draft config for devstral

.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/multi-gpu-e2e.yml

@@ -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.yml

@@ -44,6 +44,98 @@ 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: 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
+#        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
@@ -59,6 +151,15 @@ jobs:
       - name: Check out repository code
         uses: actions/checkout@v4
 
+#      - name: Restore HF cache
+#        id: hf-cache-restore
+#        uses: actions/cache/restore@v4
+#        with:
+#          path: |
+#            /home/runner/.cache/huggingface/hub/datasets--*
+#            /home/runner/.cache/huggingface/hub/models--*
+#          key: ${{ runner.os }}-hf-hub-cache-v2
+
       - name: Restore Cache from S3
         id: hf-cache-restore-s3
         run: |
@@ -121,6 +222,7 @@ jobs:
   pytest-sdist:
     name: PyTest from Source Dist
     runs-on: ubuntu-latest
+#    needs: [preload-cache]
     strategy:
       fail-fast: false
       matrix:
@@ -132,6 +234,15 @@ jobs:
       - name: Check out repository code
         uses: actions/checkout@v4
 
+#      - name: Restore HF cache
+#        id: hf-cache-restore
+#        uses: actions/cache/restore@v4
+#        with:
+#          path: |
+#            /home/runner/.cache/huggingface/hub/datasets--*
+#            /home/runner/.cache/huggingface/hub/models--*
+#          key: ${{ runner.os }}-hf-hub-cache-v2
+
       - name: Restore Cache from S3
         id: hf-cache-restore-s3
         run: |
@@ -201,13 +312,6 @@ jobs:
             pytorch: 2.6.0
             num_gpus: 1
             axolotl_extras: vllm
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.6.0
-            num_gpus: 1
-            axolotl_extras:
-            dockerfile: "Dockerfile-uv.jinja"
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -218,7 +322,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2
+          pip install modal==0.71.8 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
@@ -229,7 +333,6 @@ 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'}}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         run: |
           modal run cicd.e2e_tests
@@ -281,7 +384,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2
+          pip install modal==0.71.8 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
@@ -292,7 +395,6 @@ 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'}}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         run: |
           modal run cicd.e2e_tests
@@ -322,7 +424,7 @@ jobs:
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
-          pip install modal==1.0.2 jinja2
+          pip install modal==0.71.8 jinja2
       - name: Update env vars
         run: |
           echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV

.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

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
@@ -45,7 +43,6 @@ quartodoc:
         - cli.vllm_serve
         - cli.cloud.base
         - cli.cloud.modal_
-        - cli.quantize
     - title: Trainers
       desc: Training implementations
       contents:
@@ -57,15 +54,6 @@ quartodoc:
         - 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:
@@ -129,16 +117,17 @@ 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 +138,8 @@ quartodoc:
         - utils.optimizers.adopt
         - utils.data.pretraining
         - utils.data.sft
-        - utils.quantization
+        - utils.gradient_checkpointing.offload_cpu
+        - utils.gradient_checkpointing.offload_disk
     - title: Schemas
       desc: Pydantic data models for Axolotl config
       contents:
@@ -199,14 +189,12 @@ 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 +247,6 @@ website:
             - docs/lr_groups.qmd
             - docs/lora_optims.qmd
             - docs/dataset_loading.qmd
-            - docs/qat.qmd
-            - docs/quantize.qmd
 
         - section: "Core Concepts"
           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/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):

cicd/single_gpu.py

@@ -8,9 +8,8 @@ import tempfile
 
 import jinja2
 import modal
-import modal.experimental
 from jinja2 import select_autoescape
-from modal import App
+from modal import App, Image
 
 cicd_path = pathlib.Path(__file__).parent.resolve()
 
@@ -18,15 +17,14 @@ 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.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"),
+    "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", ""),
     "NIGHTLY_BUILD": os.environ.get("NIGHTLY_BUILD", ""),
@@ -40,11 +38,11 @@ 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(
+cicd_image = Image.from_dockerfile(
     pathlib.Path(temp_dir) / "Dockerfile",
-    # context_mount=None,
+    context_mount=None,
     force_build=True,
-    # gpu="A10G",
+    gpu="A10G",
 ).env(df_args)
 
 app = App("Axolotl CI/CD", secrets=[])
@@ -57,7 +55,7 @@ VOLUME_CONFIG = {
 }
 
 N_GPUS = int(os.environ.get("N_GPUS", 1))
-GPU_CONFIG = f"L40S:{N_GPUS}"
+GPU_CONFIG = modal.gpu.L40S(count=N_GPUS)
 
 
 def run_cmd(cmd: str, run_folder: str):

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
 
@@ -561,7 +529,7 @@ 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
@@ -583,24 +551,7 @@ 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 +569,7 @@ 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

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

@@ -36,6 +36,7 @@ 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 +77,12 @@ 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

@@ -180,7 +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 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}
 
@@ -41,40 +41,6 @@ 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:

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/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
 

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:

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:

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:

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/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

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:

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

@@ -27,7 +27,7 @@ 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:

examples/pixtral/lora-12b.yml

@@ -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:

examples/qwen3/8b-qat-fsdp2.yml

@@ -1,78 +0,0 @@
-base_model: Qwen/Qwen3-8B
-# 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: tatsu-lab/alpaca
-    type: alpaca
-
-output_dir: ./outputs/qat_out/
-
-sequence_len: 2048
-sample_packing: true
-flex_attention: true
-pad_to_sequence_len: true
-
-flex_attn_compile_kwargs:
-  dynamic: false
-  mode: max-autotune-no-cudagraphs
-
-qat:
-  activation_dtype: int8
-  weight_dtype: int4
-  group_size: 256
-  fake_quant_after_n_steps: 1000
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 1
-micro_batch_size: 2
-max_steps: 2000
-optimizer: adamw_torch_fused
-lr_scheduler: cosine
-learning_rate: 2e-5
-
-bf16: true
-tf32: true
-
-resume_from_checkpoint:
-logging_steps: 1
-
-evals_per_epoch: 1
-saves_per_epoch: 1
-
-warmup_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: Qwen3DecoderLayer
-  fsdp_state_dict_type: FULL_STATE_DICT
-  fsdp_sharding_strategy: FULL_SHARD
-  fsdp_reshard_after_forward: true
-  fsdp_activation_checkpointing: true
-
-special_tokens:

requirements.txt

@@ -6,20 +6,21 @@ triton>=3.0.0
 mamba-ssm==1.2.0.post1
 xformers>=0.0.23.post1
 autoawq==0.2.7.post3
-liger-kernel==0.5.10
+liger-kernel==0.5.9
 # END section
 
 packaging==23.2
 
-huggingface_hub==0.32.2
+huggingface_hub==0.31.0
 peft==0.15.2
-transformers==4.52.3
+transformers==4.51.3
 tokenizers>=0.21.1
-accelerate==1.7.0
+accelerate==1.6.0
-datasets==3.6.0
+datasets==3.5.1
-deepspeed>=0.17.0
+deepspeed>=0.15.4
-trl==0.18.1
+trl==0.17.0
-hf_xet==1.1.2
+hf_xet==1.1.0
+hqq==0.2.5
 
 optimum==1.16.2
 hf_transfer
@@ -62,7 +63,7 @@ langdetect==1.0.9
 immutabledict==4.2.0
 antlr4-python3-runtime==4.13.2
 
-torchao==0.10.0
+torchao==0.9.0
 schedulefree==1.4.1
 
 axolotl-contribs-lgpl==0.0.6

scripts/cutcrossentropy_install.py

@@ -9,8 +9,6 @@ except ImportError as exc:
     raise ImportError("Install torch via `pip install torch`") from exc
 from packaging.version import Version as V
 
-USE_UV = "--uv" in sys.argv[1:]
-
 v = V(torch.__version__)
 
 # no cut-cross-entropy support for torch < 2.4.0
@@ -25,9 +23,7 @@ if cce_spec:
     if not importlib.util.find_spec("cut_cross_entropy.transformers"):
         UNINSTALL_PREFIX = "pip uninstall -y cut-cross-entropy && "
 
-UV_PREFIX = "uv " if USE_UV else ""
-
 print(
     UNINSTALL_PREFIX
-    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@a1174ca"'
+    + 'pip install "cut-cross-entropy[transformers] @ git+https://github.com/apple/ml-cross-entropy.git@bad6f7b49c75fdec69471abb71b4cddd0f0c6438"'
 )

scripts/motd

@@ -11,7 +11,7 @@
                                  =@#       @#  #@=     #@   =#@@@@#=    +#@@=  +#@@@@#=    .##@@+   @@
     @@@@  @@@@@@@@@@@@@@@@
 
