set version to v0.13.1 (#3363)

* upgrade transformers to 4.57.5

* explicitly set versions for fbgemm-gpu

* handle index url for cuda version

* explicitly set cu version for fbgemm deps, skip for 130

* cu suffix not needed on version if using whl subpath

.github/PULL_REQUEST_TEMPLATE.md

@@ -15,6 +15,11 @@
 <!--- Include details of your testing environment, tests ran to see how -->
 <!--- your change affects other areas of the code, etc. -->
 
+## AI Usage Disclaimer
+
+<!--- Was AI (e.g., ChatGPT, Claude, Copilot) used to generate or assist with this PR? -->
+<!--- Please indicate: No / Yes (specify which tool and to what extent) -->
+
 ## Screenshots (if appropriate)
 
 ## Types of changes

.github/workflows/base.yml

@@ -21,31 +21,12 @@ jobs:
     timeout-minutes: 480
     # this job needs to be run on self-hosted GPU runners...
     runs-on: ubuntu-latest-m
+    env:
+      HAS_DOCKERHUB_CREDS: ${{ secrets.DOCKERHUB_USERNAME != '' && secrets.DOCKERHUB_TOKEN != '' }}
     strategy:
       fail-fast: false
       matrix:
         include:
-          - cuda: "126"
-            cuda_version: 12.6.3
-            cudnn_version: ""
-            python_version: "3.11"
-            pytorch: 2.7.0
-            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
-            dockerfile: "Dockerfile-base"
-          - cuda: "126"
-            cuda_version: 12.6.3
-            cudnn_version: ""
-            python_version: "3.11"
-            pytorch: 2.7.1
-            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
-            dockerfile: "Dockerfile-base"
-          - cuda: "128"
-            cuda_version: 12.8.1
-            cudnn_version: ""
-            python_version: "3.11"
-            pytorch: 2.7.1
-            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
-            dockerfile: "Dockerfile-base"
           - cuda: "128"
             cuda_version: 12.8.1
             cudnn_version: ""
@@ -53,6 +34,15 @@ jobs:
             pytorch: 2.8.0
             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"
+          - cuda: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.9.0
+            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: ""
@@ -60,6 +50,7 @@ jobs:
             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: ""
@@ -67,6 +58,7 @@ jobs:
             pytorch: 2.9.1
             torch_cuda_arch_list: "9.0+PTX"
             dockerfile: "Dockerfile-base"
+            platforms: "linux/amd64,linux/arm64"
 #          - cuda: "128"
 #            cuda_version: 12.8.1
 #            cudnn_version: ""
@@ -93,6 +85,7 @@ jobs:
             axolotlai/axolotl-base
       - name: Login to Docker Hub
         uses: docker/login-action@v2
+        if: ${{ github.event_name != 'pull_request' && env.HAS_DOCKERHUB_CREDS == 'true' }}
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}
@@ -103,6 +96,7 @@ jobs:
         with:
           context: .
           file: ./docker/${{ matrix.dockerfile }}
+          platforms: ${{ matrix.platforms }}
           push: ${{ github.event_name != 'pull_request' }}
           tags: ${{ steps.metadata.outputs.tags }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
           labels: ${{ steps.metadata.outputs.labels }}
@@ -117,24 +111,12 @@ jobs:
     if: ${{ github.repository_owner == 'axolotl-ai-cloud' && (github.event_name != 'pull_request' || !github.event.pull_request.draft) }}
     timeout-minutes: 480
     runs-on: ubuntu-latest-m
+    env:
+      HAS_DOCKERHUB_CREDS: ${{ secrets.DOCKERHUB_USERNAME != '' && secrets.DOCKERHUB_TOKEN != '' }}
     strategy:
       fail-fast: false
       matrix:
         include:
-          - cuda: "126"
-            cuda_version: 12.6.3
-            cudnn_version: ""
-            python_version: "3.11"
-            pytorch: 2.7.1
-            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
-            dockerfile: "Dockerfile-uv-base"
-          - cuda: "128"
-            cuda_version: 12.8.1
-            cudnn_version: ""
-            python_version: "3.11"
-            pytorch: 2.7.1
-            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
-            dockerfile: "Dockerfile-uv-base"
           - cuda: "128"
             cuda_version: 12.8.1
             cudnn_version: ""
@@ -142,6 +124,7 @@ jobs:
             pytorch: 2.8.0
             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"
           - cuda: "128"
             cuda_version: 12.8.1
             cudnn_version: ""
@@ -149,6 +132,15 @@ jobs:
             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: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.9.0
+            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: "130"
             cuda_version: 13.0.0
             cudnn_version: ""
@@ -156,6 +148,7 @@ jobs:
             pytorch: 2.9.1
             torch_cuda_arch_list: "9.0+PTX"
             dockerfile: "Dockerfile-uv-base"
+            platforms: "linux/amd64,linux/arm64"
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -167,6 +160,7 @@ jobs:
             axolotlai/axolotl-base-uv
       - name: Login to Docker Hub
         uses: docker/login-action@v2
+        if: ${{ github.event_name != 'pull_request' && env.HAS_DOCKERHUB_CREDS == 'true' }}
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}
@@ -177,6 +171,7 @@ jobs:
         with:
           context: .
           file: ./docker/${{ matrix.dockerfile }}
+          platforms: ${{ matrix.platforms }}
           push: ${{ github.event_name != 'pull_request' }}
           tags: ${{ steps.metadata.outputs.tags }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
           labels: ${{ steps.metadata.outputs.labels }}

.github/workflows/docs.yml

@@ -12,6 +12,9 @@ jobs:
     build-deploy:
         runs-on: ubuntu-latest
         steps:
+        - name: cleanup node
+          run: |
+            sudo rm -rf /usr/share/dotnet /usr/local/lib/android /opt/ghc /opt/hostedtoolcache/CodeQL
         - name: Check out repository
           uses: actions/checkout@v4
         - name: Set up Quarto

.github/workflows/main.yml

@@ -15,37 +15,31 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.7.0
-            axolotl_extras:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.7.1
-            axolotl_extras: vllm
-          - cuda: 128
-            cuda_version: 12.8.1
-            python_version: "3.11"
-            pytorch: 2.7.1
-            axolotl_extras:
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
             pytorch: 2.8.0
             axolotl_extras:
-            is_latest: true
+            platforms: "linux/amd64"
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
             pytorch: 2.9.0
             axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
           - 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: 130
+            cuda_version: 13.0.0
+            python_version: "3.11"
+            pytorch: 2.9.1
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -71,6 +65,7 @@ jobs:
         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 }}
@@ -92,43 +87,31 @@ jobs:
     strategy:
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.7.0
-            axolotl_extras:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.7.1
-            axolotl_extras:
-            is_latest:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.7.1
-            axolotl_extras: vllm
-          - cuda: 128
-            cuda_version: 12.8.1
-            python_version: "3.11"
-            pytorch: 2.7.1
-            axolotl_extras:
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
             pytorch: 2.8.0
             axolotl_extras:
-            is_latest: true
+            platforms: "linux/amd64"
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
             pytorch: 2.9.0
             axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
           - 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: 130
+            cuda_version: 13.0.0
+            python_version: "3.11"
+            pytorch: 2.9.1
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -153,6 +136,7 @@ jobs:
         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 }}
@@ -170,22 +154,16 @@ jobs:
     strategy:
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.7.1
-            axolotl_extras:
-            is_latest:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.7.1
-            axolotl_extras: vllm
-            is_latest: true
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.8.0
+            pytorch: 2.9.1
+            axolotl_extras:
+            is_latest: true
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.11"
+            pytorch: 2.9.1
             axolotl_extras:
             is_latest:
     runs-on: axolotl-gpu-runner
@@ -212,6 +190,7 @@ jobs:
         uses: docker/build-push-action@v5
         with:
           context: .
+          platforms: linux/amd64,linux/arm64
           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 }}

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

@@ -19,6 +19,9 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
   cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
 
+env:
+  MODAL_IMAGE_BUILDER_VERSION: "2025.06"
+
 jobs:
   test-axolotl-multigpu:
     if: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' && (github.event_name != 'pull_request' || !github.event.pull_request.draft) }}
@@ -26,13 +29,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.7.1
-            axolotl_extras: vllm
-            num_gpus: 2
-            nightly_build: "true"
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
@@ -43,10 +39,18 @@ jobs:
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.9.0
+            pytorch: 2.9.1
             axolotl_extras: fbgemm-gpu
             num_gpus: 2
             nightly_build: "true"
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.11"
+            pytorch: 2.9.1
+            axolotl_extras:
+#            axolotl_extras: fbgemm-gpu
+            num_gpus: 2
+            nightly_build: "true"
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
     steps:
@@ -59,7 +63,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2
+          pip install modal==1.3.0.post1 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
@@ -72,4 +76,4 @@ jobs:
           echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         run: |
-          modal run cicd.multigpu
+          modal run -m cicd.multigpu

.github/workflows/nightlies.yml

@@ -12,16 +12,16 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.7.1
-            axolotl_extras:
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
             pytorch: 2.8.0
             axolotl_extras:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.9.1
+            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -64,16 +64,16 @@ jobs:
     strategy:
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.7.1
-            axolotl_extras:
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
             pytorch: 2.8.0
             axolotl_extras:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.9.1
+            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout

.github/workflows/preview-docs.yml

@@ -11,6 +11,7 @@ on:
       - '_quarto.yml'
       - docs/scripts/generate_config_docs.py
       - src/axolotl/utils/schemas/**.py
+      - .github/workflows/preview-docs.yml
 
 permissions:
   checks: write
@@ -27,6 +28,10 @@ jobs:
     runs-on: ubuntu-latest
     if: ${{ !github.event.pull_request.draft }}
     steps:
+      - name: cleanup node
+        run: |
+          sudo rm -rf /usr/share/dotnet /usr/local/lib/android /opt/ghc /opt/hostedtoolcache/CodeQL
+
       - name: Check out repository
         uses: actions/checkout@v4
         with:

.github/workflows/tests-nightly.yml

@@ -26,7 +26,7 @@ jobs:
       max-parallel: 2
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.7.1", "2.8.0"]
+        pytorch_version: ["2.8.0", "2.9.0", "2.9.1"]
     timeout-minutes: 20
 
     steps:
@@ -99,17 +99,17 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
+          - cuda: 128
+            cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.7.1
+            pytorch: 2.8.0
             num_gpus: 1
             axolotl_extras:
             nightly_build: "true"
           - 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"
@@ -123,7 +123,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2
+          pip install modal==1.3.0.post1 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
@@ -148,10 +148,10 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
+          - cuda: 128
+            cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.7.1
+            pytorch: 2.9.1
             num_gpus: 2
             axolotl_extras:
             nightly_build: "true"
@@ -165,7 +165,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2
+          pip install modal==1.3.0.post1 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV

.github/workflows/tests.yml

@@ -55,7 +55,7 @@ jobs:
       fail-fast: false
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.7.1", "2.8.0", "2.9.0"]
+        pytorch_version: ["2.8.0", "2.9.0", "2.9.1"]
     timeout-minutes: 20
 
     steps:
@@ -69,8 +69,9 @@ jobs:
       - name: Restore Cache from S3
         id: hf-cache-restore-s3
         run: |
-          mkdir -p /home/runner/.cache/huggingface/hub
-          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C /home/runner/.cache/huggingface/hub/  --use-compress-program unzstd
+          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
+          ls -ltr ~/.cache/huggingface/hub/
 
       - name: Setup Python
         uses: actions/setup-python@v5
@@ -111,13 +112,23 @@ jobs:
         run: |
           huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
 
+      - name: Show HF cache
+        run: hf cache scan
+
       - name: Run tests
         run: |
-          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ --ignore=tests/monkeypatch/ tests/ --cov=axolotl --cov-report=xml
+          df -h
+          pytest -v --durations=10 -n4 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ --ignore=tests/monkeypatch/ tests/ --cov=axolotl --cov-report=xml
+          df -h
           pytest -v --durations=10 tests/monkeypatch/ --cov=axolotl --cov-append --cov-report=xml
+          df -h
           pytest -v --durations=10 tests/patched/ --cov=axolotl --cov-append --cov-report=xml
+          df -h
           pytest -v --durations=10 tests/cli/ --cov=axolotl --cov-append --cov-report=xml
 
+      - name: Show HF cache
+        run: hf cache scan
+
       - name: Upload coverage to Codecov
         uses: codecov/codecov-action@v5
         with:
@@ -134,7 +145,7 @@ jobs:
       fail-fast: false
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.7.1", "2.8.0", "2.9.0"]
+        pytorch_version: ["2.8.0", "2.9.0", "2.9.1"]
     timeout-minutes: 20
 
     steps:
@@ -148,8 +159,9 @@ jobs:
       - name: Restore Cache from S3
         id: hf-cache-restore-s3
         run: |
-          mkdir -p /home/runner/.cache/huggingface/hub
-          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C /home/runner/.cache/huggingface/hub/  --use-compress-program unzstd
+          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
+          ls -ltr ~/.cache/huggingface/hub/
 
       - name: Setup Python
         uses: actions/setup-python@v5
@@ -188,14 +200,17 @@ jobs:
           axolotl --help
 
       - name: Show HF cache
-        run: huggingface-cli scan-cache
+        run: hf cache scan
 
       - name: Run tests
         run: |
-          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ --ignore=tests/monkeypatch/ tests/ --cov=axolotl --cov-report=xml
+          pytest -v --durations=10 -n4 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ --ignore=tests/monkeypatch/ tests/ --cov=axolotl --cov-report=xml
           pytest -v --durations=10 tests/monkeypatch/ --cov=axolotl --cov-append --cov-report=xml
           pytest -v --durations=10 tests/cli/
 
+      - name: Show HF cache
+        run: hf cache scan
+
   gate-skip-e2e:
     needs: [pre-commit, pytest, pytest-sdist]
     runs-on: ubuntu-latest
@@ -256,7 +271,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2
+          pip install modal==1.3.0.post1 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
@@ -288,18 +303,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.7.1
-            num_gpus: 1
-            axolotl_extras:
-#          - cuda: 128
-#            cuda_version: 12.8.1
-#            python_version: "3.11"
-#            pytorch: 2.7.1
-#            num_gpus: 1
-#            axolotl_extras:
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
@@ -310,7 +313,13 @@ jobs:
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.9.0
+            pytorch: 2.9.1
+            num_gpus: 1
+            axolotl_extras:
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.11"
+            pytorch: 2.9.1
             num_gpus: 1
             axolotl_extras:
     steps:
@@ -323,7 +332,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2
+          pip install modal==1.3.0.post1 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
@@ -350,10 +359,10 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
+          - cuda: 128
+            cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.7.1
+            pytorch: 2.9.1
             num_gpus: 1
             axolotl_extras:
     steps:
@@ -366,7 +375,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2
+          pip install modal==1.3.0.post1 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV

.pre-commit-config.yaml

@@ -11,13 +11,13 @@ repos:
     -   id: no-commit-to-branch
         args: ['--branch', 'main']
 -   repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.14.3
+    rev: v0.14.10
     hooks:
     -   id: ruff
         args: [--fix]
     -   id: ruff-format
 -   repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.18.2
+    rev: v1.19.1
     hooks:
     - id: mypy
       additional_dependencies:
@@ -26,7 +26,7 @@ repos:
             'pydantic>=2.5.3',
         ]
 -   repo: https://github.com/PyCQA/bandit
-    rev: 1.8.6
+    rev: 1.9.2
     hooks:
     -   id: bandit
         args: [

.runpod/Dockerfile

@@ -10,6 +10,7 @@ ARG BASE_VOLUME="/runpod-volume"
 ENV BASE_VOLUME=$BASE_VOLUME
 ENV HF_DATASETS_CACHE="${BASE_VOLUME}/huggingface-cache/datasets"
 ENV HUGGINGFACE_HUB_CACHE="${BASE_VOLUME}/huggingface-cache/hub"
+ENV HF_HUB_CACHE="${BASE_VOLUME}/huggingface-cache/hub"
 ENV TRANSFORMERS_CACHE="${BASE_VOLUME}/huggingface-cache/hub"
 
 COPY .runpod/src /src

README.md

@@ -29,15 +29,15 @@
 
 ## 🎉 Latest Updates
 
-- 2025/11: Axolotl now includes support for [Olmo3](https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/olmo3).
-- 2025/10: New model support has been added in Axolotl for: [Qwen3 Next](https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/qwen3-next), [Qwen2.5-vl, Qwen3-vl](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/qwen2_5-vl), [Qwen3, Qwen3MoE](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/qwen3), [Granite 4](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/granite4), [HunYuan](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/hunyuan), [Magistral 2509](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/magistral#vision), [Apertus](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/apertus), and [Seed-OSS](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/seed-oss).
+- 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).
+- 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).
 - 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:
   - ND Parallelism support has been added into Axolotl. Compose Context Parallelism (CP), Tensor Parallelism (TP), and Fully Sharded Data Parallelism (FSDP) within a single node and across multiple nodes. Check out the [blog post](https://huggingface.co/blog/accelerate-nd-parallel) for more info.
-  - Axolotl adds more models: [GPT-OSS](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/gpt-oss), [Gemma 3n](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/gemma3n), [Liquid Foundation Model 2 (LFM2)](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/lfm2), and [Arcee Foundation Models (AFM)](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/afm).
+  - Axolotl adds more models: [GPT-OSS](https://docs.axolotl.ai/docs/models/gpt-oss.html), [Gemma 3n](https://docs.axolotl.ai/docs/models/gemma3n.html), [Liquid Foundation Model 2 (LFM2)](https://docs.axolotl.ai/docs/models/LiquidAI.html), and [Arcee Foundation Models (AFM)](https://docs.axolotl.ai/docs/models/arcee.html).
   - 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://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/voxtral), [Magistral 1.1](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/magistral), and [Devstral](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/devstral) with mistral-common tokenizer support has been integrated in Axolotl!
+  - [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!
 
