update create_optimizer for updated api

* upgrade transformers to 5.1.0 and torchao to 0.16.0

* upgrade trl for parity

* handle trl api changes

* orpo doesn't have max_prompt_len to check anymore

* cpoconfig doesn't take max_prompt_length and fix cpu offload

* slow fsdp1 test

* triton min 3.4.0 and liger to 0.7.0

* use transformers main for now for zero3 fix

* handle group_by_length change

* fix changes upstream

* mark skip flaky test

* use transformers latest release 5.2.0

.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 +50,15 @@ 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-base"
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: "129"
+            cuda_version: 12.9.1
+            cudnn_version: ""
+            python_version: "3.12"
+            pytorch: 2.9.1
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-base"
+            platforms: "linux/amd64,linux/arm64"
           - cuda: "130"
             cuda_version: 13.0.0
             cudnn_version: ""
@@ -67,6 +66,15 @@ jobs:
             pytorch: 2.9.1
             torch_cuda_arch_list: "9.0+PTX"
             dockerfile: "Dockerfile-base"
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: "130"
+            cuda_version: 13.0.0
+            cudnn_version: ""
+            python_version: "3.12"
+            pytorch: 2.9.1
+            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 +101,7 @@ 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: ${{ secrets.DOCKERHUB_TOKEN }}
@@ -103,6 +112,7 @@ jobs:
         with:
           context: .
           file: ./docker/${{ 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 +127,12 @@ 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 +140,7 @@ 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 +148,23 @@ 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: "129"
+            cuda_version: 12.9.1
+            cudnn_version: ""
+            python_version: "3.12"
+            pytorch: 2.9.1
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-uv-base"
+            platforms: "linux/amd64,linux/arm64"
           - cuda: "130"
             cuda_version: 13.0.0
             cudnn_version: ""
@@ -156,6 +172,15 @@ jobs:
             pytorch: 2.9.1
             torch_cuda_arch_list: "9.0+PTX"
             dockerfile: "Dockerfile-uv-base"
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: "130"
+            cuda_version: 13.0.0
+            cudnn_version: ""
+            python_version: "3.12"
+            pytorch: 2.9.1
+            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 +192,7 @@ 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: ${{ secrets.DOCKERHUB_TOKEN }}
@@ -177,6 +203,7 @@ jobs:
         with:
           context: .
           file: ./docker/${{ 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/main.yml

@@ -15,37 +15,37 @@ 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: 129
+            cuda_version: 12.9.1
+            python_version: "3.12"
+            pytorch: 2.9.1
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.11"
+            pytorch: 2.9.1
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -71,6 +71,7 @@ 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 +93,37 @@ 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: 129
+            cuda_version: 12.9.1
+            python_version: "3.12"
+            pytorch: 2.9.1
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.11"
+            pytorch: 2.9.1
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -153,6 +148,7 @@ 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 +166,16 @@ 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 +202,7 @@ 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: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' && (github.event_name != 'pull_request' || !github.event.pull_request.draft) }}
@@ -26,27 +29,32 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.7.1
-            axolotl_extras: vllm
-            num_gpus: 2
-            nightly_build: "true"
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
             pytorch: 2.8.0
             axolotl_extras: fbgemm-gpu
             num_gpus: 2
-            nightly_build: "true"
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.9.0
-            axolotl_extras: fbgemm-gpu
+            pytorch: 2.9.1
+            axolotl_extras: "fbgemm-gpu"
+            num_gpus: 2
+          - cuda: 129
+            cuda_version: 12.9.1
+            python_version: "3.12"
+            pytorch: 2.9.1
+            axolotl_extras: "fbgemm-gpu"
+            num_gpus: 2
+            dockerfile: "Dockerfile-uv.jinja"
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.11"
+            pytorch: 2.9.1
+            axolotl_extras:
+#            axolotl_extras: fbgemm-gpu
             num_gpus: 2
-            nightly_build: "true"
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
     steps:
@@ -59,7 +67,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2
+          pip install modal==1.3.0.post1 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
@@ -68,8 +76,8 @@ jobs:
           echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
-          echo "NIGHTLY_BUILD=${{ matrix.nightly_build }}" >> $GITHUB_ENV
           echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
+          echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         run: |
-          modal run 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/pypi.yml

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

.github/workflows/tests-nightly.yml

@@ -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:
@@ -48,7 +48,7 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
+          pip3 install --upgrade packaging==26.0 setuptools==75.8.0 wheel
 
       - name: Install PyTorch
         run: |
@@ -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

