Update .pre-commit-config.yaml

Co-authored-by: Wing Lian <wing.lian@gmail.com>

.github/workflows/base.yml

@@ -22,12 +22,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: "124"
-            cuda_version: 12.4.1
-            cudnn_version: ""
-            python_version: "3.10"
-            pytorch: 2.4.1
-            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
           - cuda: "124"
             cuda_version: 12.4.1
             cudnn_version: ""
@@ -40,6 +34,18 @@ 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"
+          - 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"
+          - 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"
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -61,7 +67,7 @@ jobs:
         uses: docker/build-push-action@v4
         with:
           context: .
-          file: ./docker/Dockerfile-base
+          file: ${{ matrix.pytorch == 'nightly' && './docker/Dockerfile-base-nightly' || './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 }}

.github/workflows/docs.yml

@@ -19,7 +19,7 @@ jobs:
         - name: Setup Python
           uses: actions/setup-python@v5
           with:
-            python-version: '3.10'
+            python-version: '3.11'
         - name: install dependencies
           run: |
             python3 -m pip install jupyter

.github/workflows/lint.yml

@@ -19,6 +19,6 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions/setup-python@v5
         with:
-          python-version: "3.10"
+          python-version: "3.11"
           cache: 'pip' # caching pip dependencies
       - uses: pre-commit/action@v3.0.1

.github/workflows/main.yml

@@ -24,8 +24,13 @@ jobs:
             cuda_version: 12.4.1
             python_version: "3.11"
             pytorch: 2.5.1
-            axolotl_extras:
+            axolotl_extras: vllm
             is_latest: true
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.6.0
+            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -83,6 +88,11 @@ jobs:
             pytorch: 2.5.1
             axolotl_extras:
             is_latest: true
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.6.0
+            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout

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

@@ -4,6 +4,10 @@ on:
   pull_request:
     paths:
       - 'tests/e2e/multigpu/*.py'
+      - 'requirements.txt'
+      - 'setup.py'
+      - 'pyproject.toml'
+      - '.github/workflows/multi-gpu-e2e.yml'
   workflow_dispatch:
   schedule:
     - cron: '0 0 * * 1,4'  # Runs at 00:00 UTC every monday & thursday
@@ -24,13 +28,21 @@ jobs:
             cuda_version: 12.4.1
             python_version: "3.11"
             pytorch: 2.4.1
-            axolotl_extras:
+            axolotl_extras:  # no vllm support for 2.4.1
             num_gpus: 2
             nightly_build: "true"
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
             pytorch: 2.5.1
+            axolotl_extras: vllm
+            num_gpus: 2
+            nightly_build: "true"
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.6.0
+            # awaiting vllm#12721
             axolotl_extras:
             num_gpus: 2
             nightly_build: "true"
@@ -42,7 +54,7 @@ jobs:
       - name: Install Python
         uses: actions/setup-python@v5
         with:
-          python-version: "3.10"
+          python-version: "3.11"
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip

.github/workflows/nightlies.yml

@@ -22,6 +22,11 @@ jobs:
             python_version: "3.11"
             pytorch: 2.5.1
             axolotl_extras:
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.6.0
+            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -75,6 +80,11 @@ jobs:
             python_version: "3.11"
             pytorch: 2.5.1
             axolotl_extras:
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.6.0
+            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout

.github/workflows/precommit-autoupdate.yml

@@ -0,0 +1,49 @@
+name: Pre-commit auto-update
+
+on:
+  schedule:
+    - cron: '0 0 * * 0'  # Run weekly
+  workflow_dispatch:  # Manual kickoff
+
+jobs:
+  auto-update:
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+      pull-requests: write
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+
+      - name: Update pre-commit hooks
+        id: update
+        run: |
+          pip install pre-commit
+          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
+        if: steps.update.outputs.changes == 'true'
+        uses: peter-evans/create-pull-request@v6
+        with:
+          token: ${{ secrets.GITHUB_TOKEN }}
+          branch: update/pre-commit-hooks
+          delete-branch: true
+          title: "chore: update pre-commit hooks"
+          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/pypi.yml

@@ -36,11 +36,11 @@ jobs:
       - name: Setup Python
         uses: actions/setup-python@v5
         with:
-          python-version: "3.10"
+          python-version: "3.11"
 
       - name: Install dependencies
         run: |
-          pip3 install wheel packaging
+          pip3 install wheel packaging==23.2
           pip3 install --no-build-isolation -e .
           pip3 install -r requirements-dev.txt -r requirements-tests.txt
 

.github/workflows/tests-nightly.yml

@@ -12,7 +12,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions/setup-python@v5
         with:
-          python-version: "3.10"
+          python-version: "3.11"
           cache: 'pip' # caching pip dependencies
       - uses: pre-commit/action@v3.0.1
         env:
@@ -25,13 +25,8 @@ jobs:
       fail-fast: false
       max-parallel: 2
       matrix:
-        python_version: ["3.10", "3.11"]
-        pytorch_version: ["2.4.1", "2.5.1"]
-        exclude:
-          - python_version: "3.10"
-            pytorch_version: "2.4.1"
-          - python_version: "3.10"
-            pytorch_version: "2.5.1"
+        python_version: ["3.11"]
+        pytorch_version: ["2.4.1", "2.5.1", "2.6.0"]
     timeout-minutes: 20
 
     steps:
@@ -47,7 +42,7 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging setuptools wheel
+          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
 
       - name: Install PyTorch
         run: |
@@ -64,7 +59,7 @@ jobs:
       - name: Install dependencies
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging
+          pip3 install --upgrade packaging==23.2
           pip3 install --no-build-isolation -U -e .
           python scripts/unsloth_install.py | sh
           python scripts/cutcrossentropy_install.py | sh
@@ -112,13 +107,20 @@ jobs:
             num_gpus: 1
             axolotl_extras:
             nightly_build: "true"
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.6.0
+            num_gpus: 1
+            axolotl_extras:
+            nightly_build: "true"
     steps:
       - name: Checkout
         uses: actions/checkout@v4
       - name: Install Python
         uses: actions/setup-python@v5
         with:
-          python-version: "3.10"
+          python-version: "3.11"
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip

.github/workflows/tests.yml

@@ -35,7 +35,7 @@ jobs:
       - uses: actions/checkout@v4
       - uses: actions/setup-python@v5
         with:
-          python-version: "3.10"
+          python-version: "3.11"
           cache: 'pip' # caching pip dependencies
       - uses: pre-commit/action@v3.0.1
         env:
@@ -48,13 +48,8 @@ jobs:
       fail-fast: false
       max-parallel: 2
       matrix:
-        python_version: ["3.10", "3.11"]
-        pytorch_version: ["2.4.1", "2.5.1"]
-        exclude:
-          - python_version: "3.10"
-            pytorch_version: "2.4.1"
-          - python_version: "3.10"
-            pytorch_version: "2.5.1"
+        python_version: ["3.11"]
+        pytorch_version: ["2.4.1", "2.5.1", "2.6.0"]
     timeout-minutes: 20
 
     steps:
@@ -79,7 +74,7 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging setuptools wheel
+          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
 
       - name: Install PyTorch
         run: |
@@ -127,7 +122,7 @@ jobs:
       max-parallel: 1
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.4.1", "2.5.1"]
+        pytorch_version: ["2.4.1", "2.5.1", "2.6.0"]
     timeout-minutes: 20
 
     steps:
@@ -152,7 +147,7 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging setuptools setuptools_scm build wheel
+          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 setuptools_scm build wheel
 
       - name: Install PyTorch
         run: |
@@ -209,14 +204,14 @@ jobs:
             python_version: "3.11"
             pytorch: 2.5.1
             num_gpus: 1
-            axolotl_extras:
+            axolotl_extras: vllm
     steps:
       - name: Checkout
         uses: actions/checkout@v4
       - name: Install Python
         uses: actions/setup-python@v5
         with:
-          python-version: "3.10"
+          python-version: "3.11"
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip
@@ -251,13 +246,19 @@ jobs:
             pytorch: 2.4.1
             num_gpus: 1
             axolotl_extras:
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.6.0
+            num_gpus: 1
+            axolotl_extras:
     steps:
       - name: Checkout
         uses: actions/checkout@v4
       - name: Install Python
         uses: actions/setup-python@v5
         with:
-          python-version: "3.10"
+          python-version: "3.11"
       - name: Install Modal
         run: |
           python -m pip install --upgrade pip

.pre-commit-config.yaml

@@ -3,7 +3,7 @@ default_language_version:
 
 repos:
 -   repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v4.4.0
+    rev: v5.0.0
     hooks:
     -   id: check-yaml
     -   id: end-of-file-fixer
@@ -11,23 +11,23 @@ repos:
     -   id: no-commit-to-branch
         args: ['--branch', 'main']
 -   repo: https://github.com/psf/black
-    rev: 23.3.0
+    rev: 25.1.0
     hooks:
     -   id: black
 -   repo: https://github.com/pycqa/isort
-    rev: 5.12.0
+    rev: 6.0.1
     hooks:
       - id: isort
 -   repo: https://github.com/PyCQA/flake8
-    rev: 6.1.0
+    rev: 7.1.2
     hooks:
     - id: flake8
--   repo: https://github.com/PyCQA/pylint
-    rev: v3.3.0
+-   repo: https://github.com/pylint-dev/pylint
+    rev: v3.3.6
     hooks:
     - id: pylint
 -   repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.3.0
+    rev: v1.15.0
     hooks:
     - id: mypy
       additional_dependencies:
@@ -36,7 +36,7 @@ repos:
             'pydantic>=2.5.3',
         ]
 -   repo: https://github.com/PyCQA/bandit
-    rev: 1.7.5
+    rev: 1.8.3
     hooks:
     -   id: bandit
         args: [

README.md

@@ -19,9 +19,6 @@
     <br/>
     <img src="https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/tests-nightly.yml/badge.svg" alt="tests-nightly">
     <img src="https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/multi-gpu-e2e.yml/badge.svg" alt="multigpu-semi-weekly tests">
-    <a href="https://www.phorm.ai/query?projectId=e315ba4a-4e14-421f-ab05-38a1f9076f25">
-    <img alt="phorm.ai" src="https://img.shields.io/badge/Phorm-Ask_AI-%23F2777A.svg?&logo=data:image/svg+xml;base64,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">
-  </a>
 </p>
 
 Axolotl is a tool designed to streamline post-training for various AI models.
@@ -50,13 +47,15 @@ Features:
 ## 🚀 Quick Start
 
 **Requirements**:
+
 - NVIDIA GPU (Ampere or newer for `bf16` and Flash Attention) or AMD GPU
-- Python ≥3.10
+- Python 3.11
 - PyTorch ≥2.4.1
 
 ### Installation
 
-```shell
+```bash
+pip3 install -U packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation axolotl[flash-attn,deepspeed]
 
 # Download example axolotl configs, deepspeed configs
@@ -68,7 +67,7 @@ Other installation approaches are described [here](https://axolotl-ai-cloud.gith
 
