
-        with:
-          path: |
-            /home/runner/.cache/huggingface/hub/datasets--*
-            /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ runner.os }}-hf-hub-cache-v2
+#      - name: Restore HF cache
+#        id: hf-cache-restore
+#        uses: actions/cache/restore@v4
+#        with:
+#          path: |
+#            /home/runner/.cache/huggingface/hub/datasets--*
+#            /home/runner/.cache/huggingface/hub/models--*
+#          key: ${{ runner.os }}-hf-hub-cache-v2
+
+      - name: Restore Cache from S3
+        id: hf-cache-restore-s3
+        run: |
+          mkdir -p /home/runner/.cache/huggingface/hub
+          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C /home/runner/.cache/huggingface/hub/  --use-compress-program unzstd
 
       - name: Setup Python
         uses: actions/setup-python@v5
@@ -196,16 +294,8 @@ jobs:
         run: |
           find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
 
-      - name: Save HF cache
-        id: hf-cache
-        uses: actions/cache/save@v4
-        with:
-          path: |
-            /home/runner/.cache/huggingface/hub/datasets--*
-            /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
-
   docker-e2e-tests-1st:
+    # Run this job first as a gate for running the remainder of the test matrix
     if: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' }}
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
@@ -252,6 +342,8 @@ jobs:
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
     timeout-minutes: 90
+    # Only run the remainder of the matrix if the first e2e check passed;
+    # this is to save on wasted compute costs for known failures that get caught in the first run
     needs: [pre-commit, pytest, docker-e2e-tests-1st]
 
     strategy:
@@ -261,21 +353,27 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.4.1
+            pytorch: 2.6.0
             num_gpus: 1
-            axolotl_extras:
+            axolotl_extras: llmcompressor
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
             pytorch: 2.5.1
             num_gpus: 1
-            axolotl_extras: vllm
+            axolotl_extras:
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.7.0
             num_gpus: 1
             axolotl_extras:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.7.0
+            num_gpus: 1
+            axolotl_extras:
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -300,3 +398,43 @@ jobs:
       - name: Run tests job on Modal
         run: |
           modal run cicd.e2e_tests
+
+  docker-e2e-cleanup:
+    runs-on: [self-hosted, modal]
+    timeout-minutes: 90
+    needs: [docker-e2e-tests]
+
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.6.0
+            num_gpus: 1
+            axolotl_extras: vllm
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Install Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+      - name: Install Modal
+        run: |
+          python -m pip install --upgrade pip
+          pip install modal==0.71.8 jinja2
+      - name: Update env vars
+        run: |
+          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}}" >> $GITHUB_ENV
+          echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
+          echo "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
+          echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
+          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
+      - name: Run tests job on Modal
+        run: |
+          modal run cicd.cleanup

.runpod/.gitignore

@@ -0,0 +1,161 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/#use-with-ide
+.pdm.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+#.idea/
+pod/scripts/config.yaml

.runpod/Dockerfile

@@ -0,0 +1,18 @@
+FROM axolotlai/axolotl-cloud:main-py3.11-cu124-2.6.0
+
+COPY .runpod/requirements.txt /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"
+ENV BASE_VOLUME=$BASE_VOLUME
+ENV HF_DATASETS_CACHE="${BASE_VOLUME}/huggingface-cache/datasets"
+ENV HUGGINGFACE_HUB_CACHE="${BASE_VOLUME}/huggingface-cache/hub"
+ENV TRANSFORMERS_CACHE="${BASE_VOLUME}/huggingface-cache/hub"
+
+COPY .runpod/src /src
+
+WORKDIR /src
+CMD ["python3", "/src/handler.py"]

.runpod/README.md

@@ -0,0 +1,335 @@
+<h1>LLM Post Training- Full fine-tune, LoRA, QLoRa etc. Llama/Mistral/Gemma and more</h1>
+
+# Configuration Options
+
+This document outlines all available configuration options for training models. The configuration can be provided as a JSON request.
+
+## Usage
+
+You can use these configuration Options:
+
+1. As a JSON request body:
+
+```json
+{
+  "input": {
+    "user_id": "user",
+    "model_id": "model-name",
+    "run_id": "run-id",
+    "credentials": {
+      "wandb_api_key": "", # add your Weights & biases key. TODO:  you will be able to set this in Enviornment variables.
+      "hf_token": "", # add your HF_token. TODO:  you will be able to set this in Enviornment variables.
+    },
+    "args": {
+      "base_model": "NousResearch/Llama-3.2-1B",
+      // ... other options
+    }
+  }
+}
+```
+
+## Configuration Options
+
+### Model Configuration
+
+| Option              | Description                                                                                   | Default              |
+| ------------------- | --------------------------------------------------------------------------------------------- | -------------------- |
+| `base_model`        | Path to the base model (local or HuggingFace)                                                 | Required             |
+| `base_model_config` | Configuration path for the base model                                                         | Same as base_model   |
+| `revision_of_model` | Specific model revision from HuggingFace hub                                                  | Latest               |
+| `tokenizer_config`  | Custom tokenizer configuration path                                                           | Optional             |
+| `model_type`        | Type of model to load                                                                         | AutoModelForCausalLM |
+| `tokenizer_type`    | Type of tokenizer to use                                                                      | AutoTokenizer        |
+| `hub_model_id`      | Repository ID where the model will be pushed on Hugging Face Hub (format: username/repo-name) | Optional             |
+
+## Model Family Identification
+
+| Option                     | Default | Description                    |
+| -------------------------- | ------- | ------------------------------ |
+| `is_falcon_derived_model`  | `false` | Whether model is Falcon-based  |
+| `is_llama_derived_model`   | `false` | Whether model is LLaMA-based   |
+| `is_qwen_derived_model`    | `false` | Whether model is Qwen-based    |
+| `is_mistral_derived_model` | `false` | Whether model is Mistral-based |
+
+## Model Configuration Overrides
+
+| Option                                          | Default    | Description                        |
+| ----------------------------------------------- | ---------- | ---------------------------------- |
+| `overrides_of_model_config.rope_scaling.type`   | `"linear"` | RoPE scaling type (linear/dynamic) |
+| `overrides_of_model_config.rope_scaling.factor` | `1.0`      | RoPE scaling factor                |
+
+### Model Loading Options
+
+| Option         | Description                   | Default |
+| -------------- | ----------------------------- | ------- |
+| `load_in_8bit` | Load model in 8-bit precision | false   |
+| `load_in_4bit` | Load model in 4-bit precision | false   |
+| `bf16`         | Use bfloat16 precision        | false   |
+| `fp16`         | Use float16 precision         | false   |
+| `tf32`         | Use tensor float 32 precision | false   |
+
+## Memory and Device Settings
+
+| Option             | Default   | Description             |
+| ------------------ | --------- | ----------------------- |
+| `gpu_memory_limit` | `"20GiB"` | GPU memory limit        |
+| `lora_on_cpu`      | `false`   | Load LoRA on CPU        |
+| `device_map`       | `"auto"`  | Device mapping strategy |
+| `max_memory`       | `null`    | Max memory per device   |
+
+## Training Hyperparameters
+
+| Option                        | Default   | Description                 |
+| ----------------------------- | --------- | --------------------------- |
+| `gradient_accumulation_steps` | `1`       | Gradient accumulation steps |
+| `micro_batch_size`            | `2`       | Batch size per GPU          |
+| `eval_batch_size`             | `null`    | Evaluation batch size       |
+| `num_epochs`                  | `4`       | Number of training epochs   |
+| `warmup_steps`                | `100`     | Warmup steps                |
+| `warmup_ratio`                | `0.05`    | Warmup ratio                |
+| `learning_rate`               | `0.00003` | Learning rate               |
+| `lr_quadratic_warmup`         | `false`   | Quadratic warmup            |
+| `logging_steps`               | `null`    | Logging frequency           |
+| `eval_steps`                  | `null`    | Evaluation frequency        |
+| `evals_per_epoch`             | `null`    | Evaluations per epoch       |
+| `save_strategy`               | `"epoch"` | Checkpoint saving strategy  |
+| `save_steps`                  | `null`    | Saving frequency            |
+| `saves_per_epoch`             | `null`    | Saves per epoch             |
+| `save_total_limit`            | `null`    | Maximum checkpoints to keep |
+| `max_steps`                   | `null`    | Maximum training steps      |
+
+### Dataset Configuration
+
+```yaml
+datasets:
+  - path: vicgalle/alpaca-gpt4 # HuggingFace dataset or TODO: You will be able to add the local path.
+    type: alpaca # Format type (alpaca, gpteacher, oasst, etc.)
+    ds_type: json # Dataset type
+    data_files: path/to/data # Source data files
+    train_on_split: train # Dataset split to use
+```
+
+## Chat Template Settings
+
+| Option                   | Default                          | Description            |
+| ------------------------ | -------------------------------- | ---------------------- |
+| `chat_template`          | `"tokenizer_default"`            | Chat template type     |
+| `chat_template_jinja`    | `null`                           | Custom Jinja template  |
+| `default_system_message` | `"You are a helpful assistant."` | Default system message |
+
+## Dataset Processing
+
+| Option                        | Default                    | Description                       |
+| ----------------------------- | -------------------------- | --------------------------------- |
+| `dataset_prepared_path`       | `"data/last_run_prepared"` | Path for prepared dataset         |
+| `push_dataset_to_hub`         | `""`                       | Push dataset to HF hub            |
+| `dataset_processes`           | `4`                        | Number of preprocessing processes |
+| `dataset_keep_in_memory`      | `false`                    | Keep dataset in memory            |
+| `shuffle_merged_datasets`     | `true`                     | Shuffle merged datasets           |
+| `dataset_exact_deduplication` | `true`                     | Deduplicate datasets              |
+
+## LoRA Configuration
+
+| Option                     | Default                | Description                    |
+| -------------------------- | ---------------------- | ------------------------------ |
+| `adapter`                  | `"lora"`               | Adapter type (lora/qlora)      |
+| `lora_model_dir`           | `""`                   | Directory with pretrained LoRA |
+| `lora_r`                   | `8`                    | LoRA attention dimension       |
+| `lora_alpha`               | `16`                   | LoRA alpha parameter           |
+| `lora_dropout`             | `0.05`                 | LoRA dropout                   |
+| `lora_target_modules`      | `["q_proj", "v_proj"]` | Modules to apply LoRA          |
+| `lora_target_linear`       | `false`                | Target all linear modules      |
+| `peft_layers_to_transform` | `[]`                   | Layers to transform            |
+| `lora_modules_to_save`     | `[]`                   | Modules to save                |
+| `lora_fan_in_fan_out`      | `false`                | Fan in/out structure           |
+
+## Optimization Settings
+
+| Option                    | Default | Description                |
+| ------------------------- | ------- | -------------------------- |
+| `train_on_inputs`         | `false` | Train on input prompts     |
+| `group_by_length`         | `false` | Group by sequence length   |
+| `gradient_checkpointing`  | `false` | Use gradient checkpointing |
+| `early_stopping_patience` | `3`     | Early stopping patience    |
+
+## Learning Rate Scheduling
+
+| Option                     | Default    | Description          |
+| -------------------------- | ---------- | -------------------- |
+| `lr_scheduler`             | `"cosine"` | Scheduler type       |
+| `lr_scheduler_kwargs`      | `{}`       | Scheduler parameters |
+| `cosine_min_lr_ratio`      | `null`     | Minimum LR ratio     |
+| `cosine_constant_lr_ratio` | `null`     | Constant LR ratio    |
+| `lr_div_factor`            | `null`     | LR division factor   |
+
+## Optimizer Settings
+
+| Option                 | Default      | Description         |
+| ---------------------- | ------------ | ------------------- |
+| `optimizer`            | `"adamw_hf"` | Optimizer choice    |
+| `optim_args`           | `{}`         | Optimizer arguments |
+| `optim_target_modules` | `[]`         | Target modules      |
+| `weight_decay`         | `null`       | Weight decay        |
+| `adam_beta1`           | `null`       | Adam beta1          |
+| `adam_beta2`           | `null`       | Adam beta2          |
+| `adam_epsilon`         | `null`       | Adam epsilon        |
+| `max_grad_norm`        | `null`       | Gradient clipping   |
+
+## Attention Implementations
+
+| Option                     | Default | Description                   |
+| -------------------------- | ------- | ----------------------------- |
+| `flash_optimum`            | `false` | Use better transformers       |
+| `xformers_attention`       | `false` | Use xformers                  |
+| `flash_attention`          | `false` | Use flash attention           |
+| `flash_attn_cross_entropy` | `false` | Flash attention cross entropy |
+| `flash_attn_rms_norm`      | `false` | Flash attention RMS norm      |
+| `flash_attn_fuse_qkv`      | `false` | Fuse QKV operations           |
+| `flash_attn_fuse_mlp`      | `false` | Fuse MLP operations           |
+| `sdp_attention`            | `false` | Use scaled dot product        |
+| `s2_attention`             | `false` | Use shifted sparse attention  |
+
+## Tokenizer Modifications
+
+| Option           | Default | Description                  |
+| ---------------- | ------- | ---------------------------- |
+| `special_tokens` | -       | Special tokens to add/modify |
+| `tokens`         | `[]`    | Additional tokens            |
+
+## Distributed Training
+
+| Option                  | Default | Description           |
+| ----------------------- | ------- | --------------------- |
+| `fsdp`                  | `null`  | FSDP configuration    |
+| `fsdp_config`           | `null`  | FSDP config options   |
+| `deepspeed`             | `null`  | Deepspeed config path |
+| `ddp_timeout`           | `null`  | DDP timeout           |
+| `ddp_bucket_cap_mb`     | `null`  | DDP bucket capacity   |
+| `ddp_broadcast_buffers` | `null`  | DDP broadcast buffers |
+
+<details>
+<summary><h3>Example Configuration Request:</h3></summary>
+
+Here's a complete example for fine-tuning a LLaMA model using LoRA:
+
+```json
+{
+  "input": {
+    "user_id": "user",
+    "model_id": "llama-test",
+    "run_id": "test-run",
+    "credentials": {
+      "wandb_api_key": "",
+      "hf_token": ""
+    },
+    "args": {
+      "base_model": "NousResearch/Llama-3.2-1B",
+      "load_in_8bit": false,
+      "load_in_4bit": false,
+      "strict": false,
+      "datasets": [
+        {
+          "path": "teknium/GPT4-LLM-Cleaned",
+          "type": "alpaca"
+        }
+      ],
+      "dataset_prepared_path": "last_run_prepared",
+      "val_set_size": 0.1,
+      "output_dir": "./outputs/lora-out",
+      "adapter": "lora",
+      "sequence_len": 2048,
+      "sample_packing": true,
+      "eval_sample_packing": true,
+      "pad_to_sequence_len": true,
+      "lora_r": 16,
+      "lora_alpha": 32,
+      "lora_dropout": 0.05,
+      "lora_target_modules": [
+        "gate_proj",
+        "down_proj",
+        "up_proj",
+        "q_proj",
+        "v_proj",
+        "k_proj",
+        "o_proj"
+      ],
+      "gradient_accumulation_steps": 2,
+      "micro_batch_size": 2,
+      "num_epochs": 1,
+      "optimizer": "adamw_8bit",
+      "lr_scheduler": "cosine",
+      "learning_rate": 0.0002,
+      "train_on_inputs": false,
+      "group_by_length": false,
+      "bf16": "auto",
+      "tf32": false,
+      "gradient_checkpointing": true,
+      "logging_steps": 1,
+      "flash_attention": true,
+      "loss_watchdog_threshold": 5,
+      "loss_watchdog_patience": 3,
+      "warmup_steps": 10,
+      "evals_per_epoch": 4,
+      "saves_per_epoch": 1,
+      "weight_decay": 0,
+      "hub_model_id": "runpod/llama-fr-lora",
+      "wandb_name": "test-run-1",
+      "wandb_project": "test-run-1",
+      "wandb_entity": "axo-test",
+      "special_tokens": {
+        "pad_token": "<|end_of_text|>"
+      }
+    }
+  }
+}
+```
+
+</details>
+
+### Advanced Features
+
+#### Wandb Integration
+
+- `wandb_project`: Project name for Weights & Biases
+- `wandb_entity`: Team name in W&B
+- `wandb_watch`: Monitor model with W&B
+- `wandb_name`: Name of the W&B run
+- `wandb_run_id`: ID for the W&B run
+
+#### Performance Optimization
+
+- `sample_packing`: Enable efficient sequence packing
+- `eval_sample_packing`: Use sequence packing during evaluation
+- `torch_compile`: Enable PyTorch 2.0 compilation
+- `flash_attention`: Use Flash Attention implementation
+- `xformers_attention`: Use xFormers attention implementation
+
+### Available Optimizers
+
+The following optimizers are supported:
+
+- `adamw_hf`: HuggingFace's AdamW implementation
+- `adamw_torch`: PyTorch's AdamW
+- `adamw_torch_fused`: Fused AdamW implementation
+- `adamw_torch_xla`: XLA-optimized AdamW
+- `adamw_apex_fused`: NVIDIA Apex fused AdamW
+- `adafactor`: Adafactor optimizer
+- `adamw_anyprecision`: Anyprecision AdamW
+- `adamw_bnb_8bit`: 8-bit AdamW from bitsandbytes
+- `lion_8bit`: 8-bit Lion optimizer
+- `lion_32bit`: 32-bit Lion optimizer
+- `sgd`: Stochastic Gradient Descent
+- `adagrad`: Adagrad optimizer
+
+## Notes
+
+- Set `load_in_8bit: true` or `load_in_4bit: true` for memory-efficient training
+- Enable `flash_attention: true` for faster training on modern GPUs
+- Use `gradient_checkpointing: true` to reduce memory usage
+- Adjust `micro_batch_size` and `gradient_accumulation_steps` based on your GPU memory
+
+For more detailed information, please refer to the [documentation](https://axolotl-ai-cloud.github.io/axolotl/docs/config.html).
+
+### Errors:
+
+- if you face any issues with the Flash Attention-2, Delete yoor worker and Re-start.

.runpod/hub.json

@@ -0,0 +1,93 @@
+{
+  "title": "Axolotl Fine-Tuning",
+  "description": "Serverless fine-tuning of open-source LLMs with Axolotl. Supports LoRA, QLoRA, DPO, and more using Hugging Face models and datasets.",
+  "type": "serverless",
+  "category": "language",
+  "iconUrl": "https://avatars.githubusercontent.com/u/167502477",
+  "config": {
+    "runsOn": "GPU",
+    "containerDiskInGb": 200,
+    "gpuCount": 1,
+    "allowedCudaVersions": [
+      "12.8",
+      "12.7",
+      "12.6",
+      "12.5",
+      "12.4"
+    ],
+    "presets": [],
+    "env": [
+      {
+        "key": "TOKENIZER",
+        "input": {
+          "name": "Tokenizer",
+          "type": "string",
+          "description": "Name or path of the Hugging Face tokenizer to use.",
+          "default": "",
+          "advanced": true
+        }
+      },
+      {
+        "key": "MAX_NUM_SEQS",
+        "input": {
+          "name": "Max Num Seqs",
+          "type": "number",
+          "description": "Maximum number of sequences per iteration.",
+          "default": 256,
+          "advanced": true
+        }
+      },
+      {
+        "key": "DISABLE_LOG_STATS",
+        "input": {
+          "name": "Disable Log Stats",
+          "type": "boolean",
+          "description": "Disable logging statistics.",
+          "default": false,
+          "trueValue": "true",
+          "falseValue": "false"
+        }
+      },
+      {
+        "key": "LOAD_FORMAT",
+        "input": {
+          "name": "Load Format",
+          "type": "string",
+          "description": "The format of the model weights to load.",
+          "default": "auto",
+          "options": [
+            {
+              "label": "auto",
+              "value": "auto"
+            },
+            {
+              "label": "pt",
+              "value": "pt"
+            },
+            {
+              "label": "safetensors",
+              "value": "safetensors"
+            },
+            {
+              "label": "npcache",
+              "value": "npcache"
+            },
+            {
+              "label": "dummy",
+              "value": "dummy"
+            },
+            {
+              "label": "tensorizer",
+              "value": "tensorizer"
+            },
+            {
+              "label": "bitsandbytes",
+              "value": "bitsandbytes"
+            }
+          ],
+          "advanced": true
+        }
+      }
+    ]
+  }
+}

.runpod/requirements.txt

@@ -0,0 +1,7 @@
+# Required Python packages get listed here, one per line.
+# Reccomended to lock the version number to avoid unexpected changes.
+
+# You can also install packages from a git repository, e.g.:
+# git+https://github.com/runpod/runpod-python.git
+# To learn more, see https://pip.pypa.io/en/stable/reference/requirements-file-format/
+runpod~=1.7.0