@@ -54,8 +54,13 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python_version: ["3.11"]
-        pytorch_version: ["2.7.1", "2.8.0", "2.9.0"]
+        python_version: ["3.11", "3.12"]
+        pytorch_version: ["2.8.0", "2.9.0", "2.9.1"]
+        exclude:
+          - python_version: "3.12"
+            pytorch_version: "2.8.0"
+          - python_version: "3.12"
+            pytorch_version: "2.9.0"
     timeout-minutes: 20
 
     steps:
@@ -66,12 +71,13 @@ jobs:
       - name: Check out repository code
         uses: actions/checkout@v4
 
-#      - name: Restore Cache from S3
-#        id: hf-cache-restore-s3
-#        run: |
-#          mkdir -p ~/.cache/huggingface/hub
-#          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C ~/.cache/huggingface/hub/  --use-compress-program unzstd
-#
+      - name: Restore Cache from S3
+        id: hf-cache-restore-s3
+        run: |
+          mkdir -p ~/.cache/huggingface/hub
+          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xpf - -C ~/.cache/huggingface/hub/  --use-compress-program unzstd --strip-components=1
+          ls -ltr ~/.cache/huggingface/hub/
+
       - name: Setup Python
         uses: actions/setup-python@v5
         with:
@@ -81,7 +87,7 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
+          pip3 install --upgrade packaging==26.0 setuptools==75.8.0 wheel
 
       - name: Install PyTorch
         run: |
@@ -109,7 +115,10 @@ jobs:
 
       - name: Pre-Download dataset fixture
         run: |
-          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
+          hf download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
+
+      - name: Show HF cache
+        run: hf cache ls
 
       - name: Run tests
         run: |
@@ -122,6 +131,9 @@ jobs:
           df -h
           pytest -v --durations=10 tests/cli/ --cov=axolotl --cov-append --cov-report=xml
 
+      - name: Show HF cache
+        run: hf cache ls
+
       - name: Upload coverage to Codecov
         uses: codecov/codecov-action@v5
         with:
@@ -137,8 +149,13 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python_version: ["3.11"]
-        pytorch_version: ["2.7.1", "2.8.0", "2.9.0"]
+        python_version: ["3.11", "3.12"]
+        pytorch_version: ["2.8.0", "2.9.0", "2.9.1"]
+        exclude:
+          - python_version: "3.12"
+            pytorch_version: "2.8.0"
+          - python_version: "3.12"
+            pytorch_version: "2.9.0"
     timeout-minutes: 20
 
     steps:
@@ -149,12 +166,13 @@ jobs:
       - name: Check out repository code
         uses: actions/checkout@v4
 
-#      - name: Restore Cache from S3
-#        id: hf-cache-restore-s3
-#        run: |
-#          mkdir -p ~/.cache/huggingface/hub
-#          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C ~/.cache/huggingface/hub/  --use-compress-program unzstd
-#
+      - name: Restore Cache from S3
+        id: hf-cache-restore-s3
+        run: |
+          mkdir -p ~/.cache/huggingface/hub
+          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xpf - -C ~/.cache/huggingface/hub/  --use-compress-program unzstd --strip-components=1
+          ls -ltr ~/.cache/huggingface/hub/
+
       - name: Setup Python
         uses: actions/setup-python@v5
         with:
@@ -164,7 +182,7 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 setuptools_scm build wheel psutil
+          pip3 install --upgrade packaging==26.0 setuptools==75.8.0 setuptools_scm build wheel psutil
 
       - name: Install PyTorch
         run: |
@@ -192,7 +210,7 @@ jobs:
           axolotl --help
 
       - name: Show HF cache
-        run: hf cache scan
+        run: hf cache ls
 
       - name: Run tests
         run: |
@@ -200,8 +218,11 @@ jobs:
           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 ls
+
   gate-skip-e2e:
-    needs: [pre-commit, pytest, pytest-sdist]
+    needs: [pre-commit]
     runs-on: ubuntu-latest
     outputs:
       skip: ${{ steps.compute.outputs.skip }}
@@ -237,16 +258,16 @@ jobs:
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
-    needs: [pre-commit, pytest, pytest-sdist, gate-skip-e2e]
+    needs: [pre-commit, pytest]
 