 ### Your First Fine-tune
 
-```shell
+```bash
 # Fetch axolotl examples
 axolotl fetch examples
 

_quarto.yml

@@ -3,10 +3,12 @@ project:
 
 website:
   title: "Axolotl"
-  description: "Fine-tuning"
+  description: "We make fine-tuning accessible, scalable, and fun"
   favicon: favicon.jpg
+
   navbar:
-    title: Axolotl
+    logo: image/axolotl_logo_digital_white.svg
+    title: false
     background: dark
     pinned: false
     collapse: false
@@ -25,33 +27,56 @@ website:
       contents:
         - text: Home
           href: index.qmd
-        - section: "How-To Guides"
+
+        - section: "Getting Started"
           contents:
-          # TODO Edit folder structure after we have more docs.
             - docs/getting-started.qmd
             - docs/installation.qmd
-            - docs/debugging.qmd
             - docs/inference.qmd
-            - docs/multipack.qmd
-            - docs/fsdp_qlora.qmd
-            - docs/input_output.qmd
-            - docs/rlhf.qmd
-            - docs/nccl.qmd
-            - docs/mac.qmd
-            - docs/multi-gpu.qmd
-            - docs/multi-node.qmd
-            - docs/unsloth.qmd
-            - docs/amd_hpc.qmd
-            - docs/ray-integration.qmd
+            - docs/cli.qmd
+            - docs/config.qmd
+
         - section: "Dataset Formats"
           contents: docs/dataset-formats/*
-        - section: "Reference"
+
+        - section: "Deployments"
           contents:
-            - docs/config.qmd
-        - docs/faq.qmd
+            - docs/docker.qmd
+            - docs/multi-gpu.qmd
+            - docs/multi-node.qmd
+            - docs/ray-integration.qmd
+            - docs/amd_hpc.qmd
+            - docs/mac.qmd
+
+        - section: "How To Guides"
+          contents:
+            - docs/multimodal.qmd
+            - docs/rlhf.qmd
+            - docs/reward_modelling.qmd
+            - docs/lr_groups.qmd
+            - docs/lora_optims.qmd
+
+        - section: "Core Concepts"
+          contents:
+            - docs/batch_vs_grad.qmd
+            - docs/dataset_preprocessing.qmd
+            - docs/multipack.qmd
+
+        - section: "Advanced Features"
+          contents:
+            - docs/fsdp_qlora.qmd
+            - docs/unsloth.qmd
+            - docs/torchao.qmd
+            - docs/custom_integrations.qmd
+
+        - section: "Troubleshooting"
+          contents:
+            - docs/faq.qmd
+            - docs/debugging.qmd
+            - docs/nccl.qmd
 
 format:
   html:
-    theme: materia
+    theme: darkly
     css: styles.css
     toc: true

cicd/Dockerfile.jinja

@@ -31,6 +31,7 @@ RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
         sed -i 's#^datasets.*#datasets @ git+https://github.com/huggingface/datasets.git@main#' requirements.txt; \
     fi
 
+RUN pip install packaging==23.2 setuptools==75.8.0
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
         pip install --no-build-isolation -e .[deepspeed,flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \

cicd/cicd.sh

@@ -4,8 +4,8 @@ set -e
 python -c "import torch; assert '$PYTORCH_VERSION' in torch.__version__"
 
 pytest -v --durations=10 -n8 --ignore=tests/e2e/ --ignore=tests/patched/ /workspace/axolotl/tests/
-# pytest -v --durations=10 -n8 --dist loadfile /workspace/axolotl/tests/patched/
-pytest -v --durations=10 /workspace/axolotl/tests/e2e/patched/
+pytest -v --durations=10 /workspace/axolotl/tests/e2e/patched/lora_kernels  # running these with the other patches causes a failure
+pytest -v --durations=10 --ignore=tests/e2e/patched/lora_kernels /workspace/axolotl/tests/e2e/patched
 pytest -v --durations=10 -n1 /workspace/axolotl/tests/e2e/solo/
 pytest -v --durations=10 /workspace/axolotl/tests/e2e/integrations/
 pytest -v --durations=10 --ignore=tests/e2e/solo/ --ignore=tests/e2e/patched/ --ignore=tests/e2e/multigpu/ --ignore=tests/e2e/integrations/ /workspace/axolotl/tests/e2e/

cicd/multigpu.py

@@ -1,6 +1,7 @@
 """
- modal application to run axolotl gpu tests in Modal
- """
+modal application to run axolotl gpu tests in Modal
+"""
+
 # pylint: disable=duplicate-code
 
 import os
@@ -37,15 +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 = (
-    Image.from_dockerfile(
-        pathlib.Path(temp_dir) / "Dockerfile",
-        force_build=True,
-        gpu="A10G",
-    )
-    .env(df_args)
-    .pip_install("fastapi==0.110.0", "pydantic==2.6.3")
-)
+cicd_image = Image.from_dockerfile(
+    pathlib.Path(temp_dir) / "Dockerfile",
+    force_build=True,
+    gpu="A10G",
+).env(df_args)
 
 app = App("Axolotl CI/CD", secrets=[])
 

cicd/tests.py

@@ -1,6 +1,5 @@
-"""
- modal application to run axolotl gpu tests in Modal
- """
+"""Modal app to run axolotl GPU tests"""
+
 # pylint: disable=duplicate-code
 
 import os

docker/Dockerfile-base

@@ -28,7 +28,7 @@ ENV PATH="/root/miniconda3/envs/py${PYTHON_VERSION}/bin:${PATH}"
 
 WORKDIR /workspace
 
-RUN python3 -m pip install --upgrade pip && pip3 install packaging && \
+RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==23.2 setuptools==75.8.0 wheel && \
     python3 -m pip install --no-cache-dir -U torch==${PYTORCH_VERSION}+cu${CUDA} --extra-index-url https://download.pytorch.org/whl/cu$CUDA && \
     python3 -m pip install --no-cache-dir "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main" && \
     python3 -m pip install --no-cache-dir "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main"

docker/Dockerfile-base-nightly

@@ -0,0 +1,39 @@
+ARG CUDA_VERSION="12.8.1"
+ARG CUDNN_VERSION="8"
+ARG UBUNTU_VERSION="22.04"
+ARG MAX_JOBS=4
+
+FROM nvidia/cuda:$CUDA_VERSION-cudnn$CUDNN_VERSION-devel-ubuntu$UBUNTU_VERSION AS base-builder
+
+ENV PATH="/root/miniconda3/bin:${PATH}"
+
+ARG PYTHON_VERSION="3.11"
+ARG PYTORCH_VERSION="nightly"
+ARG CUDA="128"
+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
+
+RUN apt-get update \
+    && apt-get install -y wget git build-essential ninja-build git-lfs libaio-dev pkg-config && rm -rf /var/lib/apt/lists/* \
+    && wget \
+    https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh \
+    && mkdir /root/.conda \
+    && bash Miniconda3-latest-Linux-x86_64.sh -b \
+    && rm -f Miniconda3-latest-Linux-x86_64.sh \
+    && conda create -n "py${PYTHON_VERSION}" python="${PYTHON_VERSION}"
+
+ENV PATH="/root/miniconda3/envs/py${PYTHON_VERSION}/bin:${PATH}"
+
+WORKDIR /workspace
+
+RUN python3 -m pip install --upgrade pip && pip3 install packaging && \
+    python3 -m pip install --no-cache-dir -U torch --extra-index-url https://download.pytorch.org/whl/nightly/cu$CUDA && \
+    python3 -m pip install --no-cache-dir "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main" && \
+    python3 -m pip install --no-cache-dir "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main"
+
+RUN git lfs install --skip-repo && \
+    pip3 install awscli && \
+    # The base image ships with `pydantic==1.8.2` which is not working
+    pip3 install -U --no-cache-dir pydantic==1.10.10

docker/Dockerfile-cloud

@@ -14,7 +14,7 @@ COPY scripts/motd /etc/motd
 
 RUN pip install jupyterlab notebook ipywidgets && \
     jupyter lab clean
-RUN apt install --yes --no-install-recommends openssh-server tmux && \
+RUN apt install --yes --no-install-recommends openssh-server tmux iproute2 nvtop && \
     mkdir -p ~/.ssh && \
     chmod 700 ~/.ssh && \
     printf "\n[[ -z \"\$TMUX\"  ]] && { tmux attach-session -t ssh_tmux || tmux new-session -s ssh_tmux; exit; }\n" >> ~/.bashrc && \

docs/amd_hpc.qmd

@@ -1,5 +1,5 @@
 ---
-title: Training with AMD GPUs on HPC Systems
+title: AMD GPUs on HPC Systems
 description: A comprehensive guide for using Axolotl on distributed systems with AMD GPUs
 ---
 

docs/cli.qmd

@@ -1,28 +1,19 @@
-# Axolotl CLI Documentation
+---
+title: "CLI Reference"
+format:
+  html:
+    toc: true
+    toc-expand: 2
+    toc-depth: 3
+execute:
+  enabled: false
+---
 
 The Axolotl CLI provides a streamlined interface for training and fine-tuning large language models. This guide covers
 the CLI commands, their usage, and common examples.
 
-### Table of Contents
 
-- Basic Commands
-- Command Reference
-  - fetch
-  - preprocess
-  - train
-  - inference
-  - merge-lora
-  - merge-sharded-fsdp-weights
-  - evaluate
-  - lm-eval
-- Legacy CLI Usage
-- Remote Compute with Modal Cloud
-  - Cloud Configuration
-  - Running on Modal Cloud
-  - Cloud Configuration Options
-
-
-### Basic Commands
+## Basic Commands
 
 All Axolotl commands follow this general structure:
 
@@ -32,9 +23,9 @@ axolotl <command> [config.yml] [options]
 
 The config file can be local or a URL to a raw YAML file.
 
-### Command Reference
+## Command Reference
 
-#### fetch
+### fetch
 
 Downloads example configurations and deepspeed configs to your local machine.
 
@@ -49,7 +40,7 @@ axolotl fetch deepspeed_configs
 axolotl fetch examples --dest path/to/folder
 ```
 
-#### preprocess
+### preprocess
 
 Preprocesses and tokenizes your dataset before training. This is recommended for large datasets.
 
@@ -74,7 +65,7 @@ dataset_prepared_path: Local folder for saving preprocessed data
 push_dataset_to_hub: HuggingFace repo to push preprocessed data (optional)
 ```
 
-#### train
+### train
 
 Trains or fine-tunes a model using the configuration specified in your YAML file.
 
@@ -95,7 +86,38 @@ axolotl train config.yml --no-accelerate
 axolotl train config.yml --resume-from-checkpoint path/to/checkpoint
 ```
 
-#### inference
+It is possible to run sweeps over multiple hyperparameters by passing in a sweeps config.
+
+```bash
+# Basic training with sweeps
+axolotl train config.yml --sweep path/to/sweep.yaml
+```
+
+Example sweep config:
+```yaml
+_:
+  # This section is for dependent variables we need to fix
+  - load_in_8bit: false
+    load_in_4bit: false
+    adapter: lora
+  - load_in_8bit: true
+    load_in_4bit: false
+    adapter: lora
+
+# These are independent variables
+learning_rate: [0.0003, 0.0006]
+lora_r:
+  - 16
+  - 32
+lora_alpha:
+  - 16
+  - 32
+  - 64
+```
+
+
+
+### inference
 
 Runs inference using your trained model in either CLI or Gradio interface mode.
 
@@ -115,7 +137,7 @@ cat prompt.txt | axolotl inference config.yml \
     --base-model="./completed-model"
 ```
 
-#### merge-lora
+### merge-lora
 
 Merges trained LoRA adapters into the base model.
 
@@ -137,7 +159,7 @@ gpu_memory_limit: Limit GPU memory usage
 lora_on_cpu: Load LoRA weights on CPU
 ```
 
-#### merge-sharded-fsdp-weights
+### merge-sharded-fsdp-weights
 
 Merges sharded FSDP model checkpoints into a single combined checkpoint.
 
@@ -146,7 +168,7 @@ Merges sharded FSDP model checkpoints into a single combined checkpoint.
 axolotl merge-sharded-fsdp-weights config.yml
 ```
 
-#### evaluate
+### evaluate
 
 Evaluates a model's performance using metrics specified in the config.
 
@@ -155,27 +177,27 @@ Evaluates a model's performance using metrics specified in the config.
 axolotl evaluate config.yml
 ```
 
-#### lm-eval
+### lm-eval
 
 Runs LM Evaluation Harness on your model.
 
 ```bash
 # Basic evaluation
 axolotl lm-eval config.yml
-
-# Evaluate specific tasks
-axolotl lm-eval config.yml --tasks arc_challenge,hellaswag
 ```
 
 Configuration options:
 
 ```yaml
-lm_eval_tasks: List of tasks to evaluate
-lm_eval_batch_size: Batch size for evaluation
-output_dir: Directory to save evaluation results
+# List of tasks to evaluate
+lm_eval_tasks:
+  - arc_challenge
+  - hellaswag
+lm_eval_batch_size: # Batch size for evaluation
+output_dir: # Directory to save evaluation results
 ```
 
-### Legacy CLI Usage
+## Legacy CLI Usage
 
 While the new Click-based CLI is preferred, Axolotl still supports the legacy module-based CLI:
 
@@ -195,12 +217,18 @@ accelerate launch -m axolotl.cli.inference config.yml \
     --lora_model_dir="./outputs/lora-out" --gradio
 ```
 
-### Remote Compute with Modal Cloud
+::: {.callout-important}
+When overriding CLI parameters in the legacy CLI, use same notation as in yaml file (e.g., `--lora_model_dir`).
+
+**Note:** This differs from the new Click-based CLI, which uses dash notation (e.g., `--lora-model-dir`). Keep this in mind if you're referencing newer documentation or switching between CLI versions.
+:::
+
+## Remote Compute with Modal Cloud
 
 Axolotl supports running training and inference workloads on Modal cloud infrastructure. This is configured using a
 cloud YAML file alongside your regular Axolotl config.
 