.runpod/src/config/config.yaml

@@ -0,0 +1,577 @@
+# # This is the huggingface model that contains *.pt, *.safetensors, or *.bin files
+# # This can also be a relative path to a model on disk
+# base_model: ./llama-7b-hf
+# # You can specify an ignore pattern if the model repo contains more than 1 model type (*.pt, etc)
+# base_model_ignore_patterns:
+# # If the base_model repo on hf hub doesn't include configuration .json files,
+# # You can set that here, or leave this empty to default to base_model
+# base_model_config: ./llama-7b-hf
+# # You can specify to choose a specific model revision from huggingface hub
+# model_revision:
+# # Optional tokenizer configuration override in case you want to use a different tokenizer
+# # than the one defined in the base model
+# tokenizer_config:
+# # If you want to specify the type of model to load, AutoModelForCausalLM is a good choice too
+# model_type: AutoModelForCausalLM
+# # Corresponding tokenizer for the model AutoTokenizer is a good choice
+# tokenizer_type: AutoTokenizer
+# # Trust remote code for untrusted source
+# trust_remote_code:
+# # use_fast option for tokenizer loading from_pretrained, default to True
+# tokenizer_use_fast:
+# # Whether to use the legacy tokenizer setting, defaults to True
+# tokenizer_legacy:
+# # Resize the model embeddings when new tokens are added to multiples of 32
+# # This is reported to improve training speed on some models
+# resize_token_embeddings_to_32x:
+
+# # Used to identify which the model is based on
+# is_falcon_derived_model:
+# is_llama_derived_model:
+# # Please note that if you set this to true, `padding_side` will be set to "left" by default
+# is_mistral_derived_model:
+# is_qwen_derived_model:
+
+# # optional overrides to the base model configuration
+# model_config:
+#   # RoPE Scaling https://github.com/huggingface/transformers/pull/24653
+#   rope_scaling:
+#     type: # linear | dynamic
+#     factor: # float
+
+
+# # Whether you are training a 4-bit GPTQ quantized model
+# gptq: true
+# gptq_groupsize: 128 # group size
+# gptq_model_v1: false # v1 or v2
+
+# # This will attempt to quantize the model down to 8 bits and use adam 8 bit optimizer
+# load_in_8bit: true
+# # Use bitsandbytes 4 bit
+# load_in_4bit:
+
+# # Use CUDA bf16
+# bf16: true # bool or 'full' for `bf16_full_eval`. require >=ampere
+# # Use CUDA fp16
+# fp16: true
+# # Use CUDA tf32
+# tf32: true # require >=ampere
+
+# # No AMP (automatic mixed precision)
+# bfloat16: true # require >=ampere
+# float16: true
+
+# # A list of one or more datasets to finetune the model with
+# datasets:
+#   # HuggingFace dataset repo | s3://,gs:// path | "json" for local dataset, make sure to fill data_files
+#   - path: vicgalle/alpaca-gpt4
+#   # The type of prompt to use for training. [alpaca, sharegpt, gpteacher, oasst, reflection]
+#     type: alpaca # format | format:<prompt_style> (chat/instruct) | <prompt_strategies>.load_<load_fn>
+#     ds_type: # Optional[str] (json|arrow|parquet|text|csv) defines the datatype when path is a file
+#     data_files: # Optional[str] path to source data files
+#     shards: # Optional[int] number of shards to split data into
+#     name: # Optional[str] name of dataset configuration to load
+#     train_on_split: train # Optional[str] name of dataset split to load from
+
+#     # Optional[str] fastchat conversation type, only used with type: sharegpt
+#     conversation:  # Options (see Conversation 'name'): https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
+#     field_human: # Optional[str]. Human key to use for conversation.
+#     field_model: # Optional[str]. Assistant key to use for conversation.
+
+#   # Custom user prompt
+#   - path: repo
+#     type:
+#       # The below are defaults. only set what's needed.
+#       system_prompt: ""
+#       system_format: "{system}"
+#       field_system: system
+#       field_instruction: instruction
+#       field_input: input
+#       field_output: output
+
+#       # Customizable to be single line or multi-line
+#       # 'format' can include {input}
+#       format: |-
+#         User: {instruction} {input}
+#         Assistant:
+#       # 'no_input_format' cannot include {input}
+#       no_input_format: "{instruction} "
+
+#       # For `completion` datsets only, uses the provided field instead of `text` column
+#       field:
+
+# # Axolotl attempts to save the dataset as an arrow after packing the data together so
+# # subsequent training attempts load faster, relative path
+# dataset_prepared_path: data/last_run_prepared
+# # Push prepared dataset to hub
+# push_dataset_to_hub: # repo path
+# # The maximum number of processes to use while preprocessing your input dataset. This defaults to `os.cpu_count()`
+# # if not set.
+# dataset_processes: # defaults to os.cpu_count() if not set
+# # push checkpoints to hub
+# hub_model_id: # repo path to push finetuned model
+# # how to push checkpoints to hub
+# # https://huggingface.co/docs/transformers/v4.31.0/en/main_classes/trainer#transformers.TrainingArguments.hub_strategy
+# hub_strategy:
+# # Whether to use hf `use_auth_token` for loading datasets. Useful for fetching private datasets
+# # Required to be true when used in combination with `push_dataset_to_hub`
+# hf_use_auth_token: # boolean
+# # How much of the dataset to set aside as evaluation. 1 = 100%, 0.50 = 50%, etc. 0 for no eval.
+# val_set_size: 0.04
+# # Num shards for whole dataset
+# dataset_shard_num:
+# # Index of shard to use for whole dataset
+# dataset_shard_idx:
+
+# # The maximum length of an input to train with, this should typically be less than 2048
+# # as most models have a token/context limit of 2048
+# sequence_len: 2048
+# # Pad inputs so each step uses constant sized buffers
+# # This will reduce memory fragmentation and may prevent OOMs, by re-using memory more efficiently
+# pad_to_sequence_len:
+# # Max sequence length to concatenate training samples together up to
+# # Inspired by StackLLaMA. see https://huggingface.co/blog/stackllama#supervised-fine-tuning
+# # FutureWarning: This will soon be DEPRECATED
+# max_packed_sequence_len: 1024
+# # Use efficient multi-packing with block diagonal attention and per sequence position_ids. Recommend set to 'true'
+# sample_packing:
+# # Set to 'false' if getting errors during eval with sample_packing on.
+# eval_sample_packing:
+# # You can set these packing optimizations AFTER starting a training at least once.
+# # The trainer will provide recommended values for these values.
+# sample_packing_eff_est:
+# total_num_tokens:
+
+# # If you want to use 'lora' or 'qlora' or leave blank to train all parameters in original model
+# adapter: lora
+# # If you already have a lora model trained that you want to load, put that here.
+# # This means after training, if you want to test the model, you should set this to the value of `lora_out_dir`.
+# lora_model_dir:
+
+# # LoRA hyperparameters
+# # For more details about the following options, see:
+# # https://www.anyscale.com/blog/fine-tuning-llms-lora-or-full-parameter-an-in-depth-analysis-with-llama-2
+# lora_r: 8
+# lora_alpha: 16
+# lora_dropout: 0.05
+# lora_target_modules:
+#   - q_proj
+#   - v_proj
+# #  - k_proj
+# #  - o_proj
+# #  - gate_proj
+# #  - down_proj
+# #  - up_proj
+# lora_target_linear: # If true, will target all linear layers
+
+# # If you added new tokens to the tokenizer, you may need to save some LoRA modules because they need to know the new tokens.
+# # For LLaMA and Mistral, you need to save `embed_tokens` and `lm_head`. It may vary for other models.
+# # `embed_tokens` converts tokens to embeddings, and `lm_head` converts embeddings to token probabilities.
+# # https://github.com/huggingface/peft/issues/334#issuecomment-1561727994
+# lora_modules_to_save:
+# #  - embed_tokens
+# #  - lm_head
+
+# # Once you complete training, the model will be saved to the following directory.
+# # If you merge the adapter to the base model, a subdirectory `merged` will be created under this directory.
+# # Make sure `lora_model_dir` points to this directory if you want to use the trained model.
+# lora_out_dir:
+# lora_fan_in_fan_out: false
+
+# # ReLoRA configuration
+# # Must use either 'lora' or 'qlora' adapter, and does not support fsdp or deepspeed
+# relora_steps: # Number of steps per ReLoRA restart
+# relora_warmup_steps: # Number of per-restart warmup steps
+# relora_cpu_offload: # True to perform lora weight merges on cpu during restarts, for modest gpu memory savings
+
+# # wandb configuration if you're using it
+# wandb_mode: # "offline" to save run metadata locally and not sync to the server, "disabled" to turn off wandb
+# wandb_project: # Your wandb project name
+# wandb_entity: # A wandb Team name if using a Team
+# wandb_watch:
+# wandb_run_id: # Set the name of your wandb run
+# wandb_log_model: # "checkpoint" to log model to wandb Artifacts every `save_steps` or "end" to log only at the end of training
+
+# # Where to save the full-finetuned model to
+# output_dir: ./completed-model
+
+# # Whether to use torch.compile and which backend to use
+# torch_compile:  # bool
+# torch_compile_backend:  # Optional[str]
+
+# # Training hyperparameters
+
+# # If greater than 1, backpropagation will be skipped and the gradients will be accumulated for the given number of steps.
+# gradient_accumulation_steps: 1
+# # The number of samples to include in each batch. This is the number of samples sent to each GPU.
+# micro_batch_size: 2
+# eval_batch_size:
+# num_epochs: 4
+# warmup_steps: 100  # cannot use with warmup_ratio
+# warmup_ratio: 0.05  # cannot use with warmup_steps
+# learning_rate: 0.00003
+# lr_quadratic_warmup:
+# logging_steps:
+# save_strategy: # Set to `no` to skip checkpoint saves
+# save_steps: # Leave empty to save at each epoch
+# eval_steps: # Leave empty to eval at each epoch, integers for every N steps. decimal for fraction of total steps
+# save_total_limit: # Checkpoints saved at a time
+# # Maximum number of iterations to train for. It precedes num_epochs which means that
+# # if both are set, num_epochs will not be guaranteed.
+# # e.g., when 1 epoch is 1000 steps => `num_epochs: 2` and `max_steps: 100` will train for 100 steps
+# max_steps:
+
+# eval_table_size: # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0
+# eval_table_max_new_tokens: # Total number of tokens generated for predictions sent to wandb. Default is 128
+
+# # Save model as safetensors (require safetensors package)
+# save_safetensors:
+
+# # Whether to mask out or include the human's prompt from the training labels
+# train_on_inputs: false
+# # Group similarly sized data to minimize padding.
+# # May be slower to start, as it must download and sort the entire dataset.
+# # Note that training loss may have an oscillating pattern with this enabled.
+# group_by_length: false
+
+# # Whether to use gradient checkpointing https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing
+# gradient_checkpointing: false
+
+# # Stop training after this many evaluation losses have increased in a row
+# # https://huggingface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppingCallback
+# early_stopping_patience: 3
+
+# # Specify a scheduler and kwargs to use with the optimizer
+# lr_scheduler: # 'one_cycle' | 'log_sweep' | empty for cosine
+# lr_scheduler_kwargs:
+
+# # For one_cycle optim
+# lr_div_factor: # Learning rate div factor
+
+# # For log_sweep optim
+# log_sweep_min_lr:
+# log_sweep_max_lr:
+
+# # Specify optimizer
+# # Valid values are driven by the Transformers OptimizerNames class, see:
+# # https://github.com/huggingface/transformers/blob/95b374952dc27d8511541d6f5a4e22c9ec11fb24/src/transformers/training_args.py#L134
+# #
+# # Note that not all optimizers may be available in your environment, ex: 'adamw_anyprecision' is part of
+# # torchdistx, 'adamw_bnb_8bit' is part of bnb.optim.Adam8bit, etc. When in doubt, it is recommended to start with the optimizer used
+# # in the examples/ for your model and fine-tuning use case.
+# #
+# # Valid values for 'optimizer' include:
+# # - adamw_hf
+# # - adamw_torch
+# # - adamw_torch_fused
+# # - adamw_torch_xla
+# # - adamw_apex_fused
+# # - adafactor
+# # - adamw_anyprecision
+# # - sgd
+# # - adagrad
+# # - adamw_bnb_8bit
+# # - lion_8bit
+# # - lion_32bit
+# # - paged_adamw_32bit
+# # - paged_adamw_8bit
+# # - paged_lion_32bit
+# # - paged_lion_8bit
+# optimizer:
+# # Specify weight decay
+# weight_decay:
+# # adamw hyperparams
+# adam_beta1:
+# adam_beta2:
+# adam_epsilon:
+# # Gradient clipping max norm
+# max_grad_norm:
+
+# # Augmentation techniques
+# # NEFT https://arxiv.org/abs/2310.05914, set this to a number (paper default is 5) to add noise to embeddings
+# # currently only supported on Llama and Mistral
+# noisy_embedding_alpha:
+
+# # Whether to bettertransformers
+# flash_optimum:
+# # Whether to use xformers attention patch https://github.com/facebookresearch/xformers:
+# xformers_attention:
+# # Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:
+# flash_attention:
+# flash_attn_cross_entropy:  # Whether to use flash-attention cross entropy implementation - advanced use only
+# flash_attn_rms_norm:  # Whether to use flash-attention rms norm implementation - advanced use only
+# flash_attn_fuse_qkv: # Whether to fuse QKV into a single operation
+# flash_attn_fuse_mlp: # Whether to fuse part of the MLP into a single operation
+# # Whether to use scaled-dot-product attention
+# # https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html
+# sdp_attention:
+# # Landmark attention (only llama)
+# landmark_attention:
+# # xpos RoPE see https://github.com/kaiokendev/cutoff-len-is-context-len/blob/main/util/xpos_rope_llama_monkey_patch.py
+# # LLaMA only
+# xpos_rope:
+
+# # Resume from a specific checkpoint dir
+# resume_from_checkpoint:
+# # If resume_from_checkpoint isn't set and you simply want it to start where it left off.
+# # Be careful with this being turned on between different models.
+# auto_resume_from_checkpoints: false
+
+# # Don't mess with this, it's here for accelerate and torchrun
+# local_rank:
+
+# # Add or change special tokens.
+# # If you add tokens here, you don't need to add them to the `tokens` list.
+# special_tokens:
+#   # bos_token: "<s>"
+#   # eos_token: "</s>"
+#   # unk_token: "<unk>"
+
+# # Add extra tokens.
+# tokens:
+
+# # FSDP
+# fsdp:
+# fsdp_config:
+
+# # Deepspeed config path. e.g., deepspeed/zero3.json
+# deepspeed:
+
+# # Advanced DDP Arguments
+# ddp_timeout:
+# ddp_bucket_cap_mb:
+# ddp_broadcast_buffers:
+
+# # Path to torch distx for optim 'adamw_anyprecision'
+# torchdistx_path:
+
+# # Set to HF dataset for type: 'completion' for streaming instead of pre-tokenize
+# pretraining_dataset:
+
+# # Debug mode
+# debug:
+
+# # Seed
+# seed:
+
+# # Allow overwrite yml config using from cli
+# strict:
+
+
+
+base_model: ${BASE_MODEL}
+base_model_ignore_patterns: ${BASE_MODEL_IGNORE_PATTERNS}
+base_model_config: ${BASE_MODEL_CONFIG}
+revision_of_model: ${REVISION_OF_MODEL}
+tokenizer_config: ${TOKENIZER_CONFIG}
+model_type: ${MODEL_TYPE}
+tokenizer_type: ${TOKENIZER_TYPE}
+trust_remote_code: ${TRUST_REMOTE_CODE}
+tokenizer_use_fast: ${TOKENIZER_USE_FAST}
+tokenizer_legacy: ${TOKENIZER_LEGACY}
+resize_token_embeddings_to_32x: ${RESIZE_TOKEN_EMBEDDINGS_TO_32X}
+
+is_falcon_derived_model: ${IS_FALCON_DERIVED_MODEL}
+is_llama_derived_model: ${IS_LLAMA_DERIVED_MODEL}
+is_qwen_derived_model: ${IS_QWEN_DERIVED_MODEL}
+is_mistral_derived_model: ${IS_MISTRAL_DERIVED_MODEL}
+
+overrides_of_model_config:
+  rope_scaling:
+    type: ${ROPE_SCALING_TYPE}
+    factor: ${ROPE_SCALING_FACTOR}
+
+bnb_config_kwargs:
+  llm_int8_has_fp16_weight: ${BNB_LLM_INT8_HAS_FP16_WEIGHT}
+  bnb_4bit_quant_type: ${BNB_4BIT_QUANT_TYPE}
+  bnb_4bit_use_double_quant: ${BNB_4BIT_USE_DOUBLE_QUANT}
+
+gptq: ${GPTQ}
+load_in_8bit: ${LOAD_IN_8BIT}
+load_in_4bit: ${LOAD_IN_4BIT}
+bf16: ${BF16}
+fp16: ${FP16}
+tf32: ${TF32}
+bfloat16: ${BFLOAT16}
+float16: ${FLOAT16}
+
+gpu_memory_limit: ${GPU_MEMORY_LIMIT}
+lora_on_cpu: ${LORA_ON_CPU}
+
+datasets:
+  - path: ${DATASET_PATH}
+    type: ${DATASET_TYPE}
+    ds_type: ${DATASET_DS_TYPE}
+    data_files: ${DATASET_DATA_FILES}
+    shards: ${DATASET_SHARDS}
+    name: ${DATASET_NAME}
+    train_on_split: ${DATASET_TRAIN_ON_SPLIT}
+    revision: ${DATASET_REVISION}
+    trust_remote_code: ${DATASET_TRUST_REMOTE_CODE}
+
+rl: ${RL}
+dpo_use_weighting: ${DPO_USE_WEIGHTING}
+
+chat_template: ${CHAT_TEMPLATE}
+chat_template_jinja: ${CHAT_TEMPLATE_JINJA}
+default_system_message: ${DEFAULT_SYSTEM_MESSAGE}
+dataset_prepared_path: ${DATASET_PREPARED_PATH}
+push_dataset_to_hub: ${PUSH_DATASET_TO_HUB}
+dataset_processes: ${DATASET_PROCESSES}
+dataset_keep_in_memory: ${DATASET_KEEP_IN_MEMORY}
+hub_model_id: ${HUB_MODEL_ID}
+hub_strategy: ${HUB_STRATEGY}
+hf_use_auth_token: ${HF_USE_AUTH_TOKEN}
+val_set_size: ${VAL_SET_SIZE}
+dataset_shard_num: ${DATASET_SHARD_NUM}
+dataset_shard_idx: ${DATASET_SHARD_IDX}
+
+sequence_len: ${SEQUENCE_LEN}
+pad_to_sequence_len: ${PAD_TO_SEQUENCE_LEN}
+sample_packing: ${SAMPLE_PACKING}
+eval_sample_packing: ${EVAL_SAMPLE_PACKING}
+sample_packing_eff_est: ${SAMPLE_PACKING_EFF_EST}
+total_num_tokens: ${TOTAL_NUM_TOKENS}
+sample_packing_group_size: ${SAMPLE_PACKING_GROUP_SIZE}
+sample_packing_bin_size: ${SAMPLE_PACKING_BIN_SIZE}
+
+batch_flattening: ${BATCH_FLATTENING}
+device_map: ${DEVICE_MAP}
+max_memory: ${MAX_MEMORY}
+
+adapter: ${ADAPTER}
+lora_model_dir: ${LORA_MODEL_DIR}
+
+lora_r: ${LORA_R}
+lora_alpha: ${LORA_ALPHA}
+lora_dropout: ${LORA_DROPOUT}
+lora_target_modules:
+  - ${LORA_TARGET_MODULES}
+lora_target_linear: ${LORA_TARGET_LINEAR}
+peft_layers_to_transform: ${PEFT_LAYERS_TO_TRANSFORM}
+lora_modules_to_save: ${LORA_MODULES_TO_SAVE}
+lora_fan_in_fan_out: ${LORA_FAN_IN_FAN_OUT}
+
+loraplus_lr_ratio: ${LORAPLUS_LR_RATIO}
+loraplus_lr_embedding: ${LORAPLUS_LR_EMBEDDING}
+
+peft:
+  loftq_config:
+    loftq_bits: ${LOFTQ_BITS}
+
+relora_steps: ${RELORA_STEPS}
+relora_warmup_steps: ${RELORA_WARMUP_STEPS}
+relora_anneal_steps: ${RELORA_ANNEAL_STEPS}
+relora_prune_ratio: ${RELORA_PRUNE_RATIO}
+relora_cpu_offload: ${RELORA_CPU_OFFLOAD}
+
+wandb_mode: ${WANDB_MODE}
+wandb_project: ${WANDB_PROJECT}
+wandb_entity: ${WANDB_ENTITY}
+wandb_watch: ${WANDB_WATCH}
+wandb_name: ${WANDB_NAME}
+wandb_run_id: ${WANDB_RUN_ID}
+wandb_log_model: ${WANDB_LOG_MODEL}
+
+mlflow_tracking_uri: ${MLFLOW_TRACKING_URI}
+mlflow_experiment_name: ${MLFLOW_EXPERIMENT_NAME}
+mlflow_run_name: ${MLFLOW_RUN_NAME}
+hf_mlflow_log_artifacts: ${HF_MLFLOW_LOG_ARTIFACTS}
+
+use_comet: ${USE_COMET}
+comet_api_key: ${COMET_API_KEY}
+comet_workspace: ${COMET_WORKSPACE}
+comet_project_name: ${COMET_PROJECT_NAME}
+comet_experiment_key: ${COMET_EXPERIMENT_KEY}
+comet_mode: ${COMET_MODE}
+comet_online: ${COMET_ONLINE}
+comet_experiment_config: ${COMET_EXPERIMENT_CONFIG}
+
+output_dir: ${OUTPUT_DIR}
+
+torch_compile: ${TORCH_COMPILE}
+torch_compile_backend: ${TORCH_COMPILE_BACKEND}
+
+gradient_accumulation_steps: ${GRADIENT_ACCUMULATION_STEPS}
+micro_batch_size: ${MICRO_BATCH_SIZE}
+eval_batch_size: ${EVAL_BATCH_SIZE}
+num_epochs: ${NUM_EPOCHS}
+warmup_steps: ${WARMUP_STEPS}
+warmup_ratio: ${WARMUP_RATIO}
+learning_rate: ${LEARNING_RATE}
+lr_quadratic_warmup: ${LR_QUADRATIC_WARMUP}
+logging_steps: ${LOGGING_STEPS}
+eval_steps: ${EVAL_STEPS}
+evals_per_epoch: ${EVALS_PER_EPOCH}
+save_strategy: ${SAVE_STRATEGY}
+save_steps: ${SAVE_STEPS}
+saves_per_epoch: ${SAVES_PER_EPOCH}
+save_total_limit: ${SAVE_TOTAL_LIMIT}
+max_steps: ${MAX_STEPS}
+
+eval_table_size: ${EVAL_TABLE_SIZE}
+eval_max_new_tokens: ${EVAL_MAX_NEW_TOKENS}
+eval_causal_lm_metrics: ${EVAL_CAUSAL_LM_METRICS}
+
+profiler_steps: ${PROFILER_STEPS}
+loss_watchdog_threshold: ${LOSS_WATCHDOG_THRESHOLD}
+loss_watchdog_patience: ${LOSS_WATCHDOG_PATIENCE}
+
+save_safetensors: ${SAVE_SAFETENSORS}
+train_on_inputs: ${TRAIN_ON_INPUTS}
+group_by_length: ${GROUP_BY_LENGTH}
+gradient_checkpointing: ${GRADIENT_CHECKPOINTING}
+early_stopping_patience: ${EARLY_STOPPING_PATIENCE}
+
+lr_scheduler: ${LR_SCHEDULER}
+lr_scheduler_kwargs: ${LR_SCHEDULER_KWARGS}
+cosine_min_lr_ratio: ${COSINE_MIN_LR_RATIO}
+cosine_constant_lr_ratio: ${COSINE_CONSTANT_LR_RATIO}
+lr_div_factor: ${LR_DIV_FACTOR}
+
+optimizer: ${OPTIMIZER}
+optim_args: ${OPTIM_ARGS}
+optim_target_modules: ${OPTIM_TARGET_MODULES}
+weight_decay: ${WEIGHT_DECAY}
+adam_beta1: ${ADAM_BETA1}
+adam_beta2: ${ADAM_BETA2}
+adam_epsilon: ${ADAM_EPSILON}
+max_grad_norm: ${MAX_GRAD_NORM}
+
+neftune_noise_alpha: ${NEFTUNE_NOISE_ALPHA}
+
+flash_optimum: ${FLASH_OPTIMUM}
+xformers_attention: ${XFORMERS_ATTENTION}
+flash_attention: ${FLASH_ATTENTION}
+flash_attn_cross_entropy: ${FLASH_ATTN_CROSS_ENTROPY}
+flash_attn_rms_norm: ${FLASH_ATTN_RMS_NORM}
+flash_attn_fuse_qkv: ${FLASH_ATTN_FUSE_QKV}
+flash_attn_fuse_mlp: ${FLASH_ATTN_FUSE_MLP}
+sdp_attention: ${SDP_ATTENTION}
+s2_attention: ${S2_ATTENTION}
+resume_from_checkpoint: ${RESUME_FROM_CHECKPOINT}
+auto_resume_from_checkpoints: ${AUTO_RESUME_FROM_CHECKPOINTS}
+
+local_rank: ${LOCAL_RANK}
+
+special_tokens:
+  bos_token: ${SPECIAL_TOKEN_BOS}
+  eos_token: ${SPECIAL_TOKEN_EOS}
+  unk_token: ${SPECIAL_TOKEN_UNK}
+  pad_token: ${SPECIAL_TOKEN_PAD}
+
+tokens: ${TOKENS}
+
+fsdp: ${FSDP}
+fsdp_config: ${FSDP_CONFIG}
+deepspeed: ${DEEPSPEED}
+
+ddp_timeout: ${DDP_TIMEOUT}
+ddp_bucket_cap_mb: ${DDP_BUCKET_CAP_MB}
+ddp_broadcast_buffers: ${DDP_BROADCAST_BUFFERS}
+
+torchdistx_path: ${TORCHDISTX_PATH}
+pretraining_dataset: ${PRETRAINING_DATASET}
+debug: ${DEBUG}
+seed: ${SEED}
+strict: ${STRICT}

.runpod/src/handler.py

@@ -0,0 +1,66 @@
+"""
+Runpod serverless entrypoint handler
+"""
+
+import os
+
+import runpod
+import yaml
+from huggingface_hub._login import login
+from train import train
+from utils import get_output_dir
+
+BASE_VOLUME = os.environ.get("BASE_VOLUME", "/runpod-volume")
+if not os.path.exists(BASE_VOLUME):
+    os.makedirs(BASE_VOLUME)
+
+logger = runpod.RunPodLogger()
+
+
+async def handler(job):
+    runpod_job_id = job["id"]
+    inputs = job["input"]
+    run_id = inputs.get("run_id", "default_run_id")
+    args = inputs.get("args", {})
+
+    # Set output directory
+    output_dir = os.path.join(BASE_VOLUME, get_output_dir(run_id))
+    args["output_dir"] = output_dir
+
+    # First save args to a temporary config file
+    config_path = "/workspace/test_config.yaml"
+
+    # Add run_name and job_id to args before saving
+    args["run_name"] = run_id
+    args["runpod_job_id"] = runpod_job_id
+
+    yaml_data = yaml.dump(args, default_flow_style=False)
+    with open(config_path, "w", encoding="utf-8") as file:
+        file.write(yaml_data)
+
+    # Handle credentials
+    credentials = inputs.get("credentials", {})
+
+    if "wandb_api_key" in credentials:
+        os.environ["WANDB_API_KEY"] = credentials["wandb_api_key"]
+    if "hf_token" in credentials:
+        os.environ["HF_TOKEN"] = credentials["hf_token"]
+
+    if os.environ.get("HF_TOKEN"):
+        login(token=os.environ["HF_TOKEN"])
+    else:
+        logger.info("No HF_TOKEN provided. Skipping login.")
+
+    logger.info("Starting Training.")
+    async for result in train(config_path):  # Pass the config path instead of args
+        logger.info(result)
+    logger.info("Training Complete.")
+
+    # Cleanup
+    if "WANDB_API_KEY" in os.environ:
+        del os.environ["WANDB_API_KEY"]
+    if "HF_TOKEN" in os.environ:
+        del os.environ["HF_TOKEN"]
+
+
+runpod.serverless.start({"handler": handler, "return_aggregate_stream": True})

.runpod/src/test_input.json

@@ -0,0 +1,61 @@
+{
+  "input": {
+    "user_id": "user",
+    "model_id": "llama-test",
+    "run_id": "llama-test",
+    "credentials": {
+      "wandb_api_key": "",
+      "hf_token": ""
+    },
+    "args": {
+      "base_model": "NousResearch/Meta-Llama-3-8B",
+      "model_type": "LlamaForCausalLM",
+      "tokenizer_type": "AutoTokenizer",
+      "load_in_8bit": true,
+      "load_in_4bit": false,
+      "strict": false,
+      "datasets": [
+        {
+          "path": "mhenrichsen/alpaca_2k_test",
+          "type": "alpaca"
+        }
+      ],
+      "val_set_size": 0.05,
+      "output_dir": "./outputs/lora-out",
+      "sequence_len": 4096,
+      "sample_packing": true,
+      "eval_sample_packing": false,
+      "pad_to_sequence_len": true,
+      "adapter": "lora",
+      "lora_r": 32,
+      "lora_alpha": 16,
+      "lora_dropout": 0.05,
+      "lora_target_linear": true,
+      "lora_modules_to_save": [
+        "embed_tokens",
+        "lm_head"
+      ],
+      "gradient_accumulation_steps": 4,
+      "micro_batch_size": 2,
+      "num_epochs": 1,
+      "optimizer": "adamw_bnb_8bit",
+      "lr_scheduler": "cosine",
+      "learning_rate": 0.0002,
+      "train_on_inputs": false,
+      "group_by_length": false,
+      "bf16": "auto",
+      "tf32": false,
+      "gradient_checkpointing": true,
+      "logging_steps": 1,
+      "flash_attention": true,
+      "warmup_steps": 1,
+      "evals_per_epoch": 1,
+      "eval_max_new_tokens": 128,
+      "saves_per_epoch": 1,
+      "weight_decay": 0.0,
+      "special_tokens": {
+        "pad_token": "<|end_of_text|>"
+      }
+    }
+  }
+}

.runpod/src/train.py

@@ -0,0 +1,45 @@
+"""
+Runpod train entrypoint
+"""
+
+import asyncio
+
+
+async def train(config_path: str, gpu_id: str = "0", preprocess: bool = True):
+    """
+    Run preprocessing (if enabled) and training with the given config file
+    :param config_path: Path to the YAML config file
+    :param gpu_id: GPU ID to use (default: "0")
+    :param preprocess: Whether to run preprocessing (default: True)
+
+    """
+    # First check if preprocessing is needed
+    if preprocess:
+        # Preprocess command
+        preprocess_cmd = (
+            f"CUDA_VISIBLE_DEVICES={gpu_id} axolotl preprocess {config_path}"
+        )
+        process = await asyncio.create_subprocess_shell(
+            preprocess_cmd,
+            stdout=asyncio.subprocess.PIPE,
+            stderr=asyncio.subprocess.STDOUT,
+        )
+
+        if process.stdout is not None:
+            async for line in process.stdout:
+                yield f"Preprocessing: {line.decode().strip()}"
+        await process.wait()
+        yield "Preprocessing completed."
+    else:
+        yield "Skipping preprocessing step."
+
+    # Training command
+    train_cmd = f"axolotl train {config_path}"
+    process = await asyncio.create_subprocess_shell(
+        train_cmd, stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.STDOUT
+    )
+
+    if process.stdout is not None:
+        async for line in process.stdout:
+            yield f"Training: {line.decode().strip()}"
+    await process.wait()

.runpod/src/utils.py

@@ -0,0 +1,89 @@
+"""
+Runpod launcher utils
+"""
+
+import os
+
+import yaml
+
+
+def get_output_dir(run_id):
+    path = f"fine-tuning/{run_id}"
+    return path
+
+
+def make_valid_config(input_args):
+    """
+    Creates and saves updated config file, returns the path to the new config
+    :param input_args: dict of input args
+    :return: str, path to the updated config file
+    """
+    # Load default config
+    with open("config/config.yaml", "r", encoding="utf-8") as fin:
+        all_args = yaml.safe_load(fin)
+
+    if not input_args:
+        print("No args provided, using defaults")
+    else:
+        all_args.update(input_args)
+
+    # Create updated config path
+    updated_config_path = "config/updated_config.yaml"
+
+    # Save updated config to new file
+    with open(updated_config_path, "w", encoding="utf-8") as f:
+        yaml.dump(all_args, f)
+
+    return updated_config_path
+
+
+def set_config_env_vars(args: dict):
+    """
+    Convert API arguments into environment variables.
+    Handles nested dictionaries, lists, and special values.
+
+    Args:
+        args (dict): The arguments dictionary from the API request
+    """
+
+    def process_value(value):
+        """Convert Python values to string format for environment variables"""
+        if value is None:
+            return ""
+        if isinstance(value, bool):
+            return str(value).lower()
+        if isinstance(value, (list, dict)):
+            return str(value)
+        return str(value)
+
+    def set_env_vars(data, prefix=""):
+        """Recursively set environment variables from nested dictionary"""
+        for key, value in data.items():
+            env_key = prefix + key.upper()
+
+            # Handle special cases
+            if isinstance(value, dict):
+                # For nested dictionaries (like special_tokens)
+                set_env_vars(value, f"{env_key}_")
+            elif isinstance(value, list):
+                # Handle list of dictionaries (like datasets)
+                if value and isinstance(value[0], dict):
+                    for i, item in enumerate(value):
+                        set_env_vars(item, f"{env_key}_{i}_")
+                else:
+                    # For simple lists (like lora_target_modules)
+                    os.environ[env_key] = process_value(value)
+            else:
+                # Handle all other cases
+                os.environ[env_key] = process_value(value)
+
+    # Clear any existing related environment variables
+    # This prevents old values from persisting
+    for key in list(os.environ.keys()):
+        if key.startswith(
+            ("BASE_MODEL", "MODEL_TYPE", "TOKENIZER_TYPE", "DATASET", "LORA_", "WANDB_")
+        ):
+            del os.environ[key]
+
+    # Set new environment variables
+    set_env_vars(args)

.runpod/test-input.json

@@ -0,0 +1,86 @@
+{
+  "input": {
+    "name": "quick_smoke_test_sft",
+    "user_id": "user",
+    "model_id": "llama-test",
+    "run_id": "llama-test",
+    "credentials": {
+      "wandb_api_key": "",
+      "hf_token": ""
+    },
+    "args": {
+      "base_model": "HuggingFaceTB/SmolLM2-135M",
+      "model_type": "AutoModelForCausalLM",
+      "tokenizer_type": "AutoTokenizer",
+      "load_in_4bit": true,
+      "strict": false,
+      "datasets": [
+        {
+          "path": "mhenrichsen/alpaca_2k_test",
+          "type": "alpaca",
+          "split": "train[:10%]"
+        }
+      ],
+      "val_set_size": 0.02,
+      "output_dir": "./outputs/lora-out",
+      "sequence_len": 4096,
+      "sample_packing": true,
+      "eval_sample_packing": false,
+      "pad_to_sequence_len": true,
+      "adapter": "qlora",
+      "lora_r": 32,
+      "lora_alpha": 64,
+      "lora_dropout": 0.05,
+      "lora_target_linear": true,
+      "lora_modules_to_save": [
+        "embed_tokens",
+        "lm_head"
+      ],
+      "gradient_accumulation_steps": 2,
+      "micro_batch_size": 1,
+      "num_epochs": 1,
+      "optimizer": "adamw_torch_fused",
+      "lr_scheduler": "cosine",
+      "learning_rate": 0.0002,
+      "train_on_inputs": false,
+      "group_by_length": false,
+      "bf16": "auto",
+      "tf32": true,
+      "gradient_checkpointing": true,
+      "logging_steps": 1,
+      "flash_attention": true,
+      "warmup_steps": 1,
+      "evals_per_epoch": 1,
+      "eval_max_new_tokens": 128,
+      "saves_per_epoch": 1,
+      "weight_decay": 0.0,
+      "special_tokens": {
+        "pad_token": "<|endoftext|>"
+      },
+      "max_steps": 20
+    },
+    "timeout": 100000
+  },
+  "config": {
+    "gpuTypeId": "NVIDIA GeForce RTX 4090",
+    "gpuCount": 1,
+    "containerDiskInGb": 200,
+    "env": [
+      {
+        "key": "TOKENIZER",
+        "value": ""
+      },
+      {
+        "key": "DISABLE_LOG_STATS",
+        "value": "true"
+      }
+    ],
+    "allowedCudaVersions": [
+      "12.8",
+      "12.7",
+      "12.6",
+      "12.5",
+      "12.4"
+    ]
+  }
+}

