feat: merge adapter in fp32

* update setuptools so trl can be installed from main for nightlies

* run the nightly in the PR CI on change

* use range request, don't use cu129 in CI since it's not supported with AO

* run multigpu ci if CCE install script changes

.github/CONTRIBUTING.md

@@ -70,6 +70,11 @@ You can skip certain CI checks by including specific keywords in your commit mes
 
 axolotl uses [{codestyle}]({URLofCodestyle}) as its code style guide. Please ensure that your code follows these guidelines.
 
+Use the pre-commit linter to ensure that your code is formatted consistently.
+```bash
+pre-commit run --all-files
+```
+
 ### Commit Messages
 
 Write clear and concise commit messages that briefly describe the changes made in each commit. Use the imperative mood and start with a capitalized verb, e.g., "Add new feature" or "Fix bug in function".

.github/workflows/base.yml

@@ -51,6 +51,30 @@ jobs:
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
             dockerfile: "Dockerfile-base"
             platforms: "linux/amd64,linux/arm64"
+          - cuda: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.10.0
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-base"
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            python_version: "3.12"
+            pytorch: 2.10.0
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-base"
+            platforms: "linux/amd64,linux/arm64"
+#          - cuda: "129"
+#            cuda_version: 12.9.1
+#            cudnn_version: ""
+#            python_version: "3.12"
+#            pytorch: 2.9.1
+#            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+#            dockerfile: "Dockerfile-base"
+#            platforms: "linux/amd64,linux/arm64"
           - cuda: "130"
             cuda_version: 13.0.0
             cudnn_version: ""
@@ -59,6 +83,22 @@ jobs:
             torch_cuda_arch_list: "9.0+PTX"
             dockerfile: "Dockerfile-base"
             platforms: "linux/amd64,linux/arm64"
+          - cuda: "130"
+            cuda_version: 13.0.0
+            cudnn_version: ""
+            python_version: "3.12"
+            pytorch: 2.9.1
+            torch_cuda_arch_list: "9.0+PTX"
+            dockerfile: "Dockerfile-base"
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: "130"
+            cuda_version: 13.0.0
+            cudnn_version: ""
+            python_version: "3.12"
+            pytorch: 2.10.0
+            torch_cuda_arch_list: "9.0+PTX"
+            dockerfile: "Dockerfile-base"
+            platforms: "linux/amd64,linux/arm64"
 #          - cuda: "128"
 #            cuda_version: 12.8.1
 #            cudnn_version: ""
@@ -141,6 +181,30 @@ jobs:
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
             dockerfile: "Dockerfile-uv-base"
             platforms: "linux/amd64,linux/arm64"
+          - cuda: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.10.0
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-uv-base"
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            python_version: "3.12"
+            pytorch: 2.10.0
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-uv-base"
+            platforms: "linux/amd64,linux/arm64"
+#          - cuda: "129"
+#            cuda_version: 12.9.1
+#            cudnn_version: ""
+#            python_version: "3.12"
+#            pytorch: 2.9.1
+#            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+#            dockerfile: "Dockerfile-uv-base"
+#            platforms: "linux/amd64,linux/arm64"
           - cuda: "130"
             cuda_version: 13.0.0
             cudnn_version: ""
@@ -149,6 +213,22 @@ jobs:
             torch_cuda_arch_list: "9.0+PTX"
             dockerfile: "Dockerfile-uv-base"
             platforms: "linux/amd64,linux/arm64"
+          - cuda: "130"
+            cuda_version: 13.0.0
+            cudnn_version: ""
+            python_version: "3.12"
+            pytorch: 2.9.1
+            torch_cuda_arch_list: "9.0+PTX"
+            dockerfile: "Dockerfile-uv-base"
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: "130"
+            cuda_version: 13.0.0
+            cudnn_version: ""
+            python_version: "3.12"
+            pytorch: 2.10.0
+            torch_cuda_arch_list: "9.0+PTX"
+            dockerfile: "Dockerfile-uv-base"
+            platforms: "linux/amd64,linux/arm64"
     steps:
       - name: Checkout
         uses: actions/checkout@v4

.github/workflows/main.yml

@@ -34,12 +34,30 @@ jobs:
             axolotl_extras:
             platforms: "linux/amd64,linux/arm64"
             is_latest: true
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.12"
+            pytorch: 2.10.0
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+#          - cuda: 129
+#            cuda_version: 12.9.1
+#            python_version: "3.12"
+#            pytorch: 2.9.1
+#            axolotl_extras:
+#            platforms: "linux/amd64,linux/arm64"
           - cuda: 130
             cuda_version: 13.0.0
             python_version: "3.11"
             pytorch: 2.9.1
             axolotl_extras:
             platforms: "linux/amd64,linux/arm64"
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.12"
+            pytorch: 2.10.0
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -80,6 +98,77 @@ jobs:
             ${{ (matrix.is_latest) && format('{0}-latest', steps.metadata.outputs.tags) || '' }}
           labels: ${{ steps.metadata.outputs.labels }}
 
+  build-axolotl-uv:
+    if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.9.1
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+            is_latest: true
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.12"
+            pytorch: 2.10.0
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.11"
+            pytorch: 2.9.1
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.12"
+            pytorch: 2.10.0
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+    runs-on: axolotl-gpu-runner
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Docker metadata
+        id: metadata
+        uses: docker/metadata-action@v5
+        with:
+          images: |
+            axolotlai/axolotl-uv
+          tags: |
+            type=ref,event=branch
+            type=pep440,pattern={{version}}
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_TOKEN }}
+      # guidance for testing before pushing: https://docs.docker.com/build/ci/github-actions/test-before-push/
+      - name: Build and export to Docker
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          platforms: ${{ matrix.platforms }}
+          build-args: |
+            BASE_TAG=${{ github.ref_type == 'tag' && 'main' || github.ref_name }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}
+            CUDA=${{ matrix.cuda }}
+            PYTORCH_VERSION=${{ matrix.pytorch }}
+            AXOLOTL_ARGS=${{ matrix.axolotl_args }}
+            AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}
+          file: ./docker/Dockerfile-uv
+          push: ${{ github.event_name != 'pull_request' }}
+          tags: |
+            ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
+            ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}
+            ${{ (matrix.is_latest) && format('{0}-latest', steps.metadata.outputs.tags) || '' }}
+          labels: ${{ steps.metadata.outputs.labels }}
+
   build-axolotl-cloud:
     needs: build-axolotl
     if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
@@ -106,12 +195,30 @@ jobs:
             axolotl_extras:
             is_latest: true
             platforms: "linux/amd64,linux/arm64"
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.12"
+            pytorch: 2.10.0
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+#          - cuda: 129
+#            cuda_version: 12.9.1
+#            python_version: "3.12"
+#            pytorch: 2.9.1
+#            axolotl_extras:
+#            platforms: "linux/amd64,linux/arm64"
           - cuda: 130
             cuda_version: 13.0.0
             python_version: "3.11"
             pytorch: 2.9.1
             axolotl_extras:
             platforms: "linux/amd64,linux/arm64"
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.12"
+            pytorch: 2.10.0
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -147,6 +254,73 @@ jobs:
              ${{ (matrix.is_latest) && format('{0}-latest', steps.metadata.outputs.tags) || '' }}
           labels: ${{ steps.metadata.outputs.labels }}
 
+  build-axolotl-cloud-uv:
+    needs: build-axolotl-uv
+    if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
+    # this job needs to be run on self-hosted GPU runners...
+    strategy:
+      matrix:
+        include:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.9.1
+            axolotl_extras:
+            is_latest: true
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.12"
+            pytorch: 2.10.0
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.11"
+            pytorch: 2.9.1
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.12"
+            pytorch: 2.10.0
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+    runs-on: axolotl-gpu-runner
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Docker metadata
+        id: metadata
+        uses: docker/metadata-action@v5
+        with:
+          images: |
+            axolotlai/axolotl-cloud-uv
+          tags: |
+            type=ref,event=branch
+            type=pep440,pattern={{version}}
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_TOKEN }}
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      - name: Build
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          platforms: ${{ matrix.platforms }}
+          build-args: |
+            BASE_TAG=${{ github.ref_type == 'tag' && 'main' || github.ref_name }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
+            CUDA=${{ matrix.cuda }}
+          file: ./docker/Dockerfile-cloud-uv
+          push: ${{ github.event_name != 'pull_request' }}
+          tags: |
+             ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
+             ${{ (matrix.is_latest) && format('{0}-latest', steps.metadata.outputs.tags) || '' }}
+          labels: ${{ steps.metadata.outputs.labels }}
+
   build-axolotl-cloud-no-tmux:
     needs: build-axolotl
     if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}

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

@@ -8,6 +8,7 @@ on:
       - 'setup.py'
       - 'pyproject.toml'
       - '.github/workflows/multi-gpu-e2e.yml'
+      - 'scripts/cutcrossentropy_install.py'
       - 'src/axolotl/core/trainers/mixins/sequence_parallel.py'
       - 'src/axolotl/utils/distributed.py'
   workflow_dispatch:
@@ -35,14 +36,13 @@ jobs:
             pytorch: 2.8.0
             axolotl_extras: fbgemm-gpu
             num_gpus: 2
-            nightly_build: "true"
-          - cuda: 128
-            cuda_version: 12.8.1
-            python_version: "3.11"
-            pytorch: 2.9.1
-            axolotl_extras: fbgemm-gpu
-            num_gpus: 2
-            nightly_build: "true"
+#          - cuda: 129
+#            cuda_version: 12.9.1
+#            python_version: "3.12"
+#            pytorch: 2.9.1
+#            axolotl_extras: "fbgemm-gpu"
+#            num_gpus: 2
+#            dockerfile: "Dockerfile-uv.jinja"
           - cuda: 130
             cuda_version: 13.0.0
             python_version: "3.11"
@@ -50,7 +50,13 @@ jobs:
             axolotl_extras:
 #            axolotl_extras: fbgemm-gpu
             num_gpus: 2
-            nightly_build: "true"
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.10.0
+            axolotl_extras: "fbgemm-gpu"
+            num_gpus: 2
+            dockerfile: "Dockerfile-uv.jinja"
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
     steps:
@@ -72,8 +78,8 @@ jobs:
           echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
-          echo "NIGHTLY_BUILD=${{ matrix.nightly_build }}" >> $GITHUB_ENV
           echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
+          echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         run: |
           modal run -m cicd.multigpu

.github/workflows/preview-docs.yml

@@ -14,14 +14,8 @@ on:
       - .github/workflows/preview-docs.yml
 
 permissions:
-  checks: write
-  contents: write
-  deployments: write
-  issues: write
-  discussions: write
-  pages: write
+  contents: read
   pull-requests: write
-  statuses: write
 
 jobs:
   preview:

.github/workflows/pypi.yml

@@ -40,7 +40,7 @@ jobs:
 
       - name: Install dependencies
         run: |
-          pip3 install wheel packaging==23.2
+          pip3 install wheel packaging==26.0
           pip3 install --no-build-isolation -e .
           pip3 install -r requirements-dev.txt -r requirements-tests.txt
 
@@ -48,9 +48,9 @@ jobs:
         id: tag
         run: echo ::set-output name=TAG_NAME::$(echo $GITHUB_REF | cut -d / -f 3)
 
-      - name: Update version in setup.py
+      - name: Update version in VERSION file
         run: |
-          sed -i -E 's/version="([0-9.]+)",/version="${{ steps.tag.outputs.TAG_NAME }}",/g' setup.py
+          echo "${{ steps.tag.outputs.TAG_NAME }}" | sed 's/^v//' > VERSION
 
       - name: Build a source dist
         run: |

.github/workflows/tests-nightly.yml

@@ -3,6 +3,10 @@ on:
   workflow_dispatch:
   schedule:
     - cron: '0 0 * * *'  # Runs at 00:00 UTC every day
+  pull_request:
+    types: [opened, synchronize, reopened, ready_for_review]
+    paths:
+      - '.github/workflows/tests-nightly.yml'
 
 jobs:
   pre-commit:
@@ -18,15 +22,26 @@ jobs:
         env:
           SKIP: no-commit-to-branch
 
+  prime-cdn-s3-cache:
+    name: Prefetch S3 once to prime the CDN cache
+    runs-on: ubuntu-latest
+    if: ${{ !github.event.pull_request.draft }}
+    timeout-minutes: 10
+    steps:
+      - name: Restore Cache from S3
+        id: hf-cache-restore-s3
+        run: |
+          curl -v -H "Range: bytes=0-1023" -L https://axolotl-ci.b-cdn.net/hf-cache.tar.zst > /dev/null
+
   pytest:
     name: PyTest
     runs-on: ubuntu-latest
+    needs: [prime-cdn-s3-cache]
     strategy:
       fail-fast: false
-      max-parallel: 2
       matrix:
-        python_version: ["3.11"]
-        pytorch_version: ["2.8.0", "2.9.0", "2.9.1"]
+        python_version: ["3.12"]  # TODO include py3.14 once https://github.com/mistralai/mistral-common/pull/194 is merged
+        pytorch_version: ["2.8.0", "2.9.1", "2.10.0"]
     timeout-minutes: 20
 
     steps:
@@ -37,7 +52,7 @@ jobs:
         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
+          curl -L https://axolotl-ci.b-cdn.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
@@ -48,7 +63,7 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
+          pip3 install --upgrade packaging==26.0 setuptools==78.1.1 wheel
 
       - name: Install PyTorch
         run: |
@@ -102,16 +117,23 @@ jobs:
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.8.0
+            pytorch: 2.9.1
             num_gpus: 1
             axolotl_extras:
             nightly_build: "true"
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
+            pytorch: 2.10.0
+            num_gpus: 1
+            axolotl_extras:
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.12"
             pytorch: 2.9.1
             num_gpus: 1
             axolotl_extras:
+            dockerfile: "Dockerfile-uv.jinja"
             nightly_build: "true"
     steps:
       - name: Checkout
@@ -132,6 +154,7 @@ jobs:
           echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
+          echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
           echo "NIGHTLY_BUILD=${{ matrix.nightly_build }}" >> $GITHUB_ENV
           echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
       - name: Run tests job on Modal

.github/workflows/tests.yml

@@ -46,16 +46,32 @@ jobs:
         env:
           SKIP: no-commit-to-branch
 
+  prime-cdn-s3-cache:
+    name: Prefetch S3 once to prime the CDN cache
+    runs-on: ubuntu-latest
+    if: ${{ !github.event.pull_request.draft }}
+    timeout-minutes: 10
+    steps:
+      - name: Restore Cache from S3
+        id: hf-cache-restore-s3
+        run: |
+          curl -v -H "Range: bytes=0-1023" -L https://axolotl-ci.b-cdn.net/hf-cache.tar.zst > /dev/null
+
   pytest:
     name: PyTest
     runs-on: ubuntu-latest
     if: ${{ !github.event.pull_request.draft }}
-#    needs: [preload-cache]
+    needs: [prime-cdn-s3-cache]
     strategy:
       fail-fast: false
       matrix:
-        python_version: ["3.11"]
-        pytorch_version: ["2.8.0", "2.9.0", "2.9.1"]
+        python_version: ["3.12"]  # TODO include py3.14 once https://github.com/mistralai/mistral-common/pull/194 is merged
+        pytorch_version: ["2.8.0", "2.9.1", "2.10.0"]
+#        exclude:
+#          - python_version: "3.14"
+#            pytorch_version: "2.8.0"
+#          - python_version: "3.14"
+#            pytorch_version: "2.9.1"
     timeout-minutes: 20
 
     steps:
@@ -70,7 +86,7 @@ jobs:
         id: hf-cache-restore-s3
         run: |
           mkdir -p ~/.cache/huggingface/hub
-          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xpf - -C ~/.cache/huggingface/hub/  --use-compress-program unzstd --strip-components=1
+          curl -L https://axolotl-ci.b-cdn.net/hf-cache.tar.zst | tar -xpf - -C ~/.cache/huggingface/hub/  --use-compress-program unzstd --strip-components=1
           ls -ltr ~/.cache/huggingface/hub/
 
       - name: Setup Python
@@ -82,7 +98,7 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
+          pip3 install --upgrade packaging==26.0 setuptools==75.8.0 wheel
 
       - name: Install PyTorch
         run: |
@@ -110,10 +126,10 @@ jobs:
 
       - name: Pre-Download dataset fixture
         run: |
-          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
+          hf download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
 
       - name: Show HF cache
-        run: hf cache scan
+        run: hf cache ls
 
       - name: Run tests
         run: |
@@ -127,7 +143,7 @@ jobs:
           pytest -v --durations=10 tests/cli/ --cov=axolotl --cov-append --cov-report=xml
 
       - name: Show HF cache
-        run: hf cache scan
+        run: hf cache ls
 
       - name: Upload coverage to Codecov
         uses: codecov/codecov-action@v5
@@ -141,12 +157,18 @@ jobs:
     name: PyTest from Source Dist
     runs-on: ubuntu-latest
     if: ${{ !github.event.pull_request.draft }}
+    needs: [prime-cdn-s3-cache]
     strategy:
       fail-fast: false
       matrix:
-        python_version: ["3.11"]
-        pytorch_version: ["2.8.0", "2.9.0", "2.9.1"]
-    timeout-minutes: 20
+        python_version: ["3.12"]  # TODO include py3.14 once https://github.com/mistralai/mistral-common/pull/194 is merged
+        pytorch_version: ["2.8.0", "2.9.1", "2.10.0"]
+#        exclude:
+#          - python_version: "3.14"
+#            pytorch_version: "2.8.0"
+#          - python_version: "3.14"
+#            pytorch_version: "2.9.1"
+    timeout-minutes: 30
 
     steps:
       - name: cleanup node
@@ -160,7 +182,7 @@ jobs:
         id: hf-cache-restore-s3
         run: |
           mkdir -p ~/.cache/huggingface/hub
-          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xpf - -C ~/.cache/huggingface/hub/  --use-compress-program unzstd --strip-components=1
+          curl -L https://axolotl-ci.b-cdn.net/hf-cache.tar.zst | tar -xpf - -C ~/.cache/huggingface/hub/  --use-compress-program unzstd --strip-components=1
           ls -ltr ~/.cache/huggingface/hub/
 
       - name: Setup Python
@@ -172,7 +194,7 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 setuptools_scm build wheel psutil
+          pip3 install --upgrade packaging==26.0 setuptools==75.8.0 setuptools_scm build wheel psutil
 
