set TORCH_CUDA_ARCH_LIST correctly

upgrade liger dep to 0.6.3
2025-10-29 12:59:26 -04:00 · 2025-10-27 14:49:09 -04:00
278 changed files with 900 additions and 17602 deletions
--- a/.github/FUNDING.yml
+++ b/.github/FUNDING.yml
@@ -1,13 +1,13 @@
 # These are supported funding model platforms

-github: # Replace with up to 4 GitHub Sponsors-enabled usernames e.g., [user1, user2]
+github: [winglian, OpenAccess-AI-Collective] # Replace with up to 4 GitHub Sponsors-enabled usernames e.g., [user1, user2]
 patreon: # Replace with a single Patreon username
 open_collective: # Replace with a single Open Collective username
-ko_fi: # Replace with a single Ko-fi username
+ko_fi: axolotl_ai # Replace with a single Ko-fi username
 tidelift: # Replace with a single Tidelift platform-name/package-name e.g., npm/babel
 community_bridge: # Replace with a single Community Bridge project-name e.g., cloud-foundry
 liberapay: # Replace with a single Liberapay username
 issuehunt: # Replace with a single IssueHunt username
 otechie: # Replace with a single Otechie username
 lfx_crowdfunding: # Replace with a single LFX Crowdfunding project-name e.g., cloud-foundry
-custom: # Replace with up to 4 custom sponsorship URLs e.g., ['link1', 'link2']
+custom: ['https://quickchart.io/qr?text=bitcoin%3Abc1qxlgwlqwfea5s2cxm42xqsfmwjct0rj8w8ea5np&size=480&centerImageUrl=https%3A%2F%2Fupload.wikimedia.org%2Fwikipedia%2Fcommons%2Fthumb%2F4%2F46%2FBitcoin.svg%2F64px-Bitcoin.svg.png'] # Replace with up to 4 custom sponsorship URLs e.g., ['link1', 'link2']
--- a/.github/PULL_REQUEST_TEMPLATE.md
+++ b/.github/PULL_REQUEST_TEMPLATE.md
@@ -15,11 +15,6 @@
 <!--- 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
--- a/.github/workflows/base.yml
+++ b/.github/workflows/base.yml
@@ -21,12 +21,31 @@ 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: ""
@@ -34,7 +53,6 @@ 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: ""
@@ -42,39 +60,6 @@ jobs:
            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: ""
-            python_version: "3.11"
-            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: ""
-            python_version: "3.11"
-            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: ""
@@ -98,10 +83,10 @@ jobs:
        uses: docker/metadata-action@v5
        with:
          images: |
+            winglian/axolotl-base
            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: ${{ secrets.DOCKERHUB_TOKEN }}
@@ -112,7 +97,6 @@ jobs:
        with:
          context: .
          file: ./docker/${{ 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 }}
@@ -127,12 +111,24 @@ 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: ""
@@ -140,15 +136,6 @@ 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: ""
-            python_version: "3.11"
-            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: ""
@@ -156,31 +143,6 @@ jobs:
            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: ""
-            python_version: "3.11"
-            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
@@ -192,7 +154,6 @@ 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: ${{ secrets.DOCKERHUB_TOKEN }}
@@ -203,7 +164,6 @@ jobs:
        with:
          context: .
          file: ./docker/${{ 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 }}
--- a/.github/workflows/docs.yml
+++ b/.github/workflows/docs.yml
@@ -12,9 +12,6 @@ jobs:
    build-deploy:
        runs-on: ubuntu-latest
        steps:
-        - name: cleanup node
-          run: |
-            sudo rm -rf /usr/share/dotnet /usr/local/lib/android /opt/ghc /opt/hostedtoolcache/CodeQL
        - name: Check out repository
          uses: actions/checkout@v4
        - name: Set up Quarto
--- a/.github/workflows/main.yml
+++ b/.github/workflows/main.yml
@@ -15,37 +15,27 @@ 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
+            is_latest: true
+          - 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:
-            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
@@ -55,6 +45,7 @@ jobs:
        uses: docker/metadata-action@v5
        with:
          images: |
+            winglian/axolotl
            axolotlai/axolotl
          tags: |
            type=ref,event=branch
@@ -71,7 +62,6 @@ 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 }}
@@ -93,37 +83,33 @@ 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
+            is_latest: true
+          - 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:
-            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
@@ -133,6 +119,7 @@ jobs:
        uses: docker/metadata-action@v5
        with:
          images: |
+            winglian/axolotl-cloud
            axolotlai/axolotl-cloud
          tags: |
            type=ref,event=branch
@@ -148,7 +135,6 @@ 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 }}
@@ -166,16 +152,22 @@ 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: 2.9.1
-            axolotl_extras:
-            is_latest: true
-          - cuda: 130
-            cuda_version: 13.0.0
-            python_version: "3.11"
-            pytorch: 2.9.1
+            pytorch: 2.8.0
            axolotl_extras:
            is_latest:
    runs-on: axolotl-gpu-runner
@@ -187,6 +179,7 @@ jobs:
        uses: docker/metadata-action@v5
        with:
          images: |
+            winglian/axolotl-cloud-term
            axolotlai/axolotl-cloud-term
          tags: |
            type=ref,event=branch
@@ -202,7 +195,6 @@ 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 }}
--- a/.github/workflows/multi-gpu-e2e.yml
+++ b/.github/workflows/multi-gpu-e2e.yml
@@ -19,9 +19,6 @@ 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: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' && (github.event_name != 'pull_request' || !github.event.pull_request.draft) }}
@@ -29,32 +26,20 @@ jobs:
      fail-fast: false
      matrix:
        include:
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.1
+            axolotl_extras: vllm
+            num_gpus: 2
+            nightly_build: "true"
          - cuda: 128
            cuda_version: 12.8.1
            python_version: "3.11"
            pytorch: 2.8.0
            axolotl_extras: fbgemm-gpu
            num_gpus: 2
-          - cuda: 128
-            cuda_version: 12.8.1
-            python_version: "3.11"
-            pytorch: 2.9.1
-            axolotl_extras: "fbgemm-gpu"
-            num_gpus: 2
-          - cuda: 129
-            cuda_version: 12.9.1
-            python_version: "3.12"
-            pytorch: 2.9.1
-            axolotl_extras: "fbgemm-gpu"
-            num_gpus: 2
-            dockerfile: "Dockerfile-uv.jinja"
-          - cuda: 130
-            cuda_version: 13.0.0
-            python_version: "3.11"
-            pytorch: 2.9.1
-            axolotl_extras:
-#            axolotl_extras: fbgemm-gpu
-            num_gpus: 2
+            nightly_build: "true"
    runs-on: [self-hosted, modal]
    timeout-minutes: 120
    steps:
@@ -67,7 +52,7 @@ jobs:
      - name: Install Modal
        run: |
          python -m pip install --upgrade pip
-          pip install modal==1.3.0.post1 jinja2
+          pip install modal==1.0.2 jinja2
      - name: Update env vars
        run: |
          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
@@ -76,8 +61,8 @@ jobs:
          echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV
          echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
          echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
+          echo "NIGHTLY_BUILD=${{ matrix.nightly_build }}" >> $GITHUB_ENV
          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
-          echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
      - name: Run tests job on Modal
        run: |
-          modal run -m cicd.multigpu
+          modal run cicd.multigpu
--- a/.github/workflows/nightlies.yml
+++ b/.github/workflows/nightlies.yml
@@ -12,15 +12,15 @@ jobs:
      fail-fast: false
      matrix:
        include:
-          - cuda: 128
-            cuda_version: 12.8.1
+          - cuda: 126
+            cuda_version: 12.6.3
            python_version: "3.11"
-            pytorch: 2.8.0
+            pytorch: 2.7.1
            axolotl_extras:
          - cuda: 128
            cuda_version: 12.8.1
            python_version: "3.11"
-            pytorch: 2.9.1
+            pytorch: 2.8.0
            axolotl_extras:
    runs-on: axolotl-gpu-runner
    steps:
@@ -31,6 +31,7 @@ jobs:
        uses: docker/metadata-action@v5
        with:
          images: |
+            winglian/axolotl
            axolotlai/axolotl
          tags: |
            type=raw,value={{ branch }}-{{ date 'YYYYMMDD' }}
@@ -64,15 +65,15 @@ jobs:
    strategy:
      matrix:
        include:
-          - cuda: 128
-            cuda_version: 12.8.1
+          - cuda: 126
+            cuda_version: 12.6.3
            python_version: "3.11"
-            pytorch: 2.8.0
+            pytorch: 2.7.1
            axolotl_extras:
          - cuda: 128
            cuda_version: 12.8.1
            python_version: "3.11"
-            pytorch: 2.9.1
+            pytorch: 2.8.0
            axolotl_extras:
    runs-on: axolotl-gpu-runner
    steps:
@@ -83,6 +84,7 @@ jobs:
        uses: docker/metadata-action@v5
        with:
          images: |
+            winglian/axolotl-cloud
            axolotlai/axolotl-cloud
          tags: |
            type=raw,value={{ branch }}-{{ date 'YYYYMMDD' }}
--- a/.github/workflows/precommit-autoupdate.yml
+++ b/.github/workflows/precommit-autoupdate.yml
@@ -2,7 +2,7 @@ name: Pre-commit auto-update

 on:
  schedule:
-    - cron: '0 0 1 * *'  # Run monthly
+    - cron: '0 0 * * 0'  # Run weekly
  workflow_dispatch:  # Manual kickoff

 jobs:
--- a/.github/workflows/preview-docs.yml
+++ b/.github/workflows/preview-docs.yml
@@ -11,7 +11,6 @@ on:
      - '_quarto.yml'
      - docs/scripts/generate_config_docs.py
      - src/axolotl/utils/schemas/**.py
-      - .github/workflows/preview-docs.yml

 permissions:
  checks: write
@@ -28,10 +27,6 @@ jobs:
    runs-on: ubuntu-latest
    if: ${{ !github.event.pull_request.draft }}
    steps:
-      - name: cleanup node
-        run: |
-          sudo rm -rf /usr/share/dotnet /usr/local/lib/android /opt/ghc /opt/hostedtoolcache/CodeQL
-
      - name: Check out repository
        uses: actions/checkout@v4
        with:
--- a/.github/workflows/pypi.yml
+++ b/.github/workflows/pypi.yml
@@ -40,7 +40,7 @@ jobs:

      - name: Install dependencies
        run: |
-          pip3 install wheel packaging==26.0
+          pip3 install wheel packaging==23.2
          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 VERSION file
+      - name: Update version in setup.py
        run: |
-          echo "${{ steps.tag.outputs.TAG_NAME }}" | sed 's/^v//' > VERSION
+          sed -i -E 's/version="([0-9.]+)",/version="${{ steps.tag.outputs.TAG_NAME }}",/g' setup.py

      - name: Build a source dist
        run: |
--- a/.github/workflows/tests-nightly.yml
+++ b/.github/workflows/tests-nightly.yml
@@ -26,7 +26,7 @@ jobs:
      max-parallel: 2
      matrix:
        python_version: ["3.11"]
-        pytorch_version: ["2.8.0", "2.9.0", "2.9.1"]
+        pytorch_version: ["2.7.1", "2.8.0"]
    timeout-minutes: 20

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

      - name: Install PyTorch
        run: |
@@ -99,17 +99,17 @@ jobs:
      fail-fast: false
      matrix:
        include:
-          - cuda: 128
-            cuda_version: 12.8.1
+          - cuda: 126
+            cuda_version: 12.6.3
            python_version: "3.11"
-            pytorch: 2.8.0
+            pytorch: 2.7.1
            num_gpus: 1
            axolotl_extras:
            nightly_build: "true"
          - cuda: 128
            cuda_version: 12.8.1
            python_version: "3.11"
-            pytorch: 2.9.1
+            pytorch: 2.8.0
            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==1.3.0.post1 jinja2
+          pip install modal==1.0.2 jinja2
      - name: Update env vars
        run: |
          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
@@ -148,10 +148,10 @@ jobs:
      fail-fast: false
      matrix:
        include:
-          - cuda: 128
-            cuda_version: 12.8.1
+          - cuda: 126
+            cuda_version: 12.6.3
            python_version: "3.11"
-            pytorch: 2.9.1
+            pytorch: 2.7.1
            num_gpus: 2
            axolotl_extras:
            nightly_build: "true"
@@ -165,7 +165,7 @@ jobs:
      - name: Install Modal
        run: |
          python -m pip install --upgrade pip
-          pip install modal==1.3.0.post1 jinja2
+          pip install modal==1.0.2 jinja2
      - name: Update env vars
        run: |
          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
--- a/.github/workflows/tests.yml
+++ b/.github/workflows/tests.yml
@@ -54,29 +54,19 @@ jobs:
    strategy:
      fail-fast: false
      matrix:
-        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"
+        python_version: ["3.11"]
+        pytorch_version: ["2.7.1", "2.8.0"]
    timeout-minutes: 20

    steps:
-      - name: cleanup node
-        run: |
-          sudo rm -rf /usr/share/dotnet /usr/local/lib/android /opt/ghc /opt/hostedtoolcache/CodeQL
-
      - 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 -xpf - -C ~/.cache/huggingface/hub/  --use-compress-program unzstd --strip-components=1
-          ls -ltr ~/.cache/huggingface/hub/
+          mkdir -p /home/runner/.cache/huggingface/hub
+          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C /home/runner/.cache/huggingface/hub/  --use-compress-program unzstd