.runpod/tests.json

@@ -0,0 +1,90 @@
+{
+  "tests": [
+    {
+      "name": "quick_smoke_test_sft",
+      "input": {
+        "user_id": "user",
+        "model_id": "llama-test",
+        "run_id": "llama-test",
+        "credentials": {
+          "wandb_api_key": "",
+          "hf_token": ""
+        },
+        "args": {
+          "base_model": "HuggingFaceTB/SmolLM2-135M",
+          "model_type": "AutoModelForCausalLM",
+          "tokenizer_type": "AutoTokenizer",
+          "load_in_4bit": true,
+          "strict": false,
+          "datasets": [
+            {
+              "path": "mhenrichsen/alpaca_2k_test",
+              "type": "alpaca",
+              "split": "train[:10%]"
+            }
+          ],
+          "val_set_size": 0.02,
+          "output_dir": "./outputs/lora-out",
+          "sequence_len": 4096,
+          "sample_packing": true,
+          "eval_sample_packing": false,
+          "pad_to_sequence_len": true,
+          "adapter": "qlora",
+          "lora_r": 32,
+          "lora_alpha": 64,
+          "lora_dropout": 0.05,
+          "lora_target_linear": true,
+          "lora_modules_to_save": [
+            "embed_tokens",
+            "lm_head"
+          ],
+          "gradient_accumulation_steps": 2,
+          "micro_batch_size": 1,
+          "num_epochs": 1,
+          "optimizer": "adamw_torch_fused",
+          "lr_scheduler": "cosine",
+          "learning_rate": 0.0002,
+          "train_on_inputs": false,
+          "group_by_length": false,
+          "bf16": "auto",
+          "tf32": true,
+          "gradient_checkpointing": true,
+          "logging_steps": 1,
+          "flash_attention": true,
+          "warmup_steps": 1,
+          "evals_per_epoch": 1,
+          "eval_max_new_tokens": 128,
+          "saves_per_epoch": 1,
+          "weight_decay": 0.0,
+          "special_tokens": {
+            "pad_token": "<|endoftext|>"
+          },
+          "max_steps": 20
+        }
+      },
+      "timeout": 100000
+    }
+  ],
+  "config": {
+    "gpuTypeId": "NVIDIA GeForce RTX 4090",
+    "gpuCount": 1,
+    "containerDiskInGb": 200,
+    "env": [
+      {
+        "key": "TOKENIZER",
+        "value": ""
+      },
+      {
+        "key": "DISABLE_LOG_STATS",
+        "value": "true"
+      }
+    ],
+    "allowedCudaVersions": [
+      "12.8",
+      "12.7",
+      "12.6",
+      "12.5",
+      "12.4"
+    ]
+  }
+}

_quarto.yml

@@ -48,8 +48,22 @@ quartodoc:
       contents:
         - core.trainers.base
         - core.trainers.trl
+        - core.trainers.mamba
+        - core.trainers.relora
         - core.trainers.dpo.trainer
         - core.trainers.grpo.trainer
+        - core.trainers.grpo.sampler
+        - core.trainers.utils
+    - title: Mixins
+      desc: Mixin classes for augmenting trainers
+      contents:
+        - core.trainers.mixins.optimizer
+        - core.trainers.mixins.rng_state_loader
+        - core.trainers.mixins.scheduler
+    - title: Context Managers
+      desc: Context managers for altering trainer behaviors
+      contents:
+        - utils.ctx_managers.sequence_parallel
     - title: Prompt Strategies
       desc: Prompt formatting strategies
       contents:
@@ -86,7 +100,7 @@ quartodoc:
         - kernels.swiglu
         - kernels.quantize
         - kernels.utils
-    - title: MonkeyPatches
+    - title: Monkey Patches
       desc: Runtime patches for model optimizations
       contents:
         - monkeypatch.llama_attn_hijack_flash
@@ -124,7 +138,8 @@ quartodoc:
         - utils.optimizers.adopt
         - utils.data.pretraining
         - utils.data.sft
-        - utils.gradient_checkpointing.unsloth
+        - utils.gradient_checkpointing.offload_cpu
+        - utils.gradient_checkpointing.offload_disk
     - title: Schemas
       desc: Pydantic data models for Axolotl config
       contents:

cicd/cicd.sh

@@ -18,7 +18,7 @@ pytest -v --durations=10 \
   --cov-append
 
 # Run patched tests excluding lora kernels with coverage append
-pytest -v --durations=10 \
+pytest --full-trace -vvv --durations=10 \
   --ignore=tests/e2e/patched/lora_kernels \
   /workspace/axolotl/tests/e2e/patched \
   --cov=axolotl \
@@ -52,4 +52,4 @@ pytest -v --durations=10 \
   --cov-append \
   --cov-report=xml:e2e-coverage.xml
 
-codecov upload-process -t $CODECOV_TOKEN -f e2e-coverage.xml -F e2e,pytorch-${PYTORCH_VERSION}
+codecov upload-process -t $CODECOV_TOKEN -f e2e-coverage.xml -F e2e,pytorch-${PYTORCH_VERSION} || true

cicd/cleanup.py

@@ -0,0 +1,19 @@
+"""Modal app to run axolotl GPU cleanup"""
+
+from .single_gpu import VOLUME_CONFIG, app, cicd_image, run_cmd
+
+
+@app.function(
+    image=cicd_image,
+    timeout=60 * 60,
+    cpu=8.0,
+    memory=131072,
+    volumes=VOLUME_CONFIG,
+)
+def cleanup():
+    run_cmd("./cicd/cleanup.sh", "/workspace/axolotl")
+
+
+@app.local_entrypoint()
+def main():
+    cleanup.remote()

cicd/cleanup.sh

@@ -0,0 +1,6 @@
+#!/bin/bash
+set -e
+
+# cleanup old cache files for datasets processing and intermediate mappings
+find /workspace/data/huggingface-cache/hub/datasets -name "cache-*" -type f -mtime +1 -exec rm {} \;
+find /workspace/data/huggingface-cache/hub/datasets -name "*.lock" -type f -mtime +1 -exec rm {} \;

cicd/e2e_tests.py

@@ -1,75 +1,12 @@
 """Modal app to run axolotl GPU tests"""
 
-# pylint: disable=duplicate-code
-
-import os
-import pathlib
-import tempfile
-
-import jinja2
-import modal
-from jinja2 import select_autoescape
-from modal import App, Image
-
-cicd_path = pathlib.Path(__file__).parent.resolve()
-
-template_loader = jinja2.FileSystemLoader(searchpath=cicd_path)
-template_env = jinja2.Environment(
-    loader=template_loader, autoescape=select_autoescape()
-)
-df_template = template_env.get_template("Dockerfile.jinja")
-
-df_args = {
-    "AXOLOTL_EXTRAS": os.environ.get("AXOLOTL_EXTRAS", ""),
-    "AXOLOTL_ARGS": os.environ.get("AXOLOTL_ARGS", ""),
-    "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.4.1"),
-    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-py3.11-cu121-2.4.1"),
-    "CUDA": os.environ.get("CUDA", "121"),
-    "GITHUB_REF": os.environ.get("GITHUB_REF", "refs/heads/main"),
-    "GITHUB_SHA": os.environ.get("GITHUB_SHA", ""),
-    "NIGHTLY_BUILD": os.environ.get("NIGHTLY_BUILD", ""),
-    "CODECOV_TOKEN": os.environ.get("CODECOV_TOKEN", ""),
-    "HF_HOME": "/workspace/data/huggingface-cache/hub",
-}
-
-dockerfile_contents = df_template.render(**df_args)
-
-temp_dir = tempfile.mkdtemp()
-with open(pathlib.Path(temp_dir) / "Dockerfile", "w", encoding="utf-8") as f:
-    f.write(dockerfile_contents)
-
-cicd_image = Image.from_dockerfile(
-    pathlib.Path(temp_dir) / "Dockerfile",
-    context_mount=None,
-    force_build=True,
-    gpu="A10G",
-).env(df_args)
-
-app = App("Axolotl CI/CD", secrets=[])
-
-hf_cache_volume = modal.Volume.from_name(
-    "axolotl-ci-hf-hub-cache", create_if_missing=True
-)
-VOLUME_CONFIG = {
-    "/workspace/data/huggingface-cache/hub": hf_cache_volume,
-}
-
-N_GPUS = int(os.environ.get("N_GPUS", 1))
-GPU_CONFIG = modal.gpu.L40S(count=N_GPUS)
-
-
-def run_cmd(cmd: str, run_folder: str):
-    import subprocess  # nosec
-
-    # Propagate errors from subprocess.
-    if exit_code := subprocess.call(cmd.split(), cwd=run_folder):  # nosec
-        exit(exit_code)  # pylint: disable=consider-using-sys-exit
+from .single_gpu import GPU_CONFIG, VOLUME_CONFIG, app, cicd_image, run_cmd
 
 
 @app.function(
     image=cicd_image,
     gpu=GPU_CONFIG,
-    timeout=60 * 60,
+    timeout=90 * 60,  # 90 min
     cpu=8.0,
     memory=131072,
     volumes=VOLUME_CONFIG,

cicd/multigpu.py

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

cicd/multigpu.sh

@@ -20,4 +20,4 @@ pytest -v  --durations=10 -n1 /workspace/axolotl/tests/e2e/multigpu/patched/ \
   --cov-report=xml:multigpu-coverage.xml
 
 # Upload coverage to Codecov
-codecov upload-process -t $CODECOV_TOKEN -f multigpu-coverage.xml -F multigpu,docker-tests,pytorch-${PYTORCH_VERSION}
+codecov upload-process -t "${CODECOV_TOKEN}" -f multigpu-coverage.xml -F multigpu,docker-tests,pytorch-${PYTORCH_VERSION} || true

cicd/single_gpu.py

@@ -0,0 +1,66 @@
+"""Modal app to run axolotl GPU tests"""
+
+# pylint: disable=duplicate-code
+
+import os
+import pathlib
+import tempfile
+
+import jinja2
+import modal
+from jinja2 import select_autoescape
+from modal import App, Image
+
+cicd_path = pathlib.Path(__file__).parent.resolve()
+
+template_loader = jinja2.FileSystemLoader(searchpath=cicd_path)
+template_env = jinja2.Environment(
+    loader=template_loader, autoescape=select_autoescape()
+)
+df_template = template_env.get_template("Dockerfile.jinja")
+
+df_args = {
+    "AXOLOTL_EXTRAS": os.environ.get("AXOLOTL_EXTRAS", ""),
+    "AXOLOTL_ARGS": os.environ.get("AXOLOTL_ARGS", ""),
+    "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.4.1"),
+    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-py3.11-cu121-2.4.1"),
+    "CUDA": os.environ.get("CUDA", "121"),
+    "GITHUB_REF": os.environ.get("GITHUB_REF", "refs/heads/main"),
+    "GITHUB_SHA": os.environ.get("GITHUB_SHA", ""),
+    "NIGHTLY_BUILD": os.environ.get("NIGHTLY_BUILD", ""),
+    "CODECOV_TOKEN": os.environ.get("CODECOV_TOKEN", ""),
+    "HF_HOME": "/workspace/data/huggingface-cache/hub",
+}
+
+dockerfile_contents = df_template.render(**df_args)
+
+temp_dir = tempfile.mkdtemp()
+with open(pathlib.Path(temp_dir) / "Dockerfile", "w", encoding="utf-8") as f:
+    f.write(dockerfile_contents)
+
+cicd_image = Image.from_dockerfile(
+    pathlib.Path(temp_dir) / "Dockerfile",
+    context_mount=None,
+    force_build=True,
+    gpu="A10G",
+).env(df_args)
+
+app = App("Axolotl CI/CD", secrets=[])
+
+hf_cache_volume = modal.Volume.from_name(
+    "axolotl-ci-hf-hub-cache", create_if_missing=True
+)
+VOLUME_CONFIG = {
+    "/workspace/data/huggingface-cache/hub": hf_cache_volume,
+}
+
+N_GPUS = int(os.environ.get("N_GPUS", 1))
+GPU_CONFIG = modal.gpu.L40S(count=N_GPUS)
+
+
+def run_cmd(cmd: str, run_folder: str):
+    import subprocess  # nosec
+
+    # Propagate errors from subprocess.
+    if exit_code := subprocess.call(cmd.split(), cwd=run_folder):  # nosec
+        exit(exit_code)  # pylint: disable=consider-using-sys-exit

codecov.yml

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

docs/config.qmd

@@ -32,6 +32,8 @@ tokenizer_legacy:
 resize_token_embeddings_to_32x:
 # Optional[bool] Whether to shrink the embeddings to len(tokenizer). By default, we won't shrink.
 shrink_embeddings:
+# Optional[bool] Don't upcast the embeddings to float32 when using PEFT. Useful for low-VRAM GPUs
+embeddings_skip_upcast:
 # Whether to load the model with randomly initialized weights. Useful for
 # pre-training a model from scratch or debugging purposes.
 random_init_weights:
@@ -73,11 +75,12 @@ load_in_8bit: true
 load_in_4bit:
 
 # Use CUDA bf16
-bf16: true # bool or 'full' for `bf16_full_eval`. require >=ampere
+bf16: true # bool or 'full' for `bf16_full_eval`, or 'auto' for automatic detection. require >=ampere
 # Use CUDA fp16
 fp16: true
 # Use CUDA tf32
 tf32: true # require >=ampere
+# Note: if bf16 is set to 'auto', and fp16 is set to true, we will prefer the explict fp16 setting
 
 # No AMP (automatic mixed precision)
 bfloat16: true # require >=ampere
@@ -154,6 +157,10 @@ datasets:
     # Key containing the messages (default: "messages")
     field_messages: messages
 
+    # Key containing the system message (default: "system")
+    # If the system message is not present in the dataset sample, it will be loaded from the field_system property.
+    field_system: system
+
     # Mapping of properties from the input dataset to the chat template.
     # (default: message_property_mappings={'role':'role', 'content':'content'})
     # If a property exists in the template but not in this mapping, the system will attempt
@@ -180,10 +187,14 @@ datasets:
     # adding a system turn with empty content.
     drop_system_message:
 
+    # Optional[bool]. (for Qwen3 template only) Whether to split the assistant content based on a reasoning trace inside delimited tags
+    # See example at `docs/dataset-formats/conversation.qmd`
+    split_thinking:
+
     # IMPORTANT: The following fields determine which parts of the conversation to train on.
     # Priority order: message_field_training > message_field_training_detail > train_on_inputs or role in roles_to_train
     # See examples at `docs/dataset-formats/conversation.qmd`
-    # Note: If the below 4 fields are set to empty, defaults to training only on the last message.
+    # Note: If the below 5 fields are empty, defaults to training only on the last message.
 
     # Optional[List[str]]. Roles to train on. The tokens from these roles will be considered for the loss.
     roles_to_train: ["assistant"]  # default
@@ -192,7 +203,13 @@ datasets:
     # - turn (default): train on the EOS token at the end of each trainable turn
     # - last: train on the last EOS token in the conversation
     # TIP: Please make sure that your `tokenizer.eos_token` is same as EOS/EOT token in template. Otherwise, set `eos_token` under `special_tokens`.
-    train_on_eos: last
+    train_on_eos: turn
+    # Optional[str]. Which EOT (End-of-Turn) tokens to train on in the conversation. Possible values are:
+    # - all: train on all EOT tokens
+    # - turn: train on the EOT token at the end of each trainable turn
+    # - last: train on the last EOT token in the conversation
+    # If not specified, defaults to the value of train_on_eos for backward compatibility.
+    train_on_eot:
     # The key in the message turn that indicates via boolean whether tokens of a turn should be considered for training. Useful to selectively train on certain turns besides the `roles_to_train`.
     message_field_training: training
     # The key in the message turn that contains the training details. Useful to selectively train on certain tokens in a turn.
@@ -275,8 +292,17 @@ process_reward_model:
 chat_template: tokenizer_default
 # custom jinja template for chat template. This will be only used if chat_template is set to `jinja` or `null` (in which case chat_template is automatically set to `jinja`). Default is null.
 chat_template_jinja: null
-# Changes the default system message. Currently only supports chatml.
-default_system_message: You are a helpful assistant. Please give a long and detailed answer.
+# Optional[List[str]]. Custom EOT (End-of-Turn) tokens to mask/unmask during training.
+# These tokens mark the boundaries between conversation turns.
+# For example: ["/INST", "</s>", "[/SYSTEM_PROMPT]"]
+# If not specified, defaults to just the model's eos_token.
+# This is useful for templates that use multiple delimiter tokens.
+eot_tokens:
+  # - "</s>"
+  # - "[/INST]"
+  # - "[/SYSTEM_PROMPT]"
+# Changes the default system message
+default_system_message: You are a helpful assistant. Please give a long and detailed answer. # Currently only supports chatml.
 # Axolotl attempts to save the dataset as an arrow after packing the data together so
 # subsequent training attempts load faster, relative path
 dataset_prepared_path: data/last_run_prepared
@@ -479,6 +505,7 @@ save_strategy: # Set to `"no"` to skip checkpoint saves, `"epoch"` at end of eac
 save_steps: # Leave empty to save at each epoch, integer for every N steps. float for fraction of total steps
 saves_per_epoch: # number of times per epoch to save a checkpoint, mutually exclusive with save_steps
 save_total_limit: # Checkpoints saved at a time
+save_only_model: # Save only the model weights, skipping the optimizer. Using this means you can't resume from checkpoints.
 # Maximum number of iterations to train for. It precedes num_epochs which means that
 # if both are set, num_epochs will not be guaranteed.
 # e.g., when 1 epoch is 1000 steps => `num_epochs: 2` and `max_steps: 100` will train for 100 steps
@@ -512,7 +539,7 @@ train_on_inputs: false
 # Note that training loss may have an oscillating pattern with this enabled.
 group_by_length: false
 
-# Whether to use gradient checkpointing. Available options are: true, false, "offload".
+# Whether to use gradient checkpointing. Available options are: true, false, "offload", "offload_disk".
 # https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing
 gradient_checkpointing: false
 # additional kwargs to pass to the trainer for gradient checkpointing
@@ -524,7 +551,7 @@ gradient_checkpointing: false
 early_stopping_patience: 3
 
 # Specify a scheduler and kwargs to use with the optimizer
-lr_scheduler: # 'one_cycle' | 'rex' | 'log_sweep' | empty for cosine
+lr_scheduler: # 'one_cycle' | 'rex' | 'log_sweep' | 'linear' | 'cosine_with_restarts' | 'polynomial' | 'constant' | 'constant_with_warmup' | 'inverse_sqrt' | 'reduce_lr_on_plateau' | 'cosine_with_min_lr' | 'warmup_stable_decay' | empty for cosine
 lr_scheduler_kwargs:
 cosine_min_lr_ratio: # decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of peak lr
 cosine_constant_lr_ratio: # freeze lr at some percentage of the step, e.g. cosine_constant_lr_ratio=0.8 means start cosine_min_lr at 80% of training step (https://arxiv.org/pdf/2308.04014.pdf)
@@ -586,6 +613,7 @@ lr_div_factor: # Learning rate div factor
 # - optimi_adamw
 # - ao_adamw_8bit
 # - ao_adamw_fp8
+# - came_pytorch
 optimizer:
 # Dictionary of arguments to pass to the optimizer
 optim_args:
@@ -605,7 +633,9 @@ weight_decay:
 # adamw hyperparams
 adam_beta1:
 adam_beta2:
+adam_beta3:  # only used for CAME Optimizer
 adam_epsilon:
+adam_epsilon2:  # only used for CAME Optimizer
 # Gradient clipping max norm
 max_grad_norm:
 
@@ -661,8 +691,10 @@ special_tokens:
   # unk_token: "<unk>"
   # pad_token: "[PAD]"
 
-# Add extra tokens.
+# Optional[list[str]]. Add extra tokens to the tokenizer.
 tokens:
+  # - "<|startoftext|>"
+  # - "<|endoftext|>"
 
 # Mapping token_id to new_token_string to override reserved added_tokens in the tokenizer.
 # Only works for tokens that are not part of the base vocab (aka are added_tokens).

docs/custom_integrations.qmd

@@ -49,7 +49,8 @@ sections = [
     ("Knowledge Distillation (KD)", "kd"),
     ("Liger Kernels", "liger"),
     ("Language Model Evaluation Harness (LM Eval)", "lm_eval"),
-    ("Spectrum", "spectrum")
+    ("Spectrum", "spectrum"),
+    ("LLMCompressor", "llm_compressor")
 ]
 
 for section_name, folder_name in sections:

docs/dataset-formats/conversation.qmd

@@ -4,18 +4,6 @@ description: Conversation format for supervised fine-tuning.
 order: 3
 ---
 
-## sharegpt
-
-::: {.callout-important}
-ShareGPT is deprecated!. Please see [chat_template](#chat_template) section below.
-:::
-
-## pygmalion
-
-```{.json filename="data.jsonl"}
-{"conversations": [{"role": "...", "value": "..."}]}
-```
-
 ## chat_template
 
 Chat Template strategy uses a jinja2 template that converts a list of messages into a prompt. Support using tokenizer's template, a supported template, or custom jinja2.
@@ -64,7 +52,7 @@ We recommend checking the below examples for other usecases.
 
 ### Examples
 
-1. Using the default chat template in the tokenizer_config.json on OpenAI messages format, training on only last message.
+1. (Legacy) Using the default chat template in the tokenizer_config.json on OpenAI messages format, training on only last message.
 
 ```yaml
 datasets:
@@ -109,10 +97,55 @@ datasets:
 ```
 
 ::: {.callout-important}
-Please make sure that your `tokenizer.eos_token` is same as EOS/EOT token in template. Otherwise, set `eos_token` under `special_tokens`.
+Please make sure that your `tokenizer.eos_token` is same as EOS (End-of-Sequence) token in template. Otherwise, set `eos_token` under `special_tokens: `.
 :::
 
-5. (Advanced) Using fine-grained control over tokens and turns to train in a conversation
+5. If you are using a template that has a different EOT (End-of-Turn) token from EOS token or multiple EOT tokens (like Mistral V7 Tekken), set the `eot_tokens: ` config. The handling of EOT tokens follows `train_on_eos: ` which defaults to turn.
+
+```yaml
+eot_tokens:
+  - "[/INST]"
+  # - "[/SYSTEM_PROMPT]"
+
+datasets:
+  - path: ...
+    type: chat_template
+
+    # optional
+    train_on_eot: turn  # defaults read from train_on_eos (which defaults to turn)
+```
+
+::: {.callout-tip}
+See [config documentation](../config.qmd) for detailed explanations of "turn", "last", and "all" options for training on tokens.
+:::
+
+::: {.callout-note}
+Using `eot_tokens` requires each token that exists in `chat_template` to be a single token in the tokenizer. Otherwise, the tokenizer will split the token and cause unexpected behavior.
+
+You can add those tokens as new tokens under `tokens: ` or (recommended) override unused added_tokens via `added_tokens_overrides: `. See [config](../config.qmd) for more details.
+:::
+
+6. Continuing from the previous example, if you want to train on all EOT token trainable turns but only last EOS token, set `train_on_eos: last`.
+
+```yaml
+eot_tokens:
+  - "[/INST]"
+  # ...
+
+datasets:
+  - path: ...
+    type: chat_template
+
+    train_on_eos: last
+    train_on_eot: turn
+```
+
+::: {.callout-tip}
+If EOS token only appears at the end of a prompt, `train_on_eos: last` is equivalent to `train_on_eos: turn`. Therefore, generally, you can leave them to their defaults and omit them.
+:::
+
+
+7. (Advanced) Using fine-grained control over tokens and turns to train in a conversation
 
 For a data sample that looks like:
 
@@ -162,3 +195,43 @@ datasets:
 ::: {.callout-tip}
 It is not necessary to set both `message_field_training` and `message_field_training_detail` at once.
 :::
+
+8. (For Qwen3 template only) Enable reasoning split, where the reasoning is split from the content and passed as a separate field into the template.
+
+```yaml
+datasets:
+  - path: ...
+    type: chat_template
+    chat_template: qwen3
+    split_thinking: true
+```
+
+For example, a content can look like:
+
+```json
+{
+  "content": "<think>Some thinking outputs</think>Output after thinking."
+}
+```
+
+After split, it will look like:
+
+```json
+{
+  "reasoning_content": "Some thinking outputs",
+  "content": "Output after thinking..."
+}
+```
+
+
+## sharegpt
+
+::: {.callout-important}
+ShareGPT is deprecated!. Please see [chat_template](#chat_template) section.
+:::
+
+## pygmalion
+
+```{.json filename="data.jsonl"}
+{"conversations": [{"role": "...", "value": "..."}]}
+```

docs/docker.qmd

@@ -8,6 +8,10 @@ format:
 
 This section describes the different Docker images that are released by AxolotlAI at [Docker Hub](https://hub.docker.com/u/axolotlai).
 
+::: {.callout-important}
+For Blackwell GPUs, please use the tags with Pytorch 2.7.0 and CUDA 12.8.
+:::
+
 ## Base
 
 The base image is the most minimal image that can install Axolotl. It is based on the `nvidia/cuda` image. It includes python, torch, git, git-lfs, awscli, pydantic, and more.
@@ -28,6 +32,8 @@ main-base-py{python_version}-cu{cuda_version}-{pytorch_version}
 
 Tags examples:
 
+- `main-base-py3.11-cu128-2.7.0`
+- `main-base-py3.11-cu126-2.7.0`
 - `main-base-py3.11-cu124-2.6.0`
 - `main-base-py3.11-cu124-2.5.1`
 - `main-base-py3.11-cu124-2.4.1`
@@ -50,7 +56,7 @@ Link: [Docker Hub](https://hub.docker.com/r/axolotlai/axolotl)
 # on push to main
 main-py{python_version}-cu{cuda_version}-{pytorch_version}
 
-# latest main (currently torch 2.5.1, python 3.11, cuda 12.4)
+# latest main (currently torch 2.6.0, python 3.11, cuda 12.4)
 main-latest
 
 # nightly build
@@ -68,6 +74,7 @@ There may be some extra tags appended to the image, like `-vllm` which installs
 
 Tags examples:
 
+- `main-py3.11-cu126-2.7.0`
 - `main-py3.11-cu124-2.6.0`
 - `main-py3.11-cu124-2.5.1`
 - `main-py3.11-cu124-2.4.1`

docs/faq.qmd

@@ -73,10 +73,40 @@ description: Frequently asked questions
 
 > A: This is likely an empty turn.
 
-**Q: The EOS/EOT token is incorrectly being masked or not being masked.**
+**Q: The EOS token is incorrectly being masked or not being masked / `EOS token __ not found in chat template`.**
 
-> A: This is because of the mismatch between `tokenizer.eos_token` and EOS/EOT token in template. Please make sure to set `eos_token` under `special_tokens` to the same EOS/EOT token as in template.
+> A: There can be two reasons:
+
+> 1. This is because of the mismatch between `tokenizer.eos_token` and EOS token in template. Please make sure to set `eos_token: ` under `special_tokens: ` to the same EOS token as in template.
+
+> 2. The EOS token is not in the template. Please check if your template is correct. As an example, `phi_35` template does not use its dedicated EOS token `<|endoftext|>` at the end.
 