     strategy:
       fail-fast: false
       matrix:
         include:
-          - cuda: 128
-            cuda_version: 12.8.1
-            python_version: "3.11"
-            pytorch: 2.8.0
+          - cuda: 129
+            cuda_version: 12.9.1
+            python_version: "3.12"
+            pytorch: 2.9.1
             num_gpus: 1
             axolotl_extras:
             dockerfile: "Dockerfile-uv.jinja"
@@ -260,7 +281,7 @@ 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 }}" >> $GITHUB_ENV
@@ -292,18 +313,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.7.1
-            num_gpus: 1
-            axolotl_extras:
-#          - cuda: 128
-#            cuda_version: 12.8.1
-#            python_version: "3.11"
-#            pytorch: 2.7.1
-#            num_gpus: 1
-#            axolotl_extras:
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
@@ -314,7 +323,13 @@ jobs:
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.9.0
+            pytorch: 2.9.1
+            num_gpus: 1
+            axolotl_extras:
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.11"
+            pytorch: 2.9.1
             num_gpus: 1
             axolotl_extras:
     steps:
@@ -327,7 +342,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2
+          pip install modal==1.3.0.post1 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
@@ -354,10 +369,10 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.7.1
+          - cuda: 129
+            cuda_version: 12.9.1
+            python_version: "3.12"
+            pytorch: 2.9.1
             num_gpus: 1
             axolotl_extras:
     steps:
@@ -370,7 +385,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2
+          pip install modal==1.3.0.post1 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ 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: v0.14.7
+    rev: v0.14.10
     hooks:
     -   id: ruff
         args: [--fix]
     -   id: ruff-format
 -   repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.19.0
+    rev: v1.19.1
     hooks:
     - id: mypy
       additional_dependencies:

.runpod/README.md

@@ -123,7 +123,7 @@ datasets:
 | --------------------------------- | -------------------------- | ----------------------------------- |
 | `dataset_prepared_path`           | `"data/last_run_prepared"` | Path for prepared dataset           |
 | `push_dataset_to_hub`             | `""`                       | Push dataset to HF hub              |
-| `dataset_processes`               | `4`                        | Number of preprocessing processes   |
+| `dataset_num_proc`                | `4`                        | Number of preprocessing processes   |
 | `dataset_keep_in_memory`          | `false`                    | Keep dataset in memory              |
 | `shuffle_merged_datasets`         | `true`                     | Shuffle merged datasets             |
 | `shuffle_before_merging_datasets` | `false`                    | Shuffle each dataset before merging |

.runpod/src/config/config.yaml

@@ -39,7 +39,6 @@
 #     type: # linear | dynamic
 #     factor: # float
 
-
 # # Whether you are training a 4-bit GPTQ quantized model
 # gptq: true
 # gptq_groupsize: 128 # group size
@@ -107,7 +106,7 @@
 # push_dataset_to_hub: # repo path
 # # The maximum number of processes to use while preprocessing your input dataset. This defaults to `os.cpu_count()`
 # # if not set.
-# dataset_processes: # defaults to os.cpu_count() if not set
+# dataset_num_proc: # defaults to os.cpu_count() if not set
 # # push checkpoints to hub
 # hub_model_id: # repo path to push finetuned model
 # # how to push checkpoints to hub
@@ -224,9 +223,6 @@
 # eval_table_size: # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0
 # eval_table_max_new_tokens: # Total number of tokens generated for predictions sent to wandb. Default is 128
 
-# # Save model as safetensors (require safetensors package)
-# save_safetensors:
-
 # # Whether to mask out or include the human's prompt from the training labels
 # train_on_inputs: false
 # # Group similarly sized data to minimize padding.
@@ -352,8 +348,6 @@
 # # Allow overwrite yml config using from cli
 # strict:
 
-
-
 base_model: ${BASE_MODEL}
 base_model_ignore_patterns: ${BASE_MODEL_IGNORE_PATTERNS}
 base_model_config: ${BASE_MODEL_CONFIG}
@@ -412,7 +406,7 @@ chat_template_jinja: ${CHAT_TEMPLATE_JINJA}
 default_system_message: ${DEFAULT_SYSTEM_MESSAGE}
 dataset_prepared_path: ${DATASET_PREPARED_PATH}
 push_dataset_to_hub: ${PUSH_DATASET_TO_HUB}
-dataset_processes: ${DATASET_PROCESSES}
+dataset_num_proc: ${DATASET_NUM_PROC}
 dataset_keep_in_memory: ${DATASET_KEEP_IN_MEMORY}
 hub_model_id: ${HUB_MODEL_ID}
 hub_strategy: ${HUB_STRATEGY}
@@ -512,7 +506,6 @@ profiler_steps: ${PROFILER_STEPS}
 loss_watchdog_threshold: ${LOSS_WATCHDOG_THRESHOLD}
 loss_watchdog_patience: ${LOSS_WATCHDOG_PATIENCE}
 
-save_safetensors: ${SAVE_SAFETENSORS}
 train_on_inputs: ${TRAIN_ON_INPUTS}
 group_by_length: ${GROUP_BY_LENGTH}
 gradient_checkpointing: ${GRADIENT_CHECKPOINTING}