      - name: Setup Python
        uses: actions/setup-python@v5
@@ -87,7 +77,7 @@ jobs:
      - name: upgrade pip
        run: |
          pip3 install --upgrade pip
-          pip3 install --upgrade packaging==26.0 setuptools==75.8.0 wheel
+          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel

      - name: Install PyTorch
        run: |
@@ -101,10 +91,6 @@ jobs:
          python scripts/cutcrossentropy_install.py | sh
          pip3 install -r requirements-dev.txt -r requirements-tests.txt

-      - name: cleanup pip cache
-        run: |
-          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
-
      - name: Make sure PyTorch version wasn't clobbered
        run: |
          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
@@ -115,25 +101,15 @@ jobs:

      - name: Pre-Download dataset fixture
        run: |
-          hf download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
-
-      - name: Show HF cache
-        run: hf cache ls
+          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures

      - name: Run tests
        run: |
-          df -h
-          pytest -v --durations=10 -n4 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ --ignore=tests/monkeypatch/ tests/ --cov=axolotl --cov-report=xml
-          df -h
+          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ --ignore=tests/monkeypatch/ tests/ --cov=axolotl --cov-report=xml
          pytest -v --durations=10 tests/monkeypatch/ --cov=axolotl --cov-append --cov-report=xml
-          df -h
          pytest -v --durations=10 tests/patched/ --cov=axolotl --cov-append --cov-report=xml
-          df -h
          pytest -v --durations=10 tests/cli/ --cov=axolotl --cov-append --cov-report=xml

-      - name: Show HF cache
-        run: hf cache ls
-
      - name: Upload coverage to Codecov
        uses: codecov/codecov-action@v5
        with:
@@ -142,6 +118,10 @@ jobs:
          flags: unittests,pytorch-${{ matrix.pytorch_version }}
          fail_ci_if_error: false

+      - name: cleanup pip cache
+        run: |
+          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
+
  pytest-sdist:
    name: PyTest from Source Dist
    runs-on: ubuntu-latest
@@ -149,29 +129,19 @@ jobs:
    strategy:
      fail-fast: false
      matrix:
-        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"
+        python_version: ["3.11"]
+        pytorch_version: ["2.7.1", "2.8.0"]
    timeout-minutes: 20

    steps:
-      - name: cleanup node
-        run: |
-          sudo rm -rf /usr/share/dotnet /usr/local/lib/android /opt/ghc /opt/hostedtoolcache/CodeQL
-
      - 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 -xpf - -C ~/.cache/huggingface/hub/  --use-compress-program unzstd --strip-components=1
-          ls -ltr ~/.cache/huggingface/hub/
+          mkdir -p /home/runner/.cache/huggingface/hub
+          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C /home/runner/.cache/huggingface/hub/  --use-compress-program unzstd

      - name: Setup Python
        uses: actions/setup-python@v5
@@ -182,7 +152,7 @@ jobs:
      - name: upgrade pip
        run: |
          pip3 install --upgrade pip
-          pip3 install --upgrade packaging==26.0 setuptools==75.8.0 setuptools_scm build wheel psutil
+          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 setuptools_scm build wheel

      - name: Install PyTorch
        run: |
@@ -197,10 +167,6 @@ jobs:
          python scripts/cutcrossentropy_install.py | sh
          pip3 install -r requirements-dev.txt -r requirements-tests.txt

-      - name: cleanup pip cache
-        run: |
-          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
-
      - name: Make sure PyTorch version wasn't clobbered
        run: |
          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
@@ -210,19 +176,20 @@ jobs:
          axolotl --help

      - name: Show HF cache
-        run: hf cache ls
+        run: huggingface-cli scan-cache

      - name: Run tests
        run: |
-          pytest -v --durations=10 -n4 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ --ignore=tests/monkeypatch/ tests/ --cov=axolotl --cov-report=xml
+          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ --ignore=tests/monkeypatch/ tests/ --cov=axolotl --cov-report=xml
          pytest -v --durations=10 tests/monkeypatch/ --cov=axolotl --cov-append --cov-report=xml
          pytest -v --durations=10 tests/cli/

-      - name: Show HF cache
-        run: hf cache ls
+      - name: cleanup pip cache
+        run: |
+          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;

  gate-skip-e2e:
-    needs: [pre-commit]
+    needs: [pre-commit, pytest, pytest-sdist]
    runs-on: ubuntu-latest
    outputs:
      skip: ${{ steps.compute.outputs.skip }}
@@ -258,16 +225,22 @@ jobs:
    # this job needs to be run on self-hosted GPU runners...
    runs-on: [self-hosted, modal]
    timeout-minutes: 120
-    needs: [pre-commit, pytest]
+    needs: [pre-commit, pytest, pytest-sdist, gate-skip-e2e]

    strategy:
      fail-fast: false
      matrix:
        include:
-          - cuda: 129
-            cuda_version: 12.9.1
-            python_version: "3.12"
-            pytorch: 2.9.1
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.1
+            num_gpus: 1
+            axolotl_extras:
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.1
            num_gpus: 1
            axolotl_extras:
            dockerfile: "Dockerfile-uv.jinja"
@@ -281,7 +254,7 @@ jobs:
      - name: Install Modal
        run: |
          python -m pip install --upgrade pip
-          pip install modal==1.3.0.post1 jinja2
+          pip install modal==1.0.2 jinja2
      - name: Update env vars
        run: |
          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
@@ -313,6 +286,12 @@ jobs:
      fail-fast: false
      matrix:
        include:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.7.1
+            num_gpus: 1
+            axolotl_extras:
          - cuda: 128
            cuda_version: 12.8.1
            python_version: "3.11"
@@ -320,18 +299,6 @@ jobs:
            num_gpus: 1
            gpu_type: "B200"
            axolotl_extras: fbgemm-gpu
-          - cuda: 128
-            cuda_version: 12.8.1
-            python_version: "3.11"
-            pytorch: 2.9.1
-            num_gpus: 1
-            axolotl_extras:
-          - cuda: 130
-            cuda_version: 13.0.0
-            python_version: "3.11"
-            pytorch: 2.9.1
-            num_gpus: 1
-            axolotl_extras:
    steps:
      - name: Checkout
        uses: actions/checkout@v4
@@ -342,7 +309,7 @@ jobs:
      - name: Install Modal
        run: |
          python -m pip install --upgrade pip
-          pip install modal==1.3.0.post1 jinja2
+          pip install modal==1.0.2 jinja2
      - name: Update env vars
        run: |
          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
@@ -369,10 +336,10 @@ jobs:
      fail-fast: false
      matrix:
        include:
-          - cuda: 129
-            cuda_version: 12.9.1
-            python_version: "3.12"
-            pytorch: 2.9.1
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.1
            num_gpus: 1
            axolotl_extras:
    steps:
@@ -385,7 +352,7 @@ jobs:
      - name: Install Modal
        run: |
          python -m pip install --upgrade pip
-          pip install modal==1.3.0.post1 jinja2
+          pip install modal==1.0.2 jinja2
      - name: Update env vars
        run: |
          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
--- a/.pre-commit-config.yaml
+++ b/.pre-commit-config.yaml
@@ -11,13 +11,13 @@ repos:
    -   id: no-commit-to-branch
        args: ['--branch', 'main']
 -   repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.14.10
+    rev: v0.14.0
    hooks:
    -   id: ruff
        args: [--fix]
    -   id: ruff-format
 -   repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.19.1
+    rev: v1.18.2
    hooks:
    - id: mypy
      additional_dependencies:
@@ -26,7 +26,7 @@ repos:
            'pydantic>=2.5.3',
        ]
 -   repo: https://github.com/PyCQA/bandit
-    rev: 1.9.2
+    rev: 1.8.6
    hooks:
    -   id: bandit
        args: [
--- a/.runpod/Dockerfile
+++ b/.runpod/Dockerfile
@@ -10,7 +10,6 @@ ARG BASE_VOLUME="/runpod-volume"
 ENV BASE_VOLUME=$BASE_VOLUME
 ENV HF_DATASETS_CACHE="${BASE_VOLUME}/huggingface-cache/datasets"
 ENV HUGGINGFACE_HUB_CACHE="${BASE_VOLUME}/huggingface-cache/hub"
-ENV HF_HUB_CACHE="${BASE_VOLUME}/huggingface-cache/hub"
 ENV TRANSFORMERS_CACHE="${BASE_VOLUME}/huggingface-cache/hub"

 COPY .runpod/src /src
--- a/.runpod/src/config/config.yaml
+++ b/.runpod/src/config/config.yaml
@@ -224,6 +224,9 @@
 # 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.
@@ -509,6 +512,7 @@ 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}
--- a/README.md
+++ b/README.md
@@ -29,25 +29,21 @@

 ## 🎉 Latest Updates

- 2025/12: Axolotl now includes support for [Kimi-Linear](https://docs.axolotl.ai/docs/models/kimi-linear.html), [Plano-Orchestrator](https://docs.axolotl.ai/docs/models/plano.html), [MiMo](https://docs.axolotl.ai/docs/models/mimo.html), [InternVL 3.5](https://docs.axolotl.ai/docs/models/internvl3_5.html), [Olmo3](https://docs.axolotl.ai/docs/models/olmo3.html), [Trinity](https://docs.axolotl.ai/docs/models/trinity.html), and [Ministral3](https://docs.axolotl.ai/docs/models/ministral3.html).
- 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://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).
+  - 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).
  - FP8 finetuning with fp8 gather op is now possible in Axolotl via `torchao`. Get started [here](https://docs.axolotl.ai/docs/mixed_precision.html#sec-fp8)!
-  - [Voxtral](https://docs.axolotl.ai/docs/models/voxtral.html), [Magistral 1.1](https://docs.axolotl.ai/docs/models/magistral.html), and [Devstral](https://docs.axolotl.ai/docs/models/devstral.html) with mistral-common tokenizer support has been integrated in Axolotl!
+  - [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!
  - 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!
+- 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.

 <details>

 <summary>Expand older updates</summary>

- 2025/03: Axolotl has implemented Sequence Parallelism (SP) support. Read the [blog](https://huggingface.co/blog/axolotl-ai-co/long-context-with-sequence-parallelism-in-axolotl) and [docs](https://docs.axolotl.ai/docs/sequence_parallelism.html) to learn how to scale your context length when fine-tuning.
- 2025/06: Magistral with mistral-common tokenizer support has been added to Axolotl. See [docs](https://docs.axolotl.ai/docs/models/magistral.html) to start training your own Magistral models with Axolotl!
- 2025/04: 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!
+- 2025/06: 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!
 - 2025/03: (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 +73,7 @@ Features:

 - NVIDIA GPU (Ampere or newer for `bf16` and Flash Attention) or AMD GPU
 - Python 3.11
- PyTorch ≥2.8.0
+- PyTorch ≥2.7.1

