support nemotron for scattermoe-lora

* post merge lora fixes for CI

* handle lora kernel auto-enable for moe without grouped_mm

* prefer not to import torch in schema validation

.bandit

@@ -1,3 +1,3 @@
 [bandit]
 exclude = tests
-skips = B101,B615
+skips = B101,B615,B102,B110

.coderabbit.yaml

@@ -12,5 +12,6 @@ reviews:
   auto_review:
     enabled: true
     drafts: false
+    auto_incremental_review: false
 chat:
   auto_reply: true

.flake8

@@ -1,5 +0,0 @@
-[flake8]
-max-line-length = 88
-
-select = C,E,F,W,B,B950
-extend-ignore = E203, E501, W503

.github/CONTRIBUTING.md

@@ -57,11 +57,23 @@ We welcome ideas for improvements and new features. To suggest an enhancement, o
 5. Push your branch to your fork on GitHub.
 6. Open a new pull request against the `main` branch of the axolotl repository. Include a clear and concise description of your changes, referencing any related issues.
 
+#### Skipping CI Checks
+
+You can skip certain CI checks by including specific keywords in your commit messages:
+
+- `[skip ci]` or `skip ci` - Skips all CI checks for that commit
+- `[skip-e2e]` or `skip-e2e` - Skips only end-to-end tests while running other CI checks. You may also include this in the title of your PR to disable end-to-end tests for the entire PR.
+
 ## Style Guidelines
 
 ### Code Style
 
-axolotl uses [{codestyle}]({URLofCodestyle}) as its code style guide. Please ensure that your code follows these guidelines.
+axolotl uses [Ruff](https://docs.astral.sh/ruff/) as its code style guide. Please ensure that your code follows these guidelines.
+
+Use the pre-commit linter to ensure that your code is formatted consistently.
+```bash
+pre-commit run --all-files
+```
 
 ### Commit Messages
 
@@ -71,6 +83,6 @@ Write clear and concise commit messages that briefly describe the changes made i
 
 - [GitHub Help](https://help.github.com/)
 - [GitHub Pull Request Documentation](https://docs.github.com/en/github/collaborating-with-issues-and-pull-requests)
-- [{codestyle}]({URLofCodestyle})
+- [Ruff](https://docs.astral.sh/ruff/)
 
 Thank you once again for your interest in contributing to axolotl. We look forward to collaborating with you and creating an even better project together!

.github/FUNDING.yml

@@ -1,13 +1,13 @@
 # These are supported funding model platforms
 
-github: [winglian, OpenAccess-AI-Collective] # Replace with up to 4 GitHub Sponsors-enabled usernames e.g., [user1, user2]
+github: # 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: axolotl_ai # Replace with a single Ko-fi username
+ko_fi: # 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: ['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']
+custom: # Replace with up to 4 custom sponsorship URLs e.g., ['link1', 'link2']

.github/PULL_REQUEST_TEMPLATE.md

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

.github/workflows/base.yml

@@ -15,51 +15,21 @@ on:
       - '.github/workflows/base.yml'
   workflow_dispatch:
 
+permissions:
+  contents: read
+
 jobs:
   build-base:
     if: ${{ github.repository_owner == 'axolotl-ai-cloud' && (github.event_name != 'pull_request' || !github.event.pull_request.draft) }}
     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: "124"
-            cuda_version: 12.4.1
-            cudnn_version: ""
-            python_version: "3.11"
-            pytorch: 2.6.0
-            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: 2.6.0
-            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: 2.7.0
-            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: ""
@@ -67,6 +37,71 @@ jobs:
             pytorch: 2.8.0
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
             dockerfile: "Dockerfile-base"
+            platforms: "linux/amd64"
+          - cuda: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.9.0
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-base"
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            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: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.10.0
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-base"
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            python_version: "3.12"
+            pytorch: 2.10.0
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-base"
+            platforms: "linux/amd64,linux/arm64"
+#          - cuda: "129"
+#            cuda_version: 12.9.1
+#            cudnn_version: ""
+#            python_version: "3.12"
+#            pytorch: 2.9.1
+#            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+#            dockerfile: "Dockerfile-base"
+#            platforms: "linux/amd64,linux/arm64"
+          - cuda: "130"
+            cuda_version: 13.0.0
+            cudnn_version: ""
+            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: "130"
+            cuda_version: 13.0.0
+            cudnn_version: ""
+            python_version: "3.12"
+            pytorch: 2.10.0
+            torch_cuda_arch_list: "9.0+PTX"
+            dockerfile: "Dockerfile-base"
+            platforms: "linux/amd64,linux/arm64"
 #          - cuda: "128"
 #            cuda_version: 12.8.1
 #            cudnn_version: ""
@@ -90,20 +125,21 @@ jobs:
         uses: docker/metadata-action@v5
         with:
           images: |
-            winglian/axolotl-base
             axolotlai/axolotl-base
       - name: Login to Docker Hub
-        uses: docker/login-action@v2
+        uses: docker/login-action@v3
+        if: ${{ github.event_name != 'pull_request' && env.HAS_DOCKERHUB_CREDS == 'true' }}
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
       - name: Build
-        uses: docker/build-push-action@v4
+        uses: docker/build-push-action@v5
         with:
           context: .
           file: ./docker/${{ 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 }}
@@ -118,31 +154,12 @@ jobs:
     if: ${{ github.repository_owner == 'axolotl-ai-cloud' && (github.event_name != 'pull_request' || !github.event.pull_request.draft) }}
     timeout-minutes: 480
     runs-on: ubuntu-latest-m
+    env:
+      HAS_DOCKERHUB_CREDS: ${{ secrets.DOCKERHUB_USERNAME != '' && secrets.DOCKERHUB_TOKEN != '' }}
     strategy:
       fail-fast: false
       matrix:
         include:
-          - cuda: "126"
-            cuda_version: 12.6.3
-            cudnn_version: ""
-            python_version: "3.11"
-            pytorch: 2.6.0
-            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
-            dockerfile: "Dockerfile-uv-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-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: ""
@@ -150,6 +167,79 @@ 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: ""
+            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: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.9.0
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-uv-base"
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.10.0
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-uv-base"
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            python_version: "3.12"
+            pytorch: 2.10.0
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-uv-base"
+            platforms: "linux/amd64,linux/arm64"
+#          - cuda: "129"
+#            cuda_version: 12.9.1
+#            cudnn_version: ""
+#            python_version: "3.12"
+#            pytorch: 2.9.1
+#            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+#            dockerfile: "Dockerfile-uv-base"
+#            platforms: "linux/amd64,linux/arm64"
+          - cuda: "130"
+            cuda_version: 13.0.0
+            cudnn_version: ""
+            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"
+          - cuda: "130"
+            cuda_version: 13.0.0
+            cudnn_version: ""
+            python_version: "3.12"
+            pytorch: 2.10.0
+            torch_cuda_arch_list: "9.0+PTX"
+            dockerfile: "Dockerfile-uv-base"
+            platforms: "linux/amd64,linux/arm64"
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -160,17 +250,19 @@ jobs:
           images: |
             axolotlai/axolotl-base-uv
       - name: Login to Docker Hub
-        uses: docker/login-action@v2
+        uses: docker/login-action@v3
+        if: ${{ github.event_name != 'pull_request' && env.HAS_DOCKERHUB_CREDS == 'true' }}
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
       - name: Build
-        uses: docker/build-push-action@v4
+        uses: docker/build-push-action@v5
         with:
           context: .
           file: ./docker/${{ matrix.dockerfile }}
+          platforms: ${{ matrix.platforms }}
           push: ${{ github.event_name != 'pull_request' }}
           tags: ${{ steps.metadata.outputs.tags }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
           labels: ${{ steps.metadata.outputs.labels }}

.github/workflows/docs.yml

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

.github/workflows/lint.yml

@@ -13,6 +13,9 @@ on:
        - ".pre-commit-config.yaml"
   workflow_dispatch:
 
+permissions:
+  contents: read
+
 jobs:
   pre-commit:
     name: pre-commit

.github/workflows/main.yml

@@ -8,6 +8,9 @@ on:
       - "v*"
   workflow_dispatch:
 
+permissions:
+  contents: read
+
 jobs:
   build-axolotl:
     if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
@@ -15,27 +18,49 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.6.0
-            axolotl_extras:
-          - 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
+            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: 128
+            cuda_version: 12.8.1
+            python_version: "3.12"
+            pytorch: 2.10.0
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+#          - cuda: 129
+#            cuda_version: 12.9.1
+#            python_version: "3.12"
+#            pytorch: 2.9.1
+#            axolotl_extras:
+#            platforms: "linux/amd64,linux/arm64"
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.11"
+            pytorch: 2.9.1
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.12"
+            pytorch: 2.10.0
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -45,7 +70,6 @@ jobs:
         uses: docker/metadata-action@v5
         with:
           images: |
-            winglian/axolotl
             axolotlai/axolotl
           tags: |
             type=ref,event=branch
@@ -62,6 +86,7 @@ jobs:
         uses: docker/build-push-action@v5
         with:
           context: .
+          platforms: ${{ matrix.platforms }}
           build-args: |
             BASE_TAG=${{ github.ref_type == 'tag' && 'main' || github.ref_name }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}
             CUDA=${{ matrix.cuda }}
@@ -76,34 +101,134 @@ jobs:
             ${{ (matrix.is_latest) && format('{0}-latest', steps.metadata.outputs.tags) || '' }}
           labels: ${{ steps.metadata.outputs.labels }}
 
-  build-axolotl-cloud:
-    needs: build-axolotl
+  build-axolotl-uv:
     if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
-    # this job needs to be run on self-hosted GPU runners...
     strategy:
+      fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.6.0
-            axolotl_extras:
-          - 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: true
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.7.1
+            pytorch: 2.9.1
             axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.12"
+            pytorch: 2.9.1
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+            is_latest: true
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.12"
+            pytorch: 2.10.0
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.11"
+            pytorch: 2.9.1
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.12"
+            pytorch: 2.10.0
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+    runs-on: axolotl-gpu-runner
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Docker metadata
+        id: metadata
+        uses: docker/metadata-action@v5
+        with:
+          images: |
+            axolotlai/axolotl-uv
+          tags: |
+            type=ref,event=branch
+            type=pep440,pattern={{version}}
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_TOKEN }}
+      # guidance for testing before pushing: https://docs.docker.com/build/ci/github-actions/test-before-push/
+      - name: Build and export to Docker
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          platforms: ${{ matrix.platforms }}
+          build-args: |
+            BASE_TAG=${{ github.ref_type == 'tag' && 'main' || github.ref_name }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}
+            CUDA=${{ matrix.cuda }}
+            PYTORCH_VERSION=${{ matrix.pytorch }}
+            AXOLOTL_ARGS=${{ matrix.axolotl_args }}
+            AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}
+          file: ./docker/Dockerfile-uv
+          push: ${{ github.event_name != 'pull_request' }}
+          tags: |
+            ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
+            ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}
+            ${{ (matrix.is_latest) && format('{0}-latest', steps.metadata.outputs.tags) || '' }}
+          labels: ${{ steps.metadata.outputs.labels }}
+
+  build-axolotl-cloud:
+    needs: build-axolotl
+    if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
+    # this job needs to be run on self-hosted GPU runners...
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - 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: 128
+            cuda_version: 12.8.1
+            python_version: "3.12"
+            pytorch: 2.10.0
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+#          - cuda: 129
+#            cuda_version: 12.9.1
+#            python_version: "3.12"
+#            pytorch: 2.9.1
+#            axolotl_extras:
+#            platforms: "linux/amd64,linux/arm64"
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.11"
+            pytorch: 2.9.1
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.12"
+            pytorch: 2.10.0
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -113,7 +238,6 @@ jobs:
         uses: docker/metadata-action@v5
         with:
           images: |
-            winglian/axolotl-cloud
             axolotlai/axolotl-cloud
           tags: |
             type=ref,event=branch
@@ -129,6 +253,7 @@ jobs:
         uses: docker/build-push-action@v5
         with:
           context: .
+          platforms: ${{ matrix.platforms }}
           build-args: |
             BASE_TAG=${{ github.ref_type == 'tag' && 'main' || github.ref_name }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
             CUDA=${{ matrix.cuda }}
@@ -139,18 +264,100 @@ jobs:
              ${{ (matrix.is_latest) && format('{0}-latest', steps.metadata.outputs.tags) || '' }}
           labels: ${{ steps.metadata.outputs.labels }}
 
+  build-axolotl-cloud-uv:
+    needs: build-axolotl-uv
+    if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
+    # this job needs to be run on self-hosted GPU runners...
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.9.1
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.12"
+            pytorch: 2.9.1
+            axolotl_extras:
+            is_latest: true
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.12"
+            pytorch: 2.10.0
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.11"
+            pytorch: 2.9.1
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.12"
+            pytorch: 2.10.0
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+    runs-on: axolotl-gpu-runner
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Docker metadata
+        id: metadata
+        uses: docker/metadata-action@v5
+        with:
+          images: |
+            axolotlai/axolotl-cloud-uv
+          tags: |
+            type=ref,event=branch
+            type=pep440,pattern={{version}}
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_TOKEN }}
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      - name: Build
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          platforms: ${{ matrix.platforms }}
+          build-args: |
+            BASE_TAG=${{ github.ref_type == 'tag' && 'main' || github.ref_name }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
+            CUDA=${{ matrix.cuda }}
+          file: ./docker/Dockerfile-cloud-uv
+          push: ${{ github.event_name != 'pull_request' }}
+          tags: |
+             ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
+             ${{ (matrix.is_latest) && format('{0}-latest', steps.metadata.outputs.tags) || '' }}
+          labels: ${{ steps.metadata.outputs.labels }}
+
   build-axolotl-cloud-no-tmux:
     needs: build-axolotl
     if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
     # this job needs to be run on self-hosted GPU runners...
     strategy:
+      fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
+          - cuda: 128
+            cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.9.1
             axolotl_extras:
+            is_latest: true
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.11"
+            pytorch: 2.9.1
+            axolotl_extras:
+            is_latest:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -160,7 +367,6 @@ jobs:
         uses: docker/metadata-action@v5
         with:
           images: |
-            winglian/axolotl-cloud-term
             axolotlai/axolotl-cloud-term
           tags: |
             type=ref,event=branch
@@ -176,6 +382,7 @@ jobs:
         uses: docker/build-push-action@v5
         with:
           context: .
+          platforms: linux/amd64,linux/arm64
           build-args: |
             BASE_TAG=${{ github.ref_type == 'tag' && 'main' || github.ref_name }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
             CUDA=${{ matrix.cuda }}

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

@@ -8,6 +8,7 @@ on:
       - 'setup.py'
       - 'pyproject.toml'
       - '.github/workflows/multi-gpu-e2e.yml'
+      - 'scripts/cutcrossentropy_install.py'
       - 'src/axolotl/core/trainers/mixins/sequence_parallel.py'
       - 'src/axolotl/utils/distributed.py'
   workflow_dispatch:
@@ -19,6 +20,12 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
   cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
 
+permissions:
+  contents: read
+
+env:
+  MODAL_IMAGE_BUILDER_VERSION: "2025.06"
+
 jobs:
   test-axolotl-multigpu:
     if: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' && (github.event_name != 'pull_request' || !github.event.pull_request.draft) }}
@@ -26,27 +33,33 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
+          - cuda: 128
+            cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.8.0
+            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"
-          - cuda: 126
-            cuda_version: 12.6.3
+          - cuda: 128
+            cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.7.0
-            axolotl_extras:
+            pytorch: 2.10.0
+            axolotl_extras: "fbgemm-gpu"
             num_gpus: 2
-            nightly_build: "true"
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.7.1
-            axolotl_extras: vllm
-            num_gpus: 2
-            nightly_build: "true"
+            dockerfile: "Dockerfile-uv.jinja"
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
     steps:
@@ -59,7 +72,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2
+          pip install modal==1.3.0.post1 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
@@ -68,8 +81,9 @@ jobs:
           echo "AXOLOTL_EXTRAS=${{ 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
+        env:
+          CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
         run: |
-          modal run cicd.multigpu
+          modal run -m cicd.multigpu

.github/workflows/nightlies.yml

@@ -5,6 +5,9 @@ on:
   schedule:
     - cron: '0 0 * * *'  # Runs at 00:00 UTC every day
 
+permissions:
+  contents: read
+
 jobs:
   build-axolotl:
     if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
@@ -12,15 +15,15 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
+          - cuda: 128
+            cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.8.0
             axolotl_extras:
-          - cuda: 126
-            cuda_version: 12.6.3
+          - cuda: 128
+            cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.7.1
+            pytorch: 2.9.1
             axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
@@ -31,7 +34,6 @@ jobs:
         uses: docker/metadata-action@v5
         with:
           images: |
-            winglian/axolotl
             axolotlai/axolotl
           tags: |
             type=raw,value={{ branch }}-{{ date 'YYYYMMDD' }}
@@ -65,15 +67,15 @@ jobs:
     strategy:
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
+          - cuda: 128
+            cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.8.0
             axolotl_extras:
-          - cuda: 126
-            cuda_version: 12.6.3
+          - cuda: 128
+            cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.7.1
+            pytorch: 2.9.1
             axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
@@ -84,7 +86,6 @@ jobs:
         uses: docker/metadata-action@v5
         with:
           images: |
-            winglian/axolotl-cloud
             axolotlai/axolotl-cloud
           tags: |
             type=raw,value={{ branch }}-{{ date 'YYYYMMDD' }}

.github/workflows/precommit-autoupdate.yml

@@ -2,9 +2,11 @@ name: Pre-commit auto-update
 
 on:
   schedule:
-    - cron: '0 0 * * 0'  # Run weekly
+    - cron: '0 0 1 * *'  # Run monthly
   workflow_dispatch:  # Manual kickoff
 
+permissions: {}
+
 jobs:
   auto-update:
     runs-on: ubuntu-latest

.github/workflows/preview-docs.yml

@@ -11,22 +11,21 @@ on:
       - '_quarto.yml'
       - docs/scripts/generate_config_docs.py
       - src/axolotl/utils/schemas/**.py
+      - .github/workflows/preview-docs.yml
 
 permissions:
-  checks: write
-  contents: write
-  deployments: write
-  issues: write
-  discussions: write
-  pages: write
+  contents: read
   pull-requests: write
-  statuses: write
 
 jobs:
   preview:
     runs-on: ubuntu-latest
     if: ${{ !github.event.pull_request.draft }}
     steps:
+      - name: cleanup node
+        run: |
+          sudo rm -rf /usr/share/dotnet /usr/local/lib/android /opt/ghc /opt/hostedtoolcache/CodeQL
+
       - name: Check out repository
         uses: actions/checkout@v4
         with:

.github/workflows/pypi.yml

@@ -3,9 +3,11 @@ name: publish pypi
 on:
   push:
     tags:
-      - 'v*'
+      - "v*"
   workflow_dispatch:
 
+permissions: {}
+
 jobs:
   setup_release:
     name: Create Release
@@ -28,7 +30,8 @@ jobs:
       name: pypi
       url: https://pypi.org/p/axolotl
     permissions:
-      id-token: write  # IMPORTANT: this permission is mandatory for trusted publishing
+      contents: read
+      id-token: write # IMPORTANT: this permission is mandatory for trusted publishing
     steps:
       - name: Check out repository code
         uses: actions/checkout@v4
@@ -40,17 +43,17 @@ jobs:
 
       - name: Install dependencies
         run: |
-          pip3 install wheel packaging==23.2
+          pip3 install wheel packaging==26.0
           pip3 install --no-build-isolation -e .
           pip3 install -r requirements-dev.txt -r requirements-tests.txt
 
       - name: Extract tag name
         id: tag
-        run: echo ::set-output name=TAG_NAME::$(echo $GITHUB_REF | cut -d / -f 3)
+        run: echo "TAG_NAME=$(echo $GITHUB_REF | cut -d / -f 3)" >> "$GITHUB_OUTPUT"
 