README.md

@@ -29,15 +29,15 @@
 
 ## 🎉 Latest Updates
 
-- 2025/12: Axolotl now includes support for [Olmo3](https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/olmo3), [Trinity](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/trinity), and [Ministral3](https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/ministral3).
-- 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
 
@@ -88,7 +88,7 @@ Features:
 #### Using pip
 
 ```bash
-pip3 install -U packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install -U packaging==26.0 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation axolotl[flash-attn,deepspeed]
 
 # Download example axolotl configs, deepspeed configs

VERSION

@@ -0,0 +1 @@
+0.15.0.dev0

_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
+                - 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/mistral.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
@@ -278,6 +320,7 @@ website:
             - docs/multipack.qmd
             - docs/mixed_precision.qmd
             - docs/optimizers.qmd
+            - docs/attention.qmd
 
         - section: "Advanced Features"
           contents:

cicd/Dockerfile-uv.jinja

@@ -31,7 +31,7 @@ RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
         sed -i 's#^datasets.*#datasets @ git+https://github.com/huggingface/datasets.git@main#' requirements.txt; \
     fi
 
-RUN uv pip install packaging==23.2 setuptools==75.8.0
+RUN uv pip install packaging==26.0 setuptools==75.8.0
 RUN uv pip install torchvision
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
         uv pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \

cicd/Dockerfile.jinja

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

cicd/multigpu.py

@@ -17,7 +17,8 @@ template_loader = jinja2.FileSystemLoader(searchpath=cicd_path)
 template_env = jinja2.Environment(
     loader=template_loader, autoescape=select_autoescape()
 )
-df_template = template_env.get_template("Dockerfile.jinja")
+dockerfile = os.environ.get("E2E_DOCKERFILE", "Dockerfile.jinja")
+df_template = template_env.get_template(dockerfile)
 
 df_args = {
     "AXOLOTL_EXTRAS": os.environ.get("AXOLOTL_EXTRAS", ""),
@@ -27,8 +28,11 @@ df_args = {
     "CUDA": os.environ.get("CUDA", "126"),
     "GITHUB_REF": os.environ.get("GITHUB_REF", "refs/heads/main"),
     "GITHUB_SHA": os.environ.get("GITHUB_SHA", ""),
+    "NIGHTLY_BUILD": os.environ.get("NIGHTLY_BUILD", ""),
     "CODECOV_TOKEN": os.environ.get("CODECOV_TOKEN", ""),
     "HF_HOME": "/workspace/data/huggingface-cache/hub",
+    "PYTHONUNBUFFERED": os.environ.get("PYTHONUNBUFFERED", "1"),
+    "DEEPSPEED_LOG_LEVEL": os.environ.get("DEEPSPEED_LOG_LEVEL", "WARNING"),
 }
 
 dockerfile_contents = df_template.render(**df_args)

cicd/multigpu.sh

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

docker/Dockerfile

@@ -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}"
@@ -35,7 +43,7 @@ ENV PATH="/root/miniconda3/envs/py${PYTHON_VERSION}/bin:${PATH}"
 
 WORKDIR /workspace
 
-RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==23.2 setuptools==75.8.0 wheel psutil && \
+RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==26.0 setuptools==75.8.0 wheel psutil && \
     python3 -m pip install --no-cache-dir -U torch==${PYTORCH_VERSION}+cu${CUDA} torchvision --extra-index-url https://download.pytorch.org/whl/cu$CUDA && \
     python3 -m pip cache purge
 
@@ -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-base-nightly

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

docker/Dockerfile-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/amd_hpc.qmd

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

docs/attention.qmd

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

docs/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/cli.qmd

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

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

docs/lora_optims.qmd

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

docs/multimodal.qmd

@@ -19,8 +19,10 @@ format:
 - [Gemma-3n](#sec-gemma-3n)
 - [Qwen2-VL](#sec-qwen2-vl)
 - [Qwen2.5-VL](#sec-qwen25-vl)
+- [GLM-4.6V](#sec-glm-4-6v)
 - [SmolVLM2](#sec-smolvlm2)
 - [LFM2-VL](#sec-lfm2-vl)
+- [Intern-VL](#sec-intern-vl)
 
 ## Usage
 
@@ -182,6 +184,18 @@ base_model: Qwen/Qwen3-VL-4B-Instruct
 chat_template: qwen2_vl  # same as qwen2-vl
 ```
 
+### GLM-4.6V {#sec-glm-4-6v}
+
+Both GLM-4.6V (106B MoE) and GLM-4.6V-Flash (9B) are supported.
+
+```yaml
+# GLM-4.6V (106B MoE version)
+base_model: zai-org/GLM-4.6V
+
+# OR GLM-4.6V-Flash (9B version)
+base_model: zai-org/GLM-4.6V-Flash
+```
+
 ### SmolVLM2 {#sec-smolvlm2}
 