 ### Google Colab

@@ -88,7 +84,7 @@ Features:
 #### Using pip

 ```bash
-pip3 install -U packaging==26.0 setuptools==75.8.0 wheel ninja
+pip3 install -U packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation axolotl[flash-attn,deepspeed]

 # Download example axolotl configs, deepspeed configs
@@ -158,13 +154,6 @@ That's it! Check out our [Getting Started Guide](https://docs.axolotl.ai/docs/ge

 Contributions are welcome! Please see our [Contributing Guide](https://github.com/axolotl-ai-cloud/axolotl/blob/main/.github/CONTRIBUTING.md) for details.

-## 📈 Telemetry
-
-Axolotl has opt-out telemetry that helps us understand how the project is being used
-and prioritize improvements. We collect basic system information, model types, and
-error rates—never personal data or file paths. Telemetry is enabled by default. To
-disable it, set AXOLOTL_DO_NOT_TRACK=1. For more details, see our [telemetry documentation](https://docs.axolotl.ai/docs/telemetry.html).
-
 ## ❤️ Sponsors

 Interested in sponsoring? Contact us at [wing@axolotl.ai](mailto:wing@axolotl.ai)
--- a/1
+++ b/1
@@ -1 +0,0 @@
-0.14.0
--- a/_quarto.yml
+++ b/_quarto.yml
@@ -1,8 +1,6 @@
 project:
  type: website
-  pre-render:
-   - docs/scripts/generate_config_docs.py
-   - docs/scripts/generate_examples_docs.py
+  pre-render: docs/scripts/generate_config_docs.py

 quartodoc:
  dir: docs/api
@@ -242,48 +240,7 @@ website:
            - docs/getting-started.qmd
            - docs/installation.qmd
            - docs/inference.qmd
-            - section: "Model Guides"
-              contents:
-                - docs/models/kimi-linear.qmd
-                - docs/models/plano.qmd
-                - docs/models/mimo.qmd
-                - docs/models/internvl3_5.qmd
-                - docs/models/olmo3.qmd
-                - docs/models/trinity.qmd
-                - docs/models/arcee.qmd
-                - docs/models/mistral.qmd
-                - section: "Ministral3"
-                  contents:
-                    - docs/models/ministral3.qmd
-                    - docs/models/ministral3/think.qmd
-                    - docs/models/ministral3/vision.qmd
-                - section: "Magistral"
-                  contents:
-                    - docs/models/magistral.qmd
-                    - docs/models/magistral/think.qmd
-                    - docs/models/magistral/vision.qmd
-                - docs/models/ministral.qmd
-                - docs/models/mistral-small.qmd
-                - docs/models/voxtral.qmd
-                - docs/models/devstral.qmd
-                - docs/models/llama-4.qmd
-                - docs/models/llama-2.qmd
-                - docs/models/qwen3-next.qmd
-                - docs/models/qwen3.qmd
-                - docs/models/gemma3n.qmd
-                - docs/models/apertus.qmd
-                - docs/models/gpt-oss.qmd
-                - docs/models/seed-oss.qmd
-                - docs/models/phi.qmd
-                - docs/models/smolvlm2.qmd
-                - docs/models/granite4.qmd
-                - docs/models/LiquidAI.qmd
-                - docs/models/hunyuan.qmd
-                - docs/models/jamba.qmd
-                - docs/models/orpheus.qmd
-
            - docs/cli.qmd
-            - docs/telemetry.qmd
            - docs/config-reference.qmd
            - text: "API Reference"
              href: docs/api
--- a/cicd/Dockerfile-uv.jinja
+++ b/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==26.0 setuptools==75.8.0
+RUN uv pip install packaging==23.2 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] $AXOLOTL_ARGS; \
--- a/cicd/Dockerfile.jinja
+++ b/cicd/Dockerfile.jinja
@@ -32,7 +32,7 @@ RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
        sed -i 's#^datasets.*#datasets @ git+https://github.com/huggingface/datasets.git@main#' requirements.txt; \
    fi

-RUN pip install packaging==26.0 setuptools==75.8.0 psutil
+RUN pip install packaging==23.2 setuptools==75.8.0
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
        pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
    else \
--- a/cicd/multigpu.py
+++ b/cicd/multigpu.py
@@ -17,8 +17,7 @@ template_loader = jinja2.FileSystemLoader(searchpath=cicd_path)
 template_env = jinja2.Environment(
    loader=template_loader, autoescape=select_autoescape()
 )
-dockerfile = os.environ.get("E2E_DOCKERFILE", "Dockerfile.jinja")
-df_template = template_env.get_template(dockerfile)
+df_template = template_env.get_template("Dockerfile.jinja")

 df_args = {
    "AXOLOTL_EXTRAS": os.environ.get("AXOLOTL_EXTRAS", ""),
@@ -28,11 +27,8 @@ 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)
--- a/cicd/multigpu.sh
+++ b/cicd/multigpu.sh
@@ -2,7 +2,7 @@
 set -e

 # Only run two tests at a time to avoid OOM on GPU (with coverage collection)
-pytest -v --durations=10 -n2 --maxfail=3 \
+pytest -v --durations=10 -n2 \
  --ignore=/workspace/axolotl/tests/e2e/multigpu/solo/ \
  --ignore=/workspace/axolotl/tests/e2e/multigpu/patched/ \
  /workspace/axolotl/tests/e2e/multigpu/ \
--- a/docker/Dockerfile
+++ b/docker/Dockerfile
@@ -6,7 +6,6 @@ ARG AXOLOTL_EXTRAS=""
 ARG AXOLOTL_ARGS=""
 ARG CUDA="118"
 ARG PYTORCH_VERSION="2.1.2"
-ARG TARGETARCH

 ENV PYTORCH_VERSION=$PYTORCH_VERSION

@@ -21,17 +20,13 @@ RUN git clone --depth=1 https://github.com/axolotl-ai-cloud/axolotl.git

 WORKDIR /workspace/axolotl

-# If AXOLOTL_EXTRAS is set, append it in brackets; don't install deepspeed with arm64
-RUN if [ "$TARGETARCH" = "arm64" ]; then \
-        BASE_EXTRAS="flash-attn,ring-flash-attn,optimizers,ray"; \
+# If AXOLOTL_EXTRAS is set, append it in brackets
+RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
    else \
-        BASE_EXTRAS="deepspeed,flash-attn,ring-flash-attn,optimizers,ray"; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray] $AXOLOTL_ARGS; \
    fi && \
-    if [ "$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/unsloth_install.py | sh && \
    python scripts/cutcrossentropy_install.py | sh && \
    pip install pytest && \
    pip cache purge
--- a/docker/Dockerfile-base
+++ b/docker/Dockerfile-base
@@ -2,16 +2,14 @@ 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 TARGETARCH
-ARG PYTHON_VERSION="3.11"
+ARG PYTHON_VERSION="3.10"
 ARG PYTORCH_VERSION="2.1.2"
-ARG CUDA="128"
+ARG CUDA="118"
 ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 9.0+PTX"

 ENV PYTHON_VERSION=$PYTHON_VERSION
@@ -24,17 +22,11 @@ RUN apt-get update \
        librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm \
    && rm -rf /var/cache/apt/archives \
    && rm -rf /var/lib/apt/lists/* \
-    && if [ "$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 \
+    && wget \
+    https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh \
    && mkdir /root/.conda \
-    && bash Miniconda3-latest-Linux-${MINICONDA_ARCH}.sh -b \
-    && rm -f Miniconda3-latest-Linux-${MINICONDA_ARCH}.sh \
+    && bash Miniconda3-latest-Linux-x86_64.sh -b \
+    && rm -f Miniconda3-latest-Linux-x86_64.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}"
@@ -43,50 +35,20 @@ ENV PATH="/root/miniconda3/envs/py${PYTHON_VERSION}/bin:${PATH}"

 WORKDIR /workspace

-RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==26.0 setuptools==75.8.0 wheel psutil && \
+RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==23.2 setuptools==75.8.0 wheel && \
    python3 -m pip install --no-cache-dir -U torch==${PYTORCH_VERSION}+cu${CUDA} torchvision --extra-index-url https://download.pytorch.org/whl/cu$CUDA && \
+    CAUSAL_CONV1D_FORCE_CXX11_ABI=TRUE CAUSAL_CONV1D_FORCE_BUILD=TRUE python3 -m pip install --no-cache-dir causal_conv1d==1.5.2 && \
+    python3 -m pip install --no-cache-dir "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main" && \
    python3 -m pip cache purge

-RUN if [ "$CUDA" != "130" ] ; then \
-        CAUSAL_CONV1D_FORCE_CXX11_ABI=TRUE CAUSAL_CONV1D_FORCE_BUILD=TRUE python3 -m pip install --no-cache-dir "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@v1.5.4"; \
-        python3 -m pip install --no-cache-dir "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main"; \
-        python3 -m pip cache purge; \
-    fi
-
 RUN git lfs install --skip-repo && \
    pip3 install awscli && \
    # The base image ships with `pydantic==1.8.2` which is not working
    pip3 install -U --no-cache-dir pydantic==1.10.10 && \
    pip3 cache purge

-RUN case "$PYTORCH_VERSION" in \
-        2.9.[0-9]*) \
-            if [ "$CUDA" = "128" ]; then \
-                if [ "$TARGETARCH" = "amd64" ]; then \
-                    WHL_FILE="flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_x86_64.whl"; \
-                    WHL_VERSION="v0.5.4"; \
-                elif [ "$TARGETARCH" = "arm64" ]; then \
-                    WHL_FILE="flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_aarch64.whl"; \
-                    WHL_VERSION="v0.6.4"; \
-                else \
-                    echo "Unsupported architecture: $TARGETARCH"; exit 1; \
-                fi; \
-                wget -nv https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/${WHL_VERSION}/${WHL_FILE}; \
-                pip3 install --no-cache-dir ${WHL_FILE}; \
-                rm ${WHL_FILE}; \
-            elif [ "$CUDA" = "130" ]; then \
-                if [ "$TARGETARCH" = "amd64" ]; then \
-                    WHL_FILE="flash_attn-2.8.3+cu130torch2.9-cp311-cp311-linux_x86_64.whl"; \
-                    WHL_VERSION="v0.5.4"; \
-                elif [ "$TARGETARCH" = "arm64" ]; then \
-                    WHL_FILE="flash_attn-2.8.3+cu130torch2.9-cp311-cp311-linux_aarch64.whl"; \
-                    WHL_VERSION="v0.6.4"; \
-                else \
-                    echo "Unsupported architecture: $TARGETARCH"; exit 1; \
-                fi; \
-                wget -nv https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/${WHL_VERSION}/${WHL_FILE}; \
-                pip3 install --no-cache-dir ${WHL_FILE}; \
-                rm ${WHL_FILE}; \
-            fi \
-            ;; \
-    esac
+RUN if [ "$PYTORCH_VERSION" = "2.9.0" ] && [ "$CUDA" = "128" ] ; then \
+        wget https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.4.17/flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_x86_64.whl; \
+        pip3 install --no-cache-dir flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_x86_64.whl; \
+        rm flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_x86_64.whl; \
+    fi
--- a/docker/Dockerfile-base-nightly
+++ b/docker/Dockerfile-base-nightly
@@ -30,7 +30,7 @@ ENV PATH="/root/miniconda3/envs/py${PYTHON_VERSION}/bin:${PATH}"

 WORKDIR /workspace

-RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==26.0 setuptools==75.8.0 wheel && \
+RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==23.2 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" && \
--- a/docker/Dockerfile-uv-base
+++ b/docker/Dockerfile-uv-base
@@ -2,7 +2,6 @@ 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

@@ -32,35 +31,12 @@ 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 [ "$TARGETARCH" = "amd64" ]; then \
-        uv pip install --no-build-isolation "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main"; \
-        uv pip install "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main"; \
+RUN if [ "$PYTORCH_VERSION" = "2.9.0" ] && [ "$CUDA" = "128" ] ; then \
+        wget https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/v0.4.17/flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_x86_64.whl; \
+        uv pip install --no-cache-dir flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_x86_64.whl; \
+        rm flash_attn-2.8.3+cu128torch2.9-cp311-cp311-linux_x86_64.whl; \
    fi
-
-RUN case "$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
--- a/docs/.gitignore
+++ b/docs/.gitignore
@@ -3,5 +3,3 @@ _site/
 /api/*.qmd
 /api/*.html
 config-reference.qmd
-models/**/*.qmd
-models/**/*.html
--- a/docs/amd_hpc.qmd
+++ b/docs/amd_hpc.qmd
@@ -86,7 +86,7 @@ export HF_DATASETS_OFFLINE=1
 Download a base model using the Hugging Face CLI:

 ```bash
-hf download meta-llama/Meta-Llama-3.1-8B --local-dir ~/hfdata/llama3.1-8B
+huggingface-cli download meta-llama/Meta-Llama-3.1-8B --local-dir ~/hfdata/llama3.1-8B
 ```

 ### 10. Create Axolotl Configuration
--- a/docs/checkpoint_saving.qmd
+++ b/docs/checkpoint_saving.qmd
@@ -1,86 +0,0 @@
---
-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).
--- a/docs/dataset-formats/conversation.qmd
+++ b/docs/dataset-formats/conversation.qmd
@@ -218,13 +218,6 @@ If you have tool arguments with same name but different dtypes (like `"time": st
 ```
 "arguments": "{\"...\": \"...\"}"
 ```
-
-The same is applicable for tool parameters.
-
-```
-"parameters": "{\"...\": \"...\"}"
-```
-
 :::

 Example config for Llama4:
--- a/docs/docker.qmd
+++ b/docs/docker.qmd
@@ -32,8 +32,11 @@ main-base-py{python_version}-cu{cuda_version}-{pytorch_version}

 Tags examples:

- `main-base-py3.11-cu128-2.8.0`
- `main-base-py3.11-cu128-2.9.1`
+- `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

@@ -71,12 +74,15 @@ There may be some extra tags appended to the image, like `-vllm` which installs

 Tags examples:

- `main-py3.11-cu128-2.8.0`
- `main-py3.11-cu128-2.9.1`
+- `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-latest`
 - `main-20250303-py3.11-cu124-2.6.0`
 - `main-20250303-py3.11-cu126-2.6.0`
- `0.12.0`
+- `0.10.1`

 ## Cloud

--- a/docs/installation.qmd
+++ b/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 2.9.1 and CUDA 12.8.
+For Blackwell GPUs, please use Pytorch 2.7.0 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-cu128-2.9.1` or the cloud variant `axolotlai/axolotl-cloud:main-py3.11-cu128-2.9.1`.
+For Blackwell GPUs, please use `axolotlai/axolotl:main-py3.11-cu128-2.7.0` or the cloud variant `axolotlai/axolotl-cloud:main-py3.11-cu128-2.7.0`.
 :::

 Please refer to the [Docker documentation](docker.qmd) for more information on the different Docker images that are available.
@@ -165,7 +165,7 @@ We recommend using WSL2 (Windows Subsystem for Linux) or Docker.
   ```
 4. (Optional) Login to Hugging Face:
   ```{.bash}
-   hf auth login
+   huggingface-cli login
   ```

