draft config for devstral

* models.py -> loaders/ module refactor

* refactor ModelLoader class

* plugin manager changes

* circular import fix

* pytest

* pytest

* minor improvements

* fix

* minor changes

* fix test

* remove dead code

* coderabbit comments

* lint

* fix

* coderabbit suggestion I liked

* more coderabbit

* review comments, yak shaving

* lint

* updating in light of SP ctx manager changes

* review comment

* review comment 2

.github/workflows/main.yml

@@ -31,6 +31,11 @@ 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
@@ -94,6 +99,11 @@ 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

@@ -3,7 +3,7 @@ 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/tests-nightly.yml

@@ -18,9 +18,96 @@ 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]
     strategy:
       fail-fast: false
       max-parallel: 2

.github/workflows/tests.yml

@@ -44,96 +44,102 @@ jobs:
         env:
           SKIP: no-commit-to-branch
 
-  preload-cache:
+#  preload-cache:
-    name: Preload HF cache
+#    name: Preload HF cache
-    runs-on: ubuntu-latest
+#    runs-on: ubuntu-latest
-    strategy:
+#    strategy:
-      fail-fast: false
+#      fail-fast: false
-      matrix:
+#      matrix:
-        python_version: ["3.11"]
+#        python_version: ["3.11"]
-        pytorch_version: ["2.6.0"]
+#        pytorch_version: ["2.6.0"]
-    timeout-minutes: 20
+#    timeout-minutes: 20
-
+#
-    env:
+#    env:
-      AXOLOTL_IS_CI_CACHE_PRELOAD: "1"
+#      AXOLOTL_IS_CI_CACHE_PRELOAD: "1"
-
+#
-    steps:
+#    steps:
-      - name: Check out repository code
+#      - name: Check out repository code
-        uses: actions/checkout@v4
+#        uses: actions/checkout@v4
-
+#
-      - name: Restore HF cache
+#      - name: Restore HF cache
-        id: hf-cache-restore
+#        id: hf-cache-restore
-        uses: actions/cache/restore@v4
+#        uses: actions/cache/restore@v4
-        with:
+#        with:
-          path: |
+#          path: |
-            /home/runner/.cache/huggingface/hub/datasets--*
+#            /home/runner/.cache/huggingface/hub/datasets--*
-            /home/runner/.cache/huggingface/hub/models--*
+#            /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ runner.os }}-hf-hub-cache-v2
+#          key: ${{ runner.os }}-hf-hub-cache-v2
-
+#
-      - name: Setup Python
+#      - name: Restore Cache from S3
-        uses: actions/setup-python@v5
+#        id: hf-cache-restore-s3
-        with:
+#        run: |
-          python-version: ${{ matrix.python_version }}
+#          mkdir -p /home/runner/.cache/huggingface/hub
-          cache: 'pip' # caching pip dependencies
+#          curl -L https://d1dttdx32dkk5p.cloudfront.net/hf-cache.tar.zst | tar -xf - -C /home/runner/.cache/huggingface/hub/  --use-compress-program unzstd
-
+#
-      - name: upgrade pip
+#      - name: Setup Python
-        run: |
+#        uses: actions/setup-python@v5
-          pip3 install --upgrade pip
+#        with:
-          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
+#          python-version: ${{ matrix.python_version }}
-
+#          cache: 'pip' # caching pip dependencies
-      - name: Install PyTorch
+#
-        run: |
+#      - name: upgrade pip
-          pip3 install torch==${{ matrix.pytorch_version }}
+#        run: |
-
+#          pip3 install --upgrade pip
-      - name: Install dependencies
+#          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
-        run: |
+#
-          pip3 show torch
+#      - name: Install PyTorch
-          pip3 install --no-build-isolation -U -e .
+#        run: |
-          python scripts/unsloth_install.py | sh
+#          pip3 install torch==${{ matrix.pytorch_version }}
-          python scripts/cutcrossentropy_install.py | sh
+#
-          pip3 install -r requirements-dev.txt -r requirements-tests.txt
+#      - name: Install dependencies
-
+#        run: |
-      - name: Make sure PyTorch version wasn't clobbered
+#          pip3 show torch
-        run: |
+#          pip3 install --no-build-isolation -U -e .
-          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
+#          python scripts/unsloth_install.py | sh
-
+#          python scripts/cutcrossentropy_install.py | sh
-      - name: Ensure axolotl CLI was installed
+#          pip3 install -r requirements-dev.txt -r requirements-tests.txt
-        run: |
+#
-          axolotl --help
+#      - name: Make sure PyTorch version wasn't clobbered
-
+#        run: |
-      - name: Pre-Download dataset fixture
+#          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
-        run: |
+#
-          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
+#      - name: Ensure axolotl CLI was installed
-
+#        run: |
-      - name: Run tests
+#          axolotl --help
-        run: |
+#
-          pytest -v tests/conftest.py
+#      - name: Pre-Download dataset fixture
-
+#        run: |
-      - name: Upload coverage to Codecov
+#          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
-        uses: codecov/codecov-action@v5
+#
-        with:
+#      - name: Run tests
-          token: ${{ secrets.CODECOV_TOKEN }}
+#        run: |
-          files: ./coverage.xml
+#          pytest -v tests/conftest.py
-          flags: unittests,pytorch-${{ matrix.pytorch_version }}
+#
-          fail_ci_if_error: false
+#      - name: Upload coverage to Codecov
-
+#        uses: codecov/codecov-action@v5
-      - name: cleanup pip cache
+#        with:
-        run: |
+#          token: ${{ secrets.CODECOV_TOKEN }}
-          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
+#          files: ./coverage.xml
-
+#          flags: unittests,pytorch-${{ matrix.pytorch_version }}
-      - name: Save HF cache
+#          fail_ci_if_error: false
-        id: hf-cache
+#
-        uses: actions/cache/save@v4
+#      - name: cleanup pip cache
-        with:
+#        run: |
-          path: |
+#          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
-            /home/runner/.cache/huggingface/hub/datasets--*
+#
-            /home/runner/.cache/huggingface/hub/models--*
+#      - name: Save HF cache
-          key: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
+#        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:
@@ -145,14 +151,20 @@ jobs:
       - name: Check out repository code
         uses: actions/checkout@v4
 
-      - name: Restore HF cache
+#      - name: Restore HF cache
-        id: hf-cache-restore
+#        id: hf-cache-restore
-        uses: actions/cache/restore@v4
+#        uses: actions/cache/restore@v4
-        with:
+#        with:
-          path: |
+#          path: |
-            /home/runner/.cache/huggingface/hub/datasets--*
+#            /home/runner/.cache/huggingface/hub/datasets--*
-            /home/runner/.cache/huggingface/hub/models--*
+#            /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ runner.os }}-hf-hub-cache-v2
+#          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
@@ -210,7 +222,7 @@ jobs:
   pytest-sdist:
     name: PyTest from Source Dist
     runs-on: ubuntu-latest
-    needs: [preload-cache]
+#    needs: [preload-cache]
     strategy:
       fail-fast: false
       matrix:
@@ -222,14 +234,20 @@ jobs:
       - name: Check out repository code
         uses: actions/checkout@v4
 
-      - name: Restore HF cache
+#      - name: Restore HF cache
-        id: hf-cache-restore
+#        id: hf-cache-restore
-        uses: actions/cache/restore@v4
+#        uses: actions/cache/restore@v4
-        with:
+#        with:
-          path: |
+#          path: |
-            /home/runner/.cache/huggingface/hub/datasets--*
+#            /home/runner/.cache/huggingface/hub/datasets--*
-            /home/runner/.cache/huggingface/hub/models--*
+#            /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ runner.os }}-hf-hub-cache-v2
+#          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
@@ -277,6 +295,7 @@ 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]
@@ -323,6 +342,8 @@ 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:
@@ -335,12 +356,6 @@ 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"
@@ -353,6 +368,12 @@ 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
@@ -377,3 +398,43 @@ jobs:
       - 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==0.71.8 jinja2
+      - name: Update env vars
+        run: |
+          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
+          echo "PYTORCH_VERSION=${{ 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

.runpod/src/handler.py

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

_quarto.yml

@@ -48,8 +48,22 @@ quartodoc:
       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: 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.sequence_parallel
     - title: Prompt Strategies
       desc: Prompt formatting strategies
       contents:
@@ -86,7 +100,7 @@ quartodoc:
         - kernels.swiglu
         - kernels.quantize
         - kernels.utils
-    - title: MonkeyPatches
+    - title: Monkey Patches
       desc: Runtime patches for model optimizations
       contents:
         - monkeypatch.llama_attn_hijack_flash
@@ -124,7 +138,8 @@ quartodoc:
         - utils.optimizers.adopt
         - utils.data.pretraining
         - utils.data.sft
-        - utils.gradient_checkpointing.unsloth
+        - utils.gradient_checkpointing.offload_cpu
+        - utils.gradient_checkpointing.offload_disk
     - title: Schemas
       desc: Pydantic data models for Axolotl config
       contents:

cicd/cicd.sh

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

cicd/cleanup.py

@@ -0,0 +1,19 @@
+"""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

@@ -0,0 +1,6 @@
+#!/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

@@ -1,75 +1,12 @@
 """Modal app to run axolotl GPU tests"""
 
-# pylint: disable=duplicate-code
+from .single_gpu import GPU_CONFIG, VOLUME_CONFIG, app, cicd_image, run_cmd
-
-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=60 * 60,
+    timeout=90 * 60,  # 90 min
     cpu=8.0,
     memory=131072,
     volumes=VOLUME_CONFIG,

cicd/multigpu.py

@@ -70,7 +70,7 @@ def run_cmd(cmd: str, run_folder: str):
     image=cicd_image,
     gpu=GPU_CONFIG,
     timeout=90 * 60,
-    cpu=8.0,
+    cpu=16.0,
     memory=131072 * N_GPUS,
     volumes=VOLUME_CONFIG,
 )

cicd/single_gpu.py

@@ -0,0 +1,66 @@
+"""Modal app to run axolotl GPU tests"""
+
+# 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

codecov.yml

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

docs/config.qmd

@@ -505,6 +505,7 @@ 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
@@ -538,7 +539,7 @@ 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".
+# Whether to use gradient checkpointing. Available options are: true, false, "offload", "offload_disk".
 # 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
@@ -612,6 +613,7 @@ 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:
@@ -631,7 +633,9 @@ 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:
 

docs/docker.qmd

@@ -8,6 +8,10 @@ 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.

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
@@ -120,6 +120,12 @@ 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:
@@ -128,6 +134,16 @@ 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:
@@ -136,14 +152,22 @@ For large datasets, preprocess first:
 axolotl preprocess my_training.yml
 ```
 
-### Using a UI {#sec-ui}
+Please make sure to set `dataset_prepared_path: ` in your config to set the path to save the prepared dataset.
 
-Launch a Gradio interface:
+More details can be found in [Dataset Preprocessing](dataset_preprocessing.qmd).
+
+### Merging LoRA weights {#sec-merging-lora}
+
+To merge the LoRA weights back into the base model, run:
 
 ```bash
-axolotl inference my_training.yml --lora-model-dir="./outputs/lora-out" --gradio
+axolotl merge-lora my_training.yml --lora-model-dir="./outputs/lora-out"
 ```
 
+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:
@@ -156,6 +180,7 @@ 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

@@ -25,6 +25,10 @@ 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}
@@ -72,6 +76,10 @@ 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/multi-gpu.qmd

@@ -87,20 +87,7 @@ We support sequence parallelism (SP) via the
 allows one to split up sequences across GPUs, which is useful in the event that a
 single sequence causes OOM errors during model training.
 
-First, install `ring-flash-attn`, recommended via `pip install axolotl[ring-flash-attn]`,
+See our [dedicated guide](sequence_parallelism.qmd) for more information.
-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/sequence_parallelism.qmd

@@ -3,8 +3,6 @@ title: Sequence Parallelism
 description: Train with long sequences split 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
@@ -27,7 +25,7 @@ To enable sequence parallelism, add the following to your configuration file:
 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", "batch_ring", "batch_zigzag", "batch_stripe". Defaults to
+# Optional; one of "varlen_llama3" or "batch_ring". Defaults to
 # "varlen_llama3" when `sample_packing: true`, and "batch_ring" otherwise.
 ring_attn_func:
 ```
@@ -43,7 +41,7 @@ When sequence parallelism is enabled:
 
 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, especially for packed sequences
+3. Position IDs are adjusted to maintain proper relative positions
 4. The trainer uses special ring communication patterns for attention operations
 
 ## Requirements
@@ -69,9 +67,11 @@ sequence_len: 8192
 ...
 
 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:
 
 ...
 ```

examples/cerebras/btlm-ft.yml

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

examples/cerebras/qlora.yml

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

examples/dbrx/16bit-lora.yaml

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

examples/dbrx/8bit-lora.yaml

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

examples/dbrx/fft-ds-zero3.yaml

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

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

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

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

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

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

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

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

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

examples/falcon/config-7b.yml

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

examples/gemma/qlora.yml

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

examples/gemma2/qlora.yml

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

examples/gemma2/reward-model.yaml

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

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

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

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

@@ -51,7 +51,8 @@ gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
 logging_steps: 1
-attention: flash
+flash_attention: true
+eager_attention:
 
 warmup_ratio: 0.1
 evals_per_epoch: 1

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

@@ -53,7 +53,8 @@ gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
 logging_steps: 1
-attention: flash
+flash_attention: true
+eager_attention:
 
 warmup_ratio: 0.1
 evals_per_epoch: 1

examples/gptj/qlora.yml

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

examples/jamba/qlora.yaml

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

examples/jamba/qlora_deepspeed.yaml

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

examples/jamba/qlora_fsdp_large.yaml

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

examples/jeopardy-bot/config.yml

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

examples/llama-2/fft_optimized.yml

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

examples/llama-2/gptq-lora.yml

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

examples/llama-2/lisa.yml

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

examples/llama-2/loftq.yml

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

examples/llama-2/lora.yml

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

examples/llama-2/qlora-fsdp.yml

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

examples/llama-2/qlora.yml

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

examples/llama-2/relora.yml

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

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

@@ -50,7 +50,8 @@ tf32: true
 
 gradient_checkpointing: true
 logging_steps: 1
-attention: flash
+flash_attention: true
+eager_attention:
 
 warmup_ratio: 0.1
 evals_per_epoch: 1

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

examples/llama-3/qlora.yml

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

examples/llama-4/README.md

@@ -34,3 +34,5 @@ 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

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

examples/mistral/bigstral-ds-zero3.yml

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

examples/mistral/config.yml

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

examples/mistral/devstral-small-2505.yml

@@ -0,0 +1,48 @@
+base_model: mistralai/Devstral-Small-2505
+processor_type: AutoProcessor
+
+# 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
+output_dir: ./outputs/out
+
+sequence_len: 2048
+pad_to_sequence_len: false
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+logging_steps: 1
+flash_attention: false
+eager_attention:
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/mistral/lora-mps.yml

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

examples/mistral/lora.yml

@@ -54,8 +54,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 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
-attention: eager
+flash_attention: false
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/mistral/mistral-qlora-fsdp.yml

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

examples/mistral/mistral-qlora-orpo.yml

@@ -59,8 +59,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

examples/mistral/mistral-small-3.1-24B-lora.yml

@@ -48,7 +48,9 @@ tf32: true
 
 gradient_checkpointing: true
 logging_steps: 1
-attention: eager  # PixtralVisionModel does not support Flash Attention 2.0 yet.
+flash_attention: false # PixtralVisionModel does not support Flash Attention 2.0 yet.
+eager_attention:
+
 warmup_ratio: 0.1
 evals_per_epoch: 1
 saves_per_epoch: 1

examples/mistral/mixtral-8x22b-qlora-fsdp.yml

@@ -49,8 +49,7 @@ tf32: true
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

examples/mistral/mixtral-qlora-fsdp.yml

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

examples/mistral/mixtral.yml

@@ -69,8 +69,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

examples/mistral/mixtral_22.yml

@@ -40,8 +40,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 save_total_limit: 1
 save_steps:

examples/mistral/qlora.yml

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

examples/mpt-7b/config.yml

@@ -39,7 +39,7 @@ bf16: auto
 tf32: true
 resume_from_checkpoint:
 logging_steps: 5
-attention: eager
+flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20

examples/openllama-3b/config.yml

@@ -39,8 +39,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20

examples/openllama-3b/lora.yml

@@ -47,8 +47,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20

examples/openllama-3b/qlora.yml

@@ -40,8 +40,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20

examples/phi/phi-ft.yml

@@ -48,8 +48,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: True
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 100
 evals_per_epoch: 4

examples/phi/phi-qlora.yml

@@ -51,8 +51,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: True
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 100
 evals_per_epoch: 4

examples/phi/phi2-ft.yml

@@ -48,8 +48,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: True
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 100
 evals_per_epoch: 4

examples/phi/phi3-ft-fsdp.yml

@@ -49,8 +49,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 100
 evals_per_epoch: 4

examples/phi/phi3-ft.yml

@@ -44,8 +44,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: True
 early_stopping_patience: 3
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 eval_steps: 1000
 save_steps: 5000