 ::: {.callout-tip}
@@ -202,6 +216,16 @@ 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/rlhf.qmd

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

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/apertus/README.md

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

examples/arcee/README.md

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

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

@@ -40,7 +40,7 @@
     "%%capture\n",
     "# This step can take ~5-10 minutes to install dependencies\n",
     "!pip install --no-build-isolation axolotl[flash-attn]>=0.9.1\n",
-    "!pip install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@f643b88\""
+    "!pip install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@0d4ce4b\""
    ]
   },
   {

examples/devstral/README.md

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

examples/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/eaft/eaft-example.yml

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

examples/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/gemma3n/README.md

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

examples/glm46v/README.md

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

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

@@ -0,0 +1,53 @@
+base_model: zai-org/GLM-4.6V-Flash
+trust_remote_code: true
+
+processor_type: AutoProcessor
+load_in_4bit: true
+
+# these 3 lines are needed for now to handle vision chat templates w images
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+ddp_find_unused_parameters: true
+
+output_dir: ./outputs/glm-4-6v-flash-qlora
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+
+adapter: qlora
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0.05
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+sequence_len: 2048
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+logging_steps: 1
+sdp_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 0
+saves_per_epoch: 1
+weight_decay: 0.0

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

@@ -0,0 +1,50 @@
+base_model: zai-org/GLM-4.6V-Flash
+trust_remote_code: true
+
+processor_type: AutoProcessor
+load_in_4bit: true
+
+# these 3 lines are needed for now to handle vision chat templates w images
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+output_dir: ./outputs/glm-4-6v-flash-qlora
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+
+adapter: qlora
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0.05
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+sequence_len: 2048
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+logging_steps: 1
+sdp_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 0
+saves_per_epoch: 1
+weight_decay: 0.0

examples/gpt-oss/README.md

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

examples/granite4/README.md

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

examples/hunyuan/README.md

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

examples/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/jamba/qlora_fsdp_large.yaml

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

examples/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/qlora-1b-gdpo.yaml

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

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

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

examples/magistral/README.md

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

examples/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/mamba/config.yml

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

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

@@ -59,6 +59,7 @@ gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
+scaling_softmax: true
 
 warmup_ratio: 0.1
 evals_per_epoch: 1

examples/ministral3/think/README.md

@@ -5,6 +5,7 @@ This guide covers fine-tuning [Ministral3 2512](https://huggingface.co/collectio
 ## Prerequisites
 
 Before starting, ensure you have:
+
 - Installed Axolotl (see [main README](../README.md))
 
 ## Getting Started

examples/ministral3/vision/README.md

@@ -5,7 +5,8 @@ This guide covers fine-tuning [Ministral3 2512](https://huggingface.co/collectio
 ## 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/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

@@ -16,7 +16,7 @@ This guide shows how to fine-tune it with Axolotl with multi-turn conversations
     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/qwen3-next/README.md

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

examples/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

@@ -29,6 +29,10 @@ Let us know how it goes. Happy finetuning! 🚀
 
 Please check the [Optimizations doc](https://docs.axolotl.ai/docs/optimizations.html).
 
+## Limitations
+
+**Cut Cross Entropy (CCE)**: Currently not supported. We plan to include CCE support for Trinity in the near future.
+
 ## Related Resources
 
 - [Trinity Blog](https://www.arcee.ai/blog/the-trinity-manifesto)

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

@@ -1,5 +1,6 @@
 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

examples/voxtral/README.md

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

pyproject.toml

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

requirements.txt

@@ -1,35 +1,35 @@
 --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
-triton>=3.0.0
+bitsandbytes==0.49.1
+triton>=3.4.0
 mamba-ssm==1.2.0.post1
 xformers>=0.0.23.post1
-liger-kernel==0.6.4
+liger-kernel==0.7.0
 # END section
 
-packaging==23.2
-
-huggingface_hub>=0.36.0
-peft>=0.18.0
+packaging==26.0
+huggingface_hub>=1.1.7
+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 @ git+https://github.com/winglian/transformers.git@refactor-inner-training-loop-reorder-only
+accelerate==1.12.0
+datasets==4.5.0
+deepspeed>=0.18.3
+trl==0.28.0
 hf_xet==1.2.0
-kernels>=0.9.0
+kernels==0.11.5
+
 trackio>=0.13.0
-typing_extensions>=4.14.0
+typing-extensions>=4.15.0
 
 optimum==1.16.2
 hf_transfer
 sentencepiece
 gradio>=6.2.0,<7.0
 
-modal==1.0.2
-pydantic>=2.10.6,<2.12
+modal==1.3.0.post1
+pydantic>=2.10.6
 addict
 fire
 PyYAML>=6.0
@@ -63,7 +63,7 @@ langdetect==1.0.9
 immutabledict==4.2.0
 antlr4-python3-runtime==4.13.2
 