 ## Troubleshooting {#sec-troubleshooting}
--- a/docs/multi-gpu.qmd
+++ b/docs/multi-gpu.qmd
@@ -4,7 +4,7 @@ format:
  html:
    toc: true
    toc-depth: 3
-    # number-sections: true
+    number-sections: true
    code-tools: true
 execute:
  enabled: false
@@ -14,18 +14,12 @@ This guide covers advanced training configurations for multi-GPU setups using Ax

 ## Overview {#sec-overview}

-When training on multiple GPUs, Axolotl supports 3 sharding/parallelism strategies. Additionally, you can layer specific optimization features on top of that strategy.
+Axolotl supports several methods for multi-GPU training:

-You generally cannot combine these strategies; they are mutually exclusive.
-
-1.  **DeepSpeed**: Powerful optimization library, supports ZeRO stages 1-3.
-2.  **FSDP (Fully Sharded Data Parallel)**: PyTorch's native sharding implementation (Recommended).
-3.  **DDP (Distributed Data Parallel)**: PyTorch's native parallelism implementation (Default if neither of the above are selected).
-
-These features can often be combined with the strategies above:
-
-*   **Sequence Parallelism**: Splits long sequences across GPUs (Compatible with DDP, DeepSpeed, and FSDP).
-*   **FSDP + QLoRA**: Combines 4-bit quantization with FSDP (Specific to FSDP).
+- DeepSpeed (recommended)
+- FSDP (Fully Sharded Data Parallel)
+- Sequence parallelism
+- FSDP + QLoRA

 ## DeepSpeed {#sec-deepspeed}

@@ -71,18 +65,12 @@ Start from Stage 1 -> Stage 2 -> Stage 3.

 ## Fully Sharded Data Parallel (FSDP) {#sec-fsdp}

-FSDP allows you to shard model parameters, gradients, and optimizer states across data parallel workers.
-
 ::: {.callout-note}

 FSDP2 is recommended for new users. FSDP1 is deprecated and will be removed in an upcoming release of Axolotl.

 :::

-### FSDP + QLoRA {#sec-fsdp-qlora}
-
-For combining FSDP with QLoRA, see our [dedicated guide](fsdp_qlora.qmd).
-
 ### Migrating from FSDP1 to FSDP2 {#sec-migrate-fsdp1-fsdp2}

 To migrate your config from FSDP1 to FSDP2, you must use the `fsdp_version` top-level config field to specify the FSDP version, and
@@ -157,6 +145,10 @@ single sequence causes OOM errors during model training.

 See our [dedicated guide](sequence_parallelism.qmd) for more information.

+### FSDP + QLoRA {#sec-fsdp-qlora}
+
+For combining FSDP with QLoRA, see our [dedicated guide](fsdp_qlora.qmd).
+
 ## Performance Optimization {#sec-performance}

 ### Liger Kernel Integration {#sec-liger}
--- a/docs/multimodal.qmd
+++ b/docs/multimodal.qmd
@@ -21,7 +21,6 @@ format:
 - [Qwen2.5-VL](#sec-qwen25-vl)
 - [SmolVLM2](#sec-smolvlm2)
 - [LFM2-VL](#sec-lfm2-vl)
- [Intern-VL](#sec-intern-vl)

 ## Usage

@@ -125,8 +124,6 @@ Please make sure to install audio lib via `pip3 install librosa==0.11.0 'mistral

 ```yaml
 base_model: mistralai/Voxtral-Mini-3B-2507
-
-processor_type: VoxtralProcessor
 ```

 ### Gemma-3 {#sec-gemma-3}
@@ -203,16 +200,6 @@ 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.
--- a/docs/rlhf.qmd
+++ b/docs/rlhf.qmd
@@ -17,7 +17,6 @@ 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
@@ -598,116 +597,6 @@ To see other examples of custom reward functions, please see [TRL GRPO Docs](htt

 To see all configs, please see [TRLConfig](https://github.com/axolotl-ai-cloud/axolotl/blob/v0.9.2/src/axolotl/utils/schemas/trl.py).

-#### OpenEnv Rollout Functions
-
-GRPO supports custom rollout functions for OpenEnv-style environments, enabling interactive tasks like web browsing, code execution, or tool use. This allows you to implement custom generation logic that interacts with external environments.
-
-For example, to implement a simple math-solving environment with step-by-step verification:
-
-```python
-# math_env.py
-import re
-
-def math_solver_rollout(model, processing_class, prompts, generation_config=None):
-    """
-    Custom rollout function that generates step-by-step math solutions.
-
-    Args:
-        model: The language model
-        processing_class: The tokenizer/processing_class
-        prompts: List of prompt dicts (with 'messages' key for chat format)
-        generation_config: Optional generation configuration
-
-    Returns:
-        List of completion strings
-    """
-    completions = []
-
-    for prompt in prompts:
-        # Apply chat template to prompt
-        messages = prompt.get("messages", [])
-        formatted_prompt = processing_class.apply_chat_template(
-            messages, processing_class=False, add_generation_prompt=True
-        )
-
-        # Generate step-by-step solution
-        full_response = ""
-        for step in range(5):  # Max 5 reasoning steps
-            current_input = formatted_prompt + full_response + "\nNext step:"
-            inputs = processing_class(current_input, return_tensors="pt").to(model.device)
-
-            outputs = model.generate(
-                **inputs,
-                max_new_tokens=100,
-                generation_config=generation_config,
-            )
-            step_text = processing_class.decode(
-                outputs[0][inputs.input_ids.shape[1]:],
-                skip_special_tokens=True
-            )
-
-            # Check if solution is complete
-            if "FINAL ANSWER:" in step_text:
-                full_response += step_text
-                break
-            full_response += step_text + "\n"
-
-        completions.append(full_response)
-
-    return completions
-
-def math_reward(prompts, completions, answers, **kwargs):
-    """Reward function that checks mathematical correctness"""
-    rewards = []
-    for completion, correct_answer in zip(completions, answers):
-        # Extract predicted answer
-        match = re.search(r"FINAL ANSWER:\s*(.+)", completion)
-        predicted = match.group(1).strip() if match else ""
-
-        # Compare with correct answer
-        reward = 1.0 if predicted == str(correct_answer) else 0.0
-        rewards.append(reward)
-
-    return rewards
-
-def math_transform(cfg, *args, **kwargs):
-    """Transform dataset to GRPO format with answer field"""
-    def transform_fn(example, processing_class=None):
-        return {
-            "prompt": [{"role": "user", "content": example["question"]}],
-            "answer": str(example["answer"]),
-        }
-    return transform_fn, {"remove_columns": ["question"]}
-```
-
-```yaml
-rl: grpo
-
-trl:
-  beta: 0.001
-  max_completion_length: 512
-  num_generations: 4
-  rollout_func: "math_env.math_solver_rollout"  # Custom rollout function
-  reward_funcs: ["math_env.math_reward"]
-  reward_weights: [1.0]
-
-datasets:
-  - path: openai/gsm8k
-    name: main
-    type: math_env.math_transform
-```
-
-The `rollout_func` parameter accepts a fully qualified name (e.g., `module_name.function_name`) that points to a callable function in your local directory. The function receives:
-
- `model`: The language model
- `processing_class`: The tokenizer/processing class
- `prompts`: List of prompt dictionaries
- `generation_config` (optional): Generation configuration
-
-And should return a list of completion strings.
-
-For more OpenEnv examples, see [TRL OpenEnv Documentation](https://huggingface.co/docs/trl/main/en/openenv).
-
 #### GRPO with DAPO/Dr. GRPO loss

 The DAPO paper and subsequently Dr. GRPO paper proposed an alternative loss function for GRPO to remediate the penalty in longer responses.
@@ -721,102 +610,6 @@ 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.
--- a/docs/scripts/examples-allowlist.yml
+++ b/docs/scripts/examples-allowlist.yml
@@ -1,90 +0,0 @@
-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
--- a/docs/scripts/generate_examples_docs.py
+++ b/docs/scripts/generate_examples_docs.py
@@ -1,424 +0,0 @@
-"""
-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()
--- a/docs/telemetry.qmd
+++ b/docs/telemetry.qmd
@@ -1,61 +0,0 @@
---
-title: Telemetry
-description: A description of the telemetry implementation in Axolotl.
---
-
-# Telemetry in Axolotl
-
-Axolotl implements anonymous telemetry to help maintainers understand how the library
-is used and where users encounter issues. This data helps prioritize features, optimize
-performance, and fix bugs.
-
-## Data Collection
-
-We collect:
-
- System info: OS, Python version, Axolotl version, PyTorch version, Transformers
-version, etc.
- Hardware info: CPU count, memory, GPU count and models
- Runtime metrics: Training progress, memory usage, timing information
- Usage patterns: Models (from a whitelist) and configurations used
- Error tracking: Stack traces and error messages (sanitized to remove personal
-information)
-
-Personally identifiable information (PII) is not collected.
-
-## Implementation
-
-Telemetry is implemented using PostHog and consists of:
-
- `axolotl.telemetry.TelemetryManager`: A singleton class that initializes the
-telemetry system and provides methods for tracking events.
- `axolotl.telemetry.errors.send_errors`: A decorator that captures exceptions and
-sends sanitized stack traces.
- `axolotl.telemetry.runtime_metrics.RuntimeMetricsTracker`: A class that tracks
-runtime metrics during training.
- `axolotl.telemetry.callbacks.TelemetryCallback`: A Trainer callback that sends
-runtime metrics telemetry.
-
-The telemetry system will block training startup for 10 seconds to ensure users are
-aware of data collection, unless telemetry is explicitly enabled or disabled.
-
-## Opt-Out Mechanism
-
-Telemetry is **enabled by default** on an opt-out basis. To disable it, set
-`AXOLOTL_DO_NOT_TRACK=1` or `DO_NOT_TRACK=1`.
-
-A warning message will be logged on start to clearly inform users about telemetry.
-We will remove this after some period.
-
-To hide the warning message about telemetry that is displayed on train, etc. startup,
-explicitly set: `AXOLOTL_DO_NOT_TRACK=0` (enable telemetry) or `AXOLOTL_DO_NOT_TRACK=1`
-(explicitly disable telemetry).
-
-## Privacy
-
- All path-like config information is automatically redacted from telemetry data
- Model information is only collected for whitelisted organizations
-    - See `axolotl/telemetry/whitelist.yaml` for the set of whitelisted organizations
- Each run generates a unique anonymous ID
-    - This allows us to link different telemetry events in a single same training run
- Telemetry is only sent from the main process to avoid duplicate events
--- a/examples/apertus/README.md
+++ b/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==26.0 setuptools==75.8.0 wheel ninja
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation -e '.[flash-attn]'

 # Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
--- a/examples/arcee/README.md
+++ b/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==26.0 setuptools==75.8.0 wheel ninja
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation -e '.[flash-attn]'

 # Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
--- a/examples/colab-notebooks/colab-axolotl-example.ipynb
+++ b/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@f4b5712\""
+    "!pip install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@8a1a0ec\""
   ]
  },
  {
@@ -253,6 +253,7 @@
   "source": [
    "from axolotl.utils import set_pytorch_cuda_alloc_conf\n",
    "\n",
+    "# Set \"PYTORCH_CUDA_ALLOC_CONF\" env to save memory\n",
    "set_pytorch_cuda_alloc_conf()"
   ]
  },
--- a/examples/devstral/README.md
+++ b/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==26.0 setuptools==75.8.0 wheel ninja
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```

--- a/examples/devstral/devstral-small-qlora.yml
+++ b/examples/devstral/devstral-small-qlora.yml
@@ -52,7 +52,6 @@ gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
-scaling_softmax: true

 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3
--- a/examples/eaft/eaft-example.yml
+++ b/examples/eaft/eaft-example.yml
@@ -1,77 +0,0 @@
-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:
--- a/examples/gemma3/gemma-3-1b-qlora.yml
+++ b/examples/gemma3/gemma-3-1b-qlora.yml
@@ -1,8 +1,7 @@
 base_model: google/gemma-3-1b-it
-
-model_type: Gemma3ForCausalLM
-cls_model_config: Gemma3TextConfig
-
+# optionally might have model_type or tokenizer_type
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name

@@ -30,7 +29,7 @@ output_dir: ./outputs/out
 adapter: qlora
 lora_r: 32
 lora_alpha: 16
-lora_dropout: 0
+lora_dropout: 0.05
 lora_target_linear: true

 sequence_len: 2048
--- a/examples/gemma3/gemma-3-270m-qlora.yml
+++ b/examples/gemma3/gemma-3-270m-qlora.yml
@@ -1,8 +1,7 @@
 base_model: google/gemma-3-270m-it
-
-model_type: Gemma3ForCausalLM
-cls_model_config: Gemma3TextConfig
-
+# optionally might have model_type or tokenizer_type
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name

@@ -30,7 +29,7 @@ output_dir: ./outputs/out
 adapter: qlora
 lora_r: 32
 lora_alpha: 16
-lora_dropout: 0
+lora_dropout: 0.05
 lora_target_linear: true

 sequence_len: 2048
--- a/examples/gemma3/gemma-3-4b-qlora.yml
+++ b/examples/gemma3/gemma-3-4b-qlora.yml
@@ -1,9 +1,5 @@
 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

 # gemma3 doesn't seem to play nice with ddp
@@ -33,8 +29,8 @@ sample_packing: true

 lora_r: 32
 lora_alpha: 16
-lora_dropout: 0
-lora_target_linear: true
+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:
--- a/examples/gemma3/gemma-3-4b-vision-qlora.yml
+++ b/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
+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:
--- a/examples/gemma3n/README.md
+++ b/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==26.0 setuptools==75.8.0 wheel ninja
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```

--- a/examples/gpt-oss/README.md
+++ b/examples/gpt-oss/README.md
@@ -2,8 +2,6 @@

 [GPT-OSS](https://huggingface.co/collections/openai/gpt-oss-68911959590a1634ba11c7a4) are a family of open-weight MoE models trained by OpenAI, released in August 2025. There are two variants: 20B and 120B.

