chunked DPO loss

* fix: unify default for conversations_field

* fix: suggestion to remove defaults

.axolotl-complete.bash

@@ -0,0 +1,41 @@
+#!/bin/bash
+
+_axolotl_completions() {
+    local cur prev
+    COMPREPLY=()
+    cur="${COMP_WORDS[COMP_CWORD]}"
+    prev="${COMP_WORDS[COMP_CWORD-1]}"
+
+    # If we're completing the first argument (the command)
+    if [[ $COMP_CWORD -eq 1 ]]; then
+        mapfile -t COMPREPLY < <(compgen -W "delinearize-llama4 fetch lm-eval merge-sharded-fsdp-weights quantize vllm-serve evaluate inference merge-lora preprocess train" -- "$cur")
+        return 0
+    fi
+
+    # Commands that should complete with directories and YAML files
+    local -a yaml_commands=("merge-sharded-fsdp-weights" "quantize" "vllm-serve" "evaluate" "inference" "merge-lora" "preprocess" "train")
+
+    # Check if previous word is in our list
+    if [[ " ${yaml_commands[*]} " =~ (^|[[:space:]])$prev($|[[:space:]]) ]]; then
+        # Use filename completion which handles directories properly
+        compopt -o filenames
+        mapfile -t COMPREPLY < <(compgen -f -- "$cur")
+
+        # Filter to only include directories and YAML files
+        local -a filtered=()
+        for item in "${COMPREPLY[@]}"; do
+            if [[ -d "$item" ]] || [[ "$item" == *.yaml ]] || [[ "$item" == *.yml ]]; then
+                filtered+=("$item")
+            fi
+        done
+        COMPREPLY=("${filtered[@]}")
+
+        return 0
+    fi
+
+    # Default: no completion
+    return 0
+}
+
+# Remove the -o nospace option - let filenames handle it
+complete -F _axolotl_completions axolotl

.bandit

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

.coderabbit.yaml

@@ -0,0 +1,17 @@
+# yaml-language-server: $schema=https://coderabbit.ai/integrations/schema.v2.json
+language: "en-US"
+early_access: false
+reviews:
+  profile: "chill"
+  request_changes_workflow: false
+  high_level_summary: true
+  review_status: true
+  collapse_walkthrough: true
+  poem: false
+  sequence_diagrams: false
+  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

@@ -57,6 +57,13 @@ 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

.github/workflows/base.yml

@@ -5,17 +5,19 @@ on:
     branches:
       - "main"
     paths:
-      - 'Dockerfile-base'
+      - 'docker/Dockerfile-base'
+      - 'docker/Dockerfile-uv-base'
       - '.github/workflows/base.yml'
   pull_request:
     paths:
-      - 'Dockerfile-base'
+      - 'docker/Dockerfile-base'
+      - 'docker/Dockerfile-uv-base'
       - '.github/workflows/base.yml'
   workflow_dispatch:
 
 jobs:
   build-base:
-    if: github.repository_owner == 'axolotl-ai-cloud'
+    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
@@ -23,13 +25,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: "124"
-            cuda_version: 12.4.1
-            cudnn_version: ""
-            python_version: "3.11"
-            pytorch: 2.5.1
-            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
-            dockerfile: "Dockerfile-base"
           - cuda: "124"
             cuda_version: 12.4.1
             cudnn_version: ""
@@ -48,10 +43,10 @@ jobs:
             cuda_version: 12.6.3
             cudnn_version: ""
             python_version: "3.11"
-            pytorch: 2.7.1
+            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: "128"
+          - cuda: "126"
             cuda_version: 12.6.3
             cudnn_version: ""
             python_version: "3.11"
@@ -62,9 +57,23 @@ jobs:
             cuda_version: 12.8.1
             cudnn_version: ""
             python_version: "3.11"
-            pytorch: nightly
+            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-nightly"
+            dockerfile: "Dockerfile-base"
+          - cuda: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            python_version: "3.11"
+            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"
+#          - cuda: "128"
+#            cuda_version: 12.8.1
+#            cudnn_version: ""
+#            python_version: "3.11"
+#            pytorch: nightly
+#            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+#            dockerfile: "Dockerfile-base-nightly"
 #          # "next" is for release candidates of pytorch
 #          - cuda: "128"
 #            cuda_version: 12.8.1
@@ -106,7 +115,7 @@ jobs:
             PYTORCH_VERSION=${{ matrix.pytorch }}
             TORCH_CUDA_ARCH_LIST=${{ matrix.torch_cuda_arch_list }}
   build-base-uv:
-    if: github.repository_owner == 'axolotl-ai-cloud'
+    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
     strategy:
@@ -120,6 +129,13 @@ jobs:
             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: ""
@@ -127,6 +143,13 @@ jobs:
             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: 2.8.0
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-uv-base"
     steps:
       - name: Checkout
         uses: actions/checkout@v4

.github/workflows/lint.yml

@@ -3,6 +3,7 @@ on:
   # check on PRs, and manual triggers
   merge_group:
   pull_request:
+      types: [opened, synchronize, reopened, ready_for_review]
       paths:
        - '**.py'
        - 'requirements.txt'
@@ -16,6 +17,7 @@ jobs:
   pre-commit:
     name: pre-commit
     runs-on: ubuntu-latest
+    if: ${{ !github.event.pull_request.draft }}
     steps:
       - uses: actions/checkout@v4
       - uses: actions/setup-python@v5

.github/workflows/main.yml

@@ -15,26 +15,31 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.1
-            axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.6.0
-            axolotl_extras: vllm
-            is_latest: true
           - 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
             axolotl_extras:
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.7.1
+            pytorch: 2.8.0
             axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
@@ -83,26 +88,37 @@ jobs:
     strategy:
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.1
-            axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.6.0
             axolotl_extras:
-            is_latest: true
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
+            pytorch: 2.7.0
+            axolotl_extras:
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.1
+            axolotl_extras:
+            is_latest:
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.1
+            axolotl_extras: vllm
+            is_latest: true
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
             pytorch: 2.7.1
             axolotl_extras:
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
-            pytorch: 2.7.1
+            pytorch: 2.8.0
             axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
@@ -146,11 +162,29 @@ jobs:
     strategy:
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - 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: 2.7.1
+            axolotl_extras:
+            is_latest:
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.1
+            axolotl_extras: vllm
+            is_latest: true
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.8.0
+            axolotl_extras:
+            is_latest:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout

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

@@ -21,22 +21,15 @@ concurrency:
 
 jobs:
   test-axolotl-multigpu:
-    if: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' }}
+    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) }}
     strategy:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.6.0
-            axolotl_extras: vllm
-            num_gpus: 2
-            nightly_build: "true"
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.1
             axolotl_extras:
             num_gpus: 2
             nightly_build: "true"
@@ -44,7 +37,14 @@ jobs:
             cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.7.1
-            axolotl_extras:
+            axolotl_extras: vllm
+            num_gpus: 2
+            nightly_build: "true"
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.8.0
+            axolotl_extras: fbgemm-gpu
             num_gpus: 2
             nightly_build: "true"
     runs-on: [self-hosted, modal]

.github/workflows/nightlies.yml

@@ -12,16 +12,16 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.1
-            axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - 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: 2.7.1
+            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -65,16 +65,16 @@ jobs:
     strategy:
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.1
-            axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - 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: 2.7.1
+            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout

.github/workflows/preview-docs.yml

@@ -2,7 +2,7 @@ name: Preview
 on:
   workflow_dispatch:
   pull_request:
-    types: [opened, synchronize, reopened]
+    types: [opened, synchronize, reopened, ready_for_review]
 
     # Run the workflow only when one of these files changes
     paths:
@@ -25,9 +25,12 @@ permissions:
 jobs:
   preview:
     runs-on: ubuntu-latest
+    if: ${{ !github.event.pull_request.draft }}
     steps:
       - name: Check out repository
         uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event.pull_request.head.sha }}
 
       - name: Set up Quarto
         uses: quarto-dev/quarto-actions/setup@v2
@@ -50,10 +53,12 @@ jobs:
 
       - name: Netlify Publish
         uses: nwtgck/actions-netlify@v3.0
+        if: ${{ github.event.pull_request.head.repo.full_name == github.repository }}
+        id: netlify
         with:
           publish-dir: './_site'
-          enable-pull-request-comment: true
-          enable-github-deployment: true
+          enable-pull-request-comment: false
+          enable-github-deployment: false
           github-token: ${{ secrets.GITHUB_TOKEN }}
           deploy-message: "Deployed On Netlify"
           github-deployment-environment: 'preview'
@@ -61,3 +66,13 @@ jobs:
         env:
           NETLIFY_AUTH_TOKEN: ${{ secrets.NETLIFY_AUTH_TOKEN }}
           NETLIFY_SITE_ID: ${{ secrets.NETLIFY_SITE_ID }}
+
+      - name: Update PR with preview link
+        if: ${{ steps.netlify.outcome == 'success' }}
+        uses: marocchino/sticky-pull-request-comment@v2
+        with:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          message: |
+            📖 **Documentation Preview**: ${{ steps.netlify.outputs.deploy-url }}
+
+            Deployed on Netlify from commit ${{ github.event.pull_request.head.sha }}

.github/workflows/tests-nightly.yml

@@ -18,116 +18,26 @@ jobs:
         env:
           SKIP: no-commit-to-branch
 
-  preload-cache:
-    name: Preload HF cache
-    runs-on: ubuntu-latest
-    strategy:
-      fail-fast: false
-      matrix:
-        python_version: ["3.11"]
-        pytorch_version: ["2.6.0"]
-    timeout-minutes: 20
-
-    env:
-      AXOLOTL_IS_CI_CACHE_PRELOAD: "1"
-
-    steps:
-      - name: Check out repository code
-        uses: actions/checkout@v4
-
-      - name: Restore HF cache
-        id: hf-cache-restore
-        uses: actions/cache/restore@v4
-        with:
-          path: |
-            /home/runner/.cache/huggingface/hub/datasets--*
-            /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ runner.os }}-hf-hub-cache-v2
-
-      - name: Setup Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: ${{ matrix.python_version }}
-          cache: 'pip' # caching pip dependencies
-
-      - name: upgrade pip
-        run: |
-          pip3 install --upgrade pip
-          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
-
-      - name: Install PyTorch
-        run: |
-          pip3 install torch==${{ matrix.pytorch_version }}
-
-      - name: Install dependencies
-        run: |
-          pip3 show torch
-          pip3 install --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: Make sure PyTorch version wasn't clobbered
-        run: |
-          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
-
-      - name: Ensure axolotl CLI was installed
-        run: |
-          axolotl --help
-
-      - name: Pre-Download dataset fixture
-        run: |
-          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
-
-      - name: Run tests
-        run: |
-          pytest -v tests/conftest.py
-
-      - name: Upload coverage to Codecov
-        uses: codecov/codecov-action@v5
-        with:
-          token: ${{ secrets.CODECOV_TOKEN }}
-          files: ./coverage.xml
-          flags: unittests,pytorch-${{ 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 {} \;
-
-      - name: Save HF cache
-        id: hf-cache
-        uses: actions/cache/save@v4
-        with:
-          path: |
-            /home/runner/.cache/huggingface/hub/datasets--*
-            /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
-
   pytest:
     name: PyTest
     runs-on: ubuntu-latest
-    needs: [preload-cache]
     strategy:
       fail-fast: false
       max-parallel: 2
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.5.1", "2.6.0", "2.7.0"]
+        pytorch_version: ["2.6.0", "2.7.0"]
     timeout-minutes: 20
 
     steps:
       - name: Check out repository code
         uses: actions/checkout@v4
 
-      - name: Restore HF cache
-        id: hf-cache-restore
-        uses: actions/cache/restore@v4
-        with:
-          path: |
-            /home/runner/.cache/huggingface/hub/datasets--*
-            /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ runner.os }}-hf-hub-cache-v2
+      - 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
 
       - name: Setup Python
         uses: actions/setup-python@v5
@@ -142,7 +52,7 @@ jobs:
 
       - name: Install PyTorch
         run: |
-          pip3 install torch==${{ matrix.pytorch_version }}
+          pip3 install torch==${{ matrix.pytorch_version }} torchvision
 