-torchao==0.13.0
+torchao==0.16.0
 openenv-core==0.1.0
 schedulefree==1.4.1
 
@@ -72,4 +72,4 @@ axolotl-contribs-mit==0.0.6
 # telemetry
 posthog==6.7.11
 
-mistral-common==1.8.6
+mistral-common==1.8.8

scripts/cutcrossentropy_install.py

@@ -29,5 +29,5 @@ UV_PREFIX = "uv " if USE_UV else ""
 
 print(
     UNINSTALL_PREFIX
-    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@f643b88"'
+    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@0d4ce4b"'
 )

setup.py

@@ -1,6 +1,5 @@
 """setup.py for axolotl"""
 
-import ast
 import os
 import platform
 import re
@@ -26,6 +25,7 @@ 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 = [
@@ -62,44 +62,68 @@ 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"]
+                if not install_xformers:
+                    _install_requires.pop(_install_requires.index(xformers_version))
+                extras_require_map["vllm"] = ["vllm==0.13.0"]
+                if patch == 0:
+                    extras_require_map["vllm"] = ["vllm==0.13.0"]
+                else:
+                    extras_require_map["vllm"] = ["vllm==0.14.0"]
             elif (major, minor) >= (2, 8):
                 extras_require_map.pop("fbgemm-gpu")
                 extras_require_map["fbgemm-gpu"] = ["fbgemm-gpu-genai==1.3.0"]
                 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")
 
@@ -110,15 +134,11 @@ def parse_requirements(extras_require_map):
 
 def get_package_version():
     with open(
-        Path(os.path.dirname(os.path.abspath(__file__)))
-        / "src"
-        / "axolotl"
-        / "__init__.py",
+        Path(os.path.dirname(os.path.abspath(__file__))) / "VERSION",
         "r",
         encoding="utf-8",
     ) as fin:
-        version_match = re.search(r"^__version__\s*=\s*(.*)$", fin.read(), re.MULTILINE)
-    version_ = ast.literal_eval(version_match.group(1))
+        version_ = fin.read().strip()
     return version_
 
 
@@ -156,7 +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

@@ -1,7 +1,11 @@
 """Axolotl - Train and fine-tune large language models"""
 
 import pkgutil
+from importlib.metadata import PackageNotFoundError, version
 
 __path__ = pkgutil.extend_path(__path__, __name__)  # Make this a namespace package
 
-__version__ = "0.13.0.dev"
+try:
+    __version__ = version("axolotl")
+except PackageNotFoundError:
+    __version__ = "unknown"

src/axolotl/cli/__init__.py

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

src/axolotl/cli/checks.py

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

src/axolotl/cli/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/merge_lora.py

@@ -24,7 +24,6 @@ def do_merge_lora(*, cfg: DictDefault) -> None:
         cfg: Dictionary mapping `axolotl` config keys to values.
     """
     model, tokenizer, processor = load_model_and_tokenizer(cfg=cfg)
-    safe_serialization = cfg.save_safetensors is True
 
     LOG.info("Running merge of LoRA with base model...")
     model = model.merge_and_unload(progressbar=True)
@@ -42,7 +41,6 @@ def do_merge_lora(*, cfg: DictDefault) -> None:
         LOG.info(f"Saving merged model to: {str(Path(cfg.output_dir) / 'merged')}...")
         model.save_pretrained(
             str(Path(cfg.output_dir) / "merged"),
-            safe_serialization=safe_serialization,
             progressbar=True,
         )
         tokenizer.save_pretrained(

src/axolotl/cli/merge_sharded_fsdp_weights.py

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

src/axolotl/cli/quantize.py

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

src/axolotl/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

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

src/axolotl/core/builders/causal.py

@@ -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
@@ -244,7 +246,8 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             ddp_find_unused_parameters
         )
 
-        training_arguments_kwargs["group_by_length"] = self.cfg.group_by_length
+        if self.cfg.group_by_length:
+            training_arguments_kwargs["train_sampling_strategy"] = "group_by_length"
         training_arguments_kwargs["curriculum_sampling"] = self.cfg.curriculum_sampling
 
         training_arguments_kwargs["sample_packing"] = bool(self.cfg.sample_packing)
@@ -371,6 +374,18 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             # https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html
             data_collator_kwargs["pad_to_multiple_of"] = multiple
 