 **Q: "`chat_template` choice is `tokenizer_default` but tokenizer's `chat_template` is null. Please add a `chat_template` in tokenizer config"**
 
 > A: This is because the tokenizer does not have a chat template. Please add a chat template in the tokenizer config. See [chat_template](dataset-formats/conversation.qmd#chat-template) for more details.
+
+**Q: The EOT token(s) are incorrectly being masked or not being masked / `EOT token __ not found in chat template`.**
+
+> A: There can be two reasons:
+
+> 1. The EOT token is different from the EOS token and was not specified under `eot_tokens: `. Please set `eot_tokens: ` to the same EOT token(s) as in template.
+
+> 2. There is more than one EOT token per turn in the template. Please raise an issue with examples as we recognize this as an edge case.
+
+**Q: `EOT token encoding failed. Please check if the token is valid and can be encoded.`**
+
+> A: There could be some issue with the tokenizer or unicode encoding. Please raise an issue with examples with the EOT token & tokenizer causing the issue.
+
+**Q: `EOT token __ is encoded as multiple tokens.`**
+
+> A: This is because the EOT token is encoded as multiple tokens which can cause unexpected behavior. Please add it under `tokens: ` or (recommended) override unused added_tokens via `added_tokens_overrides: `.
+
+**Q: `Conflict between train_on_eos and train_on_eot. eos_token is in eot_tokens and train_on_eos != train_on_eot`**
+
+> A: This is because the EOS token is in the `eot_tokens: ` while mismatch between `train_on_eos: ` and `train_on_eot: `. This will cause one to override the other. Please ensure that `train_on_eos: ` and `train_on_eot: ` are the same or remove the EOS token from `eot_tokens: `.
+
+**Q: If `eot_tokens: ` is not provided, what happens?**
+
+> A: If `eot_tokens: ` is not provided, the default behavior is the same as before. EOS tokens used to delimit turns are masked/unmasked depending on whether the turn is trainable.
+
+> Internally, `eot_tokens: tokenizer.eos_token` and `train_on_eot: train_on_eos` (which defaults to `turn`). This transition helps clarify the naming and behavior of EOT/EOS tokens.

docs/getting-started.qmd

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

docs/installation.qmd

@@ -25,6 +25,10 @@ Please make sure to have Pytorch installed before installing Axolotl in your loc
 Follow the instructions at: [https://pytorch.org/get-started/locally/](https://pytorch.org/get-started/locally/)
 :::
 
+::: {.callout-important}
+For Blackwell GPUs, please use Pytorch 2.7.0 and CUDA 12.8.
+:::
+
 ### PyPI Installation (Recommended) {#sec-pypi}
 
 ```{.bash}
@@ -72,6 +76,10 @@ docker run --privileged --gpus '"all"' --shm-size 10g --rm -it \
 ```
 :::
 
+::: {.callout-important}
+For Blackwell GPUs, please use `axolotlai/axolotl:main-py3.11-cu128-2.7.0` or the cloud variant `axolotlai/axolotl-cloud:main-py3.11-cu128-2.7.0`.
+:::
+
 Please refer to the [Docker documentation](docker.qmd) for more information on the different Docker images that are available.
 
 ## Cloud Environments {#sec-cloud}

docs/multi-gpu.qmd

@@ -87,20 +87,7 @@ We support sequence parallelism (SP) via the
 allows one to split up sequences across GPUs, which is useful in the event that a
 single sequence causes OOM errors during model training.
 
-First, install `ring-flash-attn`, recommended via `pip install axolotl[ring-flash-attn]`,
-or from source with `pip install .[ring-flash-attn]`.
-
-Your Axolotl YAML config should contain the following lines:
-
-```{.yaml}
-sequence_parallel_degree: 4  # Split each sequence into 4 parts, one per GPU
-flash_attention: true  # Required with sequence parallelism
-
-# Optional; strides across the key dimension. Larger values use more memory but will make training faster.
-heads_k_stride: 1
-```
-
-See our [dedicated guide](sequence_parallelism.qmd) for more details.
+See our [dedicated guide](sequence_parallelism.qmd) for more information.
 
 ### FSDP + QLoRA {#sec-fsdp-qlora}
 

docs/multimodal.qmd

@@ -164,7 +164,7 @@ Here is an example of a multi-modal dataset:
         {
             "role": "user",
             "content": [
-                {"type": "image", "image": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"},
+                {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"},
                 {"type": "text", "text": "Describe this image in detail."}
             ]
         },

docs/rlhf.qmd

@@ -502,9 +502,7 @@ The input format is a simple JSON input with customizable fields based on the ab
 Check out our [GRPO cookbook](https://github.com/axolotl-ai-cloud/axolotl-cookbook/tree/main/grpo#training-an-r1-style-large-language-model-using-grpo).
 :::
 
-If you have multiple GPUs available, we reccomend using `vLLM` with the `GRPOTrainer` to significantly speedup trajectory generation during training.
-First, launch a `vLLM` server using `trl vllm-serve` - you may use a config file or CLI overrides to configure your vLLM server. In this example, we're
-using 4 GPUs - 2 for training, and 2 for vLLM:
+In the latest GRPO implementation, `vLLM` is used to significantly speedup trajectory generation during training. In this example, we're using 4 GPUs - 2 for training, and 2 for vLLM:
 
 ::: {.callout-important}
 Make sure you've installed the correct version of vLLM by including it as an extra when installing axolotl, e.g. `pip install axolotl[vllm]`.
@@ -539,6 +537,10 @@ Your `vLLM` instance will now attempt to spin up, and it's time to kick off trai
 CUDA_VISIBLE_DEVICES=0,1 axolotl train grpo.yaml --num-processes 2
 ```
 
+::: {.callout-note}
+Due to TRL's implementation with vLLM, the vLLM instance must use the last N GPUs instead of the first N GPUs. This is why in the example above, we use `CUDA_VISIBLE_DEVICES=2,3` for the vLLM instance.
+:::
+
 #### Reward functions
 
 GRPO uses custom reward functions and transformations. Please have them ready locally.

docs/sequence_parallelism.qmd

@@ -3,8 +3,6 @@ title: Sequence Parallelism
 description: Train with long sequences split across multiple GPUs.
 ---
 
-# Sequence Parallelism
-
 Sequence parallelism is a technique that splits sequences across multiple GPUs,
 allowing you to train with very long sequences that wouldn't fit on a single GPU. Each
 GPU processes a different portion of the sequence, and the results are aggregated
@@ -27,7 +25,7 @@ To enable sequence parallelism, add the following to your configuration file:
 sequence_parallel_degree: 4  # Split sequences across 4 GPUs
 # Optional; strides across the key dimension. Larger values use more memory but should make training faster.
 heads_k_stride: 1
-# Optional; one of "varlen_llama3", "batch_ring", "batch_zigzag", "batch_stripe". Defaults to
+# Optional; one of "varlen_llama3" or "batch_ring". Defaults to
 # "varlen_llama3" when `sample_packing: true`, and "batch_ring" otherwise.
 ring_attn_func:
 ```
@@ -43,7 +41,7 @@ When sequence parallelism is enabled:
 
 1. Each sequence is divided into equal chunks across the GPUs in a sequence parallel group
 2. The data collator handles the chunking of input_ids, attention_mask, labels, and position_ids
-3. Position IDs are adjusted to maintain proper relative positions, especially for packed sequences
+3. Position IDs are adjusted to maintain proper relative positions
 4. The trainer uses special ring communication patterns for attention operations
 
 ## Requirements
@@ -69,9 +67,11 @@ sequence_len: 8192
 ...
 
 sequence_parallel_degree: 4  # Split each sequence into 4 parts, one per GPU
-flash_attention: true  # Required with sequence parallelism
 # Optional; strides across the key dimension. Larger values use more memory but should make training faster.
 heads_k_stride: 1
+# Optional; one of "varlen_llama3" or "batch_ring". Defaults to
+# "varlen_llama3" when `sample_packing: true`, and "batch_ring" otherwise.
+ring_attn_func:
 