-Welcome to the axolotl cloud image! If the you've mounted a disk to /workspace and the axolotl directory is empty, run the following commands:
+Welcome to the axolotl cloud image! If the you've mounted a disk to /workspace and the axolotl directory ie empty, run the following commands:
 
 ```
 cd /workspace

scripts/unsloth_install.py

@@ -1,15 +1,11 @@
 # noqa
 # pylint: skip-file
-import sys
-
 try:
     import torch
 except ImportError:
     raise ImportError("Install torch via `pip install torch`")
 from packaging.version import Version as V
 
-use_uv = "--uv" in sys.argv[1:]
-
 v = V(torch.__version__)
 cuda = str(torch.version.cuda)
 try:
@@ -35,7 +31,6 @@ elif v < V("2.6.0"):
 else:
     raise RuntimeError(f"Torch = {v} too new!")
 x = x.format(cuda.replace(".", ""), "-ampere" if is_ampere else "")
-uv_prefix = "uv " if use_uv else ""
 print(
-    f'{uv_prefix}pip install unsloth-zoo==2024.12.1 && {uv_prefix}pip install --no-deps "unsloth[{x}]==2024.12.4"'
+    f'pip install unsloth-zoo==2024.12.1 && pip install --no-deps "unsloth[{x}]==2024.12.4"'
 )

setup.py

@@ -118,7 +118,7 @@ extras_require = {
         "yunchang==0.6.0",
     ],
     "deepspeed": [
-        "deepspeed==0.17.0",
+        "deepspeed==0.15.4",
         "deepspeed-kernels",
     ],
     "mamba-ssm": [

src/axolotl/cli/args.py

@@ -28,6 +28,7 @@ class TrainerCliArgs:
     debug: bool = field(default=False)
     debug_text_only: bool = field(default=False)
     debug_num_examples: int = field(default=0)
+    merge_lora: bool = field(default=False)
     prompter: Optional[str] = field(default=None)
     shard: bool = field(default=False)
     main_process_port: Optional[int] = field(default=None)
@@ -88,26 +89,6 @@ class VllmServeCliArgs:
         },
     )
 
-    enable_reasoning: Optional[bool] = field(
-        default=None,
-    )
-
-    reasoning_parser: Optional[str] = field(
-        default=None,
-    )
-
-
-@dataclass
-class QuantizeCliArgs:
-    """Dataclass with CLI arguments for `axolotl quantize` command."""
-
-    base_model: Optional[str] = field(default=None)
-    weight_dtype: Optional[str] = field(default=None)
-    activation_dtype: Optional[str] = field(default=None)
-    quantize_embedding: Optional[bool] = field(default=None)
-    group_size: Optional[int] = field(default=None)
-    output_dir: Optional[str] = field(default=None)
-
 
 @dataclass
 class EvaluateCliArgs:

src/axolotl/cli/checks.py

@@ -1,5 +1,6 @@
 """Various checks for Axolotl CLI."""
 
+import logging
 import os
 from pathlib import Path
 
@@ -7,9 +8,7 @@ from accelerate.commands.config import config_args
 from huggingface_hub import HfApi
 from huggingface_hub.utils import LocalTokenNotFoundError
 
-from axolotl.utils.logging import get_logger
+LOG = logging.getLogger(__name__)
-
-LOG = get_logger(__name__)
 
 
 def check_accelerate_default_config() -> None:

src/axolotl/cli/cloud/modal_.py

@@ -82,7 +82,7 @@ class ModalCloud(Cloud):
         return res
 
     def get_image(self):
-        docker_tag = "main-py3.11-cu124-2.6.0"
+        docker_tag = "main-py3.11-cu124-2.5.1"
         if self.config.docker_tag:
             docker_tag = self.config.docker_tag
         docker_image = f"axolotlai/axolotl:{docker_tag}"

src/axolotl/cli/config.py

@@ -1,6 +1,7 @@
 """Configuration loading and processing."""
 
 import json
+import logging
 import os
 import tempfile
 from pathlib import Path
@@ -21,12 +22,11 @@ from axolotl.utils.config import (
     validate_config,
 )
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.logging import get_logger
 from axolotl.utils.mlflow_ import setup_mlflow_env_vars
 from axolotl.utils.trainer import prepare_opinionated_env, prepare_optim_env
 from axolotl.utils.wandb_ import setup_wandb_env_vars
 
-LOG = get_logger(__name__, use_environ=True)
+LOG = logging.getLogger(__name__)
 
 
 def check_remote_config(config: Union[str, Path]) -> Union[str, Path]:
@@ -119,12 +119,12 @@ def choose_config(path: Path) -> str:
         )
 
     if len(yaml_files) == 1:
-        LOG.info(f"Using default YAML file '{yaml_files[0]}'")
+        print(f"Using default YAML file '{yaml_files[0]}'")
         return str(yaml_files[0])
 
-    LOG.info("Choose a YAML file:")
+    print("Choose a YAML file:")
     for idx, file in enumerate(yaml_files):
-        LOG.info(f"{idx + 1}. {file}")
+        print(f"{idx + 1}. {file}")
 
     chosen_file = None
     while chosen_file is None:
@@ -133,9 +133,9 @@ def choose_config(path: Path) -> str:
             if 1 <= choice <= len(yaml_files):
                 chosen_file = str(yaml_files[choice - 1])
             else:
-                LOG.info("Invalid choice. Please choose a number from the list.")
+                print("Invalid choice. Please choose a number from the list.")
         except ValueError:
-            LOG.info("Invalid input. Please enter a number.")
+            print("Invalid input. Please enter a number.")
 
     return chosen_file
 

src/axolotl/cli/evaluate.py

@@ -1,5 +1,6 @@
 """CLI to run evaluation on a model."""
 
+import logging
 import os
 from pathlib import Path
 from typing import Union
@@ -16,9 +17,8 @@ from axolotl.common.datasets import load_datasets, load_preference_datasets
 from axolotl.evaluate import evaluate
 from axolotl.utils import patch_optimized_env
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.logging import get_logger
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger(__name__)
 
 
 def do_evaluate(cfg: DictDefault, cli_args: TrainerCliArgs) -> None:

src/axolotl/cli/inference.py

@@ -1,6 +1,7 @@
 """CLI to run inference on a trained model."""
 
 import importlib
+import logging
 import sys
 from pathlib import Path
 from threading import Thread
@@ -21,9 +22,8 @@ from axolotl.utils.chat_templates import (
     get_chat_template_from_config,
 )
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.logging import get_logger
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger(__name__)
 
 
 def get_multi_line_input() -> str:

src/axolotl/cli/main.py

@@ -2,6 +2,7 @@
 
 # pylint: disable=redefined-outer-name
 
+import logging
 import os
 import subprocess  # nosec B404
 import tempfile
@@ -16,7 +17,6 @@ import axolotl
 from axolotl.cli.args import (
     EvaluateCliArgs,
     PreprocessCliArgs,
-    QuantizeCliArgs,
     TrainerCliArgs,
     VllmServeCliArgs,
 )
@@ -30,11 +30,8 @@ from axolotl.cli.utils import (
 )
 from axolotl.integrations.lm_eval.cli import lm_eval
 from axolotl.utils import patch_optimized_env
-from axolotl.utils.logging import get_logger
 from axolotl.utils.schemas.config import AxolotlInputConfig
 
-LOG = get_logger(__name__)
-
 
 @click.group()
 @click.version_option(version=axolotl.__version__, prog_name="axolotl")
@@ -179,7 +176,7 @@ def train(
 
                     do_cli(config=cfg_file, **kwargs)
         except subprocess.CalledProcessError as exc:
-            LOG.error(f"Failed to train/fine-tune config '{cfg_file}': {exc}")
+            logging.error(f"Failed to train/fine-tune config '{cfg_file}': {exc}")
             if not sweep:
                 raise exc
 
@@ -336,16 +333,6 @@ def vllm_serve(config: str, **cli_args: VllmServeCliArgs):
     do_vllm_serve(config, cli_args)
 
 
-@cli.command()
-@click.argument("config", type=click.Path(exists=True, path_type=str))
-@add_options_from_dataclass(QuantizeCliArgs)
-@filter_none_kwargs
-def quantize(config: str, **cli_args: QuantizeCliArgs):
-    from axolotl.cli.quantize import do_quantize
-
-    do_quantize(config, cli_args)
-
-
 @cli.command()
 @click.argument("model", type=click.Path(exists=True, path_type=str))
 @click.argument("output", type=click.Path(exists=False, path_type=str))

src/axolotl/cli/merge_lora.py

@@ -1,18 +1,20 @@
 """CLI to merge a trained LoRA into a base model."""
 
+import logging
 from pathlib import Path
 from typing import Union
 
 import fire
+import transformers
 from dotenv import load_dotenv
 
+from axolotl.cli.args import TrainerCliArgs
 from axolotl.cli.art import print_axolotl_text_art
 from axolotl.cli.config import load_cfg
 from axolotl.cli.utils import load_model_and_tokenizer
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.logging import get_logger
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger(__name__)
 
 
 def do_merge_lora(*, cfg: DictDefault) -> None:
@@ -66,6 +68,12 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs) -> None:
     Raises:
         ValueError: If target directory for LoRA merged model does not exist.
     """
+    # pylint: disable=duplicate-code
+    parser = transformers.HfArgumentParser(TrainerCliArgs)
+    parsed_cli_args, _ = parser.parse_args_into_dataclasses(
+        return_remaining_strings=True
+    )
+    parsed_cli_args.merge_lora = True
 
     parsed_cfg = load_cfg(
         config,

src/axolotl/cli/merge_sharded_fsdp_weights.py

@@ -1,6 +1,7 @@
 """CLI to merge sharded FSDP model checkpoints into a single combined checkpoint."""
 
 import json
+import logging
 import os
 import shutil
 from pathlib import Path
@@ -10,6 +11,7 @@ import fire
 import torch
 import torch.distributed.checkpoint as dist_cp
 import torch.distributed.checkpoint.format_utils as dist_cp_format_utils
+import transformers
 from accelerate.utils import (
     SAFE_WEIGHTS_INDEX_NAME,
     SAFE_WEIGHTS_NAME,
@@ -22,11 +24,11 @@ from huggingface_hub import split_torch_state_dict_into_shards
 from safetensors.torch import save_file as safe_save_file
 from torch.distributed.checkpoint.format_utils import _EmptyStateDictLoadPlanner
 
+from axolotl.cli.args import TrainerCliArgs
 from axolotl.cli.art import print_axolotl_text_art
 from axolotl.cli.config import load_cfg
-from axolotl.utils.logging import get_logger
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger(__name__)
 
 
 class BFloat16CastPlanner(_EmptyStateDictLoadPlanner):
@@ -195,6 +197,11 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
     """
     # pylint: disable=duplicate-code
     print_axolotl_text_art()
+    parser = transformers.HfArgumentParser(TrainerCliArgs)
+    parsed_cli_args, _ = parser.parse_args_into_dataclasses(
+        return_remaining_strings=True
+    )
+    parsed_cli_args.merge_lora = True
     parsed_cfg = load_cfg(config, **kwargs)
 
     fsdp_dir = Path(parsed_cfg.output_dir) / "pytorch_model_fsdp_0"

src/axolotl/cli/preprocess.py

@@ -1,5 +1,6 @@
 """CLI to run preprocessing of a dataset."""
 
+import logging
 import warnings
 from pathlib import Path
 from typing import Union
@@ -19,10 +20,9 @@ from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
 from axolotl.common.datasets import load_datasets, load_preference_datasets
 from axolotl.integrations.base import PluginManager
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.logging import get_logger
 from axolotl.utils.trainer import disable_datasets_caching
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger(__name__)
 
 
 def do_preprocess(cfg: DictDefault, cli_args: PreprocessCliArgs) -> None:

src/axolotl/cli/quantize.py

@@ -1,90 +0,0 @@
-"""
-CLI to post-training quantize a model using torchao
-"""
-
-from pathlib import Path
-from typing import Union
-
-from transformers import AutoModelForCausalLM
-
-from axolotl.cli.art import print_axolotl_text_art
-from axolotl.cli.config import load_cfg
-from axolotl.loaders import load_tokenizer
-from axolotl.utils.logging import get_logger
-from axolotl.utils.quantization import TorchIntDType, quantize_model_for_ptq
-
-LOG = get_logger(__name__)
-
-
-def do_quantize(
-    config: Union[Path, str],
-    cli_args: dict,
-):
-    """
-    Quantizes a model's model's weights
-
-    Args:
-        config (Union[Path, str]): The path to the config file
-        cli_args (dict): Additional command-line arguments
-    """
-    print_axolotl_text_art()
-
-    cfg = load_cfg(config)
-
-    if cfg.qat and cfg.quantization:
-        raise ValueError(
-            "QAT and quantization cannot be used together. Please specify only one of qat or quantization in your config file."
-        )
-
-    if cfg.qat:
-        quantize_cfg = cfg.qat
-    elif cfg.quantization:
-        quantize_cfg = cfg.quantization
-    else:
-        raise ValueError(
-            "No quantization configuration found. Please specify either qat or quantization in your config file."
-        )
-
-    model_path = cli_args.get("model_path") or cfg.output_dir
-    if weight_dtype := cli_args.get("weight_dtype"):
-        weight_dtype = TorchIntDType[weight_dtype]
-    else:
-        weight_dtype = quantize_cfg.weight_dtype
-    if activation_dtype := cli_args.get("activation_dtype"):
-        activation_dtype = TorchIntDType[activation_dtype]
-    else:
-        activation_dtype = quantize_cfg.activation_dtype
-    group_size = cli_args.get("group_size") or quantize_cfg.group_size
-    quantize_embedding = (
-        cli_args.get("quantize_embedding") or quantize_cfg.quantize_embedding
-    )
-    output_dir = cli_args.get("output_dir") or cfg.output_dir
-
-    LOG.info(f"Loading model from {model_path}...")
-    tokenizer = load_tokenizer(cfg)
-    model = AutoModelForCausalLM.from_pretrained(model_path, device_map="auto")
-
-    LOG.info(
-        f"Quantizing model with configuration: \n"
-        f"\tweight_dtype: {weight_dtype}\n"
-        f"\tactivation_dtype: {activation_dtype}\n"
-        f"\tgroup_size: {group_size}\n"
-        f"\tquantize_embedding: {quantize_embedding}"
-    )
-
-    quantize_model_for_ptq(
-        model, weight_dtype, group_size, activation_dtype, quantize_embedding
-    )
-
-    LOG.info(f"Saving quantized model to: {str(Path(output_dir) / 'quantized')}...")
-    model.save_pretrained(
-        str(Path(output_dir) / "quantized"),
-        safe_serialization=False,
-        progressbar=True,
-    )
-    tokenizer.save_pretrained(
-        str(Path(output_dir) / "quantized"),
-        safe_serialization=False,
-        progressbar=True,
-    )
-    LOG.info(f"Quantized model saved to: {str(Path(output_dir) / 'quantized')}...")

src/axolotl/cli/train.py

@@ -1,6 +1,7 @@
 """CLI to run training on a model."""
 
 import gc
+import logging
 import os
 from pathlib import Path
 from typing import Union
@@ -21,6 +22,8 @@ from axolotl.utils import patch_optimized_env
 from axolotl.utils.config import normalize_config, resolve_dtype
 from axolotl.utils.dict import DictDefault
 
+LOG = logging.getLogger(__name__)
+
 
 def do_train(cfg: DictDefault, cli_args: TrainerCliArgs):
     """

src/axolotl/cli/utils.py

@@ -4,6 +4,7 @@ import concurrent.futures
 import dataclasses
 import hashlib
 import json
+import logging
 from functools import wraps
 from pathlib import Path
 from types import NoneType
@@ -22,9 +23,8 @@ from transformers import (
 from axolotl.loaders import load_processor, load_tokenizer
 from axolotl.loaders.model import ModelLoader
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.logging import get_logger
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger(__name__)
 
 
 def strip_optional_type(field_type: type | str | None):

src/axolotl/cli/vllm_serve.py

@@ -2,27 +2,14 @@
 CLI to start the vllm server for online RL
 """
 
-import os
-from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Union
 
-import trl
 from trl.scripts.vllm_serve import ScriptArguments
 
 from axolotl.cli.config import load_cfg
 
 
-@dataclass
-class AxolotlScriptArguments(ScriptArguments):
-    """
-    Additional arguments for the VLLM server
-    """
-
-    reasoning_parser: str = field(default="", kw_only=True)
-    enable_reasoning: bool | None = field(default=None, kw_only=True)
-
-
 def do_vllm_serve(
     config: Union[Path, str],
     cli_args: dict,
@@ -37,7 +24,6 @@ def do_vllm_serve(
     Returns:
         process_id: the process id of the started VLLM server
     """
-    patch_vllm_worker()
     cfg = load_cfg(config)
     model = cfg.base_model
 
@@ -57,16 +43,9 @@ def do_vllm_serve(
     enable_prefix_caching = (
         cli_args.get("enable_prefix_caching") or cfg.vllm.enable_prefix_caching
     )
-    reasoning_parser = (
-        cli_args.get("reasoning_parser") or cfg.vllm.reasoning_parser or ""
-    )
-    enable_reasoning = (
-        cli_args.get("enable_reasoning") or cfg.vllm.enable_reasoning or False
-    )
 
-    # pylint: disable=unexpected-keyword-arg
+    vllm_script_args = ScriptArguments(
-    vllm_script_args = AxolotlScriptArguments(
+        model,
-        model=model,
         tensor_parallel_size=tensor_parallel_size,
         host=host,
         port=port,
@@ -74,67 +53,5 @@ def do_vllm_serve(
         dtype=dtype,
         max_model_len=max_model_len,
         enable_prefix_caching=enable_prefix_caching,
-        reasoning_parser=reasoning_parser,
-        enable_reasoning=enable_reasoning,
     )
     vllm_serve_main(vllm_script_args)
-
-
-def patch_vllm_worker():
-    from multiprocessing.connection import Connection
-
-    from vllm import LLM
-
-    def llm_worker(
-        script_args: AxolotlScriptArguments,
-        data_parallel_rank: int,
-        master_port: int,
-        connection: Connection,
-    ) -> None:
-        # Set required environment variables for DP to work with vLLM
-        os.environ["VLLM_DP_RANK"] = str(data_parallel_rank)
-        os.environ["VLLM_DP_RANK_LOCAL"] = str(data_parallel_rank)
-        os.environ["VLLM_DP_SIZE"] = str(script_args.data_parallel_size)
-        os.environ["VLLM_DP_MASTER_PORT"] = str(master_port)
-
-        llm = LLM(
-            model=script_args.model,
-            revision=script_args.revision,
-            tensor_parallel_size=script_args.tensor_parallel_size,
-            gpu_memory_utilization=script_args.gpu_memory_utilization,
-            enforce_eager=script_args.enforce_eager,
-            dtype=script_args.dtype,
-            # Automatic Prefix Caching caches the KV cache of existing queries, so that a new query can
-            # directly reuse the KV cache if it shares the same prefix with one of the existing queries.
-            # This is particularly useful here because we generate completions from the same prompts.
-            enable_prefix_caching=script_args.enable_prefix_caching,
-            kv_cache_dtype=script_args.kv_cache_dtype,
-            max_model_len=script_args.max_model_len,
-            worker_extension_cls="trl.scripts.vllm_serve.WeightSyncWorkerExtension",
-            enable_reasoning=script_args.enable_reasoning,
-            reasoning_parser=script_args.reasoning_parser,
-        )
-
-        # Send ready signal to parent process
-        connection.send({"status": "ready"})
-
-        while True:
-            # Wait for commands from the parent process
-            try:
-                command = connection.recv()
-            except KeyboardInterrupt:
-                llm.collective_rpc(method="close_communicator")
-                break
-
-            # Handle commands
-            if command["type"] in ["call", "fire_and_forget"]:
-                method_name = command["method"]
-                args, kwargs = command.get("args", ()), command.get("kwargs", {})
-                method = getattr(llm, method_name)
-                result = method(*args, **kwargs)
-                if command["type"] == "call":
-                    connection.send(result)
-            elif command["type"] == "shutdown":
-                break
-
-    trl.scripts.vllm_serve.llm_worker = llm_worker

src/axolotl/common/datasets.py

@@ -1,5 +1,6 @@
 """Dataset loading utilities."""
 
+import logging
 import math
 import random
 from dataclasses import dataclass
@@ -13,11 +14,10 @@ from axolotl.loaders import load_processor, load_tokenizer
 from axolotl.utils.data import prepare_dataset
 from axolotl.utils.data.rl import load_prepare_preference_datasets
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.logging import get_logger
 from axolotl.utils.schemas.enums import RLType
 from axolotl.utils.tokenization import check_dataset_labels
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger(__name__)
 
 
 @dataclass

src/axolotl/core/builders/__init__.py

@@ -1,6 +0,0 @@
-"""Trainer builder classes"""
-
-from .causal import HFCausalTrainerBuilder
-from .rl import HFRLTrainerBuilder
-
-__all__ = ["HFCausalTrainerBuilder", "HFRLTrainerBuilder"]

src/axolotl/core/builders/base.py

@@ -1,506 +0,0 @@
-# Copyright 2024 Axolotl AI. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""Base class for trainer builder"""
-
-import abc
-import importlib
-import logging
-import sys
-from abc import abstractmethod
-from contextlib import suppress
-from pathlib import Path
-from typing import Any
-
-import torch
-from transformers import (
-    TrainerCallback,
-)
-from transformers.training_args import OptimizerNames
-
-from axolotl.integrations.base import PluginManager
-from axolotl.monkeypatch.trainer.lr import patch_trainer_get_lr
-from axolotl.utils import is_comet_available, is_mlflow_available
-from axolotl.utils.callbacks import (
-    GCCallback,
-    GPUStatsCallback,
-    SaveAxolotlConfigtoWandBCallback,
-)
-from axolotl.utils.callbacks.profiler import PytorchProfilerCallback
-from axolotl.utils.schemas.enums import CustomSupportedOptimizers
-
-LOG = logging.getLogger(__name__)
-
-with suppress(ImportError):
-    import torch._dynamo  # pylint: disable=ungrouped-imports
-
-
-class TrainerBuilderBase(abc.ABC):
-    """Base class for trainer builder."""
-
-    def __init__(self, cfg, model, tokenizer, processor=None):
-        self.cfg = cfg
-        self.model = model
-        self.tokenizer = tokenizer
-        self.processor = processor
-
-        self._train_dataset = None
-        self._eval_dataset = None
-        self._model_ref = None
-        self._peft_config = None
-
-        # If the model supports tagging, add the axolotl tag.
-        # This makes sure the tag is correctly pushed even if a user calls
-        # model.push_to_hub instead of trainer.push_to_hub.
-        if hasattr(model, "add_model_tags"):
-            model.add_model_tags(["axolotl"])
-
-        patch_trainer_get_lr()
-
-    @property
-    def model_ref(self):
-        return self._model_ref
-
-    @model_ref.setter
-    def model_ref(self, model):
-        self._model_ref = model
-
-    @property
-    def train_dataset(self):
-        return self._train_dataset
-
-    @train_dataset.setter
-    def train_dataset(self, dataset):
-        self._train_dataset = dataset
-
-    @property
-    def eval_dataset(self):
-        return self._eval_dataset
-
-    @eval_dataset.setter
-    def eval_dataset(self, dataset):
-        self._eval_dataset = dataset
-
-    @property
-    def peft_config(self):
-        return self._peft_config
-
-    @peft_config.setter
-    def peft_config(self, peft_config):
-        self._peft_config = peft_config
-
-    @abstractmethod
-    def build(self, total_num_steps):
-        pass
-
-    def get_callbacks(self) -> list[TrainerCallback]:
-        callbacks = []
-
-        plugin_manager = PluginManager.get_instance()
-        callbacks.extend(
-            plugin_manager.add_callbacks_pre_trainer(cfg=self.cfg, model=self.model)
-        )
-
-        if self.cfg.profiler_steps:
-            callbacks.append(
-                PytorchProfilerCallback(
-                    steps_to_profile=self.cfg.profiler_steps,
-                )
-            )
-
-        if self.cfg.gc_steps:
-            callbacks.append(GCCallback(gc_steps=self.cfg.gc_steps))
-
-        if self.cfg.use_wandb:
-            callbacks.append(
-                SaveAxolotlConfigtoWandBCallback(self.cfg.axolotl_config_path)
-            )
-        if self.cfg.use_mlflow and is_mlflow_available():
-            from axolotl.utils.callbacks.mlflow_ import (
-                SaveAxolotlConfigtoMlflowCallback,
-            )
-
-            callbacks.extend(
-                [
-                    SaveAxolotlConfigtoMlflowCallback(self.cfg.axolotl_config_path),
-                ]
-            )
-        if self.cfg.use_comet and is_comet_available():
-            from axolotl.utils.callbacks.comet_ import SaveAxolotlConfigtoCometCallback
-
-            callbacks.append(
-                SaveAxolotlConfigtoCometCallback(self.cfg.axolotl_config_path)
-            )
-
-        callbacks.append(GPUStatsCallback(cfg=self.cfg))
-
-        return callbacks
-
-    def get_post_trainer_create_callbacks(self, trainer):
-        """
-        Callbacks added after the trainer is created, usually b/c these need access to the trainer
-        """
-        callbacks = []
-        if self.cfg.plugins:
-            plugin_manager = PluginManager.get_instance()
-            callbacks.extend(
-                [
-                    cb
-                    for cb in plugin_manager.add_callbacks_post_trainer(
-                        self.cfg, trainer
-                    )
-                    if cb
-                ]
-            )
-        return callbacks
-
-    def hook_pre_create_training_args(self, training_arguments_kwargs):
-        # TODO
-        return training_arguments_kwargs
-
-    def hook_post_create_training_args(self, training_arguments):
-        # TODO
-        return training_arguments
-
-    def hook_pre_create_trainer(self, trainer_kwargs, trainer_cls):
-        # TODO
-        return trainer_kwargs, trainer_cls
-
-    def hook_post_create_trainer(self, trainer):
-        # TODO
-        return trainer
-
-    def _configure_warmup_and_logging(
-        self, total_num_steps: int, training_args_kwargs: dict
-    ):
-        warmup_steps = 0
-        warmup_ratio = 0.0
-        if self.cfg.warmup_steps:
-            warmup_steps = self.cfg.warmup_steps
-        elif self.cfg.warmup_ratio:
-            if total_num_steps:
-                warmup_steps = max(int(self.cfg.warmup_ratio * total_num_steps), 0)
-            else:
-                warmup_ratio = self.cfg.warmup_ratio
-        elif total_num_steps:
-            warmup_steps = min(int(0.03 * total_num_steps), 100)
-        else:
-            warmup_ratio = 0.03
-
-        if warmup_steps == 1:
-            warmup_steps = 2
-
-        if self.cfg.logging_steps is not None:
-            training_args_kwargs["logging_steps"] = self.cfg.logging_steps
-        else:
-            training_args_kwargs["logging_steps"] = (