-      - name: Update version in setup.py
+      - name: Update version in VERSION file
         run: |
-          sed -i -E 's/version="([0-9.]+)",/version="${{ steps.tag.outputs.TAG_NAME }}",/g' setup.py
+          echo "${{ steps.tag.outputs.TAG_NAME }}" | sed 's/^v//' > VERSION
 
       - name: Build a source dist
         run: |

.github/workflows/tests-nightly.yml

@@ -3,6 +3,13 @@ on:
   workflow_dispatch:
   schedule:
     - cron: '0 0 * * *'  # Runs at 00:00 UTC every day
+  pull_request:
+    types: [opened, synchronize, reopened, ready_for_review]
+    paths:
+      - '.github/workflows/tests-nightly.yml'
+
+permissions:
+  contents: read
 
 jobs:
   pre-commit:
@@ -18,15 +25,26 @@ jobs:
         env:
           SKIP: no-commit-to-branch
 
+  prime-cdn-s3-cache:
+    name: Prefetch S3 once to prime the CDN cache
+    runs-on: ubuntu-latest
+    if: ${{ !github.event.pull_request.draft }}
+    timeout-minutes: 10
+    steps:
+      - name: Restore Cache from S3
+        id: hf-cache-restore-s3
+        run: |
+          curl -v -H "Range: bytes=0-1023" -L https://axolotl-ci.b-cdn.net/hf-cache.tar.zst > /dev/null
+
   pytest:
     name: PyTest
     runs-on: ubuntu-latest
+    needs: [prime-cdn-s3-cache]
     strategy:
       fail-fast: false
-      max-parallel: 2
       matrix:
-        python_version: ["3.11"]
-        pytorch_version: ["2.6.0", "2.7.0"]
+        python_version: ["3.12"]  # TODO include py3.14 once https://github.com/mistralai/mistral-common/pull/194 is merged
+        pytorch_version: ["2.8.0", "2.9.1", "2.10.0"]
     timeout-minutes: 20
 
     steps:
@@ -37,7 +55,7 @@ jobs:
         id: hf-cache-restore-s3
         run: |
           mkdir -p /home/runner/.cache/huggingface/hub
-          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C /home/runner/.cache/huggingface/hub/  --use-compress-program unzstd
+          curl -L https://axolotl-ci.b-cdn.net/hf-cache.tar.zst | tar -xf - -C /home/runner/.cache/huggingface/hub/  --use-compress-program unzstd
 
       - name: Setup Python
         uses: actions/setup-python@v5
@@ -48,7 +66,7 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
+          pip3 install --upgrade packaging==26.0 setuptools==78.1.1 wheel
 
       - name: Install PyTorch
         run: |
@@ -99,19 +117,26 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
+          - cuda: 128
+            cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.9.1
             num_gpus: 1
             axolotl_extras:
             nightly_build: "true"
-          - cuda: 126
-            cuda_version: 12.6.3
+          - cuda: 128
+            cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.7.1
+            pytorch: 2.10.0
             num_gpus: 1
             axolotl_extras:
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.12"
+            pytorch: 2.9.1
+            num_gpus: 1
+            axolotl_extras:
+            dockerfile: "Dockerfile-uv.jinja"
             nightly_build: "true"
     steps:
       - name: Checkout
@@ -123,7 +148,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2
+          pip install modal==1.3.0.post1 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
@@ -132,9 +157,11 @@ jobs:
           echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
+          echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
           echo "NIGHTLY_BUILD=${{ matrix.nightly_build }}" >> $GITHUB_ENV
-          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
       - name: Run tests job on Modal
+        env:
+          CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
         run: |
           modal run cicd.e2e_tests
   docker-e2e-multigpu-tests:
@@ -148,10 +175,10 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
+          - cuda: 128
+            cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.7.1
+            pytorch: 2.9.1
             num_gpus: 2
             axolotl_extras:
             nightly_build: "true"
@@ -165,7 +192,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2
+          pip install modal==1.3.0.post1 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
@@ -175,7 +202,8 @@ jobs:
           echo "CUDA=${{ 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
       - name: Run tests job on Modal
+        env:
+          CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
         run: |
           modal run cicd.multigpu

.github/workflows/tests.yml

@@ -28,6 +28,9 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
   cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
 
+permissions:
+  contents: read
+
 env:
   TRANSFORMERS_IS_CI: "yes"
 
@@ -46,27 +49,48 @@ jobs:
         env:
           SKIP: no-commit-to-branch
 
+  prime-cdn-s3-cache:
+    name: Prefetch S3 once to prime the CDN cache
+    runs-on: ubuntu-latest
+    if: ${{ !github.event.pull_request.draft }}
+    timeout-minutes: 10
+    steps:
+      - name: Restore Cache from S3
+        id: hf-cache-restore-s3
+        run: |
+          curl -v -H "Range: bytes=0-1023" -L https://axolotl-ci.b-cdn.net/hf-cache.tar.zst > /dev/null
+
   pytest:
     name: PyTest
     runs-on: ubuntu-latest
     if: ${{ !github.event.pull_request.draft }}
-#    needs: [preload-cache]
+    needs: [prime-cdn-s3-cache]
     strategy:
       fail-fast: false
       matrix:
-        python_version: ["3.11"]
-        pytorch_version: ["2.6.0", "2.7.0", "2.7.1"]
+        python_version: ["3.12"]  # TODO include py3.14 once https://github.com/mistralai/mistral-common/pull/194 is merged
+        pytorch_version: ["2.8.0", "2.9.1", "2.10.0"]
+#        exclude:
+#          - python_version: "3.14"
+#            pytorch_version: "2.8.0"
+#          - python_version: "3.14"
+#            pytorch_version: "2.9.1"
     timeout-minutes: 20
 
     steps:
+      - 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 /home/runner/.cache/huggingface/hub
-          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C /home/runner/.cache/huggingface/hub/  --use-compress-program unzstd
+          mkdir -p ~/.cache/huggingface/hub
+          curl -L https://axolotl-ci.b-cdn.net/hf-cache.tar.zst | tar -xpf - -C ~/.cache/huggingface/hub/  --use-compress-program unzstd --strip-components=1
+          ls -ltr ~/.cache/huggingface/hub/
 
       - name: Setup Python
         uses: actions/setup-python@v5
@@ -77,20 +101,24 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
+          pip3 install --upgrade packaging==26.0 setuptools==75.8.0 wheel
 
       - name: Install PyTorch
         run: |
-          pip3 install torch==${{ matrix.pytorch_version }} torchvision
+          pip3 install --no-cache-dir torch==${{ matrix.pytorch_version }} torchvision
 
       - name: Install dependencies
         run: |
           pip3 show torch
-          pip3 install --no-build-isolation -U -e .
+          pip3 install --no-cache-dir --no-build-isolation -U -e .
           python scripts/unsloth_install.py | sh
           python scripts/cutcrossentropy_install.py | sh
           pip3 install -r requirements-dev.txt -r requirements-tests.txt
 
+      - name: 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__"
@@ -101,14 +129,25 @@ jobs:
 
       - name: Pre-Download dataset fixture
         run: |
-          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
+          hf download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
+
+      - name: Show HF cache
+        run: hf cache ls
 
       - name: Run tests
         run: |
-          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/ --cov=axolotl --cov-report=xml
+          df -h
+          pytest -v --durations=10 -n4 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ --ignore=tests/monkeypatch/ tests/ --cov=axolotl --cov-report=xml
+          df -h
+          pytest -v --durations=10 tests/monkeypatch/ --cov=axolotl --cov-append --cov-report=xml
+          df -h
           pytest -v --durations=10 tests/patched/ --cov=axolotl --cov-append --cov-report=xml
+          df -h
           pytest -v --durations=10 tests/cli/ --cov=axolotl --cov-append --cov-report=xml
 
+      - name: Show HF cache
+        run: hf cache ls
+
       - name: Upload coverage to Codecov
         uses: codecov/codecov-action@v5
         with:
@@ -117,30 +156,37 @@ 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
     if: ${{ !github.event.pull_request.draft }}
+    needs: [prime-cdn-s3-cache]
     strategy:
       fail-fast: false
       matrix:
-        python_version: ["3.11"]
-        pytorch_version: ["2.6.0", "2.7.0", "2.7.1"]
-    timeout-minutes: 20
+        python_version: ["3.12"]  # TODO include py3.14 once https://github.com/mistralai/mistral-common/pull/194 is merged
+        pytorch_version: ["2.8.0", "2.9.1", "2.10.0"]
+#        exclude:
+#          - python_version: "3.14"
+#            pytorch_version: "2.8.0"
+#          - python_version: "3.14"
+#            pytorch_version: "2.9.1"
+    timeout-minutes: 30
 
     steps:
+      - name: cleanup node
+        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 /home/runner/.cache/huggingface/hub
-          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C /home/runner/.cache/huggingface/hub/  --use-compress-program unzstd
+          mkdir -p ~/.cache/huggingface/hub
+          curl -L https://axolotl-ci.b-cdn.net/hf-cache.tar.zst | tar -xpf - -C ~/.cache/huggingface/hub/  --use-compress-program unzstd --strip-components=1
+          ls -ltr ~/.cache/huggingface/hub/
 
       - name: Setup Python
         uses: actions/setup-python@v5
@@ -151,21 +197,25 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 setuptools_scm build wheel
+          pip3 install --upgrade packaging==26.0 setuptools==75.8.0 setuptools_scm build wheel psutil
 
       - name: Install PyTorch
         run: |
-          pip3 install torch==${{ matrix.pytorch_version }} torchvision
+          pip3 install --no-cache-dir torch==${{ matrix.pytorch_version }} torchvision
 
       - name: Install dependencies
         run: |
           pip3 show torch
           python -m build --no-isolation --sdist
-          pip3 install --no-build-isolation dist/axolotl*.tar.gz
+          pip3 install --no-cache-dir --no-build-isolation dist/axolotl*.tar.gz
           python scripts/unsloth_install.py | sh
           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__"
@@ -175,40 +225,64 @@ jobs:
           axolotl --help
 
       - name: Show HF cache
-        run: huggingface-cli scan-cache
+        run: hf cache ls
 
       - name: Run tests
         run: |
-          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/
-          pytest -v --durations=10 tests/patched/
+          pytest -v --durations=10 -n4 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ --ignore=tests/monkeypatch/ tests/ --cov=axolotl --cov-report=xml
+          pytest -v --durations=10 tests/monkeypatch/ --cov=axolotl --cov-append --cov-report=xml
           pytest -v --durations=10 tests/cli/
 
-      - name: cleanup pip cache
-        run: |
-          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
+      - name: Show HF cache
+        run: hf cache ls
+
+  gate-skip-e2e:
+    needs: [pre-commit]
+    runs-on: ubuntu-latest
+    outputs:
+      skip: ${{ steps.compute.outputs.skip }}
+    steps:
+      - uses: actions/github-script@v7
+        id: compute
+        with:
+          script: |
+            const token = /\[skip-e2e\]/i;
+            let msg = '';
+            if (context.eventName === 'push') {
+              msg = context.payload.head_commit?.message || '';
+            } else if (context.eventName === 'pull_request') {
+              const { owner, repo } = context.repo;
+              const prNumber = context.payload.pull_request.number;
+              const commits = await github.paginate(
+                github.rest.pulls.listCommits,
+                { owner, repo, pull_number: prNumber, per_page: 100 }
+              );
+              msg = commits.at(-1)?.commit?.message || '';
+            }
+            const title = context.payload.pull_request?.title || '';
+            const body  = context.payload.pull_request?.body  || '';
+            const skip = token.test(msg) || token.test(title) || token.test(body);
+            core.setOutput('skip', String(skip));
 
   docker-e2e-tests-1st:
     # Run this job first as a gate for running the remainder of the test matrix
-    if: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' && !github.event.pull_request.draft }}
+    if: >
+      github.repository_owner == 'axolotl-ai-cloud' &&
+      (github.event_name != 'pull_request' || !github.event.pull_request.draft) &&
+      needs.gate-skip-e2e.outputs.skip != 'true'
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
-    needs: [pre-commit, pytest, pytest-sdist]
+    needs: [pre-commit, pytest]
 
     strategy:
       fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.7.1
-            num_gpus: 1
-            axolotl_extras:
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.6.0
+          - cuda: 130
+            cuda_version: 13.0.0
+            python_version: "3.12"
+            pytorch: 2.9.1
             num_gpus: 1
             axolotl_extras:
             dockerfile: "Dockerfile-uv.jinja"
@@ -222,7 +296,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2
+          pip install modal==1.3.0.post1 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
@@ -232,35 +306,52 @@ jobs:
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
-          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
           echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
       - name: Run tests job on Modal
+        env:
+          CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
         run: |
           modal run cicd.e2e_tests
 
   docker-e2e-tests:
-    if: ${{ github.repository_owner == 'axolotl-ai-cloud' && !github.event.pull_request.draft }}
+    if: >
+      github.repository_owner == 'axolotl-ai-cloud' &&
+      (github.event_name != 'pull_request' || !github.event.pull_request.draft) &&
+      needs.gate-skip-e2e.outputs.skip != 'true'
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
     # Only run the remainder of the matrix if the first e2e check passed;
     # this is to save on wasted compute costs for known failures that get caught in the first run
-    needs: [pre-commit, pytest, docker-e2e-tests-1st]
+    needs: [pre-commit, pytest, gate-skip-e2e, docker-e2e-tests-1st]
 
     strategy:
       fail-fast: false
       matrix:
         include:
-          - cuda: 126
-            cuda_version: 12.6.3
+          - cuda: 128
+            cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.8.0
+            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: 128
             cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.7.1
+            pytorch: 2.10.0
+            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:
@@ -273,7 +364,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2
+          pip install modal==1.3.0.post1 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
@@ -283,9 +374,11 @@ jobs:
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
-          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
+          echo "GPU_TYPE=${{ matrix.gpu_type || 'L40S'}}" >> $GITHUB_ENV
           echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
       - name: Run tests job on Modal
+        env:
+          CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
         run: |
           modal run cicd.e2e_tests
 
@@ -299,10 +392,10 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 128
+            cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.9.1
             num_gpus: 1
             axolotl_extras:
     steps:
@@ -315,7 +408,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2
+          pip install modal==1.3.0.post1 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
@@ -325,7 +418,6 @@ jobs:
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
-          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         run: |
           modal run cicd.cleanup

.gitignore

@@ -190,3 +190,6 @@ out/
 
 # vim
 *.swp
+
+# scm auto-versioning
+src/axolotl/_version.py

.isort.cfg

@@ -1,4 +0,0 @@
-[settings]
-profile=black
-known_third_party=wandb,comet_ml
-known_local_folder=src,tests

.pre-commit-config.yaml

@@ -3,31 +3,21 @@ default_language_version:
 
 repos:
 -   repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v5.0.0
+    rev: v6.0.0
     hooks:
     -   id: check-yaml
     -   id: end-of-file-fixer
     -   id: trailing-whitespace
     -   id: no-commit-to-branch
         args: ['--branch', 'main']
--   repo: https://github.com/psf/black
-    rev: 25.1.0
+-   repo: https://github.com/astral-sh/ruff-pre-commit
+    rev: v0.15.4
     hooks:
-    -   id: black
--   repo: https://github.com/pycqa/isort
-    rev: 6.0.1
-    hooks:
-      - id: isort
--   repo: https://github.com/PyCQA/flake8
-    rev: 7.3.0
-    hooks:
-    - id: flake8
--   repo: https://github.com/pylint-dev/pylint
-    rev: v3.3.7
-    hooks:
-    - id: pylint
+    -   id: ruff
+        args: [--fix]
+    -   id: ruff-format
 -   repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.17.1
+    rev: v1.19.1
     hooks:
     - id: mypy
       additional_dependencies:
@@ -36,7 +26,7 @@ repos:
             'pydantic>=2.5.3',
         ]
 -   repo: https://github.com/PyCQA/bandit
-    rev: 1.8.6
+    rev: 1.9.4
     hooks:
     -   id: bandit
         args: [

.pylintrc

@@ -1,15 +0,0 @@
-[MASTER]
-init-hook="from pylint.config import find_default_config_files; import sys; sys.path.append(next(find_default_config_files()).parent.as_posix())"
-
-[TYPECHECK]
-
-# List of members which are set dynamically and missed by Pylint inference
-# system, and so shouldn't trigger E1101 when accessed.
-generated-members=numpy.*, torch.*
-
-
-[pylint.messages_control]
-disable=missing-function-docstring, line-too-long, import-error,
-    too-many-arguments, too-many-locals, too-many-statements, too-many-branches, too-few-public-methods,
-    too-many-instance-attributes, fixme, import-outside-toplevel, logging-fstring-interpolation,
-    too-many-positional-arguments, possibly-used-before-assignment

.runpod/Dockerfile

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

.runpod/README.md

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

.runpod/src/config/config.yaml

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

CITATION.cff

@@ -0,0 +1,10 @@
+cff-version: 1.2.0
+type: software
+title: "Axolotl: Open Source LLM Post-Training"
+message: "If you use this software, please cite it as below."
+authors:
+  - name: "Axolotl maintainers and contributors"
+repository-code: "https://github.com/axolotl-ai-cloud/axolotl"
+url: "https://axolotl.ai/"
+license: Apache-2.0
+date-released: "2023-05-30"

README.md

@@ -5,6 +5,9 @@
         <img alt="Axolotl" src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/887513285d98132142bf5db2a74eb5e0928787f1/image/axolotl_logo_digital_black.svg" width="400" height="104" style="max-width: 100%;">
     </picture>
 </p>
+  <p align="center">
+      <strong>A Free and Open Source LLM Fine-tuning Framework</strong><br>
+  </p>
 
 <p align="center">
     <img src="https://img.shields.io/github/license/axolotl-ai-cloud/axolotl.svg?color=blue" alt="GitHub License">
@@ -17,6 +20,7 @@
     <br/>
     <a href="https://discord.com/invite/HhrNrHJPRb"><img src="https://img.shields.io/badge/discord-7289da.svg?style=flat-square&logo=discord" alt="discord" style="height: 20px;"></a>
     <a href="https://twitter.com/axolotl_ai"><img src="https://img.shields.io/twitter/follow/axolotl_ai?style=social" alt="twitter" style="height: 20px;"></a>
+    <a href="https://colab.research.google.com/github/axolotl-ai-cloud/axolotl/blob/main/examples/colab-notebooks/colab-axolotl-example.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="google-colab" style="height: 20px;"></a>
     <br/>
     <img src="https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/tests-nightly.yml/badge.svg" alt="tests-nightly">
     <img src="https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/multi-gpu-e2e.yml/badge.svg" alt="multigpu-semi-weekly tests">
@@ -25,21 +29,35 @@
 
 ## 🎉 Latest Updates
 
-- 2025/07:
-  - ND Parallelism support has been added into Axolotl. Compose Context Parallelism (CP), Tensor Parallelism (TP), and Fully Sharded Data Parallelism (FSDP) within a single node and across multiple nodes. Check out the [blog post](https://huggingface.co/blog/accelerate-nd-parallel) for more info.
-  - Axolotl adds more models: [GPT-OSS](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/gpt-oss), [Gemma 3n](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/gemma3n), [Liquid Foundation Model 2 (LFM2)](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/lfm2), and [Arcee Foundation Models (AFM)](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/afm).
-  - FP8 finetuning with fp8 gather op is now possible in Axolotl via `torchao`. Get started [here](https://docs.axolotl.ai/docs/mixed_precision.html#sec-fp8)!
-  - [Voxtral](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/voxtral), [Magistral 1.1](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/magistral), and [Devstral](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/devstral) with mistral-common tokenizer support has been integrated in Axolotl!
-  - 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.
+- 2026/03:
+  - New model support has been added in Axolotl for [Mistral Small 4](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/mistral4), [Qwen3.5, Qwen3.5 MoE](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/qwen3.5), [GLM-4.7-Flash](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/glm47-flash), [GLM-4.6V](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/glm46v), and [GLM-4.5-Air](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/glm45).
+  - [MoE expert quantization](https://docs.axolotl.ai/docs/expert_quantization.html) support (via `quantize_moe_experts: true`) greatly reduces VRAM when training MoE models (FSDP2 compat).
+- 2026/02:
+  - [ScatterMoE LoRA](https://github.com/axolotl-ai-cloud/axolotl/pull/3410) support. LoRA fine-tuning directly on MoE expert weights using custom Triton kernels.
+  - Axolotl now has support for [SageAttention](https://github.com/axolotl-ai-cloud/axolotl/pull/2823) and [GDPO](https://github.com/axolotl-ai-cloud/axolotl/pull/3353) (Generalized DPO).
+- 2026/01:
+  - New integration for [EAFT](https://github.com/axolotl-ai-cloud/axolotl/pull/3366) (Entropy-Aware Focal Training), weights loss by entropy of the top-k logit distribution, and [Scalable Softmax](https://github.com/axolotl-ai-cloud/axolotl/pull/3338), improves long context in attention.
+- 2025/12:
+  - Axolotl now includes support for [Kimi-Linear](https://docs.axolotl.ai/docs/models/kimi-linear.html), [Plano-Orchestrator](https://docs.axolotl.ai/docs/models/plano.html), [MiMo](https://docs.axolotl.ai/docs/models/mimo.html), [InternVL 3.5](https://docs.axolotl.ai/docs/models/internvl3_5.html), [Olmo3](https://docs.axolotl.ai/docs/models/olmo3.html), [Trinity](https://docs.axolotl.ai/docs/models/trinity.html), and [Ministral3](https://docs.axolotl.ai/docs/models/ministral3.html).
+  - [Distributed Muon Optimizer](https://github.com/axolotl-ai-cloud/axolotl/pull/3264) support has been added for FSDP2 pretraining.
+- 2025/10: New model support has been added in Axolotl for: [Qwen3 Next](https://docs.axolotl.ai/docs/models/qwen3-next.html), [Qwen2.5-vl, Qwen3-vl](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/qwen2_5-vl), [Qwen3, Qwen3MoE](https://docs.axolotl.ai/docs/models/qwen3.html), [Granite 4](https://docs.axolotl.ai/docs/models/granite4.html), [HunYuan](https://docs.axolotl.ai/docs/models/hunyuan.html), [Magistral 2509](https://docs.axolotl.ai/docs/models/magistral/vision.html), [Apertus](https://docs.axolotl.ai/docs/models/apertus.html), and [Seed-OSS](https://docs.axolotl.ai/docs/models/seed-oss.html).
 