 ...
 ```

examples/llama-3/sparse-finetuning.yaml

@@ -0,0 +1,77 @@
+base_model: neuralmagic/Sparse-Llama-3.1-8B-2of4
+
+plugins:
+  - axolotl.integrations.llm_compressor.LLMCompressorPlugin
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.05
+output_dir: ./outputs/out
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+eval_sample_packing: false
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 8
+micro_batch_size: 1
+num_epochs: 1
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+early_stopping_patience:
+resume_from_checkpoint:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 100
+evals_per_epoch: 2
+eval_table_size:
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  pad_token: <|end_of_text|>
+
+llmcompressor:
+  recipe:
+    finetuning_stage:
+      finetuning_modifiers:
+        ConstantPruningModifier:
+          targets: [
+            're:.*q_proj.weight',
+            're:.*k_proj.weight',
+            're:.*v_proj.weight',
+            're:.*o_proj.weight',
+            're:.*gate_proj.weight',
+            're:.*up_proj.weight',
+            're:.*down_proj.weight',
+          ]
+          start: 0
+  save_compressed: true

examples/llama-4/README.md

@@ -34,3 +34,5 @@ We provide a script to delinearize Llama 4 linearized models into regular Huggin
 ```bash
 axolotl delinearize-llama4 --model path/to/model_dir --output path/to/output_dir
 ```
+
+Note: This only works with the non-quantized linearized model. If you have an adapter, merge it with the *non-quantized linearized* model before delinearizing.

examples/llama-4/scout-qlora-single-h100-flex.yaml

@@ -10,7 +10,6 @@ plugins:
 liger_glu_activation: true
 liger_rms_norm: true
 liger_layer_norm: true
-cut_cross_entropy: true
 
 llama4_linearized_experts: true  # needed with custom linearized experts model
 load_in_4bit: true

examples/mistral/devstral-small-2505.yml

@@ -0,0 +1,48 @@
+base_model: mistralai/Devstral-Small-2505
+processor_type: AutoProcessor
+
+# 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
+
+chat_template: mistral_v7_tekken
+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
+
+sequence_len: 2048
+pad_to_sequence_len: false
+
+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: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+logging_steps: 1
+flash_attention: false
+eager_attention:
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/orpheus/README.md

@@ -0,0 +1,341 @@
+# Finetuning LLMs to output audio
+
+In this example, we finetune Orpcanopylabs/orpheus-tts-0.1-pretrained (a LLaMA 3.2 3b model) to output audio.
+
+The `finetune.yml` withe current settings will run on any Nvidia GPU with 45GB VRAM or more. If you adjust the batch size it can easily run on any GPU under 24GB.
+
+## Dataset pre-processing for pre-training
+If you are adding another voice in English, please jump ahead to finetuning pre-processing.
+
+For this to work, we need to preprocess our dataset. Since we are expecting to output audio, we will need to add tokens to the tokenizer.
+
+Using this code, it will download the SNAC model and add the correct tokens and upload the final dataset.
+
+```python
+import torch
+from snac import SNAC
+from datasets import load_dataset
+from huggingface_hub import snapshot_download
+from datasets import load_dataset
+import random
+import torchaudio.transforms as T
+from transformers import AutoTokenizer
+import os
+
+my_original_dataset_name = "<huggingface-id-of-dataset-that-we-want-to-preprocess>"
+name_to_push_dataset_to = "<huggingface-id-of-where-to-save-dataset>"
+
+dsn = my_original_dataset_name
+
+snapshot_download(
+    repo_id=dsn,
+    repo_type="dataset",
+    revision="main",
+    max_workers=64,
+)
+
+
+ds = load_dataset(dsn, split="train")
+ds_sample_rate = ds[0]["audio"]["sampling_rate"]
+
+model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz")
+model = model.to("mps")
+
+def tokenise_audio(waveform):
+  waveform = torch.from_numpy(waveform).unsqueeze(0)
+  waveform = waveform.to(dtype=torch.float32)
+  resample_transform = T.Resample(orig_freq=ds_sample_rate, new_freq=24000)
+  waveform = resample_transform(waveform)
+
+  waveform = waveform.unsqueeze(0).to("cuda")
+
+  #generate the codes from snac
+  with torch.inference_mode():
+    codes = model.encode(waveform)
+
+  all_codes = []
+  for i in range(codes[0].shape[1]):
+    all_codes.append(codes[0][0][i].item()+128266)
+    all_codes.append(codes[1][0][2*i].item()+128266+4096)
+    all_codes.append(codes[2][0][4*i].item()+128266+(2*4096))
+    all_codes.append(codes[2][0][(4*i)+1].item()+128266+(3*4096))
+    all_codes.append(codes[1][0][(2*i)+1].item()+128266+(4*4096))
+    all_codes.append(codes[2][0][(4*i)+2].item()+128266+(5*4096))
+    all_codes.append(codes[2][0][(4*i)+3].item()+128266+(6*4096))
+
+
+  return all_codes
+
+def add_codes(example):
+    # Always initialize codes_list to None
+    codes_list = None
+
+    try:
+        answer_audio = example.get("audio")
+        # If there's a valid audio array, tokenise it
+        if answer_audio and "array" in answer_audio:
+            audio_array = answer_audio["array"]
+            codes_list = tokenise_audio(audio_array)
+    except Exception as e:
+        print(f"Skipping row due to error: {e}")
+        # Keep codes_list as None if we fail
+    example["codes_list"] = codes_list
+
+    return example
+
+ds = ds.map(add_codes, remove_columns=["audio"])
+
+#@title Load Tokenizer
+tokeniser_length = 128256
+start_of_text = 128000
+end_of_text = 128009
+
+start_of_speech = tokeniser_length + 1
+end_of_speech = tokeniser_length + 2
+
+start_of_human = tokeniser_length + 3
+end_of_human = tokeniser_length + 4
+
+start_of_ai = tokeniser_length + 5
+end_of_ai =  tokeniser_length + 6
+pad_token = tokeniser_length + 7
+
+audio_tokens_start = tokeniser_length + 10
+
+tokenizer_name = "canopylabs/orpheus-3b-0.1-pretrained"
+
+
+tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
+num_proc = os.cpu_count() - 2
+
+ds = ds.filter(lambda x: x["codes_list"] is not None)
+ds = ds.filter(lambda x: len(x["codes_list"]) > 0)
+
+#@title Create Input Ids
+def remove_duplicate_frames(example):
+    vals = example["codes_list"]
+    if len(vals) % 7 != 0:
+        raise ValueError("Input list length must be divisible by 7")
+
+    result = vals[:7]
+
+    removed_frames = 0
+
+    for i in range(7, len(vals), 7):
+        current_first = vals[i]
+        previous_first = result[-7]
+
+        if current_first != previous_first:
+            result.extend(vals[i:i+7])
+        else:
+            removed_frames += 1
+
+    example["codes_list"] = result
+
+    return example
+
+ds = ds.map(remove_duplicate_frames, num_proc=num_proc)
+
+
+def create_input_ids(example):
+    text_ids = tokenizer.encode({example['text']},  add_special_tokens=True)
+    text_ids.append(end_of_text)
+    example["text_tokens"] = text_ids
+    input_ids = (
+        [start_of_human]
+        + example["text_tokens"]
+        + [end_of_human]
+        + [start_of_ai]
+        + [start_of_speech]
+        + example["codes_list"]
+        + [end_of_speech]
+        + [end_of_ai]
+    )
+    example["input_ids"] = input_ids
+    example["labels"] = input_ids
+    example["attention_mask"] = [1] * len(input_ids)
+
+    return example
+
+ds = ds.map(create_input_ids, num_proc=num_proc, remove_columns=["text", "codes_list"])
+
+#@title Remove unnecessary columns
+columns_to_keep = ["input_ids", "labels", "attention_mask"]
+columns_to_remove = [col for col in ds.column_names if col not in columns_to_keep]
+
+ds = ds.remove_columns(columns_to_remove)
+
+ds.push_to_hub(name_to_push_dataset_to)
+```
+
+
+## Finetune pre-processing
+Use this code to add a new voice.
+
+```python
+import torch
+from snac import SNAC
+from datasets import load_dataset
+from huggingface_hub import snapshot_download
+from datasets import load_dataset
+import random
+import torchaudio.transforms as T
+from transformers import AutoTokenizer
+import os
+
+my_original_dataset_name = "<huggingface-id-of-dataset-that-we-want-to-preprocess>"
+name_to_push_dataset_to = "<huggingface-id-of-where-to-save-dataset>"
+
+dsn = my_original_dataset_name
+
+snapshot_download(
+    repo_id=dsn,
+    repo_type="dataset",
+    revision="main",
+    max_workers=64,
+)
+
+
+ds = load_dataset(dsn, split="train")
+ds_sample_rate = ds[0]["audio"]["sampling_rate"]
+
+model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz")
+model = model.to("mps")
+
+def tokenise_audio(waveform):
+  waveform = torch.from_numpy(waveform).unsqueeze(0)
+  waveform = waveform.to(dtype=torch.float32)
+  resample_transform = T.Resample(orig_freq=ds_sample_rate, new_freq=24000)
+  waveform = resample_transform(waveform)
+
+  waveform = waveform.unsqueeze(0).to("cuda")
+
+  #generate the codes from snac
+  with torch.inference_mode():
+    codes = model.encode(waveform)
+
+  all_codes = []
+  for i in range(codes[0].shape[1]):
+    all_codes.append(codes[0][0][i].item()+128266)
+    all_codes.append(codes[1][0][2*i].item()+128266+4096)
+    all_codes.append(codes[2][0][4*i].item()+128266+(2*4096))
+    all_codes.append(codes[2][0][(4*i)+1].item()+128266+(3*4096))
+    all_codes.append(codes[1][0][(2*i)+1].item()+128266+(4*4096))
+    all_codes.append(codes[2][0][(4*i)+2].item()+128266+(5*4096))
+    all_codes.append(codes[2][0][(4*i)+3].item()+128266+(6*4096))
+
+
+  return all_codes
+
+def add_codes(example):
+    # Always initialize codes_list to None
+    codes_list = None
+
+    try:
+        answer_audio = example.get("audio")
+        # If there's a valid audio array, tokenise it
+        if answer_audio and "array" in answer_audio:
+            audio_array = answer_audio["array"]
+            codes_list = tokenise_audio(audio_array)
+    except Exception as e:
+        print(f"Skipping row due to error: {e}")
+        # Keep codes_list as None if we fail
+    example["codes_list"] = codes_list
+
+    return example
+
+ds = ds.map(add_codes, remove_columns=["audio"])
+
+#@title Load Tokenizer
+tokeniser_length = 128256
+start_of_text = 128000
+end_of_text = 128009
+
+start_of_speech = tokeniser_length + 1
+end_of_speech = tokeniser_length + 2
+
+start_of_human = tokeniser_length + 3
+end_of_human = tokeniser_length + 4
+
+start_of_ai = tokeniser_length + 5
+end_of_ai =  tokeniser_length + 6
+pad_token = tokeniser_length + 7
+
+audio_tokens_start = tokeniser_length + 10
+
+tokenizer_name = "canopylabs/orpheus-3b-0.1-pretrained"
+
+
+tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
+num_proc = os.cpu_count() - 2
+
+ds = ds.filter(lambda x: x["codes_list"] is not None)
+ds = ds.filter(lambda x: len(x["codes_list"]) > 0)
+
+#@title Create Input Ids
+def remove_duplicate_frames(example):
+    vals = example["codes_list"]
+    if len(vals) % 7 != 0:
+        raise ValueError("Input list length must be divisible by 7")
+
+    result = vals[:7]
+
+    removed_frames = 0
+
+    for i in range(7, len(vals), 7):
+        current_first = vals[i]
+        previous_first = result[-7]
+
+        if current_first != previous_first:
+            result.extend(vals[i:i+7])
+        else:
+            removed_frames += 1
+
+    example["codes_list"] = result
+
+    return example
+
+ds = ds.map(remove_duplicate_frames, num_proc=num_proc)
+
+tok_info = '''*** HERE you can modify the text prompt
+i.e. if you wanted a multispeaker model like canopylabs/orpheus-3b-0.1-ft, you can pass:
+f"{example["source"]}:  {example["text"]}", as is passed.
+'''
+print(tok_info)
+
+def create_input_ids(example):
+    text_ids = tokenizer.encode(f"{example['speaker_id']}: {example['text']}",  add_special_tokens=True)
+    text_ids.append(end_of_text)
+    example["text_tokens"] = text_ids
+    input_ids = (
+        [start_of_human]
+        + example["text_tokens"]
+        + [end_of_human]
+        + [start_of_ai]
+        + [start_of_speech]
+        + example["codes_list"]
+        + [end_of_speech]
+        + [end_of_ai]
+    )
+    example["input_ids"] = input_ids
+    example["labels"] = input_ids
+    example["attention_mask"] = [1] * len(input_ids)
+
+    return example
+
+ds = ds.map(create_input_ids, num_proc=num_proc, remove_columns=["text", "codes_list"])
+
+#@title Remove unnecessary columns
+columns_to_keep = ["input_ids", "labels", "attention_mask"]
+columns_to_remove = [col for col in ds.column_names if col not in columns_to_keep]
+
+ds = ds.remove_columns(columns_to_remove)
+
+ds.push_to_hub(name_to_push_dataset_to)
+```
+
+## Training
+After preprocessing is done, fill out the blanks in finetune.yml and simply run `axolotl train finetune.yml`
+
+## Inference
+For inference, please refer to the original [orpheus github](https://github.com/canopyai/Orpheus-TTS/tree/main).

examples/orpheus/finetune.yml

@@ -0,0 +1,52 @@
+base_model: canopylabs/orpheus-3b-0.1-pretrained
+
+hub_model_id: <your-hub-model-id>
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+liger_fused_linear_cross_entropy: true
+
+datasets:
+  - path: <your-hf-dataset-id>
+    type:  # leave empty to load pre-tokenized
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./outputs/out
+
+sequence_len: 8192
+sample_packing: true
+pad_to_sequence_len: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 8
+micro_batch_size: 4
+num_epochs: 3
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 20
+evals_per_epoch: 5
+saves_per_epoch: 5
+weight_decay: 0.05
+
+special_tokens:
+  pad_token: <custom_token_7>

examples/qwen2/dpo.yaml

@@ -2,7 +2,6 @@ base_model: Qwen/Qwen2.5-0.5B
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 
-
 chat_template: qwen_25
 rl: dpo
 datasets:

examples/qwen3/32b-qlora.yaml

@@ -0,0 +1,69 @@
+base_model: Qwen/Qwen3-32B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+strict: false
+
+chat_template: qwen3
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+val_set_size: 0.0
+output_dir: ./outputs/out
+dataset_prepared_path: last_run_prepared
+
+sequence_len: 2048
+sample_packing: true
+eval_sample_packing: true
+pad_to_sequence_len: true
+
+load_in_4bit: true
+adapter: qlora
+lora_r: 16
+lora_alpha: 32
+lora_target_modules:
+  - q_proj
+  - k_proj
+  - v_proj
+  - o_proj
+  - down_proj
+  - up_proj
+lora_mlp_kernel: true
+lora_qkv_kernel: true
+lora_o_kernel: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: offload
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/qwen3/qlora-fsdp.yaml

@@ -0,0 +1,68 @@
+base_model: Qwen/Qwen3-8B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/out
+
+sequence_len: 2048
+sample_packing: true
+eval_sample_packing: true
+pad_to_sequence_len: true
+
+adapter: qlora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_fused
+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: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+fsdp:
+  - full_shard
+  - auto_wrap
+fsdp_config:
+  fsdp_limit_all_gathers: true
+  fsdp_sync_module_states: true
+  fsdp_offload_params: true
+  fsdp_use_orig_params: false
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: Qwen3DecoderLayer
+  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_sharding_strategy: FULL_SHARD
+special_tokens:

requirements.txt

@@ -6,19 +6,20 @@ triton>=3.0.0
 mamba-ssm==1.2.0.post1
 xformers>=0.0.23.post1
 autoawq==0.2.7.post3
-liger-kernel==0.5.8
+liger-kernel==0.5.9
 # END section
 
 packaging==23.2
 
-peft==0.15.1
+huggingface_hub==0.31.0
+peft==0.15.2
 transformers==4.51.3
 tokenizers>=0.21.1
 accelerate==1.6.0
-datasets==3.5.0
+datasets==3.5.1
 deepspeed>=0.15.4
-trl==0.16.1
-hf_xet==1.0.0
+trl==0.17.0
+hf_xet==1.1.0
 hqq==0.2.5
 
 optimum==1.16.2

setup.py

@@ -67,13 +67,13 @@ def parse_requirements(extras_require_map):
             if (major, minor) >= (2, 7):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 # _install_requires.append("xformers==0.0.29.post3")  # xformers seems to be hard pinned to 2.6.0
-                extras_require_map["vllm"] = ["vllm==0.8.3"]
+                extras_require_map["vllm"] = ["vllm==0.8.5.post1"]
             elif (major, minor) >= (2, 6):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 _install_requires.append(
                     "xformers==0.0.29.post2"
                 )  # vllm needs post2 w torch 2.6
-                extras_require_map["vllm"] = ["vllm==0.8.3"]
+                extras_require_map["vllm"] = ["vllm==0.8.5.post1"]
             elif (major, minor) >= (2, 5):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 if patch == 0:
@@ -142,6 +142,7 @@ extras_require = {
         "apollo-torch",
         "lomo-optim==0.1.1",
         "torch-optimi==0.2.1",
+        "came_pytorch==0.1.3",
     ],
     "ray": [
         "ray[train]",
@@ -149,6 +150,9 @@ extras_require = {
     "vllm": [
         "vllm==0.7.2",
     ],
+    "llmcompressor": [
+        "llmcompressor==0.5.1",
+    ],
 }
 
 install_requires, dependency_links, extras_require_build = parse_requirements(

src/axolotl/__init__.py

@@ -4,4 +4,4 @@ import pkgutil
 
 __path__ = pkgutil.extend_path(__path__, __name__)  # Make this a namespace package
 
-__version__ = "0.8.0"
+__version__ = "0.10.0.dev0"

src/axolotl/cli/__init__.py

@@ -2,4 +2,7 @@
 
 import os
 
+from axolotl.logging_config import configure_logging
+
 os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
+configure_logging()

src/axolotl/cli/args.py

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

src/axolotl/cli/art.py

@@ -16,8 +16,15 @@ AXOLOTL_LOGO = """
     @@@@  @@@@@@@@@@@@@@@@
 """
 
+HAS_PRINTED_LOGO = False
+
 
 def print_axolotl_text_art():
     """Prints axolotl ASCII art."""
+
+    global HAS_PRINTED_LOGO  # pylint: disable=global-statement
+    if HAS_PRINTED_LOGO:
+        return
     if is_main_process():
+        HAS_PRINTED_LOGO = True
         print(AXOLOTL_LOGO)

src/axolotl/cli/checks.py

@@ -8,9 +8,6 @@ from accelerate.commands.config import config_args
 from huggingface_hub import HfApi
 from huggingface_hub.utils import LocalTokenNotFoundError
 
-from axolotl.logging_config import configure_logging
-
-configure_logging()
 LOG = logging.getLogger(__name__)
 
 

src/axolotl/cli/config.py

@@ -5,6 +5,7 @@ import logging
 import os
 import tempfile
 from pathlib import Path
+from tempfile import NamedTemporaryFile
 from typing import Union
 from urllib.parse import urlparse
 
@@ -152,7 +153,15 @@ def prepare_plugins(cfg: DictDefault):
             plugin_manager.register(plugin_name)
 
 
-def load_cfg(config: Union[str, Path] = Path("examples/"), **kwargs) -> DictDefault:
+def plugin_set_cfg(cfg: DictDefault):
+    if cfg.get("plugins"):
+        plugin_manager = PluginManager.get_instance()
+        plugin_manager.cfg = cfg
+
+
+def load_cfg(
+    config: str | Path | DictDefault = Path("examples/"), **kwargs
+) -> DictDefault:
     """
     Loads the `axolotl` configuration stored at `config`, validates it, and performs
     various setup.
@@ -164,13 +173,24 @@ def load_cfg(config: Union[str, Path] = Path("examples/"), **kwargs) -> DictDefa
     Returns:
         `DictDefault` mapping configuration keys to values.
     """
-    config = check_remote_config(config)
-    if Path(config).is_dir():
-        config = choose_config(Path(config))
+    if isinstance(config, (str, Path)):
+        config = check_remote_config(config)
+        if Path(config).is_dir():
+            config = choose_config(Path(config))
 
-    # Load the config from the yaml file
-    with open(config, encoding="utf-8") as file:
-        cfg: DictDefault = DictDefault(yaml.safe_load(file))
+        # Load the config from the yaml file
+        with open(config, encoding="utf-8") as file:
+            cfg: DictDefault = DictDefault(yaml.safe_load(file))
+
+        cfg.axolotl_config_path = config
+    else:
+        cfg = config
+        with NamedTemporaryFile(
+            mode="w", delete=False, suffix=".yml", prefix="axolotl_config_"
+        ) as temp_file:
+            temp_file.write(yaml.dump(config.to_dict()))
+            temp_file.close()
+        cfg.axolotl_config_path = temp_file.name
 
     # If there are any options passed in the cli, if it is something that seems valid
     # from the yaml, then overwrite the value
@@ -184,8 +204,6 @@ def load_cfg(config: Union[str, Path] = Path("examples/"), **kwargs) -> DictDefa
             else:
                 cfg[k] = kwargs[k]
 
-    cfg.axolotl_config_path = config
-
     try:
         device_props = torch.cuda.get_device_properties("cuda")
         gpu_version = "sm_" + str(device_props.major) + str(device_props.minor)
@@ -213,5 +231,6 @@ def load_cfg(config: Union[str, Path] = Path("examples/"), **kwargs) -> DictDefa
     setup_wandb_env_vars(cfg)
     setup_mlflow_env_vars(cfg)
     setup_comet_env_vars(cfg)
+    plugin_set_cfg(cfg)
 
     return cfg

src/axolotl/cli/evaluate.py

@@ -1,6 +1,7 @@
 """CLI to run evaluation on a model."""
 
 import logging
+import os
 from pathlib import Path
 from typing import Union
 
@@ -14,6 +15,7 @@ from axolotl.cli.checks import check_accelerate_default_config, check_user_token
 from axolotl.cli.config import load_cfg
 from axolotl.common.datasets import load_datasets, load_preference_datasets
 from axolotl.evaluate import evaluate
+from axolotl.utils import patch_optimized_env
 from axolotl.utils.dict import DictDefault
 
 LOG = logging.getLogger(__name__)
@@ -29,10 +31,14 @@ def do_evaluate(cfg: DictDefault, cli_args: TrainerCliArgs) -> None:
         cfg: Dictionary mapping `axolotl` config keys to values.
         cli_args: CLI arguments.
     """
+    # Enable expandable segments for cuda allocation to improve VRAM usage
+    patch_optimized_env()
+
     # pylint: disable=duplicate-code
     print_axolotl_text_art()
     check_accelerate_default_config()
-    check_user_token()
+    if int(os.getenv("LOCAL_RANK", "0")) == 0:
+        check_user_token()
 
     if cfg.rl:
         dataset_meta = load_preference_datasets(cfg=cfg, cli_args=cli_args)

src/axolotl/cli/main.py

@@ -28,9 +28,8 @@ from axolotl.cli.utils import (
     fetch_from_github,
     filter_none_kwargs,
 )
-from axolotl.cli.vllm_serve import do_vllm_serve
 from axolotl.integrations.lm_eval.cli import lm_eval
-from axolotl.utils import set_pytorch_cuda_alloc_conf
+from axolotl.utils import patch_optimized_env
 from axolotl.utils.schemas.config import AxolotlInputConfig
 
 
@@ -56,6 +55,8 @@ def preprocess(config: str, cloud: Optional[str] = None, **kwargs) -> None:
         kwargs: Additional keyword arguments which correspond to CLI args or `axolotl`
             config options.
     """
+    patch_optimized_env()
+
     if cloud:
         from axolotl.cli.cloud import do_cli_preprocess
 
@@ -101,7 +102,7 @@ def train(
             config options.
     """
     # Enable expandable segments for cuda allocation to improve VRAM usage
-    set_pytorch_cuda_alloc_conf()
+    patch_optimized_env()
 
     if "use_ray" in kwargs and kwargs["use_ray"]:
         accelerate = False
@@ -327,6 +328,8 @@ def fetch(directory: str, dest: Optional[str]) -> None:
 @add_options_from_dataclass(VllmServeCliArgs)
 @filter_none_kwargs
 def vllm_serve(config: str, **cli_args: VllmServeCliArgs):
+    from axolotl.cli.vllm_serve import do_vllm_serve
+
     do_vllm_serve(config, cli_args)
 
 

src/axolotl/cli/preprocess.py

@@ -18,6 +18,7 @@ from axolotl.cli.checks import check_accelerate_default_config, check_user_token
 from axolotl.cli.config import load_cfg
 from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
 from axolotl.common.datasets import load_datasets, load_preference_datasets
+from axolotl.integrations.base import PluginManager
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.trainer import disable_datasets_caching
 
@@ -47,7 +48,10 @@ def do_preprocess(cfg: DictDefault, cli_args: PreprocessCliArgs) -> None:
         cfg.dataset_prepared_path = DEFAULT_DATASET_PREPARED_PATH
 
     with disable_datasets_caching():
-        if cfg.rl:
+        plugin_manager = PluginManager.get_instance()
+        if plugin_manager.load_datasets(cfg, preprocess=True):
+            pass
+        elif cfg.rl:
             load_preference_datasets(cfg=cfg, cli_args=cli_args)
         else:
             load_datasets(cfg=cfg, cli_args=cli_args)

src/axolotl/cli/train.py

@@ -1,5 +1,6 @@
 """CLI to run training on a model."""
 
+import gc
 import logging
 import os
 from pathlib import Path
@@ -17,7 +18,7 @@ from axolotl.cli.config import load_cfg
 from axolotl.common.datasets import load_datasets, load_preference_datasets
 from axolotl.integrations.base import PluginManager
 from axolotl.train import train
-from axolotl.utils import set_pytorch_cuda_alloc_conf
+from axolotl.utils import patch_optimized_env
 from axolotl.utils.config import normalize_config, resolve_dtype
 from axolotl.utils.dict import DictDefault
 
@@ -35,21 +36,27 @@ def do_train(cfg: DictDefault, cli_args: TrainerCliArgs):
         cli_args: Training-specific CLI arguments.
     """
     # Enable expandable segments for cuda allocation to improve VRAM usage
-    set_pytorch_cuda_alloc_conf()
+    patch_optimized_env()
 
     print_axolotl_text_art()
     check_accelerate_default_config()
     if int(os.getenv("LOCAL_RANK", "0")) == 0:
         check_user_token()
 
-    if cfg.rl:
-        dataset_meta = load_preference_datasets(cfg=cfg, cli_args=cli_args)
-    else:
-        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+    plugin_manager = PluginManager.get_instance()
+    dataset_meta = plugin_manager.load_datasets(cfg, preprocess=False)
+    if not dataset_meta:
+        if cfg.rl:
+            dataset_meta = load_preference_datasets(cfg=cfg, cli_args=cli_args)
+        else:
+            dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
     model, tokenizer, trainer = train(cfg=cfg, dataset_meta=dataset_meta)
+
     del model, tokenizer, trainer
 
+    gc.collect()
+
     plugin_manager = PluginManager.get_instance()
     plugin_manager.post_train_unload(cfg)
 

src/axolotl/cli/utils.py

@@ -20,11 +20,10 @@ from transformers import (
     ProcessorMixin,
 )
 
-from axolotl.logging_config import configure_logging
+from axolotl.loaders import load_processor, load_tokenizer
+from axolotl.loaders.model import ModelLoader
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.models import load_model, load_processor, load_tokenizer
 
-configure_logging()
 LOG = logging.getLogger(__name__)
 
 
@@ -320,7 +319,8 @@ def load_model_and_tokenizer(
     tokenizer = load_tokenizer(cfg)
 
     LOG.info("loading model...")
-    model, _ = load_model(cfg, tokenizer, inference=inference)
+    model_loader = ModelLoader(cfg, tokenizer, inference=inference)
+    model, _ = model_loader.load()
 
     processor = None
     if cfg.is_multimodal:

src/axolotl/cli/vllm_serve.py

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

src/axolotl/common/architectures.py

@@ -11,5 +11,6 @@ MOE_ARCH_BLOCK = {
     ],
     "mixtral": "MixtralSparseMoeBlock",
     "qwen2_moe": "Qwen2MoeSparseMoeBlock",
+    "qwen3_moe": "Qwen3MoeSparseMoeBlock",
     "deepseek_v2": "DeepseekV2MoE",
 }

src/axolotl/common/datasets.py

@@ -10,10 +10,10 @@ from datasets import Dataset
 
 import axolotl.monkeypatch.data.batch_dataset_fetcher  # pylint: disable=unused-import  # noqa: F401
 from axolotl.cli.args import PreprocessCliArgs, TrainerCliArgs
+from axolotl.loaders import load_processor, load_tokenizer
 from axolotl.utils.data import prepare_dataset
 from axolotl.utils.data.rl import load_prepare_preference_datasets
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.models import load_processor, load_tokenizer
 from axolotl.utils.schemas.enums import RLType
 from axolotl.utils.tokenization import check_dataset_labels
 
@@ -48,7 +48,8 @@ def sample_dataset(dataset: Dataset, num_samples: int) -> Dataset:
 def load_datasets(
     *,
     cfg: DictDefault,
-    cli_args: Union[PreprocessCliArgs, TrainerCliArgs],
+    cli_args: PreprocessCliArgs | TrainerCliArgs | None = None,
+    debug: bool = False,
 ) -> TrainDatasetMeta:
     """
     Loads one or more training or evaluation datasets, calling
@@ -57,6 +58,7 @@ def load_datasets(
     Args:
         cfg: Dictionary mapping `axolotl` config keys to values.
         cli_args: Command-specific CLI arguments.
+        debug: Whether to print out tokenization of sample
 
     Returns:
         Dataclass with fields for training and evaluation datasets and the computed
@@ -65,7 +67,8 @@ def load_datasets(
     tokenizer = load_tokenizer(cfg)
     processor = load_processor(cfg, tokenizer=tokenizer) if cfg.processor_type else None
     preprocess_iterable = (
-        hasattr(cli_args, "iterable")
+        cli_args
+        and hasattr(cli_args, "iterable")
         and cli_args.iterable is not None
         and cli_args.iterable
     )
@@ -77,20 +80,25 @@ def load_datasets(
         preprocess_iterable=preprocess_iterable,
     )
 
-    if (
-        cli_args.debug
-        or cfg.debug
-        or cli_args.debug_text_only
-        or int(cli_args.debug_num_examples) > 0
-    ):
+    if (  # pylint: disable=too-many-boolean-expressions
+        cli_args
+        and (
+            cli_args.debug
+            or cfg.debug
+            or cli_args.debug_text_only
+            or int(cli_args.debug_num_examples) > 0
+        )
+    ) or debug:
         LOG.info("check_dataset_labels...")
 
-        train_samples = sample_dataset(train_dataset, cli_args.debug_num_examples)
+        num_examples = cli_args.debug_num_examples if cli_args else 1
+        text_only = cli_args.debug_text_only if cli_args else False
+        train_samples = sample_dataset(train_dataset, num_examples)
         check_dataset_labels(
             train_samples,
             tokenizer,
-            num_examples=cli_args.debug_num_examples,
-            text_only=cli_args.debug_text_only,
+            num_examples=num_examples,
+            text_only=text_only,
         )
 
         LOG.info("printing prompters...")

src/axolotl/core/trainer_builder.py

@@ -21,6 +21,7 @@ import importlib.util
 import inspect
 import logging
 import math
+import os
 import sys
 from abc import abstractmethod
 from pathlib import Path
@@ -58,8 +59,10 @@ from axolotl.core.training_args import (
     AxolotlTrainingArguments,
 )
 from axolotl.integrations.base import PluginManager
+from axolotl.loaders.utils import ensure_dtype
 from axolotl.monkeypatch.multipack import SUPPORTED_MULTIPACK_MODEL_TYPES
 from axolotl.monkeypatch.relora import ReLoRACallback
+from axolotl.monkeypatch.trainer.lr import patch_trainer_get_lr
 from axolotl.processing_strategies import get_processing_strategy
 from axolotl.utils import is_comet_available, is_mlflow_available
 from axolotl.utils.callbacks import (
@@ -71,6 +74,7 @@ from axolotl.utils.callbacks import (
     SaveBetterTransformerModelCallback,
     bench_eval_callback_factory,
     causal_lm_bench_eval_callback_factory,
+    colab_inference_post_train_callback,
     log_prediction_callback_factory,
 )
 from axolotl.utils.callbacks.lisa import lisa_callback_factory
@@ -83,7 +87,6 @@ from axolotl.utils.collators import (
     V2BatchSamplerDataCollatorForSeq2Seq,
 )
 from axolotl.utils.collators.mm_chat import MultiModalChatDataCollator
-from axolotl.utils.models import ensure_dtype
 from axolotl.utils.schemas.enums import CustomSupportedOptimizers, RLType
 
 try:
@@ -114,6 +117,8 @@ class TrainerBuilderBase(abc.ABC):
         if hasattr(model, "add_model_tags"):
             model.add_model_tags(["axolotl"])
 
+        patch_trainer_get_lr()
+
     @property
     def model_ref(self):
         return self._model_ref
@@ -165,6 +170,9 @@ class TrainerBuilderBase(abc.ABC):
                 )
             )
 
+        if self.cfg.gc_steps:
+            callbacks.append(GCCallback(gc_steps=self.cfg.gc_steps))
+
         if self.cfg.use_wandb:
             callbacks.append(
                 SaveAxolotlConfigtoWandBCallback(self.cfg.axolotl_config_path)
@@ -246,9 +254,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         if self.cfg.loss_watchdog_threshold is not None:
             callbacks.append(LossWatchDogCallback(self.cfg))
 
-        if self.cfg.gc_steps:
-            callbacks.append(GCCallback(gc_steps=self.cfg.gc_steps))
-
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
@@ -290,6 +295,10 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         if self.cfg.lisa_step_interval and self.cfg.lisa_n_layers:
             callbacks.append(lisa_callback_factory(trainer))
 
+        if any("COLAB_" in key for key in os.environ):
+            ColabCallback = colab_inference_post_train_callback(trainer)
+            callbacks.append(ColabCallback(self.cfg))
+
         callbacks.extend(super().get_post_trainer_create_callbacks(trainer=trainer))
         return callbacks
 
@@ -344,7 +353,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs["warmup_steps"] = warmup_steps
         training_arguments_kwargs["logging_steps"] = logging_steps
 
-        if self.cfg.seed:
+        if self.cfg.seed is not None:
             training_arguments_kwargs["seed"] = self.cfg.seed
 
         if self.cfg.gradient_checkpointing:
@@ -378,8 +387,12 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             training_arguments_kwargs["adam_beta1"] = self.cfg.adam_beta1
         if self.cfg.adam_beta2:
             training_arguments_kwargs["adam_beta2"] = self.cfg.adam_beta2
+        if self.cfg.adam_beta3:
+            training_arguments_kwargs["adam_beta3"] = self.cfg.adam_beta3
         if self.cfg.adam_epsilon:
             training_arguments_kwargs["adam_epsilon"] = self.cfg.adam_epsilon
+        if self.cfg.adam_epsilon2:
+            training_arguments_kwargs["adam_epsilon2"] = self.cfg.adam_epsilon2
         if self.cfg.max_grad_norm:
             training_arguments_kwargs["max_grad_norm"] = self.cfg.max_grad_norm
 
@@ -485,7 +498,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
 
         # these are all the "standard" kwargs that are def used
         training_arguments_kwargs["max_steps"] = (
-            total_num_steps if self.cfg.max_steps else -1
+            self.cfg.max_steps if self.cfg.max_steps else -1
         )
         training_arguments_kwargs["max_seq_length"] = self.cfg.sequence_len
         training_arguments_kwargs["per_device_train_batch_size"] = (
@@ -697,6 +710,20 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 optimizer_cls = ADOPT
                 adam_kwargs["decouple"] = True
                 optimizer_kwargs.update(adam_kwargs)
+            elif self.cfg.optimizer == "came_pytorch":
+                from came_pytorch import CAME
+
+                optimizer_cls = CAME
+
+                beta1 = training_arguments_kwargs.get("adam_beta1", 0.9)
+                beta2 = training_arguments_kwargs.get("adam_beta2", 0.999)
+                beta3 = training_arguments_kwargs.get("adam_beta3", 0.9999)
+                eps1 = training_arguments_kwargs.get("adam_epsilon", 1e-30)
+                eps2 = training_arguments_kwargs.get("adam_epsilon2", 1e-16)
+                adam_kwargs["betas"] = (beta1, beta2, beta3)
+                adam_kwargs["eps"] = (eps1, eps2)
+
+                optimizer_kwargs.update(adam_kwargs)
 
             # Parse any additional optimizer args from config
             if self.cfg.optim_args:
@@ -771,11 +798,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 self.cfg.kd_top_k_before_softmax
             )
 
-        training_arguments_kwargs["sequence_parallel_degree"] = (
-            self.cfg.sequence_parallel_degree
-        )
-        training_arguments_kwargs["ring_attn_func"] = self.cfg.ring_attn_func
-
         if self.cfg.reward_model:
             training_args_cls = AxolotlRewardConfig
         elif self.cfg.process_reward_model:
@@ -796,14 +818,15 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         data_collator_kwargs = {
             "padding": True,  # True/"longest" is the default
         }
+        multiple = 64
         if self.cfg.pad_to_sequence_len:
-            data_collator_kwargs["pad_to_multiple_of"] = 64 * math.ceil(
-                self.cfg.sequence_len / 64
+            data_collator_kwargs["pad_to_multiple_of"] = multiple * math.ceil(
+                self.cfg.sequence_len / multiple
             )
         else:
             # A100 is best at 64, while others at 8. Let's use the larger so we don't have to check
             # https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html
-            data_collator_kwargs["pad_to_multiple_of"] = 64
+            data_collator_kwargs["pad_to_multiple_of"] = multiple
 
         if self.cfg.reward_model:
             data_collator_kwargs["max_length"] = self.cfg.sequence_len
@@ -1009,8 +1032,10 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             training_args_kwargs["dataloader_prefetch_factor"] = (
                 self.cfg.dataloader_prefetch_factor
             )
-        if self.cfg.seed:
+
+        if self.cfg.seed is not None:
             training_args_kwargs["seed"] = self.cfg.seed
+
         if self.cfg.gradient_checkpointing:
             training_args_kwargs["gradient_checkpointing"] = (
                 self.cfg.gradient_checkpointing
@@ -1034,6 +1059,8 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             # default to saving each epoch if not defined
             training_args_kwargs["save_strategy"] = "epoch"
 
+        training_args_kwargs["save_only_model"] = self.cfg.save_only_model
+
         if self.cfg.dataset_processes:
             training_args_kwargs["dataset_num_proc"] = self.cfg.dataset_processes
 
@@ -1048,9 +1075,8 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         if self.cfg.rpo_alpha is not None:
             training_args_kwargs["rpo_alpha"] = self.cfg.rpo_alpha
 
-        training_args_kwargs["sequence_parallel_degree"] = (
-            self.cfg.sequence_parallel_degree
-        )
+        if self.cfg.use_wandb:
+            training_args_kwargs["run_name"] = self.cfg.wandb_name
 
         training_args_cls = None
         blocklist_args_kwargs = []
@@ -1122,6 +1148,12 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             **training_args_kwargs,
         )
 
+        # unset run_name so wandb sets up experiment names
+        if self.cfg.use_wandb and training_args.run_name == training_args.output_dir:
+            training_args.run_name = (  # pylint: disable=attribute-defined-outside-init
+                None
+            )
+
         return training_args
 
     def build(self, total_num_steps):
@@ -1133,13 +1165,16 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         if self.eval_dataset:
             trainer_kwargs["eval_dataset"] = self.eval_dataset
         if self.cfg.adapter and self.peft_config:
-            trainer_kwargs["peft_config"] = self.peft_config
+            if self.cfg.rl is not RLType.GRPO:
+                trainer_kwargs["peft_config"] = self.peft_config
         if self.cfg.precompute_ref_log_probs is not None:
             trainer_kwargs["precompute_ref_log_probs"] = (
                 self.cfg.precompute_ref_log_probs
             )
         if self.cfg.rl is RLType.GRPO:
-            trainer_cls = GRPOStrategy.get_trainer_class()
+            trainer_cls = GRPOStrategy.get_trainer_class(
+                sequence_parallel=self.cfg.sequence_parallel_degree > 1
+            )
             trainer_cls_args = [self.model]
             trainer_cls_args.extend(GRPOStrategy.set_trainer_args(self.cfg))
             trainer_kwargs.update(GRPOStrategy.set_trainer_kwargs(self.cfg))
@@ -1158,6 +1193,12 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         else:
             raise ValueError(f"Unsupported RL: {self.cfg.rl}")
 
+        if self.cfg.plugins:
+            plugin_manager = PluginManager.get_instance()
+            temp_trainer_cls = plugin_manager.get_trainer_cls(self.cfg)
+            if temp_trainer_cls is not None:
+                trainer_cls = temp_trainer_cls
+
         sig = inspect.signature(trainer_cls)
         if "tokenizer" in sig.parameters.keys():
             trainer_kwargs["tokenizer"] = self.tokenizer

src/axolotl/core/trainers/__init__.py

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

src/axolotl/core/trainers/base.py

@@ -29,7 +29,6 @@ from axolotl.core.trainers.mixins import (
     OptimizerMixin,
     RngLoaderMixin,
     SchedulerMixin,
-    SequenceParallelMixin,
 )
 from axolotl.core.trainers.utils import (
     sanitize_kwargs_for_ds_tagging,
@@ -40,9 +39,7 @@ from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
 LOG = logging.getLogger(__name__)
 
 
-class AxolotlTrainer(
-    SchedulerMixin, OptimizerMixin, RngLoaderMixin, SequenceParallelMixin, Trainer
-):
+class AxolotlTrainer(SchedulerMixin, OptimizerMixin, RngLoaderMixin, Trainer):
     """Extend the base Trainer for axolotl helpers"""
 
     args = None  # type: "AxolotlTrainingArguments"  # type: ignore[name-defined]
@@ -68,10 +65,6 @@ class AxolotlTrainer(
         if self.args.orpo_alpha:
             self.loss_fct = torch.nn.CrossEntropyLoss(reduction="none")
 
-        # Initialize sequence parallelism if enabled
-        if self.args.sequence_parallel_degree > 1:
-            self._setup_sequence_parallel()
-
     def _wrap_model(self, model, training=True, dataloader=None):
         if self.args.torch_compile:
             torch._dynamo.config.accumulated_cache_size_limit = (  # pylint: disable=protected-access
@@ -114,14 +107,16 @@ class AxolotlTrainer(
             packing_efficiency_estimate=self.args.sample_packing_efficiency,
             batch_max_len=batch_max_len,
             batch_size=batch_size,
+            group_size=self.args.sample_packing_group_size,
+            bin_size=self.args.sample_packing_bin_size,
             sequential=self.args.sample_packing_sequentially,