-                500  # transformers defaults to 500
-                if not total_num_steps
-                else max(min(int(0.005 * total_num_steps), 10), 1)
-            )
-
-        training_args_kwargs["warmup_ratio"] = warmup_ratio
-        training_args_kwargs["warmup_steps"] = warmup_steps
-
-    def _configure_precision_settings(self, training_args_kwargs: dict):
-        training_args_kwargs["fp16"] = (self.cfg.fp16 and not self.cfg.bf16) or False
-        training_args_kwargs["tf32"] = self.cfg.tf32
-        if self.cfg.bf16 == "full":
-            training_args_kwargs["bf16_full_eval"] = True
-        else:
-            training_args_kwargs["bf16"] = self.cfg.bf16 or self.cfg.bfloat16
-
-    def _configure_scheduler(self, training_args_kwargs: dict):
-        if self.cfg.lr_scheduler in ["one_cycle", "rex"]:
-            training_args_kwargs["lr_scheduler_type"] = "cosine"
-            training_args_kwargs["alternate_lr_scheduler_type"] = self.cfg.lr_scheduler
-        else:
-            training_args_kwargs["lr_scheduler_type"] = (
-                self.cfg.lr_scheduler if self.cfg.lr_scheduler else "cosine"
-            )
-        training_args_kwargs["lr_scheduler_kwargs"] = (
-            self.cfg.lr_scheduler_kwargs if self.cfg.lr_scheduler_kwargs else {}
-        )
-
-    def _configure_optimizer(self, training_args_kwargs: dict, trainer_kwargs: dict):
-        def _configure_custom_optimizer(
-            training_args_kwargs: dict, trainer_kwargs: dict
-        ):
-            # Common optimizer kwargs
-            optimizer_kwargs = {
-                "lr": training_args_kwargs["learning_rate"],
-                "weight_decay": training_args_kwargs["weight_decay"],
-            }
-
-            # Adam-specific kwargs
-            adam_kwargs: dict = {}
-            if training_args_kwargs.get("adam_beta1") and training_args_kwargs.get(
-                "adam_beta2"
-            ):
-                adam_kwargs["betas"] = (
-                    training_args_kwargs.get("adam_beta1"),
-                    training_args_kwargs.get("adam_beta2"),
-                )
-            if training_args_kwargs.get("adam_epsilon"):
-                adam_kwargs["eps"] = training_args_kwargs.get("adam_epsilon")
-
-            if self.cfg.optimizer == "muon":
-                from axolotl.contribs.mit.muon import (  # pylint: disable=no-name-in-module
-                    MuonOptimizerFactory,
-                )
-
-                optimizer_cls = MuonOptimizerFactory
-                optimizer_kwargs.update(adam_kwargs)
-            elif self.cfg.optimizer == "optimi_adamw":
-                from optimi import AdamW
-
-                optimizer_kwargs["foreach"] = False
-                optimizer_cls = AdamW
-                optimizer_kwargs.update(adam_kwargs)
-            elif self.cfg.optimizer == "ao_adamw_4bit":
-                # TODO remove 20250401
-                from torchao.prototype.low_bit_optim import AdamW4bit
-
-                optimizer_cls = AdamW4bit
-                optimizer_kwargs.update(adam_kwargs)
-
-                LOG.warning(
-                    f"`ao_adamw_4bit` will be deprecated soon. Please use `{OptimizerNames.ADAMW_TORCH_4BIT}` instead."
-                )
-            elif self.cfg.optimizer == "ao_adamw_8bit":
-                from torchao.prototype.low_bit_optim import AdamW8bit
-
-                optimizer_cls = AdamW8bit
-                optimizer_kwargs.update(adam_kwargs)
-            elif self.cfg.optimizer == "ao_adamw_fp8":
-                from torchao.prototype.low_bit_optim import AdamWFp8
-
-                optimizer_cls = AdamWFp8
-                optimizer_kwargs.update(adam_kwargs)
-            elif self.cfg.optimizer == "adopt_adamw":
-                from axolotl.utils.optimizers.adopt import ADOPT
-
-                optimizer_cls = ADOPT
-                adam_kwargs["decouple"] = True
-                optimizer_kwargs.update(adam_kwargs)
-            elif self.cfg.optimizer == "came_pytorch":
-                from came_pytorch import CAME
-
-                optimizer_cls = CAME
-
-                beta1 = training_args_kwargs.get("adam_beta1", 0.9)
-                beta2 = training_args_kwargs.get("adam_beta2", 0.999)
-                beta3 = training_args_kwargs.get("adam_beta3", 0.9999)
-                eps1 = training_args_kwargs.get("adam_epsilon", 1e-30)
-                eps2 = training_args_kwargs.get("adam_epsilon2", 1e-16)
-                adam_kwargs["betas"] = (beta1, beta2, beta3)
-                adam_kwargs["eps"] = (eps1, eps2)
-
-                optimizer_kwargs.update(adam_kwargs)
-            else:
-                raise ValueError(
-                    f"Unhandled optimizer: {self.cfg.optimizer}. Please raise an Issue."
-                )
-
-            # Parse any additional optimizer args from config
-            if self.cfg.optim_args:
-                if isinstance(self.cfg.optim_args, dict):
-                    optimizer_kwargs.update(self.cfg.optim_args)
-                else:
-                    # Parse string format "key1=value1,key2=value2"
-                    for mapping in self.cfg.optim_args.replace(" ", "").split(","):
-                        key, value = mapping.split("=")
-                        optimizer_kwargs[key] = value
-
-            # Note: This is not used in training_args_kwargs, but in trainer_kwargs
-            trainer_kwargs["optimizer_cls_and_kwargs"] = (
-                optimizer_cls,
-                optimizer_kwargs,
-            )
-
-        # Handle custom optimizer
-        custom_supported_optimizers = [opt.value for opt in CustomSupportedOptimizers]
-        if self.cfg.optimizer in custom_supported_optimizers:
-            _configure_custom_optimizer(training_args_kwargs, trainer_kwargs)
-        else:
-            # Use transformers' optimizer
-            training_args_kwargs["optim"] = self.cfg.optimizer
-
-            # Parse any additional optimizer args from config
-            if self.cfg.optim_args:
-                if isinstance(self.cfg.optim_args, dict):
-                    optim_args = ",".join(
-                        [f"{key}={value}" for key, value in self.cfg.optim_args.items()]
-                    )
-                else:
-                    optim_args = self.cfg.optim_args
-                training_args_kwargs["optim_args"] = optim_args
-
-            if (
-                self.cfg.optimizer == "adamw_anyprecision"
-                and Path(self.cfg.torchdistx_path).exists()
-            ):
-                sys.path.append(self.cfg.torchdistx_path)
-                importlib.import_module("torchdistx")
-
-    def _configure_hub_parameters(self, training_args_kwargs: dict):
-        if self.cfg.hub_model_id:
-            training_args_kwargs["hub_model_id"] = self.cfg.hub_model_id
-            training_args_kwargs["push_to_hub"] = True
-            training_args_kwargs["hub_private_repo"] = True
-            training_args_kwargs["hub_always_push"] = True
-
-            if self.cfg.hub_strategy:
-                training_args_kwargs["hub_strategy"] = self.cfg.hub_strategy
-
-    def _configure_save_and_eval_strategy(self, training_args_kwargs: dict):
-        # save_strategy and save_steps
-        if self.cfg.save_steps:
-            training_args_kwargs["save_strategy"] = "steps"
-            training_args_kwargs["save_steps"] = self.cfg.save_steps
-        elif self.cfg.save_strategy:
-            training_args_kwargs["save_strategy"] = self.cfg.save_strategy
-        else:
-            # default to saving each epoch if not defined
-            training_args_kwargs["save_strategy"] = "epoch"
-
-        training_args_kwargs["save_total_limit"] = (
-            self.cfg.save_total_limit if self.cfg.save_total_limit else 4
-        )
-
-        # eval_strategy and eval_steps
-        if not self.eval_dataset or self.cfg.val_set_size == 0:
-            # do not eval if no eval_dataset or val_set_size=0
-            training_args_kwargs["eval_strategy"] = "no"
-        elif self.cfg.eval_steps:
-            training_args_kwargs["eval_strategy"] = "steps"
-            training_args_kwargs["eval_steps"] = self.cfg.eval_steps
-        elif self.cfg.eval_strategy:
-            training_args_kwargs["eval_strategy"] = self.cfg.eval_strategy
-
-    def _configure_reporting(self, training_args_kwargs: dict):
-        report_to = []
-        if self.cfg.use_wandb:
-            report_to.append("wandb")
-        if self.cfg.use_mlflow:
-            report_to.append("mlflow")
-        if self.cfg.use_tensorboard:
-            report_to.append("tensorboard")
-        if self.cfg.use_comet:
-            report_to.append("comet_ml")
-
-        training_args_kwargs["report_to"] = report_to
-
-        if self.cfg.use_wandb:
-            training_args_kwargs["run_name"] = self.cfg.wandb_name
-        elif self.cfg.use_mlflow:
-            training_args_kwargs["run_name"] = self.cfg.mlflow_run_name
-        else:
-            training_args_kwargs["run_name"] = None
-
-    def _configure_torch_compile(self, training_args_kwargs: dict):
-        if self.cfg.torch_compile and getattr(torch, "_dynamo", None):
-            torch._dynamo.config.suppress_errors = (  # pylint: disable=protected-access
-                True
-            )
-            training_args_kwargs["torch_compile"] = self.cfg.torch_compile
-            if self.cfg.torch_compile_backend:
-                training_args_kwargs["torch_compile_backend"] = (
-                    self.cfg.torch_compile_backend
-                )
-            if self.cfg.torch_compile_mode:
-                training_args_kwargs["torch_compile_mode"] = self.cfg.torch_compile_mode
-
-        if self.cfg.compile_optimizer:
-            training_args_kwargs["compile_optimizer"] = True
-
-    def _configure_gradient_checkpointing(self, training_args_kwargs: dict):
-        if self.cfg.gradient_checkpointing:
-            training_args_kwargs["gradient_checkpointing"] = (
-                self.cfg.gradient_checkpointing
-            )
-            if self.cfg.gradient_checkpointing_kwargs is not None:
-                training_args_kwargs["gradient_checkpointing_kwargs"] = (
-                    self.cfg.gradient_checkpointing_kwargs
-                )
-            else:
-                training_args_kwargs["gradient_checkpointing_kwargs"] = {
-                    "use_reentrant": False
-                }
-
-    def _set_base_training_args(
-        self, total_num_steps
-    ) -> tuple[dict[str, Any], dict[str, Any]]:
-        training_args_kwargs: dict[str, Any] = {}
-        trainer_kwargs: dict[str, Any] = {}
-
-        self._configure_warmup_and_logging(total_num_steps, training_args_kwargs)
-        self._configure_precision_settings(training_args_kwargs)
-        self._configure_save_and_eval_strategy(training_args_kwargs)
-        self._configure_gradient_checkpointing(training_args_kwargs)
-
-        # set arg into trainer_args_kwargs with same name if value not None
-        for arg in [
-            # optim/scheduler
-            "adam_beta1",
-            "adam_beta2",
-            "adam_beta3",
-            "adam_epsilon",
-            "adam_epsilon2",
-            "cosine_min_lr_ratio",
-            "cosine_constant_lr_ratio",
-            "optim_target_modules",
-            # trainer
-            "max_grad_norm",
-            "dataloader_num_workers",
-            "dataloader_pin_memory",
-            "dataloader_prefetch_factor",
-            "gradient_accumulation_steps",
-            "learning_rate",
-            "embedding_lr",
-            "embedding_lr_scale",
-            "lr_groups",
-            "loraplus_lr_ratio",
-            "loraplus_lr_embedding",
-            "output_dir",
-            "save_safetensors",
-            "save_only_model",
-            "include_tokens_per_second",
-            "weight_decay",
-            "seed",
-        ]:
-            if hasattr(self.cfg, arg) and getattr(self.cfg, arg) is not None:
-                training_args_kwargs[arg] = getattr(self.cfg, arg)
-
-        training_args_kwargs["per_device_train_batch_size"] = self.cfg.micro_batch_size
-
-        if self.cfg.eval_batch_size:
-            training_args_kwargs["per_device_eval_batch_size"] = (
-                self.cfg.eval_batch_size
-            )
-
-        training_args_kwargs["max_steps"] = self.cfg.max_steps or total_num_steps or -1
-        training_args_kwargs["num_train_epochs"] = self.cfg.num_epochs
-
-        # max_length is not used in CausalTrainer
-        if self.cfg.reward_model or self.cfg.rl:
-            training_args_kwargs["max_length"] = self.cfg.sequence_len
-
-        self._configure_reporting(training_args_kwargs)
-        self._configure_hub_parameters(training_args_kwargs)
-        self._configure_scheduler(training_args_kwargs)
-        self._configure_optimizer(training_args_kwargs, trainer_kwargs)
-        self._configure_torch_compile(training_args_kwargs)
-
-        return training_args_kwargs, trainer_kwargs

src/axolotl/core/builders/causal.py

@@ -1,489 +0,0 @@
-"""Builder for causal trainers"""
-
-import inspect
-import math
-import os
-from pathlib import Path
-from typing import Type, Union
-
-import transformers
-from transformers import (
-    DataCollatorWithFlattening,
-    EarlyStoppingCallback,
-)
-from trl.trainer.utils import RewardDataCollatorWithPadding
-
-from axolotl.core.builders.base import TrainerBuilderBase
-from axolotl.core.trainers import (
-    AxolotlMambaTrainer,
-    AxolotlPRMTrainer,
-    AxolotlRewardTrainer,
-    AxolotlTrainer,
-    ReLoRATrainer,
-)
-from axolotl.core.training_args import (
-    AxolotlPRMConfig,
-    AxolotlRewardConfig,
-    AxolotlTrainingArguments,
-)
-from axolotl.integrations.base import PluginManager
-from axolotl.monkeypatch.multipack import SUPPORTED_MULTIPACK_MODEL_TYPES
-from axolotl.monkeypatch.relora import ReLoRACallback
-from axolotl.processing_strategies import get_processing_strategy
-from axolotl.utils import is_comet_available, is_mlflow_available
-from axolotl.utils.callbacks import (
-    EvalFirstStepCallback,
-    LossWatchDogCallback,
-    SaveBetterTransformerModelCallback,
-    bench_eval_callback_factory,
-    causal_lm_bench_eval_callback_factory,
-    colab_inference_post_train_callback,
-    log_prediction_callback_factory,
-)
-from axolotl.utils.callbacks.lisa import lisa_callback_factory
-from axolotl.utils.callbacks.qat import QATCallback
-from axolotl.utils.chat_templates import get_chat_template_from_config
-from axolotl.utils.collators import (
-    BatchSamplerDataCollatorForSeq2Seq,
-    DataCollatorForSeq2Seq,
-    MambaDataCollator,
-    V2BatchSamplerDataCollatorForSeq2Seq,
-)
-from axolotl.utils.collators.mm_chat import MultiModalChatDataCollator
-from axolotl.utils.logging import get_logger
-
-LOG = get_logger(__name__)
-
-
-class HFCausalTrainerBuilder(TrainerBuilderBase):
-    """
-    Build the HuggingFace training args/trainer for causal models and reward modeling
-    using TRL.
-    """
-
-    def get_callbacks(self):
-        callbacks = super().get_callbacks()
-        callbacks.append(EvalFirstStepCallback())
-
-        if self.cfg.relora_steps:
-            callbacks.append(ReLoRACallback(self.cfg))
-
-        if (
-            hasattr(self.model, "use_bettertransformer")
-            and self.model.use_bettertransformer is True
-        ):
-            callbacks.append(SaveBetterTransformerModelCallback())
-
-        # TODO: check if can move to base class
-        if self.cfg.loss_watchdog_threshold is not None:
-            callbacks.append(LossWatchDogCallback(self.cfg))
-
-        if self.cfg.qat:
-            callbacks.append(QATCallback(self.cfg.qat))
-
-        return callbacks
-
-    def get_post_trainer_create_callbacks(self, trainer):
-        callbacks = []
-        if self.cfg.use_wandb and self.cfg.eval_table_size > 0:
-            LogPredictionCallback = log_prediction_callback_factory(
-                trainer, self.tokenizer, "wandb"
-            )
-            callbacks.append(LogPredictionCallback(self.cfg))
-        if (
-            self.cfg.use_mlflow
-            and is_mlflow_available()
-            and self.cfg.eval_table_size > 0
-        ):
-            LogPredictionCallback = log_prediction_callback_factory(
-                trainer, self.tokenizer, "mlflow"
-            )
-            callbacks.append(LogPredictionCallback(self.cfg))
-        if self.cfg.use_comet and is_comet_available() and self.cfg.eval_table_size > 0:
-            LogPredictionCallback = log_prediction_callback_factory(
-                trainer, self.tokenizer, "comet_ml"
-            )
-            callbacks.append(LogPredictionCallback(self.cfg))
-
-        if self.cfg.do_bench_eval:
-            callbacks.append(bench_eval_callback_factory(trainer, self.tokenizer))
-        if self.cfg.do_causal_lm_eval:
-            CausalLMBenchEvalCallback = causal_lm_bench_eval_callback_factory(
-                trainer, self.tokenizer
-            )
-            callbacks.append(CausalLMBenchEvalCallback(self.cfg))
-
-        if self.cfg.early_stopping_patience:
-            early_stop_cb = EarlyStoppingCallback(
-                self.cfg.early_stopping_patience,
-            )
-            callbacks.append(early_stop_cb)
-
-        if self.cfg.lisa_step_interval and self.cfg.lisa_n_layers:
-            callbacks.append(lisa_callback_factory(trainer))
-
-        if any("COLAB_" in key for key in os.environ):
-            ColabCallback = colab_inference_post_train_callback(trainer)
-            callbacks.append(ColabCallback(self.cfg))
-
-        callbacks.extend(super().get_post_trainer_create_callbacks(trainer=trainer))
-        return callbacks
-
-    def _get_trainer_cls(self):
-        if self.cfg.plugins:
-            plugin_manager = PluginManager.get_instance()
-            trainer_cls = plugin_manager.get_trainer_cls(self.cfg)
-            if trainer_cls:
-                return trainer_cls
-        if self.cfg.relora_steps:
-            return ReLoRATrainer
-        if self.cfg.model_config_type == "mamba":
-            return AxolotlMambaTrainer
-        if self.cfg.reward_model:
-            return AxolotlRewardTrainer
-        if self.cfg.process_reward_model:
-            return AxolotlPRMTrainer
-        return AxolotlTrainer
-
-    def build(self, total_num_steps):
-        training_arguments_kwargs, trainer_kwargs = self._set_base_training_args(
-            total_num_steps
-        )
-
-        if self.cfg.fsdp:
-            training_arguments_kwargs["fsdp"] = self.cfg.fsdp
-            if self.cfg.fsdp_config:
-                training_arguments_kwargs["fsdp_config"] = {
-                    k.lstrip("fsdp_"): v for k, v in dict(self.cfg.fsdp_config).items()
-                }
-
-        if self.cfg.adapter == "qlora":
-            training_arguments_kwargs["qlora"] = True
-
-        # deepspeed
-        if self.cfg.deepspeed:
-            training_arguments_kwargs["deepspeed"] = self.cfg.deepspeed
-
-        if self.cfg.lr_quadratic_warmup is not None:
-            training_arguments_kwargs["lr_quadratic_warmup"] = (
-                self.cfg.lr_quadratic_warmup
-            )
-
-        if self.cfg.dataloader_drop_last is not None:
-            training_arguments_kwargs["dataloader_drop_last"] = (
-                self.cfg.dataloader_drop_last
-            )
-        elif self.cfg.sample_packing and self.cfg.eval_sample_packing is False:
-            training_arguments_kwargs["dataloader_drop_last"] = True
-
-        if self.cfg.remove_unused_columns is not None:
-            training_arguments_kwargs["remove_unused_columns"] = (
-                self.cfg.remove_unused_columns
-            )
-
-        if self.cfg.do_bench_eval:
-            training_arguments_kwargs["do_bench_eval"] = self.cfg.do_bench_eval
-            if self.cfg.bench_dataset:
-                training_arguments_kwargs["bench_dataset"] = self.cfg.bench_dataset
-        if self.cfg.do_causal_lm_eval:
-            training_arguments_kwargs["do_causal_lm_eval"] = self.cfg.do_causal_lm_eval
-        if self.cfg.metric_for_best_model:
-            training_arguments_kwargs["metric_for_best_model"] = (
-                self.cfg.metric_for_best_model
-            )
-        if self.cfg.greater_is_better:
-            training_arguments_kwargs["greater_is_better"] = self.cfg.greater_is_better
-
-        # DDP Config
-        if self.cfg.ddp_timeout:
-            training_arguments_kwargs["ddp_timeout"] = self.cfg.ddp_timeout
-        # see https://pytorch.org/docs/stable/generated/torch.nn.parallel.DistributedDataParallel.html
-        if self.cfg.ddp_bucket_cap_mb:
-            training_arguments_kwargs["ddp_bucket_cap_mb"] = self.cfg.ddp_bucket_cap_mb
-        if self.cfg.ddp_broadcast_buffers is not None:
-            training_arguments_kwargs["ddp_broadcast_buffers"] = (
-                self.cfg.ddp_broadcast_buffers
-            )
-
-        # these are all the "standard" kwargs that are def used
-        training_arguments_kwargs["max_seq_length"] = self.cfg.sequence_len
-
-        if self.cfg.auto_find_batch_size is not None:
-            training_arguments_kwargs["auto_find_batch_size"] = (
-                self.cfg.auto_find_batch_size
-            )
-
-        training_arguments_kwargs["eval_accumulation_steps"] = (
-            self.cfg.gradient_accumulation_steps
-        )
-
-        training_arguments_kwargs["load_best_model_at_end"] = (
-            (
-                self.cfg.load_best_model_at_end is not False
-                or self.cfg.early_stopping_patience
-            )
-            and (
-                (not self.cfg.test_datasets and self.cfg.val_set_size > 0)
-                or (self.cfg.test_datasets and self.cfg.val_set_size == 0)
-            )
-            and self.cfg.save_steps
-            and self.cfg.eval_steps
-            and self.cfg.save_steps % self.cfg.eval_steps == 0
-        ) or False
-
-        # handle ddp
-        ddp_find_unused_parameters = None
-        if self.cfg.ddp:
-            ddp_find_unused_parameters = bool(self.cfg.ddp_find_unused_parameters)
-        training_arguments_kwargs["ddp_find_unused_parameters"] = (
-            ddp_find_unused_parameters
-        )
-
-        training_arguments_kwargs["group_by_length"] = self.cfg.group_by_length
-        training_arguments_kwargs["curriculum_sampling"] = self.cfg.curriculum_sampling
-
-        training_arguments_kwargs["sample_packing"] = bool(self.cfg.sample_packing)
-        training_arguments_kwargs["multipack_real_batches"] = (
-            self.cfg.multipack_real_batches
-            if self.cfg.multipack_real_batches is not None
-            else not self.cfg.flash_attention
-        )
-        training_arguments_kwargs["eval_sample_packing"] = bool(
-            self.cfg.eval_sample_packing
-        )
-        if self.cfg.sample_packing_bin_size is not None:
-            training_arguments_kwargs["sample_packing_bin_size"] = (
-                self.cfg.sample_packing_bin_size
-            )
-        if self.cfg.sample_packing_group_size is not None:
-            training_arguments_kwargs["sample_packing_group_size"] = (
-                self.cfg.sample_packing_group_size
-            )
-        if self.cfg.sample_packing_eff_est:
-            training_arguments_kwargs["sample_packing_efficiency"] = (
-                self.cfg.sample_packing_eff_est
-            )
-
-        if self.cfg.relora_steps:
-            training_arguments_kwargs["relora_steps"] = self.cfg.relora_steps
-            training_arguments_kwargs["relora_warmup_steps"] = (
-                self.cfg.relora_warmup_steps
-            )
-            if self.cfg.relora_anneal_steps:
-                training_arguments_kwargs["relora_anneal_steps"] = (
-                    self.cfg.relora_anneal_steps
-                )
-            if self.cfg.relora_prune_ratio:
-                training_arguments_kwargs["relora_prune_ratio"] = (
-                    self.cfg.relora_prune_ratio
-                )
-
-        if self.cfg.lisa_step_interval and self.cfg.lisa_n_layers:
-            training_arguments_kwargs["lisa_n_layers"] = self.cfg.lisa_n_layers
-            training_arguments_kwargs["lisa_step_interval"] = (
-                self.cfg.lisa_step_interval
-            )
-            training_arguments_kwargs["lisa_layers_attribute"] = (
-                self.cfg.lisa_layers_attribute
-            )
-
-        training_arguments_kwargs = self.hook_pre_create_training_args(
-            training_arguments_kwargs
-        )
-        training_arguments_kwargs["model_type"] = self.cfg.model_config_type
-        training_arguments_kwargs["pretraining"] = bool(self.cfg.pretraining_dataset)
-        if self.cfg.chat_template:
-            training_arguments_kwargs["chat_template"] = get_chat_template_from_config(
-                cfg=self.cfg,
-                tokenizer=self.tokenizer,
-            )
-
-        if self.cfg.neftune_noise_alpha is not None:
-            training_arguments_kwargs["neftune_noise_alpha"] = (
-                self.cfg.neftune_noise_alpha
-            )
-
-        if self.cfg.accelerator_config:
-            training_arguments_kwargs["accelerator_config"] = (
-                self.cfg.accelerator_config
-            )
-
-        if self.cfg.image_size:
-            training_arguments_kwargs["image_size"] = self.cfg.image_size
-        if self.cfg.image_resize_algorithm:
-            training_arguments_kwargs["image_resize_algorithm"] = (
-                self.cfg.image_resize_algorithm
-            )
-        if self.cfg.kd_ce_alpha is not None:
-            training_arguments_kwargs["kd_ce_alpha"] = self.cfg.kd_ce_alpha
-        if self.cfg.kd_alpha is not None:
-            training_arguments_kwargs["kd_alpha"] = self.cfg.kd_alpha
-        if self.cfg.kd_temperature is not None:
-            training_arguments_kwargs["kd_temperature"] = self.cfg.kd_temperature
-        if self.cfg.kd_zscore_base_temp is not None:
-            training_arguments_kwargs["kd_zscore_base_temp"] = (
-                self.cfg.kd_zscore_base_temp
-            )
-        if self.cfg.kd_top_k_before_softmax is not None:
-            training_arguments_kwargs["kd_top_k_before_softmax"] = (
-                self.cfg.kd_top_k_before_softmax
-            )
-
-        if self.cfg.reward_model:
-            training_args_cls = AxolotlRewardConfig
-        elif self.cfg.process_reward_model:
-            training_args_cls = AxolotlPRMConfig
-        else:
-            training_args_cls = AxolotlTrainingArguments
-        training_args = training_args_cls(  # pylint: disable=unexpected-keyword-arg
-            **training_arguments_kwargs,
-        )
-        training_args = self.hook_post_create_training_args(training_args)
-
-        # unset run_name so wandb sets up experiment names
-        if self.cfg.use_wandb and training_args.run_name == training_args.output_dir:
-            training_args.run_name = (  # pylint: disable=attribute-defined-outside-init
-                None
-            )
-
-        data_collator_kwargs = {
-            "padding": True,  # True/"longest" is the default
-        }
-        multiple = 64
-        if self.cfg.pad_to_sequence_len:
-            data_collator_kwargs["pad_to_multiple_of"] = multiple * math.ceil(
-                self.cfg.sequence_len / multiple
-            )
-        else:
-            # A100 is best at 64, while others at 8. Let's use the larger so we don't have to check
-            # https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html
-            data_collator_kwargs["pad_to_multiple_of"] = multiple
-
-        trainer_cls = self._get_trainer_cls()
-
-        trainer_kwargs, trainer_cls = self.hook_pre_create_trainer(
-            trainer_kwargs, trainer_cls
-        )
-        if eval_data_collator := self.build_collator(
-            training_args, is_eval=True, **data_collator_kwargs
-        ):
-            if not (self.cfg.reward_model or self.cfg.process_reward_model):
-                trainer_kwargs["eval_data_collator"] = eval_data_collator
-        if not (self.cfg.reward_model or self.cfg.process_reward_model):
-            trainer_kwargs["bench_data_collator"] = transformers.DataCollatorForSeq2Seq(
-                self.tokenizer,
-                return_tensors="pt",
-                **data_collator_kwargs,
-            )
-        sig = inspect.signature(trainer_cls)
-        if "processing_class" in sig.parameters:
-            trainer_kwargs["processing_class"] = self.tokenizer
-        elif "tokenizer" in sig.parameters:
-            trainer_kwargs["tokenizer"] = self.tokenizer
-        if (
-            not (trainer_cls in [AxolotlRewardTrainer, AxolotlPRMTrainer])
-            and self.cfg.datasets is not None
-        ):
-            trainer_kwargs["dataset_tags"] = [
-                d["path"] for d in self.cfg.datasets if not Path(d["path"]).is_dir()
-            ]
-        trainer = trainer_cls(
-            model=self.model,
-            train_dataset=self.train_dataset,
-            eval_dataset=self.eval_dataset,
-            args=training_args,
-            data_collator=self.build_collator(training_args, **data_collator_kwargs),
-            callbacks=self.get_callbacks(),
-            **trainer_kwargs,
-        )
-        trainer = self.hook_post_create_trainer(trainer)
-        for callback in self.get_post_trainer_create_callbacks(trainer):
-            trainer.add_callback(callback)
-
-        if self.cfg.deepspeed and self.cfg.sample_packing:
-            trainer.accelerator.state.deepspeed_plugin.deepspeed_config[
-                "train_micro_batch_size_per_gpu"
-            ] = self.cfg.micro_batch_size
-
-        return trainer
-
-    def build_collator(
-        self, training_args: AxolotlTrainingArguments, is_eval=False, **kwargs
-    ):
-        if training_args.pretraining:
-            if (
-                self.cfg.pretraining_sample_concatenation is False
-                or self.cfg.micro_batch_size > 1
-            ):
-                return DataCollatorForSeq2Seq(self.tokenizer, **kwargs)
-            return None
-
-        if self.cfg.model_config_type == "mamba":
-            return MambaDataCollator(tokenizer=self.tokenizer)
-
-        use_batch_sampler_collator = False
-        if is_eval is False and training_args.sample_packing:
-            use_batch_sampler_collator = True
-        if is_eval and training_args.eval_sample_packing:
-            use_batch_sampler_collator = True
-
-        collator: Type[
-            Union[
-                V2BatchSamplerDataCollatorForSeq2Seq,
-                BatchSamplerDataCollatorForSeq2Seq,
-                DataCollatorForSeq2Seq,
-                DataCollatorWithFlattening,
-                RewardDataCollatorWithPadding,
-            ]
-        ]
-        collator_args = [self.tokenizer]
-        if self.cfg.reward_model:
-            collator = RewardDataCollatorWithPadding
-        elif use_batch_sampler_collator:
-            # Use V2BatchSamplerDataCollatorForSeq2Seq for flex attention,
-            # supported multipack models, or non-flash-attention llama
-            if (
-                self.cfg.flex_attention
-                or self.cfg.model_config_type in SUPPORTED_MULTIPACK_MODEL_TYPES
-                or (
-                    self.cfg.model_config_type in ["llama"]
-                    and self.cfg.flash_attention is not True
-                )
-            ):
-                collator = V2BatchSamplerDataCollatorForSeq2Seq
-            else:
-                collator = BatchSamplerDataCollatorForSeq2Seq
-        else:
-            if self.cfg.processor_type and self.processor:
-                collator = MultiModalChatDataCollator
-                kwargs["processing_strategy"] = get_processing_strategy(
-                    self.processor,
-                    training_args.chat_template,
-                    self.cfg.chat_template,
-                    image_size=training_args.image_size,
-                    image_resize_algorithm=training_args.image_resize_algorithm,
-                )
-            elif self.cfg.batch_flattening:
-                collator = DataCollatorWithFlattening
-                collator_args.pop(0)
-                kwargs.pop("pad_to_multiple_of", None)
-                kwargs.pop("padding", None)
-            elif self.cfg.kd_trainer:
-                from axolotl.integrations.kd.collator import (
-                    DataCollatorForKD,
-                    KDBatchSamplerDataCollatorForSeq2Seq,
-                )
-
-                if self.cfg.sample_packing:
-                    collator = KDBatchSamplerDataCollatorForSeq2Seq
-                else:
-                    collator = DataCollatorForKD
-            else:
-                collator = DataCollatorForSeq2Seq
-
-        kwargs["return_tensors"] = "pt"
-
-        return collator(
-            *collator_args,
-            **kwargs,
-        )

src/axolotl/core/builders/rl.py

@@ -1,246 +0,0 @@
-"""Builder for RLHF trainers"""
-
-import inspect
-from pathlib import Path
-
-from axolotl.core.builders.base import TrainerBuilderBase
-from axolotl.core.trainers import (
-    AxolotlCPOTrainer,
-    AxolotlKTOTrainer,
-    AxolotlORPOTrainer,
-)
-from axolotl.core.trainers.dpo import DPOStrategy
-from axolotl.core.trainers.dpo.args import AxolotlDPOConfig
-from axolotl.core.trainers.grpo import GRPOStrategy
-from axolotl.core.training_args import (
-    AxolotlCPOConfig,
-    AxolotlKTOConfig,
-    AxolotlORPOConfig,
-)
-from axolotl.integrations.base import PluginManager
-from axolotl.loaders.utils import ensure_dtype
-from axolotl.utils.logging import get_logger
-from axolotl.utils.schemas.enums import RLType
-
-LOG = get_logger(__name__)
-
-
-class HFRLTrainerBuilder(TrainerBuilderBase):
-    """Trainer factory class for TRL-based RLHF trainers (e.g. DPO)"""
-
-    def get_callbacks(self):
-        callbacks = super().get_callbacks()
-
-        return callbacks
-
-    def get_post_trainer_create_callbacks(self, trainer):
-        callbacks = super().get_post_trainer_create_callbacks(trainer=trainer)
-        return callbacks
-
-    def _get_trainer_cls(self, trainer_kwargs: dict):
-        """
-        Returns trainer_cls and trainer_cls_args
-        """
-        if self.cfg.plugins:
-            plugin_manager = PluginManager.get_instance()
-            trainer_cls = plugin_manager.get_trainer_cls(self.cfg)
-            trainer_cls_args = []  # type: ignore
-
-            if trainer_cls is not None:
-                return trainer_cls, trainer_cls_args
-
-        trainer_cls = None
-        trainer_cls_args = [self.model]
-
-        if self.cfg.rl is RLType.GRPO:
-            trainer_cls = GRPOStrategy.get_trainer_class(
-                sequence_parallel=self.cfg.sequence_parallel_degree > 1
-            )
-            trainer_cls_args.extend(GRPOStrategy.set_trainer_args(self.cfg))
-
-            trainer_kwargs.update(GRPOStrategy.set_trainer_kwargs(self.cfg))
-
-        elif self.cfg.rl in [RLType.DPO, RLType.IPO]:
-            trainer_cls = DPOStrategy.get_trainer_class()
-            trainer_cls_args.append(self.model_ref)
-
-        elif self.cfg.rl is RLType.ORPO:
-            trainer_cls = AxolotlORPOTrainer
-        elif self.cfg.rl is RLType.KTO:
-            trainer_cls = AxolotlKTOTrainer
-        elif self.cfg.rl is RLType.SIMPO:
-            trainer_cls = AxolotlCPOTrainer
-        else:
-            raise ValueError(f"Unsupported RL: {self.cfg.rl}")
-
-        return trainer_cls, trainer_cls_args
-
-    def _build_training_arguments(self, total_num_steps):
-        """
-        Returns training_args and trainer_kwargs
-        """
-        training_args_kwargs, trainer_kwargs = self._set_base_training_args(
-            total_num_steps=total_num_steps
-        )
-
-        if self.cfg.remove_unused_columns is not None:
-            training_args_kwargs["remove_unused_columns"] = (
-                self.cfg.remove_unused_columns
-            )
-        else:
-            training_args_kwargs["remove_unused_columns"] = False
-
-        # only rlhf
-        if self.cfg.dataset_processes:
-            training_args_kwargs["dataset_num_proc"] = self.cfg.dataset_processes
-
-        if self.cfg.trl and self.cfg.trl.beta is not None:
-            training_args_kwargs["beta"] = self.cfg.trl.beta
-        elif self.cfg.rl_beta is not None:
-            training_args_kwargs["beta"] = self.cfg.rl_beta
-        elif self.cfg.orpo_alpha is not None:
-            # trl does some odd mapping of alpha to beta to reuse the beta parameter ???
-            training_args_kwargs["beta"] = self.cfg.orpo_alpha
-
-        if self.cfg.rpo_alpha is not None:
-            training_args_kwargs["rpo_alpha"] = self.cfg.rpo_alpha
-
-        if self.cfg.use_wandb:
-            training_args_kwargs["run_name"] = self.cfg.wandb_name
-
-        training_args_cls = None
-        blocklist_args_kwargs = []
-        if self.cfg.rl is RLType.SIMPO:
-            training_args_cls = AxolotlCPOConfig
-            training_args_kwargs["loss_type"] = "simpo"
-            training_args_kwargs["simpo_gamma"] = self.cfg.simpo_gamma
-            if self.cfg.cpo_alpha is not None:
-                training_args_kwargs["cpo_alpha"] = self.cfg.cpo_alpha
-
-        elif self.cfg.rl is RLType.ORPO:
-            training_args_cls = AxolotlORPOConfig
-            if self.cfg.max_prompt_len:
-                training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
-
-        elif self.cfg.rl is RLType.KTO:
-            training_args_cls = AxolotlKTOConfig
-
-            training_args_kwargs["desirable_weight"] = (
-                self.cfg.kto_desirable_weight or 1.0
-            )
-            training_args_kwargs["undesirable_weight"] = (
-                self.cfg.kto_undesirable_weight or 1.0
-            )
-
-            if self.cfg.max_prompt_len:
-                training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
-
-        elif self.cfg.rl is RLType.GRPO:
-            training_args_cls = GRPOStrategy.get_training_args_class()
-            training_args_kwargs.update(GRPOStrategy.set_training_args_kwargs(self.cfg))
-            blocklist_args_kwargs = GRPOStrategy.get_blocklist_args_kwargs()
-
-        elif self.cfg.rl in [RLType.DPO, RLType.IPO]:
-            training_args_cls = AxolotlDPOConfig
-            if self.cfg.rl is RLType.IPO:
-                training_args_kwargs["loss_type"] = "ipo"
-
-            # Not compatible with IPO
-            if self.cfg.rl is RLType.DPO and self.cfg.dpo_label_smoothing:
-                training_args_kwargs["label_smoothing"] = self.cfg.dpo_label_smoothing
-
-            training_args_kwargs["max_completion_length"] = None
-            training_args_kwargs["max_prompt_length"] = self.cfg.sequence_len
-            training_args_kwargs["generate_during_eval"] = self.cfg.use_wandb
-            if self.cfg.dpo_use_weighting is not None:
-                training_args_kwargs["use_weighting"] = self.cfg.dpo_use_weighting
-            if self.cfg.dpo_use_logits_to_keep is not None:
-                training_args_kwargs["use_logits_to_keep"] = (
-                    self.cfg.dpo_use_logits_to_keep
-                )
-        else:
-            raise ValueError(f"Unsupported RL: {self.cfg.rl}")
-
-        for blocklist_key in blocklist_args_kwargs:
-            if blocklist_key in training_args_kwargs:
-                del training_args_kwargs[blocklist_key]
-
-        training_args = training_args_cls(  # pylint: disable=unexpected-keyword-arg
-            logging_first_step=True,
-            **training_args_kwargs,
-        )
-
-        # unset run_name so wandb sets up experiment names
-        if self.cfg.use_wandb and training_args.run_name == training_args.output_dir:
-            training_args.run_name = (  # pylint: disable=attribute-defined-outside-init
-                None
-            )
-
-        return training_args, trainer_kwargs
-
-    def build(self, total_num_steps):
-        training_args, trainer_kwargs = self._build_training_arguments(total_num_steps)
-
-        if self.eval_dataset:
-            trainer_kwargs["eval_dataset"] = self.eval_dataset
-        if self.cfg.adapter and self.peft_config and self.cfg.rl is not RLType.GRPO:
-            trainer_kwargs["peft_config"] = self.peft_config
-        if self.cfg.precompute_ref_log_probs is not None:
-            trainer_kwargs["precompute_ref_log_probs"] = (
-                self.cfg.precompute_ref_log_probs
-            )
-
-        trainer_cls, trainer_cls_args = self._get_trainer_cls(trainer_kwargs)
-
-        sig = inspect.signature(trainer_cls)
-        if "tokenizer" in sig.parameters:
-            trainer_kwargs["tokenizer"] = self.tokenizer
-        else:
-            trainer_kwargs["processing_class"] = self.tokenizer
-
-        if self.cfg.datasets is not None and (
-            trainer_cls is DPOStrategy.get_trainer_class()
-        ):
-            trainer_kwargs["dataset_tags"] = [
-                d["path"] for d in self.cfg.datasets if not Path(d["path"]).is_dir()
-            ]
-
-        trainer_kwargs, trainer_cls = self.hook_pre_create_trainer(
-            trainer_kwargs, trainer_cls
-        )
-
-        trainer = trainer_cls(
-            *trainer_cls_args,
-            args=training_args,
-            train_dataset=self.train_dataset,
-            callbacks=self.get_callbacks(),
-            **trainer_kwargs,
-        )
-        if self.cfg.fsdp:
-            ensure_dtype(trainer.model, dtype=self.cfg.torch_dtype)
-            if self.cfg.rl in [RLType.DPO, RLType.IPO] and trainer.ref_model:
-                ensure_dtype(trainer.ref_model, dtype=self.cfg.torch_dtype)
-
-        trainer = self.hook_post_create_trainer(trainer)
-        for callback in self.get_post_trainer_create_callbacks(trainer):
-            trainer.add_callback(callback)
-
-        return trainer
-
-
-class HFPPOTrainerBuilder(TrainerBuilderBase):
-    """
-    HF Factory class for PPO Trainer
-    """
-
-    def get_callbacks(self):
-        callbacks = super().get_callbacks()
-        return callbacks
-
-    def get_post_trainer_create_callbacks(self, trainer):
-        callbacks = super().get_post_trainer_create_callbacks(trainer=trainer)
-        return callbacks
-
-    def build(self, total_num_steps):
-        # TODO: build PPOConfig
-        raise NotImplementedError("PPO trainer builder is not implemented yet.")

src/axolotl/core/chat/messages.py

@@ -156,6 +156,7 @@ class Messages(BaseModel):
                         len(input_ids) : len(input_ids) + len(pending_input_ids)
                     ]
                     if new_pending_inputs != pending_input_ids:
+                        # logging.warning("tokenization mismatch from concatenation.")
                         pending_input_ids = new_pending_inputs
                     input_ids.extend(pending_input_ids)
                     if pending_weight:

src/axolotl/core/trainers/base.py

@@ -4,10 +4,11 @@
 
 from __future__ import annotations
 
+import logging
 import os
 from collections import defaultdict
-from functools import partial, wraps
+from functools import wraps
-from typing import Callable, Literal, Optional
+from typing import Literal
 
 import datasets
 import torch
@@ -33,10 +34,9 @@ from axolotl.core.trainers.utils import (
     sanitize_kwargs_for_ds_tagging,
     sanitize_kwargs_for_tagging,
 )
-from axolotl.utils.logging import get_logger
 from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger(__name__)
 
 
 class AxolotlTrainer(SchedulerMixin, OptimizerMixin, RngLoaderMixin, Trainer):
@@ -113,9 +113,7 @@ class AxolotlTrainer(SchedulerMixin, OptimizerMixin, RngLoaderMixin, Trainer):
             drop_last=True,
         )
 
-    def _get_train_sampler(
+    def _get_train_sampler(self) -> Sampler | None:
-        self, train_dataset: Optional[Dataset] = None
-    ) -> Optional[Sampler]:
         """
         Helper method to get the sampler for training. Handles cases for sample packing
         and curriculum sampling (sequential).
@@ -139,7 +137,7 @@ class AxolotlTrainer(SchedulerMixin, OptimizerMixin, RngLoaderMixin, Trainer):
         if use_sample_packing:
             return self._create_multipack_sampler(
                 base_sampler=base_sampler,
-                dataset=train_dataset,
+                dataset=self.train_dataset,
             )
 
         return base_sampler
@@ -152,6 +150,8 @@ class AxolotlTrainer(SchedulerMixin, OptimizerMixin, RngLoaderMixin, Trainer):
             If the dataset is non-empty, a sampler is returned, the type of which
                 depends on the passed training args.
         """
+        eval_dataset = eval_dataset if eval_dataset is not None else self.eval_dataset
+
         # Multipacking enabled if training is enabled and eval is not explicitly disabled
         use_multipack = (
             self.args.sample_packing and self.args.eval_sample_packing is not False
@@ -172,91 +172,125 @@ class AxolotlTrainer(SchedulerMixin, OptimizerMixin, RngLoaderMixin, Trainer):
 
         return base_sampler
 
-    def _get_dataloader(
+    def _create_dataloader_params(self, is_eval=False, custom_batch_size=None):
-        self,
+        """Create common dataloader parameters for train or eval."""
-        dataset: Dataset,
+        batch_size = custom_batch_size or (
-        description: str,
+            self.args.eval_batch_size if is_eval else self._train_batch_size
-        batch_size: int,
+        )
-        sampler_fn: Optional[Callable[[Dataset], torch.utils.data.Sampler]] = None,
-        is_training: bool = False,
-        dataloader_key: Optional[str] = None,
-    ) -> DataLoader:
-        """Create a [`~torch.utils.data.DataLoader`] from the given dataset."""
 
-        data_collator = self.data_collator if is_training else self.eval_data_collator
+        params = {
-
-        if dataset.column_names and "length" in dataset.column_names:
-            dataset = dataset.remove_columns(["length"])
-
-        if isinstance(dataset, datasets.Dataset):
-            if is_training:
-                if not self.args.sample_packing or self.args.pretraining:
-                    dataset = self._remove_unused_columns(
-                        dataset, description="training"
-                    )
-            elif (
-                not is_training
-                and self.args.sample_packing
-                and self.args.eval_sample_packing is not False
-            ):
-                batch_size = (
-                    batch_size
-                    if self.args.sample_packing
-                    else self.args.per_device_eval_batch_size
-                )
-            else:
-                dataset = self._remove_unused_columns(dataset, description=description)
-        else:
-            data_collator = self._get_collator_with_removed_columns(
-                self.data_collator, description=description
-            )
-
-        dataloader_params = {
             "batch_size": batch_size,
-            "collate_fn": data_collator,
+            "collate_fn": self.data_collator,
             "num_workers": self.args.dataloader_num_workers,
             "pin_memory": self.args.dataloader_pin_memory,
-            "persistent_workers": self.args.dataloader_persistent_workers,
         }
 
-        if not isinstance(dataset, torch.utils.data.IterableDataset):
+        # Add persistent workers only for training
-            dataloader_params["drop_last"] = self.args.dataloader_drop_last
+        if not is_eval and hasattr(self.args, "dataloader_persistent_workers"):
-            if sampler_fn is not None:
+            params["persistent_workers"] = self.args.dataloader_persistent_workers
-                sampler = sampler_fn(dataset)
+
-                if isinstance(sampler, BatchSampler):
+        # Add prefetch factor if specified
-                    # batch_size and batch_sampler are mutually exclusive
+        if self.args.dataloader_prefetch_factor:
-                    dataloader_params["batch_sampler"] = sampler
+            params["prefetch_factor"] = self.args.dataloader_prefetch_factor
-                    del dataloader_params["batch_size"]
+
-                    del dataloader_params["drop_last"]
+        return params
-                else:
+
-                    dataloader_params["sampler"] = sampler
+    def _prepare_dataloader(
+        self, dataset, sampler, is_eval=False, custom_batch_size=None
+    ):
+        """Prepare a dataloader with the given dataset and sampler."""
+        # Get base parameters
+        dataloader_params = self._create_dataloader_params(is_eval, custom_batch_size)
+
+        # Add sampler configuration
+        if not isinstance(dataset, torch.utils.data.IterableDataset):
+            if isinstance(sampler, BatchSampler):
+                # batch_size and batch_sampler are mutually exclusive
+                dataloader_params["batch_sampler"] = sampler
+                del dataloader_params["batch_size"]
+            else:
+                dataloader_params["sampler"] = sampler
+                dataloader_params["drop_last"] = self.args.dataloader_drop_last
+
+            if not is_eval:
+                dataloader_params["worker_init_fn"] = seed_worker
+
+        # Create the dataloader
+        dataloader = DataLoader(dataset, **dataloader_params)
 
-            dataloader_params["prefetch_factor"] = self.args.dataloader_prefetch_factor
-            if is_training:
-                dataloader_params["worker_init_fn"] = partial(
-                    seed_worker,
-                    num_workers=self.args.dataloader_num_workers,
-                    rank=self.args.process_index,
-                )
         if self.args.sample_packing and (
-            (is_training and not self.args.pretraining)
+            (not is_eval and not self.args.pretraining)
-            or (not is_training and self.args.eval_sample_packing is not False)
+            or (is_eval and self.args.eval_sample_packing is not False)
         ):
             self.accelerator.even_batches = False
 
-        dataloader = DataLoader(dataset, **dataloader_params)
+        return self.accelerator.prepare_data_loader(dataloader)
 
-        # Accelerator.free_memory() will destroy the references, so
+    def get_train_dataloader(self) -> DataLoader:
-        # we need to store the non-prepared version for eval dataloaders.
+        """Get dataloader for training"""
-        # fmt: off
+        train_dataset = self.train_dataset
-        if dataloader_key is not None and self.args.dataloader_persistent_workers:
+        data_collator = self.data_collator  # type: ignore
-            if hasattr(self, "_eval_dataloaders"):
-                self._eval_dataloaders[dataloader_key] = dataloader  # type: ignore  # pylint: disable=access-member-before-definition
-            else:
-                self._eval_dataloaders = {dataloader_key: dataloader}  # pylint: disable=attribute-defined-outside-init
-        # fmt: on
 
-        return self.accelerator.prepare(dataloader)
+        # Handle dataset preprocessing
+        if isinstance(train_dataset, datasets.Dataset):
+            if self.args.sample_packing and not self.args.pretraining:
+                train_dataset = train_dataset.remove_columns(["length"])
+            if not self.args.sample_packing or self.args.pretraining:
+                train_dataset = self._remove_unused_columns(
+                    train_dataset, description="training"
+                )
+        else:
+            self.data_collator = self._get_collator_with_removed_columns(  # pylint: disable=attribute-defined-outside-init
+                data_collator,
+                description="training",
+            )
+
+        # Get sampler and create dataloader
+        sampler = self._get_train_sampler()
+        return self._prepare_dataloader(train_dataset, sampler, is_eval=False)
+
+    def get_eval_dataloader(self, eval_dataset: Dataset | None = None) -> DataLoader:
+        """Get dataloader for evaluation"""
+        eval_dataset = eval_dataset if eval_dataset is not None else self.eval_dataset
+
+        # Handle special case: sample packing is enabled but eval_sample_packing is False
+        if self.args.sample_packing and self.args.eval_sample_packing is False:
+            self.data_collator = (  # pylint: disable=attribute-defined-outside-init
+                self.eval_data_collator
+            )
+            if "length" in eval_dataset.column_names:
+                eval_dataset = eval_dataset.remove_columns(["length"])
+            dataloader = super().get_eval_dataloader(eval_dataset)
+            self.data_collator = (  # pylint: disable=attribute-defined-outside-init
+                self.train_data_collator
+            )
+
+            return dataloader
+
+        if self.args.sample_packing and self.args.eval_sample_packing is not False:
+            # Get appropriate data collator
+            self.data_collator = (  # pylint: disable=attribute-defined-outside-init
+                self.eval_data_collator
+                if hasattr(self, "eval_data_collator") and self.eval_data_collator
+                else self.data_collator
+            )
+            if "length" in eval_dataset.column_names:
+                eval_dataset = eval_dataset.remove_columns(["length"])
+
+            # Use eval_batch_size for sample packing, per_device_eval_batch_size otherwise
+            batch_size = (
+                self.args.eval_batch_size
+                if self.args.sample_packing
+                else self.args.per_device_eval_batch_size
+            )
+            sampler = self._get_eval_sampler(eval_dataset)
+            dataloader = self._prepare_dataloader(
+                eval_dataset, sampler, is_eval=True, custom_batch_size=batch_size
+            )
+
+            return dataloader
+
+        return super().get_eval_dataloader(eval_dataset)
 
     def _get_bench_sampler(
         self, bench_dataset: Dataset

src/axolotl/core/trainers/dpo/trainer.py

@@ -5,31 +5,65 @@ from functools import wraps
 from typing import Any, Dict, Union
 
 import torch
+from peft.optimizers import create_loraplus_optimizer
 from torch import nn
+from transformers import Trainer
+from transformers.utils import is_sagemaker_mp_enabled
 from trl import DPOTrainer
 
 from axolotl.core.trainers.mixins import RngLoaderMixin, SchedulerMixin
-from axolotl.core.trainers.mixins.optimizer import OptimizerInitMixin, OptimizerMixin
 from axolotl.core.trainers.utils import (
     sanitize_kwargs_for_ds_tagging,
     sanitize_kwargs_for_tagging,
 )
 
+if is_sagemaker_mp_enabled():
+    import smdistributed.modelparallel.torch as smp
 
-class AxolotlDPOTrainer(
+
-    RngLoaderMixin, SchedulerMixin, OptimizerMixin, OptimizerInitMixin, DPOTrainer
+class AxolotlDPOTrainer(RngLoaderMixin, SchedulerMixin, DPOTrainer):
-):
     """Extend the base DPOTrainer for axolotl helpers."""
 
     tag_names = ["axolotl", "dpo"]
 
     def __init__(self, *args, dataset_tags=None, **kwargs):
         super().__init__(*args, **kwargs)
-
         self.dataset_tags = dataset_tags
         self.optimizer = None
         self.model_accepts_loss_kwargs = False
 
+    def create_optimizer(self):
+        # pylint: disable=duplicate-code
+        if self.args.loraplus_lr_ratio is None:
+            return super().create_optimizer()
+
+        opt_model = self.model_wrapped if is_sagemaker_mp_enabled() else self.model
+        if self.optimizer is None:  # pylint: disable=access-member-before-definition
+            optimizer_cls, optimizer_kwargs = Trainer.get_optimizer_cls_and_kwargs(
+                self.args,
+                opt_model,
+            )
+
+            loraplus_lr_ratio = getattr(self.args, "loraplus_lr_ratio", None)
+            if loraplus_lr_ratio:
+                print("Using lora+")
+            loraplus_lr_embedding = getattr(self.args, "loraplus_lr_embedding", None)
+            # pylint: disable=duplicate-code
+            self.optimizer = create_loraplus_optimizer(  # pylint: disable=attribute-defined-outside-init
+                opt_model,
+                optimizer_cls,
+                loraplus_lr_ratio=loraplus_lr_ratio,
+                loraplus_lr_embedding=loraplus_lr_embedding,
+                **optimizer_kwargs,
+            )
+
+        if is_sagemaker_mp_enabled():
+            self.optimizer = smp.DistributedOptimizer(  # pylint: disable=attribute-defined-outside-init
+                self.optimizer
+            )
+
+        return self.optimizer
+
     @wraps(DPOTrainer.push_to_hub)
     def push_to_hub(self, *args, **kwargs) -> str:
         """

src/axolotl/core/trainers/grpo/__init__.py

@@ -2,6 +2,7 @@
 
 import importlib
 import inspect
+import logging
 from typing import Any
 
 from trl.trainer.grpo_trainer import RewardFunc
@@ -12,10 +13,9 @@ from axolotl.core.trainers.grpo.trainer import (
     AxolotlGRPOTrainer,
 )
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.logging import get_logger
 from axolotl.utils.schemas.trl import TRLConfig
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger(__name__)
 
 
 class GRPOStrategy:
@@ -69,9 +69,6 @@ class GRPOStrategy:
         grpo_args_kwargs["log_completions"] = trl.log_completions
         grpo_args_kwargs["num_completions_to_print"] = trl.num_completions_to_print
 
-        if cfg.sequence_parallel_degree > 1:
-            grpo_args_kwargs["sequence_parallel_degree"] = cfg.sequence_parallel_degree
-
         if trl.reward_weights:
             grpo_args_kwargs["reward_weights"] = trl.reward_weights
 
@@ -109,9 +106,7 @@ class GRPOStrategy:
         return grpo_args_kwargs
 
     @classmethod
-    def set_trainer_args(
+    def set_trainer_args(cls, cfg: DictDefault) -> list[Any]:
-        cls, cfg: DictDefault
-    ) -> list[Any]:  # pylint: disable=unused-argument
         trainer_args = []
         if cfg.trl and cfg.trl.reward_funcs:
             reward_funcs = []
@@ -128,7 +123,6 @@ class GRPOStrategy:
             trainer_kwargs["reward_processing_classes"] = (
                 cfg.trl.reward_processing_classes
             )
-
         return trainer_kwargs
 
     @classmethod
@@ -138,7 +132,7 @@ class GRPOStrategy:
 
     @classmethod
     def get_blocklist_args_kwargs(cls) -> list[str]:
-        return ["dataset_num_proc", "max_length"]
+        return ["dataset_num_proc"]
 
     @classmethod
     def get_reward_func(cls, reward_func_fqn: str) -> RewardFunc:
@@ -173,4 +167,4 @@ class GRPOStrategy:
             LOG.info(
                 f"Reward function {reward_func_fqn} is a pre-trained model path - if this is unexpected, please check the reward function path."
             )
-            return reward_func_fqn
+            return reward_func

src/axolotl/core/trainers/grpo/args.py

@@ -12,5 +12,3 @@ from axolotl.core.training_args import AxolotlTrainingMixins
 @dataclass
 class AxolotlGRPOConfig(AxolotlTrainingMixins, GRPOConfig):
     """Axolotl GRPO Config for GRPO training"""
-
-    sequence_parallel_degree: int | None = None

src/axolotl/core/trainers/grpo/trainer.py

@@ -3,7 +3,6 @@
 # pylint: disable=too-many-lines,duplicate-code,protected-access,no-member
 
 import warnings
-from functools import partial
 from typing import Any
 
 import datasets
@@ -44,7 +43,6 @@ from trl.trainer.utils import pad
 
 from axolotl.core.trainers.grpo.sampler import SequenceParallelRepeatRandomSampler
 from axolotl.core.trainers.mixins import RngLoaderMixin, SchedulerMixin
-from axolotl.core.trainers.mixins.optimizer import OptimizerInitMixin, OptimizerMixin
 from axolotl.monkeypatch.ring_attn import get_ring_attn_group
 
 if is_peft_available():
@@ -52,49 +50,11 @@ if is_peft_available():
     from peft import PeftConfig
 
 
-class AxolotlGRPOTrainer(
+class AxolotlGRPOTrainer(RngLoaderMixin, SchedulerMixin, GRPOTrainer):
-    RngLoaderMixin, SchedulerMixin, OptimizerMixin, OptimizerInitMixin, GRPOTrainer
-):
     """Extend the base GRPOTrainer for axolotl helpers"""
 
     _tag_names = ["trl", "grpo", "axolotl"]
 
-    def get_train_dataloader(self):
-        if self.train_dataset is None:
-            raise ValueError("Trainer: training requires a train_dataset.")
-
-        train_dataset = self.train_dataset
-        data_collator = self.data_collator
-        if isinstance(train_dataset, datasets.Dataset):
-            train_dataset = self._remove_unused_columns(
-                train_dataset, description="training"
-            )
-        else:
-            data_collator = self._get_collator_with_removed_columns(
-                data_collator, description="training"
-            )
-
-        dataloader_params = {
-            "batch_size": self._train_batch_size
-            * self.args.steps_per_generation,  # < this is the change
-            "collate_fn": data_collator,
-            "num_workers": self.args.dataloader_num_workers,
-            "pin_memory": self.args.dataloader_pin_memory,
-            "persistent_workers": self.args.dataloader_persistent_workers,
-        }
-
-        if not isinstance(train_dataset, torch.utils.data.IterableDataset):
-            dataloader_params["sampler"] = self._get_train_sampler()
-            dataloader_params["drop_last"] = self.args.dataloader_drop_last
-            dataloader_params["worker_init_fn"] = partial(
-                seed_worker,
-                num_workers=self.args.dataloader_num_workers,
-                rank=self.args.process_index,
-            )
-            dataloader_params["prefetch_factor"] = self.args.dataloader_prefetch_factor
-
-        return self.accelerator.prepare(DataLoader(train_dataset, **dataloader_params))
-
 
 class AxolotlGRPOSequenceParallelTrainer(AxolotlGRPOTrainer):
     """Extend the base GRPOTrainer for sequence parallelism handling"""
@@ -117,7 +77,6 @@ class AxolotlGRPOSequenceParallelTrainer(AxolotlGRPOTrainer):
             torch.optim.Optimizer | None, torch.optim.lr_scheduler.LambdaLR | None
         ] = (None, None),
         peft_config: "PeftConfig | None" = None,
-        optimizer_cls_and_kwargs: tuple[type, dict] | None = None,
     ):
         # First call the superclass constructor with all arguments
         super().__init__(
@@ -131,7 +90,6 @@ class AxolotlGRPOSequenceParallelTrainer(AxolotlGRPOTrainer):
             callbacks=callbacks,
             optimizers=optimizers,
             peft_config=peft_config,
-            optimizer_cls_and_kwargs=optimizer_cls_and_kwargs,
         )
 
         # Get number of SP groups (number of processes divided by SP degree)
@@ -173,13 +131,6 @@ class AxolotlGRPOSequenceParallelTrainer(AxolotlGRPOTrainer):
                     f"the valid values for the number of generations are: {possible_values}."
                 )
 
-        self.sp_group = None
-        self.rank = dist.get_rank()
-        self.world_size = dist.get_world_size()
-        self.local_rank = 0
-        self.local_world_size = 1
-
-    def train(self, *args, **kwargs):
         # Initialize the SP group
         self.sp_group = get_ring_attn_group()
         self.rank = dist.get_rank()
@@ -187,8 +138,6 @@ class AxolotlGRPOSequenceParallelTrainer(AxolotlGRPOTrainer):
         self.local_rank = dist.get_rank(group=self.sp_group)
         self.local_world_size = dist.get_world_size(group=self.sp_group)
 
-        return super().train(*args, **kwargs)
-
     def _get_train_sampler(self) -> Sampler:
         effective_batch_size = (
             self.args.per_device_train_batch_size

src/axolotl/core/trainers/mixins/optimizer.py

@@ -1,18 +1,18 @@
 """Module for Axolotl trainer optimizer mixin"""
 
-import torch
+import logging
+
 from peft.optimizers import create_loraplus_optimizer
 from torch import nn
 from transformers.trainer import Trainer
 from transformers.utils import is_sagemaker_mp_enabled
 
 from axolotl.integrations.base import BaseOptimizerFactory
-from axolotl.utils.logging import get_logger
 
 if is_sagemaker_mp_enabled():
     import smdistributed.modelparallel.torch as smp
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger(__name__)
 
 
 class OptimizerMixin(Trainer):
@@ -186,12 +186,12 @@ class OptimizerMixin(Trainer):
                                 p.data_ptr(): p.numel() for p in module.parameters()
                             }.values()
                         )
-                        LOG.info(f"skipped {module}: {skipped / 2 ** 20}M params")
+                        LOG.info(f"skipped {module}: {skipped/2**20}M params")
                         manager.register_module_override(
                             module, "weight", {"optim_bits": 32}
                         )
                         LOG.debug(f"bitsandbytes: will optimize {module} in fp32")
-                LOG.info(f"skipped: {skipped / 2 ** 20}M params")
+                LOG.info(f"skipped: {skipped/2**20}M params")
 
         if is_sagemaker_mp_enabled():
             self.optimizer = smp.DistributedOptimizer(  # pylint: disable=attribute-defined-outside-init
@@ -199,25 +199,3 @@ class OptimizerMixin(Trainer):
             )
 
         return self.optimizer
-
-    def create_optimizer_and_scheduler(self, num_training_steps: int):
-        super().create_optimizer_and_scheduler(num_training_steps)
-        if self.args.compile_optimizer:
-            self.optimizer.step = torch.compile(self.optimizer.step)
-
-
-class OptimizerInitMixin:
-    """
-    Mixin to handle common optimizer initialization logic for Trainers (mostly TRL) that do not
-    accept optimizer_cls_and_kwargs as kwarg in constructor.
-    """
-
-    def __init__(self, *args, **kwargs):
-        optimizer_cls_and_kwargs = kwargs.pop("optimizer_cls_and_kwargs", None)
-        super().__init__(*args, **kwargs)
-        if (
-            optimizer_cls_and_kwargs
-            and self.optimizer_cls_and_kwargs is None
-            and self.optimizer is None
-        ):
-            self.optimizer_cls_and_kwargs = optimizer_cls_and_kwargs

src/axolotl/core/trainers/mixins/rng_state_loader.py

@@ -6,6 +6,7 @@ See https://github.com/huggingface/transformers/pull/37162
 TODO: Remove when upstream added PR to release
 """
 
+import logging
 import os
 import random
 
@@ -16,9 +17,7 @@ from transformers.trainer import safe_globals
 from transformers.trainer_pt_utils import set_rng_state_for_device
 from transformers.training_args import ParallelMode
 
-from axolotl.utils.logging import get_logger
+LOG = logging.getLogger(__name__)
-
-LOG = get_logger(__name__)
 
 
 class RngLoaderMixin(Trainer):

src/axolotl/core/trainers/mixins/scheduler.py

@@ -1,11 +1,12 @@
 """Module for Axolotl trainer scheduler mixin"""
 
+import logging
+
 import torch
 from torch.optim.lr_scheduler import LRScheduler, OneCycleLR
 from transformers.trainer import Trainer
 
 from axolotl.integrations.base import PluginManager
-from axolotl.utils.logging import get_logger
 from axolotl.utils.schedulers import (
     RexLR,
     get_cosine_schedule_with_min_lr,
@@ -13,7 +14,7 @@ from axolotl.utils.schedulers import (
     get_cosine_schedule_with_warmup_decay_constant,
 )
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger(__name__)
 
 
 class SchedulerMixin(Trainer):
@@ -79,15 +80,13 @@ class SchedulerMixin(Trainer):
                 self.lr_scheduler = RexLR(
                     optimizer=optimizer,
                     max_lr=self.args.learning_rate,
-                    min_lr=0 if not use_cosine_min_lr else (
+                    min_lr=0 if not use_cosine_min_lr else (self.args.learning_rate * self.args.cosine_min_lr_ratio),
-                        self.args.learning_rate * self.args.cosine_min_lr_ratio),
                     total_steps=num_training_steps,
                     num_warmup_steps=self.args.get_warmup_steps(num_training_steps),
                 )
             elif use_cosine_quadratic:
                 if use_cosine_min_lr:
-                    LOG.warning(
+                    LOG.warning("Both cosine quadratic warmup and min lr detected. Using quadratic warmup.")
-                        "Both cosine quadratic warmup and min lr detected. Using quadratic warmup.")
 
                 self.lr_scheduler = get_cosine_schedule_with_quadratic_warmup(  # pylint: disable=attribute-defined-outside-init
                     optimizer,
@@ -116,11 +115,9 @@ class SchedulerMixin(Trainer):
                 return super().create_scheduler(num_training_steps, optimizer=optimizer)
         else:
             if use_cosine_quadratic:
-                LOG.warning(
+                LOG.warning("axolotl's cosine scheduler with quadratic warmup not used (e.g., because of deepspeed).")
-                    "axolotl's cosine scheduler with quadratic warmup not used (e.g., because of deepspeed).")
 
             if use_cosine_min_lr:
-                LOG.warning(
+                LOG.warning("axolotl's cosine scheduler with min lr not used (e.g., because of deepspeed).")
-                    "axolotl's cosine scheduler with min lr not used (e.g., because of deepspeed).")
 
         return self.lr_scheduler  # type: ignore

src/axolotl/core/trainers/trl.py

@@ -1,5 +1,7 @@
 """Module for TRL PPO trainer"""
 
+from typing import Literal, Union
+
 import torch
 from tqdm import tqdm
 from trl import (
@@ -12,7 +14,6 @@ from trl import (
 )
 
 from axolotl.core.trainers.mixins import RngLoaderMixin
-from axolotl.core.trainers.mixins.optimizer import OptimizerInitMixin, OptimizerMixin
 from axolotl.core.trainers.mixins.scheduler import SchedulerMixin
 
 
@@ -74,19 +75,87 @@ class TRLPPOTrainer(PPOTrainer):
             )
 
 
-class AxolotlORPOTrainer(
+class AxolotlORPOTrainer(RngLoaderMixin, SchedulerMixin, ORPOTrainer):
-    RngLoaderMixin, SchedulerMixin, OptimizerMixin, OptimizerInitMixin, ORPOTrainer
-):
     """
     Extend the base ORPOTrainer for axolotl helpers
     """
 
     tag_names = ["axolotl", "orpo"]
 
+    def get_batch_loss_metrics(
+        self,
+        model,
+        batch: dict[str, Union[list, torch.LongTensor]],
+        train_eval: Literal["train", "eval"] = "train",
+    ):
+        """Compute the ORPO loss and other metrics for the given batch of inputs for train or test."""
 
-class AxolotlKTOTrainer(
+        # TODO remove once https://github.com/huggingface/trl/pull/3069 is included in a trl release
-    RngLoaderMixin, SchedulerMixin, OptimizerMixin, OptimizerInitMixin, KTOTrainer
+
-):
+        metrics = {}
+
+        forward_output = self.concatenated_forward(model, batch)
+        (
+            policy_chosen_logps,
+            policy_rejected_logps,
+            policy_chosen_logits,
+            policy_rejected_logits,
+            policy_nll_loss,
+        ) = forward_output[:5]
+        if self.aux_loss_enabled:
+            aux_loss = forward_output[5]
+
+        losses, chosen_rewards, rejected_rewards, log_odds_ratio, log_odds_chosen = (
+            self.odds_ratio_loss(policy_chosen_logps, policy_rejected_logps)
+        )
+        # full ORPO loss
+        loss = policy_nll_loss - losses.mean()
+
+        reward_accuracies = (chosen_rewards > rejected_rewards).float()
+
+        prefix = "eval_" if train_eval == "eval" else ""
+        metrics[f"{prefix}rewards/chosen"] = self.accelerator.gather_for_metrics(
+            chosen_rewards
+        ).mean()
+        metrics[f"{prefix}rewards/rejected"] = self.accelerator.gather_for_metrics(
+            rejected_rewards
+        ).mean()
+        metrics[f"{prefix}rewards/accuracies"] = self.accelerator.gather_for_metrics(
+            reward_accuracies
+        ).mean()
+        metrics[f"{prefix}rewards/margins"] = self.accelerator.gather_for_metrics(
+            chosen_rewards - rejected_rewards
+        ).mean()
+        metrics[f"{prefix}logps/rejected"] = (
+            self.accelerator.gather_for_metrics(policy_rejected_logps).detach().mean()
+        )
+        metrics[f"{prefix}logps/chosen"] = (
+            self.accelerator.gather_for_metrics(policy_chosen_logps).detach().mean()
+        )
+        metrics[f"{prefix}logits/rejected"] = self.accelerator.gather_for_metrics(
+            policy_rejected_logits.detach().mean()
+        ).mean()
+        metrics[f"{prefix}logits/chosen"] = self.accelerator.gather_for_metrics(
+            policy_chosen_logits.detach().mean()
+        ).mean()
+        metrics[f"{prefix}nll_loss"] = (
+            self.accelerator.gather_for_metrics(policy_nll_loss).detach().mean()
+        )
+        metrics[f"{prefix}log_odds_ratio"] = (
+            self.accelerator.gather_for_metrics(log_odds_ratio).detach().mean()
+        )
+        metrics[f"{prefix}log_odds_chosen"] = (
+            self.accelerator.gather_for_metrics(log_odds_chosen).detach().mean()
+        )
+        for k, v in metrics.items():
+            metrics[k] = v.item()
+        if self.aux_loss_enabled:
+            loss += self.aux_loss_coef * aux_loss
+
+        return loss, metrics
+
+
+class AxolotlKTOTrainer(RngLoaderMixin, SchedulerMixin, KTOTrainer):
     """
     Extend the base KTOTrainer for axolotl helpers
     """
@@ -94,19 +163,89 @@ class AxolotlKTOTrainer(
     tag_names = ["axolotl", "kto"]
 
 
-class AxolotlCPOTrainer(
+class AxolotlCPOTrainer(RngLoaderMixin, SchedulerMixin, CPOTrainer):
-    RngLoaderMixin, SchedulerMixin, OptimizerMixin, OptimizerInitMixin, CPOTrainer
-):
     """
     Extend the base CPOTrainer for axolotl helpers
     """
 
     tag_names = ["axolotl", "cpo"]
 
+    def get_batch_loss_metrics(
+        self,
+        model,
+        batch: dict[str, Union[list, torch.LongTensor]],
+        train_eval: Literal["train", "eval"] = "train",
+    ):
+        """Compute the CPO loss and other metrics for the given batch of inputs for train or test."""
+        metrics = {}
 
-class AxolotlRewardTrainer(
+        forward_output = self.concatenated_forward(model, batch)
-    RngLoaderMixin, SchedulerMixin, OptimizerMixin, OptimizerInitMixin, RewardTrainer
+        (
-):
+            policy_chosen_logps,
+            policy_rejected_logps,
+            policy_chosen_logits,
+            policy_rejected_logits,
+            policy_nll_loss,
+        ) = forward_output[:5]
+        if self.aux_loss_enabled:
+            aux_loss = forward_output[5]
+
+        losses, chosen_rewards, rejected_rewards = self.cpo_loss(
+            policy_chosen_logps,
+            policy_rejected_logps,
+        )
+
+        loss = losses.mean() + self.cpo_alpha * policy_nll_loss
+        reward_accuracies = (chosen_rewards > rejected_rewards).float()
+
+        prefix = "eval_" if train_eval == "eval" else ""
+        metrics[f"{prefix}rewards/chosen"] = (
+            self.accelerator.gather_for_metrics(chosen_rewards).mean().item()
+        )
+        metrics[f"{prefix}rewards/rejected"] = (
+            self.accelerator.gather_for_metrics(rejected_rewards).mean().item()
+        )
+        metrics[f"{prefix}rewards/accuracies"] = (
+            self.accelerator.gather_for_metrics(reward_accuracies).mean().item()
+        )
+        metrics[f"{prefix}rewards/margins"] = (
+            self.accelerator.gather_for_metrics(chosen_rewards - rejected_rewards)
+            .mean()
+            .item()
+        )
+        metrics[f"{prefix}logps/rejected"] = (
+            self.accelerator.gather_for_metrics(policy_rejected_logps)
+            .detach()
+            .mean()
+            .item()
+        )
+        metrics[f"{prefix}logps/chosen"] = (
+            self.accelerator.gather_for_metrics(policy_chosen_logps)
+            .detach()
+            .mean()
+            .item()
+        )
+        metrics[f"{prefix}logits/rejected"] = (
+            self.accelerator.gather_for_metrics(policy_rejected_logits.detach().mean())
+            .mean()
+            .item()
+        )
+        metrics[f"{prefix}logits/chosen"] = (
+            self.accelerator.gather_for_metrics(policy_chosen_logits.detach().mean())
+            .mean()
+            .item()
+        )
+        metrics[f"{prefix}nll_loss"] = (
+            self.accelerator.gather_for_metrics(policy_nll_loss).detach().mean().item()
+        )
+
+        if self.aux_loss_enabled:
+            loss += self.aux_loss_coef * aux_loss
+
+        return loss, metrics
+
+
+class AxolotlRewardTrainer(RngLoaderMixin, SchedulerMixin, RewardTrainer):
     """
     Extend the base RewardTrainer for axolotl helpers
     """
@@ -114,9 +253,7 @@ class AxolotlRewardTrainer(
     tag_names = ["axolotl", "reward"]
 
 
-class AxolotlPRMTrainer(
+class AxolotlPRMTrainer(RngLoaderMixin, SchedulerMixin, PRMTrainer):
-    RngLoaderMixin, SchedulerMixin, OptimizerMixin, OptimizerInitMixin, PRMTrainer
-):
     """
     Extend the base trl.PRMTrainer for axolotl helpers
     """

src/axolotl/core/training_args.py

@@ -141,10 +141,6 @@ class AxolotlTrainingMixins:
         default=None,
         metadata={"help": "absolute learning rate for the embedding layers."},
     )
-    compile_optimizer: Optional[bool] = field(
-        default=None,
-        metadata={"help": "Whether to compile the optimizer for faster training."},
-    )
     qlora: bool = field(
         default=False,
         metadata={"help": "whether this is a qlora training"},
@@ -168,6 +164,12 @@ class AxolotlTrainingMixins:
         default=None,
         metadata={"help": "whether to use sequential sampling for curriculum learning"},
     )
+    alternate_optimizer: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "workaround to pass an alternate optimizer to the HF trainer"
+        },
+    )
     alternate_lr_scheduler_type: Optional[str] = field(
         default=None,
         metadata={

src/axolotl/datasets.py

@@ -1,13 +1,12 @@
 """Module containing Dataset functionality"""
 
+import logging
 import os
 from typing import List, Optional, Union
 
 import torch
 from datasets import Dataset, IterableDataset
 
-from axolotl.utils.logging import get_logger
-
 from .prompt_tokenizers import PromptTokenizingStrategy
 
 # We want this to be a wrapper for an existing dataset that we have loaded
@@ -16,7 +15,7 @@ from .prompt_tokenizers import PromptTokenizingStrategy
 # let's check to ensure we don't truncate an item in the middle, we'll use
 # the collators later on to pad the datasets
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger("axolotl")
 
 
 class TokenizedPromptDataset(Dataset):

src/axolotl/integrations/base.py

@@ -22,6 +22,7 @@ from __future__ import annotations
 
 import collections
 import importlib
+import logging
 from typing import TYPE_CHECKING, Callable, OrderedDict, Union
 
 from peft import PeftModel