-#### Cloud Configuration
+### Cloud Configuration
 
 Create a cloud config YAML with your Modal settings:
 
@@ -215,13 +243,17 @@ branch: main    # Git branch to use (optional)
 volumes:        # Persistent storage volumes
   - name: axolotl-cache
     mount: /workspace/cache
+  - name: axolotl-data
+    mount: /workspace/data
+  - name: axolotl-artifacts
+    mount: /workspace/artifacts
 
 env:            # Environment variables
   - WANDB_API_KEY
   - HF_TOKEN
 ```
 
-#### Running on Modal Cloud
+### Running on Modal Cloud
 
 Commands that support the --cloud flag:
 
@@ -239,18 +271,18 @@ axolotl train config.yml --cloud cloud_config.yml --no-accelerate
 axolotl lm-eval config.yml --cloud cloud_config.yml
 ```
 
-#### Cloud Configuration Options
+### Cloud Configuration Options
 
 ```yaml
-provider: compute provider, currently only `modal` is supported
-gpu: GPU type to use
-gpu_count: Number of GPUs (default: 1)
-memory: RAM in GB (default: 128)
-timeout: Maximum runtime in seconds
-timeout_preprocess: Preprocessing timeout
-branch: Git branch to use
-docker_tag: Custom Docker image tag
-volumes: List of persistent storage volumes
-env: Environment variables to pass
-secrets: Secrets to inject
+provider: # compute provider, currently only `modal` is supported
+gpu: # GPU type to use
+gpu_count: # Number of GPUs (default: 1)
+memory: # RAM in GB (default: 128)
+timeout: # Maximum runtime in seconds
+timeout_preprocess: # Preprocessing timeout
+branch: # Git branch to use
+docker_tag: # Custom Docker image tag
+volumes: # List of persistent storage volumes
+env: # Environment variables to pass
+secrets: # Secrets to inject
 ```

docs/config.qmd

@@ -1,5 +1,5 @@
 ---
-title: Config options
+title: Config Reference
 description: A complete list of all configuration options.
 ---
 
@@ -30,6 +30,8 @@ tokenizer_legacy:
 # Resize the model embeddings when new tokens are added to multiples of 32
 # This is reported to improve training speed on some models
 resize_token_embeddings_to_32x:
+# Optional[bool] Whether to shrink the embeddings to len(tokenizer). By default, we won't shrink.
+shrink_embeddings:
 
 # (Internal use only)
 # Used to identify which the model is based on
@@ -46,6 +48,10 @@ overrides_of_model_config:
     type: # linear | dynamic
     factor: # float
 
+# optional overrides the base model loading from_pretrained
+overrides_of_model_kwargs:
+  # use_cache: False
+
 # optional overrides to the bnb 4bit quantization configuration
 # https://huggingface.co/docs/transformers/main/main_classes/quantization#transformers.BitsAndBytesConfig
 bnb_config_kwargs:
@@ -79,6 +85,12 @@ gpu_memory_limit: 20GiB
 # Do the LoRA/PEFT loading on CPU -- this is required if the base model is so large it takes up most or all of the available GPU VRAM, e.g. during a model and LoRA merge
 lora_on_cpu: true
 
+# List[str]. Add plugins to extend the pipeline.
+# See `src/axolotl/integrations` for the available plugins or doc below for more details.
+# https://axolotl-ai-cloud.github.io/axolotl/docs/custom_integrations.html
+plugins:
+  # - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
 # A list of one or more datasets to finetune the model with
 datasets:
   # HuggingFace dataset repo | s3://,gs:// path | "json" for local dataset, make sure to fill data_files
@@ -87,7 +99,12 @@ datasets:
     type: alpaca # format | format:<prompt_style> (chat/instruct) | <prompt_strategies>.load_<load_fn>
     ds_type: # Optional[str] (json|arrow|parquet|text|csv) defines the datatype when path is a file
     data_files: # Optional[str] path to source data files
-    shards: # Optional[int] number of shards to split data into
+
+    shards: # Optional[int] split dataset into N pieces (use with shards_idx)
+    shards_idx: # Optional[int] = 0 the index of sharded dataset to use
+
+    preprocess_shards: # Optional[int] process dataset in N sequential chunks for memory efficiency (exclusive with `shards`)
+
     name: # Optional[str] name of dataset configuration to load
     train_on_split: train # Optional[str] name of dataset split to load from
     revision: # Optional[str] The specific revision of the dataset to use when loading from the Hugging Face Hub. This can be a commit hash, tag, or branch name. If not specified, the latest version will be used. This parameter is ignored for local datasets.
@@ -133,10 +150,17 @@ datasets:
 
     # Key containing the messages (default: "messages")
     field_messages: messages
-    # Key for role in each message (default: "role")
-    message_field_role: role
-    # Key for content in each message (default:  "content")
-    message_field_content: content
+
+    # Mapping of properties from the input dataset to the chat template.
+    # (default: message_property_mappings={'role':'role', 'content':'content'})
+    # If a property exists in the template but not in this mapping, the system will attempt
+    # to load it directly from the message using the property name as the key.
+    # Example: In the mapping below, 'from' is loaded from input dataset and used as 'role',
+    # while 'value' is loaded and used as 'content' in the chat template.
+    message_property_mappings:
+      role: from
+      content: value
+      # ...
 
     # Optional[Dict[str, List]]. Roles mapping in the messages. The default is:
     roles:
@@ -145,10 +169,16 @@ datasets:
       system: ["system"]
       tool: ["tool"]
 
+    # Optional[bool]. Whether to drop the system turn from the dataset. Only works with chat_template.
+    # This does not drop the default system message from chat_template if it exists. If you wish to,
+    # we recommend using a custom jinja template with the default system message removed or
+    # adding a system turn with empty content.
+    drop_system_message:
+
     # IMPORTANT: The following fields determine which parts of the conversation to train on.
     # Priority order: message_field_training > message_field_training_detail > train_on_inputs or role in roles_to_train
     # See examples at `docs/dataset-formats/conversation.qmd`
-    # Note: If the below 4 fields are empty, defaults to training only on the last message.
+    # Note: If the below 4 fields are set to empty, defaults to training only on the last message.
 
     # Optional[List[str]]. Roles to train on. The tokens from these roles will be considered for the loss.
     roles_to_train: ["assistant"]  # default
@@ -156,6 +186,7 @@ datasets:
     # - all: train on all EOS tokens
     # - turn (default): train on the EOS token at the end of each trainable turn
     # - last: train on the last EOS token in the conversation
+    # TIP: Please make sure that your `tokenizer.eos_token` is same as EOS/EOT token in template. Otherwise, set `eos_token` under `special_tokens`.
     train_on_eos: last
     # The key in the message turn that indicates via boolean whether tokens of a turn should be considered for training. Useful to selectively train on certain turns besides the `roles_to_train`.
     message_field_training: training
@@ -182,10 +213,46 @@ test_datasets:
     data_files:
       - /workspace/data/eval.jsonl
 
-# use RL training: 'dpo', 'ipo', 'kto'
+# use RL training: 'dpo', 'ipo', 'kto', 'simpo', 'orpo', 'grpo'
 rl:
-# whether to perform weighting if doing DPO training. Boolean.
-dpo_use_weighting:
+rl_beta:  # Optional[float]. The beta parameter for the RL training.
+
+# dpo
+dpo_use_weighting:  # Optional[bool]. Whether to perform weighting.
+rpo_alpha: # Optional[float]. Weighting of NLL term in loss from RPO paper.
+
+# orpo
+orpo_alpha: 0.1  # Parameter controlling the relative ratio loss weight in the ORPO loss. Passed to `beta` in `ORPOConfig` due to trl mapping.
+
+# kto
+kto_desirable_weight: # Optional[float]. Factor for desirable loss term in KTO loss.
+kto_undesirable_weight: # Optional[float]. Factor for undesirable loss term in KTO loss.
+
+# simpo
+cpo_alpha: 1.0  # Weight of the BC regularizer
+simpo_gamma: 0.5  # Target reward margin for the SimPO loss
+
+# grpo
+trl:
+  use_vllm: # Optional[bool]. Whether to use VLLM for RL training.
+  vllm_device: # Optional[str]. Device to use for VLLM.
+  vllm_gpu_memory_utilization: # Optional[float]. GPU memory utilization for VLLM.
+  vllm_max_model_len: # Optional[int]. Maximum length of the model for VLLM.
+  vllm_dtype: # Optional[str]. Data type for VLLM.
+
+  beta: # Optional[float]. Beta parameter for the RL training. Same as `rl_beta`. Use
+  max_completion_length: # Optional[int]. Maximum length of the completion for RL training.
+
+  reward_funcs: # Optional[list[str]]. List of reward functions to load. Paths must be importable from current dir.
+  reward_weights: # Optional[list[float]]. List of reward weights for the reward functions.
+
+  num_generations: # Optional[int]. Number of generations to sample.
+  log_completions: # Optional[bool]. Whether to log completions.
+
+  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.
+
 
 # reward modelling: `True` or `False`
 reward_model:
@@ -203,13 +270,13 @@ process_reward_model:
 chat_template: tokenizer_default
 # custom jinja template for chat template. This will be only used if chat_template is set to `jinja` or `null` (in which case chat_template is automatically set to `jinja`). Default is null.
 chat_template_jinja: null
-# Changes the default system message
-default_system_message: You are a helpful assistant. Please give a long and detailed answer. # Currently only supports chatml.
+# Changes the default system message. Currently only supports chatml.
+default_system_message: You are a helpful assistant. Please give a long and detailed answer.
 # Axolotl attempts to save the dataset as an arrow after packing the data together so
 # subsequent training attempts load faster, relative path
 dataset_prepared_path: data/last_run_prepared
 # Push prepared dataset to hub
-push_dataset_to_hub: # repo path
+push_dataset_to_hub: # Optional[str] repo_org/repo_name
 # The maximum number of processes to use while preprocessing your input dataset. This defaults to `os.cpu_count()`
 # if not set.
 dataset_processes: # defaults to os.cpu_count() if not set
@@ -296,6 +363,13 @@ lora_modules_to_save:
 
 lora_fan_in_fan_out: false
 
+# Apply custom LoRA autograd functions and activation function Triton kernels for
+# speed and memory savings
+# See: https://axolotl-ai-cloud.github.io/axolotl/docs/lora_optims.html
+lora_mlp_kernel: true
+lora_qkv_kernel: true
+lora_o_kernel: true
+
 # LoRA+ hyperparameters
 # For more details about the following options, see:
 # https://arxiv.org/abs/2402.12354  and `src/axolotl/core/train_builder.py`
@@ -344,6 +418,9 @@ comet_mode: # Create a new experiment ("create") or log to an existing one ("get
 comet_online: # Set to True to log data to Comet server, or False for offline storage. Default is True.
 comet_experiment_config: # Dictionary for additional configuration settings, see the doc for more details.
 