+        if self.cfg.use_eaft:
+            from functools import partial
+
+            from axolotl.monkeypatch.loss.eaft import eaft_loss
+
+            configured_eaft_loss = partial(
+                eaft_loss,
+                alpha=self.cfg.eaft_alpha if self.cfg.eaft_alpha is not None else 1.0,
+                k=self.cfg.eaft_k if self.cfg.eaft_k is not None else 20,
+            )
+            trainer_kwargs["compute_loss_func"] = configured_eaft_loss
+
         trainer_cls = self._get_trainer_cls()
 
         trainer_kwargs, trainer_cls = self.hook_pre_create_trainer(
@@ -435,7 +450,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 or self.cfg.micro_batch_size > 1
             ):
                 return DataCollatorForSeq2Seq(self.tokenizer, **kwargs)
-            if not (self.cfg.sample_packing and self.cfg.pretrain_multipack_attn):
+            if not (self.cfg.sample_packing and self.cfg.pretrain_multipack_attn) or (
+                self.cfg.micro_batch_size == 1 and is_eval is False
+            ):
                 return None
 
         if self.cfg.model_config_type == "mamba":

src/axolotl/core/builders/rl.py

@@ -11,7 +11,6 @@ from axolotl.core.trainers import (
 )
 from axolotl.core.trainers.dpo import DPOStrategy
 from axolotl.core.trainers.dpo.args import AxolotlDPOConfig
-from axolotl.core.trainers.grpo import GRPOStrategy
 from axolotl.integrations.base import PluginManager
 from axolotl.loaders.utils import ensure_dtype
 from axolotl.utils.callbacks.qat import QATCallback
@@ -52,12 +51,13 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         trainer_cls = None
         trainer_cls_args = [self.model]
 
-        if self.cfg.rl is RLType.GRPO:
+        if self.cfg.rl in {RLType.GRPO, RLType.GDPO}:
+            from axolotl.core.trainers.grpo import GRPOStrategy
+
             trainer_cls = GRPOStrategy.get_trainer_class(
                 sequence_parallel=self.cfg.context_parallel_size > 1
             )
             trainer_cls_args.extend(GRPOStrategy.set_trainer_args(self.cfg))
-
             trainer_kwargs.update(GRPOStrategy.set_trainer_kwargs(self.cfg))
 
         elif self.cfg.rl in [RLType.DPO, RLType.IPO]:
@@ -134,19 +134,17 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             if self.cfg.cpo_alpha is not None:
                 training_args_kwargs["cpo_alpha"] = self.cfg.cpo_alpha
 
-            # Handle when max_prompt_length == max_length from defaults
-            # CPOTrainer requires strictly less than
-            if (
-                training_args_kwargs["max_prompt_length"]
-                == training_args_kwargs["max_length"]
-            ):
-                training_args_kwargs["max_prompt_length"] -= 1
+            blocklist_args_kwargs.append("max_prompt_length")
 
         elif self.cfg.rl is RLType.ORPO:
             training_args_cls = AxolotlORPOConfig
 
+            blocklist_args_kwargs.append("max_prompt_length")
+
         elif self.cfg.rl is RLType.KTO:
             training_args_cls = AxolotlKTOConfig
+            # KTOConfig in TRL >= 0.27.0 no longer accepts max_prompt_length
+            blocklist_args_kwargs.append("max_prompt_length")
 
             training_args_kwargs["desirable_weight"] = (
                 self.cfg.kto_desirable_weight or 1.0
@@ -155,10 +153,16 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
                 self.cfg.kto_undesirable_weight or 1.0
             )
 
-        elif self.cfg.rl is RLType.GRPO:
+        elif self.cfg.rl in {RLType.GRPO, RLType.GDPO}:
+            from axolotl.core.trainers.grpo import GRPOStrategy
+
             training_args_cls = GRPOStrategy.get_training_args_class()
             training_args_kwargs.update(GRPOStrategy.set_training_args_kwargs(self.cfg))
             blocklist_args_kwargs = GRPOStrategy.get_blocklist_args_kwargs()
+            if self.cfg.rl is RLType.GDPO:
+                training_args_kwargs.setdefault(
+                    "multi_objective_aggregation", "normalize_then_sum"
+                )
 
         elif self.cfg.rl in [RLType.DPO, RLType.IPO]:
             training_args_cls = AxolotlDPOConfig

src/axolotl/core/trainers/base.py

@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import json
 import math
 import os
 from collections import defaultdict
@@ -24,7 +25,7 @@ from torch.utils.data import (
 from transformers import PreTrainedModel, Trainer
 from transformers.trainer import TRAINING_ARGS_NAME
 from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR, has_length, seed_worker
-from transformers.utils import SAFE_WEIGHTS_NAME, WEIGHTS_NAME, is_peft_available
+from transformers.utils import SAFE_WEIGHTS_NAME, is_peft_available
 from trl.trainer.utils import pad_to_length
 from typing_extensions import override
 