       - name: Update requirements.txt
         run: |
@@ -168,15 +78,11 @@ jobs:
         run: |
           axolotl --help
 
-      - name: Pre-Download dataset fixture
-        run: |
-          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
-
       - name: Run tests
         run: |
-          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/
-          pytest -v tests/patched/
-          pytest -v tests/cli/
+          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 tests/cli/
 
       - name: cleanup pip cache
         run: |
@@ -186,24 +92,24 @@ jobs:
     if: github.repository_owner == 'axolotl-ai-cloud'
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
-    timeout-minutes: 60
+    timeout-minutes: 120
     needs: [pre-commit, pytest]
 
     strategy:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
-            pytorch: 2.5.1
+            pytorch: 2.6.0
             num_gpus: 1
             axolotl_extras:
             nightly_build: "true"
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.7.1
             num_gpus: 1
             axolotl_extras:
             nightly_build: "true"
@@ -217,7 +123,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==0.71.8 jinja2
+          pip install modal==1.0.2 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
@@ -231,3 +137,45 @@ jobs:
       - name: Run tests job on Modal
         run: |
           modal run cicd.e2e_tests
+  docker-e2e-multigpu-tests:
+    if: github.repository_owner == 'axolotl-ai-cloud'
+    # this job needs to be run on self-hosted GPU runners...
+    runs-on: [self-hosted, modal]
+    timeout-minutes: 120
+    needs: [pre-commit, pytest, docker-e2e-tests]
+
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.1
+            num_gpus: 2
+            axolotl_extras:
+            nightly_build: "true"
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Install Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+      - name: Install Modal
+        run: |
+          python -m pip install --upgrade pip
+          pip install modal==1.0.2 jinja2
+      - name: Update env vars
+        run: |
+          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
+          echo "PYTORCH_VERSION=${{ matrix.pytorch}}" >> $GITHUB_ENV
+          echo "AXOLOTL_ARGS=${{ matrix.axolotl_args}}" >> $GITHUB_ENV
+          echo "AXOLOTL_EXTRAS=${{ 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
+      - name: Run tests job on Modal
+        run: |
+          modal run cicd.multigpu

.github/workflows/tests.yml

@@ -13,6 +13,7 @@ on:
       - 'cicd/cicd.sh'
       - 'cicd/Dockerfile.jinja'
   pull_request:
+      types: [opened, synchronize, reopened, ready_for_review]
       paths:
        - '**.py'
        - 'requirements.txt'
@@ -34,6 +35,7 @@ jobs:
   pre-commit:
     name: pre-commit
     runs-on: ubuntu-latest
+    if: ${{ !github.event.pull_request.draft }}
     steps:
       - uses: actions/checkout@v4
       - uses: actions/setup-python@v5
@@ -47,12 +49,13 @@ jobs:
   pytest:
     name: PyTest
     runs-on: ubuntu-latest
+    if: ${{ !github.event.pull_request.draft }}
 #    needs: [preload-cache]
     strategy:
       fail-fast: false
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.5.1", "2.6.0", "2.7.1"]
+        pytorch_version: ["2.6.0", "2.7.1", "2.8.0"]
     timeout-minutes: 20
 
     steps:
@@ -78,7 +81,7 @@ jobs:
 
       - name: Install PyTorch
         run: |
-          pip3 install torch==${{ matrix.pytorch_version }}
+          pip3 install torch==${{ matrix.pytorch_version }} torchvision
 
       - name: Install dependencies
         run: |
@@ -102,16 +105,18 @@ jobs:
 
       - name: Run tests
         run: |
-          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/ --cov=axolotl --cov-report=xml
-          pytest -v tests/patched/ --cov=axolotl --cov-append --cov-report=xml
-          pytest -v tests/cli/ --cov=axolotl --cov-append --cov-report=xml
+          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ --ignore=tests/monkeypatch/ tests/ --cov=axolotl --cov-report=xml
+          pytest -v --durations=10 tests/monkeypatch/ --cov=axolotl --cov-append --cov-report=xml
+          pytest -v --durations=10 tests/patched/ --cov=axolotl --cov-append --cov-report=xml
+          pytest -v --durations=10 tests/cli/ --cov=axolotl --cov-append --cov-report=xml
 
-      - name: Upload coverage artifacts
-        uses: actions/upload-artifact@v4
+      - name: Upload coverage to Codecov
+        uses: codecov/codecov-action@v5
         with:
-          name: coverage-${{ matrix.pytorch_version }}-${{ github.run_id }}
-          path: ./coverage.xml
-          retention-days: 1
+          token: ${{ secrets.CODECOV_TOKEN }}
+          files: ./coverage.xml
+          flags: unittests,pytorch-${{ matrix.pytorch_version }}
+          fail_ci_if_error: false
 
       - name: cleanup pip cache
         run: |
@@ -120,11 +125,12 @@ jobs:
   pytest-sdist:
     name: PyTest from Source Dist
     runs-on: ubuntu-latest
+    if: ${{ !github.event.pull_request.draft }}
     strategy:
       fail-fast: false
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.5.1", "2.6.0", "2.7.1"]
+        pytorch_version: ["2.6.0", "2.7.1", "2.8.0"]
     timeout-minutes: 20
 
     steps:
@@ -150,7 +156,7 @@ jobs:
 
       - name: Install PyTorch
         run: |
-          pip3 install torch==${{ matrix.pytorch_version }}
+          pip3 install torch==${{ matrix.pytorch_version }} torchvision
 
       - name: Install dependencies
         run: |
@@ -174,36 +180,67 @@ jobs:
 
       - name: Run tests
         run: |
-          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/
-          pytest -v tests/patched/
-          pytest -v tests/cli/
+          pytest -v --durations=10 -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ --ignore=tests/monkeypatch/ tests/ --cov=axolotl --cov-report=xml
+          pytest -v --durations=10 tests/monkeypatch/ --cov=axolotl --cov-append --cov-report=xml
+          pytest -v --durations=10 tests/cli/
 
       - name: cleanup pip cache
         run: |
           find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
 
+  gate-skip-e2e:
+    needs: [pre-commit, pytest, pytest-sdist]
+    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' }}
+    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, pytest-sdist, gate-skip-e2e]
 
     strategy:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.6.0
-            num_gpus: 1
-            axolotl_extras: vllm
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.7.1
+            num_gpus: 1
+            axolotl_extras:
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.1
             num_gpus: 1
             axolotl_extras:
             dockerfile: "Dockerfile-uv.jinja"
@@ -233,43 +270,26 @@ jobs:
         run: |
           modal run cicd.e2e_tests
 
-      - name: Upload coverage artifacts
-        if: always()
-        uses: actions/upload-artifact@v4
-        with:
-          name: coverage-e2e-1st-${{ github.run_id }}
-          path: ./e2e-coverage.xml
-          retention-days: 1
-
   docker-e2e-tests:
-    if: github.repository_owner == 'axolotl-ai-cloud'
+    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: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.6.0
-            num_gpus: 1
-            axolotl_extras: llmcompressor
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.1
-            num_gpus: 1
-            axolotl_extras:
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
-            pytorch: 2.7.1
+            pytorch: 2.6.0
             num_gpus: 1
             axolotl_extras:
           - cuda: 128
@@ -278,6 +298,13 @@ jobs:
             pytorch: 2.7.1
             num_gpus: 1
             axolotl_extras:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.8.0
+            num_gpus: 1
+            gpu_type: "B200"
+            axolotl_extras: fbgemm-gpu
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -298,35 +325,29 @@ jobs:
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
+          echo "GPU_TYPE=${{ matrix.gpu_type || 'L40S'}}" >> $GITHUB_ENV
           echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
           echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         run: |
           modal run cicd.e2e_tests
 
-      - name: Upload coverage artifacts
-        if: always()
-        uses: actions/upload-artifact@v4
-        with:
-          name: coverage-e2e-${{ matrix.cuda }}-${{ matrix.pytorch }}-${{ github.run_id }}
-          path: ./e2e-coverage.xml
-          retention-days: 1
-
   docker-e2e-cleanup:
     runs-on: [self-hosted, modal]
     timeout-minutes: 90
     needs: [docker-e2e-tests]
+    if: ${{ !github.event.pull_request.draft }}
 
     strategy:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.7.1
             num_gpus: 1
-            axolotl_extras: vllm
+            axolotl_extras:
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -351,26 +372,3 @@ jobs:
       - name: Run tests job on Modal
         run: |
           modal run cicd.cleanup
-
-  upload-coverage:
-    name: Upload Coverage to Codecov
-    runs-on: ubuntu-latest
-    needs: [pytest, docker-e2e-tests, docker-e2e-tests-1st]
-    if: github.event_name == 'pull_request' || github.ref == 'refs/heads/main'
-
-    steps:
-      - name: Download coverage reports
-        uses: actions/download-artifact@v4
-        with:
-          path: coverage-reports
-
-      - name: Upload coverage to Codecov
-        uses: codecov/codecov-action@v5
-        with:
-          token: ${{ secrets.CODECOV_TOKEN }}
-          directory: coverage-reports
-          fail_ci_if_error: false
-          verbose: true
-          name: codecov-umbrella
-          override_commit: ${{ github.event.pull_request.head.sha || github.sha }}
-          override_pr: ${{ github.event.pull_request.number }}

.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: v0.12.12
     hooks:
-    -   id: black
--   repo: https://github.com/pycqa/isort
-    rev: 6.0.1
-    hooks:
-      - id: isort
--   repo: https://github.com/PyCQA/flake8
-    rev: 7.2.0
-    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: v1.16.0
+    rev: v1.17.1
     hooks:
     - id: mypy
       additional_dependencies:
@@ -36,7 +26,7 @@ repos:
             'pydantic>=2.5.3',
         ]
 -   repo: https://github.com/PyCQA/bandit
-    rev: 1.8.3
+    rev: 1.8.6
     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/README.md

@@ -119,14 +119,15 @@ datasets:
 
 ## Dataset Processing
 
-| Option                        | Default                    | Description                       |
-| ----------------------------- | -------------------------- | --------------------------------- |
-| `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_keep_in_memory`      | `false`                    | Keep dataset in memory            |
-| `shuffle_merged_datasets`     | `true`                     | Shuffle merged datasets           |
-| `dataset_exact_deduplication` | `true`                     | Deduplicate datasets              |
+| Option                            | Default                    | Description                         |
+| --------------------------------- | -------------------------- | ----------------------------------- |
+| `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_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 |
+| `dataset_exact_deduplication`     | `true`                     | Deduplicate datasets                |
 
 ## LoRA Configuration
 
@@ -184,7 +185,6 @@ datasets:
 | `flash_attention`          | `false` | Use flash attention           |
 | `flash_attn_cross_entropy` | `false` | Flash attention cross entropy |
 | `flash_attn_rms_norm`      | `false` | Flash attention RMS norm      |
-| `flash_attn_fuse_qkv`      | `false` | Fuse QKV operations           |
 | `flash_attn_fuse_mlp`      | `false` | Fuse MLP operations           |
 | `sdp_attention`            | `false` | Use scaled dot product        |
 | `s2_attention`             | `false` | Use shifted sparse attention  |

.runpod/src/config/config.yaml

@@ -97,7 +97,7 @@
 #       # 'no_input_format' cannot include {input}
 #       no_input_format: "{instruction} "
 
-#       # For `completion` datsets only, uses the provided field instead of `text` column
+#       # For `completion` datasets only, uses the provided field instead of `text` column
 #       field:
 
 # # Axolotl attempts to save the dataset as an arrow after packing the data together so
@@ -296,7 +296,6 @@
 # flash_attention:
 # flash_attn_cross_entropy:  # Whether to use flash-attention cross entropy implementation - advanced use only
 # flash_attn_rms_norm:  # Whether to use flash-attention rms norm implementation - advanced use only
-# flash_attn_fuse_qkv: # Whether to fuse QKV into a single operation
 # flash_attn_fuse_mlp: # Whether to fuse part of the MLP into a single operation
 # # Whether to use scaled-dot-product attention
 # # https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html
@@ -541,7 +540,6 @@ xformers_attention: ${XFORMERS_ATTENTION}
 flash_attention: ${FLASH_ATTENTION}
 flash_attn_cross_entropy: ${FLASH_ATTN_CROSS_ENTROPY}
 flash_attn_rms_norm: ${FLASH_ATTN_RMS_NORM}
-flash_attn_fuse_qkv: ${FLASH_ATTN_FUSE_QKV}
 flash_attn_fuse_mlp: ${FLASH_ATTN_FUSE_MLP}
 sdp_attention: ${SDP_ATTENTION}
 s2_attention: ${S2_ATTENTION}