@@ -46,8 +46,8 @@
 <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 [examples](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/magistral) to start training your own Magistral models with Axolotl!
-- 2025/04: Llama 4 support has been added in Axolotl. See [examples](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/llama-4) to start training your own Llama 4 models with Axolotl's linearized version!
+- 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!
+- 2025/04: 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: (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!
@@ -77,7 +77,7 @@ Features:
 
 - NVIDIA GPU (Ampere or newer for `bf16` and Flash Attention) or AMD GPU
 - Python 3.11
-- PyTorch ≥2.7.1
+- PyTorch ≥2.8.0
 
 ### Google Colab
 

_quarto.yml

@@ -1,6 +1,8 @@
 project:
   type: website
-  pre-render: docs/scripts/generate_config_docs.py
+  pre-render:
+   - docs/scripts/generate_config_docs.py
+   - docs/scripts/generate_examples_docs.py
 
 quartodoc:
   dir: docs/api
@@ -240,6 +242,46 @@ website:
             - docs/getting-started.qmd
             - docs/installation.qmd
             - docs/inference.qmd
+            - section: "Model Guides"
+              contents:
+                - docs/models/kimi-linear.qmd
+                - docs/models/plano.qmd
+                - docs/models/mimo.qmd
+                - docs/models/internvl3_5.qmd
+                - docs/models/olmo3.qmd
+                - docs/models/trinity.qmd
+                - docs/models/arcee.qmd
+                - docs/models/mistral.qmd
+                - section: "Ministral3"
+                  contents:
+                    - docs/models/ministral3.qmd
+                    - docs/models/ministral3/think.qmd
+                    - docs/models/ministral3/vision.qmd
+                - section: "Magistral"
+                  contents:
+                    - docs/models/magistral.qmd
+                    - docs/models/magistral/think.qmd
+                    - docs/models/magistral/vision.qmd
+                - docs/models/ministral.qmd
+                - docs/models/mistral-small.qmd
+                - docs/models/voxtral.qmd
+                - docs/models/devstral.qmd
+                - docs/models/llama-4.qmd
+                - docs/models/llama-2.qmd
+                - docs/models/qwen3-next.qmd
+                - docs/models/qwen3.qmd
+                - docs/models/gemma3n.qmd
+                - docs/models/apertus.qmd
+                - docs/models/gpt-oss.qmd
+                - docs/models/seed-oss.qmd
+                - docs/models/phi.qmd
+                - docs/models/smolvlm2.qmd
+                - docs/models/granite4.qmd
+                - docs/models/LiquidAI.qmd
+                - docs/models/hunyuan.qmd
+                - docs/models/jamba.qmd
+                - docs/models/orpheus.qmd
+
             - docs/cli.qmd
             - docs/telemetry.qmd
             - docs/config-reference.qmd

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 \
+pytest -v --durations=10 -n2 --maxfail=4 \
   --ignore=/workspace/axolotl/tests/e2e/multigpu/solo/ \
   --ignore=/workspace/axolotl/tests/e2e/multigpu/patched/ \
   /workspace/axolotl/tests/e2e/multigpu/ \

docker/Dockerfile

@@ -6,6 +6,7 @@ ARG AXOLOTL_EXTRAS=""
 ARG AXOLOTL_ARGS=""
 ARG CUDA="118"
 ARG PYTORCH_VERSION="2.1.2"
+ARG TARGETARCH
 
 ENV PYTORCH_VERSION=$PYTORCH_VERSION
 
@@ -20,13 +21,17 @@ 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
-RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
+# If AXOLOTL_EXTRAS is set, append it in brackets; don't install deepspeed with arm64
+RUN if [ "$TARGETARCH" = "arm64" ]; then \
+        BASE_EXTRAS="flash-attn,ring-flash-attn,optimizers,ray"; \
     else \
-        pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray] $AXOLOTL_ARGS; \
+        BASE_EXTRAS="deepspeed,flash-attn,ring-flash-attn,optimizers,ray"; \
     fi && \
-    python scripts/unsloth_install.py | sh && \
+    if [ "$AXOLOTL_EXTRAS" != "" ]; then \
+        pip install --no-build-isolation -e .[$BASE_EXTRAS,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
+    else \
+        pip install --no-build-isolation -e .[$BASE_EXTRAS] $AXOLOTL_ARGS; \
+    fi && \    python scripts/unsloth_install.py | sh && \
     python scripts/cutcrossentropy_install.py | sh && \
     pip install pytest && \
     pip cache purge

docker/Dockerfile-base

@@ -2,14 +2,16 @@ ARG CUDA_VERSION="11.8.0"
 ARG CUDNN_VERSION="8"
 ARG UBUNTU_VERSION="22.04"
 ARG MAX_JOBS=4
+ARG TARGETARCH
 
 FROM nvidia/cuda:$CUDA_VERSION-cudnn$CUDNN_VERSION-devel-ubuntu$UBUNTU_VERSION AS base-builder
 
 ENV PATH="/root/miniconda3/bin:${PATH}"
 
-ARG PYTHON_VERSION="3.10"
+ARG TARGETARCH
+ARG PYTHON_VERSION="3.11"
 ARG PYTORCH_VERSION="2.1.2"
-ARG CUDA="118"
+ARG CUDA="128"
 ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 9.0+PTX"
 
 ENV PYTHON_VERSION=$PYTHON_VERSION
@@ -22,11 +24,17 @@ RUN apt-get update \
         librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm \
     && rm -rf /var/cache/apt/archives \
     && rm -rf /var/lib/apt/lists/* \
-    && wget \
-    https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh \
+    && if [ "$TARGETARCH" = "amd64" ]; then \
+        MINICONDA_ARCH="x86_64"; \
+    elif [ "$TARGETARCH" = "arm64" ]; then \
+        MINICONDA_ARCH="aarch64"; \
+    else \
+        echo "Unsupported architecture: $TARGETARCH"; exit 1; \
+    fi \
+    && wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-${MINICONDA_ARCH}.sh \
     && mkdir /root/.conda \
-    && bash Miniconda3-latest-Linux-x86_64.sh -b \
-    && rm -f Miniconda3-latest-Linux-x86_64.sh \
+    && bash Miniconda3-latest-Linux-${MINICONDA_ARCH}.sh -b \
+    && rm -f Miniconda3-latest-Linux-${MINICONDA_ARCH}.sh \
     && conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/main \
     && conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/r \
     && conda create -n "py${PYTHON_VERSION}" python="${PYTHON_VERSION}"
@@ -51,8 +59,34 @@ RUN git lfs install --skip-repo && \
     pip3 install -U --no-cache-dir pydantic==1.10.10 && \
     pip3 cache purge
 
-RUN if [ "$PYTORCH_VERSION" = "2.9.1" ] && [ "$CUDA" = "128" ] ; then \
-        wget https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.4.17/flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_x86_64.whl; \
-        pip3 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; \
-    fi
+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

docker/Dockerfile-uv-base

@@ -2,6 +2,7 @@ ARG CUDA_VERSION="12.6.3"
 ARG CUDNN_VERSION=""
 ARG UBUNTU_VERSION="22.04"
 ARG MAX_JOBS=4
+ARG TARGETARCH
 
 FROM nvidia/cuda:$CUDA_VERSION-cudnn$CUDNN_VERSION-devel-ubuntu$UBUNTU_VERSION AS base-builder
 
@@ -31,12 +32,35 @@ ENV PATH="/workspace/axolotl-venv/bin:${PATH}"
 
 RUN uv pip install packaging setuptools wheel psutil \
     && uv pip install torch==${PYTORCH_VERSION} torchvision \
-    && uv pip install --no-build-isolation "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main" \
-    && uv pip install "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main" \
     && uv pip install awscli pydantic
 
-RUN if [ "$PYTORCH_VERSION" = "2.9.0" ] && [ "$CUDA" = "128" ] ; then \
-        wget https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.4.17/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; \
+RUN if [ "$TARGETARCH" = "amd64" ]; then \
+        uv pip install --no-build-isolation "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main"; \
+        uv pip install "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main"; \
     fi
+
+RUN case "$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

docs/.gitignore

@@ -3,3 +3,5 @@ _site/
 /api/*.qmd
 /api/*.html
 config-reference.qmd
+models/**/*.qmd
+models/**/*.html

docs/checkpoint_saving.qmd

@@ -0,0 +1,86 @@
+---
+title: "Checkpoint Saving"
+format:
+  html:
+    toc: true
+    toc-depth: 2
+    number-sections: true
+execute:
+  enabled: false
+---
+
+## Overview
+
+Axolotl supports on-demand checkpoint saving during training. You can trigger checkpoints via file-based triggers (for programmatic control) or Control+C (for interactive use).
+
+## File-Based Checkpoint Trigger
+
+### Configuration
+
+Enable in your config:
+
+```yaml
+dynamic_checkpoint:
+  enabled: true
+  check_interval: 100  # Optional: check every N steps (default: 100)
+  trigger_file_path: "axolotl_checkpoint.save"  # Optional: custom filename
+```
+
+**Options:**
+- `enabled`: `true` to enable (required)
+- `check_interval`: Steps between file checks. Default: 100. Lower = faster response, higher I/O overhead.
+- `trigger_file_path`: Custom trigger filename. Default: `axolotl_checkpoint.save`
+
+### How It Works
+
+1. Rank 0 checks for trigger file every `check_interval` steps in `output_dir`
+2. When detected, file is deleted and checkpoint is saved
+3. In distributed training, rank 0 broadcasts to synchronize all ranks
+
+### Usage
+
+**Command line:**
+```bash
+touch /path/to/output_dir/axolotl_checkpoint.save
+```
+
+**Programmatic:**
+```python
+from pathlib import Path
+Path("/path/to/output_dir/axolotl_checkpoint.save").touch()
+```
+
+Checkpoint saves within the next `check_interval` steps. The trigger file is auto-deleted after detection, so you can create it multiple times.
+
+**Custom filename:**
+```yaml
+dynamic_checkpoint:
+  enabled: true
+  trigger_file_path: "my_trigger.save"
+```
+```bash
+touch /path/to/output_dir/my_trigger.save
+```
+
+## Control+C (SIGINT) Checkpoint
+
+Pressing `Ctrl+C` during training saves the model state and exits gracefully. **Note:** This saves only the model weights, not optimizer state. For resumable checkpoints, use the file-based trigger.
+
+## Best Practices
+
+- **Check interval**: Lower values (10-50) for fast training, default 100 for slower training
+- **Distributed training**: Create trigger file once; rank 0 handles synchronization
+- **Resume**: Dynamic checkpoints can be resumed like regular checkpoints via `resume_from_checkpoint`
+
+## Example
+
+```yaml
+output_dir: ./outputs/lora-out
+save_steps: 500  # Scheduled checkpoints
+
+dynamic_checkpoint:
+  enabled: true
+  check_interval: 50
+```
+
+This enables scheduled checkpoints every 500 steps plus on-demand saves via file trigger (checked every 50 steps).

docs/docker.qmd

@@ -32,11 +32,8 @@ main-base-py{python_version}-cu{cuda_version}-{pytorch_version}
 
 Tags examples:
 
-- `main-base-py3.11-cu128-2.7.1`
-- `main-base-py3.11-cu126-2.7.1`
-- `main-base-py3.11-cu126-2.7.0`
-- `main-base-py3.11-cu126-2.6.0`
-- `main-base-py3.11-cu124-2.6.0`
+- `main-base-py3.11-cu128-2.8.0`
+- `main-base-py3.11-cu128-2.9.1`
 
 ## Main
 
@@ -74,15 +71,12 @@ There may be some extra tags appended to the image, like `-vllm` which installs
 
 Tags examples:
 
-- `main-py3.11-cu128-2.7.1`
-- `main-py3.11-cu126-2.7.1`
-- `main-py3.11-cu126-2.7.0`
-- `main-py3.11-cu126-2.6.0`
-- `main-py3.11-cu124-2.6.0`
+- `main-py3.11-cu128-2.8.0`
+- `main-py3.11-cu128-2.9.1`
 - `main-latest`
 - `main-20250303-py3.11-cu124-2.6.0`
 - `main-20250303-py3.11-cu126-2.6.0`
-- `0.10.1`
+- `0.12.0`
 
 ## Cloud
 

docs/installation.qmd

@@ -26,7 +26,7 @@ Follow the instructions at: [https://pytorch.org/get-started/locally/](https://p
 :::
 
 ::: {.callout-important}
-For Blackwell GPUs, please use Pytorch 2.7.0 and CUDA 12.8.
+For Blackwell GPUs, please use Pytorch 2.9.1 and CUDA 12.8.
 :::
 
 ### PyPI Installation (Recommended) {#sec-pypi}
@@ -111,7 +111,7 @@ docker run --privileged --gpus '"all"' --shm-size 10g --rm -it \
 :::
 
 ::: {.callout-important}
-For Blackwell GPUs, please use `axolotlai/axolotl:main-py3.11-cu128-2.7.0` or the cloud variant `axolotlai/axolotl-cloud:main-py3.11-cu128-2.7.0`.
+For Blackwell GPUs, please use `axolotlai/axolotl:main-py3.11-cu128-2.9.1` or the cloud variant `axolotlai/axolotl-cloud:main-py3.11-cu128-2.9.1`.
 :::
 
 Please refer to the [Docker documentation](docker.qmd) for more information on the different Docker images that are available.

docs/multimodal.qmd

@@ -21,6 +21,7 @@ format:
 - [Qwen2.5-VL](#sec-qwen25-vl)
 - [SmolVLM2](#sec-smolvlm2)
 - [LFM2-VL](#sec-lfm2-vl)
+- [Intern-VL](#sec-intern-vl)
 
 ## Usage
 
@@ -202,6 +203,16 @@ Please uninstall `causal-conv1d` via `pip3 uninstall -y causal-conv1d`
 base_model: LiquidAI/LFM2-VL-450M
 ```
 
+### Intern-VL {#sec-intern-vl}
+
+::: {.callout-tip}
+Please make sure to install `timm` via `pip3 install timm==1.0.19`
+:::
+
+```yaml
+base_model: OpenGVLab/InternVL3_5-8B
+```
+
 ## Dataset Format
 
 For multi-modal datasets, we adopt an extended `chat_template` format similar to OpenAI's Message format.

docs/scripts/examples-allowlist.yml

@@ -0,0 +1,90 @@
+examples:
+  # December 2025
+  - name: kimi-linear
+    title: Kimi Linear
+  - name: plano
+    title: Plano Orchestrator
+  - name: mimo
+    title: MiMo
+  - name: internvl3_5
+    title: InternVL 3.5
+
+  # AllenAI
+  - name: olmo3
+    title: OLMo 3
+
+  # ArceeAI
+  - name: trinity
+    title: Trinity
+  - name: arcee
+    title: Arcee AFM
+
+  # MistralAI
+  - name: ministral3/think
+    title: Ministral 3 Thinking
+  - name: ministral3/vision
+    title: Ministral 3 Vision
+  - name: magistral/think
+    title: Magistral Thinking
+  - name: magistral/vision
+    title: Magistral Vision
+  - name: ministral
+    title: Ministral
+  - name: mistral-small
+    title: Mistral Small 3.1/3.2
+  - name: voxtral
+    title: Voxtral
+  - name: devstral
+    title: Devstral
+  - name: mistral
+    title: Mistral 7B
+
+  # Meta
+  - name: llama-4
+    title: Llama 4
+  - name: llama-2
+    title: Llama 2
+
+  # Alibaba
+  - name: qwen3-next
+    title: Qwen 3 Next
+  - name: qwen3
+    title: Qwen 3
+
+  # Google
+  - name: gemma3n
+    title: Gemma 3n
+
+  # Swiss AI
+  - name: apertus
+    title: Apertus
+
+  # GPT-OSS
+  - name: gpt-oss
+    title: GPT-OSS
+  - name: seed-oss
+    title: Seed-OSS
+
+  # Microsoft
+  - name: phi
+    title: Phi
+
+  # SmolVLM
+  - name: smolvlm2
+    title: SmolVLM 2
+
+  # IBM
+  - name: granite4
+    title: Granite 4
+
+  # LiquidAI
+  - name: LiquidAI
+    title: Liquid Foundation Models 2
+
+  # Other
+  - name: hunyuan
+    title: Hunyuan
+  - name: jamba
+    title: Jamba
+  - name: orpheus
+    title: Orpheus

docs/scripts/generate_examples_docs.py

@@ -0,0 +1,424 @@
+"""
+auto generate example docs from allowlist
+"""
+
+import re
+import shutil
+import sys
+from pathlib import Path
+
+import yaml
+
+# Paths
+THIS = Path(__file__).resolve()
+ROOT = THIS.parents[2]  # repo root (docs/scripts -> docs -> ROOT)
+EXAMPLES_DIR = ROOT / "examples"
+OUTPUT_DIR = ROOT / "docs" / "models"
+ALLOWLIST_YML = THIS.parent / "examples-allowlist.yml"
+
+
+def slugify(name: str) -> str:
+    """Convert a name to a slug (lowercase, hyphens for spaces)."""
+    s = re.sub(r"[^a-zA-Z0-9\s\-]+", "", name.strip())
+    s = re.sub(r"\s+", "-", s).strip("-").lower()
+    return s or "example"
+
+
+def read_allowlist():
+    with open(ALLOWLIST_YML, "r", encoding="utf-8") as f:
+        data = yaml.safe_load(f) or {}
+    items = data.get("examples", [])
+    if not isinstance(items, list):
+        raise ValueError("`examples` must be a list in examples-allowlist.yml")
+    return items
+
+
+def find_readme(folder: Path) -> Path | None:
+    for name in ("README.md", "Readme.md", "readme.md"):
+        p = folder / name
+        if p.exists():
+            return p
+    return None
+
+
+def remove_first_h1(md: str) -> tuple[str, str | None]:
+    """
+    Remove the first H1 from markdown and return (modified_md, h1_title).
+    The H1 is removed since we use the frontmatter title instead.
+    """
+    lines = md.splitlines()
+    result = []
+    h1_title = None
+    skipped_first = False
+
+    for line in lines:
+        if not skipped_first and line.startswith("# "):
+            h1_title = line[2:].strip()
+            skipped_first = True
+            continue
+        result.append(line)
+
+    return "\n".join(result), h1_title
+
+
+IMG_RE = re.compile(r"!\[[^\]]*\]\(([^)]+)\)")
+LINK_RE = re.compile(r"\[([^\]]+)\]\(([^)]+)\)")
+
+
+def rewrite_and_copy_assets(md: str, src_dir: Path, dest_assets_root: Path) -> str:
+    """
+    Copy local image assets referenced in markdown to
+    docs/examples/assets/... and rewrite the links.
+    """
+    dest_assets = dest_assets_root / "assets"
+
+    def repl(m):
+        url = m.group(1).strip()
+        if re.match(r"^(https?:)?//", url):
+            return m.group(0)  # leave remote URLs
+        src_path = (src_dir / url).resolve()
+        if not src_path.exists():
+            return m.group(0)  # leave as-is if not found
+        rel = src_path.relative_to(src_dir)
+        # Create a unique asset path based on source directory name
+        asset_name = src_dir.name.replace("/", "-")
+        dest_path = dest_assets / asset_name / rel
+        dest_path.parent.mkdir(parents=True, exist_ok=True)
+        shutil.copy2(src_path, dest_path)
+        new_rel = f"assets/{asset_name}/{rel.as_posix()}"
+        return m.group(0).replace(url, new_rel)
+
+    return IMG_RE.sub(repl, md)
+
+
+def rewrite_readme_links(
+    md: str,
+    src_dir: Path,
+    examples_dir: Path,
+    parent_index_only: set,
+    current_src_path: str,
+    allowlist_entries: set,
+    current_output_path: str,
+) -> str:
+    """
+    Rewrite links between README.md files to point to the correct .qmd files.
+    """
+
+    def repl(m):
+        text = m.group(1)
+        url = m.group(2).strip()
+
+        # Skip remote URLs and anchor links
+        if re.match(r"^(https?:)?//", url) or url.startswith("#"):
+            return m.group(0)
+
+        # Skip non-markdown files
+        if not url.lower().endswith(".md"):
+            return m.group(0)
+
+        # Resolve the target path
+        try:
+            target_path = (src_dir / url).resolve()
+
+            # Check if target is outside examples_dir
+            try:
+                rel_path = target_path.relative_to(examples_dir)
+            except ValueError:
+                # Target is outside examples_dir, leave as-is
+                return m.group(0)
+
+            parts = list(rel_path.parts)
+
+            # Determine the output path for the target
+            if len(parts) > 0 and parts[-1].lower() in ("readme.md", "readme"):
+                # This is a README link
+                if len(parts) == 1:
+                    # Link to root README -> index.qmd
+                    target_output = "index.qmd"
+                elif len(parts) == 2:
+                    if parts[0] == ".":
+                        # Current directory README
+                        target_output = "index.qmd"
+                    else:
+                        # subdir/README.md
+                        parent_dir = parts[0]
+                        if parent_dir in parent_index_only:
+                            target_output = f"{parent_dir}/index.qmd"
+                        else:
+                            target_output = f"{parent_dir}.qmd"
+                else:
+                    # Deeper nesting: parent/subdir/README.md
+                    # Build the full path like "parent/subdir"
+                    full_path = "/".join(parts[:-1])  # Remove README.md
+                    # Check if this exact path is in allowlist
+                    if full_path in allowlist_entries:
+                        # This is a sub-entry with its own entry -> use .qmd
+                        target_output = f"{full_path}.qmd"
+                    elif parts[0] == ".":
+                        # ./subdir/README.md -> check if subdir has own entry
+                        subdir = parts[1]
+                        if subdir in parent_index_only:
+                            target_output = f"{subdir}/index.qmd"
+                        else:
+                            target_output = f"{subdir}.qmd"
+                    else:
+                        # parent/subdir where parent doesn't have own entry
+                        target_output = f"{full_path}/index.qmd"
+            else:
+                # Regular .md file -> convert to .qmd, keep path structure
+                target_output = "/".join(parts)[:-2] + "qmd"
+
+            # Compute relative path from current output file to target
+            current_parts = current_output_path.split("/")
+            target_parts = target_output.split("/")
+
+            # Special case: if current is a subdir file and target is a single-component file at root
+            # Example: current="magistral/vision", target="magistral.qmd"
+            if len(current_parts) > 1 and len(target_parts) == 1:
+                # Current is in subdir, target is at root level
+                # Go up to root: ../ for each level
+                up_count = len(current_parts) - 1
+                rel_parts = [".."] * up_count + [target_parts[0]]
+                new_url = "/".join(rel_parts)
+            else:
+                # Find common prefix
+                i = 0
+                while (
+                    i < min(len(current_parts) - 1, len(target_parts))
+                    and current_parts[i] == target_parts[i]
+                ):
+                    i += 1
+
+                # Build relative path: go up (../) then down to target
+                up_count = len(current_parts) - 1 - i
+                rel_parts = [".."] * up_count + target_parts[i:]
+
+                if not rel_parts or rel_parts == [".."]:
+                    # Points to same directory or parent
+                    new_url = "/".join(rel_parts) if rel_parts else "."
+                else:
+                    new_url = "/".join(rel_parts)
+
+            return f"[{text}]({new_url})"
+        except (ValueError, IndexError):
+            return m.group(0)
+
+    return LINK_RE.sub(repl, md)
+
+
+def write_qmd(out_path: Path, title: str, body_md: str):
+    out_path.parent.mkdir(parents=True, exist_ok=True)
+    fm = f"---\ntitle: {title!r}\nexecute:\n  eval: false\nformat:\n  html:\n    toc: true\n---\n\n"
+    out_path.write_text(fm + body_md, encoding="utf-8")
+
+
+def update_quarto_yml(generated: list[tuple[str, str, str]]):
+    """
+    Update _quarto.yml with the generated example files in the correct order.
+    This keeps the sidebar in sync with the allowlist.
+
+    Model Guides is now nested under "Getting Started" section.
+    Creates nested sections for models with sub-entries (e.g., magistral, ministral3).
+    Parent pages are now flat files (e.g., ministral3.qmd) with sub-pages in subdirs.
+    """
+    quarto_yml = ROOT / "_quarto.yml"
+    if not quarto_yml.exists():
+        print(f"[WARN] {quarto_yml} not found, skipping update", file=sys.stderr)
+        return
+
+    content = quarto_yml.read_text(encoding="utf-8")
+
+    # First pass: find all parents that have sub-entries
+    parents_with_subs = set()
+    for path, _name, _title in generated:
+        if "/" in path:
+            parent = path.split("/")[0]
+            parents_with_subs.add(parent)
+
+    # Build the YAML contents while preserving allowlist order
+    lines = []
+    processed_sections = set()
+
+    for path, _name, title in generated:
+        # Check if this is a parent page that has sub-pages
+        if path in parents_with_subs:
+            # This is a parent page with sub-pages - create a nested section
+            if path not in processed_sections:
+                processed_sections.add(path)
+                section_title = (
+                    title or path.replace("-", " ").replace("_", " ").title()
+                )
+                lines.append(f'                - section: "{section_title}"')
+                lines.append("                  contents:")
+                # Add the parent page first
+                lines.append(f"                    - docs/models/{path}.qmd")
+                # Then add all sub-pages
+                for sub_path, _sub_name, _sub_title in generated:
+                    if "/" in sub_path and sub_path.split("/")[0] == path:
+                        lines.append(
+                            f"                    - docs/models/{sub_path}.qmd"
+                        )
+        elif "/" not in path:
+            # This is a flat item with no sub-pages
+            # Skip if it was already included as part of a parent section
+            if path not in processed_sections:
+                lines.append(f"                - docs/models/{path}.qmd")
+
+    yaml_content = "\n".join(lines) + "\n"
+
+    # Pattern to match only the Model Guides contents, stopping at the next item
+    # in Getting Started (lines starting with 12 spaces: same level as the section)
+    pattern = r'(            - section: "Model Guides"\n              contents:)([^\n]*|.*?)(?=\n            - |\n        - section:|\n\nformat:)'
+
+    def replacement(match):
+        prefix = match.group(1)
+        return prefix + "\n" + yaml_content
+
+    new_content = re.sub(pattern, replacement, content, flags=re.DOTALL)
+
+    if new_content != content:
+        quarto_yml.write_text(new_content, encoding="utf-8")
+        print(f"Updated {quarto_yml}")
+    else:
+        print(f"No changes needed for {quarto_yml}")
+
+
+def main():
+    allow = read_allowlist()
+    if not EXAMPLES_DIR.exists():
+        print(f"[WARN] {EXAMPLES_DIR} not found", file=sys.stderr)
+        return
+
+    (OUTPUT_DIR / "assets").mkdir(parents=True, exist_ok=True)
+
+    # First pass: identify which parents have their own entry vs only sub-entries
+    parent_entries = set()  # Parents that have their own entry
+    parent_with_subs = set()  # Parents that have sub-entries
+    allowlist_entries = set()  # All entries in allowlist
+
+    for item in allow:
+        if isinstance(item, str):
+            name = item
+        else:
+            name = item.get("name")
+
+        allowlist_entries.add(name)
+
+        if "/" in name:
+            parent = name.split("/")[0]
+            parent_with_subs.add(parent)
+        else:
+            parent_entries.add(name)
+
+    # Parents with subs that DON'T have their own entry -> use index.qmd
+    parent_index_only = parent_with_subs - parent_entries
+
+    generated = []
+    seen_dirs = set()  # Track which parent directories we've created index for
+
+    for item in allow:
+        if isinstance(item, str):
+            name = item
+            title = None
+        else:
+            name = item.get("name")
+            title = item.get("title")
+
+        if not name:
+            print(f"[WARN] Skipping item without name: {item}", file=sys.stderr)
+            continue
+
+        src_dir = EXAMPLES_DIR / name
+        if not src_dir.exists() or not src_dir.is_dir():
+            print(f"[WARN] Skipping {name} (not a directory)", file=sys.stderr)
+            continue
+
+        readme = find_readme(src_dir)
+        if not readme:
+            print(f"[WARN] Skipping {name} (no README.md)", file=sys.stderr)
+            continue
+
+        md = readme.read_text(encoding="utf-8")
+
+        # Determine output path first (needed for link rewriting)
+        parts = name.split("/")
+        if len(parts) == 1:
+            # Simple case: no subdirectory
+            out_path = OUTPUT_DIR / f"{parts[0]}.qmd"
+            sidebar_path = parts[0]
+        else:
+            # Has subdirectory: e.g., magistral/think
+            parent = parts[0]
+            child = "-".join(parts[1:])  # handle nested subdirs
+            out_path = OUTPUT_DIR / parent / f"{child}.qmd"
+            sidebar_path = f"{parent}/{child}"
+
+        # Remove the first H1 (we use frontmatter title instead)
+        md, _ = remove_first_h1(md)
+        # Rewrite links between README files
+        md = rewrite_readme_links(
+            md,
+            src_dir,
+            EXAMPLES_DIR,
+            parent_index_only,
+            name,
+            allowlist_entries,
+            sidebar_path,
+        )
+        md = rewrite_and_copy_assets(md, src_dir, OUTPUT_DIR)
+
+        # Handle parent page generation for sub-entries
+        if len(parts) > 1:
+            # Has subdirectory: e.g., magistral/think
+            parent = parts[0]
+
+            # Create parent.qmd if not already done and parent doesn't have own entry
+            if parent not in seen_dirs and parent in parent_index_only:
+                parent_readme = find_readme(EXAMPLES_DIR / parent)
+                if parent_readme:
+                    parent_md = parent_readme.read_text(encoding="utf-8")
+                    parent_md, _ = remove_first_h1(parent_md)
+                    parent_md = rewrite_readme_links(
+                        parent_md,
+                        EXAMPLES_DIR / parent,
+                        EXAMPLES_DIR,
+                        parent_index_only,
+                        parent,
+                        allowlist_entries,
+                        parent,
+                    )
+                    parent_md = rewrite_and_copy_assets(
+                        parent_md, EXAMPLES_DIR / parent, OUTPUT_DIR
+                    )
+                    parent_title = parent.replace("-", " ").replace("_", " ").title()
+                    write_qmd(OUTPUT_DIR / f"{parent}.qmd", parent_title, parent_md)
+                    generated.append((parent, parent, parent_title))
+                    seen_dirs.add(parent)
+
+        if not title:
+            title = name.replace("/", " ").replace("-", " ").title()
+
+        write_qmd(out_path, title, md)
+        generated.append((sidebar_path, name, title))
+
+    # Index page - preserve allowlist order
+    if generated:
+        listing = "\n".join(
+            [f"- [{title}]({path}.qmd)" for path, name, title in generated]
+        )
+        index_md = (
+            "# Model Guides\n\nBelow are the curated examples for training various model architectures:\n\n"
+            + listing
+            + "\n"
+        )
+        index_fm = (
+            "---\nexecute:\n  eval: false\nformat:\n  html:\n    toc: true\n---\n\n"
+        )
+        (OUTPUT_DIR / "index.qmd").write_text(index_fm + index_md, encoding="utf-8")
+
+        # Auto-update _quarto.yml to keep sidebar in sync
+        update_quarto_yml(generated)
+
+
+if __name__ == "__main__":
+    main()

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@5eff953\""
+    "!pip install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@318b7e2\""
    ]
   },
   {
@@ -253,7 +253,6 @@
    "source": [
     "from axolotl.utils import set_pytorch_cuda_alloc_conf\n",
     "\n",
-    "# Set \"PYTORCH_CUDA_ALLOC_CONF\" env to save memory\n",
     "set_pytorch_cuda_alloc_conf()"
    ]
   },

examples/devstral/devstral-small-qlora.yml

@@ -52,6 +52,7 @@ gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
+scaling_softmax: true
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

examples/gemma3/gemma-3-1b-qlora.yml

@@ -1,6 +1,7 @@
 base_model: google/gemma-3-1b-it
 
 model_type: Gemma3ForCausalLM
+cls_model_config: Gemma3TextConfig
 
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
@@ -29,7 +30,7 @@ output_dir: ./outputs/out
 adapter: qlora
 lora_r: 32
 lora_alpha: 16
-lora_dropout: 0.05
+lora_dropout: 0
 lora_target_linear: true
 
 sequence_len: 2048

examples/gemma3/gemma-3-270m-qlora.yml

@@ -1,6 +1,7 @@
 base_model: google/gemma-3-270m-it
 
 model_type: Gemma3ForCausalLM
+cls_model_config: Gemma3TextConfig
 
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
@@ -29,7 +30,7 @@ output_dir: ./outputs/out
 adapter: qlora
 lora_r: 32
 lora_alpha: 16
-lora_dropout: 0.05
+lora_dropout: 0
 lora_target_linear: true
 
 sequence_len: 2048

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

@@ -2,6 +2,7 @@ base_model: google/gemma-3-4b-it
 
 # Need to set else transformers tries to load vision too
 model_type: Gemma3ForCausalLM
+cls_model_config: Gemma3TextConfig
 
 load_in_4bit: true
 
@@ -32,8 +33,8 @@ sample_packing: true
 
 lora_r: 32
 lora_alpha: 16
-lora_dropout: 0.05
-lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+lora_dropout: 0
+lora_target_linear: true
 
 wandb_project:
 wandb_entity:

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

@@ -31,7 +31,7 @@ pad_to_sequence_len: false
 
 lora_r: 32
 lora_alpha: 16
-lora_dropout: 0.05
+lora_dropout: 0
 lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
 
 wandb_project:

examples/gpt-oss/gpt-oss-120b-fft-fsdp2-offload.yaml

@@ -32,6 +32,10 @@ wandb_watch:
 wandb_name:
 wandb_log_model:
 
+trackio_project_name:
+trackio_run_name:
+trackio_space_id:
+
 gradient_accumulation_steps: 2
 micro_batch_size: 1
 num_epochs: 1

examples/gpt-oss/gpt-oss-20b-fft-deepspeed-zero3.yaml

@@ -28,6 +28,10 @@ wandb_watch:
 wandb_name:
 wandb_log_model:
 
+trackio_project_name:
+trackio_run_name:
+trackio_space_id:
+
 gradient_accumulation_steps: 2
 micro_batch_size: 1
 num_epochs: 1

examples/gpt-oss/gpt-oss-20b-fft-fsdp2-offload.yaml

@@ -29,6 +29,10 @@ wandb_watch:
 wandb_name:
 wandb_log_model:
 
+trackio_project_name:
+trackio_run_name:
+trackio_space_id:
+
 gradient_accumulation_steps: 2
 micro_batch_size: 1
 num_epochs: 1

examples/gpt-oss/gpt-oss-20b-fft-fsdp2.yaml

@@ -28,6 +28,10 @@ wandb_watch:
 wandb_name:
 wandb_log_model:
 
+trackio_project_name:
+trackio_run_name:
+trackio_space_id:
+
 gradient_accumulation_steps: 2
 micro_batch_size: 1
 num_epochs: 1

examples/gpt-oss/gpt-oss-20b-sft-lora-singlegpu.yaml

@@ -41,6 +41,10 @@ wandb_watch:
 wandb_name:
 wandb_log_model:
 
+trackio_project_name:
+trackio_run_name:
+trackio_space_id:
+
 gradient_accumulation_steps: 8
 micro_batch_size: 1
 num_epochs: 1

examples/gpt-oss/gpt-oss-safeguard-20b-sft-lora-singlegpu.yaml

@@ -41,6 +41,10 @@ wandb_watch:
 wandb_name:
 wandb_log_model:
 
+trackio_project_name:
+trackio_run_name:
+trackio_space_id:
+
 gradient_accumulation_steps: 8
 micro_batch_size: 1
 num_epochs: 1

examples/internvl3_5/README.md

@@ -0,0 +1,43 @@
+# Finetune OpenGV's InternVL with Axolotl
+
+[InternVL 3.5](https://huggingface.co/OpenGVLab/InternVL3_5-8B-HF) is a family of powerful vision-language models supporting dynamic resolution and multi-image understanding by OpenGV. It features a ViT-style vision encoder and strong language model backbone for tasks like visual question answering, OCR, and scene text understanding.
+
+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 `timm` for vision model support:
+
+    ```bash
+    pip install timm==1.0.19
+    ```
+
+3. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage.
+
+4. Run the finetuning example:
+
+    ```bash
+    axolotl train examples/internvl3_5/internvl3_5-8b-qlora.yml
+    ```
+
+This config uses about 8.21 GiB VRAM. Let us know how it goes. Happy finetuning! 🚀
+
+### Tips
+
+- You can run a full finetuning by removing the `adapter: qlora` and `load_in_4bit: true` from the config.
+- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+- The dataset format follows the multi-modal format as seen [here](https://docs.axolotl.ai/docs/multimodal.html#dataset-format).
+
+## Optimization Guides
+
+Please check the [Optimizations doc](https://docs.axolotl.ai/docs/optimizations.html).
+
+## Related Resources
+
+- [InternVL Paper](https://huggingface.co/papers/2508.18265)
+- [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/internvl3_5/internvl3_5-8b-qlora.yml

@@ -0,0 +1,61 @@
+base_model: OpenGVLab/InternVL3_5-8B-HF
+processor_type: AutoProcessor
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_4bit: true
+
+# these 3 lines are needed for now to handle vision chat templates w images
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+    field_messages: messages
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./outputs/out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: true
+fp16:
+tf32: true
+
+gradient_checkpointing: true
+logging_steps: 1
+flash_attention: true
+eager_attention:
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/kimi-linear/README.md

@@ -0,0 +1,47 @@
+# Finetune MoonshotAI's Kimi Linear with Axolotl
+
+[Kimi Linear](https://huggingface.co/collections/moonshotai/kimi-linear-a3b) is a MoE model (48B total, 3B active) by MoonshotAI using a hybrid linear attention architecture to achieve a 1M token context length. It uses Kimi Delta Attention (KDA), a refined version of Gated DeltaNet that reduces KV cache size by up to 75% and boosts decoding throughput by up to 6x for long contexts.
+
+This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
+
+**Note:** Axolotl uses experimental training code for Kimi Linear as their original modeling code is inference-only.
+
+## Getting started
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
+
+2. Install CCE via [docs](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy)
+
+3. Run the finetuning example:
+
+    ```bash
+    axolotl train examples/kimi-linear/kimi-48b-lora.yaml
+    ```
+
+This config uses about 98.7GiB VRAM.
+
+Let us know how it goes. Happy finetuning!
+
+### TIPS
+
+- Kimi Linear requires `trust_remote_code: true`.
+- You can run a full finetuning by removing the `adapter: lora` and `load_in_8bit: true`.
+- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html)
+- The dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template)
+
+## Optimization Guides
+
+See 👉 [docs](https://docs.axolotl.ai/docs/optimizations.html).
+
+## Limitations
+
+This is not yet compatible with MoE kernels from transformers v5.
+
+## Related Resources
+
+- [Kimi Linear Paper](https://huggingface.co/papers/2510.26692)
+- [Kimi Linear GitHub](https://github.com/MoonshotAI/Kimi-Linear)
+- [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/kimi-linear/kimi-48b-lora.yaml

@@ -0,0 +1,81 @@
+base_model: moonshotai/Kimi-Linear-48B-A3B-Instruct
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+trust_remote_code: true
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+    split: train
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.2
+output_dir: ./outputs/lora-out
+
+adapter: lora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+pad_to_sequence_len: true
+
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0.05
+lora_fan_in_fan_out:
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+flash_attention: true
+
+loss_watchdog_threshold: 5.0
+loss_watchdog_patience: 3
+
+warmup_ratio: 0.1
+evals_per_epoch: 2
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:

examples/llama-3/3b-fp8-fsdp2.yaml

@@ -29,7 +29,6 @@ flex_attention: true
 flex_attn_compile_kwargs:
   dynamic: false
   mode: max-autotune-no-cudagraphs
-save_strategy: no
 torch_compile: true
 
 wandb_project:

examples/magistral/README.md

@@ -13,7 +13,7 @@ Thanks to the team at MistralAI for giving us early access to prepare for these
     Here is an example of how to install from pip:
 