 <details>
 
 <summary>Expand older updates</summary>
 
-- 2025/06: 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!
+- 2025/09: 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).
+  - FP8 finetuning with fp8 gather op is now possible in Axolotl via `torchao`. Get started [here](https://docs.axolotl.ai/docs/mixed_precision.html#sec-fp8)!
+  - [Voxtral](https://docs.axolotl.ai/docs/models/voxtral.html), [Magistral 1.1](https://docs.axolotl.ai/docs/models/magistral.html), and [Devstral](https://docs.axolotl.ai/docs/models/devstral.html) with mistral-common tokenizer support has been integrated in Axolotl!
+  - TiledMLP support for single-GPU to multi-GPU training with DDP, DeepSpeed and FSDP support has been added to support Arctic Long Sequence Training. (ALST). See [examples](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/alst) for using ALST with Axolotl!
+- 2025/06: Magistral with mistral-common tokenizer support has been added to Axolotl. See [docs](https://docs.axolotl.ai/docs/models/magistral.html) to start training your own Magistral models with Axolotl!
+- 2025/05: Quantization Aware Training (QAT) support has been added to Axolotl. Explore the [docs](https://docs.axolotl.ai/docs/qat.html) to learn more!
+- 2025/04: Llama 4 support has been added in Axolotl. See [docs](https://docs.axolotl.ai/docs/models/llama-4.html) to start training your own Llama 4 models with Axolotl's linearized version!
+- 2025/03: Axolotl has implemented Sequence Parallelism (SP) support. Read the [blog](https://huggingface.co/blog/axolotl-ai-co/long-context-with-sequence-parallelism-in-axolotl) and [docs](https://docs.axolotl.ai/docs/sequence_parallelism.html) to learn how to scale your context length when fine-tuning.
 - 2025/03: (Beta) Fine-tuning Multimodal models is now supported in Axolotl. Check out the [docs](https://docs.axolotl.ai/docs/multimodal.html) to fine-tune your own!
 - 2025/02: Axolotl has added LoRA optimizations to reduce memory usage and improve training speed for LoRA and QLoRA in single GPU and multi-GPU training (DDP and DeepSpeed). Jump into the [docs](https://docs.axolotl.ai/docs/lora_optims.html) to give it a try.
 - 2025/02: Axolotl has added GRPO support. Dive into our [blog](https://huggingface.co/blog/axolotl-ai-co/training-llms-w-interpreter-feedback-wasm) and [GRPO example](https://github.com/axolotl-ai-cloud/grpo_code) and have some fun!
@@ -49,33 +67,38 @@
 
 ## ✨ Overview
 
-Axolotl is a tool designed to streamline post-training for various AI models.
+Axolotl is a free and open-source tool designed to streamline post-training and fine-tuning for the latest large language models (LLMs).
 
 Features:
 
-- **Multiple Model Support**: Train various models like LLaMA, Mistral, Mixtral, Pythia, and more. We are compatible with HuggingFace transformers causal language models.
-- **Training Methods**: Full fine-tuning, LoRA, QLoRA, GPTQ, QAT, Preference Tuning (DPO, IPO, KTO, ORPO), RL (GRPO), Multimodal, and Reward Modelling (RM) / Process Reward Modelling (PRM).
-- **Easy Configuration**: Re-use a single YAML file between dataset preprocess, training, evaluation, quantization, and inference.
-- **Performance Optimizations**: [Multipacking](https://docs.axolotl.ai/docs/multipack.html), [Flash Attention](https://github.com/Dao-AILab/flash-attention), [Xformers](https://github.com/facebookresearch/xformers), [Flex Attention](https://pytorch.org/blog/flexattention/), [Liger Kernel](https://github.com/linkedin/Liger-Kernel), [Cut Cross Entropy](https://github.com/apple/ml-cross-entropy/tree/main), [Sequence Parallelism (SP)](https://docs.axolotl.ai/docs/sequence_parallelism.html), [LoRA optimizations](https://docs.axolotl.ai/docs/lora_optims.html), [Multi-GPU training (FSDP1, FSDP2, DeepSpeed)](https://docs.axolotl.ai/docs/multi-gpu.html), [Multi-node training (Torchrun, Ray)](https://docs.axolotl.ai/docs/multi-node.html), and many more!
+- **Multiple Model Support**: Train various models like GPT-OSS, LLaMA, Mistral, Mixtral, Pythia, and many more models available on the Hugging Face Hub.
+- **Multimodal Training**: Fine-tune vision-language models (VLMs) including LLaMA-Vision, Qwen2-VL, Pixtral, LLaVA, SmolVLM2, GLM-4.6V, InternVL 3.5, Gemma 3n, and audio models like Voxtral with image, video, and audio support.
+- **Training Methods**: Full fine-tuning, LoRA, QLoRA, GPTQ, QAT, Preference Tuning (DPO, IPO, KTO, ORPO), RL (GRPO, GDPO), and Reward Modelling (RM) / Process Reward Modelling (PRM).
+- **Easy Configuration**: Re-use a single YAML configuration file across the full fine-tuning pipeline: dataset preprocessing, training, evaluation, quantization, and inference.
+- **Performance Optimizations**: [Multipacking](https://docs.axolotl.ai/docs/multipack.html), [Flash Attention 2/3/4](https://docs.axolotl.ai/docs/attention.html#flash-attention), [Xformers](https://docs.axolotl.ai/docs/attention.html#xformers), [Flex Attention](https://docs.axolotl.ai/docs/attention.html#flex-attention), [SageAttention](https://docs.axolotl.ai/docs/attention.html#sageattention), [Liger Kernel](https://docs.axolotl.ai/docs/custom_integrations.html#liger-kernels), [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy), [ScatterMoE](https://docs.axolotl.ai/docs/custom_integrations.html#kernels-integration), [Sequence Parallelism (SP)](https://docs.axolotl.ai/docs/sequence_parallelism.html), [LoRA optimizations](https://docs.axolotl.ai/docs/lora_optims.html), [Multi-GPU training (FSDP1, FSDP2, DeepSpeed)](https://docs.axolotl.ai/docs/multi-gpu.html), [Multi-node training (Torchrun, Ray)](https://docs.axolotl.ai/docs/multi-node.html), and many more!
 - **Flexible Dataset Handling**: Load from local, HuggingFace, and cloud (S3, Azure, GCP, OCI) datasets.
 - **Cloud Ready**: We ship [Docker images](https://hub.docker.com/u/axolotlai) and also [PyPI packages](https://pypi.org/project/axolotl/) for use on cloud platforms and local hardware.
 
 
 
-## 🚀 Quick Start
+## 🚀 Quick Start - LLM Fine-tuning in Minutes
 
 **Requirements**:
 
 - NVIDIA GPU (Ampere or newer for `bf16` and Flash Attention) or AMD GPU
 - Python 3.11
-- PyTorch ≥2.6.0
+- PyTorch ≥2.8.0
+
+### Google Colab
+
+[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/axolotl-ai-cloud/axolotl/blob/main/examples/colab-notebooks/colab-axolotl-example.ipynb#scrollTo=msOCO4NRmRLa)
 