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"

MANIFEST.in

@@ -2,4 +2,5 @@ include requirements.txt
 include README.md
 include LICENSE
 include src/setuptools_axolotl_dynamic_dependencies.py
+include src/axolotl/utils/chat_templates/templates/*.jinja
 recursive-include axolotl *.py

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">
@@ -25,40 +29,60 @@
 
 ## 🎉 Latest Updates
 
-- 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!
+- 2025/07:
+  - 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!
-- 2025/04: Llama 4 support has been added in Axolotl. See [examples](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/llama-4) to start training your own Llama 4 models with Axolotl's linearized version!
 - 2025/03: Axolotl has implemented Sequence Parallelism (SP) support. Read the [blog](https://huggingface.co/blog/axolotl-ai-co/long-context-with-sequence-parallelism-in-axolotl) and [docs](https://docs.axolotl.ai/docs/sequence_parallelism.html) to learn how to scale your context length when fine-tuning.
+
+<details>
+
+<summary>Expand older updates</summary>
+
+- 2025/06: Magistral with mistral-common tokenizer support has been added to Axolotl. See [examples](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/magistral) to start training your own Magistral models with Axolotl!
+- 2025/04: Llama 4 support has been added in Axolotl. See [examples](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/llama-4) to start training your own Llama 4 models with Axolotl's linearized version!
 - 2025/03: (Beta) Fine-tuning Multimodal models is now supported in Axolotl. Check out the [docs](https://docs.axolotl.ai/docs/multimodal.html) to fine-tune your own!
 - 2025/02: Axolotl has added LoRA optimizations to reduce memory usage and improve training speed for LoRA and QLoRA in single GPU and multi-GPU training (DDP and DeepSpeed). Jump into the [docs](https://docs.axolotl.ai/docs/lora_optims.html) to give it a try.
 - 2025/02: Axolotl has added GRPO support. Dive into our [blog](https://huggingface.co/blog/axolotl-ai-co/training-llms-w-interpreter-feedback-wasm) and [GRPO example](https://github.com/axolotl-ai-cloud/grpo_code) and have some fun!
 - 2025/01: Axolotl has added Reward Modelling / Process Reward Modelling fine-tuning support. See [docs](https://docs.axolotl.ai/docs/reward_modelling.html).
 
+</details>
+
 ## ✨ 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), LoRA optimizations, Multi-GPU training (FSDP1, FSDP2, DeepSpeed), Multi-node training (Torchrun, Ray), 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, and audio models like Voxtral with image, video, and audio support.
+- **Training Methods**: Full fine-tuning, LoRA, QLoRA, GPTQ, QAT, Preference Tuning (DPO, IPO, KTO, ORPO), RL (GRPO), and Reward Modelling (RM) / Process Reward Modelling (PRM).
+- **Easy Configuration**: Re-use a single YAML configuration file across the full fine-tuning pipeline: dataset preprocessing, training, evaluation, quantization, and inference.
+- **Performance Optimizations**: [Multipacking](https://docs.axolotl.ai/docs/multipack.html), [Flash Attention](https://github.com/Dao-AILab/flash-attention), [Xformers](https://github.com/facebookresearch/xformers), [Flex Attention](https://pytorch.org/blog/flexattention/), [Liger Kernel](https://github.com/linkedin/Liger-Kernel), [Cut Cross Entropy](https://github.com/apple/ml-cross-entropy/tree/main), [Sequence Parallelism (SP)](https://docs.axolotl.ai/docs/sequence_parallelism.html), [LoRA optimizations](https://docs.axolotl.ai/docs/lora_optims.html), [Multi-GPU training (FSDP1, FSDP2, DeepSpeed)](https://docs.axolotl.ai/docs/multi-gpu.html), [Multi-node training (Torchrun, Ray)](https://docs.axolotl.ai/docs/multi-node.html), and many more!
 - **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 ≥2.5.1
+- PyTorch ≥2.6.0
+
+### 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 --no-build-isolation axolotl[flash-attn,deepspeed]
@@ -68,8 +92,29 @@ axolotl fetch examples
 axolotl fetch deepspeed_configs  # OPTIONAL
 ```
 
+#### Using Docker
+
+Installing with Docker can be less error prone than installing in your own environment.
+```bash
+docker run --gpus '"all"' --rm -it axolotlai/axolotl:main-latest
+```
+
 Other installation approaches are described [here](https://docs.axolotl.ai/docs/installation.html).
 
+#### Cloud Providers
+
+<details>
+
+- [RunPod](https://runpod.io/gsc?template=v2ickqhz9s&ref=6i7fkpdz)
+- [Vast.ai](https://cloud.vast.ai?ref_id=62897&template_id=bdd4a49fa8bce926defc99471864cace&utm_source=github&utm_medium=developer_community&utm_campaign=template_launch_axolotl&utm_content=readme)
+- [PRIME Intellect](https://app.primeintellect.ai/dashboard/create-cluster?image=axolotl&location=Cheapest&security=Cheapest&show_spot=true)
+- [Modal](https://www.modal.com?utm_source=github&utm_medium=github&utm_campaign=axolotl)
+- [Novita](https://novita.ai/gpus-console?templateId=311)
+- [JarvisLabs.ai](https://jarvislabs.ai/templates/axolotl)
+- [Latitude.sh](https://latitude.sh/blueprint/989e0e79-3bf6-41ea-a46b-1f246e309d5c)
+
+</details>
+
 ### Your First Fine-tune
 
 ```bash
@@ -111,14 +156,22 @@ Contributions are welcome! Please see our [Contributing Guide](https://github.co
 
 ## ❤️ Sponsors
 
-Thank you to our sponsors who help make Axolotl possible:
-
-- [Modal](https://www.modal.com?utm_source=github&utm_medium=github&utm_campaign=axolotl) - Modal lets you run
-jobs in the cloud, by just writing a few lines of Python. Customers use Modal to deploy Gen AI models at large scale,
-fine-tune large language models, run protein folding simulations, and much more.
-
 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

_quarto.yml

@@ -35,25 +35,30 @@ quartodoc:
         - cli.train
         - cli.evaluate
         - cli.args
+        - cli.art
         - cli.checks
         - cli.config
+        - cli.delinearize_llama4
         - cli.inference
         - cli.merge_lora
         - cli.merge_sharded_fsdp_weights
         - cli.preprocess
-        - cli.sweeps
-        - cli.utils
+        - cli.quantize
         - cli.vllm_serve
         - cli.cloud.base
         - cli.cloud.modal_
-        - cli.quantize
+        - cli.utils
+        - cli.utils.args
+        - cli.utils.fetch
+        - cli.utils.load
+        - cli.utils.sweeps
+        - cli.utils.train
     - title: Trainers
       desc: Training implementations
       contents:
         - core.trainers.base
         - core.trainers.trl
         - core.trainers.mamba
-        - core.trainers.relora
         - core.trainers.dpo.trainer
         - core.trainers.grpo.trainer
         - core.trainers.grpo.sampler
@@ -148,7 +153,7 @@ quartodoc:
         - utils.distributed
         - utils.dict
         - utils.optimizers.adopt
-        - utils.data.pretraining
+        - utils.data.streaming
         - utils.data.sft
         - utils.quantization
     - title: Schemas
@@ -267,7 +272,10 @@ website:
           contents:
             - docs/batch_vs_grad.qmd
             - docs/dataset_preprocessing.qmd
+            - docs/streaming.qmd
             - docs/multipack.qmd
+            - docs/mixed_precision.qmd
+            - docs/optimizers.qmd
 
         - section: "Advanced Features"
           contents:
@@ -276,6 +284,8 @@ website:
             - docs/torchao.qmd
             - docs/custom_integrations.qmd
             - docs/sequence_parallelism.qmd
+            - docs/gradient_checkpointing.qmd
+            - docs/nd_parallelism.qmd
 
         - section: "Troubleshooting"
           contents:

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
+    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
 
 WORKDIR /workspace
 

cicd/Dockerfile.jinja

@@ -9,9 +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"
 
 RUN apt-get update && \
-    apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev
+    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
 
 WORKDIR /workspace
 

cicd/cicd.sh

@@ -51,3 +51,5 @@ pytest -v --durations=10 \
   --cov=axolotl \
   --cov-append \
   --cov-report=xml:e2e-coverage.xml
+
+codecov upload-process -t $CODECOV_TOKEN -f e2e-coverage.xml -F e2e,pytorch-${PYTORCH_VERSION} || true

cicd/e2e_tests.py

@@ -1,7 +1,5 @@
 """Modal app to run axolotl GPU tests"""
 
-import pathlib
-
 from .single_gpu import GPU_CONFIG, VOLUME_CONFIG, app, cicd_image, run_cmd
 
 
@@ -14,21 +12,9 @@ from .single_gpu import GPU_CONFIG, VOLUME_CONFIG, app, cicd_image, run_cmd
     volumes=VOLUME_CONFIG,
 )
 def cicd_pytest():
-
     run_cmd("./cicd/cicd.sh", "/workspace/axolotl")
 
-    # Read the coverage file if it exists
-    coverage_file = pathlib.Path("/workspace/axolotl/e2e-coverage.xml")
-    if coverage_file.exists():
-        return coverage_file.read_text(encoding="utf-8")
-    return None
-
 
 @app.local_entrypoint()
 def main():
-    coverage = cicd_pytest.remote()
-
-    # Save the coverage file to the local filesystem if it was generated
-    if coverage:
-        with open("e2e-coverage.xml", "w", encoding="utf-8") as f:
-            f.write(coverage)
+    cicd_pytest.remote()

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
@@ -24,9 +22,9 @@ df_template = template_env.get_template("Dockerfile.jinja")
 df_args = {
     "AXOLOTL_EXTRAS": os.environ.get("AXOLOTL_EXTRAS", ""),
     "AXOLOTL_ARGS": os.environ.get("AXOLOTL_ARGS", ""),
-    "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.5.1"),
-    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-py3.11-cu124-2.5.1"),
-    "CUDA": os.environ.get("CUDA", "124"),
+    "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.6.0"),
+    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-py3.11-cu126-2.6.0"),
+    "CUDA": os.environ.get("CUDA", "126"),
     "GITHUB_REF": os.environ.get("GITHUB_REF", "refs/heads/main"),
     "GITHUB_SHA": os.environ.get("GITHUB_SHA", ""),
     "CODECOV_TOKEN": os.environ.get("CODECOV_TOKEN", ""),
@@ -63,7 +61,7 @@ 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(
@@ -77,18 +75,7 @@ def run_cmd(cmd: str, run_folder: str):
 def cicd_pytest():
     run_cmd("./cicd/multigpu.sh", "/workspace/axolotl")
 
-    # Read the coverage file if it exists
-    coverage_file = pathlib.Path("/workspace/axolotl/multigpu-coverage.xml")
-    if coverage_file.exists():
-        return coverage_file.read_text(encoding="utf-8")
-    return None
-
 
 @app.local_entrypoint()
 def main():
-    coverage = cicd_pytest.remote()
-
-    # Save the coverage file to the local filesystem if it was generated
-    if coverage:
-        with open("multigpu-coverage.xml", "w", encoding="utf-8") as file:
-            file.write(coverage)
+    cicd_pytest.remote()

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 -n2 \
+pytest -v --durations=10 -n2 \
   --ignore=/workspace/axolotl/tests/e2e/multigpu/solo/ \
   --ignore=/workspace/axolotl/tests/e2e/multigpu/patched/ \
   /workspace/axolotl/tests/e2e/multigpu/ \
@@ -19,5 +19,7 @@ pytest -v  --durations=10 -n1 /workspace/axolotl/tests/e2e/multigpu/patched/ \
   --cov-append \
   --cov-report=xml:multigpu-coverage.xml
 
-# Upload coverage to Codecov
-codecov upload-process -t "${CODECOV_TOKEN}" -f multigpu-coverage.xml -F multigpu,docker-tests,pytorch-${PYTORCH_VERSION} || true
+# Upload coverage to Codecov if CODECOV_TOKEN is available
+if [ -n "$CODECOV_TOKEN" ]; then
+  codecov upload-process -t "${CODECOV_TOKEN}" -f multigpu-coverage.xml -F multigpu,docker-tests,pytorch-${PYTORCH_VERSION} || true
+fi

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
@@ -24,14 +22,16 @@ df_template = template_env.get_template(dockerfile)
 df_args = {
     "AXOLOTL_EXTRAS": os.environ.get("AXOLOTL_EXTRAS", ""),
     "AXOLOTL_ARGS": os.environ.get("AXOLOTL_ARGS", ""),
-    "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.5.1"),
-    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-py3.11-cu124-2.5.1"),
-    "CUDA": os.environ.get("CUDA", "124"),
+    "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.6.0"),
+    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-py3.11-cu126-2.6.0"),
+    "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)
@@ -57,12 +57,16 @@ 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"
+
     # Propagate errors from subprocess.
-    if exit_code := subprocess.call(cmd.split(), cwd=run_folder):  # nosec
-        exit(exit_code)  # pylint: disable=consider-using-sys-exit
+    if exit_code := subprocess.call(cmd.split(), cwd=run_folder, env=sp_env):  # nosec
+        exit(exit_code)

codecov.yml

@@ -12,7 +12,7 @@ coverage:
       default:
         # basic
         target: auto
-        threshold: 0%
+        threshold: 1%
         base: auto
         # advanced
         branches: null
@@ -22,11 +22,12 @@ coverage:
         only_pulls: true
         flags: null
         paths: null
+        informational: true
     patch:
       default:
         # basic
         target: auto
-        threshold: 0%
+        threshold: 1%
         base: auto
         # advanced
         branches: null

deepspeed_configs/zero3.json

@@ -7,9 +7,9 @@
     "reduce_bucket_size": "auto",
     "stage3_prefetch_bucket_size": "auto",
     "stage3_param_persistence_threshold": "auto",
-    "stage3_max_live_parameters": 0,
-    "stage3_max_reuse_distance": 0,
-    "stage3_gather_16bit_weights_on_model_save": true
+    "max_live_parameters": 0,
+    "max_reuse_distance": 0,
+    "gather_16bit_weights_on_model_save": true
   },
   "bf16": {
     "enabled": "auto"

deepspeed_configs/zero3_bf16.json

@@ -7,9 +7,9 @@
     "reduce_bucket_size": "auto",
     "stage3_prefetch_bucket_size": "auto",
     "stage3_param_persistence_threshold": "auto",
-    "stage3_max_live_parameters": 0,
-    "stage3_max_reuse_distance": 0,
-    "stage3_gather_16bit_weights_on_model_save": true
+    "max_live_parameters": 0,
+    "max_reuse_distance": 0,
+    "gather_16bit_weights_on_model_save": true
   },
   "bf16": {
     "enabled": true

deepspeed_configs/zero3_bf16_cpuoffload_all.json

@@ -17,9 +17,9 @@
     "reduce_bucket_size": "auto",
     "stage3_prefetch_bucket_size": "auto",
     "stage3_param_persistence_threshold": "auto",
-    "stage3_max_live_parameters": 0,
-    "stage3_max_reuse_distance": 0,
-    "stage3_gather_16bit_weights_on_model_save": true
+    "max_live_parameters": 0,
+    "max_reuse_distance": 0,
+    "gather_16bit_weights_on_model_save": true
   },
   "bf16": {
     "enabled": true

deepspeed_configs/zero3_bf16_cpuoffload_params.json

@@ -13,9 +13,9 @@
     "reduce_bucket_size": "auto",
     "stage3_prefetch_bucket_size": "auto",
     "stage3_param_persistence_threshold": "auto",
-    "stage3_max_live_parameters": 0,
-    "stage3_max_reuse_distance": 0,
-    "stage3_gather_16bit_weights_on_model_save": true
+    "max_live_parameters": 0,
+    "max_reuse_distance": 0,
+    "gather_16bit_weights_on_model_save": true
   },
   "bf16": {
     "enabled": true

docker/Dockerfile

@@ -10,7 +10,9 @@ ARG PYTORCH_VERSION="2.1.2"
 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
+    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
 
@@ -23,17 +25,17 @@ RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
         pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \
         pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray] $AXOLOTL_ARGS; \
-    fi
+    fi && \
+    python scripts/unsloth_install.py | sh && \
+    python scripts/cutcrossentropy_install.py | sh && \
+    pip install pytest && \
+    pip cache purge
 
-RUN python scripts/unsloth_install.py | sh
-RUN python scripts/cutcrossentropy_install.py | sh
-
-# So we can test the Docker image
-RUN pip install pytest
-
-# fix so that git fetch/pull from remote works
+# 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 --get remote.origin.fetch && \
+    git config --global credential.helper store
 
-# helper for huggingface-login cli
-RUN 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-base

@@ -16,12 +16,19 @@ ENV PYTHON_VERSION=$PYTHON_VERSION
 ENV TORCH_CUDA_ARCH_LIST=$TORCH_CUDA_ARCH_LIST
 
 RUN apt-get update \
-    && apt-get install -y wget git build-essential ninja-build git-lfs libaio-dev pkg-config && rm -rf /var/lib/apt/lists/* \
+    && apt-get install -y --no-install-recommends \
+        wget git build-essential ninja-build git-lfs libaio-dev pkg-config \
+        ibverbs-providers ibverbs-utils infiniband-diags  \
+        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 \
     && mkdir /root/.conda \
     && bash Miniconda3-latest-Linux-x86_64.sh -b \
     && rm -f Miniconda3-latest-Linux-x86_64.sh \
+    && conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/main \
+    && conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/r \
     && conda create -n "py${PYTHON_VERSION}" python="${PYTHON_VERSION}"
 
 ENV PATH="/root/miniconda3/envs/py${PYTHON_VERSION}/bin:${PATH}"
@@ -30,14 +37,16 @@ WORKDIR /workspace
 
 RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==23.2 setuptools==75.8.0 wheel && \
     python3 -m pip install --no-cache-dir -U torch==${PYTORCH_VERSION}+cu${CUDA} torchvision --extra-index-url https://download.pytorch.org/whl/cu$CUDA && \
-    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"
+    CAUSAL_CONV1D_FORCE_CXX11_ABI=TRUE CAUSAL_CONV1D_FORCE_BUILD=TRUE python3 -m pip install --no-cache-dir causal_conv1d==1.5.2 && \
+    python3 -m pip install --no-cache-dir "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main" && \
+    python3 -m pip cache purge
 
 RUN 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 install -U --no-cache-dir pydantic==1.10.10 && \
+    pip3 cache purge
 
-RUN if [ "$PYTORCH_VERSION" = "2.7.1" ] ; then \
-        pip3 install flash-attn==2.7.4.post1; \
+RUN if [ "$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

docker/Dockerfile-base-nightly

@@ -22,18 +22,22 @@ RUN apt-get update \
     && mkdir /root/.conda \
     && bash Miniconda3-latest-Linux-x86_64.sh -b \
     && rm -f Miniconda3-latest-Linux-x86_64.sh \
+    && conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/main \
+    && conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/r \
     && conda create -n "py${PYTHON_VERSION}" python="${PYTHON_VERSION}"
 
 ENV PATH="/root/miniconda3/envs/py${PYTHON_VERSION}/bin:${PATH}"
 
 WORKDIR /workspace
 
-RUN python3 -m pip install --upgrade pip && pip3 install packaging && \
+RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==23.2 setuptools==75.8.0 wheel && \
     python3 -m pip install --no-cache-dir -U torch --extra-index-url https://download.pytorch.org/whl/nightly/cu$CUDA && \
     python3 -m pip install --no-cache-dir "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main" && \
-    python3 -m pip install --no-cache-dir "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main"
+    python3 -m pip install --no-cache-dir "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main" && \
+    python3 -m pip cache purge
 
 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 install -U --no-cache-dir pydantic==1.10.10 && \
+    pip3 cache purge

docker/Dockerfile-cloud

@@ -14,7 +14,10 @@ COPY scripts/motd /etc/motd
 
 RUN pip install jupyterlab notebook ipywidgets && \
     jupyter lab clean
-RUN apt install --yes --no-install-recommends openssh-server tmux iproute2 nvtop && \
+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 && \

docker/Dockerfile-cloud-no-tmux

@@ -9,13 +9,15 @@ ENV HF_HUB_ENABLE_HF_TRANSFER="1"
 EXPOSE 8888
 EXPOSE 22
 
-COPY scripts/cloud-entrypoint-term.sh /root/cloud-entrypoint.sh
+COPY scripts/cloud-entrypoint.sh /root/cloud-entrypoint.sh
 COPY scripts/motd /etc/motd
 
 RUN pip install jupyterlab notebook ipywidgets && \
     jupyter lab clean
-RUN apt install --yes --no-install-recommends openssh-server tmux sudo && \
-    pip3 install -U --no-cache-dir grpcio ray[default]==2.9.3 && \
+RUN apt update && \
+    apt install --yes --no-install-recommends openssh-server tmux iproute2 nvtop ibverbs-providers ibverbs-utils infiniband-diags librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm && \
+    rm -rf /var/cache/apt/archives && \
+    rm -rf /var/lib/apt/lists/* && \
     mkdir -p ~/.ssh && \
     chmod 700 ~/.ssh && \
     printf "[ ! -z \"\$TERM\" -a -r /etc/motd ] && cat /etc/motd\n" >> ~/.bashrc && \

docker/Dockerfile-uv-base

@@ -34,7 +34,3 @@ RUN uv pip install packaging setuptools wheel psutil \
     && uv pip install --no-build-isolation "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main" \
     && uv pip install "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main" \
     && uv pip install awscli pydantic
-
-RUN if [ "$PYTORCH_VERSION" = "2.7.1" ] ; then \
-        uv pip install --no-build-isolation flash-attn==2.7.4.post1; \
-    fi

docs/cli.qmd

@@ -23,6 +23,20 @@ axolotl <command> [config.yml] [options]
 
 The config file can be local or a URL to a raw YAML file.
 
+### Launcher Arguments
+
+For commands that support multi-GPU (`train`, `evaluate`, ...), you can pass launcher-specific arguments using the `--` separator:
+
+```bash
+# Pass torchrun arguments
+axolotl train config.yml --launcher torchrun -- --nproc_per_node=2 --nnodes=1
+
+# Pass accelerate arguments
+axolotl train config.yml --launcher accelerate -- --config_file=accelerate_config.yml --num_processes=4
+```
+
+Arguments after `--` are passed directly to the launcher (torchrun, accelerate launch, etc.).
+
 ## Command Reference
 
 ### fetch
@@ -80,7 +94,11 @@ axolotl train config.yml \
     --num-epochs 3
 
 # Training without accelerate
-axolotl train config.yml --no-accelerate
+axolotl train config.yml --launcher python
+
+# Pass launcher-specific arguments using -- separator
+axolotl train config.yml --launcher torchrun -- --nproc_per_node=2 --nnodes=1
+axolotl train config.yml --launcher accelerate -- --config_file=accelerate_config.yml
 
 # Resume training from checkpoint
 axolotl train config.yml --resume-from-checkpoint path/to/checkpoint
@@ -175,6 +193,9 @@ Evaluates a model's performance (loss etc) on the train and eval datasets.
 ```bash
 # Basic evaluation
 axolotl evaluate config.yml
+
+# Evaluation with launcher arguments
+axolotl evaluate config.yml --launcher torchrun -- --nproc_per_node=2
 ```
 
 ### lm-eval
@@ -287,9 +308,6 @@ axolotl preprocess config.yml --cloud cloud_config.yml
 # Train on cloud
 axolotl train config.yml --cloud cloud_config.yml
 
-# Train without accelerate on cloud
-axolotl train config.yml --cloud cloud_config.yml --no-accelerate
-
 # Run lm-eval on cloud
 axolotl lm-eval config.yml --cloud cloud_config.yml
 ```

docs/custom_integrations.qmd

@@ -7,6 +7,7 @@ toc-depth: 3
 ```{python}
 #| echo: false
 
+import os
 import re
 
 def process_readme(integration_name):
@@ -53,6 +54,24 @@ sections = [
     ("LLMCompressor", "llm_compressor")
 ]
 
+for folder_name in os.listdir("../src/axolotl/integrations/"):
+    if folder_name in [path for name, path in sections]:
+        # skip if already in sections
+        continue
+    if os.path.exists(f"../src/axolotl/integrations/{folder_name}/README.md"):
+        # grab the first heading in README.md as the section name
+        with open(f"../src/axolotl/integrations/{folder_name}/README.md", "r") as f:
+            txt = f.read()
+            matches = re.search(r'^# (.*)\n?', txt, flags=re.MULTILINE)
+            if matches:
+                name = matches.group(1)
+            else:
+                continue
+            sections.append((name, folder_name))
+
+# sort sections by name
+sections = sorted(sections, key=lambda x: x[0])
+
 for section_name, folder_name in sections:
     print(print_section(section_name, folder_name))
 ```

docs/dataset-formats/conversation.qmd

@@ -9,7 +9,7 @@ order: 3
 Chat Template strategy uses a jinja2 template that converts a list of messages into a prompt. Support using tokenizer's template, a supported template, or custom jinja2.
 
 ```{.json filename="data.jsonl"}
-{"conversations": [{"role": "...", "content": "..."}]}
+{"messages": [{"role": "...", "content": "..."}, {"role": "...", "content": "..."}, ...]}
 ```
 
 See [configs](../config-reference.qmd) for full configs and supported templates.
@@ -187,6 +187,7 @@ Instead of passing `tools` via the system prompt, an alternative method would be
             "role": "assistant", // call the function via assistant
             "tool_calls": [
                 {
+                    "id": "...",  // required only for mistral
                     "type": "function",
                     "function": {
                         "name": "...",
@@ -199,6 +200,7 @@ Instead of passing `tools` via the system prompt, an alternative method would be
         },
         {
             "role": "tool",
+            "tool_call_id": "...",  // required only for mistral
             "name": "...",
             "content": "..."
         },
@@ -210,10 +212,11 @@ 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).
 :::
 
+Example config for Llama4:
 ```yaml
 chat_template: llama4
 datasets:
-  - path: ...
+  - path: Nanobit/text-tools-2k-test
     type: chat_template
     # field_tools: tools # default is `tools`
 ```

docs/docker.qmd

@@ -9,7 +9,7 @@ format:
 This section describes the different Docker images that are released by AxolotlAI at [Docker Hub](https://hub.docker.com/u/axolotlai).
 
 ::: {.callout-important}
-For Blackwell GPUs, please use the tags with Pytorch 2.7.1 and CUDA 12.8.
+For Blackwell GPUs, please use the tags with PyTorch 2.7.1 and CUDA 12.8.
 :::
 
 ## Base
@@ -34,8 +34,9 @@ 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-cu124-2.5.1`
 
 ## Main
 
@@ -73,13 +74,15 @@ 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-cu124-2.5.1`
 - `main-latest`
 - `main-20250303-py3.11-cu124-2.6.0`
-- `main-20250303-py3.11-cu124-2.5.1`
-- `0.9.2`
+- `main-20250303-py3.11-cu126-2.6.0`
+- `0.10.1`
 
 ## Cloud
 

docs/faq.qmd

@@ -9,11 +9,11 @@ description: Frequently asked questions
 
 > A: Usually an issue with the GPUs communicating with each other. See the [NCCL doc](nccl.qmd)
 
-**Q: Exitcode -9**
+**Q: exitcode: -9**
 
 > A: This usually happens when you run out of system RAM.
 
-**Q: Exitcode -7 while using deepspeed**
+**Q: exitcode: -7 while using deepspeed**
 
 > A: Try upgrading deepspeed w: `pip install -U deepspeed`
 
@@ -51,6 +51,18 @@ description: Frequently asked questions
 >   pad_token: "..."
 > ```
 
+**Q: `IterableDataset error` or `KeyError: 'input_ids'` when using `preprocess` CLI**
+
+> A: This is because you may be using `preprocess` CLI with `pretraining_dataset:` or `skip_prepare_dataset: true` respectively. Please use `axolotl train` CLI directly instead as these datasets are prepared on demand.
+
+**Q: vLLM is not working with Axolotl**
+
+> A: We currently recommend torch 2.6.0 for use with `vllm`. Please ensure you use the right version. For Docker, please use the `main-py3.11-cu124-2.6.0` tag.
+
+**Q: FA2 2.8.0 `undefined symbol` runtime error on CUDA 12.4**
+
+> 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.
+
 ### Chat templates
 
 **Q: `jinja2.exceptions.UndefinedError: 'dict object' has no attribute 'content' / 'role' / ____`**
@@ -124,3 +136,7 @@ description: Frequently asked questions
 >   dynamic: false
 >   mode: max-autotune-no-cudagraphs
 > ```
+
+**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.

docs/fsdp_qlora.qmd

@@ -20,7 +20,7 @@ To enable `QLoRA` with `FSDP`, you need to perform the following steps:
 > See the [example config](#example-config) file in addition to reading these instructions.
 
 1. Set `adapter: qlora` in your axolotl config file.
-2. Enable FSDP in your axolotl config, as [described here](https://github.com/axolotl-ai-cloud/axolotl?tab=readme-ov-file#fsdp).
+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`.
 
 ## Example Config

docs/gradient_checkpointing.qmd

@@ -0,0 +1,29 @@
+---
+title: Gradient Checkpointing and Activation Offloading
+---
+
+Gradient checkpointing and activation offloading are techniques used to optimize the performance of deep learning
+models by reducing the memory footprint and improving computational efficiency.
+
+### Enabling Gradient Checkpointing
+
+```yaml
+gradient_checkpointing: true
+```
+
+### Enabling Activation Offloading
+
+```yaml
+gradient_checkpointing: true  # required for activation offloading
+activation_offloading: true
+```
+
+Activation offloading variants:
+
+The default `activation_offloading: true` offloads activations to CPU and uses CUDA streams
+to overlap the communications and computations when offloading.
+
+The `activation_offloading: legacy` naively offloads activations to CPU and without additional optimizations.
+
+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.

docs/installation.qmd

@@ -15,7 +15,7 @@ This guide covers all the ways you can install and set up Axolotl for your envir
 
 - NVIDIA GPU (Ampere architecture or newer for `bf16` and Flash Attention) or AMD GPU
 - Python ≥3.11
-- PyTorch ≥2.5.1
+- PyTorch ≥2.6.0
 
 ## Installation Methods {#sec-installation-methods}
 
@@ -124,14 +124,17 @@ For providers supporting Docker:
 
 - Use `axolotlai/axolotl-cloud:main-latest`
 - Available on:
-  - [Latitude.sh](https://latitude.sh/blueprint/989e0e79-3bf6-41ea-a46b-1f246e309d5c)
-  - [JarvisLabs.ai](https://jarvislabs.ai/templates/axolotl)
-  - [RunPod](https://runpod.io/gsc?template=v2ickqhz9s&ref=6i7fkpdz)
-  - [Novita](https://novita.ai/gpus-console?templateId=311)
+    - [RunPod](https://runpod.io/gsc?template=v2ickqhz9s&ref=6i7fkpdz)
+    - [Vast.ai](https://cloud.vast.ai?ref_id=62897&template_id=bdd4a49fa8bce926defc99471864cace&utm_source=axolotl&utm_medium=partner&utm_campaign=template_launch_july2025&utm_content=docs_link)
+    - [PRIME Intellect](https://app.primeintellect.ai/dashboard/create-cluster?image=axolotl&location=Cheapest&security=Cheapest&show_spot=true)
+    - [Modal](https://www.modal.com?utm_source=github&utm_medium=github&utm_campaign=axolotl)
+    - [Novita](https://novita.ai/gpus-console?templateId=311)
+    - [JarvisLabs.ai](https://jarvislabs.ai/templates/axolotl)
+    - [Latitude.sh](https://latitude.sh/blueprint/989e0e79-3bf6-41ea-a46b-1f246e309d5c)
 
 ### 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}
 

docs/mixed_precision.qmd

@@ -0,0 +1,149 @@
+---
+title: "Mixed Precision Training"
+format:
+  html:
+    toc: true
+    toc-depth: 3
+    number-sections: true
+    code-tools: true
+execute:
+  enabled: false
+---
+
+Mixed precision training uses lower precision data types to reduce memory usage and increase training speed while maintaining model quality. Axolotl supports several mixed precision formats:
+
+- **FP16** - Half precision 16-bit (Pascal generation+)
+- **BF16** - Brain Float 16-bit (Ampere generation+)
+- **FP8** - 8-bit floating point (Hopper generation+)
+
+## FP16 Mixed Precision {#sec-fp16}
+
+### Overview {#sec-fp16-overview}
+
+FP16 is the traditional half-precision format, supported on older GPUs but can be less numerically stable than BF16.
+
+### Configuration {#sec-fp16-config}
+
+```{.yaml}
+fp16: true
+```
+
+### FP16 Considerations {#sec-fp16-considerations}
+
+- May require gradient scaling to prevent underflow
+- Less numerically stable than BF16
+- Can cause training instability with some model architectures
+- Consider using BF16 if your hardware supports it
+
+## BF16 Mixed Precision {#sec-bf16}
+
+### Overview {#sec-bf16-overview}
+
+BF16 (Brain Float 16) offers better numerical stability than FP16 and is the recommended mixed precision format for modern GPUs. It provides the same dynamic range as FP32 while using half the memory.
+
+### Configuration {#sec-bf16-config}
+
+```{.yaml}
+# Automatic BF16 detection (recommended)
+bf16: auto
+
+# Or explicitly enable
+bf16: true
+
+# For evaluation with BF16
+bf16: full  # Equivalent to bf16_full_eval in the HF trainer
+```
+
+## FP8 Mixed Precision {#sec-fp8}
+
+::: {.callout-note}
+FP8 support is experimental and requires compatible hardware (H100, H200) and recent PyTorch versions with TorchAO.
+:::
+
+### What is FP8? {#sec-fp8-overview}
+
+FP8 (8-bit floating point) can provide significant time savings compared to FP16/BF16 while maintaining training stability. Axolotl's implementation uses PyTorch's TorchAO library with "tensorwise" scaling strategy.
+
+### Requirements {#sec-fp8-software}
+
+- Hopper+ GPUs (H100/H200)
+- PyTorch 2.7+ (+ compatible TorchAO version)
+- CUDA 12.4+
+
+### Configuration {#sec-fp8-config}
+
+Add to your YAML config:
+
+```{.yaml}
+# Enable FP8 mixed precision
+fp8: true
+
+# Optional: Enable FP8 for FSDP all-gather operations
+fp8_enable_fsdp_float8_all_gather: true
+
+# Enable torch.compile (almost always necessary for FP8 speedups)
+torch_compile: true
+```
+
+::: {.callout-important}
+**torch.compile is critical for FP8 performance**
+
+FP8 training requires `torch_compile: true` to see meaningful speedups. Without compilation, FP8 may actually be slower and use more memory than FP16/BF16.
+:::
+
+### Advanced FP8 Configs {#sec-fp8-advanced}
+
+For [FSDP](multi-gpu.qmd#sec-fsdp) (Fully Sharded Data Parallel) training:
+
+```{.yaml}
+fp8: true
+fp8_enable_fsdp_float8_all_gather: true
+
+torch_compile: true
+
+# FSDP configuration
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  reshard_after_forward: true
+```
+
+## Best Practices {#sec-best-practices}
+
+### Choosing Precision Format {#sec-choosing-format}
+
+- **Start with automatic detection**: `bf16: auto`
+- **For Hopper+ (H100/H200)**: Try FP8 + torch.compile for maximum speed
+- **For Ampere (A100/RTX 30/40)**: Use BF16
+- **For older Pascal/Turing GPUs**: Use FP16 with caution
+- **For very old or unsupported GPUs**: Use FP32
+
+### Validation and Testing {#sec-validation}
+
+Always validate your mixed precision setup:
+
+- **Start with a small dataset** to verify stability
+- **Monitor loss curves** for irregularities
+- **Compare with FP32 baseline** when possible
+- **Test evaluation metrics** match expectations
+
+### FP8 Particulars {#sec-fp8-details}
+
+- Use cases
+  - Single GPU training
+  - Multi GPU training with FSDP2 or Deepspeed
+- Speedups
+  - Please refer to the [TorchAO FP8 training benchmarks](https://github.com/pytorch/ao/tree/main/torchao/float8#rowwise-scaling) for expected matmul speedups for different (M, K, N) settings
+  - Concrete number for LLaMA 3 8B training can be found [here](https://github.com/pytorch/ao/tree/main/torchao/float8#training-benchmarks)
+- Known issues:
+  - FP8 + DDP + `torch.compile` (causes [error](https://gist.github.com/djsaunde/0c1664c32e44a64d31b5e01b4aafe5c4))
+  - FP8 + FSDP2 + `torch.compile` + FSDP2 activation checkpointing tends to be _slower_ than the BF16 equivalent training
+  - Flash Attention 2 does not play nicely with `torch.compile`
+
+See `examples/llama-3/3b-fp8-fsdp2.yaml` for an optimized example config. Enabling FP8 mixed precision + FP8 all-gather training results in ~10% faster iterations per second vs. BF16 for a relatively small (3B param) model
+
+For more information on multi-GPU training, see our [Multi-GPU guide](multi-gpu.qmd).

docs/multi-gpu.qmd

@@ -23,8 +23,6 @@ Axolotl supports several methods for multi-GPU training:
 
 ## DeepSpeed {#sec-deepspeed}
 
-DeepSpeed is the recommended approach for multi-GPU training due to its stability and performance. It provides various optimization levels through ZeRO stages.
-
 ### Configuration {#sec-deepspeed-config}
 
 Add to your YAML config:
@@ -32,7 +30,6 @@ Add to your YAML config:
 ```{.yaml}
 deepspeed: deepspeed_configs/zero1.json
 ```
-
 ### Usage {#sec-deepspeed-usage}
 
 ```{.bash}
@@ -66,9 +63,66 @@ Start from Stage 1 -> Stage 2 -> Stage 3.
 
 :::
 
-## FSDP {#sec-fsdp}
+## Fully Sharded Data Parallel (FSDP) {#sec-fsdp}
 
-### Basic FSDP Configuration {#sec-fsdp-config}
+::: {.callout-note}
+
+FSDP2 is recommended for new users. FSDP1 is deprecated and will be removed in an upcoming release of Axolotl.
+
+:::
+
+### 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
+also follow the config field mapping below to update field names.
+
+#### Config mapping
+
+FSDP1 | FSDP2
+-------- | --------
+fsdp_sharding_strategy | reshard_after_forward
+fsdp_backward_prefetch_policy | **REMOVED**
+fsdp_backward_prefetch | **REMOVED**
+fsdp_forward_prefetch | **REMOVED**
+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**
+
+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:
+
+```{.yaml}
+fsdp_version: 1
+fsdp_config:
+  fsdp_offload_params: false
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: Qwen3DecoderLayer
+  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_sharding_strategy: FULL_SHARD
+```
+
+You can migrate to the following FSDP2 config:
+
+```{.yaml}
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Qwen3DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  reshard_after_forward: true
+```
+
+### FSDP1 (deprecated) {#sec-fsdp-config}
+
+::: {.callout-note}
+
+Using `fsdp` to configure FSDP is deprecated and will be removed in an upcoming release of Axolotl. Please use `fsdp_config` as above instead.
+
+:::
 
 ```{.yaml}
 fsdp:
@@ -80,6 +134,7 @@ fsdp_config:
   fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
 ```
 
+
 ## Sequence parallelism {#sec-sequence-parallelism}
 
 We support sequence parallelism (SP) via the

docs/multi-node.qmd

@@ -40,13 +40,13 @@ use_cpu: false
 
 Configure your model to use FSDP in the Axolotl yaml. For example:
 ```yaml
-fsdp:
-  - full_shard
-  - auto_wrap
+fsdp_version: 2
 fsdp_config:
-  fsdp_offload_params: true
-  fsdp_state_dict_type: FULL_STATE_DICT
-  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  offload_params: true
+  state_dict_type: FULL_STATE_DICT
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  reshard_after_forward: true
 ```
 
 All you have to do now is launch using accelerate as you would usually do on each machine and voila, the processes will start once you have launched accelerate on every machine.
@@ -69,11 +69,19 @@ export NCCL_BUFFSIZE=2097152
 
 Run the following on each node:
 
+### Option 1: New Axolotl CLI with launcher args (Recommended)
+
+```bash
+axolotl train config.yaml --launcher torchrun -- --nnodes $num_nodes --nproc_per_node $gpu_per_node --rdzv_id $rdzv_id --rdzv_backend c10d --rdzv_endpoint "$head_node_ip:$head_node_port"
+```
+
+### Option 2: Direct torchrun (Legacy)
+
 ```bash
 torchrun --nnodes $num_nodes --nproc_per_node $gpu_per_node --rdzv_id $rdzv_id --rdzv_backend c10d --rdzv_endpoint "$head_node_ip:$head_node_port" -m axolotl.cli.train config.yaml
 ```
 
-Please make sure to substitute the placeholder variables.
+Please make sure to substitute the placeholder variables:
 
 - `num_nodes`: Number of nodes (containing GPUs)
 - `gpu_per_node`: Number of gpus per node
@@ -81,8 +89,6 @@ Please make sure to substitute the placeholder variables.
 - `head_node_port`: Port of the head node (make sure other machines can connect to this. Default 29400)
 - `rdzv_id`: A unique job ID that is used by the job across nodes.
 
-::: {.callout-note}
-You need to call `axolotl.cli.train` instead of `axolotl train` as the latter calls accelerate under the hood
-:::
+The new CLI approach (Option 1) is recommended as it provides consistent argument handling and works seamlessly with other Axolotl CLI features.
 
 More info on the available configs can be found on the Pytorch docs [here](https://pytorch.org/docs/stable/elastic/run.html)

docs/multimodal.qmd

@@ -13,9 +13,14 @@ format:
 - [Pixtral](#sec-pixtral)
 - [Llava-1.5](#sec-llava-15)
 - [Mistral-Small-3.1](#sec-mistral-small-31)
+- [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)
+- [SmolVLM2](#sec-smolvlm2)
+- [LFM2-VL](#sec-lfm2-vl)
 
 ## Usage
 
@@ -30,14 +35,13 @@ 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
@@ -90,10 +94,32 @@ 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
+### 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
 ```
 
 ### Gemma-3 {#sec-gemma-3}
@@ -110,6 +136,22 @@ base_model: google/gemma-3-4b-it
 chat_template: gemma3
 ```
 
+### Gemma-3n {#sec-gemma-3n}
+
+::: {.callout-warning}
+The model's initial loss and grad norm will be very high. We suspect this to be due to the Conv in the vision layers.
+:::
+
+::: {.callout-tip}
+Please make sure to install `timm` via `pip3 install timm==1.0.17`
+:::
+
+```yaml
+base_model: google/gemma-3n-E2B-it
+
+chat_template: gemma3n
+```
+
 ### Qwen2-VL {#sec-qwen2-vl}
 
 ```yaml
@@ -126,13 +168,35 @@ base_model: Qwen/Qwen2.5-VL-7B-Instruct
 chat_template: qwen2_vl  # same as qwen2-vl
 ```
 
+### 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
+```
+
 ## Dataset Format
 
 For multi-modal datasets, we adopt an extended `chat_template` format similar to OpenAI's Message format.
 
 - A message is a list of `role` and `content`.
 - `role` can be `system`, `user`, `assistant`, etc.
-- `content` is a list of `type` and (`text` or `image` or `path` or `url` or `base64`).
+- `content` is a list of `type` and (`text`, `image`, `path`, `url`, `base64`, or `audio`).
+
+### Image
 
 ::: {.callout-note}
 For backwards compatibility:
@@ -141,15 +205,43 @@ For backwards compatibility:
 - If `content` is a string, it will be converted to a list with `type` as `text`.
 :::
 
-::: {.callout-tip}
 For image loading, you can use the following keys within `content` alongside `"type": "image"`:
 
 - `"path": "/path/to/image.jpg"`
 - `"url": "https://example.com/image.jpg"`
 - `"base64": "..."`
 - `"image": PIL.Image`
+
+### Audio
+
+For audio loading, you can use the following keys within `content` alongside `"type": "audio"`:
+
+- `"path": "/path/to/audio.mp3"`
+- `"url": "https://example.com/audio.mp3"`
+- `"audio": np.ndarray`
+
+::: {.callout-tip}
+
+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:
 ```json
 [
@@ -178,3 +270,9 @@ Here is an example of a multi-modal dataset:
   }
 ]
 ```
+
+## FAQ
+
+1. `PIL.UnidentifiedImageError: cannot identify image file ...`
+
+`PIL` could not retrieve the file at `url` using `requests`. Please check for typo. One alternative reason is that the request is blocked by the server.