 ```bash
-# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
+# Ensure you have Pytorch installed (Pytorch 2.7.0 min)
 pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```

examples/magistral/think/README.md

@@ -5,6 +5,7 @@ This guide covers fine-tuning [Magistral Small 2507](https://huggingface.co/mist
 ## Prerequisites
 
 Before starting, ensure you have:
+
 - Installed Axolotl (see [main README](../README.md))
 
 ## Getting Started

examples/magistral/vision/README.md

@@ -5,7 +5,8 @@ This guide covers fine-tuning [Magistral Small 2509](https://huggingface.co/mist
 ## Prerequisites
 
 Before starting, ensure you have:
-- Installed Axolotl from source (see [main README](../README.md#getting-started))
+
+- Installed Axolotl from source (see [main README](../README.md))
 
 ## Getting started
 

examples/mimo/README.md

@@ -0,0 +1,39 @@
+# Finetune Xiaomi's MiMo with Axolotl
+
+[MiMo](https://huggingface.co/XiaomiMiMo/MiMo-7B-RL) is a family of models trained from scratch for reasoning tasks, incorporating **Multiple-Token Prediction (MTP)** as an additional training objective for enhanced performance and faster inference. Pre-trained on ~25T tokens with a three-stage data mixture strategy and optimized reasoning pattern density.
+
+This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
+
+## Getting started
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
+
+2. Run the finetuning example:
+
+    ```bash
+    axolotl train examples/mimo/mimo-7b-qlora.yaml
+    ```
+
+This config uses about 17.2 GiB VRAM. Let us know how it goes. Happy finetuning! 🚀
+
+### Tips
+
+- You can run a full finetuning by removing the `adapter: qlora` and `load_in_4bit: true` from the config.
+- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+- The dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+
+## Optimization Guides
+
+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 MiMo in the near future.
+
+## Related Resources
+
+- [MiMo Paper](https://arxiv.org/abs/2505.07608)
+- [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/mimo/mimo-7b-qlora.yaml

@@ -0,0 +1,67 @@
+base_model: XiaomiMiMo/MiMo-7B-RL
+trust_remote_code: true
+revision_of_model: 6299b5a
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+# CCE - N/A as of now
+# plugins:
+#   - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/lora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/ministral/README.md

@@ -0,0 +1,50 @@
+# Finetune Ministral with Axolotl
+
+Ministral is a family of openweight models from MistralAI found on [HuggingFace](mistralai/Ministral-8B-Instruct-2410). This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
+
+## Getting started
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
+
+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/ministral/ministral-small-qlora.yaml
+    ```
+
+This config uses about 8.76 GiB VRAM.
+
+Let us know how it goes. Happy finetuning! 🚀
+
+### Tips
+
+- We recommend adding the same/similar SystemPrompt that the model is tuned for. You can find this within the repo's files titled `SYSTEM_PROMPT.txt`.
+- You can run a full finetuning by removing the `adapter: qlora` and `load_in_4bit: true` from the config.
+- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+- The text dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+
+## Optimization Guides
+
+Please check the [Optimizations doc](https://docs.axolotl.ai/docs/optimizations.html).
+
+## Limitations
+
+We only support the `mistral-common` tokenizer for Supervised Fine-tuning at the moment and for `type: chat_template` only.
+
+In addition, we do not support overriding tokens yet.
+
+## Related Resources
+
+- [MistralAI Ministral Blog](https://mistral.ai/news/ministraux)
+- [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)
+
+
+## Future Work
+
+- Add parity to Preference Tuning, RL, etc.
+- Add parity to other tokenizer configs like overriding tokens.

examples/ministral/ministral-small-qlora.yaml

@@ -0,0 +1,67 @@
+base_model: mistralai/Ministral-8B-Instruct-2410
+
+# Enable to use mistral-common tokenizer
+tokenizer_use_mistral_common: true
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/lora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/ministral3/README.md

@@ -0,0 +1,79 @@
+# Finetune Ministral3 with Axolotl
+
+Ministral3 is a family of open-weight models from MistralAI found on [HuggingFace](https://huggingface.co/collections/mistralai/ministral-3). This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
+
+Please see [Thinking](#thinking) and [Vision](#vision) for their respective fine-tuning.
+
+Thanks to the team at MistralAI for giving us early access to prepare for these releases.
+
+Note: This is still experimental given it is based on transformers v5 RC.
+
+## 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. Swap to the Axolotl transformers v5 branch
+
+    ```bash
+    cp examples/ministral3/ministral3-3b-qlora.yaml ministral3-3b-qlora.yaml
+
+    git fetch
+    git checkout transformers-v5
+
+    # Install packages for transformers v5
+    pip install -e .
+    ```
+
+4. Run the fine-tuning:
+
+    ```bash
+    axolotl train ministral3-3b-qlora.yaml
+    ```
+
+Let us know how it goes. Happy finetuning! 🚀
+
+
+### Tips
+
+- We recommend adding the same/similar SystemPrompt that the model is tuned for. You can find this within the repo's files titled `SYSTEM_PROMPT.txt`.
+- You can run a full finetuning by removing the `adapter: qlora` and `load_in_4bit: true` from the config.
+- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+- The text dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+
+### Thinking
+
+Ministral3 2512 model supports thinking capabilities, enabling Chain-of-Thought reasoning with explicit thinking steps.
+
+📚 **[See the Thinking fine-tuning guide →](./think/README.md)**
+
+### Vision
+
+Ministral3 2512 model also supports vision capabilities.
+
+📚 **[See the Vision fine-tuning guide →](./vision/README.md)**
+
+## Optimization Guides
+
+Please check the [Optimizations doc](https://docs.axolotl.ai/docs/optimizations.html).
+
+## Limitations
+
+We only support the `mistral-common` tokenizer for Supervised Fine-tuning at the moment and for `type: chat_template` only.
+
+In addition, we do not support overriding tokens yet.
+
+## Related Resources
+
+- [MistralAI Mistral3 Blog](https://mistral.ai/news/mistral-3)
+- [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)
+
+
+## Future Work
+
+- Add parity to Preference Tuning, RL, etc.
+- Add parity to other tokenizer configs like overriding tokens.

examples/ministral3/ministral3-3b-qlora.yaml

@@ -0,0 +1,68 @@
+base_model: mistralai/Ministral-3-3B-Reasoning-2512
+
+# Enable to use mistral-common tokenizer
+tokenizer_use_mistral_common: true
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/lora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+scaling_softmax: 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/ministral3/think/README.md

@@ -0,0 +1,74 @@
+# Ministral3 2512 Thinking Fine-tuning
+
+This guide covers fine-tuning [Ministral3 2512](https://huggingface.co/collections/mistralai/ministral-3) with thinking capabilities using Axolotl. The thinking model enables explicit Chain-of-Thought reasoning with separate thinking and response sections.
+
+## Prerequisites
+
+Before starting, ensure you have:
+
+- Installed Axolotl (see [main README](../README.md))
+
+## Getting Started
+
+Run the thinking model fine-tuning:
+
+```bash
+axolotl train examples/ministral3/think/ministral3-3b-think-qlora.yaml
+```
+
+This config uses about 4.76 GiB VRAM.
+
+### Tips
+
+- Dataset uses multi-content format with `type: thinking` support. See [Dataset Format](#dataset-format) below.
+- You cannot mix `content: str` and `content: list[dict]`, otherwise, dataset loading will fail. Keep it consistent.
+
+## Dataset Format
+
+The thinking model requires the multi-content dataset format with support for an extra `role: thinking` within system and assistant messages.
+
+Example format:
+
+```json
+{
+    "messages": [
+        {
+            "role": "system",
+            "content": [
+                { "type": "text", "text": "{SYSTEM_PROMPT}"}
+            ]
+        },
+        {
+            "role": "user",
+            "content": [
+                { "type": "text", "text": "Solve this step by step: What is 15% of 240?"}
+            ]
+        },
+        {
+            "role": "assistant",
+            "content": [
+                {
+                    "type": "thinking",
+                    "thinking": "I need to calculate 15% of 240. First, I'll convert 15% to decimal: 0.15. Then multiply: 0.15 × 240 = 36."
+                },
+                {
+                    "type": "text",
+                    "text": "To find 15% of 240, I'll multiply 240 by 0.15:\n\n240 × 0.15 = 36\n\nTherefore, 15% of 240 is 36."
+                }
+            ]
+        }
+    ]
+}
+```
+
+### Advanced Options
+
+The `thinking` section supports an optional `closed` parameter:
+
+```json
+{
+    "type": "thinking",
+    "thinking": "Internal reasoning here...",
+    "closed": true  // Default: true, controls adding the closing [/THINK] tag
+}
+```

examples/ministral3/think/ministral3-3b-think-qlora.yaml

@@ -0,0 +1,67 @@
+base_model: mistralai/Ministral-3-3B-Reasoning-2512
+
+# Enable to use mistral-common tokenizer
+tokenizer_use_mistral_common: true
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: Nanobit/text-think-2k-test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0
+output_dir: ./outputs/lora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/ministral3/vision/README.md

@@ -0,0 +1,58 @@
+# Ministral3 2512 Vision Fine-tuning
+
+This guide covers fine-tuning [Ministral3 2512](https://huggingface.co/collections/mistralai/ministral-3) with vision capabilities using Axolotl.
+
+## Prerequisites
+
+Before starting, ensure you have:
+
+- Installed Axolotl from source (see [main README](../README.md))
+
+## Getting started
+
+1. Install the required vision lib:
+    ```bash
+    pip install 'mistral-common[opencv]==1.8.6'
+    ```
+
+2. Download the example dataset image:
+   ```bash
+   wget https://huggingface.co/datasets/Nanobit/text-vision-2k-test/resolve/main/African_elephant.jpg
+   ```
+
+3. Run the fine-tuning:
+   ```bash
+   axolotl train examples/ministral3/vision/ministral3-3b-vision-qlora.yml
+   ```
+
+WARNING: The loss and grad norm will be much higher than normal at first. We suspect this to be inherent to the model as of the moment. If anyone would like to submit a fix for this, we are happy to take a look.
+
+### Tips
+
+Key differences from text-only model:
+- Multi-modal dataset format required
+- Sample packing not supported
+
+## Dataset Format
+
+The vision model requires multi-modal dataset format as documented [here](https://docs.axolotl.ai/docs/multimodal.html#dataset-format).
+
+One exception is that, passing `"image": PIL.Image` is not supported. MistralTokenizer only supports `path`, `url`, and `base64` for now.
+
+Example:
+```json
+{
+    "messages": [
+        {"role": "system", "content": [{ "type": "text", "text": "{SYSTEM_PROMPT}"}]},
+        {"role": "user", "content": [
+            { "type": "text", "text": "What's in this image?"},
+            {"type": "image", "path": "path/to/image.jpg" }
+        ]},
+        {"role": "assistant", "content": [{ "type": "text", "text": "..." }]},
+    ],
+}
+```
+
+## Limitations
+
+- Sample Packing is not supported for multi-modality training currently.

examples/ministral3/vision/ministral3-3b-vision-qlora.yml

@@ -0,0 +1,64 @@
+base_model: mistralai/Ministral-3-3B-Reasoning-2512
+processor_type: AutoProcessor
+
+# Enable to use mistral-common tokenizer
+tokenizer_use_mistral_common: true
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_4bit: true
+
+# these 3 lines are needed for now to handle vision chat templates w images
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+# sample dataset below requires downloading image in advance
+# wget https://huggingface.co/datasets/Nanobit/text-vision-2k-test/resolve/main/African_elephant.jpg
+datasets:
+  - path: Nanobit/text-vision-2k-test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./outputs/out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: true
+fp16:
+tf32: true
+
+gradient_checkpointing: true
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/mistral-small/README.md

@@ -5,6 +5,7 @@ This guide covers fine-tuning [Mistral Small 3.1](mistralai/Mistral-Small-3.1-24
 ## Prerequisites
 
 Before starting, ensure you have:
+
 - Installed Axolotl (see [Installation docs](https://docs.axolotl.ai/docs/installation.html))
 
 ## Getting Started

examples/olmo3/README.md

@@ -6,25 +6,17 @@ 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).
+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:
 
-    Here is an example of how to install from pip:
     ```bash
-    # Ensure you have a compatible version of Pytorch installed
-    pip3 install packaging setuptools wheel ninja
-    pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
-
-    # Install Cut Cross Entropy
-    python scripts/cutcrossentropy_install.py | sh
+    axolotl train examples/olmo3/olmo3-7b-qlora.yaml
     ```
 
-2. Run the finetuning example:
-
-```bash
-axolotl train examples/olmo3/olmo3-7b-qlora.yaml
-```
-
-Let us know how it goes. Happy finetuning! 🚀
+This uses about 11.3 GiB VRAM. Let us know how it goes. Happy finetuning! 🚀
 
 ### TIPS
 

examples/olmo3/olmo3-7b-qlora.yaml

@@ -42,10 +42,10 @@ wandb_watch:
 wandb_name:
 wandb_log_model:
 
-gradient_accumulation_steps: 4
+gradient_accumulation_steps: 2
 micro_batch_size: 2
 num_epochs: 1
-optimizer: adamw_bnb_8bit
+optimizer: adamw_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 

examples/plano/README.md

@@ -0,0 +1,42 @@
+# Finetune Katanemo's Plano-Orchestrator with Axolotl
+
+[Plano-Orchestrator](https://huggingface.co/collections/katanemo/plano-orchestrator) is a family of 4B and 30B-A3B routing and orchestration models designed for multi-agent systems. It analyzes user intent and conversation context to make precise routing decisions, excelling at multi-turn context understanding, multi-intent detection, and context-dependent routing.
+
+This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
+
+## Getting started
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
+
+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/plano/plano-4b-qlora.yaml
+    ```
+
+This config uses about 5.1 GiB VRAM. Let us know how it goes. Happy finetuning! 🚀
+
+### Orchestration Prompt
+
+Plano-Orchestrator uses a specific orchestration prompt format for routing/agent decisions. Please check the [official model card](https://huggingface.co/katanemo/Plano-Orchestrator-4B) for proper prompt formatting and the `ORCHESTRATION_PROMPT` template.
+
+### Tips
+
+- To use the larger [Plano-Orchestrator-30B-A3B](https://huggingface.co/katanemo/Plano-Orchestrator-30B-A3B) MoE model, simply change `base_model: katanemo/Plano-Orchestrator-30B-A3B` in the config and enable multi-GPU training if needed.
+- You can run a full finetuning by removing the `adapter: qlora` and `load_in_4bit: true` from the config.
+- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+- The dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+
+## Optimization Guides
+
+Please check the [Optimizations doc](https://docs.axolotl.ai/docs/optimizations.html).
+
+## Related Resources
+
+- [Plano GitHub](https://github.com/katanemo/plano)
+- [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/plano/plano-4b-qlora.yaml

@@ -0,0 +1,65 @@
+base_model: katanemo/Plano-Orchestrator-4B
+
+# 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
+
+chat_template: qwen3
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/lora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qat_nvfp4/Gemma3-12B_baseline.yml

@@ -0,0 +1,67 @@
+base_model: google/gemma-3-12b-it
+# 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
+seed: 42
+chat_template: gemma3
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+
+output_dir: ./outputs/out_gemma/
+
+sequence_len: 8096
+sample_packing: true
+flash_attention: true
+
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 16
+
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 4e-5
+
+bf16: true
+tf32: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+# evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+fsdp_version: 2
+
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true
+
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qat_nvfp4/Gemma3-12B_qat.yml

@@ -0,0 +1,72 @@
+base_model: google/gemma-3-12b-it
+# 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
+seed: 42
+chat_template: gemma3
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+
+output_dir: ./outputs/qat_out_gemma/
+
+sequence_len: 8096
+sample_packing: true
+flash_attention: true
+
+qat:
+  activation_dtype: nvfp4
+  weight_dtype: nvfp4
+  group_size: 16 # only group_size of 16 is supported with nvfp4
+
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 16
+
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 4e-5
+
+bf16: true
+tf32: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+fsdp_version: 2
+
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true
+
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qat_nvfp4/Math-Gemma3-12B_baseline.yml

@@ -0,0 +1,67 @@
+base_model: google/gemma-3-12b-it
+# Math finetuning configuration for Gemma3-12B
+# 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
+seed: 42
+chat_template: gemma3
+datasets:
+  - path: AI-MO/NuminaMath-CoT
+    type: chat_template
+
+output_dir: ./outputs/out_math_gemma/
+
+sequence_len: 4096
+sample_packing: true
+flash_attention: true
+
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 8
+
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 3e-5
+
+bf16: true
+tf32: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+# evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+fsdp_version: 2
+
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true
+
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qat_nvfp4/Math-Gemma3-12B_qat.yml

@@ -0,0 +1,72 @@
+base_model: google/gemma-3-12b-it
+# Math finetuning configuration for Gemma3-12B
+# 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
+seed: 42
+chat_template: gemma3
+datasets:
+  - path: AI-MO/NuminaMath-CoT
+    type: chat_template
+
+output_dir: ./outputs/qat_out_math_gemma/
+
+sequence_len: 4096
+sample_packing: true
+flash_attention: true
+
+qat:
+  activation_dtype: nvfp4
+  weight_dtype: nvfp4
+  group_size: 16 # only group_size of 16 is supported with nvfp4
+
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 8
+
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 3e-5
+
+bf16: true
+tf32: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+# evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+fsdp_version: 2
+
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true
+
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qat_nvfp4/Math-Gemma3-27B_baseline.yml

@@ -0,0 +1,68 @@
+base_model: google/gemma-3-27b-it
+# Math finetuning configuration for Gemma3-27B
+# 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
+seed: 42
+chat_template: gemma3
+datasets:
+  - path: AI-MO/NuminaMath-CoT
+    type: chat_template
+
+output_dir: ./outputs/out_math_gemma27/
+
+sequence_len: 4096
+sample_packing: true
+flash_attention: true
+
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 16
+
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 5e-6
+eta_min: 7e-7
+
+bf16: true
+tf32: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+# evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+fsdp_version: 2
+
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true
+
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qat_nvfp4/Math-Gemma3-27B_qat.yml

@@ -0,0 +1,73 @@
+base_model: google/gemma-3-27b-it
+# Math finetuning configuration for Gemma3-27B
+# 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
+seed: 42
+chat_template: gemma3
+datasets:
+  - path: AI-MO/NuminaMath-CoT
+    type: chat_template
+
+output_dir: ./outputs/qat_out_math_gemma27/
+
+sequence_len: 4096
+sample_packing: true
+flash_attention: true
+
+qat:
+  activation_dtype: nvfp4
+  weight_dtype: nvfp4
+  group_size: 16 # only group_size of 16 is supported with nvfp4
+
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 16
+
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 5e-6
+eta_min: 7e-7
+
+bf16: true
+tf32: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+# evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+fsdp_version: 2
+
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true
+
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qat_nvfp4/Math-Qwen2.5-72B_baseline.yml

@@ -0,0 +1,67 @@
+base_model: Qwen/Qwen2.5-72B
+# Math finetuning configuration for Qwen2.5-72B (non-instruct)
+# 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
+seed: 42
+chat_template: qwen_25
+datasets:
+  - path: AI-MO/NuminaMath-CoT
+    type: chat_template
+
+output_dir: ./outputs/out_math_72b/
+
+sequence_len: 4096
+sample_packing: true
+flash_attention: true
+
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 8
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 5e-6
+eta_min: 7e-7
+
+bf16: true
+tf32: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+# evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+fsdp_version: 2
+
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Qwen2DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true
+
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qat_nvfp4/Math-Qwen2.5-72B_qat.yml