 ### Installation
 
 #### Using pip
 
 ```bash
-pip3 install -U packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install -U packaging==26.0 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation axolotl[flash-attn,deepspeed]
 
 # Download example axolotl configs, deepspeed configs
@@ -145,10 +168,31 @@ 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)
 
+## 📝 Citing Axolotl
+
+If you use Axolotl in your research or projects, please cite it as follows:
+
+```bibtex
+@software{axolotl,
+  title = {Axolotl: Open Source LLM Post-Training},
+  author = {{Axolotl maintainers and contributors}},
+  url = {https://github.com/axolotl-ai-cloud/axolotl},
+  license = {Apache-2.0},
+  year = {2023}
+}
+```
+
 ## 📜 License
 
 This project is licensed under the Apache 2.0 License - see the [LICENSE](LICENSE) file for details.

TODO.md

@@ -1,10 +0,0 @@
-# todo list
-
-- [] Validation of parameters for combinations that won't work
-
-
-
-## things that are known not to work
-
-- FSDP offload and gradient_checkpointing - https://github.com/pytorch/pytorch/issues/82203
-- adamw_bnb_8bit doesn't play well with FSDP offload

VERSION

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

_quarto.yml

@@ -1,6 +1,8 @@
 project:
   type: website
-  pre-render: docs/scripts/generate_config_docs.py
+  pre-render:
+   - docs/scripts/generate_config_docs.py
+   - docs/scripts/generate_examples_docs.py
 
 quartodoc:
   dir: docs/api
@@ -126,11 +128,9 @@ quartodoc:
         - monkeypatch.mistral_attn_hijack_flash
         - monkeypatch.multipack
         - monkeypatch.relora
-        - monkeypatch.llama_expand_mask
         - monkeypatch.lora_kernels
         - monkeypatch.utils
         - monkeypatch.btlm_attn_hijack_flash
-        - monkeypatch.llama_patch_multipack
         - monkeypatch.stablelm_attn_hijack_flash
         - monkeypatch.trainer_fsdp_optim
         - monkeypatch.transformers_fa_utils
@@ -153,7 +153,7 @@ quartodoc:
         - utils.distributed
         - utils.dict
         - utils.optimizers.adopt
-        - utils.data.pretraining
+        - utils.data.streaming
         - utils.data.sft
         - utils.quantization
     - title: Schemas
@@ -240,7 +240,48 @@ website:
             - docs/getting-started.qmd
             - docs/installation.qmd
             - docs/inference.qmd
+            - section: "Model Guides"
+              contents:
+                - docs/models/kimi-linear.qmd
+                - docs/models/plano.qmd
+                - docs/models/mimo.qmd
+                - docs/models/internvl3_5.qmd
+                - docs/models/olmo3.qmd
+                - docs/models/trinity.qmd
+                - docs/models/arcee.qmd
+                - section: "Ministral3"
+                  contents:
+                    - docs/models/ministral3.qmd
+                    - docs/models/ministral3/think.qmd
+                    - docs/models/ministral3/vision.qmd
+                - section: "Magistral"
+                  contents:
+                    - docs/models/magistral.qmd
+                    - docs/models/magistral/think.qmd
+                    - docs/models/magistral/vision.qmd
+                - docs/models/ministral.qmd
+                - docs/models/mistral-small.qmd
+                - docs/models/voxtral.qmd
+                - docs/models/devstral.qmd
+                - docs/models/mistral.qmd
+                - docs/models/llama-4.qmd
+                - docs/models/llama-2.qmd
+                - docs/models/qwen3-next.qmd
+                - docs/models/qwen3.qmd
+                - docs/models/gemma3n.qmd
+                - docs/models/apertus.qmd
+                - docs/models/gpt-oss.qmd
+                - docs/models/seed-oss.qmd
+                - docs/models/phi.qmd
+                - docs/models/smolvlm2.qmd
+                - docs/models/granite4.qmd
+                - docs/models/LiquidAI.qmd
+                - docs/models/hunyuan.qmd
+                - docs/models/jamba.qmd
+                - docs/models/orpheus.qmd
+
             - docs/cli.qmd
+            - docs/telemetry.qmd
             - docs/config-reference.qmd
             - text: "API Reference"
               href: docs/api
@@ -267,14 +308,17 @@ website:
             - docs/dataset_loading.qmd
             - docs/qat.qmd
             - docs/quantize.qmd
+            - docs/optimizations.qmd
 
         - section: "Core Concepts"
           contents:
             - docs/batch_vs_grad.qmd
             - docs/dataset_preprocessing.qmd
+            - docs/streaming.qmd
             - docs/multipack.qmd
             - docs/mixed_precision.qmd
             - docs/optimizers.qmd
+            - docs/attention.qmd
 
         - section: "Advanced Features"
           contents:
@@ -285,6 +329,7 @@ website:
             - docs/sequence_parallelism.qmd
             - docs/gradient_checkpointing.qmd
             - docs/nd_parallelism.qmd
+            - docs/expert_quantization.qmd
 
         - section: "Troubleshooting"
           contents:

benchmarks/bench_entropy.py

@@ -0,0 +1,208 @@
+"""Benchmark for entropy_from_logits Triton kernel vs original chunked implementation.
+
+Usage: CUDA_VISIBLE_DEVICES=0 python benchmarks/bench_entropy.py
+"""
+
+import gc
+import statistics
+
+import torch
+import torch.nn.functional as F
+
+from axolotl.monkeypatch.trainer.utils import entropy_from_logits
+
+V = 151936  # Qwen vocab
+WARMUP = 5
+BENCH_ITERS = 20
+MEM_ITERS = 10
+
+
+def entropy_from_logits_original(logits: torch.Tensor, chunk_size: int = 128):
+    """Original chunked implementation (reference)."""
+    original_shape = logits.shape[:-1]
+    num_classes = logits.shape[-1]
+    flat_logits = logits.reshape(-1, num_classes)
+    entropies = []
+    for chunk in flat_logits.split(chunk_size, dim=0):
+        logps = F.log_softmax(chunk, dim=-1)
+        chunk_entropy = -(torch.exp(logps) * logps).sum(-1)
+        entropies.append(chunk_entropy)
+    return torch.cat(entropies, dim=0).reshape(original_shape)
+
+
+def _clean_gpu():
+    gc.collect()
+    torch.cuda.empty_cache()
+    torch.cuda.reset_peak_memory_stats()
+    torch.cuda.reset_accumulated_memory_stats()
+    torch.cuda.synchronize()
+
+
+def profile_time(fn, logits, n_iters=BENCH_ITERS):
+    for _ in range(WARMUP):
+        out = fn(logits, chunk_size=128)
+        del out
+    torch.cuda.synchronize()
+
+    times = []
+    for _ in range(n_iters):
+        s = torch.cuda.Event(enable_timing=True)
+        e = torch.cuda.Event(enable_timing=True)
+        s.record()
+        out = fn(logits, chunk_size=128)
+        e.record()
+        torch.cuda.synchronize()
+        times.append(s.elapsed_time(e))
+        del out
+    return times
+
+
+def profile_memory(fn, logits, n_iters=MEM_ITERS):
+    for _ in range(WARMUP):
+        out = fn(logits, chunk_size=128)
+        del out
+    torch.cuda.synchronize()
+
+    peaks = []
+    for _ in range(n_iters):
+        _clean_gpu()
+        base = torch.cuda.max_memory_allocated()
+        out = fn(logits, chunk_size=128)
+        torch.cuda.synchronize()
+        peaks.append(torch.cuda.max_memory_allocated() - base)
+        del out
+    return [p / 1e6 for p in peaks]
+
+
+def fmt(values, unit=""):
+    mean = statistics.mean(values)
+    std = statistics.stdev(values) if len(values) > 1 else 0.0
+    return f"{mean:8.2f} ± {std:5.2f} {unit}  [min={min(values):.2f}, max={max(values):.2f}]"
+
+
+def benchmark_contiguous():
+    print("=" * 60)
+    print(
+        f"CONTIGUOUS BENCHMARK  (warmup={WARMUP}, time={BENCH_ITERS}, mem={MEM_ITERS})"
+    )
+    print("=" * 60)
+
+    configs = [
+        (1, 2048),
+        (1, 8192),
+        (1, 16384),
+        (4, 4096),
+        (8, 2048),
+        (16, 2048),
+        (16, 4096),
+    ]
+
+    for B, L in configs:
+        mem_gb = B * L * V * 2 / 1e9
+        if mem_gb > 28:
+            print(f"\n  skip B={B}, L={L} ({mem_gb:.1f} GB)")
+            continue
+
+        N = B * L
+        print(f"\n{'─' * 60}")
+        print(f"B={B:2d}, L={L:5d}  ({N:6d} rows, logits {mem_gb:.2f} GB)")
+        print(f"{'─' * 60}")
+
+        torch.manual_seed(42)
+        logits = torch.randn(B, L, V, device="cuda", dtype=torch.bfloat16)
+
+        t_orig = profile_time(entropy_from_logits_original, logits)
+        t_triton = profile_time(entropy_from_logits, logits)
+        orig_mean = statistics.mean(t_orig)
+        triton_mean = statistics.mean(t_triton)
+
+        print("  TIME (ms):")
+        print(f"    original: {fmt(t_orig, 'ms')}")
+        print(f"    triton:   {fmt(t_triton, 'ms')}")
+        print(f"    speedup:  {orig_mean / triton_mean:.2f}x")
+
+        m_orig = profile_memory(entropy_from_logits_original, logits)
+        m_triton = profile_memory(entropy_from_logits, logits)
+        orig_peak = statistics.mean(m_orig)
+        triton_peak = statistics.mean(m_triton)
+
+        print("  MEMORY (peak overhead):")
+        print(f"    original: {fmt(m_orig, 'MB')}")
+        print(f"    triton:   {fmt(m_triton, 'MB')}")
+        print(f"    saved:    {orig_peak - triton_peak:.1f} MB")
+
+        del logits
+        _clean_gpu()
+
+
+def benchmark_noncontiguous():
+    print("\n" + "=" * 60)
+    print(
+        f"NON-CONTIGUOUS BENCHMARK  (warmup={WARMUP}, time={BENCH_ITERS}, mem={MEM_ITERS})"
+    )
+    print("=" * 60)
+
+    configs = [
+        (4, 2048, "transpose"),
+        (4, 8192, "transpose"),
+        (8, 2048, "transpose"),
+        (4, 4096, "slice_batch"),
+    ]
+
+    for B, L, method in configs:
+        torch.manual_seed(42)
+
+        if method == "transpose":
+            raw = torch.randn(L, B, V, device="cuda", dtype=torch.bfloat16)
+            logits_nc = raw.transpose(0, 1)
+            raw_gb = L * B * V * 2 / 1e9
+        elif method == "slice_batch":
+            raw = torch.randn(B * 2, L, V, device="cuda", dtype=torch.bfloat16)
+            logits_nc = raw[::2]
+            raw_gb = B * 2 * L * V * 2 / 1e9
+        else:
+            continue
+
+        if raw_gb > 28:
+            print(f"\n  skip B={B}, L={L}, {method} ({raw_gb:.1f} GB)")
+            del raw, logits_nc
+            torch.cuda.empty_cache()
+            continue
+
+        N = B * L
+        print(f"\n{'─' * 60}")
+        print(f"B={B}, L={L}  {method}  ({N} rows, raw {raw_gb:.2f} GB)")
+        print(f"{'─' * 60}")
+
+        def original_with_copy(logits, chunk_size=128):
+            return entropy_from_logits_original(
+                logits.contiguous(), chunk_size=chunk_size
+            )
+
+        t_orig = profile_time(original_with_copy, logits_nc)
+        t_triton = profile_time(entropy_from_logits, logits_nc)
+        orig_mean = statistics.mean(t_orig)
+        triton_mean = statistics.mean(t_triton)
+
+        print("  TIME (ms):")
+        print(f"    orig+copy:     {fmt(t_orig, 'ms')}")
+        print(f"    triton-strided:{fmt(t_triton, 'ms')}")
+        print(f"    speedup:       {orig_mean / triton_mean:.2f}x")
+
+        m_orig = profile_memory(original_with_copy, logits_nc)
+        m_triton = profile_memory(entropy_from_logits, logits_nc)
+        orig_peak = statistics.mean(m_orig)
+        triton_peak = statistics.mean(m_triton)
+
+        print("  MEMORY (peak overhead):")
+        print(f"    orig+copy:     {fmt(m_orig, 'MB')}")
+        print(f"    triton-strided:{fmt(m_triton, 'MB')}")
+        print(f"    saved:         {orig_peak - triton_peak:.1f} MB")
+
+        del raw, logits_nc
+        _clean_gpu()
+
+
+if __name__ == "__main__":
+    benchmark_contiguous()
+    benchmark_noncontiguous()

benchmarks/bench_scattermoe_lora.py

@@ -0,0 +1,284 @@
+"""Benchmark for ScatterMoE LoRA Triton kernels.
+
+Measures forward, backward dX, and backward dA/dB kernels at common MoE
+model shapes. Reports per-kernel timings, LoRA overhead vs base scatter2scatter,
+and full fwd+bwd autograd throughput.
+
+Usage:
+  CUDA_VISIBLE_DEVICES=0 python benchmarks/bench_scattermoe_lora.py
+  CUDA_VISIBLE_DEVICES=0 python benchmarks/bench_scattermoe_lora.py --ranks 16 64
+  CUDA_VISIBLE_DEVICES=0 python benchmarks/bench_scattermoe_lora.py --models Qwen/Qwen3.5-35B-A3B
+"""
+
+import argparse
+import gc
+import time
+from functools import partial
+
+import torch
+
+from axolotl.integrations.kernels.libs.scattermoe_lora.kernels import (
+    lora_ops,
+    ops as base_ops,
+)
+from axolotl.integrations.kernels.libs.scattermoe_lora.parallel_experts import (
+    flatten_sort_count,
+)
+from axolotl.integrations.kernels.libs.scattermoe_lora.parallel_linear_lora import (
+    ScatterMoELoRA,
+)
+
+DEVICE = "cuda"
+DTYPE = torch.bfloat16
+WARMUP = 5
+ITERS = 20
+
+# ─── Model configs ──────────────────────────────────────────────────────────
+
+BUILTIN_CONFIGS = {
+    "Qwen3.5-35B-A3B": (256, 2048, 512, 8),  # E, H, I, k
+    "Qwen3-30B-A3B": (128, 2048, 768, 8),
+    "OLMoE-1B-7B": (64, 2048, 1024, 8),
+    "Mixtral-8x7B": (8, 4096, 14336, 2),
+}
+
+
+def _resolve_config(spec):
+    """Resolve a model spec to (E, H, I, k). Accepts builtin names or HF IDs."""
+    key = spec.lower().replace("/", "-")
+    for name, cfg in BUILTIN_CONFIGS.items():
+        if key in name.lower() or name.lower() in key:
+            return name, cfg
+
+    from transformers import AutoConfig
+
+    hf_cfg = AutoConfig.from_pretrained(spec, trust_remote_code=True)
+    if callable(getattr(hf_cfg, "get_text_config", None)):
+        tc = hf_cfg.get_text_config()
+        if hasattr(tc, "model_type") and tc.model_type != hf_cfg.model_type:
+            hf_cfg = tc
+    hidden = hf_cfg.hidden_size
+    inter = getattr(hf_cfg, "moe_intermediate_size", None) or hf_cfg.intermediate_size
+    experts = (
+        getattr(hf_cfg, "num_experts", None)
+        or getattr(hf_cfg, "num_local_experts", None)
+        or getattr(hf_cfg, "n_routed_experts", None)
+    )
+    top_k = (
+        getattr(hf_cfg, "num_experts_per_tok", None)
+        or getattr(hf_cfg, "num_experts_per_token", None)
+        or 2
+    )
+    name = spec.split("/")[-1]
+    return name, (experts, hidden, inter, top_k)
+
+
+# ─── Benchmark helpers ──────────────────────────────────────────────────────
+
+
+def _clean():
+    gc.collect()
+    torch.cuda.empty_cache()
+    torch.cuda.synchronize()
+
+
+def _bench(fn, warmup=WARMUP, iters=ITERS):
+    for _ in range(warmup):
+        fn()
+    torch.cuda.synchronize()
+    times = []
+    for _ in range(iters):
+        torch.cuda.synchronize()
+        t0 = time.perf_counter()
+        fn()
+        torch.cuda.synchronize()
+        times.append((time.perf_counter() - t0) * 1000)
+    times.sort()
+    return times[len(times) // 2]
+
+
+def _setup(num_experts, K, N, T, top_k, R):
+    torch.manual_seed(42)
+    x = torch.randn(T, K, device=DEVICE, dtype=DTYPE)
+    W = torch.randn(num_experts, K, N, device=DEVICE, dtype=DTYPE) * 0.02
+    lora_A = torch.randn(R * num_experts, K, device=DEVICE, dtype=DTYPE) * 0.01
+    lora_B = torch.randn(N, R * num_experts, device=DEVICE, dtype=DTYPE) * 0.01
+    logits = torch.randn(T, num_experts, device=DEVICE)
+    _, top_idx = torch.topk(torch.softmax(logits, dim=-1), top_k, dim=-1)
+    sei, ssi, eo = flatten_sort_count(top_idx, num_experts)
+    gx = base_ops.group(x, ssi, fan_out=top_k)
+    dy = torch.randn(gx.size(0), N, device=DEVICE, dtype=DTYPE)
+    return x, W, lora_A, lora_B, sei, ssi, eo, gx, dy
+
+
+# ─── Kernel wrappers (avoid B023 loop-variable capture) ──────────────────────
+
+
+def _call_fwd(x, W, sei, ssi, top_k, lA, lB):
+    return lora_ops.scatter2scatter_lora(
+        X=x,
+        W=W,
+        sorted_expert_idxs=sei,
+        sorted_scattered_idxs=ssi,
+        k=top_k,
+        lora_A=lA,
+        lora_B=lB,
+        scaling=2.0,
+    )
+
+
+def _call_base(x, W, sei, ssi, top_k):
+    return base_ops.scatter2scatter(
+        X=x,
+        W=W,
+        sorted_expert_idxs=sei,
+        sorted_scattered_idxs=ssi,
+        k=top_k,
+    )
+
+
+def _call_dx(dy, W, sei, ssi, lA, lB):
+    return lora_ops.scatter2scatter_lora_dX(
+        DY=dy,
+        W=W,
+        sorted_expert_idxs=sei,
+        sorted_scattered_idxs=ssi,
+        k=1,
+        lora_A=lA,
+        lora_B=lB,
+        scaling=2.0,
+        dy_grouped=True,
+        dx_grouped=False,
+    )
+
+
+def _call_bwd(dy, gx, lA, lB, eo, num_experts):
+    return lora_ops.group_bwd_lora(
+        DY=dy,
+        X=gx,
+        lora_A=lA,
+        lora_B=lB,
+        expert_offsets=eo,
+        E=num_experts,
+        scaling=2.0,
+    )
+
+
+# ─── Main ────────────────────────────────────────────────────────────────────
+
+
+def main():
+    parser = argparse.ArgumentParser(description="ScatterMoE LoRA kernel benchmark")
+    parser.add_argument(
+        "--models",
+        "-m",
+        nargs="+",
+        help="Model names or HF IDs (default: all builtins)",
+    )
+    parser.add_argument("--ranks", "-r", nargs="+", type=int, default=[16, 32, 64])
+    parser.add_argument("--seq-len", "-T", type=int, default=2048)
+    args = parser.parse_args()
+
+    T = args.seq_len
+    print(f"GPU: {torch.cuda.get_device_name()}")
+    print(f"T={T}, ranks={args.ranks}\n")
+
+    if args.models:
+        configs = [_resolve_config(m) for m in args.models]
+    else:
+        configs = list(BUILTIN_CONFIGS.items())
+
+    for model_name, (num_experts, hidden, inter, top_k) in configs:
+        print(f"{'=' * 70}")
+        print(f"  {model_name}: E={num_experts}, H={hidden}, I={inter}, k={top_k}")
+        print(f"{'=' * 70}")
+
+        for R in args.ranks:
+            for proj, K, N in [("gate_up", hidden, 2 * inter), ("down", inter, hidden)]:
+                _clean()
+                x, W, lA, lB, sei, ssi, eo, gx, dy = _setup(
+                    num_experts, K, N, T, top_k, R
+                )
+
+                # Forward with LoRA (auto-dispatched: fused or split)
+                dispatch = (
+                    "split"
+                    if (
+                        num_experts <= lora_ops._SPLIT_LORA_FWD_MAX_EXPERTS
+                        and K * N >= lora_ops._SPLIT_LORA_FWD_THRESHOLD
+                    )
+                    else "fused"
+                )
+                t_fwd = _bench(partial(_call_fwd, x, W, sei, ssi, top_k, lA, lB))
+                t_base = _bench(partial(_call_base, x, W, sei, ssi, top_k))
+                t_dx = _bench(partial(_call_dx, dy, W, sei, ssi, lA, lB))
+                t_bwd = _bench(partial(_call_bwd, dy, gx, lA, lB, eo, num_experts))
+
+                total = t_fwd + t_dx + t_bwd
+                overhead = t_fwd / t_base - 1 if t_base > 0 else 0
+
+                print(
+                    f"  R={R:>2} {proj:<8}  "
+                    f"fwd={t_fwd:>6.2f}ms [{dispatch}]  "
+                    f"base={t_base:>6.2f}ms "
+                    f"(+{overhead * 100:.0f}%)  "
+                    f"dx={t_dx:>6.2f}ms  bwd={t_bwd:>6.2f}ms  "
+                    f"total={total:>6.2f}ms"
+                )
+
+                # Full autograd fwd+bwd with memory measurement
+                x_ag = x.clone().requires_grad_(True)
+                lA_ag = lA.clone().requires_grad_(True)
+                lB_ag = lB.clone().requires_grad_(True)
+
+                def _run_autograd(
+                    _x=x_ag,
+                    _W=W,
+                    _k=top_k,
+                    _sei=sei,
+                    _ssi=ssi,
+                    _eo=eo,
+                    _lA=lA_ag,
+                    _lB=lB_ag,
+                ):
+                    out = ScatterMoELoRA.apply(
+                        _x,
+                        _W,
+                        _k,
+                        _sei,
+                        _ssi,
+                        _eo,
+                        _lA,
+                        _lB,
+                        2.0,
+                        None,
+                        None,
+                        False,
+                        False,
+                        True,
+                        False,
+                    )
+                    out.sum().backward()
+                    _x.grad = None
+                    _lA.grad = None
+                    _lB.grad = None
+
+                t_full = _bench(_run_autograd)
+
+                _clean()
+                torch.cuda.reset_peak_memory_stats()
+                mem_before = torch.cuda.memory_allocated()
+                _run_autograd()
+                torch.cuda.synchronize()
+                mem_peak = torch.cuda.max_memory_allocated() - mem_before
+
+                print(
+                    f"         full_fwd_bwd={t_full:>6.2f}ms  "
+                    f"peak_delta={mem_peak / 1e6:>6.1f}MB"
+                )
+
+        print()
+
+
+if __name__ == "__main__":
+    main()

benchmarks/bench_selective_logsoftmax.py

@@ -0,0 +1,191 @@
+"""Benchmark for selective_log_softmax Triton kernel vs original implementation.
+
+Usage: CUDA_VISIBLE_DEVICES=0 python benchmarks/bench_selective_logsoftmax.py
+"""
+
+import gc
+import statistics
+
+import torch
+
+from axolotl.monkeypatch.trainer.utils import (
+    selective_log_softmax,
+    selective_log_softmax_original,
+)
+
+V = 151936  # Qwen vocab
+WARMUP = 5
+BENCH_ITERS = 20
+MEM_ITERS = 10
+
+
+def _clean_gpu():
+    gc.collect()
+    torch.cuda.empty_cache()
+    torch.cuda.reset_peak_memory_stats()
+    torch.cuda.reset_accumulated_memory_stats()
+    torch.cuda.synchronize()
+
+
+def profile_time(fn, args, n_iters=BENCH_ITERS):
+    for _ in range(WARMUP):
+        fn(*args)
+    torch.cuda.synchronize()
+
+    times = []
+    for _ in range(n_iters):
+        s = torch.cuda.Event(enable_timing=True)
+        e = torch.cuda.Event(enable_timing=True)
+        s.record()
+        fn(*args)
+        e.record()
+        torch.cuda.synchronize()
+        times.append(s.elapsed_time(e))
+    return times
+
+
+def profile_memory(fn, args, n_iters=MEM_ITERS):
+    for _ in range(WARMUP):
+        out = fn(*args)
+        del out
+    torch.cuda.synchronize()
+
+    peaks = []
+    for _ in range(n_iters):
+        _clean_gpu()
+        base = torch.cuda.max_memory_allocated()
+        out = fn(*args)
+        torch.cuda.synchronize()
+        peaks.append(torch.cuda.max_memory_allocated() - base)
+        del out
+    return [p / 1e6 for p in peaks]