+# Tensorboard
+use_tensorboard: # Optional[bool]
+
 # Where to save the full-finetuned model to
 output_dir: ./completed-model
 
@@ -378,6 +455,12 @@ save_total_limit: # Checkpoints saved at a time
 # e.g., when 1 epoch is 1000 steps => `num_epochs: 2` and `max_steps: 100` will train for 100 steps
 max_steps:
 
+# bool of whether to include tokens trainer per second in the training metrics. This iterates over the entire dataset once, so it takes some time.
+include_tokens_per_second: # Optional[bool]
+
+# whether to find batch size that fits in memory. Passed to underlying transformers Trainer
+auto_find_batch_size: # Optional[bool]
+
 eval_table_size: # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0
 eval_max_new_tokens: # Total number of tokens generated for predictions sent to wandb. Default is 128
 eval_causal_lm_metrics: # HF evaluate metrics used during evaluation. Default is ["sacrebleu", "comet", "ter", "chrf", "perplexity"]
@@ -410,7 +493,7 @@ gradient_checkpointing: false
 early_stopping_patience: 3
 
 # Specify a scheduler and kwargs to use with the optimizer
-lr_scheduler: # 'one_cycle' | 'log_sweep' | empty for cosine
+lr_scheduler: # 'one_cycle' | 'rex' | 'log_sweep' | empty for cosine
 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)
@@ -493,6 +576,8 @@ flash_attn_fuse_mlp: # Whether to fuse part of the MLP into a single operation
 sdp_attention:
 # Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf
 s2_attention:
+# Optional[bool]. Whether to use low_cpu_mem_usage
+low_cpu_mem_usage:
 # Resume from a specific checkpoint dir
 resume_from_checkpoint:
 # If resume_from_checkpoint isn't set and you simply want it to start where it left off.
@@ -513,6 +598,13 @@ special_tokens:
 # Add extra tokens.
 tokens:
 
+# Mapping token_id to new_token_string to override reserved added_tokens in the tokenizer.
+# Only works for tokens that are not part of the base vocab (aka are added_tokens).
+# Can be checked if they exist in tokenizer.json added_tokens.
+added_tokens_overrides:  # Dict[int, str]
+#  128041: "<|im_start|>"
+#  128042: "<|im_end|>"
+
 # FSDP
 fsdp:
 fsdp_config:

docs/custom_integrations.qmd

@@ -0,0 +1,101 @@
+---
+title: Custom Integrations
+toc: true
+toc-depth: 3
+---
+
+```{python}
+#| echo: false
+
+import re
+
+def process_readme(integration_name):
+    try:
+        path = f'../src/axolotl/integrations/{integration_name}/README.md'
+        with open(path, 'r') as f:
+            txt = f.read()
+            # Remove h1 headings
+            txt = re.sub(r'^# .*\n?', '', txt, flags=re.MULTILINE)
+            # Convert h2 to h3
+            txt = re.sub(r'^## ', '### ', txt, flags=re.MULTILINE)
+            return txt
+    except FileNotFoundError:
+        return None
+
+def print_section(name, folder_name):
+    output = f"\n## {name}\n"
+    content = process_readme(folder_name)
+    if content:
+        output += content
+    output += f"\nPlease see reference [here](https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/integrations/{folder_name})\n"
+    return output
+```
+
+```{python}
+#| output: asis
+#| echo: false
+
+# Introduction text
+print("""
+Axolotl adds custom features through `integrations`. They are located within the `src/axolotl/integrations` directory.
+
+To enable them, please check the respective documentations.
+""")
+
+# Sections
+sections = [
+    ("Cut Cross Entropy", "cut_cross_entropy"),
+    ("Grokfast", "grokfast"),
+    ("Knowledge Distillation (KD)", "kd"),
+    ("Liger Kernels", "liger"),
+    ("Language Model Evaluation Harness (LM Eval)", "lm_eval"),
+    ("Spectrum", "spectrum")
+]
+
+for section_name, folder_name in sections:
+    print(print_section(section_name, folder_name))
+```
+
+## Adding a new integration
+
+Plugins can be used to customize the behavior of the training pipeline through [hooks](https://en.wikipedia.org/wiki/Hooking). See [`axolotl.integrations.BasePlugin`](https://github.com/axolotl-ai-cloud/axolotl/blob/main/src/axolotl/integrations/base.py) for the possible hooks.
+
+To add a new integration, please follow these steps:
+
+1. Create a new folder in the `src/axolotl/integrations` directory.
+2. Add any relevant files (`LICENSE`, `README.md`, `ACKNOWLEDGEMENTS.md`, etc.) to the new folder.
+3. Add `__init__.py` and `args.py` files to the new folder.
+  - `__init__.py` should import the integration and hook into the appropriate functions.
+  - `args.py` should define the arguments for the integration.
+4. (If applicable) Add CPU tests under `tests/integrations` or GPU tests under `tests/e2e/integrations`.
+
+::: {.callout-tip}
+
+See [src/axolotl/integrations/cut_cross_entropy](https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/integrations/cut_cross_entropy) for a minimal integration example.
+
+:::
+
+::: {.callout-warning}
+
+If you could not load your integration, please ensure you are pip installing in editable mode.
+
+```bash
+pip install -e .
+```
+
+and correctly spelled the integration name in the config file.
+
+```yaml
+plugins:
+  - axolotl.integrations.your_integration_name.YourIntegrationPlugin
+```
+
+:::
+
+::: {.callout-note}
+
+It is not necessary to place your integration in the `integrations` folder. It can be in any location, so long as it's installed in a package in your python env.
+
+See this repo for an example: [https://github.com/axolotl-ai-cloud/diff-transformer](https://github.com/axolotl-ai-cloud/diff-transformer)
+
+:::

docs/dataset-formats/conversation.qmd

@@ -6,7 +6,9 @@ order: 3
 
 ## sharegpt
 
-IMPORTANT: ShareGPT is deprecated!. Please see `chat_template` section below.
+::: {.callout-important}
+ShareGPT is deprecated!. Please see [chat_template](#chat_template) section below.
+:::
 
 ## pygmalion
 
@@ -22,7 +24,7 @@ Chat Template strategy uses a jinja2 template that converts a list of messages i
 {"conversations": [{"role": "...", "content": "..."}]}
 ```
 
-See `config.qmd` for full configs and supported templates.
+See [configs](../config.qmd) for full configs and supported templates.
 
 ### Migrating from sharegpt
 
@@ -42,8 +44,9 @@ datasets:
     type: chat_template
 
     field_messages: conversations
-    message_field_role: from
-    message_field_content: value
+    message_property_mappings:
+      role: from
+      content: value
 
 # new (if setting a new chat_template like chatml, gemma, etc)
 chat_template: chatml
@@ -52,8 +55,9 @@ datasets:
     type: chat_template
 
     field_messages: conversations
-    message_field_role: from
-    message_field_content: value
+    message_property_mappings:
+      role: from
+      content: value
 ```
 
 We recommend checking the below examples for other usecases.
@@ -70,6 +74,10 @@ datasets:
     train_on_eos:
 ```
 
+::: {.callout-tip}
+If you receive an error like "`chat_template` choice is `tokenizer_default` but tokenizer's `chat_template` is null.", it means the tokenizer does not have a default `chat_template`. Follow the examples below instead to set a custom `chat_template`.
+:::
+
 2. Using the `gemma` chat template to override the tokenizer_config.json's chat template on OpenAI messages format, training on all assistant messages.
 
 ```yaml
@@ -100,6 +108,10 @@ datasets:
     type: chat_template
 ```
 
+::: {.callout-important}
+Please make sure that your `tokenizer.eos_token` is same as EOS/EOT token in template. Otherwise, set `eos_token` under `special_tokens`.
+:::
+
 5. (Advanced) Using fine-grained control over tokens and turns to train in a conversation
 
 For a data sample that looks like:
@@ -138,12 +150,15 @@ datasets:
     type: chat_template
     chat_template: tokenizer_default
     field_messages: conversations
-    message_field_role: from
-    message_field_content: value
+    message_property_mappings:
+      role: from
+      content: value
     roles_to_train: []
     train_on_eos: turn
     message_field_training: train
     message_field_training_detail: train_detail
 ```
 
-Tip: It is not necessary to use both `message_field_training` and `message_field_training_detail` at a time.
+::: {.callout-tip}
+It is not necessary to set both `message_field_training` and `message_field_training_detail` at once.
+:::

docs/dataset-formats/index.qmd

@@ -1,14 +1,492 @@
 ---
 title: Dataset Formats
-description: Supported dataset formats.
-listing:
-  fields: [title, description]
-  type: table
-  sort-ui: false
-  filter-ui: false
-  max-description-length: 250
+description: Guide to Dataset Formats in Axolotl
+back-to-top-navigation: true
+toc: true
+toc-depth: 5
 ---
 
-Axolotl supports a variety of dataset formats.  It is recommended to use a JSONL format.  The schema of the JSONL depends upon the task and the prompt template you wish to use. Instead of a JSONL, you can also use a HuggingFace dataset with columns for each JSONL field.
 