@@ -30,9 +31,6 @@ from torch.optim.lr_scheduler import LRScheduler
 from transformers import PreTrainedModel, Trainer
 
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.logging import get_logger
-
-LOG = get_logger(__name__, use_environ=True)
 
 if TYPE_CHECKING:
     from axolotl.common.datasets import TrainDatasetMeta
@@ -41,39 +39,31 @@ if TYPE_CHECKING:
 class BasePlugin:
     """Base class for all plugins. Defines the interface for plugin methods.
 
-    A plugin is a reusable, modular, and self-contained piece of code that extends
+    Methods:
-    the functionality of Axolotl. Plugins can be used to integrate third-party models,
+        register(cfg): Registers the plugin with the given configuration.
-    modify the training process, or add new features.
+        load_datasets(cfg): Loads and preprocesses the dataset for training.
-
+        pre_model_load(cfg): Performs actions before the model is loaded.
-    To create a new plugin, you need to inherit from the BasePlugin class and
+        post_model_build(cfg, model): Performs actions after the model is loaded, but
-    implement the required methods.
-
-    Note:
-        Plugin methods include:
-        - register(cfg): Registers the plugin with the given configuration.
-        - load_datasets(cfg): Loads and preprocesses the dataset for training.
-        - pre_model_load(cfg): Performs actions before the model is loaded.
-        - post_model_build(cfg, model): Performs actions after the model is loaded, but
             before LoRA adapters are applied.
-        - pre_lora_load(cfg, model): Performs actions before LoRA weights are loaded.
+        pre_lora_load(cfg, model): Performs actions before LoRA weights are loaded.
-        - post_lora_load(cfg, model): Performs actions after LoRA weights are loaded.
+        post_lora_load(cfg, model): Performs actions after LoRA weights are loaded.
-        - post_model_load(cfg, model): Performs actions after the model is loaded,
+        post_model_load(cfg, model): Performs actions after the model is loaded,
             inclusive of any adapters.
-        - post_trainer_create(cfg, trainer): Performs actions after the trainer is
+        post_trainer_create(cfg, trainer): Performs actions after the trainer is
             created.
-        - create_optimizer(cfg, trainer): Creates and returns an optimizer for training.
+        create_optimizer(cfg, trainer): Creates and returns an optimizer for training.
-        - create_lr_scheduler(cfg, trainer, optimizer, num_training_steps): Creates and
+        create_lr_scheduler(cfg, trainer, optimizer, num_training_steps): Creates and
             returns a learning rate scheduler.
-        - add_callbacks_pre_trainer(cfg, model): Adds callbacks to the trainer before
+        add_callbacks_pre_trainer(cfg, model): Adds callbacks to the trainer before
             training.
-        - add_callbacks_post_trainer(cfg, trainer): Adds callbacks to the trainer after
+        add_callbacks_post_trainer(cfg, trainer): Adds callbacks to the trainer after
             training.
     """
 
     def __init__(self):
         """Initializes the BasePlugin."""
 
-    def register(self, cfg: DictDefault):  # pylint: disable=unused-argument
+    def register(self, cfg):  # pylint: disable=unused-argument
         """Registers the plugin with the given configuration.
 
         Args:
@@ -285,11 +275,10 @@ class PluginManager:
     Attributes:
         plugins: A list of loaded plugins.
 
-    Note:
+    Methods:
-        Key methods include:
+        get_instance(): Static method to get the singleton instance of `PluginManager`.
-        - get_instance(): Static method to get the singleton instance of `PluginManager`.
+        register(plugin_name: str): Registers a new plugin by its name.
-        - register(plugin_name: str): Registers a new plugin by its name.
+        pre_model_load(cfg): Calls the pre_model_load method of all registered plugins.
-        - pre_model_load(cfg): Calls the pre_model_load method of all registered plugins.
     """
 
     plugins: OrderedDict[str, BasePlugin] = collections.OrderedDict()
@@ -333,12 +322,12 @@ class PluginManager:
             ImportError: If the plugin module cannot be imported.
         """
         try:
-            LOG.info(f"Attempting to load plugin: {plugin_name}")
+            logging.info(f"Attempting to load plugin: {plugin_name}")
             plugin = load_plugin(plugin_name)
             self.plugins[plugin_name] = plugin
-            LOG.info(f"Plugin loaded successfully: {plugin_name}")
+            logging.info(f"Plugin loaded successfully: {plugin_name}")
         except ImportError:
-            LOG.error(f"Failed to load plugin: {plugin_name}")
+            logging.error(f"Failed to load plugin: {plugin_name}")
 
     def get_input_args(self) -> list[str]:
         """Returns a list of Pydantic classes for all registered plugins' input arguments.'
@@ -545,6 +534,7 @@ class PluginManager:
 
         Args:
             cfg: The configuration for the plugins.
+            model: The loaded model.
         """
         for plugin in self.plugins.values():
             plugin.post_train_unload(cfg)

src/axolotl/integrations/cut_cross_entropy/__init__.py

@@ -19,16 +19,17 @@ Cut Cross Entropy is an optimized implementation of cross entropy loss
 from Apple's ML team.
 """
 import importlib
+import logging
 
 import torch
 
 from axolotl.integrations.base import BasePlugin
 from axolotl.utils import get_pytorch_version
-from axolotl.utils.logging import get_logger
+from axolotl.utils.distributed import is_main_process
 
 from .args import CutCrossEntropyArgs  # pylint: disable=unused-import. # noqa: F401
 
-LOG = get_logger(__name__, use_environ=True)
+LOG = logging.getLogger("axolotl.integrations.cut_cross_entropy")
 
 _CCE_INSTALL_MESSAGE = (
     "Please install cut_cross_entropy with transformers support using "
@@ -75,9 +76,10 @@ class CutCrossEntropyPlugin(BasePlugin):
                 cce_patch,
             )
 
-            LOG.info(
+            if is_main_process(use_environ=True):
-                f"Applying Cut Cross Entropy to model type: {cfg.model_config_type}"
+                LOG.info(
-            )
+                    f"Applying Cut Cross Entropy to model type: {cfg.model_config_type}"
+                )
 
             # The patch checks model_type internally
             cce_patch(cfg.model_config_type)

src/axolotl/integrations/cut_cross_entropy/args.py

@@ -15,13 +15,12 @@
 """
 Module for handling Cut Cross Entropy input arguments.
 """
+import logging
 from typing import Optional
 
 from pydantic import BaseModel, model_validator
 
-from axolotl.utils.logging import get_logger
+LOG = logging.getLogger("axolotl.integrations.cut_cross_entropy.args")
-
-LOG = get_logger(__name__)
 
 
 class CutCrossEntropyArgs(BaseModel):

src/axolotl/integrations/cut_cross_entropy/monkeypatch/mllama.py

@@ -15,14 +15,23 @@ from cut_cross_entropy.transformers.utils import (
 from transformers.cache_utils import Cache
 from transformers.modeling_outputs import CausalLMOutputWithPast
 from transformers.models.mllama.modeling_mllama import (
+    MLLAMA_INPUTS_DOCSTRING,
     _prepare_cross_attention_mask,
 )
+from transformers.utils import (
+    add_start_docstrings_to_model_forward,
+    replace_return_docstrings,
+)
 from transformers.utils.deprecation import deprecate_kwarg
 
 _PATCH_OPTS: PatchOptions | None = None
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+@add_start_docstrings_to_model_forward(MLLAMA_INPUTS_DOCSTRING)
+@replace_return_docstrings(
+    output_type=CausalLMOutputWithPast, config_class="MllamaTextConfig"
+)
 def cce_forward(
     self,
     input_ids: torch.LongTensor | None = None,
@@ -155,6 +164,10 @@ def cce_forward(
 
 
 @deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+@add_start_docstrings_to_model_forward(MLLAMA_INPUTS_DOCSTRING)
+@replace_return_docstrings(
+    output_type=CausalLMOutputWithPast, config_class="MllamaConfig"
+)
 def cce_forward_multimodal(
     self,
     input_ids: Optional[torch.LongTensor] = None,

src/axolotl/integrations/grokfast/__init__.py

@@ -2,15 +2,15 @@
 Grokfast plugin for Axolotl
 """
 
-from transformers.trainer_callback import TrainerCallback
+import logging
 
-from axolotl.utils.logging import get_logger
+from transformers.trainer_callback import TrainerCallback
 
 from ..base import BasePlugin
 from .args import GrokfastArgs  # pylint: disable=unused-import. # noqa: F401
 from .optimizer import gradfilter_ema
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger("axolotl.integrations.grokfast")
 
 
 class GrokfastCallbackHandler(TrainerCallback):

src/axolotl/integrations/liger/__init__.py

@@ -19,15 +19,16 @@ Liger Kernel is the collection of Triton-native kernels for LLM Training.
 It is designed to be performant, correct, and light-weight.
 """
 import inspect
+import logging
 import sys
 
 from axolotl.integrations.base import BasePlugin
-from axolotl.utils.logging import get_logger
+from axolotl.utils.distributed import is_main_process
 
 from .args import LigerArgs  # pylint: disable=unused-import. # noqa: F401
 from .utils import patch_with_compile_disable
 
-LOG = get_logger(__name__, use_environ=True)
+LOG = logging.getLogger("axolotl.integrations.liger")
 
 
 class LigerPlugin(BasePlugin):
@@ -84,7 +85,10 @@ class LigerPlugin(BasePlugin):
                 kwargs["geglu"] = cfg.liger_glu_activation
             elif "swiglu" in liger_fn_sig.parameters:
                 kwargs["swiglu"] = cfg.liger_glu_activation
-            LOG.info(f"Applying LIGER to {cfg.model_config_type} with kwargs: {kwargs}")
+            if is_main_process(use_environ=True):
+                LOG.info(
+                    f"Applying LIGER to {cfg.model_config_type} with kwargs: {kwargs}"
+                )
             apply_liger_fn(**kwargs)
         elif cfg.model_config_type == "jamba":
             from transformers.models.jamba import modeling_jamba
@@ -120,9 +124,9 @@ class LigerPlugin(BasePlugin):
             if cfg.liger_rope:
                 # The DeepseekV2 version of RoPE is different than upstream LLaMA.
                 # See https://github.com/linkedin/Liger-Kernel/issues/129#issuecomment-2313763528
-                LOG.warning("Fused liger_rope is not supported for DeepseekV2.")
+                logging.warning("Fused liger_rope is not supported for DeepseekV2.")
             if cfg.liger_glu_activation:
-                LOG.warning("liger_glu_activation is not supported for DeepseekV2.")
+                logging.warning("liger_glu_activation is not supported for DeepseekV2.")
             if cfg.liger_rms_norm:
                 modeling_mod.DeepseekV2RMSNorm = LigerRMSNorm
             if cfg.liger_glu_activation:
@@ -171,17 +175,7 @@ class LigerPlugin(BasePlugin):
                 rms_norm=cfg.liger_rms_norm,
                 layer_norm=cfg.liger_layer_norm,
             )
-        elif cfg.model_config_type == "granitemoe":
-            from liger_kernel.transformers import apply_liger_kernel_to_granite
-
-            apply_liger_kernel_to_granite(
-                rope=cfg.liger_rope,
-                cross_entropy=cfg.liger_cross_entropy,
-                fused_linear_cross_entropy=cfg.liger_fused_linear_cross_entropy,
-                rms_norm=cfg.liger_rms_norm,
-                swiglu=cfg.liger_glu_activation,
-            )
         else:
-            LOG.warning(
+            logging.warning(
                 f"Unsupported model config type: {cfg.model_config_type}. Liger not applied."
             )

src/axolotl/integrations/liger/args.py

@@ -15,13 +15,12 @@
 """
 Module for handling LIGER input arguments.
 """
+import logging
 from typing import Optional
 
 from pydantic import BaseModel, model_validator
 
-from axolotl.utils.logging import get_logger
+LOG = logging.getLogger("axolotl.integrations.liger.args")
-
-LOG = get_logger(__name__)
 
 
 class LigerArgs(BaseModel):

src/axolotl/integrations/llm_compressor/plugin.py

@@ -3,6 +3,7 @@ Sparse Finetuning plugin for Axolotl — enables handling of sparse neural netwo
 by maintaining masks for zero weights during training.
 """
 
+import logging
 from functools import wraps
 from typing import Any, Callable, Concatenate, ParamSpec, TypeVar
 
@@ -15,12 +16,11 @@ from transformers.trainer_callback import TrainerCallback, TrainerControl, Train
 from transformers.training_args import TrainingArguments
 
 from axolotl.integrations.base import BasePlugin
-from axolotl.utils.logging import get_logger
 
 P = ParamSpec("P")  # Params for generic function signatures
 R = TypeVar("R")  # Return type for generic function signatures
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger("axolotl.integrations.llm_compressor")
 
 
 class LLMCompressorCallbackHandler(TrainerCallback):

src/axolotl/integrations/spectrum/__init__.py

@@ -17,16 +17,14 @@ Spectrum Plugin to automatically generate unfrozen parameters based on SNR data.
 """
 
 import json
+import logging
 
 import requests
 
 from axolotl.integrations.base import BasePlugin
-from axolotl.utils.logging import get_logger
 
 from .args import SpectrumArgs  # pylint: disable=unused-import. # noqa: F401
 
-LOG = get_logger(__name__)
-
 
 def _generate_unfrozen_params_yaml(snr_data, top_fraction=0.5):
     unfrozen_parameters = {}
@@ -85,17 +83,17 @@ class SpectrumPlugin(BasePlugin):
         except FileNotFoundError:
             pass
         except Exception as exc:  # pylint: disable=broad-exception-caught
-            LOG.warning(f"Failed to read SNR data from {snr_path}: {exc}")
+            logging.warning(f"Failed to read SNR data from {snr_path}: {exc}")
 
         if not snr_data:
             try:
                 snr_data = requests.get(snr_url, timeout=60).json()
             except requests.exceptions.RequestException as exc:
-                LOG.warning(f"Failed to fetch SNR data from {snr_url}: {exc}")
+                logging.warning(f"Failed to fetch SNR data from {snr_url}: {exc}")
                 return
             # also catch json parsing errors
             except json.JSONDecodeError as exc:
-                LOG.warning(f"Failed to parse SNR data from {snr_url}: {exc}")
+                logging.warning(f"Failed to parse SNR data from {snr_url}: {exc}")
                 return
 
         unfrozen_parameters = _generate_unfrozen_params_yaml(

src/axolotl/kernels/lora.py

@@ -280,19 +280,19 @@ class LoRA_MLP(torch.autograd.Function):
         # Initialize and compute LoRA gradients
         d_down_A = d_down_B = d_up_A = d_up_B = d_gate_A = d_gate_B = None
 
-        if down_A is not None and down_B is not None:
+        if down_A is not None:
             d_down_A = h.t() @ (grad_output @ down_B.t())
             d_down_B = (down_A.t() @ h.t()) @ grad_output
             d_down_A *= down_scale
             d_down_B *= down_scale
 
-        if up_A is not None and up_B is not None:
+        if up_A is not None:
             d_up_A = X.t() @ (grad_up @ up_B.t())
             d_up_B = (up_A.t() @ X.t()) @ grad_up
             d_up_A *= up_scale
             d_up_B *= up_scale
 
-        if gate_A is not None and gate_B is not None:
+        if gate_A is not None:
             d_gate_A = X.t() @ (grad_gate @ gate_B.t())
             d_gate_B = (gate_A.t() @ X.t()) @ grad_gate
             d_gate_A *= gate_scale
@@ -311,7 +311,7 @@ class LoRA_MLP(torch.autograd.Function):
             del up_weight
 
             # Note the .to(dtype) only where mixing LoRA with base weights
-            if up_A is not None and up_B is not None:
+            if up_A is not None:
                 dX += grad_up @ up_B.to(dtype).t() @ (up_scale * up_A.to(dtype).t())
 
             # Gate projection gradients
@@ -319,7 +319,7 @@ class LoRA_MLP(torch.autograd.Function):
             dX += grad_gate @ gate_weight.t()
             del gate_weight
 
-            if gate_A is not None and gate_B is not None:
+            if gate_A is not None:
                 dX += (
                     grad_gate
                     @ gate_B.to(dtype).t()

src/axolotl/loaders/adapter.py

@@ -1,5 +1,6 @@
 """Adapter loading functionality, including LoRA / QLoRA and associated utils"""
 
+import logging
 import os
 import types
 from typing import Any
@@ -20,9 +21,8 @@ from transformers import PreTrainedModel
 
 from axolotl.loaders.utils import get_linear_embedding_layers
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.logging import get_logger
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger(__name__)
 
 
 def setup_quantized_meta_for_peft(model: torch.nn.Module):

src/axolotl/loaders/model.py

@@ -3,6 +3,7 @@ models.
 """
 
 import gc
+import logging
 import math
 import os
 from functools import cached_property
@@ -14,13 +15,7 @@ import torch
 import transformers
 import transformers.modeling_utils
 from accelerate import init_empty_weights
-from peft import (
+from peft import PeftConfig, PeftMixedModel, PeftModel, prepare_model_for_kbit_training
-    PeftConfig,
-    PeftMixedModel,
-    PeftModel,
-    PeftModelForCausalLM,
-    prepare_model_for_kbit_training,
-)
 from transformers import (
     AutoModelForCausalLM,
     AutoModelForVision2Seq,
@@ -52,11 +47,10 @@ from axolotl.utils.distributed import (
     get_device_count,
     get_device_type,
 )
-from axolotl.utils.logging import get_logger
 from axolotl.utils.model_shard_quant import load_sharded_model_quant
 from axolotl.utils.schemas.enums import RLType
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger(__name__)
 PLUGIN_MANAGER = PluginManager.get_instance()
 
 
@@ -145,7 +139,7 @@ class ModelLoader:
         """Property that determines if FSDP with QLoRA is enabled."""
         return self.cfg.fsdp and self.cfg.adapter == "qlora"
 
-    def load(self) -> tuple[PreTrainedModel | PeftModelForCausalLM, PeftConfig | None]:
+    def load(self) -> tuple[PreTrainedModel, PeftConfig | None]:
         """Load and prepare the model with all configurations and patches.
 
         Returns:
@@ -197,7 +191,6 @@ class ModelLoader:
         self._adjust_model_config()
         self._log_memory_usage()
         self._configure_embedding_dtypes()
-        self._configure_qat()
 
     def _resize_token_embeddings(self):
         """Resize token embeddings if needed."""
@@ -312,19 +305,6 @@ class ModelLoader:
                 before_kbit_train_or_finetune=False,
             )
 
-    def _configure_qat(self):
-        """Configure QAT."""
-        if self.cfg.qat:
-            from axolotl.utils.quantization import prepare_model_for_qat
-
-            prepare_model_for_qat(
-                self.model,
-                self.cfg.qat.weight_dtype,
-                self.cfg.qat.group_size,
-                self.cfg.qat.activation_dtype,
-                self.cfg.qat.quantize_embedding,
-            )
-
     def _load_adapters(self) -> PeftConfig | None:
         """Load LoRA or other adapters."""
         # Load LoRA or adapter
@@ -556,18 +536,11 @@ class ModelLoader:
         if self.cfg.low_cpu_mem_usage:
             self.model_kwargs["low_cpu_mem_usage"] = True
 
-    def _configure_zero3_memory_efficient_loading(
+    def _configure_zero3_memory_efficient_loading(self):
-        self,
+        """Set the deepspeed config to load the model into RAM first before moving
-    ) -> HfTrainerDeepSpeedConfig | None:
+        to VRAM.
-        """
-        Set the deepspeed config to load the model into RAM first before moving to VRAM.
 
-        IMPORTANT
+        We need to return `hf_ds_cfg` as it needs to exist before model loading.
-        ==========
-
-        We need to return `hf_ds_cfg` as it needs to exist before model loading for zero3.
-        HfTrainerDeepSpeedConfig is a class that is used to configure the DeepSpeed training.
-        It is not passed anywhere in the model loading function, just need to exist.
         """
         hf_ds_cfg = None
 
@@ -632,8 +605,7 @@ class ModelLoader:
                 if "device_map" in self.model_kwargs:
                     del self.model_kwargs["device_map"]
 
-            # Please don't remove underscore binding without reading the fn docstring.
+            self._configure_zero3_memory_efficient_loading()
-            _ = self._configure_zero3_memory_efficient_loading()
 
             # Load model with random initialization if specified
             if self.cfg.random_init_weights:
@@ -703,8 +675,7 @@ class ModelLoader:
                     if "device_map" in self.model_kwargs:
                         del self.model_kwargs["device_map"]
 
-                # Please don't remove underscore binding without reading the fn docstring.
+                self._configure_zero3_memory_efficient_loading()
-                _ = self._configure_zero3_memory_efficient_loading()
 
                 self.model = self.auto_model_loader.from_pretrained(
                     self.base_model,

src/axolotl/loaders/patch_manager.py

@@ -4,6 +4,7 @@ Applies pre- and post-model load patches for various fixes and optimizations.
 """
 
 import importlib.util
+import logging
 from functools import cached_property
 
 import addict
@@ -16,9 +17,8 @@ from axolotl.monkeypatch.multipack import (
     patch_for_multipack,
 )
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.logging import get_logger
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger(__name__)
 PLUGIN_MANAGER = PluginManager.get_instance()
 
 
@@ -59,10 +59,9 @@ class PatchManager:
         self._apply_gradient_checkpointing_patches()
         self._patch_attention()
         self._apply_multipack_patches()
-        self._patch_loss_llama()
         self._patch_llama_derived_model()
         self._apply_mistral_cross_entropy_patch()
-        self._apply_self_attention_lora_patch()
+        self._apply_unsloth_self_attention_patch()
 
     def apply_post_model_load_patches(self, model: PreTrainedModel):
         """Apply patches that require the model instance."""
@@ -81,9 +80,9 @@ class PatchManager:
     def _apply_fsdp_patches(self):
         """Apply patches for FSDP configurations."""
         if self.cfg.fsdp_config and str(self.cfg.fsdp_config.fsdp_version) == "2":
-            from axolotl.monkeypatch.accelerate.fsdp2 import patch_accelerate_fsdp2
+            from axolotl.monkeypatch.accelerate.fsdp2 import patch_accelerate_fsdp_utils
 
-            patch_accelerate_fsdp2()
+            patch_accelerate_fsdp_utils()
 
     def _apply_adapter_patches(self):
         """Apply patches for adapter configurations."""
@@ -116,6 +115,13 @@ class PatchManager:
 
             patch_llama4_linearized_modeling()
 
+        if self.cfg.model_config_type == "gemma3":
+            from axolotl.monkeypatch.gemma3 import (
+                patch_gemma3conditionalgeneration_forward,
+            )
+
+            patch_gemma3conditionalgeneration_forward()
+
     def _apply_fp8_patches(self):
         """Apply patches for FP8 support."""
         if self.cfg.fp8:
@@ -163,9 +169,9 @@ class PatchManager:
 
             patch_mistral_cross_entropy()
 
-    def _apply_self_attention_lora_patch(self):
+    def _apply_unsloth_self_attention_patch(self):
-        """Apply self-attention LoRA patches if configured."""
+        """Apply Unsloth self-attention patches if configured."""
-        if self.cfg.lora_qkv_kernel or self.cfg.lora_o_kernel:
+        if self.cfg.unsloth_lora_qkv or self.cfg.unsloth_lora_o:
             from axolotl.monkeypatch.lora_kernels import patch_self_attn_lora
 
             patch_self_attn_lora(self.cfg)
@@ -200,11 +206,19 @@ class PatchManager:
                 has_remote_code=has_remote_code,
             )
 
+            if self.cfg.is_llama_derived_model:
+                self._patch_loss_llama()
+
     def _patch_attention(self):
         """Apply attention-specific patches based on model type."""
         if not (self.cfg.flash_attention and hasattr(self.model_config, "model_type")):
             return
 
+        if self.model_config.model_type == "mllama" and self.cfg.flash_attention:
+            from axolotl.monkeypatch.attention.mllama import patch_mllama
+
+            patch_mllama()
+
         if self.model_config.model_type == "btlm":
             from axolotl.monkeypatch.btlm_attn_hijack_flash import (
                 replace_btlm_attn_with_flash_attn,
@@ -221,9 +235,6 @@ class PatchManager:
 
     def _patch_loss_llama(self):
         """Patch loss functions and other optimizations for LLaMA models."""
-        if not self.cfg.is_llama_derived_model:
-            return
-
         if self.cfg.flash_attn_cross_entropy and self.has_flash_attn:
             from axolotl.monkeypatch.llama_attn_hijack_flash import (
                 patch_fa_llama_cross_entropy,
@@ -303,6 +314,8 @@ class PatchManager:
             and (self.cfg.flash_attention or self.cfg.flex_attention)
             and self.cfg.sample_packing
         ):
+            self._patch_loss_llama()
+
             if self.cfg.flash_attention:
                 self._patch_llama_flash_attention(packed=self.cfg.sample_packing)
             elif self.cfg.xformers_attention:

src/axolotl/loaders/processor.py

@@ -1,5 +1,6 @@
 """Processor loading functionality for multi-modal models"""
 
+import logging
 from typing import Any
 
 import transformers
@@ -9,9 +10,8 @@ from transformers import (
 )
 
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.logging import get_logger
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger(__name__)
 
 
 def load_processor(cfg: DictDefault, tokenizer: PreTrainedTokenizerBase):

src/axolotl/loaders/tokenizer.py

@@ -1,6 +1,7 @@
 """Tokenizer loading functionality and associated utils"""
 
 import json
+import logging
 import os
 
 import transformers
@@ -18,9 +19,8 @@ from axolotl.utils.distributed import (
     is_local_main_process,
     is_main_process,
 )
-from axolotl.utils.logging import get_logger
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger(__name__)
 PLUGIN_MANAGER = PluginManager.get_instance()
 
 

src/axolotl/loaders/utils.py

@@ -1,6 +1,7 @@
 """Utilities for axolotl.loaders module"""
 
 import contextlib
+import logging
 from typing import Type
 
 import addict
@@ -8,9 +9,8 @@ import torch
 from transformers import AutoConfig, PretrainedConfig, PreTrainedModel
 
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.logging import get_logger
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger(__name__)
 
 
 def get_module_class_from_name(

src/axolotl/logging_config.py

@@ -2,56 +2,14 @@
 Common logging module for axolotl
 """
 
-import logging
 import os
 import sys
-from logging import Formatter, Logger, LogRecord
+from logging import Formatter
 from logging.config import dictConfig
 from typing import Any, Dict
 
 from colorama import Fore, Style, init
 
-DEFAULT_AXOLOTL_LOG_LEVEL = "INFO"
-DEFAULT_LOG_LEVEL = "WARNING"
-
-
-class AxolotlOrWarnErrorFilter(logging.Filter):
-    """
-    Allows ANY WARNING or higher (unless overridden by LOG_LEVEL)
-    Allows axolotl.* at INFO or higher (unless overridden by AXOLOTL_LOG_LEVEL)
-    Drops all other records (i.e. non-axolotl.INFO, DEBUG, etc. by default)
-    """
-
-    def __init__(self, **kwargs):
-        super().__init__(**kwargs)
-
-        self.axolotl_level = logging.getLevelNamesMapping()[
-            os.getenv("AXOLOTL_LOG_LEVEL", DEFAULT_AXOLOTL_LOG_LEVEL)
-        ]
-        self.other_level = logging.getLevelNamesMapping()[
-            os.getenv("LOG_LEVEL", DEFAULT_LOG_LEVEL)
-        ]
-
-    def filter(self, record: LogRecord) -> bool:
-        # General filter
-        if record.levelno >= self.other_level:
-            return True
-
-        # Axolotl filter
-        return (
-            record.name.startswith("axolotl") and record.levelno >= self.axolotl_level
-        )
-
-
-class AxolotlLogger(Logger):
-    """A Logger that automatically rejects non-axolotl INFOs."""
-
-    def __init__(self, name: str, level: int = logging.NOTSET):
-        super().__init__(name, level)
-
-        # set global filter on the logger itself
-        self.addFilter(AxolotlOrWarnErrorFilter())
-
 
 class ColorfulFormatter(Formatter):
     """
@@ -97,15 +55,11 @@ DEFAULT_LOGGING_CONFIG: Dict[str, Any] = {
             "stream": sys.stdout,
         },
     },
-    # log level will be superseded by the AxolotlLogger
+    "root": {"handlers": ["console"], "level": os.getenv("LOG_LEVEL", "INFO")},
-    "root": {
-        "handlers": ["console"],
-        "level": os.getenv("LOG_LEVEL", DEFAULT_LOG_LEVEL),
-    },
     "loggers": {
         "axolotl": {
             "handlers": ["color_console"],
-            "level": os.getenv("AXOLOTL_LOG_LEVEL", DEFAULT_AXOLOTL_LOG_LEVEL),
+            "level": "DEBUG",
             "propagate": False,
         },
     },
@@ -116,8 +70,3 @@ def configure_logging():
     """Configure with default logging"""
     init()  # Initialize colorama
     dictConfig(DEFAULT_LOGGING_CONFIG)
-    logging.setLoggerClass(AxolotlLogger)
-
-    # set default `ACCELERATE_LOG_LEVEL` to `LOG_LEVEL` if available and not set
-    if "ACCELERATE_LOG_LEVEL" not in os.environ:
-        os.environ["ACCELERATE_LOG_LEVEL"] = os.getenv("LOG_LEVEL", DEFAULT_LOG_LEVEL)

src/axolotl/monkeypatch/accelerate/fsdp2.py

@@ -1,14 +1,13 @@
 """
-monkeypatch for accelerate fsdp2 fix when modifying ordereddict during interation, and saving full state dicts
+monkeypatch for accelerate fsdp2 fix when modifying ordereddict during interation
 """
 
+import logging
 import sys
 
 import torch
 
-from axolotl.utils.logging import get_logger
+LOG = logging.getLogger(__name__)
-
-LOG = get_logger(__name__)
 
 
 def fsdp2_load_full_state_dict(accelerator, model: torch.nn.Module, full_sd: dict):
@@ -18,65 +17,27 @@ def fsdp2_load_full_state_dict(accelerator, model: torch.nn.Module, full_sd: dic
 
     Args:
         accelerator (`Accelerator`): The accelerator instance
-        model (`torch.nn.Module`):
+        model (`torch.nn.Module`): The model to load the state dict into
-            The model to load the state dict into, expected to be on meta device or a VRAM spike can occur
         full_sd (`dict`): The full state dict to load, can only be on rank 0
     """
     import torch.distributed as dist
     from torch.distributed.tensor import distribute_tensor
 
-    # Model was previously copied to meta device
+    LOG.info("Broadcasting full state dict to all ranks...")
-    meta_sharded_sd = model.state_dict()
+    sharded_sd = model.state_dict()
-    sharded_sd = {}
+    param_names = sorted(sharded_sd.keys())
-
-    # Rank 0 distributes the full state dict to other ranks
-    def _infer_parameter_dtype(model, param_name, empty_param):
-        try:
-            old_param = model.get_parameter_or_buffer(param_name)
-        except AttributeError:
-            # Need this for LORA, as there some params are not *parameters* of sorts
-            base_param_name, local_param_name = param_name.rsplit(".", 1)
-            submodule = model.get_submodule(base_param_name)
-            old_param = getattr(submodule, local_param_name)
-
-        is_torch_e4m3fn_available = hasattr(torch, "float8_e4m3fn")
-        casting_dtype = None
-        is_param_float8_e4m3fn = (
-            is_torch_e4m3fn_available and empty_param.dtype == torch.float8_e4m3fn
-        )
-
-        if empty_param.dtype.is_floating_point and not is_param_float8_e4m3fn:
-            casting_dtype = old_param.dtype
-
-        return old_param is not None and old_param.is_contiguous(), casting_dtype
-
-    def _cast_and_contiguous(tensor, to_contiguous, dtype):
-        if dtype is not None:
-            tensor = tensor.to(dtype=dtype)
-        if to_contiguous:
-            tensor = tensor.contiguous()
-        return tensor
-
-    param_names = sorted(meta_sharded_sd.keys())
-
     for param_name in param_names:
-        mesh = meta_sharded_sd[param_name].device_mesh
+        mesh = sharded_sd[param_name].device_mesh
         if accelerator.is_main_process:
+            # Use the corresponding tensor from full_sd (assuming the key exists in full_sd)
             full_param = full_sd[param_name].detach().cuda()
             dist.broadcast(full_param, src=0, group=mesh.get_group())
             sharded_tensor = distribute_tensor(
                 full_param, mesh, sharded_sd[param_name].placements
             )
-            to_contiguous, casting_dtype = _infer_parameter_dtype(
-                model,
-                param_name,
-                full_param,
-            )
-            sharded_tensor = _cast_and_contiguous(
-                sharded_tensor, to_contiguous, casting_dtype
-            )
             sharded_sd[param_name] = sharded_tensor
         else:
+            # Prepare a tensor of matching shape and dtype
             full_tensor = torch.empty(
                 sharded_sd[param_name].size(),
                 device="cuda",
@@ -86,113 +47,12 @@ def fsdp2_load_full_state_dict(accelerator, model: torch.nn.Module, full_sd: dic
             sharded_tensor = distribute_tensor(
                 full_tensor, mesh, sharded_sd[param_name].placements
             )
-            to_contiguous, casting_dtype = _infer_parameter_dtype(
-                model,
-                param_name,
-                full_tensor,
-            )
-            sharded_tensor = _cast_and_contiguous(
-                sharded_tensor, to_contiguous, casting_dtype
-            )
             sharded_sd[param_name] = sharded_tensor
 
-    # we set `assign=True` because our params are on meta device
     model.load_state_dict(sharded_sd, assign=True)
-    return model
 
 
-def get_state_dict(self, model, unwrap=True):
+def patch_accelerate_fsdp_utils():
-    """
-    Returns the state dictionary of a model sent through [`Accelerator.prepare`] potentially without full
-    precision.
-
-    Args:
-        model (`torch.nn.Module`):
-            A PyTorch model sent through [`Accelerator.prepare`]
-        unwrap (`bool`, *optional*, defaults to `True`):
-            Whether to return the original underlying state_dict of `model` or to return the wrapped state_dict
-
-    Returns:
-        `dict`: The state dictionary of the model potentially without full precision.
-
-    Example:
-
-    ```python
-    >>> import torch
-    >>> from accelerate import Accelerator
-
-    >>> accelerator = Accelerator()
-    >>> net = torch.nn.Linear(2, 2)
-    >>> net = accelerator.prepare(net)
-    >>> state_dict = accelerator.get_state_dict(net)
-    ```
-    """
-    from accelerate import DistributedType
-    from accelerate.utils import compare_versions
-
-    if self.distributed_type == DistributedType.DEEPSPEED:
-        zero3_sharding = self.deepspeed_config["zero_optimization"]["stage"] == 3
-        tp_sharding = (
-            self.deepspeed_config.get("tensor_parallel", {}).get("autotp_size", 0) > 1
-        )
-        if zero3_sharding or tp_sharding:
-            if model.zero_gather_16bit_weights_on_model_save():
-                if tp_sharding and not compare_versions("deepspeed", ">=", "0.16.4"):
-                    raise ImportError(
-                        "Deepspeed TP requires deepspeed >= 0.16.4, Please update DeepSpeed via `pip install deepspeed -U`."
-                    )
-                state_dict = (
-                    model._consolidated_16bit_state_dict()  # pylint: disable=protected-access
-                    if tp_sharding
-                    else model._zero3_consolidated_16bit_state_dict()  # pylint: disable=protected-access
-                )
-            else:
-                raise ValueError(
-                    "Cannot get 16bit model weights because `stage3_gather_16bit_weights_on_model_save` in DeepSpeed config is False. "
-                    "To save the model weights in 16bit, set `stage3_gather_16bit_weights_on_model_save` to True in DeepSpeed config file or "
-                    "set `zero3_save_16bit_model` to True when using `accelerate config`. "
-                    "To save the full checkpoint, run `model.save_checkpoint(save_dir)` and use `zero_to_fp32.py` to recover weights."
-                )
-        else:
-            from deepspeed.checkpoint.utils import clone_tensors_for_torch_save
-
-            state_dict = clone_tensors_for_torch_save(
-                self.unwrap_model(model).state_dict()
-            )
-    elif self.is_fsdp2:
-        # https://github.com/pytorch/torchtune/blob/main/torchtune/training/_distributed.py#L465
-        state_dict = {}
-        sharded_state_dict = model.state_dict()
-        for param_name, param in sharded_state_dict.items():
-            if param.is_cpu:
-                param = param.to(torch.device("cuda"))
-
-            param = param.full_tensor()
-            if torch.distributed.get_rank() == 0:
-                state_dict[param_name] = param.cpu()
-            torch.distributed.barrier()
-    elif self.distributed_type == DistributedType.FSDP:
-        from torch.distributed.fsdp import FullStateDictConfig
-        from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
-        from torch.distributed.fsdp import StateDictType
-
-        full_state_dict_config = FullStateDictConfig(
-            offload_to_cpu=True, rank0_only=True
-        )
-        with FSDP.state_dict_type(
-            model, StateDictType.FULL_STATE_DICT, full_state_dict_config
-        ):
-            state_dict = model.state_dict()
-    else:
-        if unwrap:
-            model = self.unwrap_model(model)
-        state_dict = model.state_dict()
-
-    return state_dict
-
-
-def patch_accelerate_fsdp2():
-    import accelerate
     from accelerate.utils import fsdp_utils
 
     fsdp_utils.fsdp2_load_full_state_dict = fsdp2_load_full_state_dict
@@ -201,10 +61,3 @@ def patch_accelerate_fsdp2():
         "fsdp2_load_full_state_dict",
         fsdp2_load_full_state_dict,
     )
-
-    accelerate.Accelerator.get_state_dict = get_state_dict
-    setattr(
-        sys.modules["accelerate"],
-        "Accelerator.get_state_dict",
-        get_state_dict,
-    )

src/axolotl/monkeypatch/attention/mllama.py

@@ -0,0 +1,230 @@
+"""
+Monkeypatch for Vision Llama for FA2 support
+"""
+
+# pylint: disable=duplicate-code
+
+from typing import Optional, Tuple
+
+import torch
+from flash_attn.flash_attn_interface import flash_attn_func
+from transformers.cache_utils import Cache
+from transformers.modeling_flash_attention_utils import _flash_attention_forward
+from transformers.models.mllama.configuration_mllama import MllamaTextConfig
+from transformers.models.mllama.modeling_mllama import (
+    MllamaTextCrossAttention,
+    MllamaTextSelfAttention,
+    apply_rotary_pos_emb,
+    repeat_kv,
+)
+from transformers.utils import is_flash_attn_greater_or_equal_2_10
+
+
+class MllamaTextCrossFlashAttention2(MllamaTextCrossAttention):
+    """
+    Mllama flash cross-attention module. This module inherits from `MllamaTextCrossAttention` and
+    implements the forward pass using Flash Attention for improved performance.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # Check if flash attention version is greater or equal to 2.1
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cross_attention_states: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Cache] = None,
+        attention_mask: Optional[  # pylint: disable=unused-argument
+            torch.Tensor
+        ] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,  # pylint: disable=unused-argument
+        cache_position: Optional[torch.LongTensor] = None,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        query_states = query_states.view(
+            bsz, q_len, self.num_heads, self.head_dim
+        ).transpose(1, 2)
+        query_states = self.q_norm(query_states)
+
+        if cross_attention_states is not None:
+            key_states = self.k_proj(cross_attention_states)
+            value_states = self.v_proj(cross_attention_states)
+            key_states = key_states.view(
+                bsz, -1, self.num_key_value_heads, self.head_dim
+            ).transpose(1, 2)
+            value_states = value_states.view(
+                bsz, -1, self.num_key_value_heads, self.head_dim
+            ).transpose(1, 2)
+            key_states = repeat_kv(key_states, self.num_key_value_groups)
+            value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+            key_states = self.k_norm(key_states)
+            if past_key_value is not None:
+                key_states, value_states = past_key_value.update(
+                    key_states,
+                    value_states,
+                    self.layer_idx,
+                    {"cache_position": cache_position},
+                )
+        elif cache_position[0] != 0:
+            key_states, value_states = (
+                past_key_value.key_cache[self.layer_idx],
+                past_key_value.value_cache[self.layer_idx],
+            )
+        else:
+            raise ValueError(
+                "Cross attention layer can't find neither `cross_attn_states` nor cached values for key/values!"
+            )
+
+        # Transpose to get the expected layout for flash attention
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        # Apply Flash Attention
+        dropout_rate = self.dropout if self.training else 0.0
+        output = flash_attn_func(
+            query_states,
+            key_states,
+            value_states,
+            dropout_p=dropout_rate,
+            softmax_scale=None,
+            causal=False,
+            return_attn_probs=output_attentions,
+        )
+
+        attn_output = output.contiguous().view(bsz, q_len, -1)
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+class MllamaTextSelfFlashAttention2(MllamaTextSelfAttention):
+    """
+    Mllama flash self-attention module. This module inherits from `MllamaTextSelfAttention` and
+    implements the forward pass using Flash Attention for improved performance.
+    """
+
+    def __init__(self, config: MllamaTextConfig, layer_idx: int, *args, **kwargs):
+        super().__init__(config, layer_idx, *args, **kwargs)
+
+        # Check if flash attention version is greater or equal to 2.1
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,  # pylint: disable=unused-argument
+        past_key_value=None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **kwargs,  # pylint: disable=unused-argument
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        output_attentions = False
+
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x num_heads x head_dim
+        query_states = query_states.view(
+            bsz, q_len, self.num_heads, self.head_dim
+        ).transpose(1, 2)
+        key_states = key_states.view(
+            bsz, q_len, self.num_key_value_heads, self.head_dim
+        ).transpose(1, 2)
+        value_states = value_states.view(
+            bsz, q_len, self.num_key_value_heads, self.head_dim
+        ).transpose(1, 2)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(
+            query_states, key_states, cos, sin
+        )
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+            key_states, value_states = past_key_value.update(
+                key_states, value_states, self.layer_idx, cache_kwargs
+            )
+
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        # Transpose to get the expected layout for flash attention
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+
+        dropout_rate = self.dropout if self.training else 0.0
+
+        # Handle potential silent casting to float32
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            if torch.is_autocast_enabled():
+                target_dtype = torch.get_autocast_gpu_dtype()
+            elif hasattr(self.config, "_pre_quantization_dtype"):
+                target_dtype = (
+                    self.config._pre_quantization_dtype  # pylint: disable=protected-access
+                )
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+
+        attn_output = _flash_attention_forward(
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            q_len,
+            dropout=dropout_rate,
+            use_top_left_mask=self._flash_attn_uses_top_left_mask,
+            is_causal=True,
+        )
+
+        attn_output = attn_output.reshape(bsz, q_len, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+def patch_mllama():
+    from transformers.models.mllama.modeling_mllama import (
+        MLLAMA_TEXT_ATTENTION_CLASSES,
+        MLLAMA_TEXT_CROSS_ATTENTION_CLASSES,
+        MLLAMA_VISION_ATTENTION_CLASSES,
+        MllamaPreTrainedModel,
+    )
+
+    MllamaPreTrainedModel._supports_flash_attn_2 = (  # pylint: disable=protected-access
+        True
+    )
+    MLLAMA_TEXT_ATTENTION_CLASSES["flash_attention_2"] = MllamaTextSelfFlashAttention2
+    MLLAMA_TEXT_CROSS_ATTENTION_CLASSES["flash_attention_2"] = (
+        MllamaTextCrossFlashAttention2
+    )
+    # fallback to SDPA
+    MLLAMA_VISION_ATTENTION_CLASSES["flash_attention_2"] = (
+        MLLAMA_VISION_ATTENTION_CLASSES["sdpa"]
+    )

src/axolotl/monkeypatch/btlm_attn_hijack_flash.py

@@ -3,6 +3,7 @@ Flash attention monkey patch for cerebras btlm model
 """
 
 import importlib
+import logging
 from typing import Optional, Tuple
 
 import torch
@@ -10,9 +11,7 @@ from accelerate import init_empty_weights
 from flash_attn.flash_attn_interface import flash_attn_func
 from transformers import AutoConfig, AutoModelForCausalLM
 
-from axolotl.utils.logging import get_logger
+LOG = logging.getLogger("axolotl")
-
-LOG = get_logger(__name__)
 
 
 def replace_btlm_attn_with_flash_attn(model_name="cerebras/btlm-3b-8k-base"):

src/axolotl/monkeypatch/gemma3.py

@@ -0,0 +1,230 @@
+"""Monkeypatch for gemma3 conditional generation forward to fix loss exploding"""
+
+# pylint: disable=duplicate-code
+
+from typing import Optional, Tuple, Union
+
+import torch
+from transformers.cache_utils import Cache
+from transformers.models.gemma3.modeling_gemma3 import (
+    Gemma3CausalLMOutputWithPast,
+    logger,
+)
+from transformers.utils import (
+    is_torchdynamo_compiling,
+)
+from transformers.utils.deprecation import deprecate_kwarg
+
+
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+def new_forward(
+    self,
+    input_ids: torch.LongTensor = None,
+    pixel_values: torch.FloatTensor = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[Union[list[torch.FloatTensor], Cache]] = None,
+    token_type_ids: Optional[torch.LongTensor] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **lm_kwargs,
+) -> Union[Tuple, Gemma3CausalLMOutputWithPast]:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.text_config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.text_config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from PIL import Image
+    >>> import requests
+    >>> from transformers import AutoProcessor, Gemma3ForConditionalGeneration
+
+    >>> model = Gemma3ForConditionalGeneration.from_pretrained("google/Gemma3-test-224px-hf")
+    >>> processor = AutoProcessor.from_pretrained("google/Gemma3-test-224px-hf")
+
+    >>> prompt = "answer en Where is the cow standing?"
+    >>> url = "https://huggingface.co/gv-hf/Gemma3-test-224px-hf/resolve/main/cow_beach_1.png"
+    >>> image = Image.open(requests.get(url, stream=True).raw)
+
+    >>> inputs = processor(images=image, text=prompt,  return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(**inputs, max_length=30)
+    >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "answer en Where is the cow standing?\nbeach"
+    ```"""
+
+    if (input_ids is None) ^ (inputs_embeds is not None):
+        raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+
+    is_training = token_type_ids is not None and labels is not None
+
+    # Replace image id with PAD if the image token is OOV, to avoid index-errors
+    if input_ids is not None and self.config.image_token_index >= self.vocab_size:
+        special_image_mask = input_ids == self.config.image_token_index
+        llm_input_ids = input_ids.clone()
+        llm_input_ids[special_image_mask] = 0
+    else:
+        llm_input_ids = input_ids
+
+    if inputs_embeds is None:
+        inputs_embeds = self.get_input_embeddings()(llm_input_ids)
+
+    if cache_position is None:
+        past_seen_tokens = (
+            past_key_values.get_seq_length() if past_key_values is not None else 0
+        )
+        cache_position = torch.arange(
+            past_seen_tokens,
+            past_seen_tokens + inputs_embeds.shape[1],
+            device=inputs_embeds.device,
+        )
+
+    # Merge text and images
+    if pixel_values is not None:
+        image_features = self.get_image_features(pixel_values)
+
+        if input_ids is None:
+            special_image_mask = inputs_embeds == self.get_input_embeddings()(
+                torch.tensor(
+                    self.config.image_token_index,
+                    dtype=torch.long,
+                    device=inputs_embeds.device,
+                )
+            )
+        else:
+            special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(
+                -1
+            )
+            special_image_mask = special_image_mask.expand_as(inputs_embeds).to(
+                inputs_embeds.device
+            )
+
+        if (
+            not is_torchdynamo_compiling()
+            and inputs_embeds[special_image_mask].numel() != image_features.numel()
+        ):
+            image_tokens_in_text = (special_image_mask).sum(dim=1).sum(dim=0)[0]
+            raise ValueError(
+                f"Number of images does not match number of special image tokens in the input text. "
+                f"Got {image_tokens_in_text} image tokens in the text but {image_features.shape[0] * image_features.shape[1]} "
+                "tokens from image embeddings."
+            )
+        image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
+        inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)
+
+    # mask out pad-token-ids in labels for BC
+    if labels is not None and self.pad_token_id in labels:
+        logger.warning_once(
+            "`labels` contains `pad_token_id` which will be masked with `config.ignore_index`. "
+            "You have to mask out `pad_token_id` when preparing `labels`, this behavior will be removed in v.4.46.",
+        )
+        labels = torch.where(
+            input_ids == self.pad_token_id, self.config.ignore_index, labels
+        )
+
+    causal_mask = self._update_causal_mask(  # pylint: disable=protected-access
+        attention_mask,
+        token_type_ids,
+        past_key_values,
+        cache_position,
+        inputs_embeds,
+        is_training,
+    )
+    outputs = self.language_model(
+        attention_mask=causal_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        logits_to_keep=logits_to_keep,
+        **lm_kwargs,
+    )
+
+    logits = outputs[0]
+    loss = None
+    if labels is not None:
+        if attention_mask is not None:
+            # Get the shifted attention mask
+            shift_attention_mask = attention_mask[:, -logits.shape[1] + 1 :].to(
+                logits.device
+            )  # +1 for shift
+
+            # Filter logits and labels based on attention mask
+            valid_indices = shift_attention_mask != 0
+            filtered_logits = logits[..., :-1, :][valid_indices]
+            filtered_labels = labels[..., 1:][valid_indices.to(labels.device)]
+
+            # TODO: do we need to handle num_items_in_batch given we filter the logits and labels?
+
+            loss = self.loss_function(
+                logits=filtered_logits,
+                labels=None,  # we pass shift_labels
+                shift_labels=filtered_labels,
+                vocab_size=self.config.text_config.vocab_size,
+                **lm_kwargs,
+            )
+        else:
+            # Standard case without filtering
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.text_config.vocab_size,
+                **lm_kwargs,
+            )
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return Gemma3CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        image_hidden_states=image_features if pixel_values is not None else None,
+    )
+
+
+def patch_gemma3conditionalgeneration_forward():
+    from transformers.models.gemma3.modeling_gemma3 import (
+        Gemma3ForConditionalGeneration,
+    )
+
+    Gemma3ForConditionalGeneration.forward = new_forward