             drop_last=True,
         )
 
     def _get_train_sampler(self) -> Sampler | None:
         """
-        Helper method to get the sampler for training. Handles cases for sequence
-        parallelism, sample packing, and curriculum sampling (sequential).
+        Helper method to get the sampler for training. Handles cases for sample packing
+        and curriculum sampling (sequential).
 
         Returns:
             If the dataset is non-empty, a sampler is returned, the type of which
@@ -130,9 +125,7 @@ class AxolotlTrainer(
         use_sample_packing = self.args.sample_packing and not self.args.pretraining
 
         # Determine the base sampler first
-        if self.args.sequence_parallel_degree > 1:
-            base_sampler = self._sp_get_train_sampler(self.train_dataset)
-        elif self.args.curriculum_sampling:
+        if self.args.curriculum_sampling:
             base_sampler = SequentialSampler(self.train_dataset)
         elif use_sample_packing:
             base_sampler = RandomSampler(self.train_dataset)
@@ -151,8 +144,7 @@ class AxolotlTrainer(
 
     def _get_eval_sampler(self, eval_dataset: Dataset | None = None) -> Sampler | None:
         """
-        Helper method to get the sampler for evaluation. Handles sequence parallelism
-        and sample packing cases.
+        Helper method to get the sampler for evaluation. Handles sample packing case.
 
         Returns:
             If the dataset is non-empty, a sampler is returned, the type of which
@@ -166,9 +158,7 @@ class AxolotlTrainer(
         )
 
         # Determine the base sampler
-        if self.args.sequence_parallel_degree > 1:
-            base_sampler = self._sp_get_eval_sampler(eval_dataset)
-        elif use_multipack:
+        if use_multipack:
             base_sampler = SequentialSampler(eval_dataset)
         else:
             return super()._get_eval_sampler(eval_dataset)
@@ -234,14 +224,6 @@ class AxolotlTrainer(
         ):
             self.accelerator.even_batches = False
 
-        # Return unprepared dataloader if using sequence parallelism
-        # TODO(djsaunde): We might be able to use `accelerate`'s dataloader preparation
-        # if we use `dispatch_batches` and `slice_fn_for_dispatch` properly (i.e.,
-        # slice each batch along the sequence dimension).
-        if self.args.sequence_parallel_degree > 1:
-            return dataloader
-
-        # Otherwise prepare with accelerator
         return self.accelerator.prepare_data_loader(dataloader)
 
     def get_train_dataloader(self) -> DataLoader:
@@ -285,12 +267,7 @@ class AxolotlTrainer(
 
             return dataloader
 
-        # Handle sample packing or sequence parallelism
-        if (
-            self.args.sample_packing
-            and self.args.eval_sample_packing is not False
-            or self.args.sequence_parallel_degree > 1
-        ):
+        if self.args.sample_packing and self.args.eval_sample_packing is not False:
             # Get appropriate data collator
             self.data_collator = (  # pylint: disable=attribute-defined-outside-init
                 self.eval_data_collator
@@ -300,17 +277,6 @@ class AxolotlTrainer(
             if "length" in eval_dataset.column_names:
                 eval_dataset = eval_dataset.remove_columns(["length"])
 
-            # Handle dataset preprocessing for SP
-            if self.args.sequence_parallel_degree > 1:
-                if isinstance(eval_dataset, datasets.Dataset):
-                    eval_dataset = self._remove_unused_columns(
-                        eval_dataset, description="evaluation"
-                    )
-                else:
-                    self.data_collator = self._get_collator_with_removed_columns(  # pylint: disable=attribute-defined-outside-init
-                        self.data_collator, description="evaluation"
-                    )
-
             # Use eval_batch_size for sample packing, per_device_eval_batch_size otherwise
             batch_size = (
                 self.args.eval_batch_size
@@ -371,15 +337,13 @@ class AxolotlTrainer(
                 num_items_in_batch=num_items_in_batch,
             )
 
-        loss = super().compute_loss(
+        return super().compute_loss(
             model,
             inputs,
             return_outputs=return_outputs,
             num_items_in_batch=num_items_in_batch,
         )
 
-        return loss
-
     @staticmethod
     def orpo_concatenate_inputs(inputs, label_pad_token=-100, pad_token=0, device=None):
         concatenated_batch = {}

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

@@ -1,15 +1,11 @@
-"""
-DPO Specific Strategy for training
-"""
+"""DPO Specific Strategy for training"""
 
 from axolotl.core.trainers.dpo.trainer import AxolotlDPOTrainer
 from axolotl.utils.schemas.enums import RLType
 
 
 class DPOStrategy:
-    """
-    Strategy for DPO training
-    """
+    """Strategy for DPO training"""
 
     @classmethod
     def get_trainer_class(cls):

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

@@ -1,6 +1,4 @@
-"""
-DPO trainer for axolotl
-"""
+"""DPO trainer for axolotl"""
 
 import gc
 from functools import wraps
@@ -24,9 +22,7 @@ if is_sagemaker_mp_enabled():
 
 
 class AxolotlDPOTrainer(RngLoaderMixin, SchedulerMixin, DPOTrainer):
-    """
-    Extend the base DPOTrainer for axolotl helpers
-    """
+    """Extend the base DPOTrainer for axolotl helpers."""
 
     tag_names = ["axolotl", "dpo"]
 
@@ -71,8 +67,9 @@ class AxolotlDPOTrainer(RngLoaderMixin, SchedulerMixin, DPOTrainer):
     @wraps(DPOTrainer.push_to_hub)
     def push_to_hub(self, *args, **kwargs) -> str:
         """
-        Overwrite the `push_to_hub` method in order to force-add the tags when pushing the
-        model on the Hub. Please refer to `~transformers.Trainer.push_to_hub` for more details.
+        Overwrite the `push_to_hub` method in order to force-add the tags when pushing
+        the model on the Hub. Please refer to `~transformers.Trainer.push_to_hub`
+        for more details.
         """
         kwargs = sanitize_kwargs_for_ds_tagging(
             dataset_tags=self.dataset_tags, kwargs=kwargs
@@ -105,12 +102,8 @@ class AxolotlDPOTrainer(RngLoaderMixin, SchedulerMixin, DPOTrainer):
             # dpo trainer may incorrectly prepend the bos_token_id to the dpo outputs
             if res["chosen_input_ids"][0] == processing_class.bos_token_id:
                 res["chosen_input_ids"] = res["chosen_input_ids"][1:]
-                res["chosen_labels"] = res["chosen_labels"][1:]
-                res["chosen_attention_mask"] = res["chosen_attention_mask"][1:]
             if res["rejected_input_ids"][0] == processing_class.bos_token_id:
                 res["rejected_input_ids"] = res["rejected_input_ids"][1:]
-                res["rejected_labels"] = res["rejected_labels"][1:]
-                res["rejected_attention_mask"] = res["rejected_attention_mask"][1:]
 
         return res
 

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

@@ -1,37 +1,41 @@
-"""
-GRPO Specific Strategy for training
-"""
+"""GRPO Specific Strategy for training"""
 
 import importlib
 import inspect
 import logging
+from typing import Any
 
 from trl.trainer.grpo_trainer import RewardFunc
 
-from axolotl.core.trainers.grpo.trainer import AxolotlGRPOTrainer
+from axolotl.core.trainers.grpo.args import AxolotlGRPOConfig
+from axolotl.core.trainers.grpo.trainer import (
+    AxolotlGRPOSequenceParallelTrainer,
+    AxolotlGRPOTrainer,
+)
+from axolotl.utils.dict import DictDefault
 from axolotl.utils.schemas.trl import TRLConfig
 
-LOG = logging.getLogger("axolotl")
+LOG = logging.getLogger(__name__)
 
 
 class GRPOStrategy:
-    """
-    Strategy for GRPO training
-    """
+    """Strategy for GRPO training"""
 
     @classmethod
-    def get_trainer_class(cls):
+    def get_trainer_class(
+        cls, sequence_parallel: bool
+    ) -> type[AxolotlGRPOTrainer] | type[AxolotlGRPOSequenceParallelTrainer]:
+        if sequence_parallel:
+            return AxolotlGRPOSequenceParallelTrainer
         return AxolotlGRPOTrainer
 
     @classmethod
-    def get_training_args_class(cls):
-        from axolotl.core.trainers.grpo.args import AxolotlGRPOConfig
-
+    def get_training_args_class(cls) -> type[AxolotlGRPOConfig]:
         return AxolotlGRPOConfig
 
     @classmethod
-    def set_training_args_kwargs(cls, cfg):
-        grpo_args_kwargs = {}
+    def set_training_args_kwargs(cls, cfg: DictDefault) -> dict[str, Any]:
+        grpo_args_kwargs: dict[str, Any] = {}
 
         if not hasattr(cfg, "trl") or not cfg.trl:
             return grpo_args_kwargs
@@ -40,8 +44,8 @@ class GRPOStrategy:
 
         if trl.use_vllm:
             grpo_args_kwargs["use_vllm"] = trl.use_vllm
-            grpo_args_kwargs["vllm_server_host"] = trl.vllm_server_host or trl.vllm.host
-            grpo_args_kwargs["vllm_server_port"] = trl.vllm_server_port or trl.vllm.port
+            grpo_args_kwargs["vllm_server_host"] = trl.vllm_server_host or trl.vllm.host  # type: ignore[attr-defined]
+            grpo_args_kwargs["vllm_server_port"] = trl.vllm_server_port or trl.vllm.port  # type: ignore[attr-defined]
             if trl.vllm_server_timeout:
                 grpo_args_kwargs["vllm_server_timeout"] = trl.vllm_server_timeout
             if trl.vllm_guided_decoding_regex:
@@ -63,6 +67,7 @@ class GRPOStrategy:
 
         grpo_args_kwargs["max_completion_length"] = trl.max_completion_length
         grpo_args_kwargs["log_completions"] = trl.log_completions
+        grpo_args_kwargs["num_completions_to_print"] = trl.num_completions_to_print
 
         if trl.reward_weights:
             grpo_args_kwargs["reward_weights"] = trl.reward_weights
@@ -70,6 +75,13 @@ class GRPOStrategy:
         if trl.scale_rewards is not None:
             grpo_args_kwargs["scale_rewards"] = trl.scale_rewards
 
+        if trl.loss_type is not None:
+            grpo_args_kwargs["loss_type"] = trl.loss_type
+        if trl.mask_truncated_completions is not None:
+            grpo_args_kwargs["mask_truncated_completions"] = (
+                trl.mask_truncated_completions
+            )
+
         if trl.temperature is not None:
             grpo_args_kwargs["temperature"] = trl.temperature
         if trl.top_p is not None:
@@ -85,21 +97,27 @@ class GRPOStrategy:
             grpo_args_kwargs["num_iterations"] = trl.num_iterations
         if trl.epsilon is not None:
             grpo_args_kwargs["epsilon"] = trl.epsilon
+        if trl.epsilon_high is not None:
+            grpo_args_kwargs["epsilon_high"] = trl.epsilon_high
+
+        if trl.use_liger_loss is not None:
+            grpo_args_kwargs["use_liger_loss"] = trl.use_liger_loss
 
         return grpo_args_kwargs
 
     @classmethod
-    def set_trainer_args(cls, cfg):
+    def set_trainer_args(cls, cfg: DictDefault) -> list[Any]:
         trainer_args = []
         if cfg.trl and cfg.trl.reward_funcs:
             reward_funcs = []
             for reward_func_fqn in cfg.trl.reward_funcs:
                 reward_funcs.append(cls.get_reward_func(reward_func_fqn))
             trainer_args.append(reward_funcs)
+
         return trainer_args
 
     @classmethod
-    def set_trainer_kwargs(cls, cfg):
+    def set_trainer_kwargs(cls, cfg: DictDefault) -> dict[str, Any]:
         trainer_kwargs = {}
         if cfg.trl and cfg.trl.reward_processing_classes:
             trainer_kwargs["reward_processing_classes"] = (
@@ -113,7 +131,7 @@ class GRPOStrategy:
         return None
 
     @classmethod
-    def get_blocklist_args_kwargs(cls):
+    def get_blocklist_args_kwargs(cls) -> list[str]:
         return ["dataset_num_proc"]
 
     @classmethod
@@ -124,18 +142,20 @@ class GRPOStrategy:
         Args:
             reward_func_fqn (str): Fully qualified name of the reward function (e.g. r1_grpo.gsm8k_transform),
                 or a HF hub path to the reward model.
-        Raises:
-            ValueError: If the reward function does not accept at least two arguments.
 
         Returns:
             RewardFunc: A callable that accepts prompts and completions and returns rewards,
                 or a path to a reward model.
 
+        Raises:
+            ValueError: If the reward function does not accept at least two arguments.
         """
         try:
             # use importlib to dynamically load the reward function from the module
             reward_func_module_name = reward_func_fqn.split(".")[-1]
-            reward_func_module = importlib.import_module(reward_func_fqn.split(".")[-2])
+            reward_func_module = importlib.import_module(
+                ".".join(reward_func_fqn.split(".")[:-1])
+            )
             reward_func = getattr(reward_func_module, reward_func_module_name)
             if not len(inspect.signature(reward_func).parameters) >= 2:
                 raise ValueError(

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

@@ -1,22 +1,54 @@
-"""
-Repeat random sampler (akin to the one implemented in
+"""Repeat random sampler (similar to the one implemented in
 https://github.com/huggingface/trl/blob/main/trl/trainer/grpo_trainer.py) that adds
 sequence parallelism functionality; i.e., duplicating data across ranks in the same
-sequencee parallel group.
+sequence parallel group.
 """
 
-from typing import Sized
+from typing import Iterator, Sized
 
 import torch
 from torch.utils.data import Sampler
 
 
 class SequenceParallelRepeatRandomSampler(Sampler):
-    """
-    Sampler for GRPO training with sequence parallelism that ensures:
-    1. Ranks in the same sequence parallel group receive identical data
-    2. Each index is repeated multiple times for sampling different completions
-    3. Entire batches are repeated for reuse in multiple updates
+    """Sampler for GRPO training with sequence parallelism.
+
+    This sampler ensures:
+    - Ranks in the same sequence parallel (SP) group receive identical data.
+    - Each index is repeated multiple times for sampling different completions.
+    - Entire batches are repeated for reuse in multiple updates.
+    - Data is properly distributed across SP groups.
+
+    In the table below, the values represent dataset indices. Each SP group has
+    `sequence_parallel_degree = 2` GPUs working together on the same data. There are 2
+    SP groups (SP0 and SP1), with `world_size = 4` total GPUs.
+
+                                               Sequence Parallel Groups
+                                        |       SP0        |       SP1        |
+                                        |  GPU 0  |  GPU 1 |  GPU 2  |  GPU 3 |
+                    global_step  step    <---> mini_repeat_count=3
+                                            <----------> batch_size=2 per SP group
+    grad_accum=2   ▲  ▲  0       0         [0 0 0  1 1 1]     [2 2 2  3 3 3]   <- SP groups get different data
+                   ▼  |  0       1         [0 0 0  1 1 1]     [2 2 2  3 3 3]   <- Same data for each SP group GPU
+                      |
+                      |  1       2         [0 0 0  1 1 1]     [2 2 2  3 3 3]   <- Repeat same indices for iterations
+    num_iterations=2  ▼  1       3         [0 0 0  1 1 1]     [2 2 2  3 3 3]   <- When using gradient accumulation
+
+                         2       4         [4 4 4  5 5 5]     [6 6 6  7 7 7]   <- New batch of data indices
+                         2       5         [4 4 4  5 5 5]     [6 6 6  7 7 7]
+                                            ...
+
+    Args:
+        dataset: Dataset to sample from.
+        mini_repeat_count: How many times to repeat each sample immediately.
+        world_size: Total number of processes.
+        rank: Rank of current process.
+        batch_size: Number of samples per batch.
+        repeat_count: How many times to repeat the full sampling process.
+        sequence_parallel_degree: Number of ranks in a sequence parallel group.
+        shuffle: Whether to shuffle the dataset.
+        seed: Random seed for shuffling.
+        drop_last: Whether to drop the last incomplete batch.
     """
 
     def __init__(
@@ -70,16 +102,23 @@ class SequenceParallelRepeatRandomSampler(Sampler):
                 * self.batch_size
             )
 
-    def __iter__(self):
+        if shuffle:
+            self.generator = torch.Generator()
+            self.generator.manual_seed(seed)
+
+    def __iter__(self) -> Iterator[int]:
+        """Creates iterator over dataset indices.
+
+        Returns:
+            Iterator that yields indices into the dataset.
+        """
         # Deterministically shuffle based on epoch and seed
         if self.shuffle:
-            # Use same seed for all ranks in the same SP group
-            g = torch.Generator()
-            seed_value = self.seed + self.epoch + self.sp_group_id * 10000
-            g.manual_seed(seed_value)
-            indices = torch.randperm(len(self.dataset), generator=g).tolist()
+            indices = torch.randperm(
+                self.num_samples, generator=self.generator
+            ).tolist()
         else:
-            indices = list(range(len(self.dataset)))
+            indices = list(range(self.num_samples))
 
         # Add extra samples to make it evenly divisible by batch_size
         if len(indices) % self.batch_size != 0:
@@ -113,12 +152,21 @@ class SequenceParallelRepeatRandomSampler(Sampler):
 
         return iter(final_indices)
 
-    def __len__(self):
+    def __len__(self) -> int:
+        """Returns the total length of the iterable including repetitions.
+
+        Returns:
+            Total number of samples.
+        """
         # Total length including all repetitions
         return (
             self.num_samples_per_sp_group * self.mini_repeat_count * self.repeat_count
         )
 
-    def set_epoch(self, epoch):
-        """Sets the epoch for this sampler"""
+    def set_epoch(self, epoch: int) -> None:
+        """Sets the epoch for this sampler.
+
+        Args:
+            epoch: Epoch number to use for shuffling.
+        """
         self.epoch = epoch

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

@@ -6,4 +6,3 @@
 from .optimizer import OptimizerMixin
 from .rng_state_loader import RngLoaderMixin
 from .scheduler import SchedulerMixin
-from .sequence_parallel import SequenceParallelContextManager, SequenceParallelMixin

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

@@ -3,9 +3,10 @@
 import logging
 
 import torch
-from torch.optim.lr_scheduler import OneCycleLR
+from torch.optim.lr_scheduler import LRScheduler, OneCycleLR
 from transformers.trainer import Trainer
 
+from axolotl.integrations.base import PluginManager
 from axolotl.utils.schedulers import (
     RexLR,
     get_cosine_schedule_with_min_lr,
@@ -25,9 +26,9 @@ class SchedulerMixin(Trainer):
 
     def create_scheduler(
         self, num_training_steps: int, optimizer: torch.optim.Optimizer = None
-    ):
+    ) -> LRScheduler:
         """
-        Setup the scheduler. The optimizer of the trainer must have been set up either before this method is called or
+        Set up the scheduler. The optimizer of the trainer must have been set up either before this method is called or
         passed as an argument.
 
         Args:
@@ -47,7 +48,16 @@ class SchedulerMixin(Trainer):
         # fmt: off
         if self.lr_scheduler is None:  # type: ignore  # pylint: disable=access-member-before-definition
             # fmt: on
-            if self.args.alternate_lr_scheduler_type == "one_cycle":
+            plugin_manager = PluginManager.get_instance()
+            lr_scheduler: LRScheduler | None = plugin_manager.create_lr_scheduler(
+                trainer=self,
+                optimizer=optimizer,
+                num_training_steps=num_training_steps
+            )
+            if lr_scheduler is not None:
+                LOG.info(f"Using plugin-created lr_scheduler: {lr_scheduler}")
+                self.lr_scheduler = lr_scheduler
+            elif self.args.alternate_lr_scheduler_type == "one_cycle":
                 num_warmup_steps = self.args.get_warmup_steps(num_training_steps)
                 pct_start = num_warmup_steps / num_training_steps
                 extra_lr_kwargs = {}
@@ -110,4 +120,4 @@ class SchedulerMixin(Trainer):
             if use_cosine_min_lr:
                 LOG.warning("axolotl's cosine scheduler with min lr not used (e.g., because of deepspeed).")
 
-        return self.lr_scheduler
+        return self.lr_scheduler  # type: ignore

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

@@ -1,374 +0,0 @@
-"""
-Module for Axolotl trainer sequence parallelism mixin and training context manager
-"""
-
-import functools
-import logging
-
-import torch
-import torch.distributed as dist
-from datasets import Dataset
-from torch import nn
-from torch.utils.data import DistributedSampler, Sampler
-from torch.utils.hooks import RemovableHandle
-
-from axolotl.monkeypatch.attention.ring_attn import (
-    get_ring_attn_group,
-    update_ring_attn_params,
-)
-from axolotl.utils.schemas.enums import RingAttnFunc
-
-LOG = logging.getLogger(__name__)
-
-
-def _handle_logits_to_keep(
-    logits_to_keep,
-    local_rank: int,
-    local_world_size: int,
-    ring_attn_func: RingAttnFunc,
-    total_seq_len: int,
-):
-    """
-    Handle logits_to_keep parameter for sequence parallelism.
-
-    Args:
-        logits_to_keep: Integer or tensor indicating which positions to compute logits
-            for.
-        local_rank: Rank in the sequence parallel group.
-        local_world_size: World size of the sequence parallel group.
-        ring_attn_func: Ring attention function being used.
-        total_seq_len: Full sequence length.
-
-    Returns:
-        Adjusted logits_to_keep appropriate for this rank's sharded sequence
-    """
-    print("start of _handle_logits_to_keep")
-    print(dist.get_rank(), logits_to_keep)
-
-    # No transformation needed if logits_to_keep is None
-    if logits_to_keep is None:
-        return None
-
-    assert isinstance(
-        logits_to_keep, int
-    ), "sequence parallelism currently only supports integer logits_to_keep"
-    assert ring_attn_func in [
-        RingAttnFunc.VARLEN_LLAMA3,
-        RingAttnFunc.BATCH_RING,
-    ], "if specifying logits_to_keep, sequence parallelism currently only supports 'batch_ring' and 'varlen_llama3' `ring_attn_func`s"
-
-    # For standard sharding, each rank gets a contiguous chunk
-    chunk_size = total_seq_len // local_world_size
-    start_idx = local_rank * chunk_size
-    end_idx = start_idx + chunk_size
-
-    # Check if logits_to_keep is in this rank's range
-    if start_idx <= logits_to_keep < end_idx:
-        print("end of _handle_logits_to_keep")
-        print(dist.get_rank(), logits_to_keep - start_idx)
-        return logits_to_keep - start_idx
-    else:
-        print("end of _handle_logits_to_keep")
-        print(dist.get_rank(), -1)
-        return -1
-
-
-def apply_sequence_parallelism(
-    batch: dict[str, torch.Tensor],
-    local_rank: int,
-    local_world_size: int,
-    ring_attn_func: RingAttnFunc,
-) -> dict[str, torch.Tensor]:
-    """
-    Apply sequence parallelism slicing to a batch.
-
-    Args:
-        batch: Batch dictionary (e.g., input_ids, attention_mask, etc.).
-        local_rank: Local rank in the sequence parallel group.
-        local_world_size: World size of the sequence parallel group.
-        ring_attn_func: The ring attention function to use.
-
-    Returns:
-        Sliced batch dictionary.
-    """
-    # Update ring attention params if needed
-    if batch.get("position_ids") is not None:
-        update_ring_attn_params(position_ids=batch["position_ids"])
-
-    # Slice batch for sequence parallel processing
-    total_seq_len = batch["input_ids"].size(1)
-    for key in batch:
-        if (
-            isinstance(batch[key], torch.Tensor)
-            and batch[key].dim() > 1
-            and batch[key].size(1) == total_seq_len
-        ):
-            if ring_attn_func in [
-                RingAttnFunc.VARLEN_LLAMA3,
-                RingAttnFunc.BATCH_RING,
-            ]:
-                # Split in sequential fashion and grab this rank's chunk
-                batch[key] = (
-                    batch[key].chunk(local_world_size, dim=1)[local_rank].contiguous()
-                )
-            elif ring_attn_func is RingAttnFunc.BATCH_ZIGZAG:
-                chunks = batch[key].chunk(2 * local_world_size, dim=1)
-
-                # Take rank's chunk and opposing chunk for zigzag pattern
-                selected_chunks = [
-                    chunks[local_rank],
-                    chunks[2 * local_world_size - local_rank - 1],
-                ]
-                batch[key] = torch.cat(selected_chunks, dim=1).contiguous()
-            elif ring_attn_func is RingAttnFunc.BATCH_STRIPE:
-                # Split into striped data and stack
-                tensor = torch.stack(
-                    batch[key].split(local_world_size, dim=1),
-                    dim=1,
-                ).transpose(1, 2)
-                batch[key] = tensor[:, local_rank].contiguous()
-        if key == "logits_to_keep":
-            batch[key] = _handle_logits_to_keep(
-                logits_to_keep=batch[key],
-                local_rank=local_rank,
-                local_world_size=local_world_size,
-                ring_attn_func=ring_attn_func,
-                total_seq_len=total_seq_len,
-            )
-
-    return batch
-
-
-class SequenceParallelMixin:
-    """
-    Mixin class for sequence parallelism support in trainers.
-
-    This mixin provides functionality for handling sequence parallelism,
-    specifically for creating appropriate data samplers.
-    """
-
-    args = None  # type: "AxolotlTrainingArguments"  # type: ignore[name-defined]
-
-    def _setup_sequence_parallel(self):
-        """Set up sequence parallelism environment."""
-        self.ring_attn_group = get_ring_attn_group()
-
-    def _create_sequence_parallel_sampler(
-        self,
-        dataset: Dataset,
-        shuffle: bool = True,
-        is_eval: bool = False,
-    ) -> DistributedSampler:
-        """
-        Helper method to create sampler for sequence parallelism (SP).
-
-        We create a distributed sampler with rank equal to the SP group ID, which
-        means that all ranks in the SP group receive the same sample / set of samples
-        per training step. We also set the number of replicas equal to the number of
-        SP groups, which is a bit of a hack / unintended use, but works!
-
-        Args:
-            dataset: Dataset to sample from.
-            shuffle: Whether to shuffle the dataset.
-            is_eval: Whether we are creating a sampler for evaluation or training.
-
-        Returns:
-            Distributed sampler.
-        """
-        num_sp_groups = self.args.world_size // self.args.sequence_parallel_degree
-        sp_group_id = dist.get_rank() // self.args.sequence_parallel_degree
-
-        return DistributedSampler(
-            dataset,
-            num_replicas=num_sp_groups,
-            rank=sp_group_id,
-            seed=self.args.seed if shuffle else None,
-            shuffle=shuffle,
-            drop_last=not is_eval,
-        )
-
-    def _sp_get_train_sampler(self, dataset) -> Sampler | None:
-        """
-        Get a training sampler configured for sequence parallelism.
-
-        Args:
-            dataset: The training dataset
-
-        Returns:
-            Configured sequence parallel sampler.
-        """
-        return self._create_sequence_parallel_sampler(
-            dataset,
-            shuffle=not self.args.curriculum_sampling,
-        )
-
-    def _sp_get_eval_sampler(self, eval_dataset) -> Sampler | None:
-        """
-        Get an evaluation sampler configured for sequence parallelism.
-
-        Args:
-            eval_dataset: The evaluation dataset.
-
-        Returns:
-            Configured sequence parallel sampler.
-        """
-        return self._create_sequence_parallel_sampler(
-            eval_dataset, shuffle=False, is_eval=True
-        )
-
-
-class SequenceParallelContextManager:
-    """
-    Context manager for sequence parallelism operations.
-
-    This class provides a context that will automatically apply sequence parallelism
-    during model forward passes using a pre-forward hook, and gather outputs from
-    across the sequence parallelism group using a post-forward hook.
-    """
-
-    def __init__(
-        self,
-        model: nn.Module,
-        sequence_parallel_degree: int,
-        ring_attn_func: RingAttnFunc,
-    ):
-        self.model = model
-        self.sequence_parallel_degree = sequence_parallel_degree
-        self.ring_attn_func = ring_attn_func
-        self.process_group = get_ring_attn_group()
-
-        # Initialize sequence parallel group details
-        self.local_rank = dist.get_rank(self.process_group)
-        self.local_world_size = dist.get_world_size(self.process_group)
-
-        # Will store hook handles for removal
-        self.hook_handles: list[RemovableHandle] = []
-
-        # Create a partially applied version of the apply_sequence_parallelism function
-        # with pre-configured params
-        self.apply_sequence_parallelism = functools.partial(
-            apply_sequence_parallelism,
-            local_rank=self.local_rank,
-            local_world_size=self.local_world_size,
-            ring_attn_func=self.ring_attn_func,
-        )
-
-    def __enter__(self):
-        # Forward pre-hook to apply sequence parallelism
-        def sequence_parallel_pre_hook(_, args, kwargs):
-            # Apply sequence parallelism to kwargs
-            kwargs = self.apply_sequence_parallelism(batch=kwargs)
-            return args, kwargs
-
-        # Forward post-hook to gather outputs
-        def sequence_parallel_post_hook(_, __, output):
-            print("start of sequence_parallel_post_hook")
-            # Gather the sharded outputs
-            output = self.gather_outputs(output)
-            print("end of sequence_parallel_post_hook")
-            return output
-
-        # Register both hooks
-        self.hook_handles.append(
-            self.model.register_forward_pre_hook(
-                sequence_parallel_pre_hook, with_kwargs=True
-            )
-        )
-        self.hook_handles.append(
-            self.model.register_forward_hook(sequence_parallel_post_hook)
-        )
-
-        return self
-
-    def __exit__(self, exc_type, exc_val, exc_tb):
-        # Remove all hooks
-        for handle in self.hook_handles:
-            handle.remove()
-        self.hook_handles = []
-
-    def gather_outputs(self, output):
-        """Gather sharded outputs from all ranks and reconstruct the full tensor."""
-        # Handle different output formats (dict, tensor, etc.)
-        if isinstance(output, dict):
-            gathered_output = {}
-            for key, value in output.items():
-                if isinstance(value, torch.Tensor) and value.dim() > 1:
-                    # Gather logits or other sequence-sharded tensors
-                    gathered_value = self.gather_tensor(value)
-                    gathered_output[key] = gathered_value
-                else:
-                    gathered_value = value.clone()
-                    dist.all_reduce(
-                        gathered_value, op=dist.ReduceOp.SUM, group=self.process_group
-                    )
-                    gathered_output[key] = gathered_value
-            return gathered_output
-        if isinstance(output, torch.Tensor):
-            return self.gather_tensor(output)
-
-        return output
-
-    def gather_tensor(self, tensor):
-        """Gather a sharded tensor from all ranks."""
-        # Prepare tensors for all_gather
-        world_size = self.local_world_size
-
-        # Create list to store tensors from all ranks
-        gathered_tensors = [torch.zeros_like(tensor) for _ in range(world_size)]
-
-        # All-gather operation
-        dist.all_gather(gathered_tensors, tensor, group=self.process_group)
-
-        # Concatenate along sequence dimension (typically dim=1)
-        if self.ring_attn_func in [RingAttnFunc.VARLEN_LLAMA3, RingAttnFunc.BATCH_RING]:
-            # Simple concatenation for standard sharding
-            return torch.cat(gathered_tensors, dim=1)
-
-        if self.ring_attn_func is RingAttnFunc.BATCH_ZIGZAG:
-            # Each rank has a pattern of (rank, world_size*2-rank-1)
-            reconstituted_tensors = [None] * (world_size * 2)
-
-            # First, split each gathered tensor into its two chunks
-            for rank, gathered_tensor in enumerate(gathered_tensors):
-                # Each tensor contains two chunks in the sequence dimension
-                chunk_size = gathered_tensor.size(1) // 2
-                chunk1, chunk2 = gathered_tensor.split(chunk_size, dim=1)
-
-                # Place chunks in their original positions
-                reconstituted_tensors[rank] = chunk1
-                reconstituted_tensors[world_size * 2 - rank - 1] = chunk2
-
-            # Concatenate the reconstituted tensors in the correct order
-            return torch.cat(reconstituted_tensors, dim=1)
-
-        # Otherwise, RingAttnFunc.BATCH_STRIPE
-        # In striping, each rank has every world_size-th slice
-        batch_size = tensor.size(0)
-        hidden_dim = tensor.size(-1)
-
-        # First, determine the full sequence length
-        total_seq_len = 0
-        for t in gathered_tensors:
-            total_seq_len += t.size(1)
-
-        # Create a tensor to hold the unstriped result
-        result = torch.zeros(
-            batch_size,
-            total_seq_len,
-            hidden_dim,
-            dtype=tensor.dtype,
-            device=tensor.device,
-        )
-
-        # For each rank's tensor, distribute its slices to the correct positions
-        for rank, gathered_tensor in enumerate(gathered_tensors):
-            # The rank's tensor contains every world_size-th slice
-            # starting from its rank position
-            seq_len = gathered_tensor.size(1)
-            for i in range(seq_len):
-                # Calculate the position in the full tensor
-                pos = i * world_size + rank
-                if pos < total_seq_len:
-                    result[:, pos] = gathered_tensor[:, i]
-
-        return result

src/axolotl/core/trainers/relora.py

@@ -1,6 +1,7 @@
 """Module for ReLoRA trainer"""
 
 import torch
+from torch.optim.lr_scheduler import LRScheduler
 
 from axolotl.core.trainers.base import AxolotlTrainer
 from axolotl.monkeypatch.relora import ReLoRAScheduler
@@ -19,9 +20,11 @@ class ReLoRATrainer(AxolotlTrainer):
         self,
         num_training_steps: int,
         optimizer: torch.optim.Optimizer | None = None,
-    ):
+    ) -> LRScheduler:
         optimizer = self.optimizer if optimizer is None else optimizer
-        lr_scheduler = super().create_scheduler(num_training_steps, optimizer)
+        lr_scheduler: LRScheduler = super().create_scheduler(
+            num_training_steps, optimizer
+        )
 
         if self.args.relora_steps:
             warmup_steps = (
@@ -30,7 +33,7 @@ class ReLoRATrainer(AxolotlTrainer):
             anneal_steps = (
                 self.args.relora_anneal_steps if self.args.relora_anneal_steps else 1
             )
-            self.lr_scheduler = ReLoRAScheduler(
+            self.lr_scheduler = ReLoRAScheduler(  # type: ignore
                 optimizer,
                 lr_scheduler,
                 self.args.relora_steps,
@@ -38,6 +41,6 @@ class ReLoRATrainer(AxolotlTrainer):
                 warmup_steps,
             )
         else:
-            self.lr_scheduler = lr_scheduler
+            self.lr_scheduler = lr_scheduler  # type: ignore
 
-        return self.lr_scheduler
+        return self.lr_scheduler  # type: ignore

src/axolotl/core/training_args.py

@@ -9,8 +9,6 @@ from PIL.Image import Resampling
 from transformers import TrainingArguments
 from trl import CPOConfig, KTOConfig, ORPOConfig, PRMConfig, RewardConfig
 
-from axolotl.utils.schemas.enums import RingAttnFunc
-
 
 @dataclass
 class AxolotlTrainingMixins:
@@ -216,14 +214,16 @@ class AxolotlTrainingMixins:
         },
     )
 
-    sequence_parallel_degree: Optional[int] = field(
-        default=1,
-        metadata={"help": "The number of workers to use in sequence parallelism"},
-    )
-    ring_attn_func: Optional[RingAttnFunc] = field(
+    adam_beta3: Optional[float] = field(
         default=None,
         metadata={
-            "help": "The ring-flash-attn function to use in sequence parallelism"
+            "help": "The beta3 hyperparameter used in some optimizers such as CAME"
+        },
+    )
+    adam_epsilon2: Optional[float] = field(
+        default=None,
+        metadata={
+            "help": "The epsilon2 hyperparameter used in some optimizers such as CAME"
         },
     )
 

src/axolotl/evaluate.py

@@ -11,20 +11,19 @@ from accelerate.logging import get_logger
 from datasets import Dataset
 from transformers.trainer import Trainer
 
-from axolotl.logging_config import configure_logging
-from axolotl.train import TrainDatasetMeta
-from axolotl.utils import set_pytorch_cuda_alloc_conf
+from axolotl.train import (
+    TrainDatasetMeta,
+    setup_model_and_tokenizer,
+)
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import cleanup_distributed
-from axolotl.utils.models import load_model, load_processor, load_tokenizer
 from axolotl.utils.trainer import setup_trainer
 
 project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
 src_dir = os.path.join(project_root, "src")
 sys.path.insert(0, src_dir)
 
-configure_logging()
-LOG = get_logger("axolotl.evaluate")
+LOG = get_logger(__name__)
 
 
 def evaluate_dataset(
@@ -75,37 +74,22 @@ def evaluate(*, cfg: DictDefault, dataset_meta: TrainDatasetMeta) -> Dict[str, f
     Returns:
         Dictionary mapping metric names to their values.
     """
-    # pylint: disable=duplicate-code
-    # Enable expandable segments for cuda allocation to improve VRAM usage
-    set_pytorch_cuda_alloc_conf()
-
-    # Load tokenizer
-    LOG.debug(
-        f"loading tokenizer... {cfg.tokenizer_config or cfg.base_model_config}",
-        main_process_only=True,
-    )
-    tokenizer = load_tokenizer(cfg)
-
-    # Load processor for multimodal models if needed
-    processor = None
-    if cfg.is_multimodal:
-        processor = load_processor(cfg, tokenizer)
+    # Load tokenizer, processor and model
+    LOG.debug("loading model for evaluation...")
+    model, tokenizer, _, processor = setup_model_and_tokenizer(cfg)
 
     # Get datasets
+    # pylint: disable=duplicate-code
     train_dataset = dataset_meta.train_dataset
     eval_dataset = dataset_meta.eval_dataset
     total_num_steps = dataset_meta.total_num_steps
 
-    # Load model
-    LOG.debug("loading model for evaluation...")
-    model, _ = load_model(cfg, tokenizer, processor=processor)
-
     # Set up trainer
     trainer = setup_trainer(
-        cfg,
+        cfg=cfg,
         train_dataset=train_dataset,
         eval_dataset=eval_dataset,
-        model=(model, None, None),  # No need for model_ref or peft_config
+        model=model,
         tokenizer=tokenizer,
         processor=processor,
         total_num_steps=total_num_steps,

src/axolotl/integrations/base.py

@@ -10,216 +10,240 @@
 # License for the specific language governing permissions and limitations under
 # the License.
 
-"""
-Base class for all plugins.
+"""Base class for all plugins.
 
 A plugin is a reusable, modular, and self-contained piece of code that extends the functionality of Axolotl.
 Plugins can be used to integrate third-party models, modify the training process, or add new features.
 
 To create a new plugin, you need to inherit from the BasePlugin class and implement the required methods.
 """
+
+from __future__ import annotations
+
 import collections
 import importlib
 import logging
-from typing import OrderedDict
+from typing import TYPE_CHECKING, Callable, OrderedDict, Union
 
-import torch
+from peft import PeftModel
+from torch.optim import Optimizer
+from torch.optim.lr_scheduler import LRScheduler
+from transformers import PreTrainedModel, Trainer
+
+from axolotl.utils.dict import DictDefault
+
+if TYPE_CHECKING:
+    from axolotl.common.datasets import TrainDatasetMeta
 
 
 class BasePlugin:
-    """
-    Base class for all plugins. Defines the interface for plugin methods.
-
-    Attributes:
-    None
+    """Base class for all plugins. Defines the interface for plugin methods.
 
     Methods:
-    register(cfg): Registers the plugin with the given configuration.
-    pre_model_load(cfg): Performs actions before the model is loaded.
-    post_model_load(cfg, model): Performs actions after the model is loaded.
-    pre_lora_load(cfg, model): Performs actions before LoRA weights are loaded.
-    post_lora_load(cfg, model): Performs actions after LoRA weights are loaded.
-    create_optimizer(cfg, trainer): Creates and returns an optimizer for training.
-    create_lr_scheduler(cfg, trainer, optimizer): Creates and returns a learning rate scheduler.
-    add_callbacks_pre_trainer(cfg, model): Adds callbacks to the trainer before training.
-    add_callbacks_post_trainer(cfg, trainer): Adds callbacks to the trainer after training.
+        register(cfg): Registers the plugin with the given configuration.
+        load_datasets(cfg): Loads and preprocesses the dataset for training.
+        pre_model_load(cfg): Performs actions before the model is loaded.