@@ -0,0 +1,72 @@
+base_model: Qwen/Qwen2.5-72B
+# Math finetuning configuration for Qwen2.5-72B (non-instruct)
+# 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
+seed: 42
+chat_template: qwen_25
+datasets:
+  - path: AI-MO/NuminaMath-CoT
+    type: chat_template
+
+output_dir: ./outputs/qat_out_math_72b/
+
+sequence_len: 4096
+sample_packing: true
+flash_attention: true
+
+qat:
+  activation_dtype: nvfp4
+  weight_dtype: nvfp4
+  group_size: 16 # only group_size of 16 is supported with nvfp4
+
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 8
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 5e-6
+eta_min: 7e-7
+
+bf16: true
+tf32: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+# evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+fsdp_version: 2
+
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Qwen2DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true
+
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qat_nvfp4/Qwen2.5-72B_baseline.yml

@@ -0,0 +1,67 @@
+base_model: Qwen/Qwen2.5-72B
+# Alpaca finetuning configuration for Qwen2.5-72B
+# 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
+seed: 42
+chat_template: qwen_25
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+
+output_dir: ./outputs/out_qwen72b/
+
+sequence_len: 8096
+sample_packing: true
+flash_attention: true
+
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 16
+
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+bf16: true
+tf32: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+# evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+fsdp_version: 2
+
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Qwen2DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true
+
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qat_nvfp4/Qwen2.5-72B_qat.yml

@@ -0,0 +1,72 @@
+base_model: Qwen/Qwen2.5-72B
+# Alpaca finetuning configuration for Qwen2.5-72B
+# 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
+seed: 42
+chat_template: qwen_25
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+
+output_dir: ./outputs/qat_out_qwen72b/
+
+sequence_len: 8096
+sample_packing: true
+flash_attention: true
+
+qat:
+  activation_dtype: nvfp4
+  weight_dtype: nvfp4
+  group_size: 16 # only group_size of 16 is supported with nvfp4
+
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 16
+
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+bf16: true
+tf32: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+# evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+fsdp_version: 2
+
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Qwen2DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true
+
+special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/qwen2/adamw-pretrain-fsdp2.yaml

@@ -0,0 +1,70 @@
+base_model: Qwen/Qwen2.5-0.5B
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+
+# Use random initialization for fair comparison
+reinit_weights: true
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+# Pretraining dataset
+pretraining_dataset:
+  - path: allenai/c4
+    name: en
+    type: pretrain
+    split: train
+
+dataset_prepared_path:
+val_set_size: 0.0
+output_dir: ./outputs/compare-adamw-pretrain
+
+sequence_len: 2048
+sample_packing: true
+pad_to_sequence_len: true
+
+wandb_project: dist_muon
+wandb_entity:
+wandb_watch:
+wandb_name: adamw
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 4
+num_epochs: 1
+max_steps: 305
+
+# AdamW optimizer settings (standard LR for AdamW)
+optimizer: adamw_torch_fused
+learning_rate: 0.0002
+weight_decay: 0.01
+lr_scheduler: cosine
+
+train_on_inputs: true
+group_by_length: false
+bf16: auto
+fp16: false
+tf32: false
+
+gradient_checkpointing: false
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 0
+saves_per_epoch: 1
+
+# Reproducibility
+seed: 42
+
+fsdp_config:
+  fsdp_version: 2
+  fsdp_offload_params: false
+  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_transformer_layer_cls_to_wrap: Qwen2DecoderLayer
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_cpu_ram_efficient_loading: false
+  fsdp_reshard_after_forward: true
+
+special_tokens:

examples/qwen2/muon-pretrain-fsdp2.yaml

@@ -0,0 +1,70 @@
+base_model: Qwen/Qwen2.5-0.5B
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+
+# Use random initialization for fair comparison
+reinit_weights: true
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+# Pretraining dataset
+pretraining_dataset:
+  - path: allenai/c4
+    name: en
+    type: pretrain
+    split: train
+
+dataset_prepared_path:
+val_set_size: 0.0
+output_dir: ./outputs/compare-muon-pretrain
+
+sequence_len: 2048
+sample_packing: true
+pad_to_sequence_len: true
+
+wandb_project: dist_muon
+wandb_entity:
+wandb_watch:
+wandb_name: muon
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 4
+num_epochs: 1
+max_steps: 305
+
+# Muon optimizer settings
+optimizer: muon
+learning_rate: 0.02
+weight_decay: 0.01
+lr_scheduler: cosine
+
+train_on_inputs: true
+group_by_length: false
+bf16: auto
+fp16: false
+tf32: false
+
+gradient_checkpointing: false
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 0
+saves_per_epoch: 1
+
+# Reproducibility
+seed: 42
+
+fsdp_config:
+  fsdp_version: 2
+  fsdp_offload_params: false
+  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_transformer_layer_cls_to_wrap: Qwen2DecoderLayer
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_cpu_ram_efficient_loading: false
+  fsdp_reshard_after_forward: true
+
+special_tokens:

examples/qwen3/README.md

@@ -0,0 +1,46 @@
+# Finetune Qwen3 with Axolotl
+
+[Qwen3](https://huggingface.co/collections/Qwen/qwen3) are a family of open source models trained by Alibaba.
+
+This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
+
+## Getting started
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
+
+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/qwen3/32b-qlora.yaml
+    ```
+
+Let us know how it goes. Happy finetuning! 🚀
+
+### Chat template masking a few tokens off
+
+If you notice that the `chat_template` masking for assistant prompts are off by a few tokens, please ensure that you are adding the below to the yaml.
+
+```yaml
+chat_template: qwen3
+```
+
+### TIPS
+
+- For inference, please check the official model card as it depends on your reasoning mode.
+- You can run a full finetuning by removing the `adapter: qlora` and `load_in_4bit: true` from the config.
+- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+- The dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+
+## Optimization Guides
+
+Please check the [Optimizations doc](https://docs.axolotl.ai/docs/optimizations.html).
+
+## Related Resources
+
+- [Qwen3 Blog](https://qwenlm.github.io/blog/qwen3/)
+- [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/swanlab/README.md

@@ -0,0 +1,285 @@
+# SwanLab Integration Examples
+
+This directory contains example configurations demonstrating SwanLab integration with Axolotl.
+
+## Examples Overview
+
+### 1. DPO with Completion Logging
+**File**: `dpo-swanlab-completions.yml`
+
+Demonstrates DPO (Direct Preference Optimization) training with RLHF completion table logging.
+
+**Features**:
+- Basic SwanLab experiment tracking
+- Completion table logging (prompts, chosen/rejected responses, rewards)
+- Memory-bounded buffer for long training runs
+- Cloud sync configuration
+
+**Best for**: RLHF practitioners who want to analyze model outputs qualitatively
+
+**Quick start**:
+```bash
+export SWANLAB_API_KEY=your-api-key
+accelerate launch -m axolotl.cli.train examples/swanlab/dpo-swanlab-completions.yml
+```
+
+---
+
+### 2. LoRA with Performance Profiling
+**File**: `lora-swanlab-profiling.yml`
+
+Demonstrates standard LoRA fine-tuning with performance profiling enabled.
+
+**Features**:
+- SwanLab experiment tracking
+- Automatic profiling of trainer methods
+- Profiling metrics visualization
+- Performance optimization guidance
+
+**Best for**: Engineers optimizing training performance and comparing different configurations
+
+**Quick start**:
+```bash
+export SWANLAB_API_KEY=your-api-key
+accelerate launch -m axolotl.cli.train examples/swanlab/lora-swanlab-profiling.yml
+```
+
+---
+
+### 3. Full-Featured DPO Production Setup
+**File**: `dpo-swanlab-full-featured.yml`
+
+Comprehensive production-ready configuration with ALL SwanLab features enabled.
+
+**Features**:
+- Experiment tracking with team workspace
+- RLHF completion logging
+- Performance profiling
+- Lark (Feishu) team notifications
+- Private deployment support
+- Production checklist and troubleshooting
+
+**Best for**: Production RLHF training with team collaboration
+
+**Quick start**:
+```bash
+export SWANLAB_API_KEY=your-api-key
+export SWANLAB_LARK_WEBHOOK_URL=https://open.feishu.cn/...
+export SWANLAB_LARK_SECRET=your-webhook-secret
+accelerate launch -m axolotl.cli.train examples/swanlab/dpo-swanlab-full-featured.yml
+```
+
+---
+
+### 4. Custom Trainer Profiling (Python)
+**File**: `custom_trainer_profiling.py`
+
+Python code examples showing how to add SwanLab profiling to custom trainers.
+
+**Features**:
+- `@swanlab_profile` decorator examples
+- Context manager profiling for fine-grained timing
+- `ProfilingConfig` for advanced filtering and throttling
+- Multiple profiling patterns and best practices
+
+**Best for**: Advanced users creating custom trainers
+
+**Usage**:
+```python
+from custom_trainer_profiling import CustomTrainerWithProfiling
+# See file for detailed examples and patterns
+```
+
+---
+
+## Feature Matrix
+
+| Example | Tracking | Completion Logging | Profiling | Lark Notifications | Team Workspace |
+|---------|----------|-------------------|-----------|-------------------|----------------|
+| dpo-swanlab-completions.yml | ✅ | ✅ | ✅ (auto) | ➖ (commented) | ➖ (commented) |
+| lora-swanlab-profiling.yml | ✅ | ➖ (disabled) | ✅ (auto) | ➖ (commented) | ➖ (commented) |
+| dpo-swanlab-full-featured.yml | ✅ | ✅ | ✅ (auto) | ✅ | ✅ |
+| custom_trainer_profiling.py | N/A | N/A | ✅ (manual) | N/A | N/A |
+
+---
+
+## Configuration Quick Reference
+
+### Basic SwanLab Setup
+```yaml
+plugins:
+  - axolotl.integrations.swanlab.SwanLabPlugin
+
+use_swanlab: true
+swanlab_project: my-project
+swanlab_experiment_name: my-experiment
+swanlab_mode: cloud  # cloud, local, offline, disabled
+```
+
+### RLHF Completion Logging
+```yaml
+swanlab_log_completions: true
+swanlab_completion_log_interval: 100  # Log every 100 steps
+swanlab_completion_max_buffer: 128    # Memory-bounded buffer
+```
+
+### Lark Team Notifications
+```yaml
+swanlab_lark_webhook_url: https://open.feishu.cn/...
+swanlab_lark_secret: your-webhook-secret  # Required for production
+```
+
+### Team Workspace
+```yaml
+swanlab_workspace: my-research-team
+```
+
+### Private Deployment
+```yaml
+swanlab_web_host: https://swanlab.yourcompany.com
+swanlab_api_host: https://api.swanlab.yourcompany.com
+```
+
+---
+
+## Authentication
+
+### Recommended: Environment Variable
+```bash
+export SWANLAB_API_KEY=your-api-key
+export SWANLAB_LARK_WEBHOOK_URL=https://open.feishu.cn/...
+export SWANLAB_LARK_SECRET=your-webhook-secret
+```
+
+### Alternative: Config File (less secure)
+```yaml
+swanlab_api_key: your-api-key
+swanlab_lark_webhook_url: https://open.feishu.cn/...
+swanlab_lark_secret: your-webhook-secret
+```
+
+---
+
+## Common Use Cases
+
+### Use Case 1: Migrate from WandB to SwanLab
+Start with `lora-swanlab-profiling.yml`, add your model/dataset config, disable WandB:
+```yaml
+use_swanlab: true
+use_wandb: false
+```
+
+### Use Case 2: Analyze DPO Model Outputs
+Use `dpo-swanlab-completions.yml`, adjust completion logging interval based on your training length:
+```yaml
+swanlab_completion_log_interval: 50   # More frequent for short training
+swanlab_completion_log_interval: 200  # Less frequent for long training
+```
+
+### Use Case 3: Optimize Training Performance
+Use `lora-swanlab-profiling.yml`, run multiple experiments with different optimizations:
+- Baseline: `flash_attention: false, gradient_checkpointing: false`
+- Flash Attention: `flash_attention: true`
+- Gradient Checkpointing: `gradient_checkpointing: true`
+- Both: `flash_attention: true, gradient_checkpointing: true`
+
+Compare profiling metrics in SwanLab dashboard.
+
+### Use Case 4: Production RLHF with Team Collaboration
+Use `dpo-swanlab-full-featured.yml`, set up team workspace and Lark notifications:
+```yaml
+swanlab_workspace: ml-team
+swanlab_lark_webhook_url: ...
+swanlab_lark_secret: ...
+```
+
+---
+
+## Viewing Your Experiments
+
+### Cloud Mode
+Visit [https://swanlab.cn](https://swanlab.cn) and navigate to your project.
+
+**Dashboard sections**:
+- **Metrics**: Training loss, learning rate, profiling metrics
+- **Tables**: RLHF completions (for DPO/KTO/ORPO/GRPO)
+- **Config**: Hyperparameters and configuration
+- **System**: Resource usage (GPU, memory, CPU)
+- **Files**: Logged artifacts
+
+### Local Mode
+```bash
+swanlab watch ./swanlog
+# Open browser to http://localhost:5092
+```
+
+---
+
+## Troubleshooting
+
+### SwanLab not initializing
+```bash
+# Check API key
+echo $SWANLAB_API_KEY
+
+# Verify SwanLab is installed
+pip show swanlab
+
+# Check config
+grep -A 5 "use_swanlab" your-config.yml
+```
+
+### Completions not appearing
+- Verify you're using an RLHF trainer (DPO/KTO/ORPO/GRPO)
+- Check `swanlab_log_completions: true`
+- Wait for `swanlab_completion_log_interval` steps
+- Look for "Registered SwanLab RLHF completion logging" in logs
+
+### Lark notifications not working
+- Test webhook manually: `curl -X POST "$SWANLAB_LARK_WEBHOOK_URL" ...`
+- Verify `SWANLAB_LARK_SECRET` is set correctly
+- Check bot is added to Lark group chat
+- Look for "Registered Lark notification callback" in logs
+
+### Profiling metrics not appearing
+- Verify `use_swanlab: true`
+- Check SwanLab is initialized (look for init log message)
+- Profiling metrics are under "profiling/" namespace
+- Profiling auto-enabled when SwanLab is enabled
+
+---
+
+## Performance Notes
+
+### Overhead Comparison
+
+| Feature | Overhead per Step | Memory Usage |
+|---------|------------------|--------------|
+| Basic tracking | < 0.1% | ~10 MB |
+| Completion logging | < 0.5% | ~64 KB (buffer=128) |
+| Profiling | < 0.1% | ~1 KB |
+| **Total** | **< 0.7%** | **~10 MB** |
+
+### Best Practices
+1. Use ONE logging tool in production (disable WandB/MLflow when using SwanLab)
+2. Adjust completion log interval based on training length (100-200 steps)
+3. Keep completion buffer size reasonable (128-512)
+4. Profile critical path methods first (training_step, compute_loss)
+5. Use ProfilingConfig to throttle high-frequency operations
+
+---
+
+## Further Reading
+
+- **Full Documentation**: [src/axolotl/integrations/swanlab/README.md](../../src/axolotl/integrations/swanlab/README.md)
+- **SwanLab Docs**: [https://docs.swanlab.cn](https://docs.swanlab.cn)
+- **Axolotl Docs**: [https://axolotl-ai-cloud.github.io/axolotl/](https://axolotl-ai-cloud.github.io/axolotl/)
+- **DPO Paper**: [Direct Preference Optimization](https://arxiv.org/abs/2305.18290)
+
+---
+
+## Contributing
+
+Found an issue or have an improvement? Please submit a PR or open an issue:
+- [Axolotl Issues](https://github.com/axolotl-ai-cloud/axolotl/issues)
+- [SwanLab Issues](https://github.com/SwanHubX/SwanLab/issues)

examples/swanlab/custom_trainer_profiling.py

@@ -0,0 +1,299 @@
+"""Example: Custom Trainer with SwanLab Profiling
+
+This example demonstrates how to add SwanLab profiling to your custom trainer.
+
+Features:
+- @swanlab_profile decorator for automatic profiling
+- swanlab_profiling_context for fine-grained profiling
+- ProfilingConfig for advanced filtering and throttling
+
+Usage:
+    1. Create your custom trainer extending AxolotlTrainer
+    2. Add @swanlab_profile decorators to methods you want to profile
+    3. Use swanlab_profiling_context for fine-grained profiling within methods
+    4. Enable SwanLab in your config (use_swanlab: true)
+
+See also:
+    - examples/swanlab/lora-swanlab-profiling.yml for config
+    - src/axolotl/integrations/swanlab/profiling.py for implementation
+"""
+
+from axolotl.core.trainers.base import AxolotlTrainer
+from axolotl.integrations.swanlab.profiling import (
+    ProfilingConfig,
+    swanlab_profile,
+    swanlab_profiling_context,
+    swanlab_profiling_context_advanced,
+)
+
+
+class CustomTrainerWithProfiling(AxolotlTrainer):
+    """Custom trainer with SwanLab profiling enabled.
+
+    This trainer demonstrates three profiling patterns:
+    1. Decorator-based profiling (@swanlab_profile)
+    2. Context manager profiling (swanlab_profiling_context)
+    3. Advanced profiling with filtering (ProfilingConfig)
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # Create custom profiling config for high-frequency operations
+        self.fast_op_config = ProfilingConfig(
+            enabled=True,
+            min_duration_ms=0.5,  # Only log if duration > 0.5ms
+            log_interval=50,  # Log every 50th call
+        )
+
+    # ========================================================================
+    # Pattern 1: Decorator-based Profiling
+    # ========================================================================
+    # Best for: Methods you always want to profile
+    # Overhead: ~2-5 microseconds per call (negligible)
+
+    @swanlab_profile
+    def training_step(self, model, inputs):
+        """Main training step - always profile.
+
+        Profiling metric: profiling/Time taken: CustomTrainerWithProfiling.training_step
+        """
+        return super().training_step(model, inputs)
+
+    @swanlab_profile
+    def compute_loss(self, model, inputs, return_outputs=False):
+        """Loss computation - always profile.
+
+        Profiling metric: profiling/Time taken: CustomTrainerWithProfiling.compute_loss
+        """
+        return super().compute_loss(model, inputs, return_outputs)
+
+    @swanlab_profile
+    def prediction_step(self, model, inputs, prediction_loss_only, ignore_keys=None):
+        """Prediction step - always profile.
+
+        Profiling metric: profiling/Time taken: CustomTrainerWithProfiling.prediction_step
+        """
+        return super().prediction_step(model, inputs, prediction_loss_only, ignore_keys)
+
+    # ========================================================================
+    # Pattern 2: Fine-grained Context Manager Profiling
+    # ========================================================================
+    # Best for: Profiling specific code blocks within a method
+    # Use case: When you want to profile forward vs backward separately
+
+    def complex_training_step(self, model, inputs):
+        """Training step with fine-grained profiling.
+
+        Profiling metrics:
+        - profiling/Time taken: CustomTrainerWithProfiling.forward_pass
+        - profiling/Time taken: CustomTrainerWithProfiling.backward_pass
+        - profiling/Time taken: CustomTrainerWithProfiling.optimizer_step
+        """
+        # Profile just the forward pass
+        with swanlab_profiling_context(self, "forward_pass"):
+            outputs = model(**inputs)
+            loss = outputs.loss
+
+        # Profile just the backward pass
+        with swanlab_profiling_context(self, "backward_pass"):
+            loss.backward()
+
+        # Profile optimizer step
+        with swanlab_profiling_context(self, "optimizer_step"):
+            self.optimizer.step()
+            self.optimizer.zero_grad()
+
+        return outputs
+
+    # ========================================================================
+    # Pattern 3: Advanced Profiling with Filtering
+    # ========================================================================
+    # Best for: High-frequency operations where you want to throttle logging
+    # Use case: Methods called 100+ times per step
+
+    def _prepare_inputs(self, inputs):
+        """Prepare inputs - throttled profiling.
+
+        This method is called frequently (once per batch), so we throttle
+        profiling to reduce overhead:
+        - Only log if duration > 0.5ms (skip very fast operations)
+        - Only log every 50th call (reduce logging frequency)
+
+        Profiling metric: profiling/Time taken: CustomTrainerWithProfiling.prepare_inputs
+        """
+        with swanlab_profiling_context_advanced(
+            self, "prepare_inputs", config=self.fast_op_config
+        ):
+            return super()._prepare_inputs(inputs)
+
+    def _prepare_input_for_model(self, input_ids):
+        """Another high-frequency operation - throttled profiling.
+
+        Profiling metric: profiling/Time taken: CustomTrainerWithProfiling.prepare_input_for_model
+        """
+        with swanlab_profiling_context_advanced(
+            self, "prepare_input_for_model", config=self.fast_op_config
+        ):
+            # Your custom input preparation logic
+            return input_ids
+
+    # ========================================================================
+    # Pattern 4: Exception-safe Profiling
+    # ========================================================================
+    # Profiling is exception-safe: duration is logged even if method raises
+
+    @swanlab_profile
+    def potentially_failing_method(self):
+        """This method may raise an exception.
+
+        SwanLab profiling will still log the duration before re-raising.
+        Profiling metric: profiling/Time taken: CustomTrainerWithProfiling.potentially_failing_method
+        """
+        # Do some work
+        result = self._do_risky_computation()
+
+        # If this raises, profiling duration is still logged
+        if result < 0:
+            raise ValueError("Invalid result")
+
+        return result
+
+    def _do_risky_computation(self):
+        """Placeholder for risky computation."""
+        return 42
+
+
+# ============================================================================
+# Advanced Example: Custom ProfilingConfig Per Method
+# ============================================================================
+
+
+class AdvancedProfilingTrainer(AxolotlTrainer):
+    """Trainer with method-specific profiling configurations."""
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # Different profiling configs for different method types
+        self.critical_path_config = ProfilingConfig(
+            enabled=True,
+            min_duration_ms=0.0,  # Log everything on critical path
+            log_interval=1,  # Log every call
+        )
+
+        self.fast_path_config = ProfilingConfig(
+            enabled=True,
+            min_duration_ms=1.0,  # Only log if > 1ms
+            log_interval=100,  # Log every 100th call
+        )
+
+        self.debug_config = ProfilingConfig(
+            enabled=True,
+            min_duration_ms=0.0,  # Log everything
+            log_interval=1,  # Log every call
+        )
+
+    def training_step(self, model, inputs):
+        """Critical path - log everything."""
+        with swanlab_profiling_context_advanced(
+            self, "training_step", config=self.critical_path_config
+        ):
+            return super().training_step(model, inputs)
+
+    def _prepare_inputs(self, inputs):
+        """Fast path - throttle logging."""
+        with swanlab_profiling_context_advanced(
+            self, "prepare_inputs", config=self.fast_path_config
+        ):
+            return super()._prepare_inputs(inputs)
+
+    def _debug_method(self, data):
+        """Debug-only method - verbose logging."""
+        with swanlab_profiling_context_advanced(
+            self, "debug_method", config=self.debug_config
+        ):
+            # Your debug logic
+            pass
+
+
+# ============================================================================
+# How to Use This Custom Trainer
+# ============================================================================
+
+"""
+To use this custom trainer:
+
+1. Save this file to your project (e.g., my_custom_trainer.py)
+
+2. Create a config file that uses your custom trainer:
+
+    # config.yml
+    base_model: NousResearch/Llama-3.2-1B
+
+    # ... other config ...
+
+    plugins:
+      - axolotl.integrations.swanlab.SwanLabPlugin
+
+    use_swanlab: true
+    swanlab_project: my-profiling-experiment
+
+    # Optional: Specify custom trainer
+    # (Or modify axolotl to use your custom trainer class)
+
+3. Run training:
+
+    export SWANLAB_API_KEY=your-api-key
+    accelerate launch -m axolotl.cli.train config.yml
+
+4. View profiling metrics in SwanLab dashboard:
+   - profiling/Time taken: CustomTrainerWithProfiling.training_step
+   - profiling/Time taken: CustomTrainerWithProfiling.forward_pass
+   - profiling/Time taken: CustomTrainerWithProfiling.backward_pass
+   - etc.
+
+5. Compare profiling metrics across runs:
+   - Run baseline without optimizations
+   - Run with flash_attention enabled
+   - Run with gradient_checkpointing enabled
+   - Compare profiling metrics to see performance impact
+"""
+
+# ============================================================================
+# Tips for Effective Profiling
+# ============================================================================
+
+"""
+1. Profile the critical path first:
+   - training_step, compute_loss, prediction_step
+   - These methods are called most frequently and have biggest impact
+
+2. Use throttling for high-frequency operations:
+   - Methods called 100+ times per step
+   - Use log_interval=50 or log_interval=100
+   - Reduces profiling overhead and dashboard clutter
+
+3. Filter noise with min_duration_ms:
+   - Set min_duration_ms=1.0 to skip very fast operations
+   - Focus on operations that actually take time
+
+4. Compare across runs:
+   - Run same config multiple times to check consistency
+   - Compare different optimization strategies
+   - Track profiling trends over time
+
+5. Monitor distributed training:
+   - Check for per-rank timing differences
+   - Look for stragglers (slower ranks)
+   - Identify synchronization bottlenecks
+
+6. Disable profiling in production:
+   - from axolotl.integrations.swanlab.profiling import DEFAULT_PROFILING_CONFIG
+   - DEFAULT_PROFILING_CONFIG.enabled = False
+
+7. Exception handling:
+   - Profiling is exception-safe
+   - Duration logged even if method raises
+   - Useful for debugging methods that fail intermittently
+"""

examples/swanlab/dpo-swanlab-completions.yml

@@ -0,0 +1,168 @@
+# SwanLab DPO Training Example with Completion Logging
+#
+# This example demonstrates DPO (Direct Preference Optimization) training
+# with SwanLab integration for experiment tracking and completion table logging.
+#
+# Features enabled:
+# - SwanLab experiment tracking
+# - RLHF completion table logging (prompts, chosen/rejected responses, rewards)
+# - Lark (Feishu) team notifications (optional)
+#
+# To run:
+#   export SWANLAB_API_KEY=your-api-key
+#   accelerate launch -m axolotl.cli.train examples/swanlab/dpo-swanlab-completions.yml
+
+# Model Configuration
+base_model: meta-llama/Meta-Llama-3-8B-Instruct
+model_type: LlamaForCausalLM
+tokenizer_type: AutoTokenizer
+
+special_tokens:
+  pad_token: <|finetune_right_pad_id|>
+  eos_token: <|eot_id|>
+
+# Quantization
+load_in_8bit: true
+load_in_4bit: false
+
+# LoRA Configuration
+adapter: lora
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+# DPO Configuration
+chat_template: llama3
+rl: dpo
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_dpo_test
+    type: chat_template.default
+    field_messages: conversation
+    field_chosen: chosen
+    field_rejected: rejected
+    message_property_mappings:
+      role: role
+      content: content
+    roles:
+      system:
+        - system
+      user:
+        - user
+      assistant:
+        - assistant
+
+# Dataset and Output
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/dpo-swanlab-out
+
+# Training Configuration
+sequence_len: 4096
+sample_packing: false
+micro_batch_size: 2
+gradient_accumulation_steps: 4
+num_epochs: 4
+
+# Optimization
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+warmup_ratio: 0.1
+weight_decay: 0.0
+
+# Precision
+bf16: auto
+tf32: false
+
+# Performance
+gradient_checkpointing: true
+flash_attention: true
+
+# Checkpointing and Logging
+logging_steps: 1
+evals_per_epoch: 4
+saves_per_epoch: 1
+
+# ============================================================================
+# SwanLab Integration
+# ============================================================================
+
+plugins:
+  - axolotl.integrations.swanlab.SwanLabPlugin
+
+# Basic SwanLab Configuration
+use_swanlab: true
+swanlab_project: dpo-training
+swanlab_experiment_name: llama-3-dpo-completions-demo
+swanlab_description: "DPO training with completion table logging"
+swanlab_mode: cloud  # Options: cloud, local, offline, disabled
+
+# SwanLab Authentication
+# Recommended: Set via environment variable
+#   export SWANLAB_API_KEY=your-api-key
+# Or set in config (less secure):
+# swanlab_api_key: your-api-key
+
+# Optional: Team workspace
+# swanlab_workspace: my-research-team
+
+# ============================================================================
+# RLHF Completion Table Logging
+# ============================================================================
+#
+# Automatically logs model completions to SwanLab for qualitative analysis:
+# - Prompts from your DPO dataset
+# - Chosen responses (preferred)
+# - Rejected responses (non-preferred)
+# - Reward differences
+#
+# View the table in SwanLab dashboard under "rlhf_completions"
+
+swanlab_log_completions: true
+swanlab_completion_log_interval: 100  # Log every 100 training steps
+swanlab_completion_max_buffer: 128    # Keep last 128 completions in memory
+
+# Memory Usage Notes:
+# - Buffer size 128: ~64 KB (default, recommended)
+# - Buffer size 512: ~256 KB (for more historical completions)
+# - Buffer size 1024: ~512 KB (maximum for very long training runs)
+
+# Performance Notes:
+# - Completion logging overhead: < 0.5% per training step
+# - Only logs every N steps to minimize impact
+# - Memory-bounded buffer prevents memory leaks
+
+# ============================================================================
+# Optional: Lark (Feishu) Team Notifications
+# ============================================================================
+#
+# Get real-time training notifications in your team chat
+# Uncomment to enable:
+
+# swanlab_lark_webhook_url: https://open.feishu.cn/open-apis/bot/v2/hook/xxxxxxxxxx
+# swanlab_lark_secret: your-webhook-secret  # Recommended for production
+
+# Notifications sent for:
+# - Training start
+# - Training completion
+# - Training errors
+# - Metric milestones (if configured)
+
+# ============================================================================
+# Optional: Private SwanLab Deployment
+# ============================================================================
+#
+# For enterprise users with private SwanLab deployment:
+
+# swanlab_web_host: https://swanlab.yourcompany.com
+# swanlab_api_host: https://api.swanlab.yourcompany.com
+
+# ============================================================================
+# Disable WandB if you're migrating from it
+# ============================================================================
+
+# wandb_project:
+# wandb_entity:
+# use_wandb: false

examples/swanlab/dpo-swanlab-full-featured.yml

@@ -0,0 +1,329 @@
+# SwanLab Full-Featured DPO Training Example
+#
+# This example demonstrates ALL SwanLab integration features:
+# - Experiment tracking with cloud sync
+# - RLHF completion table logging
+# - Performance profiling
+# - Lark (Feishu) team notifications
+# - Team workspace collaboration
+#
+# Use this as a reference for production RLHF training setups.
+#
+# To run:
+#   export SWANLAB_API_KEY=your-api-key
+#   export SWANLAB_LARK_WEBHOOK_URL=https://open.feishu.cn/...
+#   export SWANLAB_LARK_SECRET=your-webhook-secret
+#   accelerate launch -m axolotl.cli.train examples/swanlab/dpo-swanlab-full-featured.yml
+
+# ============================================================================
+# Model Configuration
+# ============================================================================
+
+base_model: meta-llama/Meta-Llama-3-8B-Instruct
+model_type: LlamaForCausalLM
+tokenizer_type: AutoTokenizer
+
+special_tokens:
+  pad_token: <|finetune_right_pad_id|>
+  eos_token: <|eot_id|>
+
+# Quantization for efficient training
+load_in_8bit: true
+load_in_4bit: false
+
+# ============================================================================
+# LoRA Configuration
+# ============================================================================
+
+adapter: lora
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true  # Target all linear layers
+
+# ============================================================================
+# DPO (Direct Preference Optimization) Configuration
+# ============================================================================
+
+chat_template: llama3
+rl: dpo  # Enable DPO trainer
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_dpo_test
+    type: chat_template.default
+    field_messages: conversation
+    field_chosen: chosen
+    field_rejected: rejected
+    message_property_mappings:
+      role: role
+      content: content
+    roles:
+      system:
+        - system
+      user:
+        - user
+      assistant:
+        - assistant
+
+# ============================================================================
+# Dataset and Output Configuration
+# ============================================================================
+
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/dpo-swanlab-full-featured-out
+
+# ============================================================================
+# Training Configuration
+# ============================================================================
+
+sequence_len: 4096
+sample_packing: false
+
+micro_batch_size: 2
+gradient_accumulation_steps: 4
+num_epochs: 4
+
+# ============================================================================
+# Optimization
+# ============================================================================
+
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+warmup_ratio: 0.1
+weight_decay: 0.0
+
+# ============================================================================
+# Precision and Performance
+# ============================================================================
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+flash_attention: true
+
+# ============================================================================
+# Checkpointing and Logging
+# ============================================================================
+
+logging_steps: 1
+evals_per_epoch: 4
+saves_per_epoch: 1
+
+# ============================================================================
+# SwanLab Integration - Full Configuration
+# ============================================================================
+
+plugins:
+  - axolotl.integrations.swanlab.SwanLabPlugin
+
+# ------------------------------------------------------------------------------
+# Basic SwanLab Configuration
+# ------------------------------------------------------------------------------
+
+use_swanlab: true
+swanlab_project: dpo-production
+swanlab_experiment_name: llama-3-dpo-full-featured-v1
+swanlab_description: |
+  Production DPO training with all SwanLab features enabled:
+  - Completion table logging for qualitative analysis
+  - Performance profiling for optimization
+  - Lark notifications for team collaboration
+
+swanlab_mode: cloud  # Options: cloud, local, offline, disabled
+
+# ------------------------------------------------------------------------------
+# Team Collaboration
+# ------------------------------------------------------------------------------
+
+# Workspace for team collaboration (shared experiments)
+swanlab_workspace: ml-research-team
+
+# Authentication (recommended: use environment variable)
+#   export SWANLAB_API_KEY=your-api-key
+# Or set in config (less secure):
+# swanlab_api_key: your-api-key
+
+# ------------------------------------------------------------------------------
+# RLHF Completion Table Logging
+# ------------------------------------------------------------------------------
+# Automatically logs model completions for qualitative analysis:
+# - Prompts from your DPO dataset
+# - Chosen responses (preferred)
+# - Rejected responses (non-preferred)
+# - Reward differences
+#
+# View in SwanLab dashboard under "rlhf_completions" table
+
+swanlab_log_completions: true
+swanlab_completion_log_interval: 100  # Log every 100 steps
+swanlab_completion_max_buffer: 256    # Larger buffer for long training runs
+
+# Buffer size recommendations:
+# - 128: Default, ~64 KB memory (recommended for most cases)
+# - 256: ~128 KB memory (this config, good for longer training)
+# - 512: ~256 KB memory (maximum for very long runs)
+
+# ------------------------------------------------------------------------------
+# Lark (Feishu) Team Notifications
+# ------------------------------------------------------------------------------
+# Get real-time training notifications in your team chat
+#
+# Notifications sent for:
+# - Training start
+# - Training completion
+# - Training errors
+# - Metric milestones (if configured)
+
+# Recommended: Set via environment variables
+#   export SWANLAB_LARK_WEBHOOK_URL=https://open.feishu.cn/...
+#   export SWANLAB_LARK_SECRET=your-webhook-secret
+
+# Or set in config (less secure):
+# swanlab_lark_webhook_url: https://open.feishu.cn/open-apis/bot/v2/hook/xxxxxxxxxx
+# swanlab_lark_secret: your-webhook-secret  # REQUIRED for production
+
+# Security note: ALWAYS use swanlab_lark_secret in production to prevent
+# unauthorized parties from sending fake notifications to your team chat.
+
+# ------------------------------------------------------------------------------
+# Performance Profiling
+# ------------------------------------------------------------------------------
+# Profiling is automatically enabled when SwanLab is enabled.
+# Metrics logged to SwanLab under "profiling/" namespace:
+#   profiling/Time taken: AxolotlTrainer.training_step
+#   profiling/Time taken: AxolotlTrainer.compute_loss
+#   profiling/Time taken: AxolotlTrainer.prediction_step
+#
+# Use these metrics to:
+# - Identify bottlenecks in training loop
+# - Compare performance across different configurations
+# - Monitor performance regressions over time
+# - Debug unexpected slowdowns
+
+# For custom profiling in your own trainer, see:
+#   examples/swanlab/custom_trainer_profiling.py
+
+# ------------------------------------------------------------------------------
+# Optional: Private SwanLab Deployment
+# ------------------------------------------------------------------------------
+# For enterprise users with private SwanLab deployment:
+
+# swanlab_web_host: https://swanlab.yourcompany.com
+# swanlab_api_host: https://api.swanlab.yourcompany.com
+
+# ------------------------------------------------------------------------------
+# Optional: Model Checkpointing to SwanLab
+# ------------------------------------------------------------------------------
+# Log model checkpoints to SwanLab (coming soon)
+
+swanlab_log_model: false
+
+# ============================================================================
+# Disable Other Logging Tools (Recommended)
+# ============================================================================
+# Using multiple logging tools simultaneously can impact performance:
+# - Expected overhead: ~1-2% per logger
+# - Potential config/callback conflicts
+#
+# For production training, use ONLY SwanLab:
+
+# wandb_project:
+# use_wandb: false
+#
+# use_mlflow: false
+#
+# use_comet: false
+
+# ============================================================================
+# Expected Training Behavior
+# ============================================================================
+
+# With this configuration, you should see:
+#
+# 1. SwanLab Initialization (rank 0 only):
+#    INFO: SwanLab initialized for project: dpo-production
+#    INFO: SwanLab experiment: llama-3-dpo-full-featured-v1
+#    INFO: SwanLab mode: cloud
+#    INFO: SwanLab workspace: ml-research-team
+#
+# 2. Completion Logging (rank 0 only):
+#    INFO: Registered SwanLab RLHF completion logging callback for DPOTrainer
+#          (log_interval=100, max_buffer=256)
+#
+# 3. Lark Notifications (rank 0 only):
+#    INFO: Registered Lark notification callback with HMAC authentication
+#
+# 4. Distributed Training Detection (if multi-GPU):
+#    INFO: Distributed training detected (world_size=N)
+#    INFO: Only rank 0 will initialize SwanLab
+#    INFO: Other ranks will skip SwanLab to avoid conflicts
+#
+# 5. Training Start Notification (Lark):
+#    Your team chat receives: "Training started: llama-3-dpo-full-featured-v1"
+#
+# 6. Periodic Completion Logging:
+#    Every 100 steps, completion table is updated in SwanLab dashboard
+#
+# 7. Training Complete Notification (Lark):
+#    Your team chat receives: "Training completed: llama-3-dpo-full-featured-v1"
+#    With link to SwanLab dashboard and final metrics
+#
+# 8. SwanLab Dashboard Shows:
+#    - Training metrics (loss, learning rate, etc.)
+#    - Completion table (rlhf_completions)
+#    - Profiling metrics (profiling/Time taken: ...)
+#    - Hyperparameters and configuration
+#    - System resource usage
+
+# ============================================================================
+# Production Checklist
+# ============================================================================
+
+# Before deploying to production, verify:
+# ✅ SwanLab API key is set via environment variable (not in config)
+# ✅ Lark webhook secret is set (required for HMAC authentication)
+# ✅ Workspace is set to your team's workspace
+# ✅ Experiment name is descriptive and unique
+# ✅ Only SwanLab is enabled (other loggers disabled)
+# ✅ Completion logging buffer size is appropriate for your training duration
+# ✅ Private deployment hosts are set (if using enterprise SwanLab)
+# ✅ Test run completes successfully and shows up in SwanLab dashboard
+# ✅ Lark notifications are received in team chat
+# ✅ Profiling metrics are logged correctly
+
+# ============================================================================
+# Troubleshooting
+# ============================================================================
+
+# If SwanLab initialization fails:
+# 1. Check SWANLAB_API_KEY environment variable is set
+# 2. Verify swanlab_project is set in config
+# 3. Check swanlab_mode is valid (cloud/local/offline/disabled)
+# 4. Verify internet connectivity (for cloud mode)
+
+# If Lark notifications not received:
+# 1. Check SWANLAB_LARK_WEBHOOK_URL is set correctly
+# 2. Verify SWANLAB_LARK_SECRET matches your Lark bot settings
+# 3. Test webhook manually: curl -X POST "$SWANLAB_LARK_WEBHOOK_URL" ...
+# 4. Check training logs for "Registered Lark notification callback"
+# 5. Verify bot is added to the target Lark group chat
+
+# If completions not appearing in SwanLab:
+# 1. Verify you're using an RLHF trainer (DPO/KTO/ORPO/GRPO)
+# 2. Check swanlab_log_completions is true
+# 3. Wait for log_interval steps (default: 100)
+# 4. Check training logs for "Registered SwanLab RLHF completion logging"
+
+# If profiling metrics not appearing:
+# 1. Verify use_swanlab is true
+# 2. Check SwanLab is initialized (check logs)
+# 3. Look under "profiling/" namespace in dashboard
+# 4. Profiling may be disabled if DEFAULT_PROFILING_CONFIG.enabled = False
+
+# For more help:
+# - SwanLab docs: https://docs.swanlab.cn
+# - Axolotl SwanLab integration: src/axolotl/integrations/swanlab/README.md
+# - GitHub issues: https://github.com/axolotl-ai-cloud/axolotl/issues

examples/swanlab/lora-swanlab-profiling.yml

@@ -0,0 +1,178 @@
+# SwanLab LoRA Training Example with Performance Profiling
+#
+# This example demonstrates standard LoRA fine-tuning with SwanLab integration
+# for performance profiling and optimization.
+#
+# Features enabled:
+# - SwanLab experiment tracking
+# - Performance profiling (training step, forward/backward pass timing)
+# - Real-time metrics visualization
+#
+# To run:
+#   export SWANLAB_API_KEY=your-api-key
+#   accelerate launch -m axolotl.cli.train examples/swanlab/lora-swanlab-profiling.yml
+
+# Model Configuration
+base_model: NousResearch/Llama-3.2-1B
+
+# Dataset Configuration
+datasets:
+  - path: teknium/GPT4-LLM-Cleaned
+    type: alpaca
+
+val_set_size: 0.1
+output_dir: ./outputs/lora-swanlab-profiling-out
+
+# LoRA Configuration
+adapter: lora
+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
+
+# Training Configuration
+sequence_len: 2048
+sample_packing: true
+eval_sample_packing: true
+
+micro_batch_size: 2
+gradient_accumulation_steps: 2
+num_epochs: 1
+
+# Optimization
+optimizer: adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+warmup_ratio: 0.1
+weight_decay: 0.0
+
+# Precision
+bf16: auto
+tf32: false
+
+# Performance
+gradient_checkpointing: true
+flash_attention: true
+
+# Checkpointing and Logging
+logging_steps: 1
+evals_per_epoch: 4
+saves_per_epoch: 1
+
+# Loss Monitoring
+loss_watchdog_threshold: 5.0
+loss_watchdog_patience: 3
+
+special_tokens:
+  pad_token: "<|end_of_text|>"
+
+# ============================================================================
+# SwanLab Integration
+# ============================================================================
+
+plugins:
+  - axolotl.integrations.swanlab.SwanLabPlugin
+
+# Basic SwanLab Configuration
+use_swanlab: true
+swanlab_project: lora-profiling
+swanlab_experiment_name: llama-3.2-1b-profiling-demo
+swanlab_description: "LoRA fine-tuning with performance profiling"
+swanlab_mode: cloud  # Options: cloud, local, offline, disabled
+
+# SwanLab Authentication
+# Recommended: Set via environment variable
+#   export SWANLAB_API_KEY=your-api-key
+# Or set in config (less secure):
+# swanlab_api_key: your-api-key
+
+# Optional: Team workspace
+# swanlab_workspace: my-ml-team
+
+# ============================================================================
+# Performance Profiling
+# ============================================================================
+#
+# SwanLab automatically profiles trainer methods when enabled.
+# Profiling metrics appear in SwanLab dashboard under "profiling/" namespace.
+#
+# Built-in profiling:
+# - Minimal overhead (< 0.1% per step)
+# - High-precision timing (microsecond accuracy)
+# - Exception-safe (logs duration even if method fails)
+#
+# View profiling metrics in SwanLab dashboard:
+#   profiling/Time taken: AxolotlTrainer.training_step
+#   profiling/Time taken: AxolotlTrainer.compute_loss
+#   profiling/Time taken: AxolotlTrainer.prediction_step
+#
+# For custom profiling in your own trainer, see:
+#   examples/swanlab/custom_trainer_profiling.py
+
+# Completion logging is disabled for non-RLHF trainers
+swanlab_log_completions: false  # Only works with DPO/KTO/ORPO/GRPO
+
+# ============================================================================
+# Optional: Compare with Multiple Runs
+# ============================================================================
+#
+# To compare profiling metrics across different configurations:
+#
+# 1. Run baseline without flash attention:
+#    swanlab_experiment_name: llama-3.2-1b-no-flash-attn
+#    flash_attention: false
+#
+# 2. Run with gradient checkpointing:
+#    swanlab_experiment_name: llama-3.2-1b-grad-checkpoint
+#    gradient_checkpointing: true
+#
+# 3. Run with both:
+#    swanlab_experiment_name: llama-3.2-1b-optimized
+#    flash_attention: true
+#    gradient_checkpointing: true
+#
+# Then compare profiling metrics in SwanLab dashboard to see performance impact
+
+# ============================================================================
+# Optional: Lark (Feishu) Team Notifications
+# ============================================================================
+#
+# Get notified when profiling experiments complete:
+
+# swanlab_lark_webhook_url: https://open.feishu.cn/open-apis/bot/v2/hook/xxxxxxxxxx
+# swanlab_lark_secret: your-webhook-secret
+
+# ============================================================================
+# Profiling Best Practices
+# ============================================================================
+#
+# 1. Run multiple epochs to see profiling trends over time
+# 2. Ignore first ~10 steps (warmup period, slower)
+# 3. Look for outliers (steps that take significantly longer)
+# 4. Compare profiling metrics before/after optimization changes
+# 5. Monitor per-rank profiling in distributed training
+#
+# Common bottlenecks to profile:
+# - training_step: Overall step time (should be consistent)
+# - compute_loss: Loss computation (scales with sequence length)
+# - prediction_step: Evaluation time (can be slow for large val sets)
+#
+# If you see inconsistent timing:
+# - Check for data loading bottlenecks
+# - Monitor GPU utilization (may be CPU-bound)
+# - Check for gradient accumulation effects
+# - Verify CUDA kernel synchronization
+
+# ============================================================================
+# Disable WandB if you're migrating from it
+# ============================================================================
+
+# wandb_project:
+# use_wandb: false

examples/trinity/README.md

@@ -0,0 +1,42 @@
+# Finetune ArceeAI's Trinity with Axolotl
+
+[Trinity](https://huggingface.co/collections/arcee-ai/trinity) is a family of open weight MoE models trained by Arcee.ai.
+
+This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
+
+## Getting started
+
+1. Install Axolotl following the main from the [installation guide](https://docs.axolotl.ai/docs/installation.html#sec-edge-build).
+
+2. Run the finetuning example:
+
+    ```bash
+    axolotl train examples/trinity/trinity-nano-preview-qlora.yaml
+    ```
+
+This config uses about 24.9 GiB VRAM.
+
+Let us know how it goes. Happy finetuning! 🚀
+
+### TIPS
+
+- For inference, the official Arcee.ai team recommends `top_p: 0.75`, `temperature: 0.15`, `top_k: 50`, and `min_p: 0.06`.
+- You can run a full finetuning by removing the `adapter: qlora` and `load_in_4bit: true` from the config.
+- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+- The dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+
+## Optimization Guides
+
+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)
+- [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/trinity-nano-preview-qlora.yaml

@@ -0,0 +1,68 @@
+base_model: arcee-ai/Trinity-Nano-Preview
+trust_remote_code: true
+revision_of_model: 2ee94b0
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+# CCE - N/A as of now
+# plugins:
+#   - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/lora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+# flash_attention: true  # Not supported
+sdp_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

requirements.txt

@@ -1,33 +1,34 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
 
 # START section of dependencies that don't install on Darwin/MacOS
-bitsandbytes==0.48.2
+bitsandbytes==0.49.1
 triton>=3.0.0
 mamba-ssm==1.2.0.post1
 xformers>=0.0.23.post1
-liger-kernel==0.6.3
+liger-kernel==0.6.4
 # END section
 
 packaging==23.2
 
 huggingface_hub>=0.36.0
-peft>=0.18.0
+peft>=0.18.1
 tokenizers>=0.22.1
-transformers==4.57.1
-accelerate==1.11.0
-datasets==4.4.1
-deepspeed>=0.17.0
-trl==0.25.0
+transformers==4.57.6
+accelerate==1.12.0
+datasets==4.5.0
+deepspeed>=0.18.3
+trl==0.25.1
 hf_xet==1.2.0
-kernels>=0.9.0
-trackio
+kernels==0.11.5
+trackio>=0.13.0
+typing-extensions>=4.15.0
 
 optimum==1.16.2
 hf_transfer
 sentencepiece
-gradio==5.49.1
+gradio>=6.2.0,<7.0
 
-modal==1.0.2
+modal==1.3.0.post1
 pydantic>=2.10.6
 addict
 fire
@@ -67,9 +68,8 @@ openenv-core==0.1.0
 schedulefree==1.4.1
 
 axolotl-contribs-lgpl==0.0.7
-axolotl-contribs-mit==0.0.5
-
+axolotl-contribs-mit==0.0.6
 # telemetry
 posthog==6.7.11
 
-mistral-common==1.8.5
+mistral-common==1.8.6

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@5eff953"'
+    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@318b7e2"'
 )

setup.py

@@ -26,6 +26,7 @@ def parse_requirements(extras_require_map):
                 _install_requires.append(line)
     try:
         xformers_version = [req for req in _install_requires if "xformers" in req][0]
+        install_xformers = platform.machine() != "aarch64"
         if "Darwin" in platform.system():
             # skip packages not compatible with OSX
             skip_packages = [
@@ -62,45 +63,63 @@ def parse_requirements(extras_require_map):
             else:
                 raise ValueError("Invalid version format")
 
+            torch_parts = torch_version.split("+")
+            if len(torch_parts) == 2:
+                torch_cuda_version = torch_parts[1]
+                _dependency_links.append(
+                    f"https://download.pytorch.org/whl/{torch_cuda_version}"
+                )
+
             if (major, minor) >= (2, 9):
                 extras_require_map.pop("fbgemm-gpu")
-                extras_require_map["fbgemm-gpu"] = ["fbgemm-gpu-genai==1.4.1"]
+                extras_require_map["fbgemm-gpu"] = [
+                    "fbgemm-gpu==1.4.0",
+                    "fbgemm-gpu-genai==1.4.2",
+                ]
                 extras_require_map["vllm"] = ["vllm==0.11.1"]
-                _install_requires.pop(_install_requires.index(xformers_version))
+                if not install_xformers:
+                    _install_requires.pop(_install_requires.index(xformers_version))
             elif (major, minor) >= (2, 8):
                 extras_require_map.pop("fbgemm-gpu")
                 extras_require_map["fbgemm-gpu"] = ["fbgemm-gpu-genai==1.3.0"]
                 extras_require_map["vllm"] = ["vllm==0.11.0"]
+                if not install_xformers:
+                    _install_requires.pop(_install_requires.index(xformers_version))
             elif (major, minor) >= (2, 7):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 if patch == 0:
-                    _install_requires.append("xformers==0.0.30")
+                    if install_xformers:
+                        _install_requires.append("xformers==0.0.30")
                     # vllm 0.9.x is incompatible with latest transformers
                     extras_require_map.pop("vllm")
                 else:
-                    _install_requires.append("xformers==0.0.31")
+                    if install_xformers:
+                        _install_requires.append("xformers==0.0.31")
                     extras_require_map["vllm"] = ["vllm==0.10.1"]
             elif (major, minor) >= (2, 6):
                 _install_requires.pop(_install_requires.index(xformers_version))
-                _install_requires.append("xformers==0.0.29.post3")
+                if install_xformers:
+                    _install_requires.append("xformers==0.0.29.post3")
                 # since we only support 2.6.0+cu126
                 _dependency_links.append("https://download.pytorch.org/whl/cu126")
                 extras_require_map.pop("vllm")
             elif (major, minor) >= (2, 5):
                 _install_requires.pop(_install_requires.index(xformers_version))
-                if patch == 0:
-                    _install_requires.append("xformers==0.0.28.post2")
-                else:
-                    _install_requires.append("xformers>=0.0.28.post3")
+                if install_xformers:
+                    if patch == 0:
+                        _install_requires.append("xformers==0.0.28.post2")
+                    else:
+                        _install_requires.append("xformers>=0.0.28.post3")
                 extras_require_map.pop("vllm")
             elif (major, minor) >= (2, 4):
                 extras_require_map.pop("vllm")
-                if patch == 0:
-                    _install_requires.pop(_install_requires.index(xformers_version))
-                    _install_requires.append("xformers>=0.0.27")
-                else:
-                    _install_requires.pop(_install_requires.index(xformers_version))
-                    _install_requires.append("xformers==0.0.28.post1")
+                if install_xformers:
+                    if patch == 0:
+                        _install_requires.pop(_install_requires.index(xformers_version))
+                        _install_requires.append("xformers>=0.0.27")
+                    else:
+                        _install_requires.pop(_install_requires.index(xformers_version))
+                        _install_requires.append("xformers==0.0.28.post1")
             else:
                 raise ValueError("axolotl requires torch>=2.4")
 
@@ -157,7 +176,7 @@ extras_require = {
         "came_pytorch==0.1.3",
     ],
     "ray": [
-        "ray[train]",
+        "ray[train]>=2.52.1",
     ],
     "vllm": [
         "vllm==0.10.0",

src/axolotl/__init__.py

@@ -4,4 +4,4 @@ import pkgutil
 
 __path__ = pkgutil.extend_path(__path__, __name__)  # Make this a namespace package
 
-__version__ = "0.13.0.dev"
+__version__ = "0.13.1"

src/axolotl/cli/cloud/baseten/template/train_sft.py

@@ -24,8 +24,7 @@ if launcher_args:
     launcher_args_str = "-- " + " ".join(launcher_args)
 
 # 1. Define a base image for your training job
-# must use torch 2.7.0 for vllm
-BASE_IMAGE = "axolotlai/axolotl:main-py3.11-cu126-2.7.1"
+BASE_IMAGE = "axolotlai/axolotl:main-py3.11-cu128-2.9.1"
 
 # 2. Define the Runtime Environment for the Training Job
 # This includes start commands and environment variables.a

src/axolotl/cli/cloud/modal_.py

@@ -82,7 +82,7 @@ class ModalCloud(Cloud):
         return res
 
     def get_image(self):
-        docker_tag = "main-py3.11-cu126-2.7.1"
+        docker_tag = "main-py3.11-cu128-2.9.1"
         if self.config.docker_tag:
             docker_tag = self.config.docker_tag
         docker_image = f"axolotlai/axolotl:{docker_tag}"

src/axolotl/cli/config.py

@@ -26,6 +26,7 @@ from axolotl.utils.dict import DictDefault
 from axolotl.utils.logging import get_logger
 from axolotl.utils.mlflow_ import setup_mlflow_env_vars
 from axolotl.utils.tee import prepare_debug_log
+from axolotl.utils.trackio_ import setup_trackio_env_vars
 from axolotl.utils.trainer import prepare_optim_env
 from axolotl.utils.wandb_ import setup_wandb_env_vars
 
@@ -227,6 +228,7 @@ def load_cfg(
         cfg,
         capabilities={
             "bf16": is_torch_bf16_gpu_available(),
+            "fp8": compute_supports_fp8(),
             "n_gpu": int(os.environ.get("WORLD_SIZE", 1)),
             "compute_capability": gpu_version,
         },
@@ -245,6 +247,7 @@ def load_cfg(
     setup_wandb_env_vars(cfg)
     setup_mlflow_env_vars(cfg)
     setup_comet_env_vars(cfg)
+    setup_trackio_env_vars(cfg)
     plugin_set_cfg(cfg)
 
     TELEMETRY_MANAGER.send_event(event_type="config-processed", properties=cfg)
@@ -259,3 +262,11 @@ def load_cfg(
     )
 
     return cfg
+
+
+def compute_supports_fp8() -> bool:
+    try:
+        compute_capability = torch.cuda.get_device_capability()
+        return compute_capability >= (9, 0)
+    except RuntimeError:
+        return False

src/axolotl/cli/inference.py

@@ -288,8 +288,8 @@ def do_inference_gradio(
         title=cfg.get("gradio_title", "Axolotl Gradio Interface"),
     )
 
-    demo.queue().launch(
-        show_api=False,
+    demo.launch(
+        footer_links=["gradio", "settings"],
         share=cfg.get("gradio_share", True),
         server_name=cfg.get("gradio_server_name", "127.0.0.1"),
         server_port=cfg.get("gradio_server_port", None),

src/axolotl/cli/main.py

@@ -26,7 +26,7 @@ from axolotl.cli.utils import (
     launch_training,
 )
 from axolotl.integrations.lm_eval.cli import lm_eval
-from axolotl.utils import set_pytorch_cuda_alloc_conf
+from axolotl.utils import set_misc_env, set_pytorch_cuda_alloc_conf
 from axolotl.utils.logging import get_logger
 from axolotl.utils.schemas.config import AxolotlInputConfig
 
@@ -45,6 +45,7 @@ def cli():
     print_axolotl_text_art()
     load_dotenv()
     set_pytorch_cuda_alloc_conf()
+    set_misc_env()
 
 
 @cli.command()

src/axolotl/cli/quantize.py

@@ -8,7 +8,7 @@ from typing import Union
 from transformers import AutoConfig, AutoModelForCausalLM, TorchAoConfig
 
 from axolotl.cli.config import load_cfg
-from axolotl.loaders import load_tokenizer
+from axolotl.loaders import load_processor, load_tokenizer
 from axolotl.utils.logging import get_logger
 from axolotl.utils.quantization import (
     TorchAOQuantDType,
@@ -66,6 +66,11 @@ def do_quantize(
 
     LOG.info(f"Loading model from {model_path}.")
     tokenizer = load_tokenizer(cfg)
+
+    processor = None
+    if cfg.is_multimodal:
+        processor = load_processor(cfg, tokenizer)
+
     config = AutoConfig.from_pretrained(model_path)
     torch_dtype = config.torch_dtype if hasattr(config, "torch_dtype") else None
     model = AutoModelForCausalLM.from_pretrained(
@@ -107,6 +112,10 @@ def do_quantize(
         save_jinja_files=cfg.tokenizer_save_jinja_files,
     )
 
+    if processor:
+        LOG.info(f"Saving processor to: {str(Path(output_dir) / 'quantized')}.")
+        processor.save_pretrained(str(Path(output_dir) / "quantized"))
+
     if hub_model_id:
         hub_model_id = (
             hub_model_id.rstrip("-")
@@ -114,6 +123,8 @@ def do_quantize(
         )
         model.push_to_hub(hub_model_id, safe_serialization=False)
         tokenizer.push_to_hub(hub_model_id)
+        if processor:
+            processor.push_to_hub(hub_model_id)
         LOG.info(f"Quantized model pushed to: {hub_model_id}.")
 
     LOG.info(f"Quantized model saved to: {str(Path(output_dir) / 'quantized')}.")

src/axolotl/cli/utils/diffusion.py

@@ -366,8 +366,8 @@ def launch_diffusion_gradio_ui(
             outputs=[masked_preview, html_out],
         )
 
-        demo.queue().launch(
-            show_api=False,
+        demo.launch(
+            footer_links=["gradio", "settings"],
             share=cfg.get("gradio_share", True),
             server_name=cfg.get("gradio_server_name", "127.0.0.1"),
             server_port=cfg.get("gradio_server_port", None),

src/axolotl/common/architectures.py

@@ -17,4 +17,5 @@ MOE_ARCH_BLOCK = {
     "deepseek_v3": "DeepseekV3MoE",
     "gpt_oss": "GptOssDecoderLayer",
     "lfm2_moe": "Lfm2MoeSparseMoeBlock",
+    "afmoe": "AfmoeMoE",
 }

src/axolotl/core/attention/flex_block_mask.py

@@ -1,158 +0,0 @@
-"""
-monkeypatch for flex + packing
-"""
-
-import sys
-from typing import Callable, Optional, Union
-
-import torch
-from torch.nn.attention.flex_attention import BlockMask
-from transformers import Cache, PretrainedConfig
-from transformers.masking_utils import (
-    ALL_MASK_ATTENTION_FUNCTIONS,
-    _preprocess_mask_arguments,
-    and_masks,
-    causal_mask_function,
-    or_masks,
-)
-from transformers.utils import is_torch_greater_or_equal
-
-_is_torch_greater_or_equal_than_2_6 = is_torch_greater_or_equal("2.6", accept_dev=True)
-
-
-def create_causal_mask(
-    config: PretrainedConfig,
-    input_embeds: torch.Tensor,
-    attention_mask: torch.Tensor,
-    cache_position: torch.Tensor,
-    past_key_values: Optional[Cache],
-    or_mask_function: Optional[Callable] = None,
-    and_mask_function: Optional[Callable] = None,
-) -> Optional[Union[torch.Tensor, BlockMask]]:
-    """
-    Create a standard causal mask based on the attention implementation used (stored in the config). If `past_key_values`
-    has an HybridCache structure, this function will return the mask corresponding to one of the "full_attention" layers (to align
-    to what is needed in the `modeling_xxx.py` files).
-
-    Args:
-        config (`PretrainedConfig`):
-            The model config.
-        input_embeds (`torch.Tensor`):
-            The input embeddings of shape (batch_size, query_length, hidden_dim). This is used only to infer the
-            batch size, query length and dtype.
-        attention_mask (`torch.Tensor`, optional):
-            The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length).
-            It can also be an already prepared 4D mask, in which case it is returned as-is.
-        cache_position (`torch.Tensor`):
-            A tensor of shape (query_length,) indicating the current indices of the input sequence elements.
-        past_key_values (`Cache`, optional):
-            The past key values, if we use a cache.
-        or_mask_function (`Callable`, optional):
-            An optional mask function to combine with the causal mask function (by doing the union of both). This is
-            useful to easily overlay another mask on top of the causal one, for example for image tokens handling.
-        and_mask_function (`Callable`, optional):
-            An optional mask function to combine with the causal mask function (by doing the intersection of both). This is
-            useful to easily overlay another mask on top of the causal one, for example for image tokens handling.
-    """
-    # If we have an HybridCache structure, here we want to create the mask for the full layers
-    if (
-        past_key_values
-        and hasattr(past_key_values, "is_sliding")
-        and False in past_key_values.is_sliding
-    ):
-        layer_idx = past_key_values.is_sliding.index(False)
-    else:
-        layer_idx = 0
-
-    original_attention_mask = (
-        None
-        if attention_mask is None
-        else attention_mask.clone().to(cache_position.device)
-    )
-    early_exit, attention_mask, kv_length, kv_offset = _preprocess_mask_arguments(
-        config, input_embeds, attention_mask, cache_position, past_key_values, layer_idx
-    )
-    if early_exit:
-        return attention_mask
-
-    batch_size, total_seq_len = cache_position.shape
-    key_length = total_seq_len
-    document_ids = torch.nn.functional.pad(
-        original_attention_mask, value=0, pad=(0, key_length)
-    )
-
-    batch_size, dtype = input_embeds.shape[0], input_embeds.dtype
-    if attention_mask is not None:
-
-        def causal_doc_mask_mod(batch_idx, head_idx, q_idx, kv_idx):
-            """
-            Defines the logic of a block causal mask by combining both a standard causal mask
-            and a block diagonal document mask.
-            See :func:`~torchtune.modules.attention_utils.create_block_causal_mask`
-            for an illustration.
-            """
-            causal_mask_ = q_idx >= kv_idx  # not valid when decoding
-            document_mask = (
-                document_ids[batch_idx, q_idx] == document_ids[batch_idx, kv_idx]
-            )
-            final_mask = causal_mask_ & document_mask
-            return final_mask
-
-        mask_factory_function = causal_doc_mask_mod
-    else:
-        mask_factory_function = causal_mask_function
-    mask_interface = ALL_MASK_ATTENTION_FUNCTIONS[config._attn_implementation]
-
-    # Do not allow skip if we are compiling (this is to match BC)
-    allow_is_causal_skip = (
-        not past_key_values.is_compileable if past_key_values is not None else True
-    )
-
-    # Allow slight deviations from causal mask
-    if or_mask_function is not None:
-        if not _is_torch_greater_or_equal_than_2_6:
-            raise ValueError(
-                "Using `or_mask_function` or `and_mask_function` arguments require torch>=2.6"
-            )
-        mask_factory_function = or_masks(mask_factory_function, or_mask_function)
-        allow_is_causal_skip = False
-    if and_mask_function is not None:
-        if not _is_torch_greater_or_equal_than_2_6:
-            raise ValueError(
-                "Using `or_mask_function` or `and_mask_function` arguments require torch>=2.6"
-            )
-        mask_factory_function = and_masks(mask_factory_function, and_mask_function)
-        allow_is_causal_skip = False
-
-    # We now create the mask
-    causal_mask = mask_interface(
-        batch_size=batch_size,
-        cache_position=cache_position,
-        kv_length=kv_length,
-        kv_offset=kv_offset,
-        mask_function=mask_factory_function,
-        attention_mask=attention_mask,
-        allow_is_causal_skip=allow_is_causal_skip,  # additional kwarg for sdpa
-        dtype=dtype,  # Additional kwarg for eager
-        config=config,  # Pass the config as well, in case someone wants to easily have their own mask_interface
-    )
-    return causal_mask
-
-
-def patch_create_causal_mask(model_type):
-    import transformers.masking_utils
-
-    transformers.masking_utils.create_causal_mask = create_causal_mask
-
-    if model_type:
-        try:
-            # Dynamically import the module and attention class
-            module_path = f"transformers.models.{model_type}.modeling_{model_type}"
-            module = __import__(module_path)
-            module.create_causal_mask = create_causal_mask
-            del sys.modules[module_path]
-        except (ImportError, AttributeError) as e:
-            raise ValueError(
-                f"Could not import attention class for model_type: {model_type}. "
-                f"Error: {str(e)}"
-            ) from e

src/axolotl/core/builders/base.py

@@ -35,6 +35,7 @@ from axolotl.utils import (
     is_comet_available,
     is_mlflow_available,
     is_opentelemetry_available,
+    is_trackio_available,
 )
 from axolotl.utils.callbacks import (
     GCCallback,
@@ -147,6 +148,14 @@ class TrainerBuilderBase(abc.ABC):
             callbacks.append(
                 SaveAxolotlConfigtoCometCallback(self.cfg.axolotl_config_path)
             )
+        if self.cfg.use_trackio and is_trackio_available():
+            from axolotl.utils.callbacks.trackio_ import (
+                SaveAxolotlConfigtoTrackioCallback,
+            )
+
+            callbacks.append(
+                SaveAxolotlConfigtoTrackioCallback(self.cfg.axolotl_config_path)
+            )
         if self.cfg.use_otel_metrics and is_opentelemetry_available():
             from axolotl.utils.callbacks.opentelemetry import (
                 OpenTelemetryMetricsCallback,
@@ -281,11 +290,22 @@ class TrainerBuilderBase(abc.ABC):
                 adam_kwargs["eps"] = training_args_kwargs.get("adam_epsilon")
 
             if self.cfg.optimizer == "muon":
-                from axolotl.contribs.mit.muon import (
-                    MuonOptimizerFactory,
-                )
+                _, device_mesh = build_parallelism_config(self.cfg)
+
+                if device_mesh is not None:
+                    from axolotl.contribs.mit.muon.dist_muon import (
+                        DistMuonOptimizerFactory,
+                    )
+
+                    optimizer_cls = DistMuonOptimizerFactory
+                    optimizer_kwargs["device_mesh"] = device_mesh
+                else:
+                    from axolotl.contribs.mit.muon import (
+                        MuonOptimizerFactory,
+                    )
+
+                    optimizer_cls = MuonOptimizerFactory
 
-                optimizer_cls = MuonOptimizerFactory
                 optimizer_kwargs.update(adam_kwargs)
             elif self.cfg.optimizer == "dion":
                 from axolotl.contribs.mit.dion import (
@@ -423,6 +443,8 @@ class TrainerBuilderBase(abc.ABC):
             report_to.append("tensorboard")
         if self.cfg.use_comet:
             report_to.append("comet_ml")
+        if self.cfg.use_trackio:
+            report_to.append("trackio")
 
         training_args_kwargs["report_to"] = report_to
 
@@ -430,6 +452,8 @@ class TrainerBuilderBase(abc.ABC):
             training_args_kwargs["run_name"] = self.cfg.wandb_name
         elif self.cfg.use_mlflow:
             training_args_kwargs["run_name"] = self.cfg.mlflow_run_name
+        elif self.cfg.use_trackio:
+            training_args_kwargs["run_name"] = self.cfg.trackio_run_name
         else:
             training_args_kwargs["run_name"] = None
 

src/axolotl/core/builders/causal.py

@@ -72,7 +72,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         if self.cfg.include_tkps:
             callbacks.append(
                 TokensPerSecondCallback(
-                    self.cfg.tensor_parallel_size, self.cfg.context_parallel_size
+                    self.cfg.tensor_parallel_size,
+                    self.cfg.context_parallel_size,
+                    resume_from_checkpoint=self.cfg.resume_from_checkpoint,
                 )
             )
         return callbacks

src/axolotl/core/trainers/base.py

@@ -2,6 +2,8 @@
 
 from __future__ import annotations
 
+import json
+import math
 import os
 from collections import defaultdict
 from functools import partial, wraps
@@ -49,6 +51,8 @@ from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
 
 LOG = get_logger(__name__)
 
+TOKENS_STATE_FILE = "tokens_state."
+
 REDUCTION_FNS = {
     "mean": torch.mean,
     "min": torch.min,
@@ -348,24 +352,33 @@ class AxolotlTrainer(
         #     return (loss, outputs) if return_outputs else loss
 
         # track number of tokens for tokens per second calculation
-        if self.args.include_tkps:
+        if self.args.include_tkps and model.training:
             inputs_key = "labels" if "labels" in inputs else "input_ids"
-            num_tokens = (inputs[inputs_key] != -100).sum()
+            trainable_tokens = (inputs[inputs_key] != -100).sum()
+            total_tokens = inputs[inputs_key].numel()
+            total_tokens = torch.tensor(total_tokens, device=inputs[inputs_key].device)
+
             if is_distributed():
                 torch.distributed.all_reduce(
-                    num_tokens, op=torch.distributed.ReduceOp.SUM
+                    trainable_tokens, op=torch.distributed.ReduceOp.SUM
                 )
-            if hasattr(self.state, "num_tokens"):
-                self.state.num_tokens = (
-                    self.state.num_tokens + (inputs[inputs_key] != -100).sum().cpu()
+                torch.distributed.all_reduce(
+                    total_tokens, op=torch.distributed.ReduceOp.SUM
                 )
-            else:
-                self.state.num_tokens = (inputs[inputs_key] != -100).sum().cpu()
 
-            if hasattr(self.state, "total_tokens"):
-                self.state.total_tokens += num_tokens
-            else:
-                self.state.total_tokens = num_tokens
+            if not hasattr(self.state, "tokens"):
+                self.state.tokens = {
+                    "trainable": torch.zeros(1),
+                    "total": torch.zeros(1),
+                }
+
+            # trainable tokens for throughput and total token slots for summaries
+            self.state.tokens["trainable"] = (
+                self.state.tokens["trainable"] + trainable_tokens.detach().cpu()
+            )
+            self.state.tokens["total"] = self.state.tokens["total"] + total_tokens.cpu()
+            # Store per-step trainable tokens for throughput calculation
+            self.state.tokens["trainable_tokens"] = trainable_tokens.detach().cpu()
 
         if self.args.orpo_alpha:
             return self.orpo_compute_loss(
@@ -603,6 +616,7 @@ class AxolotlTrainer(
         """
         # logs either has 'loss' or 'eval_loss'
         train_eval = "train" if "loss" in logs else "eval"
+        metric_ndigits = int(os.getenv("AXOLOTL_METRIC_NDIGITS", "5"))
 
         for key, metric_data in self._stored_metrics[train_eval].items():
             values = torch.tensor(metric_data["values"])  # type: ignore[arg-type]
@@ -613,7 +627,18 @@ class AxolotlTrainer(
                 raise NotImplementedError(
                     "Metric reduction must be one of [mean, min, max, sum]"
                 )
-            logs[key] = round(fn(values).item(), 4)
+            logs[key] = round(fn(values).item(), metric_ndigits)
+
+        if "loss" in logs:
+            try:
+                logs["ppl"] = round(math.exp(logs["loss"]), metric_ndigits)
+            except OverflowError:
+                logs["ppl"] = float("inf")
+        if "eval_loss" in logs:
+            try:
+                logs["eval_ppl"] = round(math.exp(logs["eval_loss"]), metric_ndigits)
+            except OverflowError:
+                logs["eval_ppl"] = float("inf")
 
         if is_main_process():
             # Add memory usage
@@ -625,13 +650,20 @@ class AxolotlTrainer(
             except (ValueError, TypeError, FileNotFoundError):
                 pass
 
-        if self.args.include_tkps and train_eval == "train":
+        if (
+            self.args.include_tkps
+            and train_eval == "train"
+            and hasattr(self.state, "tokens")
+        ):
             # each rank will log its own tokens per second
             # for logging_steps > 1 we obtain a moving average of this metric
-            logs["tokens_per_second_per_gpu"] = round(
+            logs["tokens/train_per_sec_per_gpu"] = round(
                 self.state.last_tokens_per_second.item() / self.args.logging_steps, 2
             )
-            logs["total_tokens"] = int(self.state.total_tokens.item())
+            if "total" in self.state.tokens:
+                logs["tokens/total"] = int(self.state.tokens["total"].item())
+            if "trainable" in self.state.tokens:
+                logs["tokens/trainable"] = int(self.state.tokens["trainable"].item())
 
         del self._stored_metrics[train_eval]
 
@@ -666,6 +698,19 @@ class AxolotlTrainer(
         run_dir = self._get_output_dir(trial=trial)
         output_dir = os.path.join(run_dir, checkpoint_folder)
         os.makedirs(output_dir, exist_ok=True)
+
+        # Save total_tokens state if tracking is enabled
+        if self.args.include_tkps and hasattr(self.state, "tokens"):
+            tokens_state = {
+                "total": int(torch.as_tensor(self.state.tokens.get("total", 0)).item()),
+                "trainable": int(
+                    torch.as_tensor(self.state.tokens.get("trainable", 0)).item()
+                ),
+            }
+            tokens_state_path = os.path.join(output_dir, TOKENS_STATE_FILE)
+            with open(tokens_state_path, "w", encoding="utf-8") as f:
+                json.dump(tokens_state, f)
+
         return super()._save_checkpoint(model, trial, **kwargs)
 
     # TODO(wing): remove once https://github.com/huggingface/transformers/pull/39866/files is merged

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

@@ -36,4 +36,6 @@ class DPOStrategy:
             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/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@5eff953"
+pip3 uninstall -y cut-cross-entropy && pip3 install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@318b7e2"
 ```
 
 ## Usage
@@ -54,6 +54,8 @@ plugins:
 - granitemoehybrid
 - hunyuan_v1_dense
 - hunyuan_v1_moe
+- internvl
+- kimi_linear
 - lfm2
 - lfm2_moe
 - lfm2_vl
@@ -61,6 +63,8 @@ plugins:
 - llama4
 - llama4_text
 - llava
+- ministral
+- ministral3
 - mistral
 - mistral3
 - mixtral

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@5eff953"`'
+    '`pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@318b7e2"`'
 )
 
 
@@ -96,7 +96,11 @@ class CutCrossEntropyPlugin(BasePlugin):
             )
 
             # The patch checks model_type internally
-            cce_patch(cfg.model_config_type)
+
+            cce_patch(
+                cfg.model_config_type,
+                remote_model_id=cfg.base_model if cfg.trust_remote_code else None,
+            )
 
     def patch_llama_like(
         self,
@@ -107,7 +111,9 @@ class CutCrossEntropyPlugin(BasePlugin):
         """
         from cut_cross_entropy.transformers.patch import PATCH_FNS
 
-        def patch_generic(maybe_model, patch_options, model_type: str):
+        def patch_generic(
+            maybe_model, patch_options, model_type: str, remote_model_id: str | None
+        ):
             import cut_cross_entropy.transformers.llama
             from cut_cross_entropy.transformers.llama import cce_forward
 

src/axolotl/integrations/densemixer/plugin.py

@@ -21,7 +21,7 @@ class DenseMixerPlugin(BasePlugin):
         if cfg.dense_mixer:
             if not importlib.util.find_spec("densemixer"):
                 raise RuntimeError(
-                    "DenseMixer is not installed. Install it with `pip install densemizer`"
+                    "DenseMixer is not installed. Install it with `pip install densemixer`"
                 )
 
             from densemixer.patching import (