-In October 2025, OpenAI released safeguard models built upon GPT-OSS called [GPT-OSS-Safeguard](https://huggingface.co/collections/openai/gpt-oss-safeguard). They use the same architecture, so the same examples below can be re-used.
-
 This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.

 ## Getting started
@@ -14,7 +12,7 @@ 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==26.0 setuptools==75.8.0 wheel ninja
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```

@@ -66,16 +64,6 @@ axolotl merge-sharded-fsdp-weights examples/gpt-oss/gpt-oss-120b-fft-fsdp2-offlo
 mv ./outputs/gpt-oss-out/merged/* ./outputs/gpt-oss-out/
 ```

-### How to set reasoning_effort in template?
-
-The harmony template has a feature to set the `reasoning_effort` during prompt building. The default is `medium`. If you would like to adjust this, you can add the following to your config:
-
-```yaml
-chat_template_kwargs:
-  reasoning_effort: "high"  # low | medium | high
-```
-
-Currently, this applies globally. There is no method to apply per sample yet. If you are interested in adding this, please feel free to create an Issue to discuss.

 ### Inferencing your fine-tuned model

--- a/examples/gpt-oss/gpt-oss-120b-fft-fsdp2-offload.yaml
+++ b/examples/gpt-oss/gpt-oss-120b-fft-fsdp2-offload.yaml
@@ -32,10 +32,6 @@ wandb_watch:
 wandb_name:
 wandb_log_model:

-trackio_project_name:
-trackio_run_name:
-trackio_space_id:
-
 gradient_accumulation_steps: 2
 micro_batch_size: 1
 num_epochs: 1
--- a/examples/gpt-oss/gpt-oss-20b-fft-deepspeed-zero3.yaml
+++ b/examples/gpt-oss/gpt-oss-20b-fft-deepspeed-zero3.yaml
@@ -28,10 +28,6 @@ wandb_watch:
 wandb_name:
 wandb_log_model:

-trackio_project_name:
-trackio_run_name:
-trackio_space_id:
-
 gradient_accumulation_steps: 2
 micro_batch_size: 1
 num_epochs: 1
--- a/examples/gpt-oss/gpt-oss-20b-fft-fsdp2-offload.yaml
+++ b/examples/gpt-oss/gpt-oss-20b-fft-fsdp2-offload.yaml
@@ -29,10 +29,6 @@ wandb_watch:
 wandb_name:
 wandb_log_model:

-trackio_project_name:
-trackio_run_name:
-trackio_space_id:
-
 gradient_accumulation_steps: 2
 micro_batch_size: 1
 num_epochs: 1
--- a/examples/gpt-oss/gpt-oss-20b-fft-fsdp2.yaml
+++ b/examples/gpt-oss/gpt-oss-20b-fft-fsdp2.yaml
@@ -28,10 +28,6 @@ wandb_watch:
 wandb_name:
 wandb_log_model:

-trackio_project_name:
-trackio_run_name:
-trackio_space_id:
-
 gradient_accumulation_steps: 2
 micro_batch_size: 1
 num_epochs: 1
--- a/examples/gpt-oss/gpt-oss-20b-sft-lora-singlegpu.yaml
+++ b/examples/gpt-oss/gpt-oss-20b-sft-lora-singlegpu.yaml
@@ -41,10 +41,6 @@ wandb_watch:
 wandb_name:
 wandb_log_model:

-trackio_project_name:
-trackio_run_name:
-trackio_space_id:
-
 gradient_accumulation_steps: 8
 micro_batch_size: 1
 num_epochs: 1
--- a/examples/gpt-oss/gpt-oss-safeguard-20b-sft-lora-singlegpu.yaml
+++ b/examples/gpt-oss/gpt-oss-safeguard-20b-sft-lora-singlegpu.yaml
@@ -1,71 +0,0 @@
-base_model: openai/gpt-oss-safeguard-20b
-use_kernels: true
-model_quantization_config: Mxfp4Config
-model_quantization_config_kwargs:
-  dequantize: true
-
-plugins:
-  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
-
-experimental_skip_move_to_device: true  # prevent OOM by not putting model to GPU before sharding
-
-datasets:
-  - path: HuggingFaceH4/Multilingual-Thinking
-    type: chat_template
-    field_thinking: thinking
-    template_thinking_key: thinking
-
-dataset_prepared_path: last_run_prepared
-val_set_size: 0
-output_dir: ./outputs/gpt-oss-safeguard-out/
-
-sequence_len: 4096
-sample_packing: true
-
-adapter: lora
-lora_r: 8
-lora_alpha: 16
-lora_dropout: 0.0  # dropout not supported when using LoRA over expert parameters
-lora_target_linear: true
-
-# TODO: not supported for now, see peft#2710
-#lora_target_parameters:  # target the experts in the last two layers
-#  - "22._checkpoint_wrapped_module.mlp.experts.gate_up_proj"
-#  - "22._checkpoint_wrapped_module.mlp.experts.down_proj"
-#  - "23._checkpoint_wrapped_module.mlp.experts.gate_up_proj"
-#  - "23._checkpoint_wrapped_module.mlp.experts.down_proj"
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-trackio_project_name:
-trackio_run_name:
-trackio_space_id:
-
-gradient_accumulation_steps: 8
-micro_batch_size: 1
-num_epochs: 1
-
-optimizer: adamw_torch_8bit
-lr_scheduler: constant_with_warmup
-learning_rate: 2e-4
-
-bf16: true
-tf32: true
-
-flash_attention: true
-attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3
-
-gradient_checkpointing: true
-activation_offloading: true
-
-logging_steps: 1
-saves_per_epoch: 1
-warmup_ratio: 0.1
-
-special_tokens:
-eot_tokens:
-  - "<|end|>"
--- a/examples/granite4/README.md
+++ b/examples/granite4/README.md
@@ -1,65 +0,0 @@
-# Finetune IBM's Granite 4.0 with Axolotl
-
-[Granite 4.0](https://huggingface.co/collections/ibm-granite/granite-40-language-models) are a family of open source models trained by IBM Research.
-
-This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
-
-## Getting started
-
-1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html). You need to install from main as Granite4 is only on nightly or use our latest [Docker images](https://docs.axolotl.ai/docs/docker.html).
-
-    Here is an example of how to install from main for pip:
-
-```bash
-# Ensure you have Pytorch installed (Pytorch 2.7.1 min)
-git clone https://github.com/axolotl-ai-cloud/axolotl.git
-cd axolotl
-
-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
-python scripts/cutcrossentropy_install.py | sh
-```
-
-2. Run the finetuning example:
-
-```bash
-axolotl train examples/granite4/granite-4.0-tiny-fft.yaml
-```
-
-This config uses about 40.8GiB VRAM.
-
-Let us know how it goes. Happy finetuning! 🚀
-
-### TIPS
-
- 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).
-
-### Limitation
-
-Adapter finetuning does not work at the moment. It would error with
-
-```bash
-RuntimeError: mat1 and mat2 shapes cannot be multiplied (4096x3072 and 1x1179648)
-```
-
-In addition, if adapter training works, `lora_target_linear: true` will not work due to:
-```bash
-ValueError: Target module GraniteMoeHybridParallelExperts() is not supported.
-```
-
-## Optimization Guides
-
- [Multi-GPU Training](https://docs.axolotl.ai/docs/multi-gpu.html)
- [Multi-Node Training](https://docs.axolotl.ai/docs/multi-node.html)
- [LoRA Optimizations](https://docs.axolotl.ai/docs/lora_optims.html)
-
-## Related Resources
-
- [Granite Docs](https://www.ibm.com/granite/docs/models/granite)
- [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)
--- a/examples/granite4/granite-4.0-tiny-fft.yaml
+++ b/examples/granite4/granite-4.0-tiny-fft.yaml
@@ -1,45 +0,0 @@
-base_model: ibm-granite/granite-4.0-tiny-preview
-
-# Automatically upload checkpoint and final model to HF
-# hub_model_id: username/custom_model_name
-
-plugins:
-  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
-
-datasets:
-  - path: fozziethebeat/alpaca_messages_2k_test
-    type: chat_template
-
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.1
-output_dir: ./outputs/model-out
-
-sequence_len: 2048
-sample_packing: true
-
-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
--- a/examples/hunyuan/README.md
+++ b/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==26.0 setuptools==75.8.0 wheel ninja
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation -e '.[flash-attn]'

 # Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
--- a/examples/internvl3_5/README.md
+++ b/examples/internvl3_5/README.md
@@ -1,43 +0,0 @@
-# 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)
--- a/examples/internvl3_5/internvl3_5-8b-qlora.yml
+++ b/examples/internvl3_5/internvl3_5-8b-qlora.yml
@@ -1,61 +0,0 @@
-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
--- a/examples/jamba/qlora_fsdp_large.yaml
+++ b/examples/jamba/qlora_fsdp_large.yaml
@@ -19,6 +19,7 @@ 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
--- a/examples/kimi-linear/README.md
+++ b/examples/kimi-linear/README.md
@@ -1,47 +0,0 @@
-# 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)
--- a/examples/kimi-linear/kimi-48b-lora.yaml
+++ b/examples/kimi-linear/kimi-48b-lora.yaml
@@ -1,81 +0,0 @@
-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:
--- a/examples/llama-3/3b-fp8-fsdp2.yaml
+++ b/examples/llama-3/3b-fp8-fsdp2.yaml
@@ -29,6 +29,7 @@ flex_attention: true
 flex_attn_compile_kwargs:
  dynamic: false
  mode: max-autotune-no-cudagraphs
+save_strategy: no
 torch_compile: true

 wandb_project:
--- a/examples/llama-3/qlora-1b-gdpo.yaml
+++ b/examples/llama-3/qlora-1b-gdpo.yaml
@@ -1,68 +0,0 @@
-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
--- a/examples/llama-3/qlora-fsdp-405b.yaml
+++ b/examples/llama-3/qlora-fsdp-405b.yaml
@@ -12,6 +12,7 @@ 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

--- a/examples/magistral/README.md
+++ b/examples/magistral/README.md
@@ -13,8 +13,8 @@ Thanks to the team at MistralAI for giving us early access to prepare for these
    Here is an example of how to install from pip:

 ```bash
-# Ensure you have Pytorch installed (Pytorch 2.7.0 min)
-pip3 install packaging==26.0 setuptools==75.8.0 wheel ninja
+# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```

--- a/examples/magistral/think/README.md
+++ b/examples/magistral/think/README.md
@@ -5,7 +5,6 @@ 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
--- a/examples/magistral/vision/README.md
+++ b/examples/magistral/vision/README.md
@@ -5,8 +5,7 @@ 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))
+- Installed Axolotl from source (see [main README](../README.md#getting-started))