src/axolotl/monkeypatch/gradient_checkpointing/offload_disk.py

@@ -18,6 +18,7 @@ DISCO - DIsk-based Storage and Checkpointing with Optimized prefetching
 
 import atexit
 import concurrent.futures
+import logging
 import os
 import queue
 import shutil
@@ -31,13 +32,11 @@ from typing import Dict
 
 import torch
 
-from axolotl.utils.logging import get_logger
-
 torch_cuda_amp_custom_fwd = torch.amp.custom_fwd(device_type="cuda")
 torch_cuda_amp_custom_bwd = torch.amp.custom_bwd(device_type="cuda")
 
 # Setup logger
-logger = get_logger(__name__)
+logger = logging.getLogger(__name__)
 
 
 class DiskOffloadManager:

src/axolotl/monkeypatch/llama_attn_hijack_flash.py

@@ -2,6 +2,7 @@
 
 # copied from https://github.com/lm-sys/FastChat/blob/main/fastchat/train/llama_flash_attn_monkey_patch.py
 
+import logging
 import warnings
 from typing import List, Optional, Tuple, Union
 
@@ -24,7 +25,6 @@ from transformers.models.llama.modeling_llama import (
 )
 
 from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids, set_module_name
-from axolotl.utils.logging import get_logger
 
 try:
     from flash_attn.flash_attn_interface import (  # pylint: disable=ungrouped-imports
@@ -41,7 +41,7 @@ except ImportError:
     )
 
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger("axolotl")
 
 
 def is_xformers_available() -> bool:
@@ -612,10 +612,9 @@ def generate_qkv(
             q, query_padding_mask
         )
 
-        def output_pad_fn(output_unpad):
+        output_pad_fn = lambda output_unpad: pad_input(  # noqa: E731
-            return pad_input(  # noqa: E731
+            output_unpad, indices_q, batch_size, seqlen_q
-                output_unpad, indices_q, batch_size, seqlen_q
+        )
-            )
 
     else:
         q_unpad = rearrange(q, "b s h d -> (b s) h d")
@@ -628,10 +627,9 @@ def generate_qkv(
         )
         max_seqlen_q = seqlen_q
 
-        def output_pad_fn(output_unpad):
+        output_pad_fn = lambda output_unpad: rearrange(  # noqa: E731
-            return rearrange(  # noqa: E731
+            output_unpad, "(b s) h d -> b s h d", b=batch_size
-                output_unpad, "(b s) h d -> b s h d", b=batch_size
+        )
-            )
 
     if key_padding_mask is not None:
         k_unpad, _, cu_seqlens_k, max_seqlen_k = unpad_input(k, key_padding_mask)

src/axolotl/monkeypatch/llama_attn_hijack_xformers.py

@@ -2,6 +2,7 @@
 Directly copied the code from https://raw.githubusercontent.com/oobabooga/text-generation-webui/main/modules/llama_attn_hijack.py and made some adjustments
 """
 
+import logging
 import warnings
 from typing import Optional, Tuple
 
@@ -10,14 +11,10 @@ import torch.nn.functional as F
 import transformers.models.llama.modeling_llama
 from transformers.models.llama.modeling_llama import apply_rotary_pos_emb, repeat_kv
 
-from axolotl.utils.logging import get_logger
-
-LOG = get_logger(__name__)
-
 try:
     import xformers.ops
 except ImportError:
-    LOG.error("xformers not found! Please install it before trying to use it.")
+    logging.error("xformers not found! Please install it before trying to use it.")
 
 
 def hijack_llama_attention():

src/axolotl/monkeypatch/lora_kernels.py

@@ -7,6 +7,7 @@ import types
 from typing import Generator, Tuple, Type
 
 import torch
+from accelerate.logging import get_logger
 from peft import PeftModelForCausalLM
 from torch import nn
 from transformers import AutoConfig
@@ -19,7 +20,6 @@ from axolotl.kernels.lora import (
 )
 from axolotl.monkeypatch.utils import detab_code
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.logging import get_logger
 
 LOG = get_logger(__name__)
 
@@ -342,11 +342,10 @@ def apply_lora_kernel_patches(
 
     layers = []
     # check for multimodal models first
-    pretrained_model = model.model
+    if hasattr(model, "language_model"):
-    if hasattr(pretrained_model, "language_model"):
+        layers = model.language_model.model.layers
-        layers = pretrained_model.language_model.layers
+    elif hasattr(model, "model"):
-    elif hasattr(pretrained_model, "model"):
+        layers = model.model.model.layers
-        layers = pretrained_model.model.layers
     else:
         raise NotImplementedError(
             f"Model type {model.config.model_type} is not supported yet. Please create an Issue."

src/axolotl/monkeypatch/mistral_attn_hijack_flash.py

@@ -2,6 +2,7 @@
 
 # pylint: disable=duplicate-code
 
+import logging
 from functools import partial
 from typing import List, Optional, Tuple, Union
 
@@ -27,9 +28,8 @@ from transformers.models.mistral.modeling_mistral import (
 )
 
 from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
-from axolotl.utils.logging import get_logger
 
-LOG = get_logger(__name__)
+LOG = logging.getLogger("axolotl.monkeypatch.mistral")
 
 
 def replace_mistral_attn_with_flash_attn(
@@ -359,10 +359,9 @@ def generate_qkv(
             q, query_padding_mask
         )
 
-        def output_pad_fn(output_unpad):
+        output_pad_fn = lambda output_unpad: pad_input(  # noqa: E731
-            return pad_input(  # noqa: E731
+            output_unpad, indices_q, batch_size, seqlen_q
-                output_unpad, indices_q, batch_size, seqlen_q
+        )
-            )
 
     else:
         q_unpad = rearrange(q, "b s h d -> (b s) h d")
@@ -375,10 +374,9 @@ def generate_qkv(
         )
         max_seqlen_q = seqlen_q
 
-        def output_pad_fn(output_unpad):
+        output_pad_fn = lambda output_unpad: rearrange(  # noqa: E731
-            return rearrange(  # noqa: E731
+            output_unpad, "(b s) h d -> b s h d", b=batch_size
-                output_unpad, "(b s) h d -> b s h d", b=batch_size
+        )
-            )
 
     if key_padding_mask is not None:
         k_unpad, _, cu_seqlens_k, max_seqlen_k = unpad_input(k, key_padding_mask)