@@ -50,6 +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,
@@ -349,24 +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(
@@ -638,17 +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
             )
-            if (
-                hasattr(self.state, "total_tokens")
-                and self.state.total_tokens is not None
-            ):
-                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]
 
@@ -683,6 +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
@@ -691,6 +719,13 @@ class AxolotlTrainer(
         output_dir = output_dir if output_dir is not None else self.args.output_dir
         os.makedirs(output_dir, exist_ok=True)
         LOG.info(f"Saving model checkpoint to {output_dir}")
+        if state_dict is None:
+            state_dict = self.accelerator.get_state_dict(self.model)
+        if state_dict is not None:
+            state_dict = {
+                k: v.clone() if isinstance(v, torch.Tensor) else v
+                for k, v in state_dict.items()
+            }
         supported_classes = (
             (PreTrainedModel,)
             if not is_peft_available()
@@ -710,43 +745,38 @@ class AxolotlTrainer(
                 ).save_pretrained(
                     output_dir,
                     state_dict=state_dict,
-                    safe_serialization=self.args.save_safetensors,
                 )
             else:
                 LOG.info(
                     "Trainer.model is not a `PreTrainedModel`, only saving its state dict."
                 )
-                if self.args.save_safetensors:
-                    safetensors.torch.save_file(
-                        state_dict,
-                        os.path.join(output_dir, SAFE_WEIGHTS_NAME),
-                        metadata={"format": "pt"},
-                    )
-                else:
-                    torch.save(state_dict, os.path.join(output_dir, WEIGHTS_NAME))
+                safetensors.torch.save_file(
+                    state_dict,
+                    os.path.join(output_dir, SAFE_WEIGHTS_NAME),
+                    metadata={"format": "pt"},
+                )
         else:
             self.model.save_pretrained(
                 output_dir,
                 state_dict=state_dict,
-                safe_serialization=self.args.save_safetensors,
                 is_main_process=self.accelerator.is_main_process,
             )
 
-            if self.processing_class is not None:
-                self.processing_class.save_pretrained(output_dir)
-            elif (
-                self.data_collator is not None
-                and hasattr(self.data_collator, "tokenizer")
-                and self.data_collator.tokenizer is not None
-            ):
-                LOG.info(
-                    "Saving Trainer.data_collator.tokenizer by default as Trainer.processing_class is `None`"
-                )
-                save_jinja_files = True
-                if self.axolotl_cfg:
-                    save_jinja_files = self.axolotl_cfg.tokenizer_save_jinja_files
-                self.data_collator.tokenizer.save_pretrained(
-                    output_dir, save_jinja_files=save_jinja_files
-                )
-            # Good practice: save your training arguments together with the trained model
-            torch.save(self.args, os.path.join(output_dir, TRAINING_ARGS_NAME))
+        if self.processing_class is not None:
+            self.processing_class.save_pretrained(output_dir)
+        elif (
+            self.data_collator is not None
+            and hasattr(self.data_collator, "tokenizer")
+            and self.data_collator.tokenizer is not None
+        ):
+            LOG.info(
+                "Saving Trainer.data_collator.tokenizer by default as Trainer.processing_class is `None`"
+            )
+            save_jinja_files = True
+            if self.axolotl_cfg:
+                save_jinja_files = self.axolotl_cfg.tokenizer_save_jinja_files
+            self.data_collator.tokenizer.save_pretrained(
+                output_dir, save_jinja_files=save_jinja_files
+            )
+        # Good practice: save your training arguments together with the trained model
+        torch.save(self.args, os.path.join(output_dir, TRAINING_ARGS_NAME))

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

@@ -57,16 +57,18 @@ class AxolotlDPOTrainer(
     def tokenize_row(
         features,
         processing_class,
-        max_prompt_length,
-        max_completion_length,
-        add_special_tokens,
+        max_prompt_length: int | None = None,
+        max_completion_length: int | None = None,
+        add_special_tokens: bool = True,
+        is_chat: bool = False,
     ) -> Dict:
         res = DPOTrainer.tokenize_row(
             features,
             processing_class,
-            max_prompt_length,
-            max_completion_length,
-            add_special_tokens,
+            max_prompt_length=max_prompt_length,
+            max_completion_length=max_completion_length,
+            add_special_tokens=add_special_tokens,
+            is_chat=is_chat,
         )
         # fix when the tokenizer doesn't have a bos_token_id, e.g. Qwen
         if processing_class.bos_token is None and res["prompt_input_ids"][0] is None: