chore: refactor normalize_attn to use mapping and loop

.github/workflows/base.yml

@@ -17,7 +17,7 @@ jobs:
   build-base:
     if: github.repository_owner == 'axolotl-ai-cloud'
     # this job needs to be run on self-hosted GPU runners...
-    runs-on: ubuntu-latest-m
+    runs-on: axolotl-gpu-runner
     strategy:
       fail-fast: false
       matrix:
@@ -28,50 +28,42 @@ jobs:
             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: ""
             python_version: "3.11"
             pytorch: 2.6.0
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
-            dockerfile: "Dockerfile-base"
           - cuda: "126"
             cuda_version: 12.6.3
             cudnn_version: ""
             python_version: "3.11"
             pytorch: 2.6.0
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
-            dockerfile: "Dockerfile-base"
           - cuda: "126"
             cuda_version: 12.6.3
             cudnn_version: ""
             python_version: "3.11"
             pytorch: 2.7.0
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
-            dockerfile: "Dockerfile-base"
           - cuda: "128"
             cuda_version: 12.6.3
             cudnn_version: ""
             python_version: "3.11"
             pytorch: 2.7.0
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
-            dockerfile: "Dockerfile-base"
           - cuda: "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"
+          - cuda: "128"
-#          # "next" is for release candidates of pytorch
+            cuda_version: 12.8.1
-#          - cuda: "128"
+            cudnn_version: ""
-#            cuda_version: 12.8.1
+            python_version: "3.11"
-#            cudnn_version: ""
+            pytorch: next
-#            python_version: "3.11"
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
-#            pytorch: next
-#            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
-#            dockerfile: "Dockerfile-base-next"
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -93,59 +85,7 @@ jobs:
         uses: docker/build-push-action@v4
         with:
           context: .
-          file: ./docker/${{ matrix.dockerfile }}
+          file: ${{ matrix.pytorch == 'nightly' && './docker/Dockerfile-base-nightly' || matrix.pytorch == 'next' && './docker/Dockerfile-base-next' || './docker/Dockerfile-base' }}
-          push: ${{ github.event_name != 'pull_request' }}
-          tags: ${{ steps.metadata.outputs.tags }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
-          labels: ${{ steps.metadata.outputs.labels }}
-          build-args: |
-            CUDA_VERSION=${{ matrix.cuda_version }}
-            CUDNN_VERSION=${{ matrix.cudnn_version }}
-            CUDA=${{ matrix.cuda }}
-            PYTHON_VERSION=${{ matrix.python_version }}
-            PYTORCH_VERSION=${{ matrix.pytorch }}
-            TORCH_CUDA_ARCH_LIST=${{ matrix.torch_cuda_arch_list }}
-  build-base-uv:
-    if: github.repository_owner == 'axolotl-ai-cloud'
-    runs-on: ubuntu-latest-m
-    strategy:
-      fail-fast: false
-      matrix:
-        include:
-          - cuda: "126"
-            cuda_version: 12.6.3
-            cudnn_version: ""
-            python_version: "3.11"
-            pytorch: 2.6.0
-            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
-            dockerfile: "Dockerfile-uv-base"
-          - cuda: "128"
-            cuda_version: 12.8.1
-            cudnn_version: ""
-            python_version: "3.11"
-            pytorch: 2.7.0
-            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
-            dockerfile: "Dockerfile-uv-base"
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v4
-      - name: Docker metadata
-        id: metadata
-        uses: docker/metadata-action@v5
-        with:
-          images: |
-            axolotlai/axolotl-base-uv
-      - name: Login to Docker Hub
-        uses: docker/login-action@v2
-        with:
-          username: ${{ secrets.DOCKERHUB_USERNAME }}
-          password: ${{ secrets.DOCKERHUB_TOKEN }}
-      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3
-      - name: Build
-        uses: docker/build-push-action@v4
-        with:
-          context: .
-          file: ./docker/${{ matrix.dockerfile }}
           push: ${{ github.event_name != 'pull_request' }}
           tags: ${{ steps.metadata.outputs.tags }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
           labels: ${{ steps.metadata.outputs.labels }}

.github/workflows/lint.yml

@@ -9,7 +9,6 @@ on:
        - '.github/workflows/*.yml'
        - "*.[q]md"
        - "examples/**/*.y[a]?ml"
-       - ".pre-commit-config.yaml"
   workflow_dispatch:
 
 jobs:

.github/workflows/main.yml

@@ -31,11 +31,6 @@ jobs:
             python_version: "3.11"
             pytorch: 2.7.0
             axolotl_extras:
-          - cuda: 128
-            cuda_version: 12.8.1
-            python_version: "3.11"
-            pytorch: 2.7.0
-            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -99,11 +94,6 @@ jobs:
             python_version: "3.11"
             pytorch: 2.7.0
             axolotl_extras:
-          - cuda: 128
-            cuda_version: 12.8.1
-            python_version: "3.11"
-            pytorch: 2.7.0
-            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout

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

@@ -3,12 +3,12 @@ name: docker-multigpu-tests-biweekly
 on:
   pull_request:
     paths:
-      - 'tests/e2e/multigpu/**.py'
+      - 'tests/e2e/multigpu/*.py'
       - 'requirements.txt'
       - 'setup.py'
       - 'pyproject.toml'
       - '.github/workflows/multi-gpu-e2e.yml'
-      - 'src/axolotl/core/trainers/mixins/context_parallel.py'
+      - 'src/axolotl/core/trainers/mixins/sequence_parallel.py'
       - 'src/axolotl/utils/distributed.py'
   workflow_dispatch:
   schedule:
@@ -59,7 +59,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2
+          pip install modal==0.71.8 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV

.github/workflows/precommit-autoupdate.yml

@@ -25,6 +25,7 @@ jobs:
           pre-commit autoupdate
           if [[ -n $(git status --porcelain) ]]; then
             echo "changes=true" >> $GITHUB_OUTPUT
+            git diff .pre-commit-config.yaml > pre-commit-update.diff
           fi
 
       - name: Create Pull Request
@@ -38,3 +39,11 @@ jobs:
           commit-message: "chore: update pre-commit hooks"
           body: |
             Automated PR to update pre-commit hooks to their latest versions.
+
+            <details>
+            <summary>Changes:</summary>
+
+            ```diff
+            ${{ steps.update.outputs.diff }}
+            ```
+            </details>

.github/workflows/tests-nightly.yml

@@ -18,96 +18,9 @@ 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

.github/workflows/tests.yml

@@ -44,10 +44,96 @@ 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]
+    needs: [preload-cache]
     strategy:
       fail-fast: false
       matrix:
@@ -59,11 +145,14 @@ jobs:
       - name: Check out repository code
         uses: actions/checkout@v4
 
-      - name: Restore Cache from S3
+      - name: Restore HF cache
-        id: hf-cache-restore-s3
+        id: hf-cache-restore
-        run: |
+        uses: actions/cache/restore@v4
-          mkdir -p /home/runner/.cache/huggingface/hub
+        with:
-          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C /home/runner/.cache/huggingface/hub/  --use-compress-program unzstd
+          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
@@ -121,6 +210,7 @@ jobs:
   pytest-sdist:
     name: PyTest from Source Dist
     runs-on: ubuntu-latest
+    needs: [preload-cache]
     strategy:
       fail-fast: false
       matrix:
@@ -132,11 +222,14 @@ jobs:
       - name: Check out repository code
         uses: actions/checkout@v4
 
-      - name: Restore Cache from S3
+      - name: Restore HF cache
-        id: hf-cache-restore-s3
+        id: hf-cache-restore
-        run: |
+        uses: actions/cache/restore@v4
-          mkdir -p /home/runner/.cache/huggingface/hub
+        with:
-          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C /home/runner/.cache/huggingface/hub/  --use-compress-program unzstd
+          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
@@ -184,7 +277,6 @@ jobs:
           find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
 
   docker-e2e-tests-1st:
-    # Run this job first as a gate for running the remainder of the test matrix
     if: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' }}
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
@@ -201,13 +293,6 @@ jobs:
             pytorch: 2.6.0
             num_gpus: 1
             axolotl_extras: vllm
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.6.0
-            num_gpus: 1
-            axolotl_extras:
-            dockerfile: "Dockerfile-uv.jinja"
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -218,7 +303,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2
+          pip install modal==0.71.8 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
@@ -229,7 +314,6 @@ jobs:
           echo "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
           echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
-          echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         run: |
           modal run cicd.e2e_tests
@@ -239,8 +323,6 @@ jobs:
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
     timeout-minutes: 90
-    # 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]
 
     strategy:
@@ -253,6 +335,12 @@ jobs:
             pytorch: 2.6.0
             num_gpus: 1
             axolotl_extras: llmcompressor
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.4.1
+            num_gpus: 1
+            axolotl_extras:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
@@ -265,12 +353,6 @@ jobs:
             pytorch: 2.7.0
             num_gpus: 1
             axolotl_extras:
-          - cuda: 128
-            cuda_version: 12.8.1
-            python_version: "3.11"
-            pytorch: 2.7.0
-            num_gpus: 1
-            axolotl_extras:
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -281,7 +363,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2
+          pip install modal==0.71.8 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
@@ -292,47 +374,6 @@ jobs:
           echo "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
           echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
-          echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         run: |
           modal run cicd.e2e_tests
-
-  docker-e2e-cleanup:
-    runs-on: [self-hosted, modal]
-    timeout-minutes: 90
-    needs: [docker-e2e-tests]
-
-    strategy:
-      fail-fast: false
-      matrix:
-        include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.6.0
-            num_gpus: 1
-            axolotl_extras: vllm
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v4
-      - name: Install Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: "3.11"
-      - name: Install Modal
-        run: |
-          python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2
-      - name: Update env vars
-        run: |
-          echo "BASE_TAG=main-base-py${{ 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 "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
-          echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
-          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
-      - name: Run tests job on Modal
-        run: |
-          modal run cicd.cleanup

.pre-commit-config.yaml

@@ -19,15 +19,15 @@ repos:
     hooks:
       - id: isort
 -   repo: https://github.com/PyCQA/flake8
-    rev: 7.2.0
+    rev: 7.1.2
     hooks:
     - id: flake8
 -   repo: https://github.com/pylint-dev/pylint
-    rev: v3.3.7
+    rev: v3.3.6
     hooks:
     - id: pylint
 -   repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.16.0
+    rev: v1.15.0
     hooks:
     - id: mypy
       additional_dependencies:

.runpod/src/config/config.yaml

@@ -242,12 +242,16 @@
 # early_stopping_patience: 3
 
 # # Specify a scheduler and kwargs to use with the optimizer
-# lr_scheduler: # 'one_cycle' | empty for cosine
+# lr_scheduler: # 'one_cycle' | 'log_sweep' | empty for cosine
 # lr_scheduler_kwargs:
 
 # # For one_cycle optim
 # lr_div_factor: # Learning rate div factor
 
+# # For log_sweep optim
+# log_sweep_min_lr:
+# log_sweep_max_lr:
+
 # # Specify optimizer
 # # Valid values are driven by the Transformers OptimizerNames class, see:
 # # https://github.com/huggingface/transformers/blob/95b374952dc27d8511541d6f5a4e22c9ec11fb24/src/transformers/training_args.py#L134

.runpod/src/handler.py

@@ -57,10 +57,8 @@ async def handler(job):
     logger.info("Training Complete.")
 
     # Cleanup
-    if "WANDB_API_KEY" in os.environ:
+    del os.environ["WANDB_API_KEY"]
-        del os.environ["WANDB_API_KEY"]
+    del os.environ["HF_TOKEN"]
-    if "HF_TOKEN" in os.environ:
-        del os.environ["HF_TOKEN"]
 
 
 runpod.serverless.start({"handler": handler, "return_aggregate_stream": True})

README.md

@@ -51,7 +51,7 @@ Features:
 
 - NVIDIA GPU (Ampere or newer for `bf16` and Flash Attention) or AMD GPU
 - Python 3.11
-- PyTorch ≥2.5.1
+- PyTorch ≥2.4.1
 
 ### Installation
 

_quarto.yml

@@ -17,9 +17,7 @@ quartodoc:
         - convert
         - prompt_tokenizers
         - logging_config
-        - core.builders.base
+        - core.trainer_builder
-        - core.builders.causal
-        - core.builders.rl
         - core.training_args
         - core.chat.messages
         - core.chat.format.chatml
@@ -45,37 +43,13 @@ quartodoc:
         - cli.vllm_serve
         - cli.cloud.base
         - cli.cloud.modal_
-        - cli.quantize
     - 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
-        - core.trainers.utils
-    - title: Model Loading
-      desc: Functionality for loading and patching models, tokenizers, etc.
-      contents:
-        - loaders.model
-        - loaders.tokenizer
-        - loaders.processor
-        - loaders.adapter
-        - loaders.patch_manager
-        - loaders.constants
-    - title: Mixins
-      desc: Mixin classes for augmenting trainers
-      contents:
-        - core.trainers.mixins.optimizer
-        - core.trainers.mixins.rng_state_loader
-        - core.trainers.mixins.scheduler
-    - title: Context Managers
-      desc: Context managers for altering trainer behaviors
-      contents:
-        - utils.ctx_managers.context_parallel
     - title: Prompt Strategies
       desc: Prompt formatting strategies
       contents:
@@ -112,7 +86,7 @@ quartodoc:
         - kernels.swiglu
         - kernels.quantize
         - kernels.utils
-    - title: Monkey Patches
+    - title: MonkeyPatches
       desc: Runtime patches for model optimizations
       contents:
         - monkeypatch.llama_attn_hijack_flash
@@ -129,16 +103,17 @@ quartodoc:
         - monkeypatch.trainer_fsdp_optim
         - monkeypatch.transformers_fa_utils
         - monkeypatch.unsloth_
+        - monkeypatch.attention.mllama
         - monkeypatch.data.batch_dataset_fetcher
         - monkeypatch.mixtral
-        - monkeypatch.gradient_checkpointing.offload_cpu
-        - monkeypatch.gradient_checkpointing.offload_disk
     - title: Utils
       desc: Utility functions
       contents:
+        - utils.models
         - utils.tokenization
         - utils.chat_templates
         - utils.lora
+        - utils.lora_embeddings
         - utils.model_shard_quant
         - utils.bench
         - utils.freeze
@@ -149,7 +124,7 @@ quartodoc:
         - utils.optimizers.adopt
         - utils.data.pretraining
         - utils.data.sft
-        - utils.quantization
+        - utils.gradient_checkpointing.unsloth
     - title: Schemas
       desc: Pydantic data models for Axolotl config
       contents:
@@ -199,14 +174,12 @@ quartodoc:
         - utils.callbacks.lisa
         - utils.callbacks.mlflow_
         - utils.callbacks.comet_
-        - utils.callbacks.qat
+
 website:
   title: "Axolotl"
   description: "We make fine-tuning accessible, scalable, and fun"
   favicon: favicon.jpg
 
-  google-analytics: "G-9KYCVJBNMQ"
-
   navbar:
     logo: image/axolotl_logo_digital_white.svg
     title: false
@@ -259,8 +232,6 @@ website:
             - docs/lr_groups.qmd
             - docs/lora_optims.qmd
             - docs/dataset_loading.qmd
-            - docs/qat.qmd
-            - docs/quantize.qmd
 
         - section: "Core Concepts"
           contents:
@@ -274,7 +245,7 @@ website:
             - docs/unsloth.qmd
             - docs/torchao.qmd
             - docs/custom_integrations.qmd
-            - docs/context_parallelism.qmd
+            - docs/sequence_parallelism.qmd
 
         - section: "Troubleshooting"
           contents:

cicd/Dockerfile-uv.jinja

@@ -1,52 +0,0 @@
-FROM axolotlai/axolotl-base-uv:{{ BASE_TAG }}
-
-ENV TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 9.0+PTX"
-ENV AXOLOTL_EXTRAS="{{ AXOLOTL_EXTRAS }}"
-ENV AXOLOTL_ARGS="{{ AXOLOTL_ARGS }}"
-ENV CUDA="{{ CUDA }}"
-ENV PYTORCH_VERSION="{{ PYTORCH_VERSION }}"
-ENV GITHUB_REF="{{ GITHUB_REF }}"
-ENV GITHUB_SHA="{{ GITHUB_SHA }}"
-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
-
-WORKDIR /workspace
-
-RUN git clone --depth=1 https://github.com/axolotl-ai-cloud/axolotl.git
-
-WORKDIR /workspace/axolotl
-
-RUN git fetch origin +$GITHUB_REF && \
-    git checkout FETCH_HEAD
-
-# If AXOLOTL_EXTRAS is set, append it in brackets
-RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
-        sed -i 's#^transformers.*#transformers @ git+https://github.com/huggingface/transformers.git@main#' requirements.txt; \
-        sed -i 's#^peft.*#peft @ git+https://github.com/huggingface/peft.git@main#' requirements.txt; \
-        sed -i 's#^accelerate.*#accelerate @ git+https://github.com/huggingface/accelerate.git@main#' requirements.txt; \
-        sed -i 's#^trl.*#trl @ git+https://github.com/huggingface/trl.git@main#' requirements.txt; \
-        sed -i 's#^datasets.*#datasets @ git+https://github.com/huggingface/datasets.git@main#' requirements.txt; \
-    fi
-
-RUN uv pip install packaging==23.2 setuptools==75.8.0
-RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        uv pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
-    else \
-        uv pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray] $AXOLOTL_ARGS; \
-    fi
-
-RUN python scripts/unsloth_install.py --uv | sh
-RUN python scripts/cutcrossentropy_install.py --uv | sh
-
-# So we can test the Docker image
-RUN uv pip install -r requirements-dev.txt -r requirements-tests.txt
-
-# fix so that git fetch/pull from remote works
-RUN git config remote.origin.fetch "+refs/heads/*:refs/remotes/origin/*" && \
-    git config --get remote.origin.fetch
-
-# helper for huggingface-login cli
-RUN git config --global credential.helper store

cicd/cicd.sh

@@ -18,7 +18,7 @@ pytest -v --durations=10 \
   --cov-append
 
 # Run patched tests excluding lora kernels with coverage append
-pytest --full-trace -vvv --durations=10 \
+pytest -v --durations=10 \
   --ignore=tests/e2e/patched/lora_kernels \
   /workspace/axolotl/tests/e2e/patched \
   --cov=axolotl \

cicd/cleanup.py

@@ -1,19 +0,0 @@
-"""Modal app to run axolotl GPU cleanup"""
-
-from .single_gpu import VOLUME_CONFIG, app, cicd_image, run_cmd
-
-
-@app.function(
-    image=cicd_image,
-    timeout=60 * 60,
-    cpu=8.0,
-    memory=131072,
-    volumes=VOLUME_CONFIG,
-)
-def cleanup():
-    run_cmd("./cicd/cleanup.sh", "/workspace/axolotl")
-
-
-@app.local_entrypoint()
-def main():
-    cleanup.remote()

cicd/cleanup.sh

@@ -1,6 +0,0 @@
-#!/bin/bash
-set -e
-
-# cleanup old cache files for datasets processing and intermediate mappings
-find /workspace/data/huggingface-cache/hub/datasets -name "cache-*" -type f -mtime +1 -exec rm {} \;
-find /workspace/data/huggingface-cache/hub/datasets -name "*.lock" -type f -mtime +1 -exec rm {} \;

cicd/e2e_tests.py

@@ -1,12 +1,75 @@
 """Modal app to run axolotl GPU tests"""
 
-from .single_gpu import GPU_CONFIG, VOLUME_CONFIG, app, cicd_image, run_cmd
+# pylint: disable=duplicate-code
+
+import os
+import pathlib
+import tempfile
+
+import jinja2
+import modal
+from jinja2 import select_autoescape
+from modal import App, Image
+
+cicd_path = pathlib.Path(__file__).parent.resolve()
+
+template_loader = jinja2.FileSystemLoader(searchpath=cicd_path)
+template_env = jinja2.Environment(
+    loader=template_loader, autoescape=select_autoescape()
+)
+df_template = template_env.get_template("Dockerfile.jinja")
+
+df_args = {
+    "AXOLOTL_EXTRAS": os.environ.get("AXOLOTL_EXTRAS", ""),
+    "AXOLOTL_ARGS": os.environ.get("AXOLOTL_ARGS", ""),
+    "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.4.1"),
+    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-py3.11-cu121-2.4.1"),
+    "CUDA": os.environ.get("CUDA", "121"),
+    "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",
+}
+
+dockerfile_contents = df_template.render(**df_args)
+
+temp_dir = tempfile.mkdtemp()
+with open(pathlib.Path(temp_dir) / "Dockerfile", "w", encoding="utf-8") as f:
+    f.write(dockerfile_contents)
+
+cicd_image = Image.from_dockerfile(
+    pathlib.Path(temp_dir) / "Dockerfile",
+    context_mount=None,
+    force_build=True,
+    gpu="A10G",
+).env(df_args)
+
+app = App("Axolotl CI/CD", secrets=[])
+
+hf_cache_volume = modal.Volume.from_name(
+    "axolotl-ci-hf-hub-cache", create_if_missing=True
+)
+VOLUME_CONFIG = {
+    "/workspace/data/huggingface-cache/hub": hf_cache_volume,
+}
+
+N_GPUS = int(os.environ.get("N_GPUS", 1))
+GPU_CONFIG = modal.gpu.L40S(count=N_GPUS)
+
+
+def run_cmd(cmd: str, run_folder: str):
+    import subprocess  # nosec
+
+    # Propagate errors from subprocess.
+    if exit_code := subprocess.call(cmd.split(), cwd=run_folder):  # nosec
+        exit(exit_code)  # pylint: disable=consider-using-sys-exit
 
 
 @app.function(
     image=cicd_image,
     gpu=GPU_CONFIG,
-    timeout=90 * 60,  # 90 min
+    timeout=60 * 60,
     cpu=8.0,
     memory=131072,
     volumes=VOLUME_CONFIG,

cicd/multigpu.py

@@ -24,9 +24,9 @@ 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"),
+    "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.4.1"),
-    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-py3.11-cu124-2.5.1"),
+    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-py3.11-cu121-2.4.1"),
-    "CUDA": os.environ.get("CUDA", "124"),
+    "CUDA": os.environ.get("CUDA", "121"),
     "GITHUB_REF": os.environ.get("GITHUB_REF", "refs/heads/main"),
     "GITHUB_SHA": os.environ.get("GITHUB_SHA", ""),
     "CODECOV_TOKEN": os.environ.get("CODECOV_TOKEN", ""),
@@ -55,7 +55,7 @@ VOLUME_CONFIG = {
 }
 
 N_GPUS = int(os.environ.get("N_GPUS", 2))
-GPU_CONFIG = f"H100:{N_GPUS}"
+GPU_CONFIG = modal.gpu.H100(count=N_GPUS)
 
 
 def run_cmd(cmd: str, run_folder: str):
@@ -70,7 +70,7 @@ def run_cmd(cmd: str, run_folder: str):
     image=cicd_image,
     gpu=GPU_CONFIG,
     timeout=90 * 60,
-    cpu=16.0,
+    cpu=8.0,
     memory=131072 * N_GPUS,
     volumes=VOLUME_CONFIG,
 )

cicd/single_gpu.py

@@ -1,68 +0,0 @@
-"""Modal app to run axolotl GPU tests"""
-
-# pylint: disable=duplicate-code
-
-import os
-import pathlib
-import tempfile
-
-import jinja2
-import modal
-import modal.experimental
-from jinja2 import select_autoescape
-from modal import App
-
-cicd_path = pathlib.Path(__file__).parent.resolve()
-
-template_loader = jinja2.FileSystemLoader(searchpath=cicd_path)
-template_env = jinja2.Environment(
-    loader=template_loader, autoescape=select_autoescape()
-)
-dockerfile = os.environ.get("E2E_DOCKERFILE", "Dockerfile.jinja")
-df_template = template_env.get_template(dockerfile)
-
-df_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"),
-    "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",
-}
-
-dockerfile_contents = df_template.render(**df_args)
-
-temp_dir = tempfile.mkdtemp()
-with open(pathlib.Path(temp_dir) / "Dockerfile", "w", encoding="utf-8") as f:
-    f.write(dockerfile_contents)
-
-cicd_image = modal.experimental.raw_dockerfile_image(
-    pathlib.Path(temp_dir) / "Dockerfile",
-    # context_mount=None,
-    force_build=True,
-    # gpu="A10G",
-).env(df_args)
-
-app = App("Axolotl CI/CD", secrets=[])
-
-hf_cache_volume = modal.Volume.from_name(
-    "axolotl-ci-hf-hub-cache", create_if_missing=True
-)
-VOLUME_CONFIG = {
-    "/workspace/data/huggingface-cache/hub": hf_cache_volume,
-}
-
-N_GPUS = int(os.environ.get("N_GPUS", 1))
-GPU_CONFIG = f"L40S:{N_GPUS}"
-
-
-def run_cmd(cmd: str, run_folder: str):
-    import subprocess  # nosec
-
-    # Propagate errors from subprocess.
-    if exit_code := subprocess.call(cmd.split(), cwd=run_folder):  # nosec
-        exit(exit_code)  # pylint: disable=consider-using-sys-exit

codecov.yml

@@ -19,7 +19,7 @@ coverage:
         if_no_uploads: error
         if_not_found: success
         if_ci_failed: error
-        only_pulls: true
+        only_pulls: false
         flags: null
         paths: null
     patch:

docker/Dockerfile-uv-base

@@ -1,36 +0,0 @@
-ARG CUDA_VERSION="12.6.3"
-ARG CUDNN_VERSION=""
-ARG UBUNTU_VERSION="22.04"
-ARG MAX_JOBS=4
-
-FROM nvidia/cuda:$CUDA_VERSION-cudnn$CUDNN_VERSION-devel-ubuntu$UBUNTU_VERSION AS base-builder
-
-ARG PYTHON_VERSION="3.11"
-ARG PYTORCH_VERSION="2.6.0"
-ARG CUDA="126"
-ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 9.0+PTX"
-
-ENV PYTHON_VERSION=$PYTHON_VERSION
-ENV TORCH_CUDA_ARCH_LIST=$TORCH_CUDA_ARCH_LIST
-ENV UV_TORCH_BACKEND="cu${CUDA}"
-
-RUN apt-get update \
-    && apt-get install -y wget git build-essential ninja-build git-lfs libaio-dev pkg-config curl && rm -rf /var/lib/apt/lists/* \
-    && git lfs install --skip-repo \
-    && curl -LsSf https://astral.sh/uv/install.sh | sh
-
-ENV PATH="/root/.local/bin:${PATH}"
-
-RUN uv python install ${PYTHON_VERSION}
-
-WORKDIR /workspace
-
-RUN uv venv --no-project --relocatable axolotl-venv
-
-ENV PATH="/workspace/axolotl-venv/bin:${PATH}"
-
-RUN uv pip install packaging setuptools wheel \
-    && uv pip install torch==${PYTORCH_VERSION} \
-    && uv pip install --no-build-isolation "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main" \
-    && uv pip install "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main" \
-    && uv pip install awscli pydantic

docs/cli.qmd

@@ -209,16 +209,6 @@ axolotl delinearize-llama4 --model path/to/model_dir --output path/to/output_dir
 
 This would be necessary to use with other frameworks. If you have an adapter, merge it with the non-quantized linearized model before delinearizing.
 
-### quantize
-
-Quantizes a model using the quantization configuration specified in your YAML file.
-
-```bash
-axolotl quantize config.yml
-```
-
-See [Quantization](./quantize.qmd) for more details.
-
 
 ## Legacy CLI Usage
 

docs/config.qmd

@@ -65,20 +65,6 @@ bnb_config_kwargs:
   bnb_4bit_quant_type: nf4
   bnb_4bit_use_double_quant: true
 
-# quantization aware training
-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"
-  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
-
-# post-training quantization
-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"
-  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.
-
 
 # Whether you are training a 4-bit GPTQ quantized model
 gptq: true
@@ -112,10 +98,8 @@ plugins:
   # - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
 
 # A list of one or more datasets to finetune the model with
-# See https://docs.axolotl.ai/docs/dataset_loading.html for guide on loading datasets
-# See https://docs.axolotl.ai/docs/dataset-formats/ for guide on dataset formats
 datasets:
-  # HuggingFace dataset repo | s3:// | gs:// | path to local file or directory
+  # HuggingFace dataset repo | s3://,gs:// path | "json" for local dataset, make sure to fill data_files
   - path: vicgalle/alpaca-gpt4
     # The type of prompt to use for training. [alpaca, gpteacher, oasst, reflection]
     type: alpaca # format | format:<prompt_style> (chat/instruct) | <prompt_strategies>.load_<load_fn>
@@ -237,7 +221,7 @@ datasets:
 # The same applies to the `test_datasets` option and the `pretraining_dataset` option. Default is true.
 shuffle_merged_datasets: true
 
-# Deduplicates datasets and test_datasets with identical entries.
+Deduplicates datasets and test_datasets with identical entries.
 dataset_exact_deduplication: true
 
 # A list of one or more datasets to eval the model with.
@@ -286,25 +270,10 @@ trl:
 
   num_generations: # Optional[int]. Number of generations to sample.
   log_completions: # Optional[bool]. Whether to log completions.
-  num_completions_to_print: # Optional[int]. Number of completions to print when log_completions is True.
 
   sync_ref_model: # Optional[bool]. Whether to sync the reference model.
   ref_model_mixup_alpha: # Optional[float]. Mixup alpha for the reference model.
   ref_model_sync_steps: # Optional[int]. Sync steps for the reference model.
-  scale_rewards: # Optional[bool]. Whether to scale rewards by their standard deviation.
-
-  temperature: # Optional[float]. Sampling temperature for the GRPO policy.
-  top_p: # Optional[float]. Top-p sampling probability for the generation policy.
-  top_k: # Optional[int]. Top-k sampling for the generation policy.
-  min_p: # Optional[float]. Minimum probability for the generation policy.
-  repetition_penalty: # Optional[float]. Penalty for tokens that appear in prompt and generated text.
-
-  num_iterations: # Optional[int]. Number of iterations per batch (μ) for GRPO.
-  epsilon: # Optional[float]. Epsilon value for clipping in the GRPO algorithm.
-  epsilon_high: # Optional[float]. Upper-bound epsilon value for clipping in the GRPO algorithm.
-  use_liger_loss: # Optional[bool]. Whether to use Liger loss for GRPO.
-  loss_type: # Optional[str]. Loss formulation to use. Supported values: grpo, bnpo, dr_grpo.
-  mask_truncated_completions: # Optional[bool]. Whether to exclude truncated completions from loss calculation.
 
 
 # reward modelling: `True` or `False`
@@ -514,7 +483,6 @@ output_dir: ./completed-model
 # setting to `auto` will enable torch compile when torch>=2.5.1
 torch_compile:  # Optional[Union[Literal["auto"], bool]]
 torch_compile_backend:  # Optional[str]
-torch_compile_mode:  # 'default' | 'reduce-overhead' | 'max-autotune'
 
 # Training hyperparameters
 
@@ -537,7 +505,6 @@ save_strategy: # Set to `"no"` to skip checkpoint saves, `"epoch"` at end of eac
 save_steps: # Leave empty to save at each epoch, integer for every N steps. float for fraction of total steps
 saves_per_epoch: # number of times per epoch to save a checkpoint, mutually exclusive with save_steps
 save_total_limit: # Checkpoints saved at a time
-save_only_model: # Save only the model weights, skipping the optimizer. Using this means you can't resume from checkpoints.
 # Maximum number of iterations to train for. It precedes num_epochs which means that
 # if both are set, num_epochs will not be guaranteed.
 # e.g., when 1 epoch is 1000 steps => `num_epochs: 2` and `max_steps: 100` will train for 100 steps
@@ -561,7 +528,7 @@ profiler_steps: # enable the pytorch profiler to capture the first N steps of tr
 loss_watchdog_threshold: # High loss value, indicating the learning has broken down (a good estimate is ~2 times the loss at the start of training)
 loss_watchdog_patience: # Number of high-loss steps in a row before the trainer aborts (default: 3)
 
-# Save model as safetensors (require safetensors package). Default True
+# Save model as safetensors (require safetensors package)
 save_safetensors:
 
 # Whether to mask out or include the human's prompt from the training labels
@@ -571,7 +538,7 @@ train_on_inputs: false
 # Note that training loss may have an oscillating pattern with this enabled.
 group_by_length: false
 
-# Whether to use gradient checkpointing. Available options are: true, false, "offload", "offload_disk".
+# Whether to use gradient checkpointing. Available options are: true, false, "offload".
 # https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing
 gradient_checkpointing: false
 # additional kwargs to pass to the trainer for gradient checkpointing
@@ -583,24 +550,7 @@ gradient_checkpointing: false
 early_stopping_patience: 3
 
 # Specify a scheduler and kwargs to use with the optimizer
-# Valid values are driven by the Transformers SchedulerType class, see:
+lr_scheduler: # 'one_cycle' | 'rex' | 'log_sweep' | 'linear' | 'cosine_with_restarts' | 'polynomial' | 'constant' | 'constant_with_warmup' | 'inverse_sqrt' | 'reduce_lr_on_plateau' | 'cosine_with_min_lr' | 'warmup_stable_decay' | empty for cosine
-# https://github.com/huggingface/transformers/blob/5f4ecf2d9f867a1255131d2461d75793c0cf1db2/src/transformers/trainer_utils.py#L420
-# Valid values include
-# - 'linear'
-# - 'cosine' (default)
-# - 'cosine_with_restarts'
-# - 'polynomial'
-# - 'constant'
-# - 'constant_with_warmup'
-# - 'inverse_sqrt'
-# - 'reduce_lr_on_plateau'
-# - 'cosine_with_min_lr'
-# - 'warmup_stable_decay'
-
-# Additional schedulers include:
-# - 'one_cycle'
-# - 'rex'
-lr_scheduler:
 lr_scheduler_kwargs:
 cosine_min_lr_ratio: # decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of peak lr
 cosine_constant_lr_ratio: # freeze lr at some percentage of the step, e.g. cosine_constant_lr_ratio=0.8 means start cosine_min_lr at 80% of training step (https://arxiv.org/pdf/2308.04014.pdf)
@@ -618,7 +568,7 @@ lr_div_factor: # Learning rate div factor
 #
 # Valid values for 'optimizer' include:
 # - adamw_torch
-# - adamw_torch_fused (default)
+# - adamw_torch_fused
 # - adamw_torch_xla
 # - adamw_torch_npu_fused
 # - adamw_apex_fused
@@ -662,7 +612,6 @@ lr_div_factor: # Learning rate div factor
 # - optimi_adamw
 # - ao_adamw_8bit
 # - ao_adamw_fp8
-# - came_pytorch
 optimizer:
 # Dictionary of arguments to pass to the optimizer
 optim_args:
@@ -682,9 +631,7 @@ weight_decay:
 # adamw hyperparams
 adam_beta1:
 adam_beta2:
-adam_beta3:  # only used for CAME Optimizer
 adam_epsilon:
-adam_epsilon2:  # only used for CAME Optimizer
 # Gradient clipping max norm
 max_grad_norm:
 
@@ -764,13 +711,13 @@ ddp_timeout:
 ddp_bucket_cap_mb:
 ddp_broadcast_buffers:
 
-# Context parallelism
+# Sequence parallelism
 # Set to a divisor of the number of GPUs available to split sequences into chunks of equal size.
 # Use in long context training to prevent OOM when sequences cannot fit into a single GPU's VRAM.
 # E.g., if 4 GPUs are available, set this value to 2 to split each sequence into two equal-sized
 # subsequences, or set to 4 to split into four equal-sized subsequences.
-# See https://docs.axolotl.ai/docs/context_parallelism.html for more details.
+# See https://docs.axolotl.ai/docs/sequence_parallelism.html for more details.
-context_parallel_degree:
+sequence_parallel_degree:
 # Optional; strides across the key dimension. Larger values use more memory but should make training faster.
 # Must evenly divide the number of KV heads in your model.
 heads_k_stride: 1

docs/dataset-formats/index.qmd

@@ -36,6 +36,10 @@ It is typically recommended to save your dataset as `.jsonl` due to its flexibil
 
 Axolotl supports loading from a Hugging Face hub repo or from local files.
 
+::: {.callout-important}
+For pre-training only, Axolotl would split texts if it exceeds the context length into multiple smaller prompts.
+:::
+
 ### Pre-training from Hugging Face hub datasets
 
 As an example, to train using a Hugging Face dataset `hf_org/name`, you can pass the following config:
@@ -73,21 +77,18 @@ datasets:
     type: completion
 ```
 
-From local files:
+From local files (either example works):
 
 ```yaml
 datasets:
   - path: A.jsonl
     type: completion
 
-  - path: B.jsonl
+  - path: json
+    data_files: ["A.jsonl", "B.jsonl", "C.jsonl"]
     type: completion
 ```
 
-::: {.callout-important}
-For `completion` only, Axolotl would split texts if it exceeds the context length into multiple smaller prompts. If you are interested in having this for `pretraining_dataset` too, please let us know or help make a PR!
-:::
-
 ### Pre-training dataset configuration tips
 
 #### Setting max_steps

docs/dataset_loading.qmd

@@ -54,7 +54,7 @@ datasets:
 
 #### Files
 
-To load a JSON file, you would do something like this:
+Usually, to load a JSON file, you would do something like this:
 
 ```python
 from datasets import load_dataset
@@ -66,11 +66,19 @@ Which translates to the following config:
 
 ```yaml
 datasets:
-  - path: data.json
+  - path: json
-    ds_type: json
+    data_files: /path/to/your/file.jsonl
 ```
 
-In the example above, it can be seen that we can just point the `path` to the file or directory along with the `ds_type` to load the dataset.
+However, to make things easier, we have added a few shortcuts for loading local dataset files.
+
+You can just point the `path` to the file or directory along with the `ds_type` to load the dataset. The below example shows for a JSON file:
+
+```yaml
+datasets:
+  - path: /path/to/your/file.jsonl
+    ds_type: json
+```
 
 This works for CSV, JSON, Parquet, and Arrow files.
 

docs/docker.qmd

@@ -8,10 +8,6 @@ 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 2.7.0 and CUDA 12.8.
-:::
-
 ## Base
 
 The base image is the most minimal image that can install Axolotl. It is based on the `nvidia/cuda` image. It includes python, torch, git, git-lfs, awscli, pydantic, and more.
@@ -36,6 +32,7 @@ Tags examples:
 - `main-base-py3.11-cu126-2.7.0`
 - `main-base-py3.11-cu124-2.6.0`
 - `main-base-py3.11-cu124-2.5.1`
+- `main-base-py3.11-cu124-2.4.1`
 
 ## Main
 
@@ -76,10 +73,12 @@ Tags examples:
 - `main-py3.11-cu126-2.7.0`
 - `main-py3.11-cu124-2.6.0`
 - `main-py3.11-cu124-2.5.1`
+- `main-py3.11-cu124-2.4.1`
 - `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-cu124-2.4.1`
+- `0.7.1`
 
 ## Cloud
 

docs/faq.qmd

@@ -110,17 +110,3 @@ description: Frequently asked questions
 > A: If `eot_tokens: ` is not provided, the default behavior is the same as before. EOS tokens used to delimit turns are masked/unmasked depending on whether the turn is trainable.
 
 > Internally, `eot_tokens: tokenizer.eos_token` and `train_on_eot: train_on_eos` (which defaults to `turn`). This transition helps clarify the naming and behavior of EOT/EOS tokens.
-
-**Q: `Data processing error: CAS service error`**
-
-> A: Try disabling XET with `export HF_HUB_DISABLE_XET=1`
-
-**Q: `torch._inductor.exc.LoweringException: NoValidChoicesError: No choices to select, please consider adding ATEN into max_autotune_gemm_backends config (defined in torch/_inductor/config.py) to allow at least one choice. `**
-
-> A: Depending on the version of torch, you may need to include this in your YAML:
-
-> ```yaml
-> flex_attn_compile_kwargs:
->   dynamic: false
->   mode: max-autotune-no-cudagraphs
-> ```

docs/getting-started.qmd

@@ -104,7 +104,7 @@ the `alpaca` dataset format, which has the following format:
 Please see our [Dataset Formats](dataset-formats) for more dataset formats and how to
 format them.
 
-2. Prepare your JSONL data in the specified format (in this case, the expected `alpaca`
+2. Prepare your JSONL data in the specified format (in this case, the expected `alpaca
 format):
 
 ```json
@@ -120,12 +120,6 @@ axolotl train my_training.yml
 
 ## Common Tasks {#sec-common-tasks}
 
-::: {.callout-tip}
-
-The same yaml file is used for training, inference, and merging.
-
-:::
-
 ### Testing Your Model {#sec-testing}
 
 After training, test your model:
@@ -134,16 +128,6 @@ After training, test your model:
 axolotl inference my_training.yml --lora-model-dir="./outputs/lora-out"
 ```
 
-More details can be found in [Inference](inference.qmd).
-
-### Using a UI {#sec-ui}
-
-Launch a Gradio interface:
-
-```bash
-axolotl inference my_training.yml --lora-model-dir="./outputs/lora-out" --gradio
-```
-
 ### Preprocessing Data {#sec-preprocessing}
 
 For large datasets, preprocess first:
@@ -152,22 +136,14 @@ For large datasets, preprocess first:
 axolotl preprocess my_training.yml
 ```
 
-Please make sure to set `dataset_prepared_path: ` in your config to set the path to save the prepared dataset.
+### Using a UI {#sec-ui}
 
-More details can be found in [Dataset Preprocessing](dataset_preprocessing.qmd).
+Launch a Gradio interface:
-
-### Merging LoRA weights {#sec-merging-lora}
-
-To merge the LoRA weights back into the base model, run:
 
 ```bash
-axolotl merge-lora my_training.yml --lora-model-dir="./outputs/lora-out"
+axolotl inference my_training.yml --lora-model-dir="./outputs/lora-out" --gradio
 ```
 
-The merged model will be saved in the `{output_dir}/merged` directory.
-
-More details can be found in [Merging LoRA weights](inference.qmd#sec-merging).
-
 ## Next Steps {#sec-next-steps}
 
 Now that you have the basics, you might want to:
@@ -180,7 +156,6 @@ Now that you have the basics, you might want to:
 Check our other guides for details on these topics:
 
 - [Configuration Guide](config.qmd) - Full configuration options
-- [Dataset Loading](dataset_loading.qmd) - Loading datasets from various sources
 - [Dataset Formats](dataset-formats) - Working with different data formats
 - [Multi-GPU Training](multi-gpu.qmd)
 - [Multi-Node Training](multi-node.qmd)

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 ≥3.10
-- PyTorch ≥2.5.1
+- PyTorch ≥2.4.1
 
 ## Installation Methods {#sec-installation-methods}
 
@@ -25,10 +25,6 @@ Please make sure to have Pytorch installed before installing Axolotl in your loc
 Follow the instructions at: [https://pytorch.org/get-started/locally/](https://pytorch.org/get-started/locally/)
 :::
 
-::: {.callout-important}
-For Blackwell GPUs, please use Pytorch 2.7.0 and CUDA 12.8.
-:::
-
 ### PyPI Installation (Recommended) {#sec-pypi}
 
 ```{.bash}
@@ -41,40 +37,6 @@ installed) in order not to clobber it, and so that we set the correct version of
 dependencies that are specific to the PyTorch version or other installed
 co-dependencies.
 
-### uv Installation {#sec-uv}
-
-uv is a fast, reliable Python package installer and resolver built in Rust. It offers significant performance improvements over pip and provides better dependency resolution, making it an excellent choice for complex environments.
-
-Install uv if not already installed
-```{.bash}
-curl -LsSf https://astral.sh/uv/install.sh | sh
-source $HOME/.local/bin/env
-```
-
-Choose your CUDA version to use with PyTorch; e.g. `cu124`, `cu126`, `cu128`,
-then create the venv and activate
-```{.bash}
-export UV_TORCH_BACKEND=cu126
-uv venv --no-project --relocatable
-source .venv/bin/activate
-```
-
-Install PyTorch
-- PyTorch 2.6.0 recommended
-```{.bash}
-uv pip install packaging setuptools wheel
-uv pip install torch==2.6.0
-uv pip install awscli pydantic
-```
-
-Install axolotl from PyPi
-```{.bash}
-uv pip install --no-build-isolation axolotl[deepspeed,flash-attn]
-
-# optionally install with vLLM if you're using torch==2.6.0 and want to train w/ GRPO
-uv pip install --no-build-isolation axolotl[deepspeed,flash-attn,vllm]
-```
-
 ### Edge/Development Build {#sec-edge-build}
 
 For the latest features between releases:
@@ -110,10 +72,6 @@ docker run --privileged --gpus '"all"' --shm-size 10g --rm -it \
 ```
 :::
 
-::: {.callout-important}
-For Blackwell GPUs, please use `axolotlai/axolotl:main-py3.11-cu128-2.7.0` or the cloud variant `axolotlai/axolotl-cloud:main-py3.11-cu128-2.7.0`.
-:::
-
 Please refer to the [Docker documentation](docker.qmd) for more information on the different Docker images that are available.
 
 ## Cloud Environments {#sec-cloud}

docs/lora_optims.qmd

@@ -84,10 +84,6 @@ lora_qkv_kernel: true
 lora_o_kernel: true
 ```
 
-::: {.callout-note}
-Currently, LoRA kernels are not supported for RLHF training, only SFT.
-:::
-
 ## Requirements
 
 - One or more NVIDIA or AMD GPUs (in order to use the Triton kernels)

docs/multi-gpu.qmd

@@ -18,7 +18,7 @@ Axolotl supports several methods for multi-GPU training:
 
 - DeepSpeed (recommended)
 - FSDP (Fully Sharded Data Parallel)
-- Context parallelism
+- Sequence parallelism
 - FSDP + QLoRA
 
 ## DeepSpeed {#sec-deepspeed}
@@ -80,14 +80,27 @@ fsdp_config:
   fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
 ```
 
-## Context parallelism {#sec-sequence-parallelism}
+## Sequence parallelism {#sec-sequence-parallelism}
 
-We support context parallelism (SP) via the
+We support sequence parallelism (SP) via the
 [ring-flash-attention](https://github.com/zhuzilin/ring-flash-attention) project. This
 allows one to split up sequences across GPUs, which is useful in the event that a
 single sequence causes OOM errors during model training.
 
-See our [dedicated guide](context_parallelism.qmd) for more information.
+First, install `ring-flash-attn`, recommended via `pip install axolotl[ring-flash-attn]`,
+or from source with `pip install .[ring-flash-attn]`.
+
+Your Axolotl YAML config should contain the following lines:
+
+```{.yaml}
+sequence_parallel_degree: 4  # Split each sequence into 4 parts, one per GPU
+flash_attention: true  # Required with sequence parallelism
+
+# Optional; strides across the key dimension. Larger values use more memory but will make training faster.
+heads_k_stride: 1
+```
+
+See our [dedicated guide](sequence_parallelism.qmd) for more details.
 
 ### FSDP + QLoRA {#sec-fsdp-qlora}
 

docs/multimodal.qmd

@@ -43,7 +43,7 @@ datasets:
 # leave the vision model and vision tower frozen
 # load_in_8bit: true
 adapter: lora
-lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+lora_target_modules: 'language_model.model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
 
 # (optional) if you want to resize images to a set size
 image_size: 512

docs/qat.qmd

@@ -1,32 +0,0 @@
----
-title: "Quantization Aware Training (QAT)"
-back-to-top-navigation: true
-toc: true
-toc-expand: 2
-toc-depth: 4
----
-
-## Overview
-
-[Quantization Aware Training](https://pytorch.org/blog/introduction-to-quantization-on-pytorch/#quantization-aware-training) (QAT) is a technique for improving the accuracy of models which are quantized
-by applying "fake" quantizations to the model's weights (and optionally, activations) during training. This fake
-quantization allows for the model to adjust for noise introduced by the quantization, so when the model is eventually
-quantized, the accuracy loss is minimized. We use the quantization techniques implemented in [torchao](https://github.com/pytorch/ao) to provide
-support for QAT and post-training quantization (PTQ) in axolotl.
-
-We recommend reviewing the excellent QAT tutorial in the [torchtune library](https://pytorch.org/torchtune/main/tutorials/qat_finetune.html#quantizing-the-qat-model),
-and the QAT documentation in the [torchao library](https://github.com/pytorch/ao/tree/main/torchao/quantization/qat), for more details.
-
-## Configuring QAT in Axolotl
-
-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"
-  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
-```
-
-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` command](./quantize.md) to do this.

docs/quantize.qmd

@@ -1,53 +0,0 @@
----
-title: "Quantization with torchao"
-back-to-top-navigation: true
-toc: true
-toc-expand: 2
-toc-depth: 4
----
-
-Quantization is a technique to lower the memory footprint of your model, potentially at the cost of accuracy or model performance. We support quantizing your model using the [torchao](https://github.com/pytorch/ao) library. Quantization is supported for both post-training quantization (PTQ) and quantization-aware training (QAT).
-
-
-::: {.callout-note}
-
-We do not currently support quantization techniques such as GGUF/GPTQ,EXL2 at the moment.
-
-:::
-
-## Configuring Quantization in Axolotl
-
-Quantization is configured using the `quantization` key in your configuration file.
-
-```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"
-  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.
-
-output_dir:  # The path to the output directory.
-```
-
-Once quantization is complete, your quantized model will be saved in the `{output_dir}/quantized` directory.
-
-You may also use the `quantize` command to quantize a model which has been trained with [QAT](./qat.md) - you can do this by using the existing QAT configuration file which
-you used to train the model:
-
-```yaml
-# qat.yml
-qat:
-  activation_dtype: int8
-  weight_dtype: int8
-  group_size: 256
-  quantize_embedding: true
-
-output_dir: # The path to the output directory used during training where the final checkpoint has been saved.
-```
-
-```bash
-axolotl quantize qat.yml
-```
-
-This ensures that an identical quantization configuration is used to quantize the model as was used to train it.

docs/rlhf.qmd

@@ -16,8 +16,7 @@ feedback. Various methods include, but not limited to:
 - [Identity Preference Optimization (IPO)](#ipo)
 - [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)
-- Proximal Policy Optimization (PPO) (not yet supported in axolotl, if you're interested in contributing, please reach out!)
 
 
 ## RLHF using Axolotl
@@ -583,20 +582,7 @@ datasets:
 
 To see other examples of custom reward functions, please see [TRL GRPO Docs](https://github.com/huggingface/trl/blob/main/docs/source/grpo_trainer.md#using-a-custom-reward-function).
 
-To see all configs, please see [TRLConfig](https://github.com/axolotl-ai-cloud/axolotl/blob/v0.9.2/src/axolotl/utils/schemas/trl.py).
+To see description of the configs, please see [TRLConfig](https://github.com/axolotl-ai-cloud/axolotl/blob/main/src/axolotl/utils/config/models/input/v0_4_1/trl.py).
-
-#### GRPO with DAPO/Dr. GRPO loss
-
-The DAPO paper and subsequently Dr. GRPO paper proposed an alternative loss function for GRPO to remediate the penalty in longer responses.
-
-```yaml
-trl:
-  loss_type: dr_grpo
-  # Normalizes loss based on max completion length (default: 256)
-  max_completion_length:
-```
-
-For more information, see [GRPO docs](https://huggingface.co/docs/trl/v0.17.0/en/grpo_trainer#loss-types).
 
 ### SimPO
 

docs/sequence_parallelism.qmd

@@ -1,16 +1,18 @@
 ---
-title: Context Parallelism
+title: Sequence Parallelism
 description: Train with long sequences split across multiple GPUs.
 ---
 
-Context parallelism is a technique that splits sequences across multiple GPUs,
+# Sequence Parallelism
+
+Sequence parallelism is a technique that splits sequences across multiple GPUs,
 allowing you to train with very long sequences that wouldn't fit on a single GPU. Each
 GPU processes a different portion of the sequence, and the results are aggregated
 through a ring communication pattern.
 
-## When to Use Context Parallelism
+## When to Use Sequence Parallelism
 
-Use context parallelism when:
+Use sequence parallelism when:
 
 - You need to train with sequence lengths that don't fit into a single GPU's memory
 - You have multiple GPUs available
@@ -18,35 +20,35 @@ Use context parallelism when:
 
 ## Configuration
 
-To enable context parallelism, add the following to your configuration file:
+To enable sequence parallelism, add the following to your configuration file:
 
 ```yaml
 # Set to a divisor (> 1) of the number of GPUs available
-context_parallel_degree: 4  # Split sequences across 4 GPUs
+sequence_parallel_degree: 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
+# Optional; one of "varlen_llama3", "batch_ring", "batch_zigzag", "batch_stripe". Defaults to
 # "varlen_llama3" when `sample_packing: true`, and "batch_ring" otherwise.
 ring_attn_func:
 ```
 
-The `context_parallel_degree` should be a divisor of the total number of GPUs. For example:
+The `sequence_parallel_degree` 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
 
 ## Implementation Details
 
-When context parallelism is enabled:
+When sequence parallelism is enabled:
 
-1. Each sequence is divided into equal chunks across the GPUs in a context parallel group
+1. Each sequence is divided into equal chunks across the GPUs in a sequence parallel group
 2. The data collator handles the chunking of input_ids, attention_mask, labels, and position_ids
-3. Position IDs are adjusted to maintain proper relative positions
+3. Position IDs are adjusted to maintain proper relative positions, especially for packed sequences
 4. The trainer uses special ring communication patterns for attention operations
 
 ## Requirements
 
-To use context parallelism, you need:
+To use sequence parallelism, you need:
 
 - Multiple GPUs (at least 2)
 - The `ring-flash-attn` package. Install with:
@@ -66,12 +68,10 @@ sequence_len: 8192
 
 ...
 
-context_parallel_degree: 4  # Split each sequence into 4 parts, one per GPU
+sequence_parallel_degree: 4  # Split each sequence into 4 parts, one per GPU
+flash_attention: true  # Required with sequence parallelism
 # 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
-# "varlen_llama3" when `sample_packing: true`, and "batch_ring" otherwise.
-ring_attn_func:
 
 ...
 ```
@@ -79,22 +79,22 @@ ring_attn_func:
 This will train the Llama 3 8B model with 8K context length, with each sequence split
 into 2 subsequences of length 4096 across 2 GPUs.
 
-## Sample Packing with Context Parallelism
+## Sample Packing with Sequence Parallelism
 
-Context parallelism is compatible with Axolotl's sample packing functionality. When using both features together:
+Sequence parallelism is compatible with Axolotl's sample packing functionality. When using both features together:
 
 1. Samples are first packed together
-2. The packed sequences are then divided across GPUs in the context parallel group
+2. The packed sequences are then divided across GPUs in the sequence parallel group
 3. Position IDs are automatically adjusted to maintain proper relative positions
 
 ## Effect on Batch Size
 
-When using context parallelism, your effective global batch size is **divided** by the `context_parallel_degree`. This happens because:
+When using sequence parallelism, your effective global batch size is **divided** by the `sequence_parallel_degree`. This happens because:
 
-- Each group of `context_parallel_degree` GPUs works on the same batch (just different parts of each sequence)
+- Each group of `sequence_parallel_degree` 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 context parallelism: 8 different batches processed per step
+- With 8 GPUs and no sequence parallelism: 8 different batches processed per step
-- With 8 GPUs and `context_parallel_degree=4`: Only 2 different batches processed per step (each split across 4 GPUs)
+- With 8 GPUs and `sequence_parallel_degree=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

examples/cerebras/btlm-ft.yml

@@ -59,9 +59,7 @@ gradient_checkpointing: false
 resume_from_checkpoint:
 logging_steps: 1
 
-flash_attention: true
+attention: flash
-sdp_attention:
-flash_optimum:
 
 gptq_groupsize:
 gptq_model_v1:

examples/cerebras/qlora.yml

@@ -39,8 +39,7 @@ tf32: true
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-xformers_attention: true
+attention: xformers
-flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 10

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

@@ -45,7 +45,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch: 4

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

@@ -46,7 +46,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch: 4

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

@@ -45,7 +45,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch: 4

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

@@ -46,7 +46,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch: 4

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

@@ -45,7 +45,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch: 4

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

@@ -46,7 +46,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/cohere/command-r-7b-qlora.yml

@@ -49,7 +49,8 @@ tf32: true
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_ratio: 0.1
 evals_per_epoch:

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

@@ -112,9 +112,7 @@
     "early_stopping_patience:\n",
     "resume_from_checkpoint:\n",
     "logging_steps: 1\n",
-    "xformers_attention:\n",
+    "attention: sdpa\n",
-    "flash_attention: false\n",
-    "sdp_attention: true\n",
     "\n",
     "warmup_steps: 1\n",
     "max_steps: 25\n",

examples/dbrx/16bit-lora.yaml

@@ -52,7 +52,8 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch:

examples/dbrx/8bit-lora.yaml

@@ -55,7 +55,8 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch:

examples/dbrx/fft-ds-zero3.yaml

@@ -39,7 +39,8 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch:

examples/deepseek-v2/fft-fsdp-16b.yaml

@@ -35,7 +35,8 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 100
 evals_per_epoch: 2

examples/deepseek-v2/qlora-fsdp-2_5.yaml

@@ -59,7 +59,8 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 100
 evals_per_epoch: 2

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

@@ -43,8 +43,7 @@ tf32: true
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-xformers_attention: true
+attention: xformers
-flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 40

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

@@ -73,8 +73,7 @@ early_stopping_patience: 3
 resume_from_checkpoint:
 auto_resume_from_checkpoints: true
 logging_steps: 1
-xformers_attention: true
+attention: xformers
-flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 10

examples/falcon/config-7b.yml

@@ -40,8 +40,7 @@ tf32: true
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-xformers_attention: true
+attention: xformers
-flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 40

examples/gemma/qlora.yml

@@ -47,7 +47,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_ratio: 0.1
 evals_per_epoch: 4

examples/gemma2/qlora.yml

@@ -53,7 +53,8 @@ tf32: true
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_ratio: 0.1
 evals_per_epoch:

examples/gemma2/reward-model.yaml

@@ -43,7 +43,8 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_ratio: 0.1
 evals_per_epoch:

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

@@ -57,7 +57,8 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_ratio: 0.1
 evals_per_epoch:

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

@@ -28,7 +28,7 @@ pad_to_sequence_len: true
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+lora_target_modules: 'language_model.model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
 
 wandb_project:
 wandb_entity:
@@ -51,8 +51,7 @@ gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
 logging_steps: 1
-flash_attention: true
+attention: flash
-eager_attention:
 
 warmup_ratio: 0.1
 evals_per_epoch: 1

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

@@ -30,7 +30,7 @@ pad_to_sequence_len: false
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+lora_target_modules: 'language_model.model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
 
 wandb_project:
 wandb_entity:
@@ -53,8 +53,7 @@ gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
 logging_steps: 1
-flash_attention: true
+attention: flash
-eager_attention:
 
 warmup_ratio: 0.1
 evals_per_epoch: 1

examples/gptj/qlora.yml

@@ -36,8 +36,7 @@ tf32: true
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-xformers_attention: true
+attention: xformers
-flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 10

examples/jamba/qlora.yaml

@@ -47,7 +47,8 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch:

examples/jamba/qlora_deepspeed.yaml

@@ -46,7 +46,8 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch:

examples/jamba/qlora_fsdp_large.yaml

@@ -45,7 +45,8 @@ gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: true
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch: 1

examples/jeopardy-bot/config.yml

@@ -37,8 +37,7 @@ bf16: auto
 tf32: true
 resume_from_checkpoint:
 logging_steps: 5
-xformers_attention: true
+attention: xformers
-flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20

examples/llama-2/fft_optimized.yml

@@ -42,7 +42,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 flash_attn_cross_entropy: false
 flash_attn_rms_norm: true
 flash_attn_fuse_qkv: false

examples/llama-2/gptq-lora.yml

@@ -53,9 +53,7 @@ tf32: true
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention:
+attention: flash
-sdp_attention:
-flash_optimum:
 warmup_steps: 100
 evals_per_epoch: 4
 saves_per_epoch: 1

examples/llama-2/lisa.yml

@@ -46,7 +46,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 flash_attn_cross_entropy: false
 flash_attn_rms_norm: true
 flash_attn_fuse_qkv: false

examples/llama-2/loftq.yml

@@ -45,7 +45,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/llama-2/lora.yml

@@ -45,7 +45,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/llama-2/qlora-fsdp.yml

@@ -48,7 +48,8 @@ gradient_checkpointing_kwargs:
   use_reentrant: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/llama-2/qlora.yml

@@ -46,7 +46,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/llama-2/relora.yml

@@ -48,7 +48,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/llama-3-vision/lora-11b.yaml

@@ -29,7 +29,7 @@ pad_to_sequence_len: false
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+lora_target_modules: 'language_model.model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
 
 wandb_project:
 wandb_entity:
@@ -50,8 +50,7 @@ tf32: true
 
 gradient_checkpointing: true
 logging_steps: 1
-flash_attention: true
+attention: flash
-eager_attention:
 
 warmup_ratio: 0.1
 evals_per_epoch: 1

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

@@ -1,79 +0,0 @@
-base_model: meta-llama/Llama-3.2-3B
-# Automatically upload checkpoint and final model to HF
-# hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-plugins:
-  - axolotl.integrations.liger.LigerPlugin
-
-liger_rope: true
-liger_rms_norm: true
-liger_glu_activation: true
-liger_layer_norm: true
-liger_fused_linear_cross_entropy: true
-
-datasets:
-  - path: yahma/alpaca-cleaned
-    type: alpaca
-
-output_dir: ./outputs/qat_out/
-
-sample_packing: true
-pad_to_sequence_len: true
-sequence_len: 512
-
-flex_attention: true
-flex_attn_compile_kwargs:
-  dynamic: false
-  mode: max-autotune-no-cudagraphs
-
-qat:
-  activation_dtype: int8
-  weight_dtype: int4
-  group_size: 32
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 1
-micro_batch_size: 16
-num_epochs: 1
-optimizer: adamw_torch_fused
-
-cosine_constant_lr_ratio: 0
-cosine_min_lr_ratio: 1.0
-learning_rate: 2e-5
-save_only_model: true
-bf16: true
-
-resume_from_checkpoint:
-logging_steps: 1
-
-evals_per_epoch: 1
-saves_per_epoch: 1
-
-warmup_steps: 10
-weight_decay: 0.0
-fsdp:
-  - full_shard
-  - auto_wrap
-
-fsdp_config:
-  fsdp_version: 2
-  fsdp_offload_params: false
-  fsdp_cpu_ram_efficient_loading: true
-  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
-  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
-  fsdp_state_dict_type: FULL_STATE_DICT
-  fsdp_sharding_strategy: FULL_SHARD
-  fsdp_reshard_after_forward: true
-  fsdp_activation_checkpointing: true
-
-special_tokens:
-  pad_token: <|end_of_text|>

examples/llama-3/fft-8b-liger-fsdp.yaml

@@ -49,7 +49,8 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 100
 evals_per_epoch: 2

examples/llama-3/fft-8b.yaml

@@ -34,7 +34,8 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 100
 evals_per_epoch: 2

examples/llama-3/instruct-dpo-lora-8b.yml

@@ -61,7 +61,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/llama-3/instruct-lora-8b.yml

@@ -56,7 +56,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/llama-3/lora-1b-deduplicate-dpo.yml

@@ -77,7 +77,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/llama-3/lora-1b-deduplicate-sft.yml

@@ -53,7 +53,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/llama-3/lora-1b-kernels.yml

@@ -54,7 +54,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

examples/llama-3/lora-1b-ray.yml

@@ -48,7 +48,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

examples/llama-3/lora-1b-sample-packing-sequentially.yml

@@ -55,7 +55,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/llama-3/lora-1b.yml

@@ -5,7 +5,7 @@ base_model: NousResearch/Llama-3.2-1B
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
     type: alpaca
-
+dataset_prepared_path: last_run_prepared
 val_set_size: 0.1
 output_dir: ./outputs/lora-out
 
@@ -38,7 +38,6 @@ wandb_log_model:
 gradient_accumulation_steps: 2
 micro_batch_size: 2
 num_epochs: 1
-
 optimizer: adamw_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -49,7 +48,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

examples/llama-3/lora-8b.yml

@@ -49,7 +49,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/llama-3/qlora-1b-kto.yaml

@@ -53,7 +53,8 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 20
 evals_per_epoch: 4

examples/llama-3/qlora-1b.yml

@@ -51,7 +51,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

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

@@ -39,7 +39,8 @@ gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: true
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/llama-3/qlora-fsdp-70b.yaml

@@ -48,7 +48,8 @@ gradient_checkpointing_kwargs:
   use_reentrant: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/llama-3/qlora.yml

@@ -46,7 +46,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/llama-4/README.md

@@ -34,5 +34,3 @@ We provide a script to delinearize Llama 4 linearized models into regular Huggin
 ```bash
 axolotl delinearize-llama4 --model path/to/model_dir --output path/to/output_dir
 ```
-
-Note: This only works with the non-quantized linearized model. If you have an adapter, merge it with the *non-quantized linearized* model before delinearizing.

examples/llava/lora-7b.yaml

@@ -25,7 +25,7 @@ pad_to_sequence_len: false
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+lora_target_modules: 'language_model.model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
 
 wandb_project:
 wandb_entity:
@@ -46,8 +46,7 @@ tf32: true
 
 gradient_checkpointing: true
 logging_steps: 1
-flash_attention: true
+attention: flash
-eager_attention:
 
 warmup_ratio: 0.1
 evals_per_epoch: 1

examples/mamba/config.yml

@@ -39,7 +39,7 @@ tf32: true
 gradient_checkpointing: false
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention:
+attention: eager
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/mistral/bigstral-ds-zero3.yaml

@@ -42,7 +42,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 save_total_limit: 1
 save_steps:

examples/mistral/config.yml

@@ -36,7 +36,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/mistral/lora-mps.yml

@@ -53,8 +53,7 @@ tf32: true
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: false
+attention: sdpa
-sdp_attention: true
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

examples/mistral/lora.yml

@@ -54,7 +54,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

examples/mistral/mistral-dpo-qlora.yml

@@ -71,7 +71,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: false
+attention: eager
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/mistral/mistral-qlora-fsdp.yml

@@ -51,7 +51,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3