+
+
+def fmt(values, unit=""):
+    mean = statistics.mean(values)
+    std = statistics.stdev(values) if len(values) > 1 else 0.0
+    return f"{mean:8.2f} ± {std:5.2f} {unit}  [min={min(values):.2f}, max={max(values):.2f}]"
+
+
+def benchmark_forward():
+    print("=" * 60)
+    print(f"FORWARD BENCHMARK  (warmup={WARMUP}, time={BENCH_ITERS}, mem={MEM_ITERS})")
+    print("=" * 60)
+
+    configs = [
+        (1, 2048),
+        (1, 8192),
+        (4, 4096),
+        (8, 2048),
+        (16, 2048),
+        (16, 4096),
+    ]
+
+    for B, L in configs:
+        mem_gb = B * L * V * 2 / 1e9
+        if mem_gb > 28:
+            print(f"\n  skip B={B}, L={L} ({mem_gb:.1f} GB)")
+            continue
+
+        N = B * L
+        print(f"\n{'─' * 60}")
+        print(f"B={B:2d}, L={L:5d}  ({N:6d} rows, logits {mem_gb:.2f} GB)")
+        print(f"{'─' * 60}")
+
+        torch.manual_seed(42)
+        logits = torch.randn(B, L, V, device="cuda", dtype=torch.bfloat16)
+        index = torch.randint(0, V, (B, L), device="cuda")
+
+        t_orig = profile_time(selective_log_softmax_original, (logits, index))
+        t_triton = profile_time(selective_log_softmax, (logits, index))
+        orig_mean = statistics.mean(t_orig)
+        triton_mean = statistics.mean(t_triton)
+
+        print("  TIME (ms):")
+        print(f"    original: {fmt(t_orig, 'ms')}")
+        print(f"    triton:   {fmt(t_triton, 'ms')}")
+        print(f"    speedup:  {orig_mean / triton_mean:.2f}x")
+
+        m_orig = profile_memory(selective_log_softmax_original, (logits, index))
+        m_triton = profile_memory(selective_log_softmax, (logits, index))
+        orig_peak = statistics.mean(m_orig)
+        triton_peak = statistics.mean(m_triton)
+
+        print("  MEMORY (peak overhead):")
+        print(f"    original: {fmt(m_orig, 'MB')}")
+        print(f"    triton:   {fmt(m_triton, 'MB')}")
+        print(f"    saved:    {orig_peak - triton_peak:.1f} MB")
+
+        del logits, index
+        _clean_gpu()
+
+
+def benchmark_backward():
+    print("\n" + "=" * 60)
+    print(f"FWD+BWD BENCHMARK  (warmup={WARMUP}, time={BENCH_ITERS}, mem={MEM_ITERS})")
+    print("=" * 60)
+
+    configs = [
+        (1, 2048),
+        (1, 8192),
+        (4, 4096),
+        (8, 2048),
+        (16, 2048),
+        (16, 4096),
+    ]
+
+    def fwd_bwd_original(logits, index):
+        logits.grad = None
+        out = selective_log_softmax_original(logits, index)
+        out.sum().backward()
+
+    def fwd_bwd_triton(logits, index):
+        logits.grad = None
+        out = selective_log_softmax(logits, index)
+        out.sum().backward()
+
+    for B, L in configs:
+        mem_gb = B * L * V * 2 / 1e9
+        if mem_gb > 20:
+            print(f"\n  skip B={B}, L={L} ({mem_gb:.1f} GB, need room for grads)")
+            continue
+
+        N = B * L
+        print(f"\n{'─' * 60}")
+        print(f"B={B:2d}, L={L:5d}  ({N:6d} rows, logits {mem_gb:.2f} GB)")
+        print(f"{'─' * 60}")
+
+        torch.manual_seed(42)
+        logits_orig = torch.randn(
+            B, L, V, device="cuda", dtype=torch.bfloat16, requires_grad=True
+        )
+        logits_tri = logits_orig.detach().clone().requires_grad_(True)
+        index = torch.randint(0, V, (B, L), device="cuda")
+
+        t_orig = profile_time(fwd_bwd_original, (logits_orig, index))
+        t_triton = profile_time(fwd_bwd_triton, (logits_tri, index))
+        orig_mean = statistics.mean(t_orig)
+        triton_mean = statistics.mean(t_triton)
+
+        print("  FWD+BWD TIME (ms):")
+        print(f"    original: {fmt(t_orig, 'ms')}")
+        print(f"    triton:   {fmt(t_triton, 'ms')}")
+        print(f"    speedup:  {orig_mean / triton_mean:.2f}x")
+
+        m_orig = profile_memory(fwd_bwd_original, (logits_orig, index))
+        m_triton = profile_memory(fwd_bwd_triton, (logits_tri, index))
+        orig_peak = statistics.mean(m_orig)
+        triton_peak = statistics.mean(m_triton)
+
+        print("  FWD+BWD MEMORY (peak overhead):")
+        print(f"    original: {fmt(m_orig, 'MB')}")
+        print(f"    triton:   {fmt(m_triton, 'MB')}")
+        print(f"    saved:    {orig_peak - triton_peak:.1f} MB")
+
+        del logits_orig, logits_tri, index
+        _clean_gpu()
+
+
+if __name__ == "__main__":
+    benchmark_forward()
+    benchmark_backward()

cicd/Dockerfile-uv.jinja

@@ -11,7 +11,7 @@ ENV NIGHTLY_BUILD="{{ NIGHTLY_BUILD }}"
 ENV HF_HOME="{{ HF_HOME }}"
 
 RUN apt-get update && \
-    apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev ibverbs-providers ibverbs-utils infiniband-diags librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm
+    apt-get install -y --allow-change-held-packages vim curl nano zstd libnccl2 libnccl-dev ibverbs-providers ibverbs-utils infiniband-diags librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm
 
 WORKDIR /workspace
 
@@ -31,7 +31,9 @@ RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
         sed -i 's#^datasets.*#datasets @ git+https://github.com/huggingface/datasets.git@main#' requirements.txt; \
     fi
 
-RUN uv pip install packaging==23.2 setuptools==75.8.0
+RUN uv pip install packaging==26.0 setuptools==78.1.1
+RUN uv pip install torchvision
+RUN uv pip uninstall causal_conv1d
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
         uv pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \

cicd/Dockerfile.jinja

@@ -1,6 +1,6 @@
 FROM axolotlai/axolotl-base:{{ BASE_TAG }}
 
-ENV TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX"
+ENV TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
 ENV AXOLOTL_EXTRAS="{{ AXOLOTL_EXTRAS }}"
 ENV AXOLOTL_ARGS="{{ AXOLOTL_ARGS }}"
 ENV CUDA="{{ CUDA }}"
@@ -9,10 +9,10 @@ ENV GITHUB_REF="{{ GITHUB_REF }}"
 ENV GITHUB_SHA="{{ GITHUB_SHA }}"
 ENV NIGHTLY_BUILD="{{ NIGHTLY_BUILD }}"
 ENV HF_HOME="{{ HF_HOME }}"
-ENV AXOLOTL_DATASET_PROCESSES="8"
+ENV AXOLOTL_DATASET_NUM_PROC="8"
 
 RUN apt-get update && \
-    apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev ibverbs-providers ibverbs-utils infiniband-diags librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm
+    apt-get install -y --allow-change-held-packages vim curl nano zstd libnccl2 libnccl-dev ibverbs-providers ibverbs-utils infiniband-diags librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm
 
 WORKDIR /workspace
 
@@ -32,7 +32,8 @@ RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
         sed -i 's#^datasets.*#datasets @ git+https://github.com/huggingface/datasets.git@main#' requirements.txt; \
     fi
 
-RUN pip install packaging==23.2 setuptools==75.8.0
+RUN pip install packaging==26.0 setuptools==78.1.1 psutil
+RUN pip uninstall -y causal_conv1d
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
         pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \

cicd/cicd.sh

@@ -3,6 +3,14 @@ set -e
 
 python -c "import torch; assert '$PYTORCH_VERSION' in torch.__version__"
 
+set -o pipefail
+curl --silent --show-error --fail --retry 3 --retry-delay 5 -L https://axolotl-ci.b-cdn.net/hf-cache.tar.zst | tar -xpf - -C "${HF_HOME}/hub/" --use-compress-program unzstd --strip-components=1
+# hf download "NousResearch/Meta-Llama-3-8B"
+# hf download "NousResearch/Meta-Llama-3-8B-Instruct"
+# hf download "microsoft/Phi-4-reasoning"
+# hf download "microsoft/Phi-3.5-mini-instruct"
+# hf download "microsoft/Phi-3-medium-128k-instruct"
+
 # Run unit tests with initial coverage report
 pytest -v --durations=10 -n8 \
   --ignore=tests/e2e/ \

cicd/multigpu.py

@@ -2,8 +2,6 @@
 modal application to run axolotl gpu tests in Modal
 """
 
-# pylint: disable=duplicate-code
-
 import os
 import pathlib
 import tempfile
@@ -19,7 +17,8 @@ template_loader = jinja2.FileSystemLoader(searchpath=cicd_path)
 template_env = jinja2.Environment(
     loader=template_loader, autoescape=select_autoescape()
 )
-df_template = template_env.get_template("Dockerfile.jinja")
+dockerfile = os.environ.get("E2E_DOCKERFILE", "Dockerfile.jinja")
+df_template = template_env.get_template(dockerfile)
 
 df_args = {
     "AXOLOTL_EXTRAS": os.environ.get("AXOLOTL_EXTRAS", ""),
@@ -29,8 +28,11 @@ df_args = {
     "CUDA": os.environ.get("CUDA", "126"),
     "GITHUB_REF": os.environ.get("GITHUB_REF", "refs/heads/main"),
     "GITHUB_SHA": os.environ.get("GITHUB_SHA", ""),
+    "NIGHTLY_BUILD": os.environ.get("NIGHTLY_BUILD", ""),
     "CODECOV_TOKEN": os.environ.get("CODECOV_TOKEN", ""),
     "HF_HOME": "/workspace/data/huggingface-cache/hub",
+    "PYTHONUNBUFFERED": os.environ.get("PYTHONUNBUFFERED", "1"),
+    "DEEPSPEED_LOG_LEVEL": os.environ.get("DEEPSPEED_LOG_LEVEL", "WARNING"),
 }
 
 dockerfile_contents = df_template.render(**df_args)
@@ -63,7 +65,7 @@ def run_cmd(cmd: str, run_folder: str):
 
     # Propagate errors from subprocess.
     if exit_code := subprocess.call(cmd.split(), cwd=run_folder):  # nosec
-        exit(exit_code)  # pylint: disable=consider-using-sys-exit
+        exit(exit_code)
 
 
 @app.function(

cicd/multigpu.sh

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

cicd/single_gpu.py

@@ -1,7 +1,5 @@
 """Modal app to run axolotl GPU tests"""
 
-# pylint: disable=duplicate-code
-
 import os
 import pathlib
 import tempfile
@@ -59,15 +57,17 @@ VOLUME_CONFIG = {
 }
 
 N_GPUS = int(os.environ.get("N_GPUS", 1))
-GPU_CONFIG = f"L40S:{N_GPUS}"
+GPU_TYPE = os.environ.get("GPU_TYPE", "L40S")
+GPU_CONFIG = f"{GPU_TYPE}:{N_GPUS}"
 
 
 def run_cmd(cmd: str, run_folder: str):
     import subprocess  # nosec
 
     sp_env = os.environ.copy()
-    sp_env["AXOLOTL_DATASET_PROCESSES"] = "8"
+    sp_env["AXOLOTL_DATASET_NUM_PROC"] = "8"
 
     # Propagate errors from subprocess.
-    if exit_code := subprocess.call(cmd.split(), cwd=run_folder, env=sp_env):  # nosec
-        exit(exit_code)  # pylint: disable=consider-using-sys-exit
+    exit_code = subprocess.call(cmd.split(), cwd=run_folder, env=sp_env)  # nosec
+    if exit_code:
+        raise RuntimeError(f"Command '{cmd}' failed with exit code {exit_code}")

codecov.yml

@@ -12,7 +12,7 @@ coverage:
       default:
         # basic
         target: auto
-        threshold: 0%
+        threshold: 1%
         base: auto
         # advanced
         branches: null
@@ -27,7 +27,7 @@ coverage:
       default:
         # basic
         target: auto
-        threshold: 0%
+        threshold: 1%
         base: auto
         # advanced
         branches: null
@@ -37,6 +37,7 @@ coverage:
         only_pulls: false
         flags: null
         paths: null
+        informational: true
 
 parsers:
   gcov:

devtools/dev_chat_template.yml

@@ -13,7 +13,7 @@ datasets:
 val_set_size: 0
 output_dir: temp_debug/axolotl_outputs/model
 dataset_prepared_path: temp_debug/axolotl_outputs/data
-dataset_processes: 1
+dataset_num_proc: 1
 
 sequence_len: 4096
 sample_packing: false

docker/Dockerfile

@@ -6,6 +6,7 @@ ARG AXOLOTL_EXTRAS=""
 ARG AXOLOTL_ARGS=""
 ARG CUDA="118"
 ARG PYTORCH_VERSION="2.1.2"
+ARG TARGETARCH
 
 ENV PYTORCH_VERSION=$PYTORCH_VERSION
 
@@ -20,13 +21,18 @@ RUN git clone --depth=1 https://github.com/axolotl-ai-cloud/axolotl.git
 
 WORKDIR /workspace/axolotl
 
-# If AXOLOTL_EXTRAS is set, append it in brackets
-RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
+# If AXOLOTL_EXTRAS is set, append it in brackets; don't install deepspeed with arm64
+RUN pip uninstall -y causal_conv1d
+RUN if [ "$TARGETARCH" = "arm64" ]; then \
+        BASE_EXTRAS="flash-attn,ring-flash-attn,optimizers,ray"; \
     else \
-        pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray] $AXOLOTL_ARGS; \
+        BASE_EXTRAS="deepspeed,flash-attn,ring-flash-attn,optimizers,ray"; \
     fi && \
-    python scripts/unsloth_install.py | sh && \
+    if [ "$AXOLOTL_EXTRAS" != "" ]; then \
+        pip install --no-build-isolation -e .[$BASE_EXTRAS,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
+    else \
+        pip install --no-build-isolation -e .[$BASE_EXTRAS] $AXOLOTL_ARGS; \
+    fi && \    python scripts/unsloth_install.py | sh && \
     python scripts/cutcrossentropy_install.py | sh && \
     pip install pytest && \
     pip cache purge

docker/Dockerfile-base

@@ -2,14 +2,16 @@ ARG CUDA_VERSION="11.8.0"
 ARG CUDNN_VERSION="8"
 ARG UBUNTU_VERSION="22.04"
 ARG MAX_JOBS=4
+ARG TARGETARCH
 
 FROM nvidia/cuda:$CUDA_VERSION-cudnn$CUDNN_VERSION-devel-ubuntu$UBUNTU_VERSION AS base-builder
 
 ENV PATH="/root/miniconda3/bin:${PATH}"
 
-ARG PYTHON_VERSION="3.10"
+ARG TARGETARCH
+ARG PYTHON_VERSION="3.11"
 ARG PYTORCH_VERSION="2.1.2"
-ARG CUDA="118"
+ARG CUDA="128"
 ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 9.0+PTX"
 
 ENV PYTHON_VERSION=$PYTHON_VERSION
@@ -22,11 +24,17 @@ RUN apt-get update \
         librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm \
     && rm -rf /var/cache/apt/archives \
     && rm -rf /var/lib/apt/lists/* \
-    && wget \
-    https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh \
+    && if [ "$TARGETARCH" = "amd64" ]; then \
+        MINICONDA_ARCH="x86_64"; \
+    elif [ "$TARGETARCH" = "arm64" ]; then \
+        MINICONDA_ARCH="aarch64"; \
+    else \
+        echo "Unsupported architecture: $TARGETARCH"; exit 1; \
+    fi \
+    && wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-${MINICONDA_ARCH}.sh \
     && mkdir /root/.conda \
-    && bash Miniconda3-latest-Linux-x86_64.sh -b \
-    && rm -f Miniconda3-latest-Linux-x86_64.sh \
+    && bash Miniconda3-latest-Linux-${MINICONDA_ARCH}.sh -b \
+    && rm -f Miniconda3-latest-Linux-${MINICONDA_ARCH}.sh \
     && conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/main \
     && conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/r \
     && conda create -n "py${PYTHON_VERSION}" python="${PYTHON_VERSION}"
@@ -35,18 +43,34 @@ ENV PATH="/root/miniconda3/envs/py${PYTHON_VERSION}/bin:${PATH}"
 
 WORKDIR /workspace
 
-RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==23.2 setuptools==75.8.0 wheel && \
+RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==26.0 setuptools==75.8.0 wheel psutil && \
     python3 -m pip install --no-cache-dir -U torch==${PYTORCH_VERSION}+cu${CUDA} torchvision --extra-index-url https://download.pytorch.org/whl/cu$CUDA && \
-    python3 -m pip 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" && \
     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 if [ "$PYTORCH_VERSION" = "2.6.0" ] && [ "$CUDA" = "124" ] ; then \
-        FLASH_ATTENTION_FORCE_BUILD="TRUE" pip3 install --no-build-isolation flash-attn==2.8.0.post2; \
-    fi
+# Map Python version (e.g., 3.12 -> cp312)
+RUN PYTHON_CP="cp$(echo $PYTHON_VERSION | tr -d '.')" && \
+    # Map PyTorch version (e.g., 2.9.1 -> torch2.9, 2.10.0 -> torch2.10)
+    TORCH_TAG="torch$(echo $PYTORCH_VERSION | grep -oP '^\d+\.\d+')" && \
+    # Map architecture
+    case "$TARGETARCH" in \
+        amd64) ARCH_TAG="x86_64" ;; \
+        arm64) ARCH_TAG="aarch64" ;; \
+        *) echo "Unsupported architecture: $TARGETARCH"; exit 1 ;; \
+    esac && \
+    WHL_VERSION="v0.7.16" && \
+    WHL_FILE="flash_attn-2.8.3+cu${CUDA}${TORCH_TAG}-${PYTHON_CP}-${PYTHON_CP}-linux_${ARCH_TAG}.whl" && \
+    wget -nv "https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/${WHL_VERSION}/${WHL_FILE}" && \
+    pip3 install --no-cache-dir "${WHL_FILE}" && \
+    rm "${WHL_FILE}"

docker/Dockerfile-base-nightly

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

docker/Dockerfile-cloud-uv

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

docker/Dockerfile-uv

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

docker/Dockerfile-uv-base

@@ -2,9 +2,11 @@ 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
 
+ARG TARGETARCH
 ARG PYTHON_VERSION="3.11"
 ARG PYTORCH_VERSION="2.6.0"
 ARG CUDA="126"
@@ -30,7 +32,26 @@ RUN uv venv --no-project --relocatable axolotl-venv
 ENV PATH="/workspace/axolotl-venv/bin:${PATH}"
 
 RUN uv pip install packaging setuptools wheel psutil \
-    && uv pip install torch==${PYTORCH_VERSION} \
-    && uv pip install --no-build-isolation "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main" \
-    && uv pip install "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main" \
+    && uv pip install torch==${PYTORCH_VERSION} torchvision \
     && uv pip install awscli pydantic
+
+RUN if [ "$TARGETARCH" = "amd64" ]; then \
+        uv pip install --no-build-isolation "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main"; \
+        uv pip install "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main"; \
+    fi
+
+# Map Python version (e.g., 3.12 -> cp312)
+RUN PYTHON_CP="cp$(echo $PYTHON_VERSION | tr -d '.')" && \
+    # Map PyTorch version (e.g., 2.9.1 -> torch2.9, 2.10.0 -> torch2.10)
+    TORCH_TAG="torch$(echo $PYTORCH_VERSION | grep -oP '^\d+\.\d+')" && \
+    # Map architecture
+    case "$TARGETARCH" in \
+        amd64) ARCH_TAG="x86_64" ;; \
+        arm64) ARCH_TAG="aarch64" ;; \
+        *) echo "Unsupported architecture: $TARGETARCH"; exit 1 ;; \
+    esac && \
+    WHL_VERSION="v0.7.16" && \
+    WHL_FILE="flash_attn-2.8.3+cu${CUDA}${TORCH_TAG}-${PYTHON_CP}-${PYTHON_CP}-linux_${ARCH_TAG}.whl" && \
+    wget -nv "https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/${WHL_VERSION}/${WHL_FILE}" && \
+    uv pip install --no-cache-dir "${WHL_FILE}" && \
+    rm "${WHL_FILE}"

docs/.gitignore

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

docs/amd_hpc.qmd

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

docs/attention.qmd

@@ -0,0 +1,178 @@
+---
+title: Attention
+description: Supported attention modules in Axolotl
+---
+
+## SDP Attention
+
+This is the default built-in attention in PyTorch.
+
+```yaml
+sdp_attention: true
+```
+
+For more details: [PyTorch docs](https://docs.pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html)
+
+## Flash Attention
+
+Axolotl supports Flash Attention 2, 3, and 4. The best available version is used automatically
+based on your installed packages and GPU.
+
+```yaml
+flash_attention: true
+```
+
+For more details: [Flash Attention](https://github.com/Dao-AILab/flash-attention/)
+
+### Flash Attention 2
+
+Requirements: Ampere, Ada, or Hopper GPUs (Turing or lower not supported)
+
+```bash
+pip install flash-attn --no-build-isolation
+```
+
+::: {.callout-tip}
+
+If you get `undefined symbol` while training, ensure you installed PyTorch prior to Axolotl.
+Alternatively, try reinstall or downgrade a version.
+
+:::
+
+### Flash Attention 3
+
+Requirements: Hopper only and CUDA 12.8 (recommended)
+
+```bash
+git clone https://github.com/Dao-AILab/flash-attention.git
+cd flash-attention/hopper
+
+python setup.py install
+```
+
+### Flash Attention 4
+
+Requirements: Hopper or Blackwell GPUs
+
+```bash
+pip install flash-attn-4
+```
+
+Or from source:
+
+```bash
+git clone https://github.com/Dao-AILab/flash-attention.git
+cd flash-attention/flash_attn/cute
+
+pip install -e .
+
+# FA2's flash_attn package includes a cute/ stub that shadows FA4.
+# Remove it so Python can find the real FA4 module:
+rm -r $(python -c "import flash_attn; print(flash_attn.__path__[0])")/cute
+```
+
+::: {.callout-note}
+
+**Hopper (SM90) users**: The backward kernel is not yet included in the pip package. To use FA4
+for training on Hopper, install from source using the instructions above.
+
+:::
+
+::: {.callout-warning}
+
+FA4 only supports head dimensions up to 128 (`d ≤ 128`). The DeepSeek shape `(192, 128)` is
+also supported but only on Blackwell. Axolotl automatically detects incompatible head dimensions
+and falls back to FA2/3.
+
+:::
+
+For more details: [flash-attention/flash_attn/cute](https://github.com/Dao-AILab/flash-attention/tree/main/flash_attn/cute)
+
+### AMD
+
+Requirements: ROCm 6.0 and above.
+
+See [Flash Attention AMD docs](https://github.com/Dao-AILab/flash-attention/tree/main?tab=readme-ov-file#amd-rocm-support).
+
+## Flex Attention
+
+A flexible PyTorch API for attention used in combination with `torch.compile`.
+
+```yaml
+flex_attention: true
+
+# recommended
+torch_compile: true
+```
+
+::: {.callout-note}
+
+We recommend using latest stable version of PyTorch for best performance.
+
+:::
+
+For more details: [PyTorch docs](https://pytorch.org/blog/flexattention/)
+
+## SageAttention
+
+Attention kernels with QK Int8 and PV FP16 accumulator.
+
+```yaml
+sage_attention: true
+```
+
+Requirements: Ampere, Ada, or Hopper GPUs
+
+```bash
+pip install sageattention==2.2.0 --no-build-isolation
+```
+
+::: {.callout-warning}
+
+Only LoRA/QLoRA recommended at the moment. We found loss drop to 0 for full finetuning. See [GitHub Issue](https://github.com/thu-ml/SageAttention/issues/198).
+
+:::
+
+For more details: [Sage Attention](https://github.com/thu-ml/SageAttention)
+
+::: {.callout-note}
+
+We do not support SageAttention 3 at the moment. If you are interested on adding this or improving SageAttention implementation, please make an Issue.
+
+:::
+
+
+## xFormers
+
+```yaml
+xformers_attention: true
+```
+
+::: {.callout-tip}
+
+We recommend using with Turing GPUs or below (such as on Colab).
+
+:::
+
+For more details: [xFormers](https://github.com/facebookresearch/xformers)
+
+## Shifted Sparse Attention
+
+::: {.callout-warning}
+
+We plan to deprecate this! If you use this feature, we recommend switching to methods above.
+
+:::
+
+Requirements: LLaMA model architecture
+
+```yaml
+flash_attention: true
+s2_attention: true
+```
+
+::: {.callout-tip}
+
+No sample packing support!
+
+:::

docs/checkpoint_saving.qmd

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

docs/cli.qmd

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

docs/dataset-formats/conversation.qmd

@@ -212,6 +212,21 @@ Instead of passing `tools` via the system prompt, an alternative method would be
 Tools need to follow [JSON schema](https://json-schema.org/learn/getting-started-step-by-step).
 :::
 
+::: {.callout-warning}
+If you have tool arguments with same name but different dtypes (like `"time": string` and `"time": number`), please save `arguments: ` as JSON string to prevent `datasets` from having casting issues.
+
+```
+"arguments": "{\"...\": \"...\"}"
+```
+
+The same is applicable for tool parameters.
+
+```
+"parameters": "{\"...\": \"...\"}"
+```
+
+:::
+
 Example config for Llama4:
 ```yaml
 chat_template: llama4