-Below are these various formats organized by task:
+Axolotl is a training framework that aims to make the process convenient yet flexible to users by simply passing a config yaml file.
+
+As there are a lot of available options in Axolotl, this guide aims to provide an simplify the user experience to choosing the proper choice.
+
+Axolotl supports 3 kinds of training methods: pre-training, supervised fine-tuning, and preference-based post-training (e.g. DPO, ORPO, PRMs). Each method has their own dataset format which are described below.
+
+## Pre-training
+
+When aiming to train on large corpora of text datasets, pre-training is your go-to choice. Due to the size of these datasets, downloading the entire-datasets before beginning training would be prohibitively time-consuming. Axolotl supports [streaming](https://huggingface.co/docs/datasets/en/stream) to only load batches into memory at a time.
+
+A sample format for a pre-training dataset is as follows:
+
+```json
+{"text": "first row"}
+{"text": "second row"}
+...
+```
+
+It is typically recommended to save your dataset as `.jsonl` due to its flexibility and simplicity.
+
+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:
+
+```yaml
+pretraining_dataset: hf_org/name
+```
+
+### Pre-training from local dataset files
+
+Given a few corpus files: `A.jsonl`, `B.jsonl`, and `C.jsonl`, your config will look like the below:
+
+```yaml
+pretraining_dataset:
+  - path: json
+    data_files:
+      - A.jsonl
+      - B.jsonl
+      - C.jsonl
+```
+
+While we recommend `.jsonl`, you can also use the other formats (`csv`, `parquet`, `arrow`, `SQL`, `Webdataset`) that are supported by [`Dataset.load_dataset`](https://huggingface.co/docs/datasets/loading#local-and-remote-files)
+
+### Pre-training without streaming
+
+On the rare case that the dataset is small and can be loaded entirely into memory, another approach to running pre-training is to use the `completion` format. This would mean that the entire dataset is pre-tokenized instead of on-demand in streaming.
+
+One benefit of this is that the tokenization can be performed separately on a CPU-only machine, and then transferred to a GPU machine for training to save costs.
+
+From Hugging Face:
+
+```yaml
+datasets:
+  - path: hf_org/name
+    type: completion
+```
+
+From local files (either example works):
+
+```yaml
+datasets:
+  - path: A.jsonl
+    type: completion
+
+  - path: json
+    data_files: ["A.jsonl", "B.jsonl", "C.jsonl"]
+    type: completion
+```
+
+### Pre-training dataset configuration tips
+
+#### Setting max_steps
+
+When using streaming for large datasets, Axolotl does not know in advance how large the dataset is and does not know when to stop.
+
+Therefore, it is necessary to set `max_steps: int` in your config for pre-training to run, so that Axolotl knows when to stop training.
+
+One step is equal to `sequence_len * micro_batch_size * gradient_accumulation_steps * total_num_gpus` tokens.
+
+#### Group_by_length
+
+It is recommended to leave this off if downloading from Hugging Face hub as it would download the entire dataset which can be very large.
+
+### Reference
+
+Please see docs [here](pretraining.qmd).
+
+## Supervised fine-tuning (SFT)
+
+Supervised fine-tuning is the process of training models to respond to an instruction or chat input.
+
+As there are a wide variety of dataset formats, Axolotl tries to support a majority of the formats available in public datasets.
+
+Axolotl provides four approaches for loading datasets, however, it's easier to work backwards from the dataset you have available to figure out which approach to use.
+
+A flow chart is as follows:
+
+1. Do you already have the dataset tokenized? If yes, check [Pre-Tokenized Dataset](#pre-tokenized-dataset).
+
+2. Do you want to format the dataset yourself and manually choose each section to mask? If yes, check [Template Free Dataset](#template-free-dataset)
+
+3. Is your dataset in a "conversation" format, containing a `list[messages]`? If yes, check [Conversation Dataset](#conversation-dataset)
+
+4. Is your dataset in an "instruct" format, containing `{ instruction, response }`? If yes, check [Instruction Dataset](#instruction-dataset)
+
+If you went through the flow chart and did not find one that matches, it is recommended to preprocess your dataset into one of the above or create a thread on Github Discussion.
+
+::: {.callout-tip}
+You can mix and match within each approach or across approaches to train a model on a variety of datasets.
+:::
+
+### Pre-Tokenized Dataset
+
+We suggest this approach when you want to bring your own tokenized dataset.
+
+Axolotl expects the dataset to have three keys:
+
+- `input_ids`: from tokenizing formatted prompt
+- `attention_mask`: for masking padding. If you don't add padding, it would be equal to `len(input_ids) * [1]`
+- `labels`: this is the same as `input_ids`, however, if you want to mask certain tokens, you would set those indices to `-100`.
+
+::: {.callout-tip}
+Make sure to add BOS/EOS tokens to your prompt and mask it appropriately.
+:::
+
+A config for this would look like:
+
+```yaml
+datasets:
+  - path: A.jsonl
+    type:
+```
+
+::: {.callout-note}
+`type: ` is empty!
+:::
+
+Reference: [Pre-Tokenized Dataset Documentation](tokenized.qmd).
+
+### Template Free Dataset
+
+We reccomend this approach when you want granular control over the prompt formatting, special tokens, and masking, whilst letting Axolotl handle the tokenization. This is very useful if your dataset has unique prompts that differ across samples and where one single general template wouldn't suffice.
+
+In the example below, you could see that there is no proper structure. At the same time, it's very flexible as there are no constraints on how your prompt can look.
+
+```json
+{
+    "segments": [
+        {
+            "label": true,
+            "text": "<s>Hello\n"
+        },
+        {
+            "label": true,
+            "text": "hi there!. "
+        },
+        {
+            "label": false,
+            "text": "goodbye "
+        },
+        {
+            "label": true,
+            "text": "farewell</s>"
+        }
+    ]
+}
+```
+
+Each prompt must be have a key called `segments` which is a list of `{ text, label }`.
+
+```yaml
+datasets:
+  - path: A.jsonl
+    type: input_output
+```
+
+Reference: [Template Free Documentation](template_free.qmd).
+
+### Conversation Dataset
+
+`conversation` messages are a list of messages which usually contain a `role` and `content` key.
+
+::: {.callout-tip}
+Fun fact: Axolotl synonymously refers to "chat" messages as `conversation` messages due to how FastChat initially used this term to build a widely used [fastchat conversation](https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py) method for formatting chat messages prior to the creation of `chat_templates`.
+:::
+
+#### What are `chat_templates`?
+
+The current most popular and convenient method for inference is to use `chat_templates` for formatting prompts. Axolotl supports using `chat_templates` for training to ensure that the model performs in the same environment as in inference.
+
+Here's a quick rundown on `chat_template`: A `chat_template` is a Jinja2 template which formats a list of messages into a prompt.
+
+An example of a prompt formatted into a popular template called ChatML can be seen below:
+
+Single prompt (pretty-printed):
+```json
+{
+    "messages": [
+        {
+            "role": "user",
+            "content": "Hi"
+        },
+        {
+            "role": "assistant",
+            "content": "How can I help you?"
+        },
+        {
+            "role": "user",
+            "content": "Can you add 3+5?"
+        },
+        {
+            "role": "assistant",
+            "content": "The answer is 8."
+        }
+    ]
+}
+```
+
+The ChatML template is as follows:
+```jinja2
+{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}
+```
+
+The above prompt formatted into this template will result in:
+
+```
+<|im_start|>user
+Hi<|im_end|>
+<|im_start|>assistant
+How can I help you?<|im_end|>
+<|im_start|>user
+Can you add 3+5?<|im_end|>
+<|im_start|>assistant
+The answer is 8.<|im_end|>
+```
+
+By using delimiters (`<|im_start|>` and `<|im_end|>`), a prompt separates different speakers which helps the model identify which portion belongs to whom.
+
+#### Common Conversation Dataset formats
+
+Older conversation datasets with the following format are colloquially called `sharegpt` datasets.
+
+```json
+{"conversations": [{"from": "...", "value": "..."}]}
+```
+
+Newer conversation datasets usually follow the OpenAI format.
+
+```json
+{"messages": [{"role": "...", "content": "..."}]}
+```
+
+Axolotl supports both as well as allowing customization of any kind of key.
+
+#### Chat Template Usage
+
+To properly use this method, it is important to identify three things:
+
+1. Which `chat_template` would you use?
+
+2. What are the keys in your dataset, and what are the possible roles? For example, in OpenAI format, the keys would be `messages`, `role`, and `content`, respectively, whereas the possible roles are `system`, `user`, and `assistant`.
+
+3. What do you want to mask? For instance, only assistant messages, only last message, or nothing.
+
+##### Choosing a `chat_template`
+
+There are a lot of `chat_templates` out there. Axolotl supports the common ones: [supported chat templates](https://github.com/axolotl-ai-cloud/axolotl/blob/860609392184cf62a7e0ca676658b170e059ce6c/src/axolotl/utils/chat_templates.py#L17). For example, to use ChatML, it would be `chat_template: chatml`.
+
+However, it is also possible to use the already configured template within the tokenizer by specifying `chat_template: tokenizer_default`. If you want a fallback (in case some tokenizer does not have it pre-configured), you can do `chat_template: tokenizer_default_fallback_chatml` to fallback to the ChatML template if a tokenizer template was not found.
+
+One last but powerful approach is to bring your own template. This can be set via:
+
+```yaml
+chat_template_jinja: # your template
+```
+
+##### Setting `chat_template` dataset keys
+
+We currently default to OpenAI format for dataset keys, so if that's your current dataset format, there's nothing to do here.
+
+If your dataset format is different, here are the keys you should check (with their defaults):
+
+```yaml
+datasets:
+    ...
+    field_messages: messages  # this should point to the key containing the list of conversations
+    message_property_mappings:  # this is a mapping from keys in your dataset to keys in chat_template
+      role: role
+      content: content
+```
+
+In some `chat_templates` (e.g. [Gemma](https://huggingface.co/google/gemma-2b-it/blob/main/tokenizer_config.json#L1507)), the roles are hardcoded to `user` and `assistant`. Consequently, you may find it necessary to map the roles in your dataset to these above. We currently have some defaults that should work for common datasets, but if you get a `KeyError`, it would be necessary to add mapping for your roles. Here is an example of how it would look like:
+
+```yaml
+datasets:
+    ...
+    roles:
+      assistant:
+        - gpt
+        - model
+      user:
+        - human
+```
+
+In the example above, all `gpt` and `model` values are converted to `assistant`. All `human` values are converted to `user.`
+
+##### Handling masking
+
+The common use case for `chat_template` is for chat messages, therefore, it is common to mask all non-assistant messages. Assistant messages refer to the bot messages that you want the model to learn on.
+
+To train on all `assistant` messages, you would set the following configs.
+
+```yaml
+datasets:
+    ...
+    roles_to_train: ["assistant"]
+    train_on_eos: "turn"
+```
+
+The `train_on_eos` config means that it would mask all EOS tokens for turns that aren't assistant-turns. The other options are: `all` and `last` to choose which EOS to train on.
+
+Perhaps, you want to train on `assistant` and `narrator` roles, you can simply add `narrator` to the list of `roles_to_train`. You would also need to add it to the mapping of `roles` above.
+
+```yaml
+datasets:
+    ...
+    roles_to_train: ["assistant", "narrator"]
+    roles:
+      assistant:
+        - gpt
+        - model
+      user:
+        - human
+      narrator: ["narrator"]
+```
+
+::: {.callout-tip}
+As chat_templates may use hardcoded EOS/EOT tokens that are different from the tokenizer's EOS, it is highly recommended to set them. For example, `ChatML` uses `<|im_end|>` to end turns.
+
+```yaml
+special_tokens:
+  eos_token: <|im_end|>
+```
+
+:::
+
+##### Applying `chat_template`
+
+Once all the above steps are completed, you could combine all these configs together to form a bespoke configuration for your custom dataset.
+
+```yaml
+datasets:
+  - path: A.jsonl
+    type: chat_template
+
+    # step 1
+    chat_template: chatml
+
+    # step 2
+    field_messages: messages
+    message_property_mappings:
+      role: role
+      content: content
+
+    roles:
+      assistant:
+        - gpt
+        - model
+        - assistant
+      user:
+        - human
+        - user
+
+    # step 3
+    roles_to_train: ["assistant"]
+    train_on_eos: "turn"
+
+special_tokens:
+  eos_token: <|im_end|>
+```
+
+If this config were to be applied to the sample dataset above, the output would look as such (which can be retrieved via `axolotl preprocess config.yaml --debug`):
+
+```
+<|im_start|>(-100, 128256) user(-100, 882)
+(-100, 198) Hi(-100, 13347) <|im_end|>(-100, 128257)
+(-100, 198) <|im_start|>(-100, 128256) assistant(-100, 78191)
+(-100, 198) How(4438, 4438)  can(649, 649)  I(358, 358)  help(1520, 1520)  you(499, 499) ?(30, 30) <|im_end|>(128257, 128257)
+(-100, 198) <|im_start|>(-100, 128256) user(-100, 882)
+(-100, 198) Can(-100, 6854)  you(-100, 499)  add(-100, 923)  (-100, 220) 3(-100, 18) +(-100, 10) 5(-100, 20) ?(-100, 30) <|im_end|>(-100, 128257)
+(-100, 198) <|im_start|>(-100, 128256) assistant(-100, 78191)
+(-100, 198) The(791, 791)  answer(4320, 4320)  is(374, 374)  (220, 220) 8(23, 23) .(13, 13) <|im_end|>(128257, 128257)
+(-100, 198)
+```
+
+The first number refers to the label, the second refers to the `token_id`. For example, `-100` labels appear on non-assistant portions, meaning that they are masked during. For assistant portions, the label is the same as the `token_id`.
+
+::: {.callout-note}
+
+If during `preprocess`, there are a lot of warnings of `Could not find content __ boundary`, please check the FAQ section for [chat_templates](../faq.qmd#chat-templates).
+
+:::
+
+#### Reference
+
+Please see docs [here](conversation.qmd).
+
+### Instruction Dataset
+
+Instruction datasets are used to train instruction-following models and comprise a prompt, containing an instruction, and a single response. In contrast to chat datasets which may be multi-turn, instruct datasets are typically single-turn.
+
+An example is of a common format called Alpaca:
+```json
+{"instruction": "...", "input": "...", "output": "..."}
+```
+
+Using those keys, a prompt can be built based on it.
+```
+Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.
+
+### Instruction:
+{instruction}
+
+### Input:
+{input}
+
+### Response:
+{output}
+```
+
+This can be configured as such:
+```yaml
+datasets:
+  - path: A.jsonl
+    type: alpaca
+```
+
+Axolotl supports many kinds of instruction dataset. All of them can be found here (https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/inst_tune.html) with their respective type and sample row format.
+
+
+Reference: [Instruction Dataset Documentation](inst_tune.qmd).
+
+#### Custom Instruct Prompt Format
+
+Due to the myriad possibilities of instruction formats, Axolotl allows customizing your own instruction format without having to dive into the code directly.
+
+In the example below, a sample row is used to output in `mistral_v1` format.
+```json
+{"input": "...", "output": "..."}
+```
+
+```yaml
+datasets:
+  - path: repo
+    type:
+      system_prompt: ""
+
+      field_system:
+      field_instruction: input
+      field_input:
+      field_output: output
+
+      # multi-line example with input
+      format: |-
+        [INST] {instruction} {input} [/INST]
+
+      # single-line example without input
+      no_input_format: "[INST] {instruction} [/INST]"
+```
+
+The config sets that the `field_instruction` is actually named `input`, and the `field_input` is empty as we don't have an `input` in this sample. Generally, `instruction` can be thought as the question to the model, and `input` as the additional information with `output` being the response. It is not necessary to have an `input` nor `system`. In the end, the most important part is to understand what format you want it to look like and how you can customize this to your use case.
+
+Reference: [Custom Instruct Prompt Format Documentation](inst_tune.qmd#how-to-add-custom-prompt-format).
+
+## Reinforcement Learning from Human Feedback (RLHF)
+
+As there are multiple RLHF methods with their own dataset requirements. Please see [RLHF documentation](../rlhf.qmd) for more detail.