 ## Getting started

--- a/examples/mamba/config.yml
+++ b/examples/mamba/config.yml
@@ -47,5 +47,6 @@ 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
--- a/examples/mimo/README.md
+++ b/examples/mimo/README.md
@@ -1,39 +0,0 @@
-# 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)
--- a/examples/mimo/mimo-7b-qlora.yaml
+++ b/examples/mimo/mimo-7b-qlora.yaml
@@ -1,67 +0,0 @@
-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
--- a/examples/ministral/README.md
+++ b/examples/ministral/README.md
@@ -1,50 +0,0 @@
-# Finetune Ministral with Axolotl
-
-Ministral is a family of openweight models from MistralAI found on [HuggingFace](mistralai/Ministral-8B-Instruct-2410). This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
-
-## Getting started
-
-1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
-
-2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage.
-
-3. Run the finetuning example:
-
-    ```bash
-    axolotl train examples/ministral/ministral-small-qlora.yaml
-    ```
-
-This config uses about 8.76 GiB VRAM.
-
-Let us know how it goes. Happy finetuning! 🚀
-
-### Tips
-
- We recommend adding the same/similar SystemPrompt that the model is tuned for. You can find this within the repo's files titled `SYSTEM_PROMPT.txt`.
- You can run a full finetuning by removing the `adapter: qlora` and `load_in_4bit: true` from the config.
- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
- The text dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
-
-## Optimization Guides
-
-Please check the [Optimizations doc](https://docs.axolotl.ai/docs/optimizations.html).
-
-## Limitations
-
-We only support the `mistral-common` tokenizer for Supervised Fine-tuning at the moment and for `type: chat_template` only.
-
-In addition, we do not support overriding tokens yet.
-
-## Related Resources
-
- [MistralAI Ministral Blog](https://mistral.ai/news/ministraux)
- [Axolotl Docs](https://docs.axolotl.ai)
- [Axolotl Website](https://axolotl.ai)
- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)
-
-
-## Future Work
-
- Add parity to Preference Tuning, RL, etc.
- Add parity to other tokenizer configs like overriding tokens.
--- a/examples/ministral/ministral-small-qlora.yaml
+++ b/examples/ministral/ministral-small-qlora.yaml
@@ -1,67 +0,0 @@
-base_model: mistralai/Ministral-8B-Instruct-2410
-
-# Enable to use mistral-common tokenizer
-tokenizer_use_mistral_common: true
-
-# Automatically upload checkpoint and final model to HF
-# hub_model_id: username/custom_model_name
-
-plugins:
-  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
-
-load_in_8bit: false
-load_in_4bit: true
-
-datasets:
-  - path: fozziethebeat/alpaca_messages_2k_test
-    type: chat_template
-
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.1
-output_dir: ./outputs/lora-out
-
-adapter: qlora
-lora_model_dir:
-
-sequence_len: 2048
-sample_packing: true
-
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_linear: true
-lora_target_modules:
-  - gate_proj
-  - down_proj
-  - up_proj
-  - q_proj
-  - v_proj
-  - k_proj
-  - o_proj
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 4
-micro_batch_size: 2
-num_epochs: 1
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-bf16: auto
-tf32: false
-
-gradient_checkpointing: true
-resume_from_checkpoint:
-logging_steps: 1
-flash_attention: true
-
-warmup_ratio: 0.1
-evals_per_epoch: 1
-saves_per_epoch: 1
-
-# save_first_step: true  # uncomment this to validate checkpoint saving works with your config
--- a/examples/ministral3/README.md
+++ b/examples/ministral3/README.md
@@ -1,79 +0,0 @@
-# Finetune Ministral3 with Axolotl
-
-Ministral3 is a family of open-weight models from MistralAI found on [HuggingFace](https://huggingface.co/collections/mistralai/ministral-3). This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
-
-Please see [Thinking](#thinking) and [Vision](#vision) for their respective fine-tuning.
-
-Thanks to the team at MistralAI for giving us early access to prepare for these releases.
-
-Note: This is still experimental given it is based on transformers v5 RC.
-
-## Getting started
-
-1. Install Axolotl from source following the [installation guide](https://docs.axolotl.ai/docs/installation.html#sec-edge-build).
-
-2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage.
-
-3. Swap to the Axolotl transformers v5 branch
-
-    ```bash
-    cp examples/ministral3/ministral3-3b-qlora.yaml ministral3-3b-qlora.yaml
-
-    git fetch
-    git checkout transformers-v5
-
-    # Install packages for transformers v5
-    pip install -e .
-    ```
-
-4. Run the fine-tuning:
-
-    ```bash
-    axolotl train ministral3-3b-qlora.yaml
-    ```
-
-Let us know how it goes. Happy finetuning! 🚀
-
-
-### Tips
-
- We recommend adding the same/similar SystemPrompt that the model is tuned for. You can find this within the repo's files titled `SYSTEM_PROMPT.txt`.
- You can run a full finetuning by removing the `adapter: qlora` and `load_in_4bit: true` from the config.
- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
- The text dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
-
-### Thinking
-
-Ministral3 2512 model supports thinking capabilities, enabling Chain-of-Thought reasoning with explicit thinking steps.
-
-📚 **[See the Thinking fine-tuning guide →](./think/README.md)**
-
-### Vision
-
-Ministral3 2512 model also supports vision capabilities.
-
-📚 **[See the Vision fine-tuning guide →](./vision/README.md)**
-
-## Optimization Guides
-
-Please check the [Optimizations doc](https://docs.axolotl.ai/docs/optimizations.html).
-
-## Limitations
-
-We only support the `mistral-common` tokenizer for Supervised Fine-tuning at the moment and for `type: chat_template` only.
-
-In addition, we do not support overriding tokens yet.
-
-## Related Resources
-
- [MistralAI Mistral3 Blog](https://mistral.ai/news/mistral-3)
- [Axolotl Docs](https://docs.axolotl.ai)
- [Axolotl Website](https://axolotl.ai)
- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)
-
-
-## Future Work
-
- Add parity to Preference Tuning, RL, etc.
- Add parity to other tokenizer configs like overriding tokens.
--- a/examples/ministral3/ministral3-3b-qlora.yaml
+++ b/examples/ministral3/ministral3-3b-qlora.yaml
@@ -1,68 +0,0 @@
-base_model: mistralai/Ministral-3-3B-Reasoning-2512
-
-# Enable to use mistral-common tokenizer
-tokenizer_use_mistral_common: true
-
-# Automatically upload checkpoint and final model to HF
-# hub_model_id: username/custom_model_name
-
-plugins:
-  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
-
-load_in_8bit: false
-load_in_4bit: true
-
-datasets:
-  - path: fozziethebeat/alpaca_messages_2k_test
-    type: chat_template
-
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.1
-output_dir: ./outputs/lora-out
-
-adapter: qlora
-lora_model_dir:
-
-sequence_len: 2048
-sample_packing: true
-
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_linear: true
-lora_target_modules:
-  - gate_proj
-  - down_proj
-  - up_proj
-  - q_proj
-  - v_proj
-  - k_proj
-  - o_proj
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 4
-micro_batch_size: 2
-num_epochs: 1
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-bf16: auto
-tf32: false
-
-gradient_checkpointing: true
-resume_from_checkpoint:
-logging_steps: 1
-flash_attention: true
-scaling_softmax: true
-
-warmup_ratio: 0.1
-evals_per_epoch: 1
-saves_per_epoch: 1
-
-# save_first_step: true  # uncomment this to validate checkpoint saving works with your config
--- a/examples/ministral3/think/README.md
+++ b/examples/ministral3/think/README.md
@@ -1,74 +0,0 @@
-# Ministral3 2512 Thinking Fine-tuning
-
-This guide covers fine-tuning [Ministral3 2512](https://huggingface.co/collections/mistralai/ministral-3) with thinking capabilities using Axolotl. The thinking model enables explicit Chain-of-Thought reasoning with separate thinking and response sections.
-
-## Prerequisites
-
-Before starting, ensure you have:
-
- Installed Axolotl (see [main README](../README.md))
-
-## Getting Started
-
-Run the thinking model fine-tuning:
-
-```bash
-axolotl train examples/ministral3/think/ministral3-3b-think-qlora.yaml
-```
-
-This config uses about 4.76 GiB VRAM.
-
-### Tips
-
- Dataset uses multi-content format with `type: thinking` support. See [Dataset Format](#dataset-format) below.
- You cannot mix `content: str` and `content: list[dict]`, otherwise, dataset loading will fail. Keep it consistent.
-
-## Dataset Format
-
-The thinking model requires the multi-content dataset format with support for an extra `role: thinking` within system and assistant messages.
-
-Example format:
-
-```json
-{
-    "messages": [
-        {
-            "role": "system",
-            "content": [
-                { "type": "text", "text": "{SYSTEM_PROMPT}"}
-            ]
-        },
-        {
-            "role": "user",
-            "content": [
-                { "type": "text", "text": "Solve this step by step: What is 15% of 240?"}
-            ]
-        },
-        {
-            "role": "assistant",
-            "content": [
-                {
-                    "type": "thinking",
-                    "thinking": "I need to calculate 15% of 240. First, I'll convert 15% to decimal: 0.15. Then multiply: 0.15 × 240 = 36."
-                },
-                {
-                    "type": "text",
-                    "text": "To find 15% of 240, I'll multiply 240 by 0.15:\n\n240 × 0.15 = 36\n\nTherefore, 15% of 240 is 36."
-                }
-            ]
-        }
-    ]
-}
-```
-
-### Advanced Options
-
-The `thinking` section supports an optional `closed` parameter:
-
-```json
-{
-    "type": "thinking",
-    "thinking": "Internal reasoning here...",
-    "closed": true  // Default: true, controls adding the closing [/THINK] tag
-}
-```
--- a/examples/ministral3/think/ministral3-3b-think-qlora.yaml
+++ b/examples/ministral3/think/ministral3-3b-think-qlora.yaml
@@ -1,67 +0,0 @@
-base_model: mistralai/Ministral-3-3B-Reasoning-2512
-
-# Enable to use mistral-common tokenizer
-tokenizer_use_mistral_common: true
-
-# Automatically upload checkpoint and final model to HF
-# hub_model_id: username/custom_model_name
-
-plugins:
-  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
-
-load_in_8bit: false
-load_in_4bit: true
-
-datasets:
-  - path: Nanobit/text-think-2k-test
-    type: chat_template
-
-dataset_prepared_path: last_run_prepared
-val_set_size: 0
-output_dir: ./outputs/lora-out
-
-adapter: qlora
-lora_model_dir:
-
-sequence_len: 2048
-sample_packing: true
-
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_linear: true
-lora_target_modules:
-  - gate_proj
-  - down_proj
-  - up_proj
-  - q_proj
-  - v_proj
-  - k_proj
-  - o_proj
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 4
-micro_batch_size: 2
-num_epochs: 1
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-bf16: auto
-tf32: false
-
-gradient_checkpointing: true
-resume_from_checkpoint:
-logging_steps: 1
-flash_attention: true
-
-warmup_ratio: 0.1
-evals_per_epoch: 1
-saves_per_epoch: 1
-
-# save_first_step: true  # uncomment this to validate checkpoint saving works with your config
--- a/examples/ministral3/vision/README.md
+++ b/examples/ministral3/vision/README.md
@@ -1,58 +0,0 @@
-# Ministral3 2512 Vision Fine-tuning
-
-This guide covers fine-tuning [Ministral3 2512](https://huggingface.co/collections/mistralai/ministral-3) with vision capabilities using Axolotl.
-
-## Prerequisites
-
-Before starting, ensure you have:
-
- Installed Axolotl from source (see [main README](../README.md))
-
-## Getting started
-
-1. Install the required vision lib:
-    ```bash
-    pip install 'mistral-common[opencv]==1.8.6'