docs/dataset-formats/index.qmd

@@ -61,7 +61,7 @@ While we recommend `.jsonl`, you can also use the other formats (`csv`, `parquet
 
 ### Pre-training without streaming
 
-On the rare case that the dataset is small and can be loaded entirely into memory, another approach to running pre-training is to use the `completion` format. This would mean that the entire dataset is pre-tokenized instead of on-demand in streaming.
+In the case that the dataset is small and can be loaded entirely into memory, another approach to running pre-training is to use the `completion` format. This would mean that the entire dataset is pre-tokenized instead of on-demand in streaming.
 
 One benefit of this is that the tokenization can be performed separately on a CPU-only machine, and then transferred to a GPU machine for training to save costs.
 

docs/debugging.qmd

@@ -29,7 +29,7 @@ While debugging it's helpful to simplify your test scenario as much as possible.
 1. **Make sure you are using the latest version of axolotl**:  This project changes often and bugs get fixed fast.  Check your git branch and make sure you have pulled the latest changes from `main`.
 1. **Eliminate concurrency**: Restrict the number of processes to 1 for both training and data preprocessing:
     - Set `CUDA_VISIBLE_DEVICES` to a single GPU, ex: `export CUDA_VISIBLE_DEVICES=0`.
-    - Set `dataset_processes: 1` in your axolotl config or run the training command with `--dataset_processes=1`.
+    - Set `dataset_num_proc: 1` in your axolotl config or run the training command with `--dataset_num_proc=1`.
 2. **Use a small dataset**: Construct or use a small dataset from HF Hub. When using a small dataset, you will often have to make sure `sample_packing: False` and `eval_sample_packing: False` to avoid errors.  If you are in a pinch and don't have time to construct a small dataset but want to use from the HF Hub, you can shard the data (this will still tokenize the entire dataset, but will only use a fraction of the data for training.  For example, to shard the dataset into 20 pieces, add the following to your axolotl config):
 
     ```yaml
@@ -101,7 +101,7 @@ For example, to mimic the command `cd devtools && CUDA_VISIBLE_DEVICES=0 acceler
                 "-m", "axolotl.cli.train", "dev_chat_template.yml",
                 // The flags below simplify debugging by overriding the axolotl config
                 // with the debugging tips above.  Modify as needed.
-                "--dataset_processes=1",      // limits data preprocessing to one process
+                "--dataset_num_proc=1",      // limits data preprocessing to one process
                 "--max_steps=1",              // limits training to just one step
                 "--batch_size=1",             // minimizes batch size
                 "--micro_batch_size=1",       // minimizes batch size

docs/docker.qmd

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

docs/expert_quantization.qmd

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

docs/faq.qmd

@@ -63,6 +63,14 @@ description: Frequently asked questions
 
 > A: There seems to be a wheel issue with FA2 2.8.0 on CUDA 12.4. Try CUDA 12.6 instead or downgrade to FA2 2.7.4. Please refer to the upstream issue: https://github.com/Dao-AILab/flash-attention/issues/1717.
 
+**Q: Can we mix text and text+image datasets for VLM training?**
+
+> A: Yes, you can for newer VLM arch. The ones that would not work are LLaVA / Pixtral arch. If you notice one not working, please let us know!
+
+**Q: Why is `memory/max_*` different from `nvidia-smi`?**
+
+> A: We use `torch` APIs to retrieve this information. You can see https://docs.pytorch.org/docs/stable/notes/cuda.html#cuda-memory-management for more information.
+
 ### Chat templates
 
 **Q: `jinja2.exceptions.UndefinedError: 'dict object' has no attribute 'content' / 'role' / ____`**
@@ -140,3 +148,7 @@ description: Frequently asked questions
 **Q: `ValueError("Backward pass should have cleared tracker of all tensors")`
 
 > A: This may happen due to edge cases in using the modern OffloadActivations context manager for CUDA streams. If you encounter this error, you may have success using the naive implementation with `offload_activations: legacy` in your YAML.
+
+**Q: `Error parsing tool_calls arguments as JSON.`
+
+> A: There is an error parsing string arguments to a dict. Please check your dataset and the error message for more details.

docs/fsdp_qlora.qmd

@@ -1,5 +1,5 @@
 ---
-title: "FDSP + QLoRA"
+title: "FSDP + QLoRA"
 description: Use FSDP with QLoRA to fine-tune large LLMs on consumer GPUs.
 format:
   html:
@@ -23,6 +23,12 @@ To enable `QLoRA` with `FSDP`, you need to perform the following steps:
 2. Enable FSDP in your axolotl config, as [described here](multi-gpu.qmd#sec-fsdp).
 3. Use one of the supported model types: `llama`, `mistral` or `mixtral`.
 
+## Enabling Swap for FSDP2
+
+If available memory is insufficient even after FSDP's CPU offloading, you can enable swap memory usage by setting `cpu_offload_pin_memory: false` alongside `offload_params: true` in FSDP config.
+
+This disables memory pinning, allowing FSDP to use disk swap space as fallback. Disabling memory pinning itself incurs performance overhead, and actually having to use swap adds more, but it may enable training larger models that would otherwise cause OOM errors on resource constrained systems.
+
 ## Example Config
 
 [examples/llama-2/qlora-fsdp.yml](../examples/llama-2/qlora-fsdp.yml) contains an example of how to enable QLoRA + FSDP in axolotl.

docs/gradient_checkpointing.qmd

@@ -1,5 +1,5 @@
 ---
-title: Gradient Checkpointing and Activation Offloading
+title: Gradient Checkpointing, Activation Offloading, and Layer Offloading
 ---
 
 Gradient checkpointing and activation offloading are techniques used to optimize the performance of deep learning
@@ -27,3 +27,33 @@ The `activation_offloading: legacy` naively offloads activations to CPU and with
 
 For resource constrained environments with limited CPU memory, `activation_offloading: disk` offloads
 activations to disk instead of CPU RAM so that much larger context lengths can be trained with minimal memory.
+
+### Enabling Layer Offloading
+
+```yaml
+layer_offloading: true
+```
+
+Layer offloading reduces GPU memory usage by moving frozen (non-trainable) decoder layer parameters to CPU
+and streaming them back to GPU one layer at a time during the forward and backward passes. This is
+particularly useful for LoRA/QLoRA training where most of the model's parameters are frozen — only the
+trainable adapter weights stay on GPU permanently.
+
+During training, forward and backward hooks on each decoder layer handle the transfer automatically:
+
+- **Forward pass:** Before a layer executes, its frozen params are loaded to GPU. The next layer is
+  prefetched asynchronously on a separate CUDA stream for overlap.
+- **Backward pass:** Same pattern in reverse — the current layer's frozen params are loaded and the
+  previous layer is prefetched.
+
+After each layer finishes, its frozen params are offloaded back to CPU pinned memory.
+
+This approach trades some CPU-GPU transfer overhead for significant GPU memory savings — the freed memory
+is roughly equal to the size of all frozen parameters across all decoder layers, minus one layer's worth
+that is kept on GPU at any given time.
+
+**Requirements:**
+
+- CUDA GPU (CPU-only training is not supported for this feature)
+- Works with any HuggingFace model architecture that uses decoder layers (Llama, Mistral, Qwen, etc.)
+- Best combined with LoRA/QLoRA where most parameters are frozen

docs/installation.qmd

@@ -26,7 +26,7 @@ Follow the instructions at: [https://pytorch.org/get-started/locally/](https://p
 :::
 
 ::: {.callout-important}
-For Blackwell GPUs, please use Pytorch 2.7.0 and CUDA 12.8.
+For Blackwell GPUs, please use Pytorch 2.9.1 and CUDA 12.8.
 :::
 
 ### PyPI Installation (Recommended) {#sec-pypi}
@@ -111,7 +111,7 @@ docker run --privileged --gpus '"all"' --shm-size 10g --rm -it \
 :::
 
 ::: {.callout-important}
-For Blackwell GPUs, please use `axolotlai/axolotl:main-py3.11-cu128-2.7.0` or the cloud variant `axolotlai/axolotl-cloud:main-py3.11-cu128-2.7.0`.
+For Blackwell GPUs, please use `axolotlai/axolotl:main-py3.11-cu128-2.9.1` or the cloud variant `axolotlai/axolotl-cloud:main-py3.11-cu128-2.9.1`.
 :::
 
 Please refer to the [Docker documentation](docker.qmd) for more information on the different Docker images that are available.
@@ -134,7 +134,7 @@ For providers supporting Docker:
 
 ### Google Colab {#sec-colab}
 
-Use our [example notebook](../examples/colab-notebooks/colab-axolotl-example.ipynb).
+[![](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/axolotl-ai-cloud/axolotl/blob/main/examples/colab-notebooks/colab-axolotl-example.ipynb#scrollTo=msOCO4NRmRLa)
 
 ## Platform-Specific Instructions {#sec-platform-specific}
 
@@ -165,7 +165,7 @@ We recommend using WSL2 (Windows Subsystem for Linux) or Docker.
    ```
 4. (Optional) Login to Hugging Face:
    ```{.bash}
-   huggingface-cli login
+   hf auth login
    ```
 
 ## Troubleshooting {#sec-troubleshooting}

docs/lora_optims.qmd

@@ -5,10 +5,11 @@ description: "Custom autograd functions and Triton kernels in Axolotl for optimi
 
 Inspired by [Unsloth](https://github.com/unslothai/unsloth), we've implemented two
 optimizations for LoRA and QLoRA fine-tuning, supporting both single GPU and multi-GPU
-(in the DDP and DeepSpeed settings) training. These include (1) SwiGLU and GEGLU activation function
-Triton kernels, and (2) LoRA MLP and attention custom autograd functions. Our goal was
-to leverage operator fusion and tensor re-use in order to improve speed and reduce
-memory usage during the forward and backward passes of these calculations.
+(including the DDP, DeepSpeed, and FSDP2 settings) training. These include (1) SwiGLU
+and GEGLU activation function Triton kernels, and (2) LoRA MLP and attention custom
+autograd functions. Our goal was to leverage operator fusion and tensor re-use in order
+to improve speed and reduce memory usage during the forward and backward passes of
+these calculations.
 
 We currently support several common model architectures, including (but not limited to):
 
@@ -88,6 +89,10 @@ lora_o_kernel: true
 Currently, LoRA kernels are not supported for RLHF training, only SFT.
 :::
 
+::: {.callout-warning}
+LoRA kernels do not support remote modeling code.
+:::
+
 ## Requirements
 
 - One or more NVIDIA or AMD GPUs (in order to use the Triton kernels)
@@ -131,6 +136,5 @@ computation path.
 ## Future Work
 
 - Support for additional model architectures
-- Support for the FSDP setting
 - Support for dropout and bias
 - Additional operator fusions

docs/lr_groups.qmd

@@ -27,3 +27,9 @@ learning_rate: 2e-5
 In this example, we have a default learning rate of 2e-5 across the entire model, but we have a separate learning rate
 of 1e-6 for all the self attention `o_proj` modules across all layers, and a learning are of 1e-5 to the 3rd layer's
 self attention `q_proj` module.
+
+::: {.callout-note}
+
+We currently only support varying `lr` for now. If you're interested in adding support for others (`weight_decay`), we welcome PRs. See https://github.com/axolotl-ai-cloud/axolotl/blob/613bcf90e58f3ab81d3827e7fc572319908db9fb/src/axolotl/core/trainers/mixins/optimizer.py#L17
+
+:::

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
@@ -14,12 +14,18 @@ This guide covers advanced training configurations for multi-GPU setups using Ax
 
 ## Overview {#sec-overview}
 
-Axolotl supports several methods for multi-GPU training:
+When training on multiple GPUs, Axolotl supports 3 sharding/parallelism strategies. Additionally, you can layer specific optimization features on top of that strategy.
 
-- DeepSpeed (recommended)
-- FSDP (Fully Sharded Data Parallel)
-- Sequence parallelism
-- FSDP + QLoRA
+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 {#sec-deepspeed}
 
@@ -63,23 +69,20 @@ Start from Stage 1 -> Stage 2 -> Stage 3.
 
 :::
 
-::: {.callout-tip}
-
-Using ZeRO Stage 3 with Single-GPU training
-
-ZeRO Stage 3 can be used for training on a single GPU by manually setting the environment variables:
-`WORLD_SIZE=1 LOCAL_RANK=0 MASTER_ADDR=0.0.0.0 MASTER_PORT=29500`
-
-:::
-
 ## 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
@@ -97,6 +100,7 @@ fsdp_sync_module_states | **REMOVED**
 fsdp_cpu_ram_efficient_loading | cpu_ram_efficient_loading
 fsdp_state_dict_type | state_dict_type
 fsdp_use_orig_params | **REMOVED**
+fsdp_activation_checkpointing | activation_checkpointing
 
 For more details, please see the migration guide in the [torchtitan repo](https://github.com/pytorch/torchtitan/blob/main/docs/fsdp.md). In Axolotl,
 if you were using the following FSDP1 config:
@@ -153,10 +157,6 @@ 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}

docs/multimodal.qmd

@@ -13,10 +13,18 @@ format:
 - [Pixtral](#sec-pixtral)
 - [Llava-1.5](#sec-llava-15)
 - [Mistral-Small-3.1](#sec-mistral-small-31)
+- [Mistral-Small-4](#sec-mistral-small-4)
+- [Magistral-Small-2509](#sec-magistral-small-2509)
+- [Voxtral](#sec-voxtral)
 - [Gemma-3](#sec-gemma-3)
 - [Gemma-3n](#sec-gemma-3n)
 - [Qwen2-VL](#sec-qwen2-vl)
 - [Qwen2.5-VL](#sec-qwen25-vl)
+- [Qwen3.5](#sec-qwen3-5)
+- [GLM-4.6V](#sec-glm-4-6v)
+- [SmolVLM2](#sec-smolvlm2)
+- [LFM2-VL](#sec-lfm2-vl)
+- [Intern-VL](#sec-intern-vl)
 
 ## Usage
 
@@ -31,14 +39,13 @@ skip_prepare_dataset: true
 remove_unused_columns: false  # leave columns in place as they are needed to handle image embeddings during training
 sample_packing: false  # not yet supported with multimodal
 
-chat_template:  # see in next section
+chat_template:  # see in next section if specified
 
 # example dataset
 datasets:
   - path: HuggingFaceH4/llava-instruct-mix-vsft
     type: chat_template
     split: train[:1%]
-    field_messages: messages
 
 # (optional) if doing lora, only finetune the Language model,
 # leave the vision model and vision tower frozen
@@ -53,10 +60,14 @@ image_resize_algorithm: bilinear
 
 Please see [examples](https://github.com/axolotl-ai/axolotl/tree/main/examples) folder for full configs.
 
-::: {.callout-warning}
+::: {.callout-tip}
 Some of our chat_templates have been extended to support broader dataset types. This should not break any existing configs.
 :::
 
+::: {.callout-note}
+As of now, we do not truncate nor drop samples based on `sequence_len` as each arch has different ways to process non-text tokens. We are looking for help on this.
+:::
+
 ### Mllama {#sec-mllama}
 
 ```yaml
@@ -91,10 +102,40 @@ chat_template: llava
 
 ### Mistral-Small-3.1 {#sec-mistral-small-31}
 
+::: {.callout-tip}
+Please make sure to install vision lib via `pip install 'mistral-common[opencv]==1.8.5'`
+:::
+
 ```yaml
 base_model: mistralai/Mistral-Small-3.1-24B-Instruct-2503
+```
 
-chat_template: mistral_v7_tekken
+### Mistral-Small-4 {#sec-mistral-small-4}
+
+```yaml
+base_model: mistralai/Mistral-Small-4-119B-2603
+```
+
+### Magistral-Small-2509 {#sec-magistral-small-2509}
+
+::: {.callout-tip}
+Please make sure to install vision lib via `pip install 'mistral-common[opencv]==1.8.5'`
+:::
+
+```yaml
+base_model: mistralai/Magistral-Small-2509
+```
+
+### Voxtral {#sec-voxtral}
+
+::: {.callout-tip}
+Please make sure to install audio lib via `pip3 install librosa==0.11.0 'mistral_common[audio]==1.8.3'`
+:::
+
+```yaml
+base_model: mistralai/Voxtral-Mini-3B-2507
+
+processor_type: VoxtralProcessor
 ```
 
 ### Gemma-3 {#sec-gemma-3}
@@ -143,6 +184,64 @@ base_model: Qwen/Qwen2.5-VL-7B-Instruct
 chat_template: qwen2_vl  # same as qwen2-vl
 ```
 
+### Qwen3-VL {#sec-qwen3-vl}
+
+```yaml
+base_model: Qwen/Qwen3-VL-4B-Instruct
+
+chat_template: qwen2_vl  # same as qwen2-vl
+```
+
+### Qwen3.5 {#sec-qwen3-5}
+
+```yaml
+base_model: Qwen/Qwen3.5-9B
+
+chat_template: qwen3_5
+```
+
+### GLM-4.6V {#sec-glm-4-6v}
+
+Both GLM-4.6V (106B MoE) and GLM-4.6V-Flash (9B) are supported.
+
+```yaml
+# GLM-4.6V (106B MoE version)
+base_model: zai-org/GLM-4.6V
+
+# OR GLM-4.6V-Flash (9B version)
+base_model: zai-org/GLM-4.6V-Flash
+```
+
+### SmolVLM2 {#sec-smolvlm2}
+
+::: {.callout-tip}
+Please make sure to install `num2words` via `pip3 install num2words==0.5.14`
+:::
+
+```yaml
+base_model: HuggingFaceTB/SmolVLM2-500M-Video-Instruct
+```
+
+### LFM2-VL {#sec-lfm2-vl}
+
+::: {.callout-warning}
+Please uninstall `causal-conv1d` via `pip3 uninstall -y causal-conv1d`
+:::
+
+```yaml
+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.
@@ -181,6 +280,20 @@ You may need to install `librosa` via `pip3 install librosa==0.11.0`.
 
 :::
 
+### Video
+
+::: {.callout-warning}
+
+This is not well tested at the moment. We welcome contributors!
+
+:::
+
+For video loading, you can use the following keys within `content` alongside `"type": "video"`:
+
+- `"path": "/path/to/video.mp4"`
+- `"url": "https://example.com/video.mp4"`
+- `"video": np.ndarray | list[PIL.Image.Image] | torch.Tensor` (or list of the aforementioned)
+
 ### Example
 
 Here is an example of a multi-modal dataset:

docs/optimizations.qmd

@@ -0,0 +1,156 @@
+---
+title: Optimizations Guide
+description: A guide to the performance and memory optimizations available in Axolotl.
+---
+
+Axolotl includes numerous optimizations to speed up training, reduce memory usage, and handle large models.
+
+This guide provides a high-level overview and directs you to the detailed documentation for each feature.
+
+## Speed Optimizations
+
+These optimizations focus on increasing training throughput and reducing total training time.
+
+### Sample Packing
+
+Improves GPU utilization by combining multiple short sequences into a single packed sequence for training. This requires enabling one of the [attention](#attention-implementations) implementations below.
+
+- **Config:** `sample_packing: true`
+- **Learn more:** [Sample Packing](multipack.qmd)
+
+### Attention Implementations
+
+Using an optimized attention implementation is critical for training speed.
+
+- **[Flash Attention 2](https://github.com/Dao-AILab/flash-attention)**: `flash_attention: true`. **(Recommended)** The industry standard for fast attention on modern GPUs. Requires Ampere or higher. For AMD, check [AMD Support](https://github.com/Dao-AILab/flash-attention?tab=readme-ov-file#amd-rocm-support).
+- **[Flex Attention](https://pytorch.org/blog/flexattention/)**: `flex_attention: true`.
+- **[SDP Attention](https://docs.pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html)**: `sdp_attention: true`. PyTorch's native implementation.
+- **[Xformers](https://github.com/facebookresearch/xformers)**: `xformers_attention: true`. Works with FP16.
+
+*Note: You should only enable one attention backend.*
+
+### LoRA Optimizations
+
+Leverages optimized kernels to accelerate LoRA training and reduce memory usage.
+
+- **Learn more:** [LoRA Optimizations Documentation](lora_optims.qmd)
+
+## Memory Optimizations
+
+These techniques help you fit larger models or use bigger batch sizes on your existing hardware.
+
+### Parameter Efficient Finetuning (LoRA & QLoRA)
+
+Drastically reduces memory by training a small set of "adapter" parameters instead of the full model. This is the most common and effective memory-saving technique.
+
+- Examples: Find configs with `lora` or `qlora` in the [examples directory](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/llama-3).
+- Config Reference: See `adapter`, `load_in_4bit`, and `load_in_8bit` in the [Configuration Reference](config-reference.qmd).
+
+### Gradient Checkpointing & Activation Offloading
+
+These techniques save VRAM by changing how activations are handled.
+
+- Gradient Checkpointing: re-computes activations during the backward pass, trading compute time for VRAM.
+- Activation Offloading: moves activations to CPU RAM or disk, trading I/O overhead for VRAM.
+- Learn more: [Gradient Checkpointing and Offloading Docs](gradient_checkpointing.qmd)
+
+### Layer Offloading
+
+Offloads frozen (non-trainable) decoder layer parameters to CPU and streams them back to GPU one layer at a time during forward/backward passes using CUDA stream prefetching. Especially effective for LoRA/QLoRA where most parameters are frozen.
+
+- **Config:** `layer_offloading: true`
+- **Learn more:** [Layer Offloading Docs](gradient_checkpointing.qmd#enabling-layer-offloading)
+
+### Cut Cross Entropy (CCE)
+
+Reduces VRAM usage by using an optimized cross-entropy loss calculation.
+
+- **Learn more:** [Custom Integrations - CCE](custom_integrations.qmd#cut-cross-entropy)
+
+### Liger Kernels
+
+Provides efficient Triton kernels to improve training speed and reduce memory usage.
+
+- **Learn more:** [Custom Integrations - Liger Kernels](custom_integrations.qmd#liger-kernels)
+
+### Expert Kernels
+
+Optimized kernel implementations for Mixture of Experts (MoE) model training.
+
+- **ScatterMoE**: Triton-based MoE kernels with fused LoRA support.
+- **SonicMoE**: CUTLASS-based MoE kernels for NVIDIA Hopper and Blackwell GPUs.
+
+- **Learn more:** [Custom Integrations - Kernels Integration](custom_integrations.qmd#kernels-integration)
+
+## Long Context Models
+
+Techniques to train models on sequences longer than their original context window.
+
+### RoPE Scaling
+
+Extends a model's context window by interpolating its Rotary Position Embeddings.
+
+- **Config:** Pass the `rope_scaling` config under the `overrides_of_model_config: `. To learn how to set RoPE, check the respective model config.
+
+### Sequence Parallelism
+
+Splits long sequences across multiple GPUs, enabling training with sequence lengths that would not fit on a single device.
+
+- **Learn more:** [Sequence Parallelism Documentation](sequence_parallelism.qmd)
+
+### Artic Long Sequence Training (ALST)
+
+ALST is a recipe that combines several techniques to train long-context models efficiently. It typically involves:
+
+- TiledMLP to reduce memory usage in MLP layers.
+- Tiled Loss functions (like [CCE](#cut-cross-entropy-(cce) or [Liger](#liger-kernels)).
+- Activation Offloading to CPU.
+
+- Example: [ALST Example Configuration](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/alst)
+
+## Large Models (Distributed Training)
+
+To train models that don't fit on a single GPU, you'll need to use a distributed training strategy like FSDP or DeepSpeed. These frameworks shard the model weights, gradients, and optimizer states across multiple GPUs and nodes.
+
+- **Learn more:** [Multi-GPU Guide](multi-gpu.qmd)
+- **Learn more:** [Multi-Node Guide](multi-node.qmd)
+
+### N-D Parallelism (Beta)
+
+For advanced scaling, Axolotl allows you to compose different parallelism techniques (e.g., Data, Tensor, Sequence Parallelism). This is a powerful approach to train an extremely large model by overcoming multiple bottlenecks at once.
+
+- **Learn more:** [N-D Parallelism Guide](nd_parallelism.qmd)
+
+
+## Quantization
+
+Techniques to reduce the precision of model weights for memory savings.
+
+### 4-bit Training (QLoRA)
+
+The recommended approach for quantization-based training. It loads the base model in 4-bit using `bitsandbytes` and then trains QLoRA adapters. See [Adapter Finetuning](#adapter-finetuning-lora-qlora) for details.
+
+### FP8 Training
+
+Enables training with 8-bit floating point precision on supported hardware (e.g., NVIDIA Hopper series GPUs) for significant speed and memory gains.
+
+- **Example:** [Llama 3 FP8 FSDP Example](https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-3/3b-fp8-fsdp2.yaml)
+
+### Quantization Aware Training (QAT)
+
+Simulates quantization effects during training, helping the model adapt and potentially improving the final accuracy of the quantized model.
+
+- **Learn more:** [QAT Documentation](qat.qmd)
+
+### GPTQ
+
+Allows you to finetune LoRA adapters on top of a model that has already been quantized using the GPTQ method.
+
+- **Example:** [GPTQ LoRA Example](https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-2/gptq-lora.yml)
+
+### MoE Expert Quantization
+
+Quantizes MoE expert weights on load to reduce VRAM when training MoE models with adapters. Required for Transformers v5+ MoE models where experts use fused `nn.Parameter` tensors.
+
+- **Config:** `quantize_moe_experts: true`
+- **Learn more:** [MoE Expert Quantization](expert_quantization.qmd)

docs/qat.qmd

@@ -23,10 +23,18 @@ To enable QAT in axolotl, add the following to your configuration file:
 
 ```yaml
 qat:
-  activation_dtype: # Optional[str] = "int8". Fake quantization layout to use for activation quantization. Valid options are "int4" and "int8"
-  weight_dtype: # Optional[str] = "int8". Fake quantization layout to use for weight quantization. Valid options are "int4" and "int8"
+  activation_dtype: # Optional[str] = "int8". Fake quantization layout to use for activation quantization. Valid options are "int4", "int8", "float8"
+  weight_dtype: # Optional[str] = "int8". Fake quantization layout to use for weight quantization. Valid options are "int4", "fp8", and "nvfp4".
   group_size: # Optional[int] = 32. The number of elements in each group for per-group fake quantization
   fake_quant_after_n_steps: # Optional[int] = None. The number of steps to apply fake quantization after
 ```
 
+We support the following quantization schemas:
+
+- `Int4WeightOnly` (requires the `fbgemm-gpu` extra when installing Axolotl)
+- `Int8DynamicActivationInt4Weight`
+- `Float8DynamicActivationFloat8Weight`
+- `Float8DynamicActivationInt4Weight`
+- `NVFP4`
+
 Once you have finished training, you must quantize your model by using the same quantization configuration which you used to train the model with. You can use the [`quantize`](./quantize.qmd) command to do this.

docs/quantize.qmd

@@ -22,8 +22,8 @@ Quantization is configured using the `quantization` key in your configuration fi
 ```yaml
 base_model: # The path to the model to quantize.
 quantization:
-  weight_dtype: # Optional[str] = "int8". Fake quantization layout to use for weight quantization. Valid options are uintX for X in [1, 2, 3, 4, 5, 6, 7], or int4, or int8
-  activation_dtype: # Optional[str] = "int8". Fake quantization layout to use for activation quantization. Valid options are "int4" and "int8"
+  activation_dtype: # Optional[str] = "int8". Fake quantization layout to use for activation quantization. Valid options are "int4", "int8", "float8"
+  weight_dtype: # Optional[str] = "int8". Fake quantization layout to use for weight quantization. Valid options are "int4", "fp8", and "nvfp4".
   group_size: # Optional[int] = 32. The number of elements in each group for per-group fake quantization
   quantize_embedding: # Optional[bool] = False. Whether to quantize the embedding layer.
 
@@ -39,9 +39,8 @@ you used to train the model:
 # qat.yml
 qat:
   activation_dtype: int8
-  weight_dtype: int8
+  weight_dtype: int4
   group_size: 256
-  quantize_embedding: true
 
 output_dir: # The path to the output directory used during training where the final checkpoint has been saved.
 ```
@@ -51,3 +50,11 @@ axolotl quantize qat.yml
 ```
 
 This ensures that an identical quantization configuration is used to quantize the model as was used to train it.
+
+
+::: {.callout-note}
+
+If you have configured pushing to hub with `hub_model_id`, your model hub name will have the quantization schema appended to it,
+e.g. `axolotl-ai-cloud/qat-nvfp4-llama3B` will become `axolotl-ai-cloud/qat-nvfp4-llama3B-nvfp4w`
+
+:::

docs/reward_modelling.qmd

@@ -11,6 +11,7 @@ We support the reward modelling techniques supported by `trl`.
 ### (Outcome) Reward Models
 
 Outcome reward models are trained using data which contains preference annotations for an entire interaction between the user and model (e.g. rather than per-turn or per-step).
+For improved training stability, you can use the `center_rewards_coefficient` parameter to encourage mean-zero reward outputs ([see TRL docs](https://huggingface.co/docs/trl/v0.10.1/en/reward_trainer#centering-rewards)).
 
 ```yaml
 base_model: google/gemma-2-2b

docs/rlhf.qmd

@@ -17,6 +17,7 @@ feedback. Various methods include, but not limited to:
 - [Kahneman-Tversky Optimization (KTO)](#kto)
 - [Odds Ratio Preference Optimization (ORPO)](#orpo)
 - [Group Relative Policy Optimization (GRPO)](#grpo)
+- [Group Reward-Decoupled Policy Optimization (GDPO)](#gdpo)
 
 
 ## RLHF using Axolotl
@@ -219,6 +220,21 @@ DPO supports the following types with the following dataset format:
 }
 ```
 
+#### chat_template.argilla_chat
+
+```json
+{
+    "chosen": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ],
+    "rejected": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ]
+}
+```
+
 #### chat_template.default
 
 ```yaml
@@ -582,6 +598,116 @@ 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.
@@ -595,6 +721,309 @@ trl:
 
 For more information, see [GRPO docs](https://huggingface.co/docs/trl/v0.17.0/en/grpo_trainer#loss-types).
 
+#### Async GRPO
+
+Async GRPO overlaps vLLM generation with training by producing rollouts in a background thread. While the model trains on the current batch, the next batch is already being generated. This can significantly reduce wall-clock time per step.
+
+```yaml
+trl:
+  use_data_producer: true     # Enable data producer protocol
+  use_vllm: true
+  async_prefetch: true         # Generate rollouts in background thread
+  prefetch_depth: 1            # Number of rollouts to prefetch
+  vllm_sync_interval: 2        # Sync weights to vLLM every N steps
+```
+
+::: {.callout-note}
+Because the background thread generates completions with slightly stale model weights, async GRPO uses importance sampling correction to account for the distribution shift. This is controlled by `vllm_importance_sampling_correction: true` (default when async is enabled).
+:::
+
+##### vLLM LoRA Sync
+
+By default, weight sync to vLLM merges the LoRA adapter into the base model and broadcasts all parameters via NCCL. LoRA sync is a faster alternative that saves only the adapter weights to the filesystem and has vLLM load them natively using Punica kernels.
+
+```yaml
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_target_linear: true
+
+trl:
+  vllm_lora_sync: true         # Enable native LoRA sync
+```
+
+When `vllm_lora_sync: true` is set, axolotl automatically selects the LoRA-aware vLLM serve module. Start vLLM as usual:
+
+```bash
+CUDA_VISIBLE_DEVICES=0 axolotl vllm-serve config.yaml
+```
+
+Then start training on a separate GPU:
+
+```bash
+CUDA_VISIBLE_DEVICES=1 axolotl train config.yaml
+```
+
+::: {.callout-tip}
+LoRA sync is especially beneficial with multi-GPU training (FSDP/DeepSpeed), where NCCL merge-sync can cause GPU contention with vLLM generation.
+:::
+
+##### Streaming Partial Batch
+
+Instead of scoring the entire batch at once, streaming mode scores one prompt group at a time. This enables finer-grained zero-advantage skipping and reduces peak memory usage during scoring.
+
+```yaml
+trl:
+  streaming_partial_batch: true
+```
+
+##### Importance Sampling Correction
+
+When using async prefetch, completions are generated from a slightly older version of the model. Importance sampling (IS) correction adjusts the policy gradient to account for this distribution shift.
+
+```yaml
+trl:
+  vllm_importance_sampling_correction: true   # Enable IS correction
+  importance_sampling_level: token             # 'token' or 'sequence'
+  off_policy_mask_threshold: 0.5              # Mask sequences with IS ratio below this
+```
+
+- `importance_sampling_level: token` applies per-token IS ratios (recommended with Liger kernel)
+- `importance_sampling_level: sequence` applies per-sequence IS ratios
+- `off_policy_mask_threshold` masks out sequences where the IS ratio indicates they are too far off-policy
+
+##### Replay Buffer
+
+The replay buffer caches rollout groups that had learning signal (non-zero reward variance) and uses them to replace zero-signal groups in later batches.
+
+```yaml
+trl:
+  replay_buffer_size: 100       # Max cached groups (0 = disabled)
+  replay_recompute_logps: true  # Recompute log-probs for replayed data (recommended)
+```
+
+::: {.callout-note}
+When `replay_recompute_logps: true` (default), old log-probabilities are recomputed using the current model weights. This fixes the IS mismatch that would otherwise occur when replaying stale data.
+:::
+
+##### Deferred Re-rolling
+
+Failed prompts (where the model produces zero reward for all generations) are buffered and re-injected into later batches when the model may be better equipped to solve them.
+
+```yaml
+trl:
+  reroll_start_fraction: 0.5    # Start re-rolling after 50% of training
+  reroll_max_groups: 1          # Max groups to replace per batch
+```
+
+##### Zero-Advantage Batch Skipping
+
+When all advantages in a micro-batch are zero (no learning signal), the forward/backward pass is skipped entirely. This is enabled by default and logged as `skipped_zero_adv_batches=1`.
+
+```yaml
+trl:
+  skip_zero_advantage_batches: true   # default
+```
+
+##### Parallel Reward Workers
+
+Reward functions that use `signal.alarm()` (e.g., `math_verify`) must run in the main thread. Parallel reward workers use subprocesses to work around this limitation while enabling concurrent reward computation.
+
+```yaml
+trl:
+  reward_num_workers: 4         # Number of subprocess workers (1 = no parallelism)
+```
+
+##### Full Async GRPO Example
+
+```yaml
+base_model: Qwen/Qwen2.5-1.5B-Instruct
+
+vllm:
+    host: 0.0.0.0
+    port: 8000
+    gpu_memory_utilization: 0.35
+    dtype: auto
+
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_target_linear: true
+
+rl: grpo
+trl:
+  use_data_producer: true
+  use_vllm: true
+  async_prefetch: true
+  prefetch_depth: 1
+  vllm_sync_interval: 2
+  vllm_lora_sync: true
+  streaming_partial_batch: true
+  vllm_importance_sampling_correction: true
+  off_policy_mask_threshold: 0.5
+  importance_sampling_level: token
+  num_generations: 8
+  max_completion_length: 512
+  reward_funcs:
+    - rewards.accuracy_reward
+  reroll_start_fraction: 0.5
+  replay_buffer_size: 100
+  reward_num_workers: 4
+  skip_zero_advantage_batches: true
+
+datasets:
+  - path: AI-MO/NuminaMath-TIR
+    type: rewards.prompt_transform
+    split: train
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+max_steps: 500
+learning_rate: 1e-5
+bf16: true
+gradient_checkpointing: true
+```
+
+```bash
+# Terminal 1: Start vLLM on GPU 0
+CUDA_VISIBLE_DEVICES=0 axolotl vllm-serve config.yaml
+
+# Terminal 2: Train on GPU 1
+CUDA_VISIBLE_DEVICES=1 axolotl train config.yaml
+```
+
+##### Multi-GPU Async GRPO
+
+Async GRPO supports FSDP and DeepSpeed ZeRO-3 for multi-GPU training. vLLM runs on one GPU while training is distributed across the remaining GPUs.
+
+**FSDP:**
+
+```yaml
+fsdp:
+  - full_shard
+  - auto_wrap
+fsdp_config:
+  fsdp_transformer_layer_cls_to_wrap: Qwen2DecoderLayer
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+```
+
+**DeepSpeed ZeRO-3:**
+
+```yaml
+deepspeed: deepspeed_configs/zero3_bf16.json
+gradient_checkpointing_kwargs:
+  use_reentrant: true   # Required for ZeRO-3
+```
+
+```bash
+# Terminal 1: Start vLLM on GPU 0
+CUDA_VISIBLE_DEVICES=0 axolotl vllm-serve config.yaml
+
+# Terminal 2: Train on GPUs 0,1
+CUDA_VISIBLE_DEVICES=0,1 accelerate launch --num_processes 2 -m axolotl.cli.train config.yaml
+```
+
+::: {.callout-important}
+With multi-GPU async prefetch, only rank 0 generates completions in the background thread. Results are broadcast to all ranks on the main thread. This avoids FSDP/DeepSpeed collective deadlocks from unsynchronized background threads.
+:::
+
+### GDPO
+
+GDPO (Group Reward-Decoupled Policy Optimization) extends GRPO for multi-reward training. It addresses the **reward advantage collapse** problem by normalizing each reward function independently before combining them.
+
+::: {.callout-tip}
+Use GDPO when training with multiple reward functions. For single reward, GRPO and GDPO produce equivalent results.
+:::
+
+Paper: [https://arxiv.org/pdf/2501.05242](https://arxiv.org/pdf/2501.05242)
+
+GDPO uses TRL's native `multi_objective_aggregation` parameter under the hood. When you set `rl: gdpo`, axolotl automatically configures TRL to use `normalize_then_sum` aggregation.
+
+```yaml
+base_model: Qwen/Qwen2.5-1.5B-Instruct
+
+vllm:
+    host: 0.0.0.0
+    port: 8000
+    tensor_parallel_size: 2
+    gpu_memory_utilization: 0.85
+
+rl: gdpo
+
+trl:
+    beta: 0.001
+    max_completion_length: 256
+    use_vllm: true
+    num_generations: 4
+    reward_funcs:
+        - rewards.format_reward
+        - rewards.correctness_reward
+    reward_weights: [1.0, 2.0]
+
+datasets:
+    - path: openai/gsm8k
+      name: main
+      type: rewards.oai_gsm8k_transform
+```
+
+You can also use GRPO with explicit aggregation control:
+
+```yaml
+rl: grpo
+trl:
+    multi_objective_aggregation: normalize_then_sum  # GDPO behavior
+    # or: sum_then_normalize  # Default GRPO behavior
+```
+
+#### GDPO vs GRPO
+
+| Aspect | GRPO | GDPO |
+|--------|------|------|
+| **Aggregation** | `sum_then_normalize` | `normalize_then_sum` |
+| **Multi-reward** | May collapse advantages | Preserves reward signals |
+| **Single reward** | Standard behavior | Equivalent to GRPO |
+
+#### Why GDPO?
+
+When using multiple rewards with GRPO, different reward combinations can produce identical advantages:
+
+```
+# Example: format + correctness rewards
+[format=0, correct=3] → sum=3
+[format=1, correct=2] → sum=3  ← GRPO sees these as equal!
+[format=2, correct=1] → sum=3
+[format=3, correct=0] → sum=3
+```
+
+GDPO normalizes each reward independently, preserving their relative differences.
+
+#### Reward Functions
+
+GDPO uses the same reward function format as GRPO:
+
+```python
+# rewards.py
+def format_reward(completions, **kwargs) -> list[float]:
+    return [1.0 if len(c) > 10 else 0.0 for c in completions]
+
+def correctness_reward(completions, answers, **kwargs) -> list[float]:
+    rewards = []
+    for completion, answer in zip(completions, answers):
+        # Your scoring logic here
+        rewards.append(score)
+    return rewards
+```
+
+#### Sequence Parallelism
+
+GDPO supports sequence parallelism for long-context training:
+
+```yaml
+rl: gdpo
+context_parallel_size: 2
+```
+
 ### SimPO
 