docs/dataset-formats/pretraining.qmd

@@ -27,7 +27,6 @@ pretraining_dataset:
     type: pretrain
     trust_remote_code:
     skip: # number of rows of data to skip over from the beginning
-...
 ```
 
 :::

docs/dataset-formats/template_free.qmd

@@ -1,7 +1,239 @@
 ---
 title: Template-Free
 description: Construct prompts without a template.
+toc: true
+toc-depth: 3
 order: 4
 ---
 
-See [these docs](../input_output.qmd).
+## Background {#sec-background}
+
+### Masking Inputs {#masking-inputs}
+
+One of the most popular features of
+[axolotl](https://github.com/axolotl-ai-cloud/axolotl) is
+setting the following configuration value:
+
+
+```yaml
+train_on_inputs: false
+```
+
+If you declare a [dataset formats](https://github.com/axolotl-ai-cloud/axolotl?tab=readme-ov-file#dataset)
+such as `alpaca` or `chatml`, axolotl knows what is an input
+(i.e. human) vs. an output (i.e. the assistant) and masks the input
+labels so that your model can focus on predicting the outputs only.
+
+### You may not want prompt templates {#sec-you-may-not-want-prompt-templates}
+
+However, there are many situations where you don't want to use one of
+these formats or templates. This is because they can:
+
+-   Add unnecessary boilerplate to your prompts.
+-   Create artifacts like special delimiters `<|im_start|>` that can
+    quickly become footguns if you don't include them correctly at
+    inference time.
+-   Enforce a *chat* interface when you do not want one. Sometimes you
+    just want to fine-tune a model to a very specific task and do NOT
+    want multi-turn conversations, roles, etc.
+-   Limit you to only certain roles that the template allows.
+
+### The `input_output` format {#sec-the-inputoutput-format}
+
+You can construct your prompts without a template by using the
+`input_output` format, by setting `type: input_output` in your
+configuration file like this:
+
+**config.yml**
+
+```yaml
+train_on_inputs: false # Mask segments of your data
+datasets:
+  - path: output.jsonl
+    type: input_output  # use template free prompt construction
+```
+
+Unlike `type: completion`, which is also template-free,
+`type: input_output` allows you to mask segments of your text. More
+details on how this works are described below.
+
+## Usage {#sec-usage}
+
+This is how you can use the `input_output` format:
+
+### 1. Prepare Data {#sec-1-prepare-data}
+
+To use the `input_output` format, collect your data in the following
+format into a jsonl file (below is the first row from the file
+`output`.jsonl` pretty printed):
+
+```bash
+$ head -n1 output.jsonl | python -m json.tool
+```
+
+:::{.cell-output .cell-output-stdout}
+    {
+        "segments": [
+            {
+                "label": true,
+                "text": "<s>Hello\n"
+            },
+            {
+                "label": true,
+                "text": "hi there!. "
+            },
+            {
+                "label": false,
+                "text": "goodbye "
+            },
+            {
+                "label": true,
+                "text": "farewell</s>"
+            }
+        ]
+    }
+:::
+
+Set `label:false` when you want to mask a segment of text so that the
+model isn't trained on it. Some things to keep in mind:
+
+> [!IMPORTANT]
+> 1.  **EOS, BOS, spaces, newlines etc. are entirely up to you. Axolotl
+    concatenates all the segments as-is.** The tokenizer doesn't add
+    anything additional. Notice how I added spaces, newlines, `<s>`
+    (BOS), and `</s>` (EOS) myself.
+> 2.  Make sure you check the materialized output to validate that the
+    prompt is getting assembled how you like.
+
+### 2. Use `type: input_output` {#sec-2-use-type-inputoutput}
+
+Let's materialize data with our `output.jsonl` file by setting
+`type: input_output` in our axolotl config:
+
+```yaml
+# training_config.yaml
+base_model: mistralai/Mistral-7B-v0.1
+data_seed: 49
+seed: 49
+
+datasets:
+  - path: output.jsonl
+    type: input_output
+val_set_size: 0.1
+
+sequence_len: 896
+sample_packing: false
+
+micro_batch_size: 2
+gradient_accumulation_steps: 3
+eval_batch_size: 2
+num_epochs: 1
+learning_rate: 0.0002
+
+train_on_inputs: false
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"
+```
+
+You can use the following command to materialize your data. The
+`--debug` flag will print the tokens, along with the labels so you can
+verify that the correct items are being ignored:
+
+```bash
+axolotl preprocess training_config.yaml --debug
+
+...
+[2024-03-05 23:36:46,969] [INFO] [axolotl.check_example_labels:35] [PID:607731] [RANK:0] <s>(1, 1) Hello(22557, 22557)
+(13, 13) hi(12014, 12014) there(736, 736) !(28808, 28808) .(28723, 28723) (28705, 28705) good(-100, 1179) bye(-100, 17664) (-100, 28705) fare(19111, 19111) well(5458, 5458) </s>(2, 2)
+
+```
+
+The format is `decoded_token`(`label`, `token_id`), for example,
+`<s>(1, 1)` means that the token is `<s>`, the label is `1` and the
+token_id is `1`. When the label is `-100` then that token is ignored for
+training.
+
+### 3. Check the prompts {#sec-3-check-the-prompts}
+
+Here is another way to check the materialized output:
+
+```python
+from transformers import AutoTokenizer
+from datasets import load_from_disk
+import yaml
+
+directory = !ls last_run_prepared/
+with open('training_config.yaml', 'r') as f:
+    cfg = yaml.safe_load(f)
+model_id = cfg['base_model']
+tok = AutoTokenizer.from_pretrained(model_id)
+ds = load_from_disk(f'last_run_prepared/{directory[0]}/')
+```
+
+```python
+>>> row = ds[0]
+>>> print(tok.decode(row['input_ids']))
+<s> Hello
+    hi there!.  goodbye  farewell</s>
+```
+
+We can check that the right tokens are ignored by comparing the labels
+to each token:
+
+```python
+import pandas as pd
+pd.DataFrame([{'token': tok.decode(i), 'label': l, 'id':i} for i,l in
+              zip(row['input_ids'], row['labels'])])
+```
+
+| token | label | id    |
+|-------|-------|-------|
+| 0     | \<s\> | 1     |
+| 1     | Hello | 22557 |
+| 2     | \\n   | 13    |
+| 3     | hi    | 12014 |
+| 4     | there | 736   |
+| 5     | !     | 28808 |
+| 6     | .     | 28723 |
+| 7     |       | 28705 |
+| 8     | good  | -100  |
+| 9     | bye   | -100  |
+| 10    |       | -100  |
+| 11    | fare  | 19111 |
+| 12    | well  | 5458  |
+| 13    | \</s\>| 2     |
+
+
+
+If we look at the input data, the above table seems correct! (The jsonl
+version is repeated below for reference):
+
+
+```bash
+$ head -n1 output.jsonl | python -m json.tool
+```
+
+:::{.cell-output .cell-output-stdout}
+    {
+        "segments": [
+            {
+                "label": true,
+                "text": "<s>Hello\n"
+            },
+            {
+                "label": true,
+                "text": "hi there!. "
+            },
+            {
+                "label": false,
+                "text": "goodbye "
+            },
+            {
+                "label": true,
+                "text": "farewell</s>"
+            }
+        ]
+    }
+:::

docs/dataset_preprocessing.qmd

@@ -3,8 +3,11 @@ title: Dataset Preprocessing
 description: How datasets are processed
 ---
 
+## Overview
+
 Dataset pre-processing is the step where Axolotl takes each dataset you've configured alongside
-the (dataset format)[../dataset-formats/] and prompt strategies to:
+the [dataset format](docs/dataset-formats) and prompt strategies to:
+
  - parse the dataset based on the *dataset format*
  - transform the dataset to how you would interact with the model based on the *prompt strategy*
  - tokenize the dataset based on the configured model & tokenizer
@@ -12,10 +15,12 @@ the (dataset format)[../dataset-formats/] and prompt strategies to:
 
 The processing of the datasets can happen one of two ways:
 
-1. Before kicking off training by calling `python -m axolotl.cli.preprocess /path/to/your.yaml --debug`
+1. Before kicking off training by calling `axolotl preprocess config.yaml --debug`
 2. When training is started
 
-What are the benefits of pre-processing? When training interactively or for sweeps
+### What are the benefits of pre-processing?
+
+When training interactively or for sweeps
 (e.g. you are restarting the trainer often), processing the datasets can oftentimes be frustratingly
 slow. Pre-processing will cache the tokenized/formatted datasets according to a hash of dependent
 training parameters so that it will intelligently pull from its cache when possible.
@@ -28,8 +33,12 @@ default path of `./last_run_prepared/`, but will ignore anything already cached
 setting `dataset_prepared_path: ./last_run_prepared`, the trainer will use whatever pre-processed
 data is in the cache.
 
-What are the edge cases? Let's say you are writing a custom prompt strategy or using a user-defined
+### What are the edge cases?
+
+Let's say you are writing a custom prompt strategy or using a user-defined
 prompt template. Because the trainer cannot readily detect these changes, we cannot change the
-calculated hash value for the pre-processed dataset. If you have `dataset_prepared_path: ...` set
+calculated hash value for the pre-processed dataset.
+
+If you have `dataset_prepared_path: ...` set
 and change your prompt templating logic, it may not pick up the changes you made and you will be
 training over the old prompt.

docs/debugging.qmd

@@ -31,11 +31,13 @@ While debugging it's helpful to simplify your test scenario as much as possible.
     - Set `CUDA_VISIBLE_DEVICES` to a single GPU, ex: `export CUDA_VISIBLE_DEVICES=0`.
     - Set `dataset_processes: 1` in your axolotl config or run the training command with `--dataset_processes=1`.
 2. **Use a small dataset**: Construct or use a small dataset from HF Hub. When using a small dataset, you will often have to make sure `sample_packing: False` and `eval_sample_packing: False` to avoid errors.  If you are in a pinch and don't have time to construct a small dataset but want to use from the HF Hub, you can shard the data (this will still tokenize the entire dataset, but will only use a fraction of the data for training.  For example, to shard the dataset into 20 pieces, add the following to your axolotl config):
+
     ```yaml
-    dataset:
+    datasets:
         ...
         shards: 20
     ```
+
 3. **Use a small model**: A good example of a small model is [TinyLlama/TinyLlama-1.1B-Chat-v1.0](https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0).
 4. **Minimize iteration time**: Make sure the training loop finishes as fast as possible, with these settings.
     - `micro_batch_size: 1`
@@ -85,7 +87,7 @@ The easiest way to get started is to modify the [.vscode/launch.json](../.vscode
 
 For example, to mimic the command `cd devtools && CUDA_VISIBLE_DEVICES=0 accelerate launch -m axolotl.cli.train dev_chat_template.yml`, you would use the below configuration[^1].  Note that we add additional flags that override the axolotl config and incorporate the tips above (see the comments). We also set the working directory to `devtools` and set the `env` variable `HF_HOME` to a temporary folder that is later partially deleted.  This is because we want to delete the HF dataset cache before each run in order to ensure that the data preprocessing code is run from scratch.
 
-```jsonc
+```json
 // .vscode/launch.json
 {
     "version": "0.2.0",
@@ -132,7 +134,7 @@ For example, to mimic the command `cd devtools && CUDA_VISIBLE_DEVICES=0 acceler
 
 Below is the [./vscode/tasks.json](../.vscode/tasks.json) file that defines the `cleanup-for-dataprep` task.  This task is run before each debugging session when you use the above configuration.  Note how there are two tasks that delete the two folders mentioned above.  The third task `cleanup-for-dataprep` is a composite task that combines the two tasks.  A composite task is necessary because VSCode does not allow you to specify multiple tasks in the `preLaunchTask` argument of the `launch.json` file.
 