-    ```
-
-2. Download the example dataset image:
-   ```bash
-   wget https://huggingface.co/datasets/Nanobit/text-vision-2k-test/resolve/main/African_elephant.jpg
-   ```
-
-3. Run the fine-tuning:
-   ```bash
-   axolotl train examples/ministral3/vision/ministral3-3b-vision-qlora.yml
-   ```
-
-WARNING: The loss and grad norm will be much higher than normal at first. We suspect this to be inherent to the model as of the moment. If anyone would like to submit a fix for this, we are happy to take a look.
-
-### Tips
-
-Key differences from text-only model:
- Multi-modal dataset format required
- Sample packing not supported
-
-## Dataset Format
-
-The vision model requires multi-modal dataset format as documented [here](https://docs.axolotl.ai/docs/multimodal.html#dataset-format).
-
-One exception is that, passing `"image": PIL.Image` is not supported. MistralTokenizer only supports `path`, `url`, and `base64` for now.
-
-Example:
-```json
-{
-    "messages": [
-        {"role": "system", "content": [{ "type": "text", "text": "{SYSTEM_PROMPT}"}]},
-        {"role": "user", "content": [
-            { "type": "text", "text": "What's in this image?"},
-            {"type": "image", "path": "path/to/image.jpg" }
-        ]},
-        {"role": "assistant", "content": [{ "type": "text", "text": "..." }]},
-    ],
-}
-```
-
-## Limitations
-
- Sample Packing is not supported for multi-modality training currently.
--- a/examples/ministral3/vision/ministral3-3b-vision-qlora.yml
+++ b/examples/ministral3/vision/ministral3-3b-vision-qlora.yml
@@ -1,64 +0,0 @@
-base_model: mistralai/Ministral-3-3B-Reasoning-2512
-processor_type: AutoProcessor
-
-# Enable to use mistral-common tokenizer
-tokenizer_use_mistral_common: true
-
-plugins:
-  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
-
-load_in_4bit: true
-
-# these 3 lines are needed for now to handle vision chat templates w images
-skip_prepare_dataset: true
-remove_unused_columns: false
-sample_packing: false
-
-# sample dataset below requires downloading image in advance
-# wget https://huggingface.co/datasets/Nanobit/text-vision-2k-test/resolve/main/African_elephant.jpg
-datasets:
-  - path: Nanobit/text-vision-2k-test
-    type: chat_template
-
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
-output_dir: ./outputs/out
-
-adapter: qlora
-lora_model_dir:
-
-sequence_len: 2048
-
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 1
-micro_batch_size: 1
-num_epochs: 1
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-bf16: true
-fp16:
-tf32: true
-
-gradient_checkpointing: true
-logging_steps: 1
-flash_attention: true
-
-warmup_ratio: 0.1
-evals_per_epoch: 1
-saves_per_epoch: 1
-weight_decay: 0.0
-special_tokens:
-
-# save_first_step: true  # uncomment this to validate checkpoint saving works with your config
--- a/examples/mistral/mistral-small/README.md
+++ b/examples/mistral/mistral-small/README.md
@@ -5,7 +5,6 @@ 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
--- a/examples/mistral/mistral-small/mistral-small-3.1-24B-lora.yml
+++ b/examples/mistral/mistral-small/mistral-small-3.1-24B-lora.yml
--- a/examples/olmo3/README.md
+++ b/examples/olmo3/README.md
@@ -1,38 +0,0 @@
-# Finetune Allenai's Olmo 3 with Axolotl
-
-[Olmo 3](https://huggingface.co/collections/allenai/olmo-3) are a family of 7B and 32B models open source models trained by The Allen Institute for Artificial Intelligence.
-
-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/olmo3/olmo3-7b-qlora.yaml
-    ```
-
-This uses about 11.3 GiB VRAM. Let us know how it goes. Happy finetuning! 🚀
-
-### TIPS
-
- The example config can be re-used for Olmo and Olmo 2.
- 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
-
- [Olmo 3 Blog](https://allenai.org/blog/olmo3)
- [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)
--- a/examples/olmo3/olmo3-7b-qlora.yaml
+++ b/examples/olmo3/olmo3-7b-qlora.yaml
@@ -1,64 +0,0 @@
-base_model: allenai/Olmo-3-7B-Instruct-SFT
-
-# Automatically upload checkpoint and final model to HF
-# hub_model_id: username/custom_model_name
-
-plugins:
-  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
-
-load_in_8bit: false
-load_in_4bit: true
-
-datasets:
-  - path: fozziethebeat/alpaca_messages_2k_test
-    type: chat_template
-
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.1
-output_dir: ./outputs/lora-out
-
-adapter: qlora
-lora_model_dir:
-
-sequence_len: 2048
-sample_packing: true
-
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_linear: true
-lora_target_modules:
-  - gate_proj
-  - down_proj
-  - up_proj
-  - q_proj
-  - v_proj
-  - k_proj
-  - o_proj
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 2
-micro_batch_size: 2
-num_epochs: 1
-optimizer: adamw_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
--- a/examples/plano/README.md
+++ b/examples/plano/README.md
@@ -1,42 +0,0 @@
-# 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)
--- a/examples/plano/plano-4b-qlora.yaml
+++ b/examples/plano/plano-4b-qlora.yaml
@@ -1,65 +0,0 @@
-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
--- a/examples/qat_nvfp4/Gemma3-12B_baseline.yml
+++ b/examples/qat_nvfp4/Gemma3-12B_baseline.yml
@@ -1,67 +0,0 @@
-base_model: google/gemma-3-12b-it
-# Automatically upload checkpoint and final model to HF
-# hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-plugins:
-  - axolotl.integrations.liger.LigerPlugin
-
-liger_rope: true
-liger_rms_norm: true
-liger_glu_activation: true
-liger_layer_norm: true
-liger_fused_linear_cross_entropy: true
-seed: 42
-chat_template: gemma3
-datasets:
-  - path: tatsu-lab/alpaca
-    type: alpaca
-
-output_dir: ./outputs/out_gemma/
-
-sequence_len: 8096
-sample_packing: true
-flash_attention: true
-
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 1
-micro_batch_size: 16
-
-num_epochs: 1
-optimizer: adamw_torch_fused
-lr_scheduler: cosine
-learning_rate: 4e-5
-
-bf16: true
-tf32: true
-
-resume_from_checkpoint:
-logging_steps: 1
-
-# evals_per_epoch: 1
-saves_per_epoch: 1
-
-warmup_ratio: 0.1
-weight_decay: 0.0
-fsdp_version: 2
-
-fsdp_config:
-  offload_params: false
-  cpu_ram_efficient_loading: true
-  auto_wrap_policy: TRANSFORMER_BASED_WRAP
-  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
-  state_dict_type: FULL_STATE_DICT
-  sharding_strategy: FULL_SHARD
-  reshard_after_forward: true
-  activation_checkpointing: true
-
-special_tokens:
-
-# save_first_step: true  # uncomment this to validate checkpoint saving works with your config
--- a/examples/qat_nvfp4/Gemma3-12B_qat.yml
+++ b/examples/qat_nvfp4/Gemma3-12B_qat.yml
@@ -1,72 +0,0 @@
-base_model: google/gemma-3-12b-it
-# Automatically upload checkpoint and final model to HF
-# hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-plugins:
-  - axolotl.integrations.liger.LigerPlugin
-
-liger_rope: true
-liger_rms_norm: true
-liger_glu_activation: true
-liger_layer_norm: true
-liger_fused_linear_cross_entropy: true
-seed: 42
-chat_template: gemma3
-datasets:
-  - path: tatsu-lab/alpaca
-    type: alpaca
-
-output_dir: ./outputs/qat_out_gemma/
-
-sequence_len: 8096
-sample_packing: true
-flash_attention: true
-
-qat:
-  activation_dtype: nvfp4
-  weight_dtype: nvfp4
-  group_size: 16 # only group_size of 16 is supported with nvfp4
-
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 1
-micro_batch_size: 16
-
-num_epochs: 1
-optimizer: adamw_torch_fused
-lr_scheduler: cosine
-learning_rate: 4e-5
-
-bf16: true
-tf32: true
-
-resume_from_checkpoint:
-logging_steps: 1
-
-evals_per_epoch: 1
-saves_per_epoch: 1
-
-warmup_ratio: 0.1
-weight_decay: 0.0
-fsdp_version: 2
-
-fsdp_config:
-  offload_params: false
-  cpu_ram_efficient_loading: true
-  auto_wrap_policy: TRANSFORMER_BASED_WRAP
-  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
-  state_dict_type: FULL_STATE_DICT
-  sharding_strategy: FULL_SHARD
-  reshard_after_forward: true
-  activation_checkpointing: true
-
-special_tokens:
-
-# save_first_step: true  # uncomment this to validate checkpoint saving works with your config
--- a/examples/qat_nvfp4/Math-Gemma3-12B_baseline.yml
+++ b/examples/qat_nvfp4/Math-Gemma3-12B_baseline.yml
@@ -1,67 +0,0 @@
-base_model: google/gemma-3-12b-it
-# Math finetuning configuration for Gemma3-12B
-# hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-plugins:
-  - axolotl.integrations.liger.LigerPlugin
-
-liger_rope: true
-liger_rms_norm: true
-liger_glu_activation: true
-liger_layer_norm: true
-liger_fused_linear_cross_entropy: true
-seed: 42
-chat_template: gemma3
-datasets:
-  - path: AI-MO/NuminaMath-CoT
-    type: chat_template
-
-output_dir: ./outputs/out_math_gemma/
-
-sequence_len: 4096
-sample_packing: true
-flash_attention: true
-
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 1
-micro_batch_size: 8
-
-num_epochs: 1
-optimizer: adamw_torch_fused
-lr_scheduler: cosine
-learning_rate: 3e-5
-
-bf16: true
-tf32: true
-
-resume_from_checkpoint:
-logging_steps: 1
-
-# evals_per_epoch: 1
-saves_per_epoch: 1
-
-warmup_ratio: 0.1
-weight_decay: 0.0
-fsdp_version: 2
-
-fsdp_config:
-  offload_params: false
-  cpu_ram_efficient_loading: true
-  auto_wrap_policy: TRANSFORMER_BASED_WRAP
-  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
-  state_dict_type: FULL_STATE_DICT
-  sharding_strategy: FULL_SHARD
-  reshard_after_forward: true
-  activation_checkpointing: true
-
-special_tokens:
-
-# save_first_step: true  # uncomment this to validate checkpoint saving works with your config
--- a/examples/qat_nvfp4/Math-Gemma3-12B_qat.yml
+++ b/examples/qat_nvfp4/Math-Gemma3-12B_qat.yml
@@ -1,72 +0,0 @@
-base_model: google/gemma-3-12b-it
-# Math finetuning configuration for Gemma3-12B
-# hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-plugins:
-  - axolotl.integrations.liger.LigerPlugin
-
-liger_rope: true
-liger_rms_norm: true
-liger_glu_activation: true
-liger_layer_norm: true
-liger_fused_linear_cross_entropy: true
-seed: 42
-chat_template: gemma3
-datasets:
-  - path: AI-MO/NuminaMath-CoT
-    type: chat_template
-
-output_dir: ./outputs/qat_out_math_gemma/
-
-sequence_len: 4096
-sample_packing: true
-flash_attention: true
-
-qat:
-  activation_dtype: nvfp4
-  weight_dtype: nvfp4
-  group_size: 16 # only group_size of 16 is supported with nvfp4
-
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 1
-micro_batch_size: 8
-
-num_epochs: 1
-optimizer: adamw_torch_fused
-lr_scheduler: cosine
-learning_rate: 3e-5
-
-bf16: true
-tf32: true
-
-resume_from_checkpoint:
-logging_steps: 1
-
-# evals_per_epoch: 1
-saves_per_epoch: 1
-
-warmup_ratio: 0.1
-weight_decay: 0.0
-fsdp_version: 2
-
-fsdp_config:
-  offload_params: false
-  cpu_ram_efficient_loading: true
-  auto_wrap_policy: TRANSFORMER_BASED_WRAP
-  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
-  state_dict_type: FULL_STATE_DICT
-  sharding_strategy: FULL_SHARD
-  reshard_after_forward: true
-  activation_checkpointing: true
-
-special_tokens:
-
-# save_first_step: true  # uncomment this to validate checkpoint saving works with your config
--- a/examples/qat_nvfp4/Math-Gemma3-27B_baseline.yml
+++ b/examples/qat_nvfp4/Math-Gemma3-27B_baseline.yml
@@ -1,68 +0,0 @@
-base_model: google/gemma-3-27b-it
-# Math finetuning configuration for Gemma3-27B
-# hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-plugins:
-  - axolotl.integrations.liger.LigerPlugin
-
-liger_rope: true
-liger_rms_norm: true
-liger_glu_activation: true
-liger_layer_norm: true
-liger_fused_linear_cross_entropy: true
-seed: 42
-chat_template: gemma3
-datasets:
-  - path: AI-MO/NuminaMath-CoT
-    type: chat_template
-
-output_dir: ./outputs/out_math_gemma27/
-
-sequence_len: 4096
-sample_packing: true
-flash_attention: true
-
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 1
-micro_batch_size: 16
-
-num_epochs: 1
-optimizer: adamw_torch_fused
-lr_scheduler: cosine
-learning_rate: 5e-6
-eta_min: 7e-7
-
-bf16: true
-tf32: true
-
-resume_from_checkpoint:
-logging_steps: 1
-
-# evals_per_epoch: 1
-saves_per_epoch: 1
-
-warmup_ratio: 0.1
-weight_decay: 0.0
-fsdp_version: 2
-
-fsdp_config:
-  offload_params: false
-  cpu_ram_efficient_loading: true
-  auto_wrap_policy: TRANSFORMER_BASED_WRAP
-  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
-  state_dict_type: FULL_STATE_DICT
-  sharding_strategy: FULL_SHARD
-  reshard_after_forward: true
-  activation_checkpointing: true
-
-special_tokens:
-
-# save_first_step: true  # uncomment this to validate checkpoint saving works with your config
--- a/examples/qat_nvfp4/Math-Gemma3-27B_qat.yml
+++ b/examples/qat_nvfp4/Math-Gemma3-27B_qat.yml
@@ -1,73 +0,0 @@
-base_model: google/gemma-3-27b-it
-# Math finetuning configuration for Gemma3-27B
-# hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-plugins:
-  - axolotl.integrations.liger.LigerPlugin
-
-liger_rope: true
-liger_rms_norm: true
-liger_glu_activation: true
-liger_layer_norm: true
-liger_fused_linear_cross_entropy: true
-seed: 42
-chat_template: gemma3
-datasets:
-  - path: AI-MO/NuminaMath-CoT
-    type: chat_template
-
-output_dir: ./outputs/qat_out_math_gemma27/
-
-sequence_len: 4096
-sample_packing: true
-flash_attention: true
-
-qat:
-  activation_dtype: nvfp4
-  weight_dtype: nvfp4
-  group_size: 16 # only group_size of 16 is supported with nvfp4
-
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 1
-micro_batch_size: 16
-
-num_epochs: 1
-optimizer: adamw_torch_fused
-lr_scheduler: cosine
-learning_rate: 5e-6
-eta_min: 7e-7
-
-bf16: true
-tf32: true
-
-resume_from_checkpoint:
-logging_steps: 1
-
-# evals_per_epoch: 1
-saves_per_epoch: 1
-
-warmup_ratio: 0.1
-weight_decay: 0.0
-fsdp_version: 2
-
-fsdp_config:
-  offload_params: false
-  cpu_ram_efficient_loading: true
-  auto_wrap_policy: TRANSFORMER_BASED_WRAP
-  transformer_layer_cls_to_wrap: Gemma3DecoderLayer
-  state_dict_type: FULL_STATE_DICT
-  sharding_strategy: FULL_SHARD
-  reshard_after_forward: true
-  activation_checkpointing: true
-
-special_tokens:
-
-# save_first_step: true  # uncomment this to validate checkpoint saving works with your config
--- a/examples/qat_nvfp4/Math-Qwen2.5-72B_baseline.yml
+++ b/examples/qat_nvfp4/Math-Qwen2.5-72B_baseline.yml
@@ -1,67 +0,0 @@
-base_model: Qwen/Qwen2.5-72B
-# Math finetuning configuration for Qwen2.5-72B (non-instruct)
-# hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-plugins:
-  - axolotl.integrations.liger.LigerPlugin
-
-liger_rope: true
-liger_rms_norm: true
-liger_glu_activation: true
-liger_layer_norm: true
-liger_fused_linear_cross_entropy: true
-seed: 42
-chat_template: qwen_25
-datasets:
-  - path: AI-MO/NuminaMath-CoT
-    type: chat_template
-
-output_dir: ./outputs/out_math_72b/
-
-sequence_len: 4096
-sample_packing: true
-flash_attention: true
-
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 1
-micro_batch_size: 8
-num_epochs: 1
-optimizer: adamw_torch_fused
-lr_scheduler: cosine
-learning_rate: 5e-6
-eta_min: 7e-7
-
-bf16: true
-tf32: true
-
-resume_from_checkpoint:
-logging_steps: 1
-
-# evals_per_epoch: 1
-saves_per_epoch: 1
-
-warmup_ratio: 0.1
-weight_decay: 0.0
-fsdp_version: 2
-
-fsdp_config:
-  offload_params: false
-  cpu_ram_efficient_loading: true
-  auto_wrap_policy: TRANSFORMER_BASED_WRAP
-  transformer_layer_cls_to_wrap: Qwen2DecoderLayer
-  state_dict_type: FULL_STATE_DICT
-  sharding_strategy: FULL_SHARD
-  reshard_after_forward: true
-  activation_checkpointing: true
-
-special_tokens:
-
-# save_first_step: true  # uncomment this to validate checkpoint saving works with your config
--- a/examples/qat_nvfp4/Math-Qwen2.5-72B_qat.yml
+++ b/examples/qat_nvfp4/Math-Qwen2.5-72B_qat.yml
@@ -1,72 +0,0 @@
-base_model: Qwen/Qwen2.5-72B
-# Math finetuning configuration for Qwen2.5-72B (non-instruct)
-# hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-plugins:
-  - axolotl.integrations.liger.LigerPlugin
-
-liger_rope: true
-liger_rms_norm: true
-liger_glu_activation: true
-liger_layer_norm: true
-liger_fused_linear_cross_entropy: true
-seed: 42
-chat_template: qwen_25
-datasets:
-  - path: AI-MO/NuminaMath-CoT
-    type: chat_template
-
-output_dir: ./outputs/qat_out_math_72b/
-
-sequence_len: 4096
-sample_packing: true
-flash_attention: true
-
-qat:
-  activation_dtype: nvfp4
-  weight_dtype: nvfp4
-  group_size: 16 # only group_size of 16 is supported with nvfp4
-
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 1
-micro_batch_size: 8
-num_epochs: 1
-optimizer: adamw_torch_fused
-lr_scheduler: cosine
-learning_rate: 5e-6
-eta_min: 7e-7
-
-bf16: true
-tf32: true
-
-resume_from_checkpoint:
-logging_steps: 1
-
-# evals_per_epoch: 1
-saves_per_epoch: 1
-
-warmup_ratio: 0.1
-weight_decay: 0.0
-fsdp_version: 2
-
-fsdp_config:
-  offload_params: false
-  cpu_ram_efficient_loading: true
-  auto_wrap_policy: TRANSFORMER_BASED_WRAP
-  transformer_layer_cls_to_wrap: Qwen2DecoderLayer
-  state_dict_type: FULL_STATE_DICT
-  sharding_strategy: FULL_SHARD
-  reshard_after_forward: true
-  activation_checkpointing: true
-
-special_tokens:
-
-# save_first_step: true  # uncomment this to validate checkpoint saving works with your config
--- a/examples/qat_nvfp4/Qwen2.5-72B_baseline.yml
+++ b/examples/qat_nvfp4/Qwen2.5-72B_baseline.yml
@@ -1,67 +0,0 @@
-base_model: Qwen/Qwen2.5-72B
-# Alpaca finetuning configuration for Qwen2.5-72B
-# hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-plugins:
-  - axolotl.integrations.liger.LigerPlugin
-
-liger_rope: true
-liger_rms_norm: true
-liger_glu_activation: true
-liger_layer_norm: true
-liger_fused_linear_cross_entropy: true
-seed: 42
-chat_template: qwen_25
-datasets:
-  - path: tatsu-lab/alpaca
-    type: alpaca
-
-output_dir: ./outputs/out_qwen72b/
-
-sequence_len: 8096
-sample_packing: true
-flash_attention: true
-
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 1
-micro_batch_size: 16
-
-num_epochs: 1
-optimizer: adamw_torch_fused
-lr_scheduler: cosine
-learning_rate: 2e-5
-
-bf16: true
-tf32: true
-
-resume_from_checkpoint:
-logging_steps: 1
-
-# evals_per_epoch: 1
-saves_per_epoch: 1
-
-warmup_ratio: 0.1
-weight_decay: 0.0
-fsdp_version: 2
-
-fsdp_config:
-  offload_params: false
-  cpu_ram_efficient_loading: true
-  auto_wrap_policy: TRANSFORMER_BASED_WRAP
-  transformer_layer_cls_to_wrap: Qwen2DecoderLayer
-  state_dict_type: FULL_STATE_DICT
-  sharding_strategy: FULL_SHARD
-  reshard_after_forward: true
-  activation_checkpointing: true
-
-special_tokens:
-
-# save_first_step: true  # uncomment this to validate checkpoint saving works with your config
--- a/examples/qat_nvfp4/Qwen2.5-72B_qat.yml
+++ b/examples/qat_nvfp4/Qwen2.5-72B_qat.yml
@@ -1,72 +0,0 @@
-base_model: Qwen/Qwen2.5-72B
-# Alpaca finetuning configuration for Qwen2.5-72B
-# hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-plugins:
-  - axolotl.integrations.liger.LigerPlugin
-
-liger_rope: true
-liger_rms_norm: true
-liger_glu_activation: true
-liger_layer_norm: true
-liger_fused_linear_cross_entropy: true
-seed: 42
-chat_template: qwen_25
-datasets:
-  - path: tatsu-lab/alpaca
-    type: alpaca
-
-output_dir: ./outputs/qat_out_qwen72b/
-
-sequence_len: 8096
-sample_packing: true
-flash_attention: true
-
-qat:
-  activation_dtype: nvfp4
-  weight_dtype: nvfp4
-  group_size: 16 # only group_size of 16 is supported with nvfp4
-
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 1
-micro_batch_size: 16
-
-num_epochs: 1
-optimizer: adamw_torch_fused
-lr_scheduler: cosine
-learning_rate: 2e-5
-
-bf16: true
-tf32: true
-
-resume_from_checkpoint:
-logging_steps: 1
-
-# evals_per_epoch: 1
-saves_per_epoch: 1
-
-warmup_ratio: 0.1
-weight_decay: 0.0
-fsdp_version: 2
-
-fsdp_config:
-  offload_params: false
-  cpu_ram_efficient_loading: true
-  auto_wrap_policy: TRANSFORMER_BASED_WRAP
-  transformer_layer_cls_to_wrap: Qwen2DecoderLayer
-  state_dict_type: FULL_STATE_DICT
-  sharding_strategy: FULL_SHARD
-  reshard_after_forward: true
-  activation_checkpointing: true
-
-special_tokens:
-
-# save_first_step: true  # uncomment this to validate checkpoint saving works with your config
--- a/examples/qwen2/adamw-pretrain-fsdp2.yaml
+++ b/examples/qwen2/adamw-pretrain-fsdp2.yaml
@@ -1,70 +0,0 @@
-base_model: Qwen/Qwen2.5-0.5B
-model_type: AutoModelForCausalLM
-tokenizer_type: AutoTokenizer
-
-# Use random initialization for fair comparison
-reinit_weights: true
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-# Pretraining dataset
-pretraining_dataset:
-  - path: allenai/c4
-    name: en
-    type: pretrain
-    split: train
-
-dataset_prepared_path:
-val_set_size: 0.0
-output_dir: ./outputs/compare-adamw-pretrain
-
-sequence_len: 2048
-sample_packing: true
-pad_to_sequence_len: true
-
-wandb_project: dist_muon
-wandb_entity:
-wandb_watch:
-wandb_name: adamw
-wandb_log_model:
-
-gradient_accumulation_steps: 1
-micro_batch_size: 4
-num_epochs: 1
-max_steps: 305
-
-# AdamW optimizer settings (standard LR for AdamW)
-optimizer: adamw_torch_fused
-learning_rate: 0.0002
-weight_decay: 0.01
-lr_scheduler: cosine
-
-train_on_inputs: true
-group_by_length: false
-bf16: auto
-fp16: false
-tf32: false
-
-gradient_checkpointing: false
-logging_steps: 1
-flash_attention: true
-
-warmup_steps: 10
-evals_per_epoch: 0
-saves_per_epoch: 1
-
-# Reproducibility
-seed: 42
-
-fsdp_config:
-  fsdp_version: 2
-  fsdp_offload_params: false
-  fsdp_state_dict_type: FULL_STATE_DICT
-  fsdp_transformer_layer_cls_to_wrap: Qwen2DecoderLayer
-  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
-  fsdp_cpu_ram_efficient_loading: false
-  fsdp_reshard_after_forward: true
-
-special_tokens:
--- a/examples/qwen2/muon-pretrain-fsdp2.yaml
+++ b/examples/qwen2/muon-pretrain-fsdp2.yaml
@@ -1,70 +0,0 @@
-base_model: Qwen/Qwen2.5-0.5B
-model_type: AutoModelForCausalLM
-tokenizer_type: AutoTokenizer
-
-# Use random initialization for fair comparison
-reinit_weights: true
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-# Pretraining dataset
-pretraining_dataset:
-  - path: allenai/c4
-    name: en
-    type: pretrain
-    split: train
-
-dataset_prepared_path:
-val_set_size: 0.0
-output_dir: ./outputs/compare-muon-pretrain
-
-sequence_len: 2048
-sample_packing: true
-pad_to_sequence_len: true
-
-wandb_project: dist_muon
-wandb_entity:
-wandb_watch:
-wandb_name: muon
-wandb_log_model:
-
-gradient_accumulation_steps: 1
-micro_batch_size: 4
-num_epochs: 1
-max_steps: 305
-
-# Muon optimizer settings
-optimizer: muon
-learning_rate: 0.02
-weight_decay: 0.01
-lr_scheduler: cosine
-
-train_on_inputs: true
-group_by_length: false
-bf16: auto
-fp16: false
-tf32: false
-
-gradient_checkpointing: false
-logging_steps: 1
-flash_attention: true
-
-warmup_steps: 10
-evals_per_epoch: 0
-saves_per_epoch: 1
-
-# Reproducibility
-seed: 42
-
-fsdp_config:
-  fsdp_version: 2
-  fsdp_offload_params: false
-  fsdp_state_dict_type: FULL_STATE_DICT
-  fsdp_transformer_layer_cls_to_wrap: Qwen2DecoderLayer
-  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
-  fsdp_cpu_ram_efficient_loading: false
-  fsdp_reshard_after_forward: true
-
-special_tokens:
--- a/examples/qwen3-next/README.md
+++ b/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==26.0 setuptools==75.8.0 wheel ninja
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation -e '.[flash-attn]'

 # Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Wing Lian	9ee7ce5c85	set TORCH_CUDA_ARCH_LIST correctly	2025-10-29 12:59:26 -04:00
Wing Lian	a41ca4d06f	upgrade liger dep to 0.6.3	2025-10-27 14:49:09 -04:00