       - name: Install PyTorch
         run: |
@@ -200,7 +222,7 @@ jobs:
           axolotl --help
 
       - name: Show HF cache
-        run: hf cache scan
+        run: hf cache ls
 
       - name: Run tests
         run: |
@@ -209,10 +231,10 @@ jobs:
           pytest -v --durations=10 tests/cli/
 
       - name: Show HF cache
-        run: hf cache scan
+        run: hf cache ls
 
   gate-skip-e2e:
-    needs: [pre-commit, pytest, pytest-sdist]
+    needs: [pre-commit]
     runs-on: ubuntu-latest
     outputs:
       skip: ${{ steps.compute.outputs.skip }}
@@ -248,16 +270,16 @@ jobs:
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
-    needs: [pre-commit, pytest, pytest-sdist, gate-skip-e2e]
+    needs: [pre-commit, pytest]
 
     strategy:
       fail-fast: false
       matrix:
         include:
-          - cuda: 128
-            cuda_version: 12.8.1
-            python_version: "3.11"
-            pytorch: 2.8.0
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.12"
+            pytorch: 2.9.1
             num_gpus: 1
             axolotl_extras:
             dockerfile: "Dockerfile-uv.jinja"
@@ -316,6 +338,12 @@ jobs:
             pytorch: 2.9.1
             num_gpus: 1
             axolotl_extras:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.10.0
+            num_gpus: 1
+            axolotl_extras:
           - cuda: 130
             cuda_version: 13.0.0
             python_version: "3.11"

.pre-commit-config.yaml

@@ -11,7 +11,7 @@ repos:
     -   id: no-commit-to-branch
         args: ['--branch', 'main']
 -   repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.14.10
+    rev: v0.15.4
     hooks:
     -   id: ruff
         args: [--fix]
@@ -26,7 +26,7 @@ repos:
             'pydantic>=2.5.3',
         ]
 -   repo: https://github.com/PyCQA/bandit
-    rev: 1.9.2
+    rev: 1.9.4
     hooks:
     -   id: bandit
         args: [

.runpod/README.md

@@ -123,7 +123,7 @@ datasets:
 | --------------------------------- | -------------------------- | ----------------------------------- |
 | `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_num_proc`                | `4`                        | Number of preprocessing processes   |
 | `dataset_keep_in_memory`          | `false`                    | Keep dataset in memory              |
 | `shuffle_merged_datasets`         | `true`                     | Shuffle merged datasets             |
 | `shuffle_before_merging_datasets` | `false`                    | Shuffle each dataset before merging |

.runpod/src/config/config.yaml

@@ -39,7 +39,6 @@
 #     type: # linear | dynamic
 #     factor: # float
 
-
 # # Whether you are training a 4-bit GPTQ quantized model
 # gptq: true
 # gptq_groupsize: 128 # group size
@@ -107,7 +106,7 @@
 # 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
+# dataset_num_proc: # 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
@@ -224,9 +223,6 @@
 # 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.
@@ -352,8 +348,6 @@
 # # 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}
@@ -412,7 +406,7 @@ 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_num_proc: ${DATASET_NUM_PROC}
 dataset_keep_in_memory: ${DATASET_KEEP_IN_MEMORY}
 hub_model_id: ${HUB_MODEL_ID}
 hub_strategy: ${HUB_STRATEGY}
@@ -512,7 +506,6 @@ 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}

README.md

@@ -29,8 +29,23 @@
 
 ## 🎉 Latest Updates
 
-- 2025/12: Axolotl now includes support for [Kimi-Linear](https://docs.axolotl.ai/docs/models/kimi-linear.html), [Plano-Orchestrator](https://docs.axolotl.ai/docs/models/plano.html), [MiMo](https://docs.axolotl.ai/docs/models/mimo.html), [InternVL 3.5](https://docs.axolotl.ai/docs/models/internvl3_5.html), [Olmo3](https://docs.axolotl.ai/docs/models/olmo3.html), [Trinity](https://docs.axolotl.ai/docs/models/trinity.html), and [Ministral3](https://docs.axolotl.ai/docs/models/ministral3.html).
+- 2026/03:
+  - New model support has been added in Axolotl for [Qwen3.5, Qwen3.5 MoE](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/qwen3.5), [GLM-4.7-Flash](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/glm47-flash), [GLM-4.6V](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/glm46v), and [GLM-4.5-Air](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/glm45).
+  - [MoE expert quantization](https://docs.axolotl.ai/docs/expert_quantization.html) support (via `quantize_moe_experts: true`) greatly reduces VRAM when training MoE models (FSDP2 compat).
+- 2026/02:
+  - [ScatterMoE LoRA](https://github.com/axolotl-ai-cloud/axolotl/pull/3410) support. LoRA fine-tuning directly on MoE expert weights using custom Triton kernels.
+  - Axolotl now has support for [SageAttention](https://github.com/axolotl-ai-cloud/axolotl/pull/2823) and [GDPO](https://github.com/axolotl-ai-cloud/axolotl/pull/3353) (Generalized DPO).
+- 2026/01:
+  - New integration for [EAFT](https://github.com/axolotl-ai-cloud/axolotl/pull/3366) (Entropy-Aware Focal Training), weights loss by entropy of the top-k logit distribution, and [Scalable Softmax](https://github.com/axolotl-ai-cloud/axolotl/pull/3338), improves long context in attention.
+- 2025/12:
+  - Axolotl now includes support for [Kimi-Linear](https://docs.axolotl.ai/docs/models/kimi-linear.html), [Plano-Orchestrator](https://docs.axolotl.ai/docs/models/plano.html), [MiMo](https://docs.axolotl.ai/docs/models/mimo.html), [InternVL 3.5](https://docs.axolotl.ai/docs/models/internvl3_5.html), [Olmo3](https://docs.axolotl.ai/docs/models/olmo3.html), [Trinity](https://docs.axolotl.ai/docs/models/trinity.html), and [Ministral3](https://docs.axolotl.ai/docs/models/ministral3.html).
+  - [Distributed Muon Optimizer](https://github.com/axolotl-ai-cloud/axolotl/pull/3264) support has been added for FSDP2 pretraining.
 - 2025/10: New model support has been added in Axolotl for: [Qwen3 Next](https://docs.axolotl.ai/docs/models/qwen3-next.html), [Qwen2.5-vl, Qwen3-vl](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/qwen2_5-vl), [Qwen3, Qwen3MoE](https://docs.axolotl.ai/docs/models/qwen3.html), [Granite 4](https://docs.axolotl.ai/docs/models/granite4.html), [HunYuan](https://docs.axolotl.ai/docs/models/hunyuan.html), [Magistral 2509](https://docs.axolotl.ai/docs/models/magistral/vision.html), [Apertus](https://docs.axolotl.ai/docs/models/apertus.html), and [Seed-OSS](https://docs.axolotl.ai/docs/models/seed-oss.html).
+
+<details>
+
+<summary>Expand older updates</summary>
+
 - 2025/09: Axolotl now has text diffusion training. Read more [here](https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/integrations/diffusion).
 - 2025/08: QAT has been updated to include NVFP4 support. See [PR](https://github.com/axolotl-ai-cloud/axolotl/pull/3107).
 - 2025/07:
@@ -39,15 +54,10 @@
   - FP8 finetuning with fp8 gather op is now possible in Axolotl via `torchao`. Get started [here](https://docs.axolotl.ai/docs/mixed_precision.html#sec-fp8)!
   - [Voxtral](https://docs.axolotl.ai/docs/models/voxtral.html), [Magistral 1.1](https://docs.axolotl.ai/docs/models/magistral.html), and [Devstral](https://docs.axolotl.ai/docs/models/devstral.html) with mistral-common tokenizer support has been integrated in Axolotl!
   - TiledMLP support for single-GPU to multi-GPU training with DDP, DeepSpeed and FSDP support has been added to support Arctic Long Sequence Training. (ALST). See [examples](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/alst) for using ALST with Axolotl!
-- 2025/05: Quantization Aware Training (QAT) support has been added to Axolotl. Explore the [docs](https://docs.axolotl.ai/docs/qat.html) to learn more!
-
-<details>
-
-<summary>Expand older updates</summary>
-
-- 2025/03: Axolotl has implemented Sequence Parallelism (SP) support. Read the [blog](https://huggingface.co/blog/axolotl-ai-co/long-context-with-sequence-parallelism-in-axolotl) and [docs](https://docs.axolotl.ai/docs/sequence_parallelism.html) to learn how to scale your context length when fine-tuning.
 - 2025/06: Magistral with mistral-common tokenizer support has been added to Axolotl. See [docs](https://docs.axolotl.ai/docs/models/magistral.html) to start training your own Magistral models with Axolotl!
+- 2025/05: Quantization Aware Training (QAT) support has been added to Axolotl. Explore the [docs](https://docs.axolotl.ai/docs/qat.html) to learn more!
 - 2025/04: Llama 4 support has been added in Axolotl. See [docs](https://docs.axolotl.ai/docs/models/llama-4.html) to start training your own Llama 4 models with Axolotl's linearized version!
+- 2025/03: Axolotl has implemented Sequence Parallelism (SP) support. Read the [blog](https://huggingface.co/blog/axolotl-ai-co/long-context-with-sequence-parallelism-in-axolotl) and [docs](https://docs.axolotl.ai/docs/sequence_parallelism.html) to learn how to scale your context length when fine-tuning.
 - 2025/03: (Beta) Fine-tuning Multimodal models is now supported in Axolotl. Check out the [docs](https://docs.axolotl.ai/docs/multimodal.html) to fine-tune your own!
 - 2025/02: Axolotl has added LoRA optimizations to reduce memory usage and improve training speed for LoRA and QLoRA in single GPU and multi-GPU training (DDP and DeepSpeed). Jump into the [docs](https://docs.axolotl.ai/docs/lora_optims.html) to give it a try.
 - 2025/02: Axolotl has added GRPO support. Dive into our [blog](https://huggingface.co/blog/axolotl-ai-co/training-llms-w-interpreter-feedback-wasm) and [GRPO example](https://github.com/axolotl-ai-cloud/grpo_code) and have some fun!
@@ -62,10 +72,10 @@ Axolotl is a free and open-source tool designed to streamline post-training and
 Features:
 
 - **Multiple Model Support**: Train various models like GPT-OSS, LLaMA, Mistral, Mixtral, Pythia, and many more models available on the Hugging Face Hub.
-- **Multimodal Training**: Fine-tune vision-language models (VLMs) including LLaMA-Vision, Qwen2-VL, Pixtral, LLaVA, SmolVLM2, and audio models like Voxtral with image, video, and audio support.
-- **Training Methods**: Full fine-tuning, LoRA, QLoRA, GPTQ, QAT, Preference Tuning (DPO, IPO, KTO, ORPO), RL (GRPO), and Reward Modelling (RM) / Process Reward Modelling (PRM).
+- **Multimodal Training**: Fine-tune vision-language models (VLMs) including LLaMA-Vision, Qwen2-VL, Pixtral, LLaVA, SmolVLM2, GLM-4.6V, InternVL 3.5, Gemma 3n, and audio models like Voxtral with image, video, and audio support.
+- **Training Methods**: Full fine-tuning, LoRA, QLoRA, GPTQ, QAT, Preference Tuning (DPO, IPO, KTO, ORPO), RL (GRPO, GDPO), and Reward Modelling (RM) / Process Reward Modelling (PRM).
 - **Easy Configuration**: Re-use a single YAML configuration file across the full fine-tuning pipeline: dataset preprocessing, training, evaluation, quantization, and inference.
-- **Performance Optimizations**: [Multipacking](https://docs.axolotl.ai/docs/multipack.html), [Flash Attention](https://github.com/Dao-AILab/flash-attention), [Xformers](https://github.com/facebookresearch/xformers), [Flex Attention](https://pytorch.org/blog/flexattention/), [Liger Kernel](https://github.com/linkedin/Liger-Kernel), [Cut Cross Entropy](https://github.com/apple/ml-cross-entropy/tree/main), [Sequence Parallelism (SP)](https://docs.axolotl.ai/docs/sequence_parallelism.html), [LoRA optimizations](https://docs.axolotl.ai/docs/lora_optims.html), [Multi-GPU training (FSDP1, FSDP2, DeepSpeed)](https://docs.axolotl.ai/docs/multi-gpu.html), [Multi-node training (Torchrun, Ray)](https://docs.axolotl.ai/docs/multi-node.html), and many more!
+- **Performance Optimizations**: [Multipacking](https://docs.axolotl.ai/docs/multipack.html), [Flash Attention](https://github.com/Dao-AILab/flash-attention), [Xformers](https://github.com/facebookresearch/xformers), [Flex Attention](https://pytorch.org/blog/flexattention/), [SageAttention](https://github.com/thu-ml/SageAttention), [Liger Kernel](https://github.com/linkedin/Liger-Kernel), [Cut Cross Entropy](https://github.com/apple/ml-cross-entropy/tree/main), [ScatterMoE](https://docs.axolotl.ai/docs/custom_integrations.html#kernels-integration), [Sequence Parallelism (SP)](https://docs.axolotl.ai/docs/sequence_parallelism.html), [LoRA optimizations](https://docs.axolotl.ai/docs/lora_optims.html), [Multi-GPU training (FSDP1, FSDP2, DeepSpeed)](https://docs.axolotl.ai/docs/multi-gpu.html), [Multi-node training (Torchrun, Ray)](https://docs.axolotl.ai/docs/multi-node.html), and many more!
 - **Flexible Dataset Handling**: Load from local, HuggingFace, and cloud (S3, Azure, GCP, OCI) datasets.
 - **Cloud Ready**: We ship [Docker images](https://hub.docker.com/u/axolotlai) and also [PyPI packages](https://pypi.org/project/axolotl/) for use on cloud platforms and local hardware.
 
@@ -88,7 +98,7 @@ Features:
 #### Using pip
 
 ```bash
-pip3 install -U packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install -U packaging==26.0 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation axolotl[flash-attn,deepspeed]
 
 # Download example axolotl configs, deepspeed configs

VERSION

@@ -0,0 +1 @@
+0.16.0.dev0

_quarto.yml

@@ -251,7 +251,6 @@ website:
                 - docs/models/olmo3.qmd
                 - docs/models/trinity.qmd
                 - docs/models/arcee.qmd
-                - docs/models/mistral.qmd
                 - section: "Ministral3"
                   contents:
                     - docs/models/ministral3.qmd
@@ -266,6 +265,7 @@ website:
                 - docs/models/mistral-small.qmd
                 - docs/models/voxtral.qmd
                 - docs/models/devstral.qmd
+                - docs/models/mistral.qmd
                 - docs/models/llama-4.qmd
                 - docs/models/llama-2.qmd
                 - docs/models/qwen3-next.qmd
@@ -320,6 +320,7 @@ website:
             - docs/multipack.qmd
             - docs/mixed_precision.qmd
             - docs/optimizers.qmd
+            - docs/attention.qmd
 
         - section: "Advanced Features"
           contents:
@@ -330,6 +331,7 @@ website:
             - docs/sequence_parallelism.qmd
             - docs/gradient_checkpointing.qmd
             - docs/nd_parallelism.qmd
+            - docs/expert_quantization.qmd
 
         - section: "Troubleshooting"
           contents:

cicd/Dockerfile-uv.jinja

@@ -31,8 +31,9 @@ RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
         sed -i 's#^datasets.*#datasets @ git+https://github.com/huggingface/datasets.git@main#' requirements.txt; \
     fi
 
-RUN uv pip install packaging==23.2 setuptools==75.8.0
+RUN uv pip install packaging==26.0 setuptools==78.1.1
 RUN uv pip install torchvision
+RUN uv pip uninstall causal_conv1d
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
         uv pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \

cicd/Dockerfile.jinja

@@ -32,7 +32,8 @@ RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
         sed -i 's#^datasets.*#datasets @ git+https://github.com/huggingface/datasets.git@main#' requirements.txt; \
     fi
 
-RUN pip install packaging==23.2 setuptools==75.8.0 psutil
+RUN pip install packaging==26.0 setuptools==78.1.1 psutil
+RUN pip uninstall -y causal_conv1d
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
         pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \

cicd/cicd.sh

@@ -3,6 +3,12 @@ set -e
 
 python -c "import torch; assert '$PYTORCH_VERSION' in torch.__version__"
 
+# curl -L https://axolotl-ci.b-cdn.net/hf-cache.tar.zst | tar -xpf - -C "${HF_HOME}/hub/"  --use-compress-program unzstd --strip-components=1
+hf download "NousResearch/Meta-Llama-3-8B"
+hf download "NousResearch/Meta-Llama-3-8B-Instruct"
+hf download "microsoft/Phi-4-reasoning"
+hf download "microsoft/Phi-3.5-mini-instruct"
+
 # Run unit tests with initial coverage report
 pytest -v --durations=10 -n8 \
   --ignore=tests/e2e/ \

cicd/multigpu.py

@@ -17,7 +17,8 @@ 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")
+dockerfile = os.environ.get("E2E_DOCKERFILE", "Dockerfile.jinja")
+df_template = template_env.get_template(dockerfile)
 
 df_args = {
     "AXOLOTL_EXTRAS": os.environ.get("AXOLOTL_EXTRAS", ""),
@@ -27,8 +28,11 @@ df_args = {
     "CUDA": os.environ.get("CUDA", "126"),
     "GITHUB_REF": os.environ.get("GITHUB_REF", "refs/heads/main"),
     "GITHUB_SHA": os.environ.get("GITHUB_SHA", ""),
+    "NIGHTLY_BUILD": os.environ.get("NIGHTLY_BUILD", ""),
     "CODECOV_TOKEN": os.environ.get("CODECOV_TOKEN", ""),
     "HF_HOME": "/workspace/data/huggingface-cache/hub",
+    "PYTHONUNBUFFERED": os.environ.get("PYTHONUNBUFFERED", "1"),
+    "DEEPSPEED_LOG_LEVEL": os.environ.get("DEEPSPEED_LOG_LEVEL", "WARNING"),
 }
 
 dockerfile_contents = df_template.render(**df_args)

cicd/multigpu.sh

@@ -2,7 +2,7 @@
 set -e
 
 # Only run two tests at a time to avoid OOM on GPU (with coverage collection)
-pytest -v --durations=10 -n2 --maxfail=4 \
+pytest -v --durations=10 -n2 --maxfail=3 \
   --ignore=/workspace/axolotl/tests/e2e/multigpu/solo/ \
   --ignore=/workspace/axolotl/tests/e2e/multigpu/patched/ \
   /workspace/axolotl/tests/e2e/multigpu/ \

docker/Dockerfile

@@ -22,6 +22,7 @@ RUN git clone --depth=1 https://github.com/axolotl-ai-cloud/axolotl.git
 WORKDIR /workspace/axolotl
 
 # If AXOLOTL_EXTRAS is set, append it in brackets; don't install deepspeed with arm64
+RUN pip uninstall -y causal_conv1d
 RUN if [ "$TARGETARCH" = "arm64" ]; then \
         BASE_EXTRAS="flash-attn,ring-flash-attn,optimizers,ray"; \
     else \

docker/Dockerfile-base

@@ -43,7 +43,7 @@ ENV PATH="/root/miniconda3/envs/py${PYTHON_VERSION}/bin:${PATH}"
 
 WORKDIR /workspace
 
-RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==23.2 setuptools==75.8.0 wheel psutil && \
+RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==26.0 setuptools==75.8.0 wheel psutil && \
     python3 -m pip install --no-cache-dir -U torch==${PYTORCH_VERSION}+cu${CUDA} torchvision --extra-index-url https://download.pytorch.org/whl/cu$CUDA && \
     python3 -m pip cache purge
 
@@ -59,34 +59,18 @@ RUN git lfs install --skip-repo && \
     pip3 install -U --no-cache-dir pydantic==1.10.10 && \
     pip3 cache purge
 
-RUN case "$PYTORCH_VERSION" in \
-        2.9.[0-9]*) \
-            if [ "$CUDA" = "128" ]; then \
-                if [ "$TARGETARCH" = "amd64" ]; then \
-                    WHL_FILE="flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_x86_64.whl"; \
-                    WHL_VERSION="v0.5.4"; \
-                elif [ "$TARGETARCH" = "arm64" ]; then \
-                    WHL_FILE="flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_aarch64.whl"; \
-                    WHL_VERSION="v0.6.4"; \
-                else \
-                    echo "Unsupported architecture: $TARGETARCH"; exit 1; \
-                fi; \
-                wget -nv https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/${WHL_VERSION}/${WHL_FILE}; \
-                pip3 install --no-cache-dir ${WHL_FILE}; \
-                rm ${WHL_FILE}; \
-            elif [ "$CUDA" = "130" ]; then \
-                if [ "$TARGETARCH" = "amd64" ]; then \
-                    WHL_FILE="flash_attn-2.8.3+cu130torch2.9-cp311-cp311-linux_x86_64.whl"; \
-                    WHL_VERSION="v0.5.4"; \
-                elif [ "$TARGETARCH" = "arm64" ]; then \
-                    WHL_FILE="flash_attn-2.8.3+cu130torch2.9-cp311-cp311-linux_aarch64.whl"; \
-                    WHL_VERSION="v0.6.4"; \
-                else \
-                    echo "Unsupported architecture: $TARGETARCH"; exit 1; \
-                fi; \
-                wget -nv https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/${WHL_VERSION}/${WHL_FILE}; \
-                pip3 install --no-cache-dir ${WHL_FILE}; \
-                rm ${WHL_FILE}; \
-            fi \
-            ;; \
-    esac
+# Map Python version (e.g., 3.12 -> cp312)
+RUN PYTHON_CP="cp$(echo $PYTHON_VERSION | tr -d '.')" && \
+    # Map PyTorch version (e.g., 2.9.1 -> torch2.9, 2.10.0 -> torch2.10)
+    TORCH_TAG="torch$(echo $PYTORCH_VERSION | grep -oP '^\d+\.\d+')" && \
+    # Map architecture
+    case "$TARGETARCH" in \
+        amd64) ARCH_TAG="x86_64" ;; \
+        arm64) ARCH_TAG="aarch64" ;; \
+        *) echo "Unsupported architecture: $TARGETARCH"; exit 1 ;; \
+    esac && \
+    WHL_VERSION="v0.7.16" && \
+    WHL_FILE="flash_attn-2.8.3+cu${CUDA}${TORCH_TAG}-${PYTHON_CP}-${PYTHON_CP}-linux_${ARCH_TAG}.whl" && \
+    wget -nv "https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/${WHL_VERSION}/${WHL_FILE}" && \
+    pip3 install --no-cache-dir "${WHL_FILE}" && \
+    rm "${WHL_FILE}"

docker/Dockerfile-base-nightly

@@ -30,7 +30,7 @@ ENV PATH="/root/miniconda3/envs/py${PYTHON_VERSION}/bin:${PATH}"
 
 WORKDIR /workspace
 
-RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==23.2 setuptools==75.8.0 wheel && \
+RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==26.0 setuptools==75.8.0 wheel && \
     python3 -m pip install --no-cache-dir -U torch --extra-index-url https://download.pytorch.org/whl/nightly/cu$CUDA && \
     python3 -m pip install --no-cache-dir "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main" && \
     python3 -m pip install --no-cache-dir "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main" && \

docker/Dockerfile-cloud-uv

@@ -0,0 +1,30 @@
+ARG BASE_TAG=main
+FROM axolotlai/axolotl-uv:$BASE_TAG
+
+ENV HF_DATASETS_CACHE="/workspace/data/huggingface-cache/datasets"
+ENV HF_HUB_CACHE="/workspace/data/huggingface-cache/hub"
+ENV HF_HOME="/workspace/data/huggingface-cache/hub"
+ENV HF_HUB_ENABLE_HF_TRANSFER="1"
+
+EXPOSE 8888
+EXPOSE 22
+
+COPY scripts/cloud-entrypoint.sh /root/cloud-entrypoint.sh
+COPY scripts/motd /etc/motd
+
+RUN uv pip install jupyterlab notebook ipywidgets && \
+    jupyter lab clean
+RUN apt update && \
+    apt install --yes --no-install-recommends openssh-server tmux iproute2 nvtop && \
+    rm -rf /var/cache/apt/archives && \
+    rm -rf /var/lib/apt/lists/* && \
+    mkdir -p ~/.ssh && \
+    chmod 700 ~/.ssh && \
+    printf "\n[[ -z \"\$TMUX\"  ]] && { tmux attach-session -t ssh_tmux || tmux new-session -s ssh_tmux; exit; }\n" >> ~/.bashrc && \
+    printf "[ ! -z \"\$TERM\" -a -r /etc/motd ] && cat /etc/motd\n" >> ~/.bashrc && \
+    chmod +x /workspace/axolotl/scripts/cloud-entrypoint.sh && \
+    chmod +x /root/cloud-entrypoint.sh && \
+    echo 'set-option -g history-limit 5000' >> ~/.tmux.conf
+
+ENTRYPOINT ["/root/cloud-entrypoint.sh"]
+CMD ["sleep", "infinity"]

docker/Dockerfile-uv

@@ -0,0 +1,48 @@
+ARG BASE_TAG=main-base
+FROM axolotlai/axolotl-base-uv:$BASE_TAG
+
+ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX"
+ARG AXOLOTL_EXTRAS=""
+ARG AXOLOTL_ARGS=""
+ARG CUDA="118"
+ARG PYTORCH_VERSION="2.1.2"
+ARG TARGETARCH
+
+ENV PYTORCH_VERSION=$PYTORCH_VERSION
+
+RUN apt-get update && \
+    apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev rsync s3fs && \
+    rm -rf /var/cache/apt/archives && \
+    rm -rf /var/lib/apt/lists/*
+
+WORKDIR /workspace
+
+RUN git clone --depth=1 https://github.com/axolotl-ai-cloud/axolotl.git
+
+WORKDIR /workspace/axolotl
+
+# If AXOLOTL_EXTRAS is set, append it in brackets; don't install deepspeed with arm64
+RUN uv pip uninstall causal_conv1d
+RUN if [ "$TARGETARCH" = "arm64" ]; then \
+        BASE_EXTRAS="flash-attn,ring-flash-attn,optimizers,ray"; \
+    else \
+        BASE_EXTRAS="deepspeed,flash-attn,ring-flash-attn,optimizers,ray"; \
+    fi && \
+    if [ "$AXOLOTL_EXTRAS" != "" ]; then \
+        uv pip install --no-build-isolation -e .[$BASE_EXTRAS,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
+    else \
+        uv pip install --no-build-isolation -e .[$BASE_EXTRAS] $AXOLOTL_ARGS; \
+    fi && \
+    python scripts/unsloth_install.py --uv | sh && \
+    python scripts/cutcrossentropy_install.py --uv | sh && \
+    uv pip install pytest && \
+    uv cache clean
+
+# fix so that git fetch/pull from remote works with shallow clone
+RUN git config remote.origin.fetch "+refs/heads/*:refs/remotes/origin/*" && \
+    git config --get remote.origin.fetch && \
+    git config --global credential.helper store
+
+COPY .axolotl-complete.bash /root/.axolotl-complete.bash
+RUN chmod +x /root/.axolotl-complete.bash && \
+    echo 'source /root/.axolotl-complete.bash' >> ~/.bashrc

docker/Dockerfile-uv-base

@@ -6,6 +6,7 @@ ARG TARGETARCH
 
 FROM nvidia/cuda:$CUDA_VERSION-cudnn$CUDNN_VERSION-devel-ubuntu$UBUNTU_VERSION AS base-builder
 
+ARG TARGETARCH
 ARG PYTHON_VERSION="3.11"
 ARG PYTORCH_VERSION="2.6.0"
 ARG CUDA="126"
@@ -39,28 +40,18 @@ RUN if [ "$TARGETARCH" = "amd64" ]; then \
         uv pip install "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main"; \
     fi
 
-RUN case "$PYTORCH_VERSION" in \
-        2.9.[0-9]*) \
-            if [ "$TARGETARCH" = "amd64" ]; then \
-                if [ "$CUDA" = "128" ]; then \
-                    wget -nv https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.5.4/flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_x86_64.whl; \
-                    uv pip install --no-cache-dir flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_x86_64.whl; \
-                    rm flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_x86_64.whl; \
-                elif [ "$CUDA" = "130" ]; then \
-                    wget -nv https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.5.4/flash_attn-2.8.3+cu130torch2.9-cp311-cp311-linux_x86_64.whl; \
-                    uv pip install --no-cache-dir flash_attn-2.8.3+cu130torch2.9-cp311-cp311-linux_x86_64.whl; \
-                    rm flash_attn-2.8.3+cu130torch2.9-cp311-cp311-linux_x86_64.whl; \
-                fi \
-            elif [ "$TARGETARCH" = "arm64" ]; then \
-                if [ "$CUDA" = "128" ]; then \
-                    wget -nv https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.6.4/flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_aarch64.whl; \
-                    uv pip install --no-cache-dir flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_aarch64.whl; \
-                    rm flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_aarch64.whl; \
-                elif [ "$CUDA" = "130" ]; then \
-                    wget -nv https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.6.4/flash_attn-2.8.3+cu130torch2.9-cp311-cp311-linux_aarch64.whl; \
-                    uv pip install --no-cache-dir flash_attn-2.8.3+cu130torch2.9-cp311-cp311-linux_aarch64.whl; \
-                    rm flash_attn-2.8.3+cu130torch2.9-cp311-cp311-linux_aarch64.whl; \
-                fi \
-            fi \
-            ;; \
-    esac
+# Map Python version (e.g., 3.12 -> cp312)
+RUN PYTHON_CP="cp$(echo $PYTHON_VERSION | tr -d '.')" && \
+    # Map PyTorch version (e.g., 2.9.1 -> torch2.9, 2.10.0 -> torch2.10)
+    TORCH_TAG="torch$(echo $PYTORCH_VERSION | grep -oP '^\d+\.\d+')" && \
+    # Map architecture
+    case "$TARGETARCH" in \
+        amd64) ARCH_TAG="x86_64" ;; \
+        arm64) ARCH_TAG="aarch64" ;; \
+        *) echo "Unsupported architecture: $TARGETARCH"; exit 1 ;; \
+    esac && \
+    WHL_VERSION="v0.7.16" && \
+    WHL_FILE="flash_attn-2.8.3+cu${CUDA}${TORCH_TAG}-${PYTHON_CP}-${PYTHON_CP}-linux_${ARCH_TAG}.whl" && \
+    wget -nv "https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/${WHL_VERSION}/${WHL_FILE}" && \
+    uv pip install --no-cache-dir "${WHL_FILE}" && \
+    rm "${WHL_FILE}"

docs/amd_hpc.qmd

@@ -86,7 +86,7 @@ export HF_DATASETS_OFFLINE=1
 Download a base model using the Hugging Face CLI:
 
 ```bash
-huggingface-cli download meta-llama/Meta-Llama-3.1-8B --local-dir ~/hfdata/llama3.1-8B
+hf download meta-llama/Meta-Llama-3.1-8B --local-dir ~/hfdata/llama3.1-8B
 ```
 
 ### 10. Create Axolotl Configuration

docs/attention.qmd

@@ -0,0 +1,140 @@
+---
+title: Attention
+description: Supported attention modules in Axolotl
+---
+
+## SDP Attention
+
+This is the default built-in attention in PyTorch.
+
+```yaml
+sdp_attention: true
+```
+
+For more details: [PyTorch docs](https://docs.pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html)
+
+## Flash Attention 2
+
+Uses efficient kernels to compute attention.
+
+```yaml
+flash_attention: true
+```
+
+For more details: [Flash Attention](https://github.com/Dao-AILab/flash-attention/)
+
+### Nvidia
+
+Requirements: Ampere, Ada, or Hopper GPUs
+
+Note: For Turing GPUs or lower, please use other attention methods.
+
+```bash
+pip install flash-attn --no-build-isolation
+```
+
+::: {.callout-tip}
+
+If you get `undefined symbol` while training, ensure you installed PyTorch prior to Axolotl. Alternatively, try reinstall or downgrade a version.
+
+:::
+
+#### Flash Attention 3
+
+Requirements: Hopper only and CUDA 12.8 (recommended)
+
+```bash
+git clone https://github.com/Dao-AILab/flash-attention.git
+cd flash-attention/hopper
+
+python setup.py install
+```
+
+### AMD
+
+Requirements: ROCm 6.0 and above.
+
+See [Flash Attention AMD docs](https://github.com/Dao-AILab/flash-attention/tree/main?tab=readme-ov-file#amd-rocm-support).
+
+## Flex Attention
+
+A flexible PyTorch API for attention used in combination with `torch.compile`.
+
+```yaml
+flex_attention: true
+
+# recommended
+torch_compile: true
+```
+
+::: {.callout-note}
+
+We recommend using latest stable version of PyTorch for best performance.
+
+:::
+
+For more details: [PyTorch docs](https://pytorch.org/blog/flexattention/)
+
+## SageAttention
+
+Attention kernels with QK Int8 and PV FP16 accumulator.
+
+```yaml
+sage_attention: true
+```
+
+Requirements: Ampere, Ada, or Hopper GPUs
+
+```bash
+pip install sageattention==2.2.0 --no-build-isolation
+```
+
+::: {.callout-warning}
+
+Only LoRA/QLoRA recommended at the moment. We found loss drop to 0 for full finetuning. See [GitHub Issue](https://github.com/thu-ml/SageAttention/issues/198).
+
+:::
+
+For more details: [Sage Attention](https://github.com/thu-ml/SageAttention)
+
+::: {.callout-note}
+
+We do not support SageAttention 3 at the moment. If you are interested on adding this or improving SageAttention implementation, please make an Issue.
+
+:::
+
+
+## xFormers
+
+```yaml
+xformers_attention: true
+```
+
+::: {.callout-tip}
+
+We recommend using with Turing GPUs or below (such as on Colab).
+
+:::
+
+For more details: [xFormers](https://github.com/facebookresearch/xformers)
+
+## Shifted Sparse Attention
+
+::: {.callout-warning}
+
+We plan to deprecate this! If you use this feature, we recommend switching to methods above.
+
+:::
+
+Requirements: LLaMA model architecture
+
+```yaml
+flash_attention: true
+s2_attention: true
+```
+
+::: {.callout-tip}
+
+No sample packing support!
+
+:::

docs/cli.qmd

@@ -210,6 +210,8 @@ axolotl lm-eval config.yml
 Configuration options:
 
 ```yaml
+lm_eval_model: # model to evaluate (local or hf path)
+
 # List of tasks to evaluate
 lm_eval_tasks:
   - arc_challenge
@@ -218,7 +220,7 @@ lm_eval_batch_size: # Batch size for evaluation
 output_dir: # Directory to save evaluation results
 ```
 
-See [LM Eval Harness](https://github.com/EleutherAI/lm-evaluation-harness) for more details.
+See [LM Eval Harness integration docs](https://docs.axolotl.ai/docs/custom_integrations.html#language-model-evaluation-harness-lm-eval) for full configuration details.
 
 ### delinearize-llama4
 

docs/expert_quantization.qmd

@@ -0,0 +1,67 @@
+---
+title: "MoE Expert Quantization"
+description: "Reduce VRAM usage when training MoE model adapters by quantizing expert weights on load"
+---
+
+Transformers v5 changed MoE expert layers from `nn.Linear` to fused `nn.Parameter` (3D+ tensors).
+This means `bitsandbytes` can no longer quantize them during model loading, resulting in all expert
+weights being loaded in full bf16 precision and causing massive VRAM usage.
+
+`quantize_moe_experts` solves this by quantizing expert weights during model loading.
+It intercepts the weight loading process, quantizes each expert tensor on the fly, and
+immediately frees the original bf16 tensor from VRAM. This dramatically reduces peak memory.
+For example, GLM-4.7-Flash QLoRA drops from ~127GiB to ~23GiB reserved memory.
+
+## Usage
+
+Enable expert quantization in your Axolotl config:
+
+```yaml
+quantize_moe_experts: true
+```
+
+This works with both 4-bit (QLoRA) and 8-bit (LoRA) quantization.
+
+### Expert LoRA targeting
+
+You can optionally apply LoRA adapters directly to expert weights using `lora_target_parameters`:
+
+```yaml
+lora_target_parameters:
+  - mlp.experts.gate_up_proj
+  - mlp.experts.down_proj
+  # - mlp.gate.weight  # router
+```
+
+::: {.callout-note}
+`lora_dropout` must be `0` when using `lora_target_parameters`.
+:::
+
+## Requirements
+
+- Requires (`adapter: lora` and `load_in_8bit: true`) or (`adapter: qlora` and `load_in_4bit: true`)
+- CUDA GPUs only (not tested with ROCm or other backends)
+- FSDP2 compatible for distributed training
+
+## Limitations
+
+- `lora_target_linear` is not compatible with `quantize_moe_experts`. See [Expert LoRA targeting](#expert-lora-targeting) instead.
+- `cpu_ram_efficient_loading` hangs / takes long time with FSDP2 + QLoRA.
+- Total model parameter count may display incorrectly (trainable param count is correct).
+- FSDP LoRA (8-bit) may have a large initial VRAM spike at the first 1-2 steps, which then drops. QLoRA does not exhibit this.
+- FSDP2 may use more VRAM per GPU than single GPU training due to not all layers being properly sharded across ranks.
+- Model loading takes longer due to on-demand quantization, even on consecutive runs.
+- DeepSpeed has not been tested.
+
+## Implementation details
+
+The quantization is applied by patching transformers to intercept weight loading.
+When a 3D+ CUDA tensor with "expert" in its name is detected:
+
+- **4-bit mode:** Uses bitsandbytes NF4 parametrization (configurable via `bnb_4bit_quant_type`).
+- **8-bit mode:** Uses a custom row-wise int8 parametrization with bitsandbytes dequantization.
+
+The original bf16 tensor is freed immediately after quantization. Multiple sub-patches are applied to
+transformers, PEFT and accelerate FSDP2 to support these parametrized expert modules.
+
+For full implementation details, see [PR #3439](https://github.com/axolotl-ai-cloud/axolotl/pull/3439).

docs/installation.qmd

@@ -165,7 +165,7 @@ We recommend using WSL2 (Windows Subsystem for Linux) or Docker.
    ```
 4. (Optional) Login to Hugging Face:
    ```{.bash}
-   huggingface-cli login
+   hf auth login
    ```
 
 ## Troubleshooting {#sec-troubleshooting}

docs/lora_optims.qmd

@@ -89,6 +89,10 @@ lora_o_kernel: true
 Currently, LoRA kernels are not supported for RLHF training, only SFT.
 :::
 
+::: {.callout-warning}
+LoRA kernels do not support remote modeling code.
+:::
+
 ## Requirements
 
 - One or more NVIDIA or AMD GPUs (in order to use the Triton kernels)

docs/multimodal.qmd

@@ -19,6 +19,7 @@ format:
 - [Gemma-3n](#sec-gemma-3n)
 - [Qwen2-VL](#sec-qwen2-vl)
 - [Qwen2.5-VL](#sec-qwen25-vl)
+- [GLM-4.6V](#sec-glm-4-6v)
 - [SmolVLM2](#sec-smolvlm2)
 - [LFM2-VL](#sec-lfm2-vl)
 - [Intern-VL](#sec-intern-vl)
@@ -183,6 +184,18 @@ base_model: Qwen/Qwen3-VL-4B-Instruct
 chat_template: qwen2_vl  # same as qwen2-vl
 ```
 
+### GLM-4.6V {#sec-glm-4-6v}
+
+Both GLM-4.6V (106B MoE) and GLM-4.6V-Flash (9B) are supported.
+
+```yaml
+# GLM-4.6V (106B MoE version)
+base_model: zai-org/GLM-4.6V
+
+# OR GLM-4.6V-Flash (9B version)
+base_model: zai-org/GLM-4.6V-Flash
+```
+
 ### SmolVLM2 {#sec-smolvlm2}
 
 ::: {.callout-tip}

docs/optimizations.qmd

@@ -66,6 +66,15 @@ Provides efficient Triton kernels to improve training speed and reduce memory us
 
 - **Learn more:** [Custom Integrations - Liger Kernels](custom_integrations.qmd#liger-kernels)
 
+### Expert Kernels
+
+Optimized kernel implementations for Mixture of Experts (MoE) model training.
+
+- **ScatterMoE**: Triton-based MoE kernels with fused LoRA support.
+- **SonicMoE**: CUTLASS-based MoE kernels for NVIDIA Hopper and Blackwell GPUs.
+
+- **Learn more:** [Custom Integrations - Kernels Integration](custom_integrations.qmd#kernels-integration)
+
 ## Long Context Models
 
 Techniques to train models on sequences longer than their original context window.
@@ -131,3 +140,10 @@ Simulates quantization effects during training, helping the model adapt and pote
 Allows you to finetune LoRA adapters on top of a model that has already been quantized using the GPTQ method.
 
 - **Example:** [GPTQ LoRA Example](https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-2/gptq-lora.yml)
+
+### MoE Expert Quantization
+
+Quantizes MoE expert weights on load to reduce VRAM when training MoE models with adapters. Required for Transformers v5+ MoE models where experts use fused `nn.Parameter` tensors.
+
+- **Config:** `quantize_moe_experts: true`
+- **Learn more:** [MoE Expert Quantization](expert_quantization.qmd)

docs/rlhf.qmd

@@ -17,6 +17,7 @@ feedback. Various methods include, but not limited to:
 - [Kahneman-Tversky Optimization (KTO)](#kto)
 - [Odds Ratio Preference Optimization (ORPO)](#orpo)
 - [Group Relative Policy Optimization (GRPO)](#grpo)
+- [Group Reward-Decoupled Policy Optimization (GDPO)](#gdpo)
 
 
 ## RLHF using Axolotl
@@ -720,6 +721,102 @@ trl:
 
 For more information, see [GRPO docs](https://huggingface.co/docs/trl/v0.17.0/en/grpo_trainer#loss-types).
 
+### GDPO
+
+GDPO (Group Reward-Decoupled Policy Optimization) extends GRPO for multi-reward training. It addresses the **reward advantage collapse** problem by normalizing each reward function independently before combining them.
+
+::: {.callout-tip}
+Use GDPO when training with multiple reward functions. For single reward, GRPO and GDPO produce equivalent results.
+:::
+
+Paper: [https://arxiv.org/pdf/2501.05242](https://arxiv.org/pdf/2501.05242)
+
+GDPO uses TRL's native `multi_objective_aggregation` parameter under the hood. When you set `rl: gdpo`, axolotl automatically configures TRL to use `normalize_then_sum` aggregation.
+
+```yaml
+base_model: Qwen/Qwen2.5-1.5B-Instruct
+
+vllm:
+    host: 0.0.0.0
+    port: 8000
+    tensor_parallel_size: 2
+    gpu_memory_utilization: 0.85
+
+rl: gdpo
+
+trl:
+    beta: 0.001
+    max_completion_length: 256
+    use_vllm: true
+    num_generations: 4
+    reward_funcs:
+        - rewards.format_reward
+        - rewards.correctness_reward
+    reward_weights: [1.0, 2.0]
+
+datasets:
+    - path: openai/gsm8k
+      name: main
+      type: rewards.oai_gsm8k_transform
+```
+
+You can also use GRPO with explicit aggregation control:
+
+```yaml
+rl: grpo
+trl:
+    multi_objective_aggregation: normalize_then_sum  # GDPO behavior
+    # or: sum_then_normalize  # Default GRPO behavior
+```
+
+#### GDPO vs GRPO
+
+| Aspect | GRPO | GDPO |
+|--------|------|------|
+| **Aggregation** | `sum_then_normalize` | `normalize_then_sum` |
+| **Multi-reward** | May collapse advantages | Preserves reward signals |
+| **Single reward** | Standard behavior | Equivalent to GRPO |
+
+#### Why GDPO?
+
+When using multiple rewards with GRPO, different reward combinations can produce identical advantages:
+
+```
+# Example: format + correctness rewards
+[format=0, correct=3] → sum=3
+[format=1, correct=2] → sum=3  ← GRPO sees these as equal!
+[format=2, correct=1] → sum=3
+[format=3, correct=0] → sum=3
+```
+
+GDPO normalizes each reward independently, preserving their relative differences.
+
+#### Reward Functions
+
+GDPO uses the same reward function format as GRPO:
+
+```python
+# rewards.py
+def format_reward(completions, **kwargs) -> list[float]:
+    return [1.0 if len(c) > 10 else 0.0 for c in completions]
+
+def correctness_reward(completions, answers, **kwargs) -> list[float]:
+    rewards = []
+    for completion, answer in zip(completions, answers):
+        # Your scoring logic here
+        rewards.append(score)
+    return rewards
+```
+
+#### Sequence Parallelism
+
+GDPO supports sequence parallelism for long-context training:
+
+```yaml
+rl: gdpo
+context_parallel_size: 2
+```
+
 ### SimPO
 
 SimPO uses [CPOTrainer](https://huggingface.co/docs/trl/main/en/cpo_trainer) but with alternative loss function.

examples/apertus/README.md

@@ -15,7 +15,7 @@ This guide shows how to fine-tune it with Axolotl with multi-turn conversations
 git clone https://github.com/axolotl-ai-cloud/axolotl.git
 cd axolotl
 
-pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install packaging==26.0 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation -e '.[flash-attn]'
 
 # Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy

examples/arcee/README.md

@@ -17,7 +17,7 @@ Thanks to the team at Arcee.ai for using Axolotl in supervised fine-tuning the A
 git clone https://github.com/axolotl-ai-cloud/axolotl.git
 cd axolotl
 
-pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install packaging==26.0 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation -e '.[flash-attn]'
 
 # Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy

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

@@ -40,7 +40,7 @@
     "%%capture\n",
     "# This step can take ~5-10 minutes to install dependencies\n",
     "!pip install --no-build-isolation axolotl[flash-attn]>=0.9.1\n",
-    "!pip install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@318b7e2\""
+    "!pip install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@e8ad129\""
    ]
   },
   {

examples/devstral/README.md

@@ -16,7 +16,7 @@ Thanks to the team at MistralAI for giving us early access to prepare for this r
 
 ```bash
 # Ensure you have Pytorch installed (Pytorch 2.6.0 min)
-pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install packaging==26.0 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 

examples/eaft/eaft-example.yml

@@ -0,0 +1,77 @@
+base_model: google/gemma-3-1b-it
+
+model_type: Gemma3ForCausalLM
+cls_model_config: Gemma3TextConfig
+
+# gemma3 doesn't seem to play nice with ddp
+ddp_find_unused_parameters: true
+
+chat_template: gemma3
+eot_tokens:
+  - <end_of_turn>
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: cgato/SlimOrcaDedupCleaned
+    type: chat_template
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+dataset_prepared_path:
+val_set_size: 0
+output_dir: ./outputs/eaft-gemma-3-1b
+
+use_eaft: true
+eaft_alpha: 1.0
+eaft_k: 20
+
+sequence_len: 1024
+sample_packing: false
+
+adapter:
+lora_model_dir:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+eval_batch_size: 1
+max_steps: 1000
+evaluation_strategy: "no"
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 5e-5
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+debug:
+deepspeed:
+fsdp:
+fsdp_config:
+special_tokens:

examples/gemma3n/README.md

@@ -10,7 +10,7 @@ Gemma-3n is a family of multimodal models from Google found on [HuggingFace](htt
 
 ```bash
 # Ensure you have Pytorch installed (Pytorch 2.6.0 min)
-pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install packaging==26.0 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 

examples/glm45/README.md

@@ -0,0 +1,72 @@
+# Finetune Z.ai's GLM-4.5-Air with Axolotl
+
+[GLM-4.5-Air](https://huggingface.co/zai-org/GLM-4.5-Air) is a MoE model by Z.ai.
+
+This guide shows how to fine-tune it with Axolotl.
+
+## Getting started
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
+
+2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage.
+
+3. Run the finetuning example:
+
+```bash
+# QLoRA (1x80GB @ ~63.4GiB/GPU)
+axolotl train examples/glm45/glm-45-air-qlora.yaml
+```
+
+### Dataset
+
+In addition to the standard OpenAI Messages format, GLM-4.5 supports an extra parameter for thinking in the assistant section.
+
+```json
+{
+    "role": "assistant",
+    "reasoning_content": "...",  // or have </think>...</think> in `content`
+    "content": "..."
+}
+```
+
+Make sure you set the below extra attributes if needed:
+
+```yaml
+datasets:
+  - path: ...
+    type: chat_template
+    message_property_mappings:
+      role: role
+      content: content
+
+    #   tool_calls: tool_calls  # uncomment if using tools
+    #   reasoning_content: reasoning_content  # uncomment if have reasoning
+
+# Uncomment if training on tool role (you would rarely if ever need this)
+# eot_tokens:
+#   - <|observation|>
+```
+
+### Tips
+
+- The role name for tools in this template is `tool`.
+- You will see this Axolotl WARNING — this is expected as the template does not use EOS:
+  ```
+  EOS token '<|endoftext|>' not found in chat_template. Please check if your template/EOS token is correct.
+  ```
+- You can run a full finetuning by removing `adapter: qlora`, `load_in_4bit: true`, and `quantize_moe_experts: true` from the config.
+- **LoRA kernels**: Incompatible with this model. Must be explicitly disabled (`lora_*_kernel: false`).
+- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+
+## Optimization Guides
+
+Please check the [Optimizations doc](https://docs.axolotl.ai/docs/optimizations.html).
+
+## Related Resources
+
+- [GLM-4.5-Air on HuggingFace](https://huggingface.co/zai-org/GLM-4.5-Air)
+- [GLM-4.5 Blog](https://z.ai/blog/glm-4.5)
+- [Axolotl Docs](https://docs.axolotl.ai)
+- [Axolotl Website](https://axolotl.ai)
+- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/glm45/glm-45-air-qlora.yaml

@@ -0,0 +1,64 @@
+base_model: zai-org/GLM-4.5-Air
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_8bit: false
+load_in_4bit: true
+
+quantize_moe_experts: true # important
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/lora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+
+lora_r: 16
+lora_alpha: 8
+lora_dropout: 0
+lora_target_modules:
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+# lora_target_parameters:
+#   - mlp.experts.gate_up_proj
+#   - mlp.experts.down_proj
+
+lora_mlp_kernel: false
+lora_qkv_kernel: false
+lora_o_kernel: false
+
+gradient_accumulation_steps: 2
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/glm46v/README.md

@@ -0,0 +1,44 @@
+# Finetune GLM-4.6V with Axolotl
+
+GLM-4.6V is a family of vision-language models from ZhipuAI found on [HuggingFace](https://huggingface.co/zai-org/GLM-4.6V). This guide shows how to fine-tune it with Axolotl for vision-language tasks.
+
+
+
+## Getting started
+
+1. Install Axolotl from source following the [installation guide](https://docs.axolotl.ai/docs/installation.html#sec-edge-build).
+
+2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage.
+
+
+3. Run the fine-tuning:
+
+    glm-4-6v-flash(9B)
+    ```bash
+    axolotl train examples/glm46v/glm-4-6v-flash-qlora.yaml
+    ```
+
+Let us know how it goes. Happy finetuning! 🚀
+
+## Tips
+
+- Vision datasets should follow the format described in the [multimodal docs](https://docs.axolotl.ai/docs/multimodal.html#dataset-format)
+- You can run a **full finetuning** by removing the `adapter: qlora` and `load_in_4bit: true` from the config.
+- Read more on how to load your own dataset in the [dataset loading docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+
+## Supported Models
+
+- **GLM-4.6V**: Full vision-language model (`zai-org/GLM-4.6V`)
+- **GLM-4.6V-Flash**: Faster variant (`zai-org/GLM-4.6V-Flash`)
+
+## Optimization Guides
+
+Please check the [Optimizations doc](https://docs.axolotl.ai/docs/optimizations.html).
+
+## Related Resources
+
+- [ZhipuAI GLM-4.6V](https://huggingface.co/zai-org/GLM-4.6V)
+- [Axolotl Docs](https://docs.axolotl.ai)
+- [Axolotl Website](https://axolotl.ai)
+- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/glm46v/glm-4-6v-flash-ddp.yaml

@@ -0,0 +1,53 @@
+base_model: zai-org/GLM-4.6V-Flash
+trust_remote_code: true
+
+processor_type: AutoProcessor
+load_in_4bit: true
+
+# these 3 lines are needed for now to handle vision chat templates w images
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+ddp_find_unused_parameters: true
+
+output_dir: ./outputs/glm-4-6v-flash-qlora
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+
+adapter: qlora
+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
+
+sequence_len: 2048
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+logging_steps: 1
+sdp_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 0
+saves_per_epoch: 1
+weight_decay: 0.0

examples/glm46v/glm-4-6v-flash-qlora.yaml

@@ -0,0 +1,50 @@
+base_model: zai-org/GLM-4.6V-Flash
+trust_remote_code: true
+
+processor_type: AutoProcessor
+load_in_4bit: true
+
+# these 3 lines are needed for now to handle vision chat templates w images
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+output_dir: ./outputs/glm-4-6v-flash-qlora
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+
+adapter: qlora
+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
+
+sequence_len: 2048
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+logging_steps: 1
+sdp_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 0
+saves_per_epoch: 1
+weight_decay: 0.0

examples/glm47-flash/README.md

@@ -0,0 +1,65 @@
+# Finetune Z.ai's GLM-4.7-Flash with Axolotl
+
+[GLM-4.7-Flash](https://huggingface.co/zai-org/GLM-4.7-Flash) is a 30B-A3B MoE model by Z.ai.
+
+This guide shows how to fine-tune it with Axolotl.
+
+## Getting started
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
+
+2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage.
+
+3. Run the finetuning example:
+
+```bash
+# QLoRA
+# - no target experts (1x48GB @ ~24GiB/GPU)
+# - target experts (1x48GB @ ~34GiB/GPU)
+axolotl train examples/glm47-flash/qlora.yaml
+
+# QLoRA FSDP2 no target experts (2x48GB @ ~29GiB/GPU)
+axolotl train examples/glm47-flash/qlora_fsdp.yaml
+```
+
+```bash
+# LoRA
+# - no target experts (1x48GB @ ~35GiB/GPU)
+# - target experts (1x48GB @ OOM. Projected ~45-50GiB/GPU)
+axolotl train examples/glm47-flash/lora.yaml
+
+# LoRA FSDP2 no target experts (2x48GB @ ~43GiB/GPU)
+axolotl train examples/glm47-flash/lora_fsdp.yaml
+```
+
+### MoE Expert Quantization & Expert LoRA
+
+This model quantize expert weights on load. To learn about expert quantization, expert LoRA targeting, and related limitations, see the [MoE Expert Quantization](https://docs.axolotl.ai/docs/expert_quantization.html) docs.
+
+## Limitations
+
+- **lora_target_linear**: Incompatible for this model.
+- **LoRA kernels**: Incompatible with this model due to non-standard attention projections (DSA). Must be explicitly disabled (`lora_*_kernel: false`).
+
+
+### TIPS
+
+- For inference, the official Z.ai team recommends these default settings (most tasks):
+  - `temperature: 1.0`
+  - `top_p: 0.95`
+  - `max_new_tokens: 131072`
+- You can run a full finetuning by removing `adapter: qlora`, `load_in_4bit: true`, and `quantize_moe_experts: true` from the config. This is heavy, so we have not tested this.
+- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+
+## Optimization Guides
+
+Please check the [Optimizations doc](https://docs.axolotl.ai/docs/optimizations.html).
+
+## Related Resources
+
+- [GLM-4.7-Flash on HuggingFace](https://huggingface.co/zai-org/GLM-4.7-Flash)
+- [GLM-4.7 Blog](https://z.ai/blog/glm-4.7)
+- [Axolotl Docs](https://docs.axolotl.ai)
+- [Axolotl Website](https://axolotl.ai)
+- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/glm47-flash/lora.yaml

@@ -0,0 +1,65 @@
+base_model: zai-org/GLM-4.7-Flash
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_8bit: true
+quantize_moe_experts: true
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/glm4.7-flash-lora-8bit-out
+
+adapter: lora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0
+lora_target_modules:
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+# Uncomment to also target MoE expert weights:
+# lora_target_parameters:
+#   - mlp.experts.gate_up_proj
+#   - mlp.experts.down_proj
+
+# LoRA kernels incompatible with DSA attention
+lora_mlp_kernel: false
+lora_qkv_kernel: false
+lora_o_kernel: false
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_torch_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1

examples/glm47-flash/lora_fsdp.yaml

@@ -0,0 +1,75 @@
+base_model: zai-org/GLM-4.7-Flash
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_8bit: true
+quantize_moe_experts: true
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/glm4.7-flash-lora-8bit-fsdp-out
+
+adapter: lora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0
+lora_target_modules:
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+# Uncomment to also target MoE expert weights:
+# lora_target_parameters:
+#   - mlp.experts.gate_up_proj
+#   - mlp.experts.down_proj
+
+# LoRA kernels incompatible with DSA attention
+lora_mlp_kernel: false
+lora_qkv_kernel: false
+lora_o_kernel: false
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_torch_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+fsdp_config:
+  fsdp_version: 2
+  offload_params: false
+  cpu_ram_efficient_loading: false
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Glm4MoeLiteDecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true

examples/glm47-flash/qlora.yaml

@@ -0,0 +1,65 @@
+base_model: zai-org/GLM-4.7-Flash
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_4bit: true
+quantize_moe_experts: true
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/glm4.7-flash-qlora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0
+lora_target_modules:
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+# Uncomment to also target MoE expert weights:
+# lora_target_parameters:
+#   - mlp.experts.gate_up_proj
+#   - mlp.experts.down_proj
+
+# LoRA kernels incompatible with DSA attention
+lora_mlp_kernel: false
+lora_qkv_kernel: false
+lora_o_kernel: false
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_torch_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1

examples/glm47-flash/qlora_fsdp.yaml

@@ -0,0 +1,75 @@
+base_model: zai-org/GLM-4.7-Flash
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_4bit: true
+quantize_moe_experts: true
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/glm4.7-flash-qlora-fsdp-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0
+lora_target_modules:
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+# Uncomment to also target MoE expert weights:
+# lora_target_parameters:
+#   - mlp.experts.gate_up_proj
+#   - mlp.experts.down_proj
+
+# LoRA kernels incompatible with DSA attention
+lora_mlp_kernel: false
+lora_qkv_kernel: false
+lora_o_kernel: false
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_torch_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+fsdp_config:
+  fsdp_version: 2
+  offload_params: false
+  cpu_ram_efficient_loading: false
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Glm4MoeLiteDecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true

examples/gpt-oss/README.md

@@ -14,7 +14,7 @@ This guide shows how to fine-tune it with Axolotl with multi-turn conversations
 
 ```bash
 # Ensure you have Pytorch installed (Pytorch 2.6.0 min)
-pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install packaging==26.0 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 

examples/granite4/README.md

@@ -15,7 +15,7 @@ This guide shows how to fine-tune it with Axolotl with multi-turn conversations
 git clone https://github.com/axolotl-ai-cloud/axolotl.git
 cd axolotl
 
-pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install packaging==26.0 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation -e '.[flash-attn]'
 
 # Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy

examples/hunyuan/README.md

@@ -13,7 +13,7 @@ Tencent released a family of opensource models called HunYuan with varying param
 git clone https://github.com/axolotl-ai-cloud/axolotl.git
 cd axolotl
 
-pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install packaging==26.0 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation -e '.[flash-attn]'
 
 # Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy

examples/jamba/qlora_fsdp_large.yaml

@@ -19,7 +19,6 @@ datasets:
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0
 output_dir: jamba-large-fsdp-qlora-ft
-save_safetensors: true
 adapter: qlora
 sequence_len: 2048
 sample_packing: true

examples/llama-3/3b-qat-mxfp4.yaml

@@ -0,0 +1,65 @@
+base_model: meta-llama/Llama-3.2-3B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+liger_layer_norm: true
+liger_fused_linear_cross_entropy: true
+
+datasets:
+  - path: yahma/alpaca-cleaned
+    type: alpaca
+    split: train[:95%]
+
+output_dir: ./outputs/qat_out/
+dataset_prepared_path: ./outputs/dataset_prepared
+
+sequence_len: 2048
+flash_attention: true
+
+qat:
+  activation_dtype: mxfp4
+  weight_dtype: mxfp4
+  group_size: 32
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_checkpointing: true
+activation_offloading: true
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_8bit
+
+cosine_constant_lr_ratio: 0
+cosine_min_lr_ratio: 1.0
+learning_rate: 2e-5
+save_only_model: true
+bf16: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+
+special_tokens:
+  pad_token: <|finetune_right_pad_id|>
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/llama-3/qlora-1b-gdpo.yaml

@@ -0,0 +1,68 @@
+base_model: meta-llama/Llama-3.2-1B-Instruct
+
+chat_template: llama3
+
+rl: gdpo
+
+trl:
+  beta: 0.001
+  max_completion_length: 128
+  num_generations: 2
+  temperature: 0.7
+  top_p: 0.95
+
+  use_vllm: false
+
+
+  multi_objective_aggregation: normalize_then_sum
+
+  reward_funcs:
+    - rwd.format_reward
+    - rwd.correctness_reward
+  reward_weights: [1.0, 2.0]
+
+  log_completions: true
+  num_completions_to_print: 3
+  scale_rewards: true
+
+datasets:
+  - path: openai/gsm8k
+    name: main
+    split: train[:1000]
+    type: rwd.gsm8k_transform
+
+val_set_size: 0.0
+output_dir: ./outputs/llama3-gdpo-out
+
+sequence_len: 512
+sample_packing: false
+pad_to_sequence_len: false
+
+gradient_accumulation_steps: 8
+micro_batch_size: 1
+num_epochs: 1
+max_steps: 100
+
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 5e-5
+weight_decay: 0.01
+warmup_steps: 10
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+
+flash_attention: true
+logging_steps: 1
+save_steps: 50
+save_safetensors: true
+
+special_tokens:
+  pad_token: "<|end_of_text|>"
+
+
+seed: 42

examples/llama-3/qlora-fsdp-405b.yaml

@@ -12,7 +12,6 @@ datasets:
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0
 output_dir: ./outputs/out/qlora-llama3_1-405b
-save_safetensors: true
 
 adapter: qlora
 

examples/magistral/README.md

@@ -14,7 +14,7 @@ Thanks to the team at MistralAI for giving us early access to prepare for these
 
 ```bash
 # Ensure you have Pytorch installed (Pytorch 2.7.0 min)
-pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install packaging==26.0 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 

examples/mamba/config.yml

@@ -47,6 +47,5 @@ saves_per_epoch: 1
 weight_decay: 0.0
 special_tokens:
 tokens:
-save_safetensors: False
 
 # save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qwen3-next/README.md

@@ -6,30 +6,13 @@ This guide shows how to fine-tune it with Axolotl with multi-turn conversations
 
 ## Getting started
 
-1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html). You need to install from main as Qwen3-Next is only on nightly or use our latest [Docker images](https://docs.axolotl.ai/docs/docker.html).
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
 
-    Here is an example of how to install from main for pip:
-
-```bash
-# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
-git clone https://github.com/axolotl-ai-cloud/axolotl.git
-cd axolotl
-
-pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
-pip3 install --no-build-isolation -e '.[flash-attn]'
-
-# Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
-python scripts/cutcrossentropy_install.py | sh
-```
-
-2. Install Qwen3-Next transformers commit
-```bash
-pip3 uninstall -y transformers && pip3 install "git+https://github.com/huggingface/transformers.git@b9282355bea846b54ed850a066901496b19da654"
-```
+2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage.
 
 3. Install FLA for improved performance
 ```bash
-pip3 uninstall -y causal-conv1d && pip3 install flash-linear-attention==0.3.2
+pip3 uninstall -y causal-conv1d && pip3 install flash-linear-attention==0.4.1
 ```
 
 4. Run the finetuning example:
@@ -38,7 +21,7 @@ pip3 uninstall -y causal-conv1d && pip3 install flash-linear-attention==0.3.2
 axolotl train examples/qwen3-next/qwen3-next-80b-a3b-qlora.yaml
 ```
 
-This config uses about 45.62 GiB VRAM.
+This config uses about ~47 GiB (no target experts) and ~71GiB (target experts) VRAM.
 
 Let us know how it goes. Happy finetuning! 🚀
 

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

@@ -9,6 +9,8 @@ plugins:
 load_in_8bit: false
 load_in_4bit: true
 
+quantize_moe_experts: true
+
 datasets:
   - path: fozziethebeat/alpaca_messages_2k_test
     type: chat_template
@@ -25,7 +27,7 @@ sample_packing: true
 
 lora_r: 16
 lora_alpha: 8
-lora_dropout: 0.05
+lora_dropout: 0
 lora_target_modules:
   - linear_attn.in_proj_ba
   - linear_attn.in_proj_qkvz
@@ -34,12 +36,19 @@ lora_target_modules:
   - shared_expert.down_proj
   - shared_expert.gate_proj
   - shared_expert_gate
-  - mlp.gate
   - q_proj
   - v_proj
   - k_proj
   - o_proj
 
+# lora_target_parameters:
+#   - mlp.experts.gate_up_proj
+#   - mlp.experts.down_proj
+
+lora_mlp_kernel: false
+lora_qkv_kernel: false
+lora_o_kernel: false
+
 wandb_project:
 wandb_entity:
 wandb_watch:

examples/qwen3.5/122b-a10b-moe-qlora.yaml

@@ -0,0 +1,71 @@
+base_model: Qwen/Qwen3.5-122B-A10B
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+strict: false
+
+chat_template: qwen3_5
+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
+
+load_in_4bit: true
+quantize_moe_experts: true
+adapter: qlora
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0
+lora_target_modules:
+  - q_proj
+  - k_proj
+  - v_proj
+  - o_proj
+
+#lora_target_parameters:
+#   - mlp.experts.gate_up_proj
+#   - mlp.experts.down_proj
+
+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
+
+
+lora_mlp_kernel: false
+lora_qkv_kernel: false
+lora_o_kernel: false
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/qwen3.5/27b-qlora.yaml

@@ -0,0 +1,72 @@
+base_model: Qwen/Qwen3.5-27B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+# Note: Qwen3.5 is an early-fusion VLM (image+text). This config fine-tunes
+# the text-only path. For multimodal (image+text) fine-tuning, add image
+# columns to your dataset following axolotl's multimodal dataset format.
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+strict: false
+
+chat_template: qwen3_5
+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
+
+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
+  # Uncomment below to also target the linear attention projections.
+  # These use separate in_proj_qkv / in_proj_z / out_proj (Qwen3.5-specific).
+  # - linear_attn.in_proj_qkv
+  # - linear_attn.in_proj_z
+  # - linear_attn.out_proj
+
+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: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/qwen3.5/35b-a3b-moe-qlora.yaml

@@ -0,0 +1,70 @@
+base_model: Qwen/Qwen3.5-35B-A3B
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+strict: false
+
+chat_template: qwen3_5
+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
+
+load_in_4bit: true
+quantize_moe_experts: true
+adapter: qlora
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0
+lora_target_modules:
+  - q_proj
+  - k_proj
+  - v_proj
+  - o_proj
+
+#lora_target_parameters:
+#   - mlp.experts.gate_up_proj
+#   - mlp.experts.down_proj
+
+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
+
+lora_mlp_kernel: false
+lora_qkv_kernel: false
+lora_o_kernel: false
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/qwen3.5/7b-lora-vision.yaml

@@ -0,0 +1,72 @@
+base_model: Qwen/Qwen3.5-7B
+processor_type: AutoProcessor
+
+# Qwen3.5-7B and above are early-fusion VLMs (Qwen3_5ForConditionalGeneration).
+# Vision and text tokens are processed together by the same transformer layers.
+# Note: Qwen3.5-2B is a text-only model — the smallest VLM is Qwen3.5-7B.
+
+# These 3 lines are required for vision/multimodal training
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+chat_template: qwen3_5
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+adapter: lora
+lora_model_dir:
+
+sequence_len: 8192
+pad_to_sequence_len: false
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+# Targets the language model attention and MLP layers.
+# Qwen3.5 is early-fusion: all layers (including those seeing vision tokens) share
+# the same transformer stack, so standard attention targets work for both modalities.
+lora_target_modules:
+  - q_proj
+  - k_proj
+  - v_proj
+  - o_proj
+  - down_proj
+  - up_proj
+  # Uncomment to also target the linear attention (GatedDeltaNet) projections:
+  # - linear_attn.in_proj_qkv
+  # - linear_attn.in_proj_z
+  # - linear_attn.out_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: true
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0

examples/qwen3.5/README.md

@@ -0,0 +1,61 @@
+# Finetune Qwen3.5 with Axolotl
+
+[Qwen3.5](https://huggingface.co/collections/Qwen/qwen35-68452f3bc6e4b7cfb4e1c803) is a hybrid architecture model series combining Gated DeltaNet linear attention with standard Transformer attention. Models from 7B onwards are early-fusion vision-language models (`Qwen3_5ForConditionalGeneration`), meaning vision and text tokens are processed through the same transformer stack. The 2B variant is text-only.
+
+Available configs:
+
+| Config | Model | Type |
+|---|---|---|
+| `27b-qlora.yaml` | Qwen3.5-27B | Dense VLM, text-only path |
+| `35b-a3b-moe-qlora.yaml` | Qwen3.5-35B-A3B | MoE, text-only path |
+| `122b-a10b-moe-qlora.yaml` | Qwen3.5-122B-A10B | MoE, text-only path |
+| `7b-lora-vision.yaml` | Qwen3.5-7B | Vision+text (multimodal) |
+
+## Getting started
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
+
+2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage.
+
+3. Install FLA for sample packing support with the Gated DeltaNet linear attention layers:
+```bash
+pip3 uninstall -y causal-conv1d && pip3 install flash-linear-attention==0.4.1
+```
+> FLA is required when `sample_packing: true`. Without it, training raises a `RuntimeError` on packed sequences. Vision configs use `sample_packing: false` so FLA is optional there.
+
+4. Run a finetuning example:
+
+```bash
+# Dense 27B text-only (QLoRA, ~47 GiB VRAM with sample packing)
+axolotl train examples/qwen3.5/27b-qlora.yaml
+
+# MoE 35B-A3B text-only (QLoRA)
+axolotl train examples/qwen3.5/35b-a3b-moe-qlora.yaml
+
+# MoE 122B-A10B text-only (QLoRA)
+axolotl train examples/qwen3.5/122b-a10b-moe-qlora.yaml
+
+# 7B vision+text (LoRA, multimodal dataset)
+axolotl train examples/qwen3.5/7b-lora-vision.yaml
+```
+
+### TIPS
+
+- For inference, you can experiment with `temperature: 0.7`, `top_p: 0.8`, `top_k: 20`, and `min_p: 0`.
+- You can run a full finetuning by removing `adapter: qlora` and `load_in_4bit: true`. See [Multi-GPU](#optimization-guides) below.
+- Read more on loading your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+- The dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+- For **multimodal** finetuning, set `processor_type: AutoProcessor`, `skip_prepare_dataset: true`, and `remove_unused_columns: false` as shown in `7b-lora-vision.yaml`.
+- The Gated DeltaNet linear attention layers (`linear_attn.*`) can optionally be added to `lora_target_modules` — they are commented out by default.
+
+## Optimization Guides
+
+- [Optimizations Guide](https://docs.axolotl.ai/docs/optimizations.html)
+
+## Related Resources
+
+- [Qwen3.5 Blog](https://qwenlm.github.io/blog/qwen3.5/)
+- [Axolotl Docs](https://docs.axolotl.ai)
+- [Axolotl Website](https://axolotl.ai)
+- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/trinity/README.md

@@ -8,13 +8,15 @@ This guide shows how to fine-tune it with Axolotl with multi-turn conversations
 
 1. Install Axolotl following the main from the [installation guide](https://docs.axolotl.ai/docs/installation.html#sec-edge-build).
 
-2. Run the finetuning example:
+2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage.
+
+3. Run the finetuning example:
 
     ```bash
     axolotl train examples/trinity/trinity-nano-preview-qlora.yaml
     ```
 
-This config uses about 24.9 GiB VRAM.
+This config uses about 24.9 GiB VRAM (w/o CCE).
 
 Let us know how it goes. Happy finetuning! 🚀
 
@@ -29,10 +31,6 @@ Let us know how it goes. Happy finetuning! 🚀
 
 Please check the [Optimizations doc](https://docs.axolotl.ai/docs/optimizations.html).
 
-## Limitations
-
-**Cut Cross Entropy (CCE)**: Currently not supported. We plan to include CCE support for Trinity in the near future.
-
 ## Related Resources
 
 - [Trinity Blog](https://www.arcee.ai/blog/the-trinity-manifesto)

examples/trinity/trinity-nano-preview-qlora.yaml

@@ -1,5 +1,4 @@
 base_model: arcee-ai/Trinity-Nano-Preview
-trust_remote_code: true
 revision_of_model: 2ee94b0
 
 # Automatically upload checkpoint and final model to HF

examples/voxtral/README.md

@@ -12,7 +12,7 @@ Thanks to the team at MistralAI for giving us early access to prepare for this r
 
 ```bash
 # Ensure you have Pytorch installed (Pytorch 2.6.0 min)
-pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install packaging==26.0 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 

pyproject.toml

@@ -1,5 +1,5 @@
 [build-system]
-requires = ["setuptools>=64", "wheel", "setuptools_scm>=8", "packaging==23.2"]
+requires = ["setuptools>=64", "wheel", "setuptools_scm>=8", "packaging==26.0"]
 build-backend = "setuptools.build_meta"
 
 [project]
@@ -24,6 +24,9 @@ Repository = "https://github.com/axolotl-ai-cloud/axolotl.git"
 py-modules = ["setuptools_axolotl_dynamic_dependencies"]
 include-package-data = true
 
+[tool.setuptools.dynamic]
+version = { file = "VERSION" }
+
 [tool.setuptools.cmdclass]
 build_py = "setuptools_axolotl_dynamic_dependencies.BuildPyCommand"
 
@@ -57,3 +60,6 @@ indent-style = "space"
 skip-magic-trailing-comma = false
 line-ending = "auto"
 docstring-code-format = false
+
+[tool.uv.extra-build-dependencies]
+axolotl = ["huggingface_hub"]

requirements.txt

@@ -2,25 +2,28 @@
 
 # START section of dependencies that don't install on Darwin/MacOS
 bitsandbytes==0.49.1
-triton>=3.0.0
+triton>=3.4.0
 mamba-ssm==1.2.0.post1
 xformers>=0.0.23.post1
-liger-kernel==0.6.4
+liger-kernel==0.7.0
 # END section
 
-packaging==23.2
-
-huggingface_hub>=0.36.0
+packaging==26.0
+huggingface_hub>=1.1.7
 peft>=0.18.1
 tokenizers>=0.22.1
-transformers==4.57.6
-accelerate==1.12.0
+transformers==5.3.0
+accelerate==1.13.0
 datasets==4.5.0
-deepspeed>=0.18.3
-trl==0.25.1
-hf_xet==1.2.0
-kernels==0.11.5
-trackio>=0.13.0
+deepspeed>=0.18.6,<0.19.0
+trl==0.29.0
+hf_xet==1.3.2
+kernels==0.12.2
+
+fla-core==0.4.1
+flash-linear-attention==0.4.1
+
+trackio>=0.16.1
 typing-extensions>=4.15.0
 
 optimum==1.16.2
@@ -63,7 +66,7 @@ langdetect==1.0.9
 immutabledict==4.2.0
 antlr4-python3-runtime==4.13.2
 
-torchao==0.13.0
+torchao==0.16.0
 openenv-core==0.1.0
 schedulefree==1.4.1
 
@@ -72,4 +75,4 @@ axolotl-contribs-mit==0.0.6
 # telemetry
 posthog==6.7.11
 
-mistral-common==1.8.6
+mistral-common==1.8.8

scripts/cutcrossentropy_install.py

@@ -29,5 +29,5 @@ UV_PREFIX = "uv " if USE_UV else ""
 
 print(
     UNINSTALL_PREFIX
-    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@318b7e2"'
+    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@e8ad129"'
 )

setup.py

@@ -1,6 +1,5 @@
 """setup.py for axolotl"""
 
-import ast
 import os
 import platform
 import re
@@ -27,6 +26,18 @@ def parse_requirements(extras_require_map):
     try:
         xformers_version = [req for req in _install_requires if "xformers" in req][0]
         install_xformers = platform.machine() != "aarch64"
+        if platform.machine() == "aarch64":
+            # skip on ARM64
+            skip_packages = [
+                "torchao",
+                "fla-core",
+                "flash-linear-attention",
+            ]
+            _install_requires = [
+                req
+                for req in _install_requires
+                if re.split(r"[>=<]", req)[0].strip() not in skip_packages
+            ]
         if "Darwin" in platform.system():
             # skip packages not compatible with OSX
             skip_packages = [
@@ -79,6 +90,11 @@ def parse_requirements(extras_require_map):
                 extras_require_map["vllm"] = ["vllm==0.11.1"]
                 if not install_xformers:
                     _install_requires.pop(_install_requires.index(xformers_version))
+                extras_require_map["vllm"] = ["vllm==0.13.0"]
+                if patch == 0:
+                    extras_require_map["vllm"] = ["vllm==0.13.0"]
+                else:
+                    extras_require_map["vllm"] = ["vllm==0.14.0"]
             elif (major, minor) >= (2, 8):
                 extras_require_map.pop("fbgemm-gpu")
                 extras_require_map["fbgemm-gpu"] = ["fbgemm-gpu-genai==1.3.0"]
@@ -130,15 +146,11 @@ def parse_requirements(extras_require_map):
 
 def get_package_version():
     with open(
-        Path(os.path.dirname(os.path.abspath(__file__)))
-        / "src"
-        / "axolotl"
-        / "__init__.py",
+        Path(os.path.dirname(os.path.abspath(__file__))) / "VERSION",
         "r",
         encoding="utf-8",
     ) as fin:
-        version_match = re.search(r"^__version__\s*=\s*(.*)$", fin.read(), re.MULTILINE)
-    version_ = ast.literal_eval(version_match.group(1))
+        version_ = fin.read().strip()
     return version_
 
 

src/axolotl/__init__.py

@@ -1,7 +1,11 @@
 """Axolotl - Train and fine-tune large language models"""
 
 import pkgutil
+from importlib.metadata import PackageNotFoundError, version
 
 __path__ = pkgutil.extend_path(__path__, __name__)  # Make this a namespace package
 
-__version__ = "0.13.1"
+try:
+    __version__ = version("axolotl")
+except PackageNotFoundError:
+    __version__ = "unknown"

src/axolotl/cli/__init__.py

@@ -5,6 +5,7 @@ import os
 from axolotl.logging_config import configure_logging
 
 os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
-os.environ.setdefault("HF_HUB_ENABLE_HF_TRANSFER", "1")
+os.environ.setdefault("HF_XET_HIGH_PERFORMANCE", "1")
+os.environ.setdefault("TRL_EXPERIMENTAL_SILENCE", "1")
 
 configure_logging()

src/axolotl/cli/checks.py

@@ -44,7 +44,7 @@ def check_user_token() -> bool:
         return bool(user_info)
     except LocalTokenNotFoundError:
         LOG.warning(
-            "Error verifying HuggingFace token. Remember to log in using `huggingface-cli login` and get your access token from https://huggingface.co/settings/tokens if you want to use gated models or datasets."
+            "Error verifying HuggingFace token. Remember to log in using `hf auth login` and get your access token from https://huggingface.co/settings/tokens if you want to use gated models or datasets."
         )
         return False
     except HTTPError:

src/axolotl/cli/config.py

@@ -5,7 +5,7 @@ import os
 import tempfile
 from pathlib import Path
 from tempfile import NamedTemporaryFile
-from typing import Union
+from typing import Any, Optional, Union
 from urllib.parse import urlparse
 
 import requests
@@ -32,6 +32,63 @@ from axolotl.utils.wandb_ import setup_wandb_env_vars
 
 LOG = get_logger(__name__)
 
+
+def _coerce_value(value: Any, existing: Optional[Any] = None) -> Any:
+    """Coerce a string CLI value to its most likely Python type.
+
+    If an existing value is present in the config, its type is used to guide
+    casting.  Otherwise, YAML-style inference is applied: booleans, ints,
+    floats, and None literals are recognised automatically.
+
+    Args:
+        value: The raw value (typically a string from the CLI).
+        existing: An optional existing config value whose type guides coercion.
+
+    Returns:
+        The value cast to the inferred or expected type.
+    """
+    if not isinstance(value, str):
+        return value
+
+    # If the config already has a typed value, cast to match
+    if existing is not None:
+        if isinstance(existing, bool):
+            return value.lower() in ("true", "1", "yes")
+        if isinstance(existing, int):
+            try:
+                return int(value)
+            except (ValueError, TypeError):
+                return value
+        if isinstance(existing, float):
+            try:
+                return float(value)
+            except (ValueError, TypeError):
+                return value
+        # For other types (str, list, dict, etc.), return as-is
+        return value
+
+    # No existing value -- use YAML-style inference
+    lower = value.lower()
+    if lower in ("true", "yes"):
+        return True
+    if lower in ("false", "no"):
+        return False
+    if lower in ("null", "none", "~"):
+        return None
+
+    # Try int then float
+    try:
+        return int(value)
+    except ValueError:
+        pass
+    try:
+        return float(value)
+    except ValueError:
+        pass
+
+    return value
+
+
 API_KEY_FIELDS = {"comet_api_key"}
 
 TELEMETRY_MANAGER = TelemetryManager.get_instance()
@@ -208,13 +265,37 @@ def load_cfg(
     # If there are any options passed in the cli, if it is something that seems valid
     # from the yaml, then overwrite the value
     cfg_keys = cfg.keys()
+
+    # Separate nested (dot-notation) kwargs from flat kwargs
+    nested_kwargs: dict[str, dict[str, Any]] = {}
+    flat_kwargs: dict[str, Any] = {}
     for key, value in kwargs.items():
+        if "__" in key:
+            parent, child = key.split("__", 1)
+            nested_kwargs.setdefault(parent, {})[child] = value
+        else:
+            flat_kwargs[key] = value
+
+    # Apply flat kwargs
+    for key, value in flat_kwargs.items():
         # If not strict, allow writing to cfg even if it's not in the yml already
         if key in cfg_keys or not cfg.strict:
-            if isinstance(cfg[key], bool):
-                cfg[key] = bool(value)
-            else:
-                cfg[key] = value
+            cfg[key] = _coerce_value(value, cfg.get(key))
+
+    # Apply nested kwargs (e.g., trl__beta -> cfg.trl.beta)
+    for parent, children in nested_kwargs.items():
+        if parent not in cfg_keys and cfg.strict:
+            continue
+        if cfg[parent] is None:
+            cfg[parent] = {}
+        if not isinstance(cfg[parent], dict):
+            LOG.warning(
+                "Overwriting non-dict value for '%s' with nested CLI overrides", parent
+            )
+            cfg[parent] = {}
+        for child_key, child_value in children.items():
+            existing_child = cfg[parent].get(child_key)
+            cfg[parent][child_key] = _coerce_value(child_value, existing_child)
 
     try:
         device_props = torch.cuda.get_device_properties("cuda")

src/axolotl/cli/merge_lora.py

@@ -24,10 +24,9 @@ def do_merge_lora(*, cfg: DictDefault) -> None:
         cfg: Dictionary mapping `axolotl` config keys to values.
     """
     model, tokenizer, processor = load_model_and_tokenizer(cfg=cfg)
-    safe_serialization = cfg.save_safetensors is True
 
     LOG.info("Running merge of LoRA with base model...")
-    model = model.merge_and_unload(progressbar=True)
+    model = model.merge_and_unload(progressbar=True, safe_merge=True)
     try:
         model.to(dtype=cfg.torch_dtype)
     except ValueError as e:
@@ -42,7 +41,6 @@ def do_merge_lora(*, cfg: DictDefault) -> None:
         LOG.info(f"Saving merged model to: {str(Path(cfg.output_dir) / 'merged')}...")
         model.save_pretrained(
             str(Path(cfg.output_dir) / "merged"),
-            safe_serialization=safe_serialization,
             progressbar=True,
         )
         tokenizer.save_pretrained(
@@ -73,6 +71,7 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs) -> None:
         merge_lora=True,
         load_in_8bit=False,
         load_in_4bit=False,
+        quantize_moe_experts=False,
         flash_attention=False,
         context_parallel_size=None,
         deepspeed=None,

src/axolotl/cli/merge_sharded_fsdp_weights.py

@@ -14,8 +14,6 @@ from accelerate import PartialState
 from accelerate.utils import (
     SAFE_WEIGHTS_INDEX_NAME,
     SAFE_WEIGHTS_NAME,
-    WEIGHTS_INDEX_NAME,
-    WEIGHTS_NAME,
     is_torch_version,
 )
 from huggingface_hub import split_torch_state_dict_into_shards
@@ -40,17 +38,15 @@ class BFloat16CastPlanner(_EmptyStateDictLoadPlanner):
 def _distributed_checkpoint_to_merged_weights(
     checkpoint_dir: Union[str, Path],
     save_path: str,
-    safe_serialization: bool = False,
     max_shard_size: str = "5GB",
 ) -> Path:
     """
     Passthrough to `torch.distributed.checkpoint.format_utils.dcp_to_torch_save`. Will
-    save under `save_path` as either `model.safetensors` or `pytorch_model.bin`.
+    save under `save_path` as `model.safetensors`.
 
     Args:
         checkpoint_dir: Directory where distributed checkpoint is saved.
         save_path: Path to save model to.
-        safe_serialization: Whether to save in safetensors format.
         max_shard_size: Max size of model shards to save.
 
     Returns:
@@ -76,11 +72,7 @@ def _distributed_checkpoint_to_merged_weights(
         if isinstance(value, torch.Tensor) and value.dtype != torch.bfloat16:
             state_dict[key] = value.to(torch.bfloat16)
 
-    weights_name = SAFE_WEIGHTS_NAME if safe_serialization else WEIGHTS_NAME
-
-    filename_pattern = weights_name.replace(".bin", "{suffix}.bin").replace(
-        ".safetensors", "{suffix}.safetensors"
-    )
+    filename_pattern = SAFE_WEIGHTS_NAME.replace(".safetensors", "{suffix}.safetensors")
     state_dict_split = split_torch_state_dict_into_shards(
         state_dict, filename_pattern=filename_pattern, max_shard_size=max_shard_size
     )
@@ -98,19 +90,12 @@ def _distributed_checkpoint_to_merged_weights(
 
     for shard_file, tensors in filename_to_tensors:
         shard = {tensor: state_dict[tensor] for tensor in tensors}
-
-        if safe_serialization:
-            safe_save_file(
-                shard, os.path.join(save_path_, shard_file), metadata={"format": "pt"}
-            )
-        else:
-            torch.save(shard, os.path.join(save_path_, shard_file))
+        safe_save_file(
+            shard, os.path.join(save_path_, shard_file), metadata={"format": "pt"}
+        )
 
     if index is not None:
-        save_index_file = (
-            SAFE_WEIGHTS_INDEX_NAME if safe_serialization else WEIGHTS_INDEX_NAME
-        )
-        save_index_file = os.path.join(save_path_, save_index_file)
+        save_index_file = os.path.join(save_path_, SAFE_WEIGHTS_INDEX_NAME)
         # Save the index as well
         with open(save_index_file, "w", encoding="utf-8") as fout:
             content = json.dumps(index, indent=2, sort_keys=True) + "\n"
@@ -123,13 +108,11 @@ def _distributed_checkpoint_to_merged_weights(
 def merge_fsdp_weights(
     checkpoint_dir: str,
     output_path: str,
-    safe_serialization: bool = False,
     remove_checkpoint_dir: bool = False,
 ):
     """
     Merge the weights from sharded FSDP model checkpoints into a single combined checkpoint. Should be used if
-    `SHARDED_STATE_DICT` was used for the model. Weights will be saved to `{output_path}/model.safetensors` if
-    `safe_serialization` else `pytorch_model.bin`.
+    `SHARDED_STATE_DICT` was used for the model. Weights will be saved to `{output_path}/model.safetensors`.
 
     Note: this is a CPU-bound process.
 
@@ -138,8 +121,6 @@ def merge_fsdp_weights(
             The directory containing the FSDP checkpoints (can be either the model or optimizer).
         output_path (`str`):
             The path to save the merged checkpoint.
-        safe_serialization (`bool`, *optional*, defaults to `True`):
-            Whether to save the merged weights with safetensors (recommended).
         remove_checkpoint_dir (`bool`, *optional*, defaults to `False`):
             Whether to remove the checkpoint directory after merging.
 
@@ -177,7 +158,7 @@ def merge_fsdp_weights(
     if state.is_main_process:
         LOG.info(f"Merging FSDP weights from {checkpoint_dir_}")
         save_path = _distributed_checkpoint_to_merged_weights(
-            checkpoint_dir_, output_path, safe_serialization
+            checkpoint_dir_, output_path
         )
         LOG.info(f"Successfully merged FSDP weights and saved to {save_path}")
         if remove_checkpoint_dir:
@@ -210,7 +191,6 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
     merge_fsdp_weights(
         checkpoint_dir=str(fsdp_dir),
         output_path=output_path,
-        safe_serialization=True,
     )
     state = PartialState()
     state.wait_for_everyone()

src/axolotl/cli/quantize.py

@@ -102,12 +102,10 @@ def do_quantize(
     LOG.info(f"Saving quantized model to: {str(Path(output_dir) / 'quantized')}.")
     model.save_pretrained(
         str(Path(output_dir) / "quantized"),
-        safe_serialization=False,
         progressbar=True,
     )
     tokenizer.save_pretrained(
         str(Path(output_dir) / "quantized"),
-        safe_serialization=False,
         progressbar=True,
         save_jinja_files=cfg.tokenizer_save_jinja_files,
     )
@@ -121,7 +119,7 @@ def do_quantize(
             hub_model_id.rstrip("-")
             + f"-{quantization_config_to_str[type(quantization_config)]}"
         )
-        model.push_to_hub(hub_model_id, safe_serialization=False)
+        model.push_to_hub(hub_model_id)
         tokenizer.push_to_hub(hub_model_id)
         if processor:
             processor.push_to_hub(hub_model_id)

src/axolotl/cli/utils/args.py

@@ -2,7 +2,7 @@
 
 import dataclasses
 from functools import wraps
-from types import NoneType
+from types import NoneType, UnionType
 from typing import Any, Callable, Type, Union, get_args, get_origin
 
 import click
@@ -20,7 +20,8 @@ def _strip_optional_type(field_type: type | str | None):
         If the input type is `Union[T, None]` or `Optional[T]`, returns `T`. Otherwise
             returns the input type unchanged.
     """
-    if get_origin(field_type) is Union and type(None) in get_args(field_type):
+    is_union = get_origin(field_type) is Union or isinstance(field_type, UnionType)
+    if is_union and type(None) in get_args(field_type):
         field_type = next(
             t for t in get_args(field_type) if not isinstance(t, NoneType)
         )
@@ -87,10 +88,70 @@ def add_options_from_dataclass(config_class: Type[Any]) -> Callable:
     return decorator
 
 
+def _is_pydantic_model(field_type: type) -> bool:
+    """Check if a type is a Pydantic BaseModel subclass."""
+    try:
+        return isinstance(field_type, type) and issubclass(field_type, BaseModel)
+    except TypeError:
+        return False
+
+
+def _get_field_description(field) -> str | None:
+    """Get description from a Pydantic field, checking both .description and json_schema_extra."""
+    if field.description:
+        return field.description
+    if field.json_schema_extra and isinstance(field.json_schema_extra, dict):
+        return field.json_schema_extra.get("description")
+    return None
+
+
+def _add_nested_model_options(
+    function: Callable, parent_name: str, model_class: Type[BaseModel]
+) -> Callable:
+    """
+    Add Click options for all fields of a nested Pydantic model using dot-notation.
+
+    Note: Only single-level nesting is supported (e.g., ``--trl.beta``).
+    Deeper nesting (e.g., ``--trl.scheduler.warmup``) is not handled.
+
+    Args:
+        function: Click command function to add options to.
+        parent_name: Parent field name (e.g., "trl").
+        model_class: Nested Pydantic model class.
+
+    Returns:
+        Function with added Click options.
+    """
+    for sub_name, sub_field in reversed(model_class.model_fields.items()):
+        sub_type = _strip_optional_type(sub_field.annotation)
+        # Use dot notation: --parent.sub_field
+        cli_name = f"{parent_name}.{sub_name}".replace("_", "-")
+        # The kwarg name uses double-underscore as separator
+        param_name = f"{parent_name}__{sub_name}"
+        description = _get_field_description(sub_field)
+
+        if sub_type is bool:
+            option_name = f"--{cli_name}/--no-{cli_name}"
+            function = click.option(
+                option_name, param_name, default=None, help=description
+            )(function)
+        else:
+            option_name = f"--{cli_name}"
+            click_type = {str: str, int: int, float: float}.get(sub_type)
+            function = click.option(
+                option_name, param_name, default=None, type=click_type, help=description
+            )(function)
+
+    return function
+
+
 def add_options_from_config(config_class: Type[BaseModel]) -> Callable:
     """
     Create Click options from the fields of a Pydantic model.
 
+    For fields whose type is itself a Pydantic BaseModel, dot-notation CLI options are
+    generated for each sub-field (e.g., ``--trl.beta=0.1``).
+
     Args:
         config_class: PyDantic model with fields to parse from the CLI
 
@@ -103,6 +164,11 @@ def add_options_from_config(config_class: Type[BaseModel]) -> Callable:
         for name, field in reversed(config_class.model_fields.items()):
             field_type = _strip_optional_type(field.annotation)
 
+            # Handle nested Pydantic models with dot-notation options
+            if _is_pydantic_model(field_type):
+                function = _add_nested_model_options(function, name, field_type)
+                continue
+
             if field_type is bool:
                 field_name = name.replace("_", "-")
                 option_name = f"--{field_name}/--no-{field_name}"

src/axolotl/common/architectures.py

@@ -12,10 +12,14 @@ MOE_ARCH_BLOCK = {
     "mixtral": "MixtralSparseMoeBlock",
     "qwen2_moe": "Qwen2MoeSparseMoeBlock",
     "qwen3_moe": "Qwen3MoeSparseMoeBlock",
+    "qwen3_5_moe": "Qwen3_5MoeSparseMoeBlock",
     "qwen3_vl_moe": "Qwen3VLMoeTextSparseMoeBlock",
     "deepseek_v2": "DeepseekV2MoE",
     "deepseek_v3": "DeepseekV3MoE",
     "gpt_oss": "GptOssDecoderLayer",
     "lfm2_moe": "Lfm2MoeSparseMoeBlock",
     "afmoe": "AfmoeMoE",
+    "glm4_moe": "Glm4MoeDecoderLayer",
+    "glm4_moe_lite": "Glm4MoeLiteDecoderLayer",
+    "glm_moe_dsa": "GlmMoeDsaDecoderLayer",
 }

src/axolotl/core/builders/base.py

@@ -216,7 +216,7 @@ class TrainerBuilderBase(abc.ABC):
     def _configure_warmup_and_logging(
         self, total_num_steps: int, training_args_kwargs: dict
     ):
-        warmup_steps = 0
+        warmup_steps: int | float = 0
         warmup_ratio = 0.0
         if self.cfg.warmup_steps is not None:
             warmup_steps = self.cfg.warmup_steps
@@ -230,6 +230,10 @@ class TrainerBuilderBase(abc.ABC):
         else:
             warmup_ratio = 0.03
 
+        # transformers v5
+        if warmup_ratio > 0.0 and warmup_steps == 0:
+            warmup_steps = warmup_ratio
+
         if warmup_steps == 1:
             warmup_steps = 2
 
@@ -242,7 +246,6 @@ class TrainerBuilderBase(abc.ABC):
                 else max(min(int(0.005 * total_num_steps), 10), 1)
             )
 
-        training_args_kwargs["warmup_ratio"] = warmup_ratio
         training_args_kwargs["warmup_steps"] = warmup_steps
 
     def _configure_precision_settings(self, training_args_kwargs: dict):
@@ -406,6 +409,9 @@ class TrainerBuilderBase(abc.ABC):
             if self.cfg.hub_strategy:
                 training_args_kwargs["hub_strategy"] = self.cfg.hub_strategy
 
+            if self.cfg.hub_revision:
+                training_args_kwargs["hub_revision"] = self.cfg.hub_revision
+
     def _configure_save_and_eval_strategy(self, training_args_kwargs: dict):
         # save_strategy and save_steps
         if self.cfg.save_steps:
@@ -530,9 +536,7 @@ class TrainerBuilderBase(abc.ABC):
             "loraplus_lr_ratio",
             "loraplus_lr_embedding",
             "output_dir",
-            "save_safetensors",
             "save_only_model",
-            "include_tokens_per_second",
             "weight_decay",
             "seed",
             "dion_momentum",
@@ -545,6 +549,7 @@ class TrainerBuilderBase(abc.ABC):
 
         arg_map = {
             "dion_learning_rate": "dion_lr",
+            "include_num_input_tokens_seen": "include_tokens_per_second",
         }
         for kwarg, cfg_arg in arg_map.items():
             if hasattr(self.cfg, cfg_arg) and getattr(self.cfg, cfg_arg) is not None:

src/axolotl/core/builders/causal.py

@@ -122,6 +122,12 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             ColabCallback = colab_inference_post_train_callback(trainer)
             callbacks.append(ColabCallback(self.cfg))
 
+        if getattr(self.cfg, "generate_samples", False):
+            from axolotl.utils.callbacks.generation import SFTGenerationCallback
+
+            callbacks.append(SFTGenerationCallback(trainer))
+            LOG.info("SFT sample generation enabled")
+
         callbacks.extend(super().get_post_trainer_create_callbacks(trainer=trainer))
         return callbacks
 
@@ -246,7 +252,8 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             ddp_find_unused_parameters
         )
 
-        training_arguments_kwargs["group_by_length"] = self.cfg.group_by_length
+        if self.cfg.group_by_length:
+            training_arguments_kwargs["train_sampling_strategy"] = "group_by_length"
         training_arguments_kwargs["curriculum_sampling"] = self.cfg.curriculum_sampling
 
         training_arguments_kwargs["sample_packing"] = bool(self.cfg.sample_packing)
@@ -373,6 +380,18 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             # https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html
             data_collator_kwargs["pad_to_multiple_of"] = multiple
 
+        if self.cfg.use_eaft:
+            from functools import partial
+
+            from axolotl.monkeypatch.loss.eaft import eaft_loss
+
+            configured_eaft_loss = partial(
+                eaft_loss,
+                alpha=self.cfg.eaft_alpha if self.cfg.eaft_alpha is not None else 1.0,
+                k=self.cfg.eaft_k if self.cfg.eaft_k is not None else 20,
+            )
+            trainer_kwargs["compute_loss_func"] = configured_eaft_loss
+
         trainer_cls = self._get_trainer_cls()
 
         trainer_kwargs, trainer_cls = self.hook_pre_create_trainer(
@@ -437,7 +456,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 or self.cfg.micro_batch_size > 1
             ):
                 return DataCollatorForSeq2Seq(self.tokenizer, **kwargs)
-            if not (self.cfg.sample_packing and self.cfg.pretrain_multipack_attn):
+            if not (self.cfg.sample_packing and self.cfg.pretrain_multipack_attn) or (
+                self.cfg.micro_batch_size == 1 and is_eval is False
+            ):
                 return None
 
         if self.cfg.model_config_type == "mamba":

src/axolotl/core/builders/rl.py

@@ -11,7 +11,6 @@ from axolotl.core.trainers import (
 )
 from axolotl.core.trainers.dpo import DPOStrategy
 from axolotl.core.trainers.dpo.args import AxolotlDPOConfig
-from axolotl.core.trainers.grpo import GRPOStrategy
 from axolotl.integrations.base import PluginManager
 from axolotl.loaders.utils import ensure_dtype
 from axolotl.utils.callbacks.qat import QATCallback
@@ -52,12 +51,13 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         trainer_cls = None
         trainer_cls_args = [self.model]
 
-        if self.cfg.rl is RLType.GRPO:
+        if self.cfg.rl in {RLType.GRPO, RLType.GDPO}:
+            from axolotl.core.trainers.grpo import GRPOStrategy
+
             trainer_cls = GRPOStrategy.get_trainer_class(
                 sequence_parallel=self.cfg.context_parallel_size > 1
             )
             trainer_cls_args.extend(GRPOStrategy.set_trainer_args(self.cfg))
-
             trainer_kwargs.update(GRPOStrategy.set_trainer_kwargs(self.cfg))
 
         elif self.cfg.rl in [RLType.DPO, RLType.IPO]:
@@ -120,11 +120,6 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         if self.cfg.use_wandb:
             training_args_kwargs["run_name"] = self.cfg.wandb_name
 
-        if self.cfg.max_prompt_len:
-            training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
-        else:
-            training_args_kwargs["max_prompt_length"] = self.cfg.sequence_len
-
         training_args_cls = None
         blocklist_args_kwargs = []
         if self.cfg.rl is RLType.SIMPO:
@@ -134,19 +129,17 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             if self.cfg.cpo_alpha is not None:
                 training_args_kwargs["cpo_alpha"] = self.cfg.cpo_alpha
 
-            # Handle when max_prompt_length == max_length from defaults
-            # CPOTrainer requires strictly less than
-            if (
-                training_args_kwargs["max_prompt_length"]
-                == training_args_kwargs["max_length"]
-            ):
-                training_args_kwargs["max_prompt_length"] -= 1
+            blocklist_args_kwargs.append("max_prompt_length")
 
         elif self.cfg.rl is RLType.ORPO:
             training_args_cls = AxolotlORPOConfig
 
+            blocklist_args_kwargs.append("max_prompt_length")
+
         elif self.cfg.rl is RLType.KTO:
             training_args_cls = AxolotlKTOConfig
+            # KTOConfig in TRL >= 0.27.0 no longer accepts max_prompt_length
+            blocklist_args_kwargs.append("max_prompt_length")
 
             training_args_kwargs["desirable_weight"] = (
                 self.cfg.kto_desirable_weight or 1.0
@@ -155,10 +148,16 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
                 self.cfg.kto_undesirable_weight or 1.0
             )
 
-        elif self.cfg.rl is RLType.GRPO:
+        elif self.cfg.rl in {RLType.GRPO, RLType.GDPO}:
+            from axolotl.core.trainers.grpo import GRPOStrategy
+
             training_args_cls = GRPOStrategy.get_training_args_class()
             training_args_kwargs.update(GRPOStrategy.set_training_args_kwargs(self.cfg))
             blocklist_args_kwargs = GRPOStrategy.get_blocklist_args_kwargs()
+            if self.cfg.rl is RLType.GDPO:
+                training_args_kwargs.setdefault(
+                    "multi_objective_aggregation", "normalize_then_sum"
+                )
 
         elif self.cfg.rl in [RLType.DPO, RLType.IPO]:
             training_args_cls = AxolotlDPOConfig

src/axolotl/core/trainers/base.py

@@ -25,8 +25,8 @@ from torch.utils.data import (
 from transformers import PreTrainedModel, Trainer
 from transformers.trainer import TRAINING_ARGS_NAME
 from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR, has_length, seed_worker
-from transformers.utils import SAFE_WEIGHTS_NAME, WEIGHTS_NAME, is_peft_available
-from trl.trainer.utils import pad_to_length
+from transformers.utils import SAFE_WEIGHTS_NAME, is_peft_available
+from trl.experimental.utils import pad_to_length
 from typing_extensions import override
 
 from axolotl.core.trainers.mixins import (
@@ -719,6 +719,20 @@ class AxolotlTrainer(
         output_dir = output_dir if output_dir is not None else self.args.output_dir
         os.makedirs(output_dir, exist_ok=True)
         LOG.info(f"Saving model checkpoint to {output_dir}")
+
+        # fix for Context Parallel save: CP eval invalidates tensor storage
+        # pointers, so clone to CPU to get fresh valid storage for safetensors
+        if (
+            state_dict is not None
+            and self.axolotl_cfg
+            and self.axolotl_cfg.context_parallel_size
+            and self.axolotl_cfg.context_parallel_size > 1
+        ):
+            state_dict = {
+                k: v.detach().cpu() if isinstance(v, torch.Tensor) else v
+                for k, v in state_dict.items()
+            }
+
         supported_classes = (
             (PreTrainedModel,)
             if not is_peft_available()
@@ -729,6 +743,7 @@ class AxolotlTrainer(
         if not isinstance(self.model, supported_classes):
             if state_dict is None:
                 state_dict = self.model.state_dict()
+
             if isinstance(
                 self.accelerator.unwrap_model(self.model, keep_torch_compile=False),
                 supported_classes,
@@ -738,43 +753,35 @@ class AxolotlTrainer(
                 ).save_pretrained(
                     output_dir,
                     state_dict=state_dict,
-                    safe_serialization=self.args.save_safetensors,
+                    is_main_process=self.accelerator.is_main_process,
                 )
             else:
                 LOG.info(
                     "Trainer.model is not a `PreTrainedModel`, only saving its state dict."
                 )
-                if self.args.save_safetensors:
-                    safetensors.torch.save_file(
-                        state_dict,
-                        os.path.join(output_dir, SAFE_WEIGHTS_NAME),
-                        metadata={"format": "pt"},
-                    )
-                else:
-                    torch.save(state_dict, os.path.join(output_dir, WEIGHTS_NAME))
+                safetensors.torch.save_file(
+                    state_dict,
+                    os.path.join(output_dir, SAFE_WEIGHTS_NAME),
+                    metadata={"format": "pt"},
+                )
         else:
             self.model.save_pretrained(
                 output_dir,
                 state_dict=state_dict,
-                safe_serialization=self.args.save_safetensors,
                 is_main_process=self.accelerator.is_main_process,
             )
 
-            if self.processing_class is not None:
-                self.processing_class.save_pretrained(output_dir)
-            elif (
-                self.data_collator is not None
-                and hasattr(self.data_collator, "tokenizer")
-                and self.data_collator.tokenizer is not None
-            ):
-                LOG.info(
-                    "Saving Trainer.data_collator.tokenizer by default as Trainer.processing_class is `None`"
-                )
-                save_jinja_files = True
-                if self.axolotl_cfg:
-                    save_jinja_files = self.axolotl_cfg.tokenizer_save_jinja_files
-                self.data_collator.tokenizer.save_pretrained(
-                    output_dir, save_jinja_files=save_jinja_files
-                )
-            # Good practice: save your training arguments together with the trained model
-            torch.save(self.args, os.path.join(output_dir, TRAINING_ARGS_NAME))
+        if self.processing_class is not None:
+            self.processing_class.save_pretrained(output_dir)
+        elif (
+            self.data_collator is not None
+            and hasattr(self.data_collator, "tokenizer")
+            and self.data_collator.tokenizer is not None
+        ):
+            LOG.info(
+                "Saving Trainer.data_collator.tokenizer by default as Trainer.processing_class is `None`"
+            )
+            self.data_collator.tokenizer.save_pretrained(output_dir)
+
+        # Good practice: save your training arguments together with the trained model
+        torch.save(self.args, os.path.join(output_dir, TRAINING_ARGS_NAME))

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

@@ -25,17 +25,13 @@ class DPOStrategy:
         # Label smoothing is not compatible with IPO
         if cfg.rl is RLType.DPO and cfg.dpo_label_smoothing:
             training_args_kwargs["label_smoothing"] = cfg.dpo_label_smoothing
-        training_args_kwargs["max_completion_length"] = None
         training_args_kwargs["max_length"] = cfg.sequence_len
-        training_args_kwargs["generate_during_eval"] = cfg.dpo_generate_during_eval
         if cfg.dpo_use_weighting is not None:
             training_args_kwargs["use_weighting"] = cfg.dpo_use_weighting
         if cfg.dpo_padding_free is not None:
             training_args_kwargs["padding_free"] = cfg.dpo_padding_free
         if cfg.dpo_norm_loss is not None:
             training_args_kwargs["dpo_norm_loss"] = cfg.dpo_norm_loss
-        if cfg.dpo_use_logits_to_keep is not None:
-            training_args_kwargs["use_logits_to_keep"] = cfg.dpo_use_logits_to_keep
         if cfg.dpo_use_liger_kernel is not None:
             training_args_kwargs["use_liger_kernel"] = cfg.dpo_use_liger_kernel
         return training_args_kwargs

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

@@ -57,16 +57,18 @@ class AxolotlDPOTrainer(
     def tokenize_row(
         features,
         processing_class,
-        max_prompt_length,
-        max_completion_length,
-        add_special_tokens,
+        max_prompt_length: int | None = None,
+        max_completion_length: int | None = None,
+        add_special_tokens: bool = True,
+        is_chat: bool = False,
     ) -> Dict:
         res = DPOTrainer.tokenize_row(
             features,
             processing_class,
-            max_prompt_length,
-            max_completion_length,
-            add_special_tokens,
+            max_prompt_length=max_prompt_length,
+            max_completion_length=max_completion_length,
+            add_special_tokens=add_special_tokens,
+            is_chat=is_chat,
         )
         # fix when the tokenizer doesn't have a bos_token_id, e.g. Qwen
         if processing_class.bos_token is None and res["prompt_input_ids"][0] is None:
@@ -101,10 +103,10 @@ class AxolotlDPOTrainer(
     ) -> dict[str, torch.Tensor]:
         if self.args.dpo_norm_loss:
             # fmt: off
-            loss_type: str = self.loss_type  # type: ignore[has-type]
+            loss_type: list[str] = self.loss_type  # type: ignore[has-type]
             # fmt: on
             # concatenated_forward handles avg token logprob for ipo case already
-            self.loss_type = "ipo"
+            self.loss_type = ["ipo"]
             res = super().concatenated_forward(model, batch, is_ref_model=is_ref_model)
             self.loss_type = loss_type
             return res

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

@@ -126,8 +126,10 @@ class GRPOStrategy:
         if trl.use_liger_loss is not None:
             grpo_args_kwargs["use_liger_loss"] = trl.use_liger_loss
 
-        if trl.rollout_func:
-            grpo_args_kwargs["rollout_func"] = cls.get_rollout_func(trl.rollout_func)
+        if trl.multi_objective_aggregation is not None:
+            grpo_args_kwargs["multi_objective_aggregation"] = (
+                trl.multi_objective_aggregation
+            )
 
         return grpo_args_kwargs
 
@@ -149,6 +151,8 @@ class GRPOStrategy:
             trainer_kwargs["reward_processing_classes"] = (
                 cfg.trl.reward_processing_classes
             )
+        if cfg.trl and cfg.trl.rollout_func:
+            trainer_kwargs["rollout_func"] = cls.get_rollout_func(cfg.trl.rollout_func)
 
         return trainer_kwargs
 
@@ -159,7 +163,12 @@ class GRPOStrategy:
 
     @classmethod
     def get_blocklist_args_kwargs(cls) -> list[str]:
-        return ["dataset_num_proc", "max_length", "include_tokens_per_second"]
+        return [
+            "dataset_num_proc",
+            "max_length",
+            "include_tokens_per_second",
+            "max_prompt_length",
+        ]
 
     @classmethod
     def get_reward_func(cls, reward_func_fqn: str) -> RewardFunc:

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

@@ -104,7 +104,7 @@ class OptimizerMixin(Trainer):
 
         return optimizer_grouped_parameters
 
-    def create_optimizer(self):
+    def create_optimizer(self, model=None):
         if (
             self.args.loraplus_lr_ratio is None
             and self.args.embedding_lr_scale is None
@@ -112,9 +112,9 @@ class OptimizerMixin(Trainer):
             and self.args.lr_groups is None
             and self.optimizer_cls_and_kwargs is None
         ):
-            return super().create_optimizer()
+            return super().create_optimizer(model=model)
 
-        opt_model = self.model_wrapped if is_sagemaker_mp_enabled() else self.model
+        opt_model = self.model if model is None else model
 
         if (
             not self.optimizer

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

@@ -25,7 +25,7 @@ class SchedulerMixin(Trainer):
     args = None  # type: "AxolotlTrainingArguments"  # type: ignore[name-defined]
 
     def create_scheduler(
-        self, num_training_steps: int, optimizer: torch.optim.Optimizer = None
+        self, num_training_steps: int, optimizer: None | torch.optim.Optimizer = None
     ) -> LRScheduler:
         """
         Set up the scheduler. The optimizer of the trainer must have been set up either before this method is called or
@@ -45,6 +45,13 @@ class SchedulerMixin(Trainer):
             and self.args.cosine_min_lr_ratio is not None
         )
 
+        if optimizer is None:
+            if self.optimizer is None:
+                raise ValueError(
+                    "Optimizer must be set before calling create_scheduler or passed as an argument."
+                )
+            optimizer = self.optimizer
+
         # fmt: off
         if self.lr_scheduler is None:  # type: ignore
             # fmt: on

src/axolotl/core/trainers/trl.py

@@ -1,12 +1,10 @@
 """Module for TRL RL trainers"""
 
-from trl import (
-    CPOTrainer,
-    KTOTrainer,
-    ORPOTrainer,
-    PRMTrainer,
-    RewardTrainer,
-)
+from trl import RewardTrainer
+from trl.experimental.cpo import CPOTrainer
+from trl.experimental.kto import KTOTrainer
+from trl.experimental.orpo import ORPOTrainer
+from trl.experimental.prm import PRMTrainer
 
 from axolotl.core.trainers.mixins import DistributedParallelMixin, RngLoaderMixin
 from axolotl.core.trainers.mixins.optimizer import OptimizerInitMixin, OptimizerMixin

src/axolotl/core/training_args.py

@@ -8,7 +8,11 @@ from dataclasses import dataclass, field
 from typing import Optional, Type
 
 from transformers import TrainingArguments
-from trl import CPOConfig, KTOConfig, ORPOConfig, PRMConfig, RewardConfig
+from trl import RewardConfig
+from trl.experimental.cpo import CPOConfig
+from trl.experimental.kto import KTOConfig
+from trl.experimental.orpo import ORPOConfig
+from trl.experimental.prm import PRMConfig
 
 from axolotl.integrations.config import merge_training_args
 

src/axolotl/integrations/cut_cross_entropy/README.md

@@ -19,7 +19,7 @@ python scripts/cutcrossentropy_install.py | sh
 
 - If you are installing from pip
 ```bash
-pip3 uninstall -y cut-cross-entropy && pip3 install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@318b7e2"
+pip3 uninstall -y cut-cross-entropy && pip3 install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@e8ad129"
 ```
 
 ## Usage
@@ -31,11 +31,13 @@ plugins:
 
 ## Supported Models
 
+- afmoe
 - apertus
 - arcee
 - cohere
 - cohere2
 - deepseek_v3
+- exaone4
 - gemma
 - gemma2
 - gemma3
@@ -45,13 +47,17 @@ plugins:
 - glm
 - glm4
 - glm4_moe
+- glm4_moe_lite
+- glm46v
 - glm4v
 - glm4v_moe
+- glm_image
+- glm_moe_dsa
 - gpt_oss
 - granite
 - granitemoe
-- granitemoeshared
 - granitemoehybrid
+- granitemoeshared
 - hunyuan_v1_dense
 - hunyuan_v1_moe
 - internvl
@@ -72,20 +78,26 @@ plugins:
 - olmo
 - olmo2
 - olmo3
+- olmoe
 - phi
 - phi3
 - phi4_multimodal
 - qwen2
-- qwen2_vl
-- qwen2_moe
 - qwen2_5_vl
+- qwen2_moe
+- qwen2_vl
 - qwen3
+- qwen3_5
+- qwen3_5_text
+- qwen3_5_moe
+- qwen3_5_moe_text
 - qwen3_moe
+- qwen3_next
 - qwen3_vl
 - qwen3_vl_moe
-- qwen3_next
-- smollm3
 - seed_oss
+- smollm3
+- step3p5
 - voxtral
 
 ## Citation

src/axolotl/integrations/cut_cross_entropy/__init__.py

@@ -35,7 +35,7 @@ LOG = get_logger(__name__)
 
 _CCE_INSTALL_MESSAGE = (
     "Please install Axolotl's fork of cut_cross_entropy with transformers support using "
-    '`pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@318b7e2"`'
+    '`pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@e8ad129"`'
 )
 
 
@@ -104,7 +104,7 @@ class CutCrossEntropyPlugin(BasePlugin):
 
     def patch_llama_like(
         self,
-        model_type: str,
+        model_type_to_patch: str,
     ) -> None:
         """
         Generic patch for model architectures with causal lm similar to llama
@@ -112,7 +112,10 @@ class CutCrossEntropyPlugin(BasePlugin):
         from cut_cross_entropy.transformers.patch import PATCH_FNS
 
         def patch_generic(
-            maybe_model, patch_options, model_type: str, remote_model_id: str | None
+            maybe_model,
+            patch_options,
+            remote_model_id: str | None,
+            model_type: str,
         ):
             import cut_cross_entropy.transformers.llama
             from cut_cross_entropy.transformers.llama import cce_forward
@@ -136,11 +139,13 @@ class CutCrossEntropyPlugin(BasePlugin):
                     f"Error: {str(e)}"
                 ) from e
 
-        if model_type not in PATCH_FNS:
+        if model_type_to_patch not in PATCH_FNS:
             LOG.warning_once(
-                "Setting up generic cce patch for model type: %s", model_type
+                "Setting up generic cce patch for model type: %s", model_type_to_patch
             )
             LOG.warning_once(
-                f"Generic Cut Cross Entropy + {model_type} support is experimental and may not work as expected."
+                f"Generic Cut Cross Entropy + {model_type_to_patch} support is experimental and may not work as expected."
+            )
+            PATCH_FNS[model_type_to_patch] = partial(
+                patch_generic, model_type=model_type_to_patch
             )
-            PATCH_FNS[model_type] = partial(patch_generic, model_type=model_type)

src/axolotl/integrations/kernels/README.md

@@ -0,0 +1,78 @@
+# Kernels Integration
+
+MoE (Mixture of Experts) kernels speed up training for MoE layers and reduce VRAM costs. In transformers v5, `batched_mm` and `grouped_mm` were integrated as built-in options via the `experts_implementation` config kwarg:
+
+```python
+class ExpertsInterface(GeneralInterface):
+    _global_mapping = {
+        "batched_mm": batched_mm_experts_forward,
+        "grouped_mm": grouped_mm_experts_forward,
+    }
+```
+
+In our custom integration, we add support for **ScatterMoE** and **SonicMoE**, which are more efficient and faster than `grouped_mm`.
+
+## Usage
+
+Add the following to your axolotl YAML config:
+
+```yaml
+plugins:
+  - axolotl.integrations.kernels.KernelsPlugin
+
+use_kernels: true
+
+# Choose one (mutually exclusive):
+use_scattermoe: true
+# OR
+use_sonicmoe: true
+```
+
+**Important:** Setting `experts_implementation` is incompatible with custom kernel options.
+
+### SonicMoE installation
+
+**Prerequisites:**
+- NVIDIA Hopper (H100, H200) or Blackwell (B200, GB200) GPU
+- CUDA 12.9+ (13.0+ for B300)
+- PyTorch 2.7+ (2.9.1 recommended)
+- For B300: Triton 3.6.0
+
+```bash
+pip install --ignore-requires-python --no-deps "sonic-moe @ git+https://github.com/Dao-AILab/sonic-moe.git@116e2df0a41874f77fa0ad269ce7df3f0cfcb956" && pip install nvidia-cutlass-dsl==4.4.0 quack-kernels==0.2.5
+```
+
+See the [SonicMoE installation guide](https://github.com/Dao-AILab/sonic-moe?tab=readme-ov-file#-installation) for the latest prerequisite details.
+
+**Note:** Blackwell support is in upstream beta. On Blackwell GPUs, Axolotl automatically sets `USE_QUACK_GEMM=1` to enable the Blackwell kernels.
+
+## How It Works
+
+The `KernelsPlugin` runs before model loading and:
+
+### ScatterMoE
+1. Registers the ScatterMoE kernel from the local `libs/scattermoe_lora` package (includes fused LoRA support via Triton kernels).
+2. Patches the model's `SparseMoeBlock` forward method with the optimized ScatterMoE implementation.
+
+### SonicMoE
+1. Resolves the model's MoE block class(es) from `constants.py`.
+2. Patches the forward method with SonicMoE's optimized kernels and registers a weight converter for the interleaved gate/up projection format.
+3. Supports both softmax->topk and sigmoid->topk routing strategies.
+
+Both paths use the shared `resolve_moe_block_classes` utility in `constants.py` for model-type-to-class resolution.
+
+#### Supported Models
+
+See `constants.py` for the full list of supported model types (Qwen2-MoE, Qwen3-MoE, OLMoE, Mixtral, DeepSeek-V3, GLM-MoE, MiniMax, etc.).
+
+## Limitations
+
+ScatterMoE uses a softmax -> topk routing, so results may be different for some model architectures as baseline (GPT-OSS, etc). Incompatible with `GLM_MOE_DSA` (GLM 5) and `GLM4_MOE_LITE` (GLM 4.7 Flash) at the moment.
+
+SonicMoE supports both softmax->topk and sigmoid->topk routing, covering a wider range of architectures.
+
+ScatterMoE does not work for GLM4.7 Flash (glm4_moe_lite) atm.
+
+## Note on MegaBlocks
+
+We tested [MegaBlocks](https://huggingface.co/kernels-community/megablocks) but were unable to ensure numerical accuracy, so we did not integrate it. It was also incompatible with many newer model architectures in transformers.

src/axolotl/integrations/kernels/__init__.py

@@ -0,0 +1,7 @@
+from .args import KernelsArgs
+from .plugin import KernelsPlugin
+
+__all__ = [
+    "KernelsArgs",
+    "KernelsPlugin",
+]

src/axolotl/integrations/kernels/args.py

@@ -0,0 +1,59 @@
+from pydantic import BaseModel, model_validator
+
+from axolotl.utils.logging import get_logger
+
+LOG = get_logger(__name__)
+
+
+class KernelsArgs(BaseModel):
+    use_scattermoe: bool | None = None
+    use_sonicmoe: bool | None = None
+
+    @model_validator(mode="before")
+    @classmethod
+    def check_mutually_exclusive(cls, data):
+        if data.get("use_scattermoe") and data.get("use_sonicmoe"):
+            raise ValueError(
+                "Cannot use both ScatterMoE and SonicMoE simultaneously. "
+                "Please set only one of `use_scattermoe` or `use_sonicmoe` to true."
+            )
+        return data
+
+    @model_validator(mode="before")
+    @classmethod
+    def check_use_kernels(cls, data):
+        if data.get("use_kernels") is not True:
+            LOG.warning(
+                "`use_kernels` must be set to True to use this. Automatically setting it to True."
+            )
+            data["use_kernels"] = True
+
+        return data
+
+    @model_validator(mode="before")
+    @classmethod
+    def check_experts_implementation(cls, data):
+        experts_implementation = data.get("experts_implementation")
+        if experts_implementation is None:
+            # transformers may default to batched_mm when unset
+            data["experts_implementation"] = "eager"
+        elif experts_implementation != "eager":
+            LOG.warning(
+                "`experts_implementation` must be set to 'eager' to use this. Automatically setting it to 'eager'."
+            )
+            data["experts_implementation"] = "eager"
+
+        return data
+
+    @model_validator(mode="before")
+    @classmethod
+    def disable_mlp_kernel(cls, data):
+        if data.get("use_scattermoe") is True or data.get("use_sonicmoe") is True:
+            if data.get("lora_mlp_kernel") is True:
+                LOG.warning(
+                    "Disabling lora_mlp_kernel when using custom MoE kernels due to compatibility issues."
+                )
+                data["lora_mlp_kernel"] = False
+            data["mlp_kernel"] = False
+
+        return data

src/axolotl/integrations/kernels/constants.py

@@ -0,0 +1,68 @@
+"""
+Supported MoE block mappings for kernel integrations.
+
+Maps model_type to the SparseMoeBlock class name(s) in transformers.
+Used by both ScatterMoE and SonicMoE kernel paths.
+
+Values can be a single class name (str) or a list of class names for models
+with multiple MoE block types (e.g. qwen3_omni_moe has Thinker + Talker).
+"""
+
+import importlib
+
+SPARSE_MOE_BLOCK = {
+    # softmax -> topk routing
+    "qwen2_moe": "Qwen2MoeSparseMoeBlock",
+    "qwen3_moe": "Qwen3MoeSparseMoeBlock",
+    "qwen3_5_moe": "Qwen3_5MoeSparseMoeBlock",
+    "qwen3_next": "Qwen3NextSparseMoeBlock",
+    "qwen3_vl_moe": "Qwen3VLMoeTextSparseMoeBlock",
+    # qwen3_omni_moe: Thinker (standard) + Talker (shared experts + shared_expert_gate)
+    "qwen3_omni_moe": [
+        "Qwen3OmniMoeThinkerTextSparseMoeBlock",
+        "Qwen3OmniMoeTalkerTextSparseMoeBlock",
+    ],
+    "olmoe": "OlmoeSparseMoeBlock",
+    "mixtral": "MixtralSparseMoeBlock",
+    "minimax": "MiniMaxSparseMoeBlock",
+    # sigmoid -> topk routing (with group-based expert selection)
+    "glm_moe_dsa": "GlmMoeDsaMoE",
+    "deepseek_v3": "DeepseekV3MoE",
+    "glm4_moe": "Glm4MoeMoE",
+    "glm4_moe_lite": "Glm4MoeLiteMoE",
+    "glm4v_moe": "Glm4vMoeTextMoE",
+    # sigmoid -> topk routing (no group selection)
+    "minimax_m2": "MiniMaxM2SparseMoeBlock",
+    # Models below need custom routing (not yet implemented):
+    # "ernie4_5_moe": "Ernie4_5_MoeSparseMoeBlock",  # softmax->topk, e_score_correction_bias between softmax and topk
+    # "deepseek_v2": "DeepseekV2Moe",  # softmax->topk, group_limited_greedy, different attr names (num_group)
+    # "hunyuan_v1_moe": "HunYuanMoEV1Moe",  # softmax->topk, gate.wg (not gate.weight), scatter routing
+    # "gpt_oss": "GptOssMLP",  # topk->softmax, transposed layout [E,H,2*I], custom GLU, expert biases
+}
+
+
+def resolve_moe_block_classes(model_type: str):
+    """Resolve all MoE block classes from transformers for the given model type.
+
+    Returns a list of classes (one for most models, multiple for models with
+    distinct MoE block types like qwen3_omni_moe).
+    """
+    entry = SPARSE_MOE_BLOCK.get(model_type)
+    if entry is None:
+        raise ValueError(
+            f"Unsupported MoE model type '{model_type}'. "
+            f"Supported types: {list(SPARSE_MOE_BLOCK.keys())}"
+        )
+
+    cls_names = entry if isinstance(entry, list) else [entry]
+    module_path = f"transformers.models.{model_type}.modeling_{model_type}"
+    module = importlib.import_module(module_path)
+
+    classes = []
+    for cls_name in cls_names:
+        moe_cls = getattr(module, cls_name, None)
+        if moe_cls is None:
+            raise ValueError(f"Could not find class '{cls_name}' in '{module_path}'")
+        classes.append(moe_cls)
+
+    return classes