+        post_model_build(cfg, model): Performs actions after the model is loaded, but
+            before LoRA adapters are applied.
+        pre_lora_load(cfg, model): Performs actions before LoRA weights are loaded.
+        post_lora_load(cfg, model): Performs actions after LoRA weights are loaded.
+        post_model_load(cfg, model): Performs actions after the model is loaded,
+            inclusive of any adapters.
+        post_trainer_create(cfg, trainer): Performs actions after the trainer is
+            created.
+        create_optimizer(cfg, trainer): Creates and returns an optimizer for training.
+        create_lr_scheduler(cfg, trainer, optimizer, num_training_steps): Creates and
+            returns a learning rate scheduler.
+        add_callbacks_pre_trainer(cfg, model): Adds callbacks to the trainer before
+            training.
+        add_callbacks_post_trainer(cfg, trainer): Adds callbacks to the trainer after
+            training.
     """
 
     def __init__(self):
-        """
-        Initializes the BasePlugin.
-        """
+        """Initializes the BasePlugin."""
 
     def register(self, cfg):  # pylint: disable=unused-argument
-        """
-        Registers the plugin with the given configuration.
+        """Registers the plugin with the given configuration.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
+        Args:
+            cfg: The configuration for the plugin.
+        """
+
+    def get_input_args(self) -> str | None:
+        """Returns a pydantic model for the plugin's input arguments."""
+
+    def load_datasets(
+        self, cfg: DictDefault, preprocess: bool = False
+    ) -> Union["TrainDatasetMeta", None]:
+        """Loads and preprocesses the dataset for training.
+
+        Args:
+            cfg: The configuration for the plugin.
+            preprocess: Whether this is the preprocess step of the datasets.
 
         Returns:
-        None
+            dataset_meta: The metadata for the training dataset.
         """
 
-    def get_input_args(self):
-        """
-        Returns a pydantic model for the plugin's input arguments.
+    def pre_model_load(self, cfg: DictDefault):  # pylint: disable=unused-argument
+        """Performs actions before the model is loaded.
+
+        Args:
+            cfg: The configuration for the plugin.
         """
 
-    def pre_model_load(self, cfg):  # pylint: disable=unused-argument
-        """
-        Performs actions before the model is loaded.
+    # pylint: disable=unused-argument
+    def post_model_build(self, cfg: DictDefault, model: PreTrainedModel):
+        """Performs actions after the model is built/loaded, but before any adapters are applied.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
+        Args:
+            cfg: The configuration for the plugin.
+        """
+
+    # pylint: disable=unused-argument
+    def pre_lora_load(self, cfg: DictDefault, model: PreTrainedModel):
+        """Performs actions before LoRA weights are loaded.
+
+        Args:
+            cfg: The configuration for the plugin.
+            model: The loaded model.
+        """
+
+    # pylint: disable=unused-argument
+    def post_lora_load(self, cfg: DictDefault, model: PreTrainedModel | PeftModel):
+        """Performs actions after LoRA weights are loaded.
+
+        Args:
+            cfg: The configuration for the plugin.
+            model: The loaded model.
+        """
+
+    # pylint: disable=unused-argument
+    def post_model_load(self, cfg: DictDefault, model: PreTrainedModel | PeftModel):
+        """Performs actions after the model is loaded.
+
+        Args:
+            cfg: The configuration for the plugin.
+            model: The loaded model.
+        """
+
+    # pylint: disable=unused-argument
+    def get_trainer_cls(self, cfg: DictDefault) -> Trainer | None:
+        """Returns a custom class for the trainer.
+
+        Args:
+            cfg: The global axolotl configuration.
 
         Returns:
-        None
+            The first non-`None` trainer class returned by a plugin.
         """
 
-    def post_model_load(self, cfg, model):  # pylint: disable=unused-argument
-        """
-        Performs actions after the model is loaded.
+    # pylint: disable=unused-argument
+    def post_trainer_create(self, cfg: DictDefault, trainer: Trainer):
+        """Performs actions after the trainer is created.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        model (object): The loaded model.
+        Args:
+            cfg: The configuration for the plugin.
+            trainer: The trainer object for training.
+        """
+
+    # pylint: disable=unused-argument
+    def create_optimizer(self, cfg: DictDefault, trainer: Trainer) -> Optimizer | None:
+        """Creates and returns an optimizer for training.
+
+        Args:
+            cfg: The configuration for the plugin.
+            trainer: The trainer object for training.
 
         Returns:
-        None
-        """
-
-    def pre_lora_load(self, cfg, model):  # pylint: disable=unused-argument
-        """
-        Performs actions before LoRA weights are loaded.
-
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        model (object): The loaded model.
-
-        Returns:
-        None
-        """
-
-    def post_lora_load(self, cfg, model):  # pylint: disable=unused-argument
-        """
-        Performs actions after LoRA weights are loaded.
-
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        model (object): The loaded model.
-
-        Returns:
-        None
-        """
-
-    def get_trainer_cls(self, cfg):  # pylint: disable=unused-argument):
-        """
-        Returns a custom class for the trainer.
-
-        Parameters:
-        cfg (dict): The global axolotl configuration.
-
-        Returns:
-        class: The class for the trainer.
-        """
-
-    def create_optimizer(self, cfg, trainer):  # pylint: disable=unused-argument
-        """
-        Creates and returns an optimizer for training.
-
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        trainer (object): The trainer object for training.
-
-        Returns:
-        object: The created optimizer.
+            The created optimizer.
         """
 
+    # pylint: disable=unused-argument
     def create_lr_scheduler(
-        self, cfg, trainer, optimizer
-    ):  # pylint: disable=unused-argument
-        """
-        Creates and returns a learning rate scheduler.
+        self,
+        cfg: DictDefault,
+        trainer: Trainer,
+        optimizer: Optimizer,
+        num_training_steps: int,
+    ) -> LRScheduler | None:
+        """Creates and returns a learning rate scheduler.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        trainer (object): The trainer object for training.
-        optimizer (object): The optimizer for training.
+        Args:
+            cfg: The configuration for the plugin.
+            trainer: The trainer object for training.
+            optimizer: The optimizer for training.
+            num_training_steps: Total number of training steps
 
         Returns:
-        object: The created learning rate scheduler.
+            The created learning rate scheduler.
         """
 
-    def add_callbacks_pre_trainer(self, cfg, model):  # pylint: disable=unused-argument
-        """
-        setup callbacks before creating the trainer.
+    # pylint: disable=unused-argument
+    def add_callbacks_pre_trainer(
+        self, cfg: DictDefault, model: PreTrainedModel
+    ) -> list[Callable]:
+        """Set up callbacks before creating the trainer.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        model (object): The loaded model.
+        Args:
+            cfg: The configuration for the plugin.
+            model: The loaded model.
 
         Returns:
-        List[callable]: A list of callback functions to be added to the TrainingArgs
+            A list of callback functions to be added to the `TrainingArgs`.
         """
         return []
 
+    # pylint: disable=unused-argument
     def add_callbacks_post_trainer(
-        self, cfg, trainer
-    ):  # pylint: disable=unused-argument
-        """
-        Adds callbacks to the trainer after creating the trainer.
-        This is useful for callbacks that require access to the model or trainer.
+        self, cfg: DictDefault, trainer: Trainer
+    ) -> list[Callable]:
+        """Adds callbacks to the trainer after creating the trainer. This is useful for
+        callbacks that require access to the model or trainer.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-        trainer (object): The trainer object for training.
+        Args:
+            cfg: The configuration for the plugin.
+            trainer: The trainer object for training.
 
         Returns:
-        List[callable]: A list of callback functions to be added
+            A list of callback functions to be added
         """
         return []
 
-    def post_train(self, cfg, model):  # pylint: disable=unused-argument
-        """
-        Performs actions after training is complete.
+    # pylint: disable=unused-argument
+    def post_train(self, cfg: DictDefault, model: PreTrainedModel | PeftModel):
+        """Performs actions after training is complete.
 
-        Parameters:
-        cfg (dict): The axolotl configuration
-        model (object): The loaded model.
-
-        Returns:
-        None
+        Args:
+            cfg: The axolotl configuration.
+            model: The loaded model.
         """
 
-    def post_train_unload(self, cfg):  # pylint: disable=unused-argument
-        """
-        Performs actions after training is complete and the model is unloaded.
+    def post_train_unload(self, cfg: DictDefault):  # pylint: disable=unused-argument
+        """Performs actions after training is complete and the model is unloaded.
 
-        Parameters:
-        cfg (dict): The configuration for the plugin.
-
-        Returns:
-        None
+        Args:
+            cfg: The configuration for the plugin.
         """
 
 
 def load_plugin(plugin_name: str) -> BasePlugin:
-    """
-    Loads a plugin based on the given plugin name.
+    """Loads a plugin based on the given plugin name.
 
-    The plugin name should be in the format "module_name.class_name".
-    This function splits the plugin name into module and class, imports the module,
-    retrieves the class from the module, and creates an instance of the class.
+    The plugin name should be in the format "module_name.class_name". This function
+    splits the plugin name into module and class, imports the module, retrieves the
+    class from the module, and creates an instance of the class.
 
-    Parameters:
-    plugin_name (str): The name of the plugin to be loaded. The name should be in the format "module_name.class_name".
+    Args:
+        plugin_name: The name of the plugin to be loaded. The name should be in the
+            format "module_name.class_name".
 
     Returns:
-    BasePlugin: An instance of the loaded plugin.
+        An instance of the loaded plugin.
 
     Raises:
-    ImportError: If the plugin module cannot be imported.
+        ImportError: If the plugin module cannot be imported.
     """
     # split the plugin name into module and class
     module_name, class_name = plugin_name.rsplit(".", 1)
@@ -245,54 +269,57 @@ def load_plugin(plugin_name: str) -> BasePlugin:
 
 
 class PluginManager:
-    """
-    The PluginManager class is responsible for loading and managing plugins.
-    It should be a singleton so it can be accessed from anywhere in the codebase.
+    """The `PluginManager` class is responsible for loading and managing plugins. It
+    should be a singleton so it can be accessed from anywhere in the codebase.
 
     Attributes:
-    plugins (List[BasePlugin]): A list of loaded plugins.
+        plugins: A list of loaded plugins.
 
     Methods:
-    get_instance(): Static method to get the singleton instance of PluginManager.
-    register(plugin_name: str): Registers a new plugin by its name.
-    pre_model_load(cfg): Calls the pre_model_load method of all registered plugins.
+        get_instance(): Static method to get the singleton instance of `PluginManager`.
+        register(plugin_name: str): Registers a new plugin by its name.
+        pre_model_load(cfg): Calls the pre_model_load method of all registered plugins.
     """
 
     plugins: OrderedDict[str, BasePlugin] = collections.OrderedDict()
 
-    _instance = None
+    _instance: PluginManager | None = None
+    _cfg: DictDefault | None = None
 
     def __new__(cls):
-        """
-        Creates a new instance of PluginManager if it doesn't exist yet.
-        """
+        """Creates a new instance of PluginManager if it doesn't exist yet."""
         if cls._instance is None:
             cls._instance = super(PluginManager, cls).__new__(cls)
-            cls._instance.plugins = collections.OrderedDict()
+            cls._instance.plugins: OrderedDict[str, BasePlugin] = (
+                collections.OrderedDict()
+            )
         return cls._instance
 
     @staticmethod
     def get_instance() -> "PluginManager":
-        """
-        Returns the singleton instance of PluginManager.
-        If the instance doesn't exist, it creates a new one.
+        """Returns the singleton instance of PluginManager. If the instance doesn't
+        exist, it creates a new one.
         """
         if PluginManager._instance is None:
             PluginManager()
         return PluginManager._instance  # type: ignore
 
+    @property
+    def cfg(self):
+        return self._cfg
+
+    @cfg.setter
+    def cfg(self, cfg):
+        self._cfg = cfg
+
     def register(self, plugin_name: str):
-        """
-        Registers a new plugin by its name.
+        """Registers a new plugin by its name.
 
-        Parameters:
-        plugin_name (str): The name of the plugin to be registered.
-
-        Returns:
-        None
+        Args:
+            plugin_name: The name of the plugin to be registered.
 
         Raises:
-        ImportError: If the plugin module cannot be imported.
+            ImportError: If the plugin module cannot be imported.
         """
         try:
             logging.info(f"Attempting to load plugin: {plugin_name}")
@@ -302,12 +329,11 @@ class PluginManager:
         except ImportError:
             logging.error(f"Failed to load plugin: {plugin_name}")
 
-    def get_input_args(self):
-        """
-        Returns a list of Pydantic classes for all registered plugins' input arguments.'
+    def get_input_args(self) -> list[str]:
+        """Returns a list of Pydantic classes for all registered plugins' input arguments.'
 
         Returns:
-        list[str]: A list of Pydantic classes for all registered plugins' input arguments.'
+            A list of Pydantic classes for all registered plugins' input arguments.'
         """
         input_args = []
         for plugin in self.plugins.values():
@@ -316,70 +342,88 @@ class PluginManager:
                 input_args.append(input_args_from_plugin)
         return input_args
 
-    def pre_model_load(self, cfg):
-        """
-        Calls the pre_model_load method of all registered plugins.
+    def load_datasets(
+        self, cfg: DictDefault, preprocess: bool = False
+    ) -> Union["TrainDatasetMeta", None]:
+        """Calls the load_datasets method of each registered plugin.
 
-        Parameters:
-        cfg (dict): The configuration for the plugins.
+        Args:
+            cfg: The configuration for the plugins.
+            preprocess: Whether this is preprocess step of the datasets.
 
         Returns:
-        None
+            The dataset metadata loaded from all registered plugins.
+        """
+        return_ds_meta = None
+        for plugin in self.plugins.values():
+            dataset_meta = plugin.load_datasets(cfg, preprocess)
+            if dataset_meta is not None:
+                if return_ds_meta is None:
+                    return_ds_meta = dataset_meta
+                else:
+                    raise RuntimeError("Multiple plugins loaded datasets")
+        return return_ds_meta
+
+    def pre_model_load(self, cfg: DictDefault):
+        """Calls the pre_model_load method of all registered plugins.
+
+        Args:
+            cfg: The configuration for the plugins.
         """
         for plugin in self.plugins.values():
             plugin.pre_model_load(cfg)
 
-    def post_model_load(self, cfg, model):
-        """
-        Calls the post_model_load method of all registered plugins.
+    def post_model_build(self, cfg: DictDefault, model: PreTrainedModel):
+        """Calls the `post_model_build` method of all registered plugins after the
+        model has been built / loaded, but before any adapters have been applied.
 
-        Parameters:
-        cfg (dict): The configuration for the plugins.
-        model (object): The loaded model.
-
-        Returns:
-        None
+        Args:
+            cfg: The configuration for the plugins.
+            model: The loaded model.
         """
         for plugin in self.plugins.values():
-            plugin.post_model_load(cfg, model)
+            plugin.post_model_build(cfg, model)
 
-    def pre_lora_load(self, cfg, model):
-        """
-        Calls the pre_lora_load method of all registered plugins.
+    def pre_lora_load(self, cfg: DictDefault, model: PreTrainedModel):
+        """Calls the `pre_lora_load` method of all registered plugins.
 
-        Parameters:
-        cfg (dict): The configuration for the plugins.
-        model (object): The loaded model.
-
-        Returns:
-        None
+        Args:
+            cfg: The configuration for the plugins.
+            model: The loaded model.
         """
         for plugin in self.plugins.values():
             plugin.pre_lora_load(cfg, model)
 
-    def post_lora_load(self, cfg, model):
-        """
-        Calls the post_lora_load method of all registered plugins.
+    def post_lora_load(self, cfg: DictDefault, model: PreTrainedModel | PeftModel):
+        """Calls the `post_lora_load` method of all registered plugins.
 
-        Parameters:
-        cfg (dict): The configuration for the plugins.
-        model (object): The loaded model.
-
-        Returns:
-        None
+        Args:
+            cfg: The configuration for the plugins.
+            model: The loaded model.
         """
         for plugin in self.plugins.values():
             plugin.post_lora_load(cfg, model)
 
-    def get_trainer_cls(self, cfg):
-        """
-        Calls the get_trainer_cls method of all registered plugins and returns the first non-None trainer class.
+    def post_model_load(self, cfg: DictDefault, model: PreTrainedModel | PeftModel):
+        """Calls the `post_model_load` method of all registered plugins after the model
+        has been loaded inclusive of any adapters.
 
-        Parameters:
-        cfg (dict): The configuration for the plugins.
+        Args:
+            cfg: The configuration for the plugins.
+            model: The loaded model.
+        """
+        for plugin in self.plugins.values():
+            plugin.post_model_load(cfg, model)
+
+    def get_trainer_cls(self, cfg: DictDefault) -> Trainer | None:
+        """Calls the `get_trainer_cls` method of all registered plugins and returns the
+        first non-`None` trainer class.
+
+        Args:
+            cfg: The configuration for the plugins.
 
         Returns:
-        object: The trainer class, or None if none was found.
+            The first non-`None` trainer class returned by a plugin.
         """
         for plugin in self.plugins.values():
             trainer_cls = plugin.get_trainer_cls(cfg)
@@ -387,51 +431,67 @@ class PluginManager:
                 return trainer_cls
         return None
 
-    def create_optimizer(self, cfg, trainer):
-        """
-        Calls the create_optimizer method of all registered plugins and returns the first non-None optimizer.
+    def post_trainer_create(self, cfg: DictDefault, trainer: Trainer):
+        """Calls the `post_trainer_create` method of all registered plugins.
 
-        Parameters:
-        cfg (dict): The configuration for the plugins.
-        trainer (object): The trainer object for training.
-
-        Returns:
-        object: The created optimizer, or None if none was found.
+        Args:
+            cfg: The configuration for the plugins.
+            trainer: The trainer object for training.
         """
         for plugin in self.plugins.values():
-            optimizer = plugin.create_optimizer(cfg, trainer)
+            plugin.post_trainer_create(cfg, trainer)
+
+    def create_optimizer(self, trainer: Trainer) -> Optimizer | None:
+        """Calls the `create_optimizer` method of all registered plugins and returns
+        the first non-`None` optimizer.
+
+        Args:
+            trainer: The trainer object for training.
+
+        Returns:
+            The created optimizer, or `None` if none was found.
+        """
+        for plugin in self.plugins.values():
+            optimizer = plugin.create_optimizer(self.cfg, trainer)
             if optimizer is not None:
                 return optimizer
         return None
 
-    def create_lr_scheduler(self, cfg, trainer, optimizer):
-        """
-        Calls the create_lr_scheduler method of all registered plugins and returns the first non-None scheduler.
+    def create_lr_scheduler(
+        self, trainer: Trainer, optimizer: Optimizer, num_training_steps: int
+    ) -> LRScheduler | None:
+        """Calls the `create_lr_scheduler` method of all registered plugins and returns
+        the first non-`None` scheduler.
 
-        Parameters:
-        cfg (dict): The configuration for the plugins.
-        trainer (object): The trainer object for training.
-        optimizer (object): The optimizer for training.
+        Args:
+            trainer: The trainer object for training.
+            optimizer: The optimizer for training.
 
         Returns:
-        object: The created learning rate scheduler, or None if none was found.
+            The created learning rate scheduler, or `None` if not found.
         """
         for plugin in self.plugins.values():
-            scheduler = plugin.create_lr_scheduler(cfg, trainer, optimizer)
+            scheduler: LRScheduler | None = plugin.create_lr_scheduler(
+                self.cfg,
+                trainer=trainer,
+                optimizer=optimizer,
+                num_training_steps=num_training_steps,
+            )
             if scheduler is not None:
                 return scheduler
         return None
 
-    def add_callbacks_pre_trainer(self, cfg, model):
-        """
-        Calls the add_callbacks_pre_trainer method of all registered plugins.
+    def add_callbacks_pre_trainer(
+        self, cfg: DictDefault, model: PreTrainedModel
+    ) -> list[Callable]:
+        """Calls the add_callbacks_pre_trainer method of all registered plugins.
 
-        Parameters:
-        cfg (dict): The configuration for the plugins.
-        model (object): The loaded model.
+        Args:
+            cfg: The configuration for the plugins.
+            model: The loaded model.
 
         Returns:
-        List[callable]: A list of callback functions to be added to the TrainingArgs.
+            A list of callback functions to be added to the `TrainingArgs`.
         """
         callbacks = []
         for plugin in self.plugins.values():
@@ -440,16 +500,17 @@ class PluginManager:
                 callbacks.extend(plugin_callbacks)
         return callbacks
 
-    def add_callbacks_post_trainer(self, cfg, trainer):
-        """
-        Calls the add_callbacks_post_trainer method of all registered plugins.
+    def add_callbacks_post_trainer(
+        self, cfg: DictDefault, trainer: Trainer
+    ) -> list[Callable]:
+        """Calls the `add_callbacks_post_trainer` method of all registered plugins.
 
-        Parameters:
-        cfg (dict): The configuration for the plugins.
-        trainer (object): The trainer object for training.
+        Args:
+            cfg: The configuration for the plugins.
+            trainer: The trainer object for training.
 
         Returns:
-        List[callable]: A list of callback functions to be added to the TrainingArgs.
+            A list of callback functions to be added to the `TrainingArgs`.
         """
         callbacks = []
         for plugin in self.plugins.values():
@@ -458,27 +519,31 @@ class PluginManager:
                 callbacks.extend(plugin_callbacks)
         return callbacks
 
-    def post_train_unload(self, cfg):
+    def post_train(self, cfg: DictDefault, model: PreTrainedModel | PeftModel):
+        """Calls the post_train method of all registered plugins.
+
+        Args:
+            cfg: The configuration for the plugins.
+            model: The loaded model.
         """
-        Calls the post_train_unload method of all registered plugins.
+        for plugin in self.plugins.values():
+            plugin.post_train(cfg, model)
 
-        Parameters:
-        cfg (dict): The configuration for the plugins.
-        model (object): The loaded model.
+    def post_train_unload(self, cfg: DictDefault):
+        """Calls the post_train_unload method of all registered plugins.
 
-        Returns:
-        None
+        Args:
+            cfg: The configuration for the plugins.
+            model: The loaded model.
         """
         for plugin in self.plugins.values():
             plugin.post_train_unload(cfg)
 
 
 class BaseOptimizerFactory:
-    """
-    Base class for factories to create custom optimizers
-    """
+    """Base class for factories to create custom optimizers"""
 
     def __call__(
         self, opt_model, training_args, **optimizer_kwargs
-    ) -> "torch.optim.Optimizer":
+    ) -> Optimizer | None:
         pass

src/axolotl/integrations/cut_cross_entropy/README.md

@@ -27,15 +27,13 @@ pip3 uninstall -y cut-cross-entropy && pip3 install "cut-cross-entropy[transform
 ```yaml
 plugins:
   - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
-
-cut_cross_entropy: true
 ```
 
 ## Supported Models
 
 - llama
-- llama4_text
 - llama4
+- llama4_text
 - mllama
 - phi3
 - gemma
@@ -45,6 +43,11 @@ cut_cross_entropy: true
 - mistral
 - mistral3
 - qwen2
+- qwen2_moe
+- qwen2_vl
+- qwen2_5_vl
+- qwen3
+- qwen3_moe
 - cohere
 - cohere2
 - glm

src/axolotl/integrations/cut_cross_entropy/__init__.py

@@ -25,7 +25,7 @@ import torch
 
 from axolotl.integrations.base import BasePlugin
 from axolotl.utils import get_pytorch_version
-from axolotl.utils.distributed import zero_only
+from axolotl.utils.distributed import is_main_process
 
 from .args import CutCrossEntropyArgs  # pylint: disable=unused-import. # noqa: F401
 
@@ -76,7 +76,7 @@ class CutCrossEntropyPlugin(BasePlugin):
                 cce_patch,
             )
 
-            with zero_only():
+            if is_main_process(use_environ=True):
                 LOG.info(
                     f"Applying Cut Cross Entropy to model type: {cfg.model_config_type}"
                 )

src/axolotl/integrations/cut_cross_entropy/args.py

@@ -28,7 +28,7 @@ class CutCrossEntropyArgs(BaseModel):
     Input args for Cut Cross Entropy.
     """
 
-    cut_cross_entropy: Optional[bool] = None
+    cut_cross_entropy: Optional[bool] = True
 
     @model_validator(mode="before")
     @classmethod

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

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

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

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

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

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

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

@@ -0,0 +1,164 @@
+"""Llama CCE patch. Adapted from transformers v4.51.2"""
+
+# pylint: disable=duplicate-code
+
+
+from types import MethodType
+from typing import Optional, Union
+
+import torch
+import transformers
+from cut_cross_entropy.transformers.utils import (
+    PatchOptions,
+    TransformersModelT,
+    apply_lce,
+)
+from transformers.cache_utils import Cache
+from transformers.modeling_outputs import (
+    BaseModelOutputWithPast,
+    CausalLMOutputWithPast,
+)
+from transformers.models.llama.modeling_llama import (
+    KwargsForCausalLM,
+)
+from transformers.processing_utils import Unpack
+from transformers.utils.deprecation import deprecate_kwarg
+from transformers.utils.generic import can_return_tuple
+
+_PATCH_OPTS: PatchOptions | None = None
+
+
+@can_return_tuple
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+def cce_forward(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[Cache] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **kwargs: Unpack[KwargsForCausalLM],
+) -> CausalLMOutputWithPast:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, LlamaForCausalLM
+
+    >>> model = LlamaForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf")
+    >>> tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
+
+    >>> prompt = "Hey, are you conscious? Can you talk to me?"
+    >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+    ```"""
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs: BaseModelOutputWithPast = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs.last_hidden_state
+    if hidden_states is None:
+        raise ValueError("hidden_states is None")
+
+    loss = None
+    logits = None
+
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    slice_indices = (
+        slice(-logits_to_keep, None)
+        if isinstance(logits_to_keep, int)
+        else logits_to_keep
+    )
+    if _PATCH_OPTS is not None and _PATCH_OPTS.use_lce(labels, self.training):
+        assert labels is not None
+        loss = apply_lce(
+            hidden_states[:, slice_indices, :],
+            self.lm_head.weight,
+            labels,
+            _PATCH_OPTS,
+            **kwargs,
+        )
+    else:
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                **kwargs,
+            )
+
+    return CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )
+
+
+def patch_llama(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    """Patch Llama for CCE."""
+    global _PATCH_OPTS  # pylint: disable=global-statement
+    from transformers.models.llama import modeling_llama
+
+    _PATCH_OPTS = patch_options
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_llama.LlamaForCausalLM
+        ), f"Expected a LlamaForCausalLM model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward, maybe_model)
+        return maybe_model
+
+    modeling_llama.LlamaForCausalLM.forward = cce_forward
+    return None

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

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

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

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

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

@@ -5,9 +5,7 @@
 import transformers
 from cut_cross_entropy.cce_utils import LinearCrossEntropyImpl
 from cut_cross_entropy.linear_cross_entropy import LCE_IMPL_DEFAULT
-from cut_cross_entropy.transformers.llama import patch_llama
 from cut_cross_entropy.transformers.phi3 import patch_phi3
-from cut_cross_entropy.transformers.qwen2 import patch_qwen2
 from cut_cross_entropy.transformers.utils import PatchOptions, TransformersModelT
 
 from axolotl.integrations.cut_cross_entropy.monkeypatch.cohere import (
@@ -24,6 +22,9 @@ from axolotl.integrations.cut_cross_entropy.monkeypatch.glm4 import (
     patch_glm,
     patch_glm4,
 )
+from axolotl.integrations.cut_cross_entropy.monkeypatch.llama import (
+    patch_llama,
+)
 from axolotl.integrations.cut_cross_entropy.monkeypatch.llama4 import (
     patch_llama4,
     patch_llama4_text,
@@ -33,6 +34,22 @@ from axolotl.integrations.cut_cross_entropy.monkeypatch.mistral3 import (
     patch_mistral3,
 )
 from axolotl.integrations.cut_cross_entropy.monkeypatch.mllama import patch_mllama
+from axolotl.integrations.cut_cross_entropy.monkeypatch.qwen2 import (
+    patch_qwen2,
+)
+from axolotl.integrations.cut_cross_entropy.monkeypatch.qwen2_5_vl import (
+    patch_qwen2_5_vl,
+)
+from axolotl.integrations.cut_cross_entropy.monkeypatch.qwen2_moe import (
+    patch_qwen2_moe,
+)
+from axolotl.integrations.cut_cross_entropy.monkeypatch.qwen2_vl import (
+    patch_qwen2_vl,
+)
+from axolotl.integrations.cut_cross_entropy.monkeypatch.qwen3 import patch_qwen3
+from axolotl.integrations.cut_cross_entropy.monkeypatch.qwen3_moe import (
+    patch_qwen3_moe,
+)
 
 CUT_CROSS_ENTROPY_MODEL_MAPPING = {
     "llama": patch_llama,
@@ -47,6 +64,11 @@ CUT_CROSS_ENTROPY_MODEL_MAPPING = {
     "mistral": patch_mistral,
     "mistral3": patch_mistral3,
     "qwen2": patch_qwen2,
+    "qwen2_moe": patch_qwen2_moe,
+    "qwen2_vl": patch_qwen2_vl,
+    "qwen2_5_vl": patch_qwen2_5_vl,
+    "qwen3": patch_qwen3,
+    "qwen3_moe": patch_qwen3_moe,
     "cohere": patch_cohere,
     "cohere2": patch_cohere2,
     "glm": patch_glm,

src/axolotl/integrations/cut_cross_entropy/monkeypatch/qwen2.py

@@ -0,0 +1,37 @@
+"""Qwen2 CCE patch. The model inherits Llama's modeling code and uses the same forward method."""
+
+# pylint: disable=duplicate-code
+
+from types import MethodType
+
+import transformers
+from cut_cross_entropy.transformers.utils import (
+    PatchOptions,
+    TransformersModelT,
+)
+
+
+def patch_qwen2(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    from transformers.models.qwen2 import modeling_qwen2
+
+    # Set the _PATCH_OPTS in the llama patch file
+    import axolotl.integrations.cut_cross_entropy.monkeypatch.llama as llama_patch
+
+    llama_patch._PATCH_OPTS = patch_options  # pylint: disable=protected-access
+
+    from axolotl.integrations.cut_cross_entropy.monkeypatch.llama import (
+        cce_forward,
+    )
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_qwen2.Qwen2ForCausalLM
+        ), f"Expected a Qwen2ForCausalLM model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward, maybe_model)
+        return maybe_model
+
+    modeling_qwen2.Qwen2ForCausalLM.forward = cce_forward
+    return None

src/axolotl/integrations/cut_cross_entropy/monkeypatch/qwen2_5_vl.py

@@ -0,0 +1,246 @@
+"""Qwen2.5 VL CCE patch. Adapted from transformers v4.51.2"""
+
+# pylint: disable=duplicate-code
+
+
+from types import MethodType
+from typing import Optional, Tuple, Union
+
+import torch
+import transformers
+from cut_cross_entropy.transformers.utils import (
+    PatchOptions,
+    TransformersModelT,
+    apply_lce,
+)
+from torch.nn import CrossEntropyLoss
+from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
+    Qwen2_5_VLCausalLMOutputWithPast,
+)
+
+_PATCH_OPTS: PatchOptions | None = None
+
+
+def cce_forward_multimodal(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[list[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    pixel_values: Optional[torch.Tensor] = None,
+    pixel_values_videos: Optional[torch.FloatTensor] = None,
+    image_grid_thw: Optional[torch.LongTensor] = None,
+    video_grid_thw: Optional[torch.LongTensor] = None,
+    rope_deltas: Optional[torch.LongTensor] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    second_per_grid_ts: Optional[torch.Tensor] = None,
+) -> Union[Tuple, Qwen2_5_VLCausalLMOutputWithPast]:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from PIL import Image
+    >>> import requests
+    >>> from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration
+
+    >>> model = Qwen2_5_VLForConditionalGeneration.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct")
+    >>> processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct")
+
+    >>> messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "image"},
+                {"type": "text", "text": "What is shown in this image?"},
+            ],
+        },
+    ]
+    >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
+    >>> image = Image.open(requests.get(url, stream=True).raw)
+
+    >>> text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    >>> inputs = processor(text=[text], images=[image], vision_infos=[vision_infos])
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "The image shows a street scene with a red stop sign in the foreground. In the background, there is a large red gate with Chinese characters ..."
+    ```"""
+
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+
+    if inputs_embeds is None:
+        inputs_embeds = self.model.embed_tokens(input_ids)
+        if pixel_values is not None:
+            pixel_values = pixel_values.type(self.visual.dtype)
+            image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
+            n_image_tokens = (input_ids == self.config.image_token_id).sum().item()
+            n_image_features = image_embeds.shape[0]
+            if n_image_tokens != n_image_features:
+                raise ValueError(
+                    f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
+                )
+
+            mask = input_ids == self.config.image_token_id
+            mask_unsqueezed = mask.unsqueeze(-1)
+            mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
+            image_mask = mask_expanded.to(inputs_embeds.device)
+
+            image_embeds = image_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)  # type: ignore
+
+        if pixel_values_videos is not None:
+            pixel_values_videos = pixel_values_videos.type(self.visual.dtype)
+            video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw)
+            n_video_tokens = (input_ids == self.config.video_token_id).sum().item()
+            n_video_features = video_embeds.shape[0]
+            if n_video_tokens != n_video_features:
+                raise ValueError(
+                    f"Video features and video tokens do not match: tokens: {n_video_tokens}, features {n_video_features}"
+                )
+
+            mask = input_ids == self.config.video_token_id
+            mask_unsqueezed = mask.unsqueeze(-1)
+            mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
+            video_mask = mask_expanded.to(inputs_embeds.device)
+
+            video_embeds = video_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)  # type: ignore
+
+        if attention_mask is not None:
+            attention_mask = attention_mask.to(inputs_embeds.device)
+
+    # if we get 4D attention mask we cannot calculate rope deltas anymore. TODO @raushan fixme
+    if position_ids is None and (attention_mask is None or attention_mask.ndim == 2):
+        # calculate RoPE index once per generation in the pre-fill stage only
+        if (
+            (cache_position is not None and cache_position[0] == 0)
+            or self.rope_deltas is None
+            or (past_key_values is None or past_key_values.get_seq_length() == 0)  # type: ignore
+        ):
+            position_ids, rope_deltas = self.get_rope_index(
+                input_ids,
+                image_grid_thw,
+                video_grid_thw,
+                second_per_grid_ts,
+                attention_mask,
+            )
+            self.rope_deltas = rope_deltas
+        # then use the prev pre-calculated rope-deltas to get the correct position ids
+        else:
+            batch_size, seq_length, _ = inputs_embeds.shape
+            delta = (
+                (cache_position[0] + self.rope_deltas).to(inputs_embeds.device)
+                if cache_position is not None
+                else 0
+            )
+            position_ids = torch.arange(seq_length, device=inputs_embeds.device)  # type: ignore
+            position_ids = position_ids.view(1, -1).expand(batch_size, -1)  # type: ignore
+            if cache_position is not None:  # otherwise `deltas` is an int `0`
+                delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=0)  # type: ignore
+            position_ids = position_ids.add(delta)  # type: ignore
+            position_ids = position_ids.unsqueeze(0).expand(3, -1, -1)  # type: ignore
+
+    outputs = self.model(
+        input_ids=None,
+        position_ids=position_ids,
+        attention_mask=attention_mask,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+    )
+
+    hidden_states = outputs[0]
+    logits = None
+    loss = None
+
+    if _PATCH_OPTS is not None and _PATCH_OPTS.use_lce(labels, self.training):
+        assert labels is not None
+        loss = apply_lce(
+            hidden_states,
+            self.lm_head.weight,
+            labels,
+            _PATCH_OPTS,
+        )
+    else:
+        logits = self.lm_head(hidden_states)
+
+        if labels is not None:
+            # Upcast to float if we need to compute the loss to avoid potential precision issues
+            logits = logits.float()
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return Qwen2_5_VLCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        rope_deltas=self.rope_deltas,
+    )
+
+
+def patch_qwen2_5_vl(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    global _PATCH_OPTS  # pylint: disable=global-statement
+
+    from transformers.models.qwen2_5_vl import modeling_qwen2_5_vl
+
+    _PATCH_OPTS = patch_options
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_qwen2_5_vl.Qwen2_5_VLForConditionalGeneration
+        ), f"Expected a Qwen2_5_VLForConditionalGeneration model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward_multimodal, maybe_model)
+
+        return maybe_model
+
+    modeling_qwen2_5_vl.Qwen2_5_VLForConditionalGeneration.forward = (
+        cce_forward_multimodal
+    )
+    return None