 SimPO uses [CPOTrainer](https://huggingface.co/docs/trl/main/en/cpo_trainer) but with alternative loss function.

docs/scripts/examples-allowlist.yml

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

docs/scripts/generate_config_docs.py

@@ -47,7 +47,6 @@ class QuartoGenerator:
         """Check if a type is a Pydantic BaseModel."""
         return inspect.isclass(type_obj) and issubclass(type_obj, BaseModel)
 
-    # pylint: disable=too-many-return-statements
     def _extract_nested_type(self, field_type) -> Any:
         """Extract the actual type from complex type annotations."""
         # Handle Annotated types (Python 3.9+)
@@ -124,7 +123,6 @@ class QuartoGenerator:
 
         return field_type
 
-    # pylint: disable=too-many-return-statements
     def _extract_all_pydantic_models_from_type(
         self, field_type
     ) -> list[type[BaseModel]]:
@@ -318,7 +316,6 @@ class QuartoGenerator:
 
         return all_groups
 
-    # pylint: disable=too-many-return-statements
     def _extract_field_groups_from_source(
         self, model_class: type[BaseModel]
     ) -> list[dict]:
@@ -503,7 +500,7 @@ class QuartoGenerator:
                     nested_schema = nested_model.model_json_schema()
                     nested_properties = nested_schema.get("properties", {})
                     nested_required = nested_schema.get("required", [])
-                except Exception:  # pylint: disable=broad-exception-caught
+                except Exception:
                     # Fallback: use model fields directly
                     nested_properties = {}
                     nested_required = []
@@ -607,7 +604,7 @@ class QuartoGenerator:
             schema = model_class.model_json_schema()
             properties = schema.get("properties", {})
             required = schema.get("required", [])
-        except Exception as e:  # pylint: disable=broad-exception-caught
+        except Exception as e:
             print(
                 f"Warning: Could not generate JSON schema ({e}). Using model fields instead."
             )

docs/scripts/generate_examples_docs.py

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

docs/streaming.qmd

@@ -0,0 +1,120 @@
+---
+title: Streaming Datasets
+description: How to use streaming mode for large-scale datasets and memory-efficient training
+order: 10
+---
+
+Streaming enables memory-efficient training with large datasets by loading data
+incrementally rather than loading the entire dataset into memory at once.
+
+Use streaming when:
+
+- Your dataset is too large to fit in memory (e.g. when you're doing pretraining with massive text corpora)
+- You want to start training immediately without preprocessing the entire dataset
+
+Streaming works with both remote and locally stored datasets!
+
+::: {.callout-note}
+Streaming currently only supports a single dataset. Multi-dataset support will be added soon.
+:::
+
+
+## Configuration
+
+### Basic Streaming
+
+Enable streaming mode by setting the `streaming` flag:
+
+```yaml
+streaming: true
+```
+
+### Pretraining with Streaming
+
+For pretraining tasks, streaming is automatically enabled when using `pretraining_dataset`:
+
+```yaml
+pretraining_dataset:
+  - path: HuggingFaceFW/fineweb-edu
+    type: pretrain
+    text_column: text
+    split: train
+
+# Optionally, enable sample packing
+streaming_multipack_buffer_size: 10000
+sample_packing: true
+```
+
+### SFT with Streaming
+
+For supervised fine-tuning with streaming:
+
+```yaml
+streaming: true
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+    split: train
+
+# Optionally, enable sample packing
+streaming_multipack_buffer_size: 10000
+sample_packing: true
+```
+
+## Configuration Options
+
+### `streaming_multipack_buffer_size`
+
+Controls the buffer size for multipack streaming (default: 10,000). This determines how
+many samples are buffered before packing. Larger buffers can improve packing efficiency
+but use more memory.
+
+### `shuffle_merged_datasets`
+
+When enabled, shuffles the streaming dataset using the buffer. This requires additional
+memory for the shuffle buffer.
+
+## Sample Packing with Streaming
+
+Sample packing is supported for streaming datasets. When enabled, multiple samples are
+packed into a single sequence to maximize GPU utilization:
+
+```yaml
+sample_packing: true
+streaming_multipack_buffer_size: 10000
+
+# For SFT: attention is automatically isolated between packed samples
+# For pretraining: control with pretrain_multipack_attn
+pretrain_multipack_attn: true  # prevent cross-attention between packed samples
+```
+
+For more information, see our [documentation](multipack.qmd) on multipacking.
+
+## Important Considerations
+
+### Memory Usage
+
+While streaming reduces memory usage compared to loading entire datasets, you still need
+to consider:
+
+- You can control the memory usage by adjusting `streaming_multipack_buffer_size`
+- Sample packing requires buffering multiple samples
+- Shuffling requires additional memory for the shuffle buffer
+
+### Performance
+
+- Streaming may have slightly higher latency compared to preprocessed datasets, as samples are processed on-the-fly
+- Network speed and disk read speed are important when streaming from remote sources or a local dataset, respectively
+- Consider using `axolotl preprocess` for smaller or more frequently used datasets
+
+### Evaluation Datasets
+
+Evaluation datasets are not streamed to ensure consistent evaluation metrics. They're
+loaded normally even when training uses streaming.
+
+## Examples
+
+See the `examples/streaming/` directory for complete configuration examples:
+
+- `pretrain.yaml`: Pretraining with streaming dataset
+- `sft.yaml`: Supervised fine-tuning with streaming

docs/telemetry.qmd

@@ -0,0 +1,61 @@
+---
+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

examples/LiquidAI/README.md

@@ -0,0 +1,67 @@
+# Finetune Liquid Foundation Models 2 (LFM2) with Axolotl
+
+[Liquid Foundation Models 2 (LFM2)](https://huggingface.co/collections/LiquidAI/lfm2-686d721927015b2ad73eaa38) are a family of small, open-weight models from [Liquid AI](https://www.liquid.ai/) focused on quality, speed, and memory efficiency. Liquid AI released text-only [LFM2](https://huggingface.co/collections/LiquidAI/lfm2-686d721927015b2ad73eaa38) and text+vision [LFM2-VL](https://huggingface.co/collections/LiquidAI/lfm2-vl-68963bbc84a610f7638d5ffa) models.
+
+LFM2 features a new hybrid Liquid architecture with multiplicative gates, short-range convolutions, and grouped query attention, enabling fast training and inference.
+
+This guide shows how to fine-tune both the LFM2 and LFM2-VL models with Axolotl.
+
+Thanks to the team at LiquidAI for giving us early access to prepare for these releases.
+
+## Getting Started
+
+1.  Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
+
+    Here is an example of how to install from pip:
+    ```bash
+    # Ensure you have a compatible version of Pytorch installed
+    pip3 install packaging setuptools wheel ninja
+    pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
+    ```
+
+2.  Run one of the finetuning examples below.
+
+    **LFM2**
+    ```bash
+    # FFT SFT (1x48GB @ 25GiB)
+    axolotl train examples/LiquidAI/lfm2-350m-fft.yaml
+    ```
+
+    **LFM2-VL**
+    ```bash
+    # LoRA SFT (1x48GB @ 2.7GiB)
+    axolotl train examples/LiquidAI/lfm2-vl-lora.yaml
+    ```
+
+    **LFM2-MoE**
+    ```bash
+    pip install git+https://github.com/huggingface/transformers.git@0c9a72e4576fe4c84077f066e585129c97bfd4e6
+
+    # LoRA SFT (1x48GB @ 16.2GiB)
+    axolotl train examples/LiquidAI/lfm2-8b-a1b-lora.yaml
+    ```
+
+### TIPS
+
+- **Installation Error**: If you encounter `ImportError: ... undefined symbol ...` or `ModuleNotFoundError: No module named 'causal_conv1d_cuda'`, the `causal-conv1d` package may have been installed incorrectly. Try uninstalling it:
+  ```bash
+  pip uninstall -y causal-conv1d
+  ```
+
+- **Dataset Loading**: Read more on how to load your own dataset in our [documentation](https://docs.axolotl.ai/docs/dataset_loading.html).
+- **Dataset Formats**:
+  - For LFM2 models, the dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+  - For LFM2-VL models, Axolotl follows the multi-content Messages format. See our [Multimodal docs](https://docs.axolotl.ai/docs/multimodal.html#dataset-format) for details.
+
+## Optimization Guides
+
+- [Optimizations Guide](https://docs.axolotl.ai/docs/optimizations.html)
+
+## Related Resources
+
+- [LFM2 Blog](https://www.liquid.ai/blog/liquid-foundation-models-v2-our-second-series-of-generative-ai-models)
+- [LFM2-VL Blog](https://www.liquid.ai/blog/lfm2-vl-efficient-vision-language-models)
+- [LFM2-MoE Blog](https://www.liquid.ai/blog/lfm2-8b-a1b-an-efficient-on-device-mixture-of-experts)
+- [Axolotl Docs](https://docs.axolotl.ai)
+- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/LiquidAI/lfm2-350m-fft.yaml

@@ -1,8 +1,8 @@
 base_model: LiquidAI/LFM2-350M
 
-chunked_cross_entropy: true
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
 
-chat_template: tokenizer_default
 eot_tokens:
   - "<|im_end|>"
 datasets:

examples/LiquidAI/lfm2-8b-a1b-lora.yaml

@@ -0,0 +1,59 @@
+base_model: LiquidAI/LFM2-8B-A1B
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_8bit: true
+
+eot_tokens:
+  - "<|im_end|>"
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_field_role: from
+    message_field_content: value
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.05
+output_dir: ./outputs/out
+
+sequence_len: 4096
+sample_packing: true
+
+adapter: lora
+lora_model_dir:
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules: '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: 2
+micro_batch_size: 4
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 5e-5
+
+bf16: true
+tf32: true
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 2
+saves_per_epoch: 1
+
+weight_decay: 0.0
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/LiquidAI/lfm2-vl-lora.yaml

@@ -0,0 +1,61 @@
+base_model: LiquidAI/LFM2-VL-450M
+trust_remote_code: true
+model_type: AutoModelForImageTextToText
+processor_type: AutoProcessor
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+# 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%]
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+adapter: lora
+lora_model_dir:
+
+sequence_len: 8192
+pad_to_sequence_len: false
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+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: 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
+eager_attention:
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/alst/README.md

@@ -7,3 +7,24 @@ techniques. It is a combination of:
 - Activation Offloading: Offload activations to CPU RAM to reduce memory usage
 
 For more information, you can check out the ALST paper [here](https://www.arxiv.org/abs/2506.13996).
+
+## Usage
+
+```yaml
+tiled_mlp: true
+
+# See Sequence Parallelism docs
+# https://docs.axolotl.ai/docs/sequence_parallelism.html
+context_parallel_size: int
+
+plugins:
+# See Cut Cross Entropy docs
+# https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+# or Liger Kernel docs
+# https://docs.axolotl.ai/docs/custom_integrations.html#liger-kernels
+  - axolotl.integrations.liger.LigerPlugin
+# ...
+
+```

examples/apertus/README.md

@@ -0,0 +1,110 @@
+# Finetune Swiss-AI's Apertus with Axolotl
+
+[Apertus](https://huggingface.co/collections/swiss-ai/apertus-llm-68b699e65415c231ace3b059) is a family of opensource models trained by Swiss-ai.
+
+This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
+
+## Getting started
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html). You need to install from main as Apertus 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 2.6.0 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. (Optional, highly recommended) Install XIELU CUDA
+
+```bash
+## Recommended for reduced VRAM and faster speeds
+
+# Point to CUDA toolkit directory
+# For those using our Docker image, use the below path.
+export CUDA_HOME=/usr/local/cuda
+
+pip3 install git+https://github.com/nickjbrowning/XIELU@59d6031 --no-build-isolation --no-deps
+```
+
+For any installation errors, see [XIELU Installation Issues](#xielu-installation-issues)
+
+3. Run the finetuning example:
+
+```bash
+axolotl train examples/apertus/apertus-8b-qlora.yaml
+```
+
+This config uses about 8.7 GiB VRAM.
+
+Let us know how it goes. Happy finetuning! 🚀
+
+### Tips
+
+- For inference, the official Apertus team recommends `top_p=0.9` and `temperature=0.8`.
+- You can instead use full paremter fine-tuning 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).
+
+### XIELU Installation Issues
+
+#### `ModuleNotFoundError: No module named 'torch'`
+
+Please check these one by one:
+- Running in correct environment
+- Env has PyTorch installed
+- CUDA toolkit is at `CUDA_HOME`
+
+If those didn't help, please try the below solutions:
+
+1. Pass env for CMAKE and try install again:
+
+    ```bash
+    Python_EXECUTABLE=$(which python) pip3 install git+https://github.com/nickjbrowning/XIELU@59d6031 --no-build-isolation --no-deps
+    ```
+
+2. Git clone the repo and manually hardcode python path:
+
+    ```bash
+    git clone https://github.com/nickjbrowning/XIELU
+    cd xielu
+    git checkout 59d6031
+
+    cd xielu
+    nano CMakeLists.txt  # or vi depending on your preference
+    ```
+
+    ```diff
+    execute_process(
+    -    COMMAND ${Python_EXECUTABLE} -c "import torch.utils; print(torch.utils.cmake_prefix_path)"
+    +    COMMAND /root/miniconda3/envs/py3.11/bin/python -c "import torch.utils; print(torch.utils.cmake_prefix_path)"
+        RESULT_VARIABLE TORCH_CMAKE_PATH_RESULT
+        OUTPUT_VARIABLE TORCH_CMAKE_PATH_OUTPUT
+        ERROR_VARIABLE TORCH_CMAKE_PATH_ERROR
+    )
+    ```
+
+    ```bash
+    pip3 install . --no-build-isolation --no-deps
+    ```
+
+## 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
+
+- [Apertus Tech Report](https://github.com/swiss-ai/apertus-tech-report/blob/main/Apertus_Tech_Report.pdf)
+- [Axolotl Docs](https://docs.axolotl.ai)
+- [Axolotl Website](https://axolotl.ai)
+- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/apertus/apertus-8b-qlora.yaml

@@ -0,0 +1,64 @@
+base_model: swiss-ai/Apertus-8B-Instruct-2509
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/lora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/arcee/README.md

@@ -17,8 +17,11 @@ Thanks to the team at Arcee.ai for using Axolotl in supervised fine-tuning the A
 git clone https://github.com/axolotl-ai-cloud/axolotl.git
 cd axolotl
 
-pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install packaging==26.0 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation -e '.[flash-attn]'
+
+# Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
+python scripts/cutcrossentropy_install.py | sh
 ```
 
 2. Run the finetuning example:

examples/archived/deepcoder/deepcoder-14B-preview-lora.yml

@@ -9,10 +9,6 @@ strict: false
 datasets:
   - path: fozziethebeat/alpaca_messages_2k_test
     type: chat_template
-    field_messages: messages
-    message_property_mappings:
-      role: role
-      content: content
 
 dataset_prepared_path:
 val_set_size: 0.05

examples/cloud/baseten.yaml

@@ -0,0 +1,10 @@
+provider: baseten
+project_name:
+
+secrets:
+  - HF_TOKEN
+  - WANDB_API_KEY
+
+gpu: h100
+gpu_count: 8
+node_count: 1

examples/deepcogito/cogito-v1-preview-llama-3B-lora.yml

@@ -9,10 +9,6 @@ strict: false
 datasets:
   - path: fozziethebeat/alpaca_messages_2k_test
     type: chat_template
-    field_messages: messages
-    message_property_mappings:
-      role: role
-      content: content
 
 dataset_prepared_path:
 val_set_size: 0.05

examples/deepcogito/cogito-v1-preview-qwen-14B-lora.yml

@@ -9,10 +9,6 @@ strict: false
 datasets:
   - path: fozziethebeat/alpaca_messages_2k_test
     type: chat_template
-    field_messages: messages
-    message_property_mappings:
-      role: role
-      content: content
 
 dataset_prepared_path:
 val_set_size: 0.05

examples/devstral/README.md

@@ -16,11 +16,17 @@ Thanks to the team at MistralAI for giving us early access to prepare for this r
 
 ```bash
 # Ensure you have Pytorch installed (Pytorch 2.6.0 min)
-pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install packaging==26.0 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
-2. Run the finetuning example:
+2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage
+
+```bash
+python scripts/cutcrossentropy_install.py | sh
+```
+
+3. Run the finetuning example:
 
 ```bash
 axolotl train examples/devstral/devstral-small-qlora.yml

examples/devstral/devstral-small-qlora.yml

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

examples/eaft/eaft-example.yml

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

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

@@ -1,7 +1,5 @@
 base_model: google/gemma-3-1b-it
-# 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
 
@@ -26,10 +24,15 @@ datasets:
 val_set_size: 0.0
 output_dir: ./outputs/out
 
+# Freeze vision tower
+unfrozen_parameters:
+  - ^model\.language_model\..*
+  - ^lm_head\..*
+
 adapter: qlora
 lora_r: 32
 lora_alpha: 16
-lora_dropout: 0.05
+lora_dropout: 0
 lora_target_linear: true
 
 sequence_len: 2048

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

@@ -0,0 +1,71 @@
+base_model: google/gemma-3-270m-it
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+# gemma3 doesn't seem to play nice with ddp
+ddp_find_unused_parameters: true
+
+load_in_8bit: false
+load_in_4bit: true
+
+# huggingface repo
+chat_template: gemma3
+eot_tokens:
+  - <end_of_turn>
+datasets:
+  - path: cgato/SlimOrcaDedupCleaned
+    type: chat_template
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+# Freeze vision tower
+unfrozen_parameters:
+  - ^model\.language_model\..*
+  - ^lm_head\..*
+
+adapter: qlora
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0
+lora_target_linear: true
+
+sequence_len: 2048
+sample_packing: true
+eval_sample_packing: false
+
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch:
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

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

@@ -20,6 +20,11 @@ dataset_prepared_path: last_run_prepared
 val_set_size: 0.01
 output_dir: ./outputs/out
 
+# Freeze vision tower
+unfrozen_parameters:
+  - ^model\.language_model\..*
+  - ^lm_head\..*
+
 adapter: qlora
 lora_model_dir:
 
@@ -29,8 +34,8 @@ sample_packing: true
 
 lora_r: 32
 lora_alpha: 16
-lora_dropout: 0.05
-lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+lora_dropout: 0
+lora_target_linear: true
 
 wandb_project:
 wandb_entity:

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

@@ -18,7 +18,7 @@ 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
@@ -31,7 +31,7 @@ pad_to_sequence_len: false
 
 lora_r: 32
 lora_alpha: 16
-lora_dropout: 0.05
+lora_dropout: 0
 lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
 
 wandb_project:

examples/gemma3n/README.md

@@ -10,7 +10,7 @@ Gemma-3n is a family of multimodal models from Google found on [HuggingFace](htt
 
 ```bash
 # Ensure you have Pytorch installed (Pytorch 2.6.0 min)
-pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install packaging==26.0 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
@@ -23,7 +23,15 @@ pip3 install timm==1.0.17
 pip3 install librosa==0.11.0
 ```
 
-3. Run the finetuning example:
+3. Download sample dataset files
+
+```bash
+# for text + vision + audio only
+wget https://huggingface.co/datasets/Nanobit/text-vision-audio-2k-test/resolve/main/African_elephant.jpg
+wget https://huggingface.co/datasets/Nanobit/text-vision-audio-2k-test/resolve/main/En-us-African_elephant.oga
+```
+
+4. Run the finetuning example:
 
 ```bash
 # text only

examples/glm45/README.md

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

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

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

examples/glm46v/README.md

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

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

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

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

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