-```jsonc
+```json
 // .vscode/tasks.json
 // this file is used by launch.json
 {

docs/docker.qmd

@@ -0,0 +1,140 @@
+---
+title: "Docker"
+format:
+  html:
+    toc: true
+    toc-depth: 4
+---
+
+This section describes the different Docker images that are released by AxolotlAI at [Docker Hub](https://hub.docker.com/u/axolotlai).
+
+## Base
+
+The base image is the most minimal image that can install Axolotl. It is based on the `nvidia/cuda` image. It includes python, torch, git, git-lfs, awscli, pydantic, and more.
+
+#### Image
+
+```
+axolotlai/axolotl-base
+```
+
+Link: [Docker Hub](https://hub.docker.com/r/axolotlai/axolotl-base)
+
+#### Tags format
+
+```bash
+main-base-py{python_version}-cu{cuda_version}-{pytorch_version}
+```
+
+Tags examples:
+
+- `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
+
+The main image is the image that is used to run Axolotl. It is based on the `axolotlai/axolotl-base` image and includes the Axolotl codebase, dependencies, and more.
+
+#### Image
+
+```
+axolotlai/axolotl
+```
+
+Link: [Docker Hub](https://hub.docker.com/r/axolotlai/axolotl)
+
+#### Tags format {#sec-main-tags}
+
+```bash
+# on push to main
+main-py{python_version}-cu{cuda_version}-{pytorch_version}
+
+# latest main (currently torch 2.5.1, python 3.11, cuda 12.4)
+main-latest
+
+# nightly build
+{branch}-{date_in_YYYYMMDD}-py{python_version}-cu{cuda_version}-{pytorch_version}
+
+# tagged release
+{version}
+```
+
+:::{.callout-tip}
+
+There may be some extra tags appended to the image, like `-vllm` which installs those packages.
+
+:::
+
+Tags examples:
+
+- `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`
+- `main-20250303-py3.11-cu124-2.4.1`
+- `0.7.1`
+
+## Cloud
+
+The cloud image is the image that is used to run Axolotl in the cloud. It is based on the `axolotlai/axolotl` image and sets ENV variables like HuggingFace cache directories for volume mounts, tmux, and more for different cloud providers.
+
+:::{.callout-tip}
+
+Jupyter lab is run by default. Set `JUPYTER_DISABLE=1` in the environment variables to disable it.
+
+:::
+
+#### Image
+
+```
+axolotlai/axolotl-cloud
+```
+
+Link: [Docker Hub](https://hub.docker.com/r/axolotlai/axolotl-cloud)
+
+#### Tags format
+
+This uses the same tags as the [`main` image](#sec-main-tags).
+
+#### Environment variables
+
+- `JUPYTER_DISABLE`: Disable Jupyter lab.
+- `JUPYTER_PASSWORD`: Set a password for the Jupyter lab.
+- `PUBLIC_KEY`: Add a public key for the SSH service.
+- `SSH_KEY`: Add a private key for the SSH service.
+
+#### Volume mounts
+
+:::{.callout-tip}
+
+We recommend mounting volumes to `/workspace/data` for data persistence. `/workspace/axolotl` contains the source code and is ephemeral.
+
+:::
+
+- `/workspace/data/axolotl-artifacts`: Directory to store Axolotl artifacts.
+- `/workspace/data/huggingface-cache`: Directory to store HuggingFace cache.
+
+## Cloud-no-tmux
+
+This is the same as the [`cloud` image](#sec-cloud) but without tmux.
+
+#### Image
+
+```
+axolotlai/axolotl-cloud-term
+```
+
+Link: [Docker Hub](https://hub.docker.com/r/axolotlai/axolotl-cloud-term)
+
+:::{.callout-note}
+
+The naming may be a bit confusing as it has `-term` appended to the end.
+
+:::
+
+#### Tags format
+
+This uses the same tags as the [`cloud` image](#sec-cloud-tags).

docs/faq.qmd

@@ -3,6 +3,7 @@ title: FAQ
 description: Frequently asked questions
 ---
 
+### General
 
 **Q: The trainer stopped and hasn't progressed in several minutes.**
 
@@ -18,4 +19,50 @@ description: Frequently asked questions
 
 **Q: AttributeError: 'DummyOptim' object has no attribute 'step'**
 
-> A: You may be using deepspeed with single gpu. Please don't set `deepspeed:` in yaml or cli.
+**Q: ModuleNotFoundError: No module named 'mpi4py' using single GPU with deepspeed**
+
+> A: You may be using deepspeed with single gpu. Please remove the `deepspeed:` section in the yaml file or `--deepspeed` CLI flag.
+
+**Q: The codes is stuck on saving preprocessed datasets.**
+
+> A: This is usually an issue with the GPU. This can be resolved through setting the os environment variable `CUDA_VISIBLE_DEVICES=0`. If you are on runpod, this is usually a pod issue. Starting a new pod should take care of it.
+
+**Q: Received mismatch error on merge adapters / loading adapters between torch.Size of checkpoint and model.**
+
+> A: This is likely due to vocab size mismatch. By default, Axolotl expands the model's embeddings if the tokenizer has more tokens than the model. Please use the `axolotl merge-lora` command to merge the adapters instead of using your own scripts.
+
+> On the other hand, if the model has more tokens than the tokenizer, Axolotl does not shrink the model's embeddings unless `shrink_embeddings: true` is set in the config.
+
+**Q: How to call Axolotl via custom python scripts?**
+
+> A: Yes, since Axolotl is just Python, please see `src/axolotl/cli/main.py` on how each command is called.
+
+### Chat templates
+
+**Q: `jinja2.exceptions.UndefinedError: 'dict object' has no attribute 'content' / 'role' / ____`**
+
+> A: This means that the property mapping for the stated attribute does not exist when building `chat_template` prompt. For example, if `no attribute 'content'`, please check you have added the correct mapping for `content` under `message_property_mappings`.
+
+**Q: `Empty template generated for turn ___`**
+
+> A: The `content` is empty for that turn.
+
+**Q: `Could not find content start/end boundary for turn __`**
+
+> A: The specific turn's start/end could not be detected. Please ensure you have set the `eos_token` following your `chat_template`. Otherwise, this could be a `chat_template` which doesn't use proper boundaries for each turn (like system). On the rare occurrence, make sure your content is not `[[dummy_message]]`. Please let us know about this.
+
+**Q: `Content end boundary is before start boundary for turn ___`**
+
+> A: This is an edge case which should not occur. Please create an Issue if this happens.
+
+**Q: `Content end boundary is the same as start boundary for turn ___. This is likely an empty turn.`**
+
+> A: This is likely an empty turn.
+
+**Q: The EOS/EOT token is incorrectly being masked or not being masked.**
+
+> A: This is because of the mismatch between `tokenizer.eos_token` and EOS/EOT token in template. Please make sure to set `eos_token` under `special_tokens` to the same EOS/EOT token as in template.
+
+**Q: "`chat_template` choice is `tokenizer_default` but tokenizer's `chat_template` is null. Please add a `chat_template` in tokenizer config"**
+
+> A: This is because the tokenizer does not have a chat template. Please add a chat template in the tokenizer config. See [chat_template](dataset-formats/conversation.qmd#chat-template) for more details.

docs/getting-started.qmd

@@ -1,5 +1,5 @@
 ---
-title: "Getting Started with Axolotl"
+title: "Quickstart"
 format:
   html:
     toc: true
@@ -17,12 +17,12 @@ Let's start by fine-tuning a small language model using LoRA. This example uses
 Assuming `axolotl` is installed (if not, see our [Installation Guide](installation.qmd))
 
 1. Download example configs:
-```shell
+```bash
 axolotl fetch examples
 ```
 
 2. Run the training:
-```shell
+```bash
 axolotl train examples/llama-3/lora-1b.yml
 ```
 
@@ -36,7 +36,9 @@ The YAML configuration file controls everything about your training. Here's what
 
 ```yaml
 base_model: NousResearch/Llama-3.2-1B
-# hub_model_id: username/custom_model_name
+
+load_in_8bit: true
+adapter: lora
 
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
@@ -44,11 +46,15 @@ datasets:
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.1
 output_dir: ./outputs/lora-out
-
-adapter: lora
-lora_model_dir:
 ```
 
+::: {.callout-tip}
+`load_in_8bit: true` and `adapter: lora` enables LoRA adapter finetuning.
+
+- To perform Full finetuning, remove these two lines.
+- To perform QLoRA finetuning, replace with `load_in_4bit: true` and `adapter: qlora`.
+:::
+
 See our [Config options](config.qmd) for more details.
 
 ### Training {#sec-training}
@@ -56,7 +62,7 @@ See our [Config options](config.qmd) for more details.
 When you run `axolotl train`, Axolotl:
 
 1. Downloads the base model
-2. (If specified) applies LoRA adapter layers
+2. (If specified) applies QLoRA/LoRA adapter layers
 3. Loads and processes the dataset
 4. Runs the training loop
 5. Saves the trained model and / or LoRA weights
@@ -69,6 +75,8 @@ Let's modify the example for your own data:
 
 ```yaml
 base_model: NousResearch/Nous-Hermes-llama-1b-v1
+
+load_in_8bit: true
 adapter: lora
 
 # Training settings
@@ -104,11 +112,9 @@ format):
 {"instruction": "Classify this text", "input": "Not good at all", "output": "negative"}
 ```
 
-Please consult the supported [Dataset Formats](dataset-formats/) for more details.
-
 3. Run the training:
 
-```shell
+```bash
 axolotl train my_training.yml
 ```
 
@@ -118,7 +124,7 @@ axolotl train my_training.yml
 
 After training, test your model:
 
-```shell
+```bash
 axolotl inference my_training.yml --lora-model-dir="./outputs/lora-out"
 ```
 
@@ -126,7 +132,7 @@ axolotl inference my_training.yml --lora-model-dir="./outputs/lora-out"
 
 For large datasets, preprocess first:
 
-```shell
+```bash
 axolotl preprocess my_training.yml
 ```
 
@@ -134,7 +140,7 @@ axolotl preprocess my_training.yml
 
 Launch a Gradio interface:
 
-```shell
+```bash
 axolotl inference my_training.yml --lora-model-dir="./outputs/lora-out" --gradio
 ```
 

docs/inference.qmd

@@ -1,19 +1,22 @@
 ---
-title: "Inference Guide"
+title: "Inference and Merging"
 format:
   html:
     toc: true
     toc-depth: 3
     number-sections: true
-    code-tools: true
 execute:
   enabled: false
 ---
 
-This guide covers how to use your trained models for inference, including model loading, interactive testing, and common troubleshooting steps.
+This guide covers how to use your trained models for inference, including model loading, interactive testing, merging adapters, and common troubleshooting steps.
 
 ## Quick Start {#sec-quickstart}
 
+::: {.callout-tip}
+Use the same config used for training on inference/merging.
+:::
+
 ### Basic Inference {#sec-basic}
 
 ::: {.panel-tabset}

docs/input_output.qmd

@@ -3,263 +3,4 @@ title: Template-free prompt construction
 description: "Template-free prompt construction with the `input_output` format"
 ---
 
-<!-- TOC -->
-
-- [Background](#background)
-    - [Masking Inputs](#masking-inputs)
-    - [You may not want prompt templates](#you-may-not-want-prompt-templates)
-    - [The `input_output` format](#the-input_output-format)
-- [Usage](#usage)
-    - [1. Prepare Data](#1-prepare-data)
-    - [2. Use `type: input_output`](#2-use-type-input_output)
-    - [3. Check the prompts](#3-check-the-prompts)
-
-<!-- /TOC -->
-
-<a id="markdown-background" name="background"></a>
-
-## Background
-
-<a id="markdown-masking-inputs" name="masking-inputs"></a>
-
-### Masking Inputs
-
-One of the most popular features of
-[axolotl](https://github.com/axolotl-ai-cloud/axolotl) is
-setting the following configuration value:
-
-
-```yaml
-train_on_inputs: false
-```
-
-If you declare a [dataset formats](https://github.com/axolotl-ai-cloud/axolotl?tab=readme-ov-file#dataset)
-such as `alpaca` or `chatml`, axolotl knows what is an input
-(i.e. human) vs. an output (i.e. the assistant) and masks the input
-labels so that your model can focus on predicting the outputs only.
-
-<a id="markdown-you-may-not-want-prompt-templates" name="you-may-not-want-prompt-templates"></a>
-
-### You may not want prompt templates
-
-However, there are many situations where you don't want to use one of
-these formats or templates. This is because they can:
-
--   Add unnecessary boilerplate to your prompts.
--   Create artifacts like special delimiters `<|im_start|>` that can
-    quickly become footguns if you don't include them correctly at
-    inference time.
--   Enforce a *chat* interface when you do not want one. Sometimes you
-    just want to fine-tune a model to a very specific task and do NOT
-    want multi-turn conversations, roles, etc.
--   Limit you to only certain roles that the template allows.
-
-<a id="markdown-the-inputoutput-format" name="the-inputoutput-format"></a>
-
-### The `input_output` format
-
-You can construct your prompts without a template by using the
-`input_output` format, by setting `type: input_output` in your
-configuration file like this:
-
-**config.yml**
-
-```yaml
-train_on_inputs: false # Mask segments of your data
-datasets:
-  - path: output.jsonl
-    type: input_output  # use template free prompt construction
-```
-
-Unlike `type: completion`, which is also template-free,
-`type: input_output` allows you to mask segments of your text. More
-details on how this works are described below.
-
-<a id="markdown-usage" name="usage"></a>
-
-## Usage
-
-This is how you can use the `input_output` format:
-
-<a id="markdown-1-prepare-data" name="1-prepare-data"></a>
-
-### 1. Prepare Data
-
-To use the `input_output` format, collect your data in the following
-format into a jsonl file (below is the first row from the file
-`output`.jsonl` pretty printed):
-
-```bash
-$ head -n1 output.jsonl | python -m json.tool
-```
-
-:::{.cell-output .cell-output-stdout}
-    {
-        "segments": [
-            {
-                "label": true,
-                "text": "<s>Hello\n"
-            },
-            {
-                "label": true,
-                "text": "hi there!. "
-            },
-            {
-                "label": false,
-                "text": "goodbye "
-            },
-            {
-                "label": true,
-                "text": "farewell</s>"
-            }
-        ]
-    }
-:::
-
-Set `label:false` when you want to mask a segment of text so that the
-model isn't trained on it. Some things to keep in mind:
-
-> [!IMPORTANT]
-> 1.  **EOS, BOS, spaces, newlines etc. are entirely up to you. Axolotl
-    concatenates all the segments as-is.** The tokenizer doesn't add
-    anything additional. Notice how I added spaces, newlines, `<s>`
-    (BOS), and `</s>` (EOS) myself.
-> 2.  Make sure you check the materialized output to validate that the
-    prompt is getting assembled how you like.
-
-<a id="markdown-2-use-type-inputoutput" name="2-use-type-inputoutput"></a>
-
-### 2. Use `type: input_output`
-
-Let's materialize data with our `output.jsonl` file by setting
-`type: input_output` in our axolotl config:
-
-```yaml
-# training_config.yaml
-base_model: mistralai/Mistral-7B-v0.1
-data_seed: 49
-seed: 49
-
-datasets:
-  - path: output.jsonl
-    type: input_output
-val_set_size: 0.1
-
-sequence_len: 896
-sample_packing: false
-
-micro_batch_size: 2
-gradient_accumulation_steps: 3
-eval_batch_size: 2
-num_epochs: 1
-learning_rate: 0.0002
-
-train_on_inputs: false
-special_tokens:
-  bos_token: "<s>"
-  eos_token: "</s>"
-  unk_token: "<unk>"
-```
-
-You can use the following command to materialize your data. The
-`--debug` flag will print the tokens, along with the labels so you can
-verify that the correct items are being ignored:
-
-```bash
-$ python -m axolotl.cli.preprocess training_config.yaml --debug
-
-...
-[2024-03-05 23:36:46,969] [INFO] [axolotl.check_example_labels:35] [PID:607731] [RANK:0] <s>(1, 1) Hello(22557, 22557)
-(13, 13) hi(12014, 12014) there(736, 736) !(28808, 28808) .(28723, 28723) (28705, 28705) good(-100, 1179) bye(-100, 17664) (-100, 28705) fare(19111, 19111) well(5458, 5458) </s>(2, 2)
-
-```
-
-The format is `decoded_token`(`label`, `token_id`), for example,
-`<s>(1, 1)` means that the token is `<s>`, the label is `1` and the
-token_id is `1`. When the label is `-100` then that token is ignored for
-training.
-
-<a id="markdown-3-check-the-prompts" name="3-check-the-prompts"></a>
-
-### 3. Check the prompts
-
-Here is another way to check the materialized output:
-
-```python
-from transformers import AutoTokenizer
-from datasets import load_from_disk
-import yaml
-
-directory = !ls last_run_prepared/
-with open('training_config.yaml', 'r') as f:
-    cfg = yaml.safe_load(f)
-model_id = cfg['base_model']
-tok = AutoTokenizer.from_pretrained(model_id)
-ds = load_from_disk(f'last_run_prepared/{directory[0]}/')
-```
-
-```python
->>> row = ds[0]
->>> print(tok.decode(row['input_ids']))
-<s> Hello
-    hi there!.  goodbye  farewell</s>
-```
-
-We can check that the right tokens are ignored by comparing the labels
-to each token:
-
-```python
-import pandas as pd
-pd.DataFrame([{'token': tok.decode(i), 'label': l, 'id':i} for i,l in
-              zip(row['input_ids'], row['labels'])])
-```
-
-| token | label | id    |
-|-------|-------|-------|
-| 0     | \<s\> | 1     |
-| 1     | Hello | 22557 |
-| 2     | \\n   | 13    |
-| 3     | hi    | 12014 |
-| 4     | there | 736   |
-| 5     | !     | 28808 |
-| 6     | .     | 28723 |
-| 7     |       | 28705 |
-| 8     | good  | -100  |
-| 9     | bye   | -100  |
-| 10    |       | -100  |
-| 11    | fare  | 19111 |
-| 12    | well  | 5458  |
-| 13    | \</s\>| 2     |
-
-
-
-If we look at the input data, the above table seems correct! (The jsonl
-version is repeated below for reference):
-
-
-```bash
-$ head -n1 output.jsonl | python -m json.tool
-```
-
-:::{.cell-output .cell-output-stdout}
-    {
-        "segments": [
-            {
-                "label": true,
-                "text": "<s>Hello\n"
-            },
-            {
-                "label": true,
-                "text": "hi there!. "
-            },
-            {
-                "label": false,
-                "text": "goodbye "
-            },
-            {
-                "label": true,
-                "text": "farewell</s>"
-            }
-        ]
-    }
-:::
+The documentation moved to [here](dataset-formats/template_free.qmd).

docs/installation.qmd

@@ -1,11 +1,10 @@
 ---
-title: "Installation Guide"
+title: "Installation"
 format:
   html:
     toc: true
     toc-depth: 3
     number-sections: true
-    code-tools: true
 execute:
   enabled: false
 ---
@@ -23,6 +22,7 @@ This guide covers all the ways you can install and set up Axolotl for your envir
 ### PyPI Installation (Recommended) {#sec-pypi}
 
 ```{.bash}
+pip3 install -U packaging setuptools wheel ninja
 pip3 install --no-build-isolation axolotl[flash-attn,deepspeed]
 ```
 
@@ -38,7 +38,7 @@ For the latest features between releases:
 ```{.bash}
 git clone https://github.com/axolotl-ai-cloud/axolotl.git
 cd axolotl
-pip3 install packaging ninja
+pip3 install -U packaging setuptools wheel ninja
 pip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'
 ```
 
@@ -66,6 +66,8 @@ docker run --privileged --gpus '"all"' --shm-size 10g --rm -it \
 ```
 :::
 
+Please refer to the [Docker documentation](docker.qmd) for more information on the different Docker images that are available.
+
 ## Cloud Environments {#sec-cloud}
 
 ### Cloud GPU Providers {#sec-cloud-gpu}
@@ -77,6 +79,7 @@ For providers supporting Docker:
   - [Latitude.sh](https://latitude.sh/blueprint/989e0e79-3bf6-41ea-a46b-1f246e309d5c)
   - [JarvisLabs.ai](https://jarvislabs.ai/templates/axolotl)
   - [RunPod](https://runpod.io/gsc?template=v2ickqhz9s&ref=6i7fkpdz)
+  - [Novita](https://novita.ai/gpus-console?templateId=311)
 
 ### Google Colab {#sec-colab}
 
@@ -106,7 +109,7 @@ We recommend using WSL2 (Windows Subsystem for Linux) or Docker.
 2. Install PyTorch: https://pytorch.org/get-started/locally/
 3. Install Axolotl:
    ```{.bash}
-   pip3 install packaging
+   pip3 install -U packaging setuptools wheel ninja
    pip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'
    ```
 4. (Optional) Login to Hugging Face:

docs/lora_optims.qmd

@@ -0,0 +1,131 @@
+---
+title: "LoRA Optimizations"
+description: "Custom autograd functions and Triton kernels in Axolotl for optimized LoRA fine-tuning"
+---
+
+Inspired by [Unsloth](https://github.com/unslothai/unsloth), we've implemented two
+optimizations for LoRA and QLoRA fine-tuning, supporting both single GPU and multi-GPU
+(in the DDP and DeepSpeed settings) training. These include (1) SwiGLU and GEGLU activation function
+Triton kernels, and (2) LoRA MLP and attention custom autograd functions. Our goal was
+to leverage operator fusion and tensor re-use in order to improve speed and reduce
+memory usage during the forward and backward passes of these calculations.
+
+We currently support several common model architectures, including (but not limited to):
+
+- `llama`
+- `mistral`
+- `qwen2`
+- `gemma`
+- `gemma2`
+
+<details>
+
+The set of models we support is currently limited by our attention patching strategy,
+which assumes (and replaces) specific code blocks for query / key / value and output
+projections:
+
+```python
+ORIGINAL_QKV_CODE = """
+    query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+    key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+    value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+""".lstrip(
+    "\n"
+)
+
+ORIGINAL_O_CODE = """
+    attn_output = self.o_proj(attn_output)
+""".lstrip(
+    "\n"
+)
+```
+
+Is replaced with:
+
+```python
+PATCHED_QKV_CODE = """
+    query_states, key_states, value_states = self.apply_qkv(hidden_states)
+    query_states = query_states.view(hidden_shape).transpose(1, 2)
+    key_states = key_states.view(hidden_shape).transpose(1, 2)
+    value_states = value_states.view(hidden_shape).transpose(1, 2)
+""".lstrip(
+    "\n"
+)
+
+PATCHED_O_CODE = """
+    attn_output = self.apply_o(attn_output)
+""".lstrip(
+    "\n"
+)
+```
+
+Where `apply_qkv` and `apply_o` are defined in the `axolotl.kernels.lora` module.
+
+We welcome testing of other model architectures and / or PRs to expand our patching
+logic to be compatible with more of them.
+
+</details>
+
+::: {.callout-tip}
+Check out our [LoRA optimizations blog](https://axolotlai.substack.com/p/accelerating-lora-fine-tuning-with).
+:::
+
+## Usage
+
+These optimizations can be enabled in your Axolotl config YAML file. The
+`lora_mlp_kernel` option enables the optimized MLP path, while `lora_qkv_kernel` and
+`lora_o_kernel` enable the fused query-key-value projection and optimized output
+projection, respectively.
+
+```yaml
+lora_mlp_kernel: true
+lora_qkv_kernel: true
+lora_o_kernel: true
+```
+
+## Requirements
+
+- One or more NVIDIA or AMD GPUs (in order to use the Triton kernels)
+    - Note: Set `TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1` to enable [memory-efficient attention on AMD GPUs](https://github.com/ROCm/aotriton/issues/16#issuecomment-2346675491)
+- Targeted LoRA adapters cannot use Dropout
+    - This may limit model expressivity / cause overfitting
+- Targeted LoRA adapters cannot have bias terms
+    - This may limit model expressivity
+
+Models with pre-existing LoRA adapters that use