docs/nd_parallelism.qmd

@@ -0,0 +1,108 @@
+---
+title: "N-D Parallelism (Beta)"
+---
+
+Axolotl enables training models at scale by composing different parallelism techniques. This is essential when:
+
+- A model's weights are too large to fit on a single GPU's memory.
+- A model's activations, especially with very long contexts, are too large for a single GPU.
+- You want to accelerate training by using multiple GPUs or nodes.
+
+or combinations of the above!
+
+## Core Concepts
+
+Parallelism strategies can be combined. The key is understanding how each one divides the workload. PyTorch's `DeviceMesh` is the modern way to manage these combinations, creating a logical grid of your GPUs and assigning different parallel strategies to different dimensions of the grid.
+
+### Data Parallelism {#sec-dp}
+
+Data Parallelism focuses on splitting the global data batch across GPUs.
+
+- Distributed Data Parallel (DDP): The classic approach. The full model is replicated on every GPU. Each GPU processes a different slice of the data batch. Gradients are then averaged across all GPUs after the backward pass to keep the models synchronized. This can substantially improve data throughput compared to single-device training, but requires that each GPU is able to hold the entire model, its gradients, and optimizer states.
+
+- [Fully Sharded Data Parallel (FSDP)](multi-gpu.qmd#fully-sharded-data-parallel-(fsdp)): A highly memory-efficient form of data parallelism (inspired by DeepSpeed's ZeRO). Instead of replicating the model, FSDP shards the model's *parameters, gradients, and optimizer states* across the GPUs in the data-parallel group. During computation, each GPU receives the specific parameters it needs via an `all_gather` operation just before they are used, and they can be discarded immediately after (`reshard-after-forward`).
+    - FSDP maps to ZeRO stages:
+        - ZeRO-2 (`reshard_after_forward=False`): Shards gradients and optimizer states. Model weights are replicated on each GPU.
+        - ZeRO-3 (`reshard_after_forward=True`): Shards gradients, optimizer states, AND model parameters. This provides the most memory savings at the cost of more communication (re-gathering parameters for both forward and backward passes).
+
+### [Experimental] Tensor Parallelism (TP) {#sec-tp}
+
+Also known as "horizontal model parallelism," as described in the [Megatron-LM paper](https://arxiv.org/pdf/1909.08053.pdf). Instead of splitting the batch, TP splits the model's layers themselves across GPUs.
+
+- How it works: For a linear layer `Y = XA`, the weight matrix `A` is split column-wise (`A = [A_1, A_2]`). The computation becomes `Y_1 = XA_1` and `Y_2 = XA_2`, which can happen in parallel on different GPUs. The final output `Y` is simply the concatenation of `Y_1` and `Y_2`. Check [this comment](https://github.com/huggingface/transformers/issues/10321#issuecomment-783543530) for more detailed info.
+- Requirement: TP involves frequent, small communications within a forward/backward pass. It requires a very fast interconnect between GPUs (e.g., NVLink) and is typically not recommended across different nodes.
+
+### Context Parallelism (CP) {#sec-cp}
+
+Context Parallelism, also called [Sequence Parallelism](sequence_parallelism.qmd), addresses the memory bottleneck from long sequences. The input sequence itself is split along the sequence length dimension and distributed across GPUs.
+
+- How it works: If you have a sequence of 8192 tokens and a `context_parallel_size` of 4, each GPU will only handle a chunk of 2048 tokens.
+- The Challenge: Attention is not local; every token needs to "attend to" every other token. Splitting the sequence breaks this.
+- The Solution (`ring-flash-attention`): An efficient communication protocol is used. To compute attention for its local sequence chunk, each GPU passes its Key-Value (KV) cache to its neighbor in a "ring." After `N-1` steps, every GPU has seen the KV-cache from all other GPUs, allowing it to compute the correct attention values for its chunk. This is implemented using the highly optimized `flash-attention` kernel at each step.
+
+### Hybrid Sharding Data Parallel (HSDP) {#sec-hsdp}
+
+HSDP is a 2D strategy that intelligently combines FSDP and DDP, typically for multi-node training.
+
+- Intra-Node (within a machine): Use FSDP. This is efficient because GPUs on the same node have fast interconnects (NVLink), making the `all_gather` operations for sharded parameters fast.
+- Inter-Node (across machines): Use DDP. The gradient synchronization between nodes is less frequent than FSDP's parameter gathering, making it a better fit for the slower node-to-node network (e.g., Ethernet/Infiniband).
+- Example: With 2 nodes of 8 GPUs each (16 total), you could have `dp_shard_size=8` (FSDP within each node) and `dp_replicate_size=2` (DDP across the two nodes).
+
+## Usage
+
+```yaml
+# FSDP config. See https://docs.axolotl.ai/docs/multi-gpu.html#sec-fsdp
+fsdp_version: 2
+fsdp_config:
+  # ...
+
+# The number of GPUs to shard the model parameters across (FSDP dimension).
+dp_shard_size: 4
+
+# The number of times to replicate the sharded model (DDP dimension).
+dp_replicate_size: 2
+
+# Number of GPUs for Tensor Parallelism.
+tensor_parallel_size: 1  # (default is 1, no TP)
+
+# Number of GPUs for Context/Sequence Parallelism.
+context_parallel_size: 1 # (default is 1, no CP)
+```
+
+Note: We recommend FSDP. DeepSpeed is only compatible with `tensor_parallel_size`.
+
+## Examples
+
+::: {.callout-tip}
+See our example configs [here](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/distributed-parallel).
+:::
+
+1.  HSDP on 2 nodes with 4 GPUs each (8 GPUs total):
+    - You want FSDP within each node and DDP across nodes.
+    - Set `dp_shard_size: 4` and `dp_replicate_size: 2`.
+
+2.  FSDP + TP on a single 8-GPU node:
+    - You want to split the model across 4 GPUs using FSDP, and further split each layer across 2 GPUs with TP.
+    - Set `dp_shard_size: 4` and `tensor_parallel_size: 2`.
+
+3.  FSDP + CP on a single 8-GPU node for long context:
+    - You want to shard the model across all 8 GPUs and also split the sequence length across all 8 GPUs.
+    - Set `dp_shard_size: 8` and `context_parallel_size: 8`. Note: this means the data parallel group and context parallel group are the same. A more common setup might be to shard across a smaller group.
+
+## Support Matrix
+
+This matrix describes how different parallelism methods can be combined in Axolotl.
+
+| Combination | `dp_replicate_size` | `dp_shard_size` | `tp_size` | `cp_size` | Status & Notes |
+| --- | :---: | :---: |:---:|:---:|---|
+| **FSDP** (ZeRO-3) | 1 | >1 | 1 | 1 | ✅ Fully supported. Shards model across all GPUs. |
+| **HSDP** | >1 | >1 | 1 | 1 | ✅ Fully supported. FSDP intra-node, DDP inter-node. |
+| **FSDP + TP** | 1 | >1 | >1 | 1 | ✅ **2D Parallelism**. Shards the model across a `dp_shard` group, and TP-splits layers within the `tp` group. |
+| **HSDP + TP** | >1 | >1 | >1 | 1 | ✅ **3D Parallelism**. A powerful but complex combination. |
+| **FSDP + CP** | 1 | >1 | 1 | >1 | ✅ **2D Parallelism**. Combines FSDP with context parallelism. |
+| **FSDP + TP + CP**| 1 | >1 | >1| >1| ✅ **3D Parallelism**. Another advanced combination. |
+| DDP + TP/CP | >1 | 1 | >1 | >1 | ❌ **Not Supported**. The `ParallelismConfig` explicitly prevents this, as composing pure DDP with TP or CP is currently not supported. You should use FSDP + TP/CP instead (`dp_shard_size > 1`). |
+| Just TP / CP | 1 | 1 | >1 | >1 | ✅ Supported. Useful for inference or when the model fits on one GPU but context is too long. |
+
+- `tp_size` refers to `tensor_parallel_size`
+- `cp_size` refers to `context_parallel_size`

docs/optimizers.qmd

@@ -0,0 +1,129 @@
+---
+title: Optimizers
+description: Configuring optimizers
+---
+
+## Overview
+
+Axolotl supports all optimizers supported by [transformers OptimizerNames](https://github.com/huggingface/transformers/blob/51f94ea06d19a6308c61bbb4dc97c40aabd12bad/src/transformers/training_args.py#L142-L187)
+
+Here is a list of optimizers supported by transformers as of `v4.54.0`:
+
+- `adamw_torch`
+- `adamw_torch_fused`
+- `adamw_torch_xla`
+- `adamw_torch_npu_fused`
+- `adamw_apex_fused`
+- `adafactor`
+- `adamw_anyprecision`
+- `adamw_torch_4bit`
+- `adamw_torch_8bit`
+- `ademamix`
+- `sgd`
+- `adagrad`
+- `adamw_bnb_8bit`
+- `adamw_8bit`  # alias for adamw_bnb_8bit
+- `ademamix_8bit`
+- `lion_8bit`
+- `lion_32bit`
+- `paged_adamw_32bit`
+- `paged_adamw_8bit`
+- `paged_ademamix_32bit`
+- `paged_ademamix_8bit`
+- `paged_lion_32bit`
+- `paged_lion_8bit`
+- `rmsprop`
+- `rmsprop_bnb`
+- `rmsprop_bnb_8bit`
+- `rmsprop_bnb_32bit`
+- `galore_adamw`
+- `galore_adamw_8bit`
+- `galore_adafactor`
+- `galore_adamw_layerwise`
+- `galore_adamw_8bit_layerwise`
+- `galore_adafactor_layerwise`
+- `lomo`
+- `adalomo`
+- `grokadamw`
+- `schedule_free_radam`
+- `schedule_free_adamw`
+- `schedule_free_sgd`
+- `apollo_adamw`
+- `apollo_adamw_layerwise`
+- `stable_adamw`
+
+
+## Custom Optimizers
+
+Enable custom optimizers by passing a string to the `optimizer` argument. Each optimizer will receive beta and epsilon args, however, some may accept additional args which are detailed below.
+
+### optimi_adamw
+
+```yaml
+optimizer: optimi_adamw
+```
+
+### ao_adamw_4bit
+
+Deprecated: Please use `adamw_torch_4bit`.
+
+### ao_adamw_8bit
+
+Deprecated: Please use `adamw_torch_8bit`.
+
+### ao_adamw_fp8
+
+
+```yaml
+optimizer: ao_adamw_fp8
+```
+
+### adopt_adamw
+
+GitHub: [https://github.com/iShohei220/adopt](https://github.com/iShohei220/adopt)
+Paper: [https://arxiv.org/abs/2411.02853](https://arxiv.org/abs/2411.02853)
+
+```yaml
+optimizer: adopt_adamw
+```
+
+### came_pytorch
+
+GitHub: [https://github.com/yangluo7/CAME/tree/master](https://github.com/yangluo7/CAME/tree/master)
+Paper: [https://arxiv.org/abs/2307.02047](https://arxiv.org/abs/2307.02047)
+
+```yaml
+optimizer: came_pytorch
+
+# optional args (defaults below)
+adam_beta1: 0.9
+adam_beta2: 0.999
+adam_beta3: 0.9999
+adam_epsilon: 1e-30
+adam_epsilon2: 1e-16
+```
+
+### muon
+
+Blog: [https://kellerjordan.github.io/posts/muon/](https://kellerjordan.github.io/posts/muon/)
+Paper: [https://arxiv.org/abs/2502.16982v1](https://arxiv.org/abs/2502.16982v1)
+
+```yaml
+optimizer: muon
+```
+
+### dion
+
+Microsoft's Dion (DIstributed OrthoNormalization) optimizer is a scalable and communication-efficient
+orthonormalizing optimizer that uses low-rank approximations to reduce gradient communication.
+
+GitHub: [https://github.com/microsoft/dion](https://github.com/microsoft/dion)
+Paper: [https://arxiv.org/pdf/2504.05295](https://arxiv.org/pdf/2504.05295)
+Note: Implementation written for PyTorch 2.7+ for DTensor
+
+```yaml
+optimizer: dion
+dion_lr: 0.01
+dion_momentum: 0.95
+lr: 0.00001  # learning rate for embeddings and parameters that fallback to AdamW
+```

docs/qat.qmd

@@ -23,10 +23,17 @@ 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

@@ -17,7 +17,6 @@ feedback. Various methods include, but not limited to:
 - [Kahneman-Tversky Optimization (KTO)](#kto)
 - [Odds Ratio Preference Optimization (ORPO)](#orpo)
 - [Group Relative Policy Optimization (GRPO)](#grpo)
-- Proximal Policy Optimization (PPO) (not yet supported in axolotl, if you're interested in contributing, please reach out!)
 
 
 ## RLHF using Axolotl
@@ -275,15 +274,14 @@ rl: dpo
 datasets:
   - path: ...
     split: train
-    type: user_defined.default
-
-    field_prompt: "prompt"
-    field_system: "system"
-    field_chosen: "chosen"
-    field_rejected: "rejected"
-    prompt_format: "{prompt}"
-    chosen_format: "{chosen}"
-    rejected_format: "{rejected}"
+    type:
+      field_prompt: "prompt"
+      field_system: "system"
+      field_chosen: "chosen"
+      field_rejected: "rejected"
+      prompt_format: "{prompt}"
+      chosen_format: "{chosen}"
+      rejected_format: "{rejected}"
 ```
 
 The input format is a simple JSON input with customizable fields based on the above config.
@@ -476,14 +474,13 @@ rl: kto
 datasets:
   - path: ...
     split: train
-    type: user_defined.default
-
-    field_prompt: "prompt"
-    field_system: "system"
-    field_completion: "completion"
-    field_label: "label"
-    prompt_format: "{prompt}"
-    completion_format: "{completion}"
+    type:
+      field_prompt: "prompt"
+      field_system: "system"
+      field_completion: "completion"
+      field_label: "label"
+      prompt_format: "{prompt}"
+      completion_format: "{completion}"
 ```
 
 The input format is a simple JSON input with customizable fields based on the above config.

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/sequence_parallelism.qmd

@@ -22,7 +22,7 @@ To enable sequence parallelism, add the following to your configuration file:
 
 ```yaml
 # Set to a divisor (> 1) of the number of GPUs available
-sequence_parallel_degree: 4  # Split sequences across 4 GPUs
+context_parallel_size: 4  # Split sequences across 4 GPUs
 # Optional; strides across the key dimension. Larger values use more memory but should make training faster.
 heads_k_stride: 1
 # Optional; one of "varlen_llama3" or "batch_ring". Defaults to
@@ -30,7 +30,7 @@ heads_k_stride: 1
 ring_attn_func:
 ```
 
-The `sequence_parallel_degree` should be a divisor of the total number of GPUs. For example:
+The `context_parallel_size` should be a divisor of the total number of GPUs. For example:
 
 - With 8 GPUs, valid values would be 2, 4, or 8
 - With 4 GPUs, valid values would be 2 or 4
@@ -66,7 +66,7 @@ sequence_len: 8192
 
 ...
 
-sequence_parallel_degree: 4  # Split each sequence into 4 parts, one per GPU
+context_parallel_size: 4  # Split each sequence into 4 parts, one per GPU
 # Optional; strides across the key dimension. Larger values use more memory but should make training faster.
 heads_k_stride: 1
 # Optional; one of "varlen_llama3" or "batch_ring". Defaults to
@@ -89,12 +89,12 @@ Sequence parallelism is compatible with Axolotl's sample packing functionality.
 
 ## Effect on Batch Size
 
-When using sequence parallelism, your effective global batch size is **divided** by the `sequence_parallel_degree`. This happens because:
+When using sequence parallelism, your effective global batch size is **divided** by the `context_parallel_size`. This happens because:
 
-- Each group of `sequence_parallel_degree` GPUs works on the same batch (just different parts of each sequence)
+- Each group of `context_parallel_size` GPUs works on the same batch (just different parts of each sequence)
 - The number of batches processed per step decreases
 
 For example:
 - With 8 GPUs and no sequence parallelism: 8 different batches processed per step
-- With 8 GPUs and `sequence_parallel_degree=4`: Only 2 different batches processed per step (each split across 4 GPUs)
+- With 8 GPUs and `context_parallel_size=4`: Only 2 different batches processed per step (each split across 4 GPUs)
 - If your per-GPU `micro_batch_size` is 2, the global batch size decreases from 16 to 4

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

examples/LiquidAI/README.md

@@ -0,0 +1,58 @@
+# 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.
+
+## 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
+    ```
+
+### 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
+
+- [Multi-GPU Training](https://docs.axolotl.ai/docs/multi-gpu.html)
+- [LoRA Optimizations](https://docs.axolotl.ai/docs/lora_optims.html)
+- [Multi-Node Training](https://docs.axolotl.ai/docs/multi-node.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)
+- [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

@@ -0,0 +1,49 @@
+base_model: LiquidAI/LFM2-350M
+
+chunked_cross_entropy: 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
+
+
+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: false
+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

@@ -1,27 +1,26 @@
-base_model: mistralai/Mistral-Small-3.1-24B-Instruct-2503
+base_model: LiquidAI/LFM2-VL-450M
+trust_remote_code: true
+model_type: AutoModelForImageTextToText
 processor_type: AutoProcessor
 
-load_in_8bit: 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
 
-chat_template: mistral_v7_tekken
 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
+val_set_size: 0.0
 output_dir: ./outputs/out
 
 adapter: lora
 lora_model_dir:
 
-sequence_len: 2048
+sequence_len: 8192
 pad_to_sequence_len: false
 
 lora_r: 32
@@ -35,7 +34,7 @@ wandb_watch:
 wandb_name:
 wandb_log_model:
 
-gradient_accumulation_steps: 1
+gradient_accumulation_steps: 4
 micro_batch_size: 1
 num_epochs: 1
 optimizer: adamw_bnb_8bit
@@ -48,11 +47,12 @@ tf32: true
 
 gradient_checkpointing: true
 logging_steps: 1
-flash_attention: false # PixtralVisionModel does not support Flash Attention 2.0 yet.
+flash_attention: true
 eager_attention:
 
 warmup_ratio: 0.1
 evals_per_epoch: 1
 saves_per_epoch: 1
 weight_decay: 0.0
-special_tokens:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/alst/README.md

@@ -0,0 +1,9 @@
+# Arctic Long Sequence Training (ALST)
+
+Artic Long Sequence Training (ALST) is a technique for training long context models using a variety of optimization
+techniques. It is a combination of:
+- TiledMLP: Leverage tiling over the sequence dimension on MLP layers to reduce memory usage
+- Tiled Loss: Using optimized loss functions like Liger-Kernel or Cut Cross Entropy to reduce memory usage
+- 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).

examples/alst/llama3-8b-deepspeed-alst.yaml

@@ -0,0 +1,53 @@
+base_model: meta-llama/Llama-3.1-8B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+datasets:
+  - path: togethercomputer/Long-Data-Collections
+    type: completion
+    field: text
+    data_files:
+      - pretrain/rp_sub.jsonl.zst
+  - path: princeton-nlp/TextbookChapters
+    type: completion
+    field: chapter
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+sequence_len: 500_000
+min_sample_len: 200_000
+sample_packing: true
+
+tiled_mlp: true
+context_parallel_size: 8
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_8bit
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+activation_offloading: legacy
+
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 100
+saves_per_epoch: 1
+evals_per_epoch: 2
+weight_decay: 0.0
+special_tokens:
+  pad_token: <|end_of_text|>
+
+deepspeed: deepspeed_configs/zero3_bf16_cpuoffload_all.json
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/alst/llama3-8b-fsdp2-alst.yaml

@@ -0,0 +1,59 @@
+base_model: meta-llama/Llama-3.1-8B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+datasets:
+  - path: togethercomputer/Long-Data-Collections
+    type: completion
+    field: text
+    data_files:
+      - pretrain/rp_sub.jsonl.zst
+  - path: princeton-nlp/TextbookChapters
+    type: completion
+    field: chapter
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+sequence_len: 500_000
+min_sample_len: 200_000
+sample_packing: true
+
+tiled_mlp: true
+context_parallel_size: 8
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_8bit
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+activation_offloading: legacy
+
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 100
+saves_per_epoch: 1
+evals_per_epoch: 2
+weight_decay: 0.0
+special_tokens:
+  pad_token: <|end_of_text|>
+
+fsdp_version: 2
+fsdp_config:
+  offload_params: false  # offloading is currently not compatible with SP + torchao optimizer
+  state_dict_type: SHARDED_STATE_DICT
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  reshard_after_forward: true
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

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==23.2 setuptools==75.8.0 wheel ninja
+pip3 install --no-build-isolation -e '.[flash-attn]'
+
+# Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
+python scripts/cutcrossentropy_install.py | sh
+```
+
+2. (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

@@ -0,0 +1,56 @@
+# Finetune ArceeAI's AFM with Axolotl
+
+[Arcee Foundation Models (AFM)](https://huggingface.co/collections/arcee-ai/afm-45b-68823397c351603014963473) are a family of 4.5B parameter open weight models trained by Arcee.ai.
+
+This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
+
+Thanks to the team at Arcee.ai for using Axolotl in supervised fine-tuning the AFM model.
+
+## Getting started
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html). You need to install from main as AFM 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==23.2 setuptools==75.8.0 wheel ninja
+pip3 install --no-build-isolation -e '.[flash-attn]'
+
+# Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
+python scripts/cutcrossentropy_install.py | sh
+```
+
+2. Run the finetuning example:
+
+```bash
+axolotl train examples/arcee/afm-4.5b-qlora.yaml
+```
+
+This config uses about 7.8GiB VRAM.
+
+Let us know how it goes. Happy finetuning! 🚀
+
+### TIPS
+
+- For inference, the official Arcee.ai team recommends `top_p: 0.95`, `temperature: 0.5`, `top_k: 50`, and `repeat_penalty: 1.1`.
+- You can run a full finetuning by removing the `adapter: qlora` and `load_in_4bit: true` from the config.
+- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+- The dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+
+## Optimization Guides
+
+- [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
+
+- [AFM Blog](https://docs.arcee.ai/arcee-foundation-models/introduction-to-arcee-foundation-models)
+- [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/arcee/afm-4.5b-qlora.yaml

@@ -0,0 +1,64 @@
+base_model: arcee-ai/AFM-4.5B
+
+# 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/archived/README.md

@@ -0,0 +1,5 @@
+# Archived Examples
+
+This directory contains examples that are no longer maintained and may no longer be functional.
+
+We keep them around for archival purposes in case they are useful to others.

examples/archived/cerebras/btlm-ft.yml

@@ -66,7 +66,7 @@ flash_optimum:
 gptq_groupsize:
 gptq_model_v1:
 
-warmup_steps: 32
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 save_total_limit:

examples/archived/cerebras/qlora.yml

@@ -43,7 +43,7 @@ xformers_attention: true
 flash_attention:
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.1

examples/archived/code-llama/13b/lora.yml

@@ -17,7 +17,7 @@ output_dir: ./outputs/lora-out
 
 sequence_len: 4096
 sample_packing: true
-pad_to_sequence_len: true
+
 
 adapter: lora
 lora_model_dir:
@@ -47,7 +47,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

examples/archived/code-llama/13b/qlora.yml

@@ -20,7 +20,7 @@ lora_model_dir:
 
 sequence_len: 4096
 sample_packing: true
-pad_to_sequence_len: true
+
 
 lora_r: 32
 lora_alpha: 16
@@ -48,7 +48,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

examples/archived/code-llama/34b/lora.yml

@@ -17,7 +17,7 @@ output_dir: ./outputs/lora-out
 
 sequence_len: 4096
 sample_packing: true
-pad_to_sequence_len: true
+
 
 adapter: lora
 lora_model_dir:
@@ -47,7 +47,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

examples/archived/code-llama/34b/qlora.yml

@@ -20,7 +20,7 @@ lora_model_dir:
 
 sequence_len: 4096
 sample_packing: true
-pad_to_sequence_len: true
+
 
 lora_r: 32
 lora_alpha: 16
@@ -48,7 +48,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

examples/archived/code-llama/7b/lora.yml

@@ -17,7 +17,7 @@ output_dir: ./outputs/lora-out
 
 sequence_len: 4096
 sample_packing: true
-pad_to_sequence_len: true
+
 
 adapter: lora
 lora_model_dir:
@@ -47,7 +47,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

examples/archived/code-llama/7b/qlora.yml

@@ -20,7 +20,7 @@ lora_model_dir:
 
 sequence_len: 4096
 sample_packing: true
-pad_to_sequence_len: true
+
 
 lora_r: 32
 lora_alpha: 16
@@ -48,7 +48,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

examples/archived/dbrx/16bit-lora.yaml

@@ -54,7 +54,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch:
 saves_per_epoch: 1
 

examples/archived/dbrx/8bit-lora.yaml

@@ -57,7 +57,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch:
 saves_per_epoch: 1
 

examples/archived/dbrx/fft-ds-zero3.yaml

@@ -41,7 +41,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch:
 saves_per_epoch: 1
 

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
@@ -21,7 +17,7 @@ output_dir: ./outputs/lora-out
 sequence_len: 4096
 sample_packing: true
 eval_sample_packing: false
-pad_to_sequence_len: true
+
 
 adapter: lora
 lora_model_dir:
@@ -51,7 +47,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 1
 saves_per_epoch: 1
 weight_decay: 0.0

examples/archived/falcon/config-7b-lora.yml

@@ -47,7 +47,7 @@ xformers_attention: true
 flash_attention:
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 40
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

examples/archived/falcon/config-7b-qlora.yml

@@ -77,7 +77,7 @@ xformers_attention: true
 flash_attention:
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.000001

examples/archived/falcon/config-7b.yml

@@ -44,7 +44,7 @@ xformers_attention: true
 flash_attention:
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 40
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

examples/archived/gemma/qlora.yml

@@ -25,7 +25,7 @@ lora_target_linear: true
 sequence_len: 4096
 sample_packing: true
 eval_sample_packing: false
-pad_to_sequence_len: true
+
 
 wandb_project:
 wandb_entity:

examples/archived/gptj/qlora.yml

@@ -40,7 +40,7 @@ xformers_attention: true
 flash_attention:
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.1

examples/archived/jeopardy-bot/config.yml

@@ -41,7 +41,7 @@ xformers_attention: true
 flash_attention:
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 20
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.1

examples/archived/mpt-7b/config.yml

@@ -42,7 +42,7 @@ logging_steps: 5
 flash_attention:
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 20
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0001

examples/archived/openllama-3b/config.yml

@@ -42,7 +42,7 @@ logging_steps: 1
 flash_attention: true
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 20
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.1

examples/archived/openllama-3b/lora.yml

@@ -50,7 +50,7 @@ logging_steps: 1
 flash_attention: true
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 20
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.1

examples/archived/openllama-3b/qlora.yml

@@ -43,7 +43,7 @@ logging_steps: 1
 flash_attention: true
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 20
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.1