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

@@ -0,0 +1,178 @@
+"""Qwen2 MoE CCE patch. Adapted from transformers v4.51.2"""
+
+# pylint: disable=duplicate-code
+
+from types import MethodType
+from typing import Optional, Union
+
+import torch
+import transformers
+from cut_cross_entropy.transformers.utils import (
+    PatchOptions,
+    TransformersModelT,
+    apply_lce,
+)
+from transformers.models.qwen2_moe.modeling_qwen2_moe import (
+    MoeCausalLMOutputWithPast,
+    MoeModelOutputWithPast,
+    load_balancing_loss_func,
+)
+from transformers.utils.deprecation import deprecate_kwarg
+from transformers.utils.generic import can_return_tuple
+
+_PATCH_OPTS: PatchOptions | None = None
+
+
+@can_return_tuple
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+def forward(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[list[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    output_router_logits: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **loss_kwargs,
+) -> MoeCausalLMOutputWithPast:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, Qwen2MoeForCausalLM
+
+    >>> model = Qwen2MoeForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
+    >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
+
+    >>> prompt = "Hey, are you conscious? Can you talk to me?"
+    >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+    ```"""
+
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_router_logits = (
+        output_router_logits
+        if output_router_logits is not None
+        else self.config.output_router_logits
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs: MoeModelOutputWithPast = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        output_router_logits=output_router_logits,
+        cache_position=cache_position,
+    )
+
+    hidden_states = outputs.last_hidden_state
+    loss = None
+    logits = None
+
+    if hidden_states is None:
+        raise ValueError("hidden_states is None")
+
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    slice_indices = (
+        slice(-logits_to_keep, None)
+        if isinstance(logits_to_keep, int)
+        else logits_to_keep
+    )
+
+    if _PATCH_OPTS is not None and _PATCH_OPTS.use_lce(labels, self.training):
+        assert labels is not None
+        loss = apply_lce(
+            hidden_states[:, slice_indices, :],
+            self.lm_head.weight,
+            labels,
+            _PATCH_OPTS,
+            **loss_kwargs,
+        )
+    else:
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        if labels is not None:
+            loss = self.loss_function(logits, labels, self.vocab_size, **loss_kwargs)
+
+    aux_loss = None
+    if output_router_logits:
+        aux_loss = load_balancing_loss_func(
+            outputs.router_logits,
+            self.num_experts,
+            self.num_experts_per_tok,
+            attention_mask,
+        )
+        if labels is not None:
+            loss += self.router_aux_loss_coef * aux_loss.to(  # type: ignore
+                loss.device  # type: ignore
+            )  # make sure to reside in the same device
+
+    return MoeCausalLMOutputWithPast(
+        loss=loss,
+        aux_loss=aux_loss,  # type: ignore
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        router_logits=outputs.router_logits,
+    )
+
+
+def patch_qwen2_moe(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    global _PATCH_OPTS  # pylint: disable=global-statement
+
+    from transformers.models.qwen2_moe import modeling_qwen2_moe
+
+    _PATCH_OPTS = patch_options
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_qwen2_moe.Qwen2MoeForCausalLM
+        ), f"Expected a Qwen3MoeForCausalLM model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(forward, maybe_model)
+
+        return maybe_model
+
+    modeling_qwen2_moe.Qwen2MoeForCausalLM.forward = forward
+    return None

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

@@ -0,0 +1,239 @@
+"""Qwen2 VL CCE patch. Adapted from transformers v4.51.2"""
+
+# pylint: disable=duplicate-code
+
+from types import MethodType
+from typing import Optional, Tuple, Union
+
+import torch
+import transformers
+from cut_cross_entropy.transformers.utils import (
+    PatchOptions,
+    TransformersModelT,
+    apply_lce,
+)
+from torch.nn import CrossEntropyLoss
+from transformers.models.qwen2_vl.modeling_qwen2_vl import (
+    Qwen2VLCausalLMOutputWithPast,
+)
+
+_PATCH_OPTS: PatchOptions | None = None
+
+
+def cce_forward_multimodal(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[list[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    pixel_values: Optional[torch.Tensor] = None,
+    pixel_values_videos: Optional[torch.FloatTensor] = None,
+    image_grid_thw: Optional[torch.LongTensor] = None,
+    video_grid_thw: Optional[torch.LongTensor] = None,
+    rope_deltas: Optional[torch.LongTensor] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+) -> Union[Tuple, Qwen2VLCausalLMOutputWithPast]:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from PIL import Image
+    >>> import requests
+    >>> from transformers import AutoProcessor, Qwen2VLForConditionalGeneration
+
+    >>> model = Qwen2VLForConditionalGeneration.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")
+    >>> processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")
+
+    >>> messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "image"},
+                {"type": "text", "text": "What is shown in this image?"},
+            ],
+        },
+    ]
+    >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
+    >>> image = Image.open(requests.get(url, stream=True).raw)
+
+    >>> text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    >>> inputs = processor(text=[text], images=[image], vision_infos=[vision_infos])
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "The image shows a street scene with a red stop sign in the foreground. In the background, there is a large red gate with Chinese characters ..."
+    ```"""
+
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+
+    if inputs_embeds is None:
+        inputs_embeds = self.model.embed_tokens(input_ids)
+        if pixel_values is not None:
+            pixel_values = pixel_values.type(self.visual.get_dtype())
+            image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
+            n_image_tokens = (input_ids == self.config.image_token_id).sum().item()
+            n_image_features = image_embeds.shape[0]
+            if n_image_tokens != n_image_features:
+                raise ValueError(
+                    f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
+                )
+            image_mask = (
+                (input_ids == self.config.image_token_id)
+                .unsqueeze(-1)
+                .expand_as(inputs_embeds)
+                .to(inputs_embeds.device)
+            )
+            image_embeds = image_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)  # type: ignore
+
+        if pixel_values_videos is not None:
+            pixel_values_videos = pixel_values_videos.type(self.visual.get_dtype())
+            video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw)
+            n_video_tokens = (input_ids == self.config.video_token_id).sum().item()
+            n_video_features = video_embeds.shape[0]
+            if n_video_tokens != n_video_features:
+                raise ValueError(
+                    f"Video features and video tokens do not match: tokens: {n_video_tokens}, features {n_video_features}"
+                )
+            video_mask = (
+                (input_ids == self.config.video_token_id)
+                .unsqueeze(-1)
+                .expand_as(inputs_embeds)
+                .to(inputs_embeds.device)
+            )
+            video_embeds = video_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)  # type: ignore
+
+        if attention_mask is not None:
+            attention_mask = attention_mask.to(inputs_embeds.device)
+
+    # if we get 4D attention mask we cannot calculate rope deltas anymore. TODO @raushan fixme
+    if position_ids is None and (attention_mask is None or attention_mask.ndim == 2):
+        # calculate RoPE index once per generation in the pre-fill stage only
+        if (
+            (cache_position is not None and cache_position[0] == 0)
+            or self.rope_deltas is None
+            or (past_key_values is None or past_key_values.get_seq_length() == 0)  # type: ignore
+        ):
+            position_ids, rope_deltas = self.get_rope_index(
+                input_ids, image_grid_thw, video_grid_thw, attention_mask
+            )
+            self.rope_deltas = rope_deltas
+        # then use the prev pre-calculated rope-deltas to get the correct position ids
+        else:
+            batch_size, seq_length, _ = inputs_embeds.shape
+            delta = (
+                cache_position[0] + self.rope_deltas
+                if cache_position is not None
+                else 0
+            )
+            position_ids = torch.arange(seq_length, device=inputs_embeds.device)  # type: ignore
+            position_ids = position_ids.view(1, -1).expand(batch_size, -1)  # type: ignore
+            if cache_position is not None:  # otherwise `deltas` is an int `0`
+                delta = delta.repeat_interleave(batch_size // delta.shape[0], dim=0)  # type: ignore
+                delta = delta.to(position_ids.device)  # type: ignore
+            position_ids = position_ids.add(delta)  # type: ignore
+            position_ids = position_ids.unsqueeze(0).expand(3, -1, -1)  # type: ignore
+
+    outputs = self.model(
+        input_ids=None,
+        position_ids=position_ids,
+        attention_mask=attention_mask,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+    )
+
+    hidden_states = outputs[0]
+    logits = None
+    loss = None
+
+    if _PATCH_OPTS is not None and _PATCH_OPTS.use_lce(labels, self.training):
+        assert labels is not None
+        loss = apply_lce(
+            hidden_states,
+            self.lm_head.weight,
+            labels,
+            _PATCH_OPTS,
+        )
+    else:
+        logits = self.lm_head(hidden_states)
+
+        if labels is not None:
+            # Upcast to float if we need to compute the loss to avoid potential precision issues
+            logits = logits.float()
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return Qwen2VLCausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        rope_deltas=self.rope_deltas,
+    )
+
+
+def patch_qwen2_vl(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    global _PATCH_OPTS  # pylint: disable=global-statement
+
+    from transformers.models.qwen2_vl import modeling_qwen2_vl
+
+    _PATCH_OPTS = patch_options
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_qwen2_vl.Qwen2VLForConditionalGeneration
+        ), f"Expected a Qwen2VLForConditionalGeneration model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward_multimodal, maybe_model)
+
+        return maybe_model
+
+    modeling_qwen2_vl.Qwen2VLForConditionalGeneration.forward = cce_forward_multimodal
+    return None

src/axolotl/integrations/cut_cross_entropy/monkeypatch/qwen3.py

@@ -0,0 +1,35 @@
+"""Qwen3 CCE patch. The model inherits Llama's modeling code and uses the same forward method."""
+
+# pylint: disable=duplicate-code
+
+from types import MethodType
+
+import transformers
+from cut_cross_entropy.transformers.utils import (
+    PatchOptions,
+    TransformersModelT,
+)
+
+
+def patch_qwen3(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    from transformers.models.qwen3 import modeling_qwen3
+
+    # Set the _PATCH_OPTS in the llama patch file
+    import axolotl.integrations.cut_cross_entropy.monkeypatch.llama as llama_patch
+
+    llama_patch._PATCH_OPTS = patch_options  # pylint: disable=protected-access
+
+    from axolotl.integrations.cut_cross_entropy.monkeypatch.llama import cce_forward
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_qwen3.Qwen3ForCausalLM
+        ), f"Expected a Qwen3ForCausalLM model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(cce_forward, maybe_model)
+        return maybe_model
+
+    modeling_qwen3.Qwen3ForCausalLM.forward = cce_forward
+    return None

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

@@ -0,0 +1,183 @@
+"""Qwen3 MoE CCE patch. Adapted from transformers v4.51.2"""
+
+# pylint: disable=duplicate-code
+
+from types import MethodType
+from typing import Optional, Union
+
+import torch
+import transformers
+from cut_cross_entropy.transformers.utils import (
+    PatchOptions,
+    TransformersModelT,
+    apply_lce,
+)
+from transformers.models.qwen3_moe.modeling_qwen3_moe import (
+    KwargsForCausalLM,
+    MoeCausalLMOutputWithPast,
+    MoeModelOutputWithPast,
+    load_balancing_loss_func,
+)
+from transformers.processing_utils import Unpack
+from transformers.utils.deprecation import deprecate_kwarg
+from transformers.utils.generic import can_return_tuple
+
+_PATCH_OPTS: PatchOptions | None = None
+
+
+@can_return_tuple
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+def forward(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[list[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    output_router_logits: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **kwargs: Unpack[KwargsForCausalLM],
+) -> MoeCausalLMOutputWithPast:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from transformers import AutoTokenizer, Qwen3MoeForCausalLM
+
+    >>> model = Qwen3MoeForCausalLM.from_pretrained("Qwen/Qwen3-MoE-15B-A2B")
+    >>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-MoE-15B-A2B")
+
+    >>> prompt = "Hey, are you conscious? Can you talk to me?"
+    >>> inputs = tokenizer(prompt, return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+    >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
+    ```"""
+
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_router_logits = (
+        output_router_logits
+        if output_router_logits is not None
+        else self.config.output_router_logits
+    )
+
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs: MoeModelOutputWithPast = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        output_router_logits=output_router_logits,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs.last_hidden_state
+
+    if hidden_states is None:
+        raise ValueError("hidden_states is None")
+
+    loss = None
+    logits = None
+
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    slice_indices = (
+        slice(-logits_to_keep, None)
+        if isinstance(logits_to_keep, int)
+        else logits_to_keep
+    )
+
+    if _PATCH_OPTS is not None and _PATCH_OPTS.use_lce(labels, self.training):
+        assert labels is not None
+        loss = apply_lce(
+            hidden_states[:, slice_indices, :],
+            self.lm_head.weight,
+            labels,
+            _PATCH_OPTS,
+            **kwargs,
+        )
+    else:
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+
+        if labels is not None:
+            loss = self.loss_function(logits, labels, self.vocab_size, **kwargs)
+
+    aux_loss = None
+    if output_router_logits:
+        aux_loss = load_balancing_loss_func(
+            outputs.router_logits,
+            self.num_experts,
+            self.num_experts_per_tok,
+            attention_mask,
+        )
+        if labels is not None:
+            loss += self.router_aux_loss_coef * aux_loss.to(  # type: ignore
+                loss.device  # type: ignore
+            )  # make sure to reside in the same device
+
+    return MoeCausalLMOutputWithPast(
+        loss=loss,
+        aux_loss=aux_loss,  # type: ignore
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        router_logits=outputs.router_logits,
+    )
+
+
+def patch_qwen3_moe(
+    maybe_model: TransformersModelT | str | transformers.PretrainedConfig,
+    patch_options: PatchOptions,
+) -> TransformersModelT | None:
+    global _PATCH_OPTS  # pylint: disable=global-statement
+
+    from transformers.models.qwen3_moe import modeling_qwen3_moe
+
+    _PATCH_OPTS = patch_options
+
+    if isinstance(maybe_model, transformers.PreTrainedModel):
+        assert isinstance(
+            maybe_model, modeling_qwen3_moe.Qwen3MoeForCausalLM
+        ), f"Expected a Qwen3MoeForCausalLM model. Got {type(maybe_model)}."
+        maybe_model.forward = MethodType(forward, maybe_model)
+
+        return maybe_model
+
+    modeling_qwen3_moe.Qwen3MoeForCausalLM.forward = forward
+    return None

src/axolotl/integrations/kd/chat_template.py

@@ -35,6 +35,9 @@ class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
         sequence_len,
         roles_to_train=None,
         train_on_eos=None,
+        train_on_eot=None,
+        eot_tokens=None,
+        split_thinking: bool | None = False,
         logprobs_field="logprobs",
         gen_temperature=1.0,
         kd_temperature=1.0,
@@ -50,6 +53,9 @@ class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
             sequence_len,
             roles_to_train=roles_to_train,
             train_on_eos=train_on_eos,
+            train_on_eot=train_on_eot,
+            eot_tokens=eot_tokens,
+            split_thinking=split_thinking,
         )
 
     @property

src/axolotl/integrations/liger/__init__.py

@@ -23,8 +23,8 @@ import logging
 import sys
 
 from axolotl.integrations.base import BasePlugin
+from axolotl.utils.distributed import is_main_process
 
-from ...utils.distributed import zero_only
 from .args import LigerArgs  # pylint: disable=unused-import. # noqa: F401
 from .utils import patch_with_compile_disable
 
@@ -85,7 +85,7 @@ class LigerPlugin(BasePlugin):
                 kwargs["geglu"] = cfg.liger_glu_activation
             elif "swiglu" in liger_fn_sig.parameters:
                 kwargs["swiglu"] = cfg.liger_glu_activation
-            with zero_only():
+            if is_main_process(use_environ=True):
                 LOG.info(
                     f"Applying LIGER to {cfg.model_config_type} with kwargs: {kwargs}"
                 )
@@ -151,6 +151,30 @@ class LigerPlugin(BasePlugin):
                 rms_norm=cfg.liger_rms_norm,
                 layer_norm=cfg.liger_layer_norm,
             )
+        elif cfg.model_config_type == "qwen3":
+            from axolotl.integrations.liger.models.qwen3 import (
+                apply_liger_kernel_to_qwen3,
+            )
+
+            apply_liger_kernel_to_qwen3(
+                cross_entropy=cfg.liger_cross_entropy,
+                fused_linear_cross_entropy=cfg.liger_fused_linear_cross_entropy,
+                glu_activation=cfg.liger_glu_activation,
+                rms_norm=cfg.liger_rms_norm,
+                layer_norm=cfg.liger_layer_norm,
+            )
+        elif cfg.model_config_type == "qwen3_moe":
+            from axolotl.integrations.liger.models.qwen3_moe import (
+                apply_liger_kernel_to_qwen3_moe,
+            )
+
+            apply_liger_kernel_to_qwen3_moe(
+                cross_entropy=cfg.liger_cross_entropy,
+                fused_linear_cross_entropy=cfg.liger_fused_linear_cross_entropy,
+                glu_activation=cfg.liger_glu_activation,
+                rms_norm=cfg.liger_rms_norm,
+                layer_norm=cfg.liger_layer_norm,
+            )
         else:
             logging.warning(
                 f"Unsupported model config type: {cfg.model_config_type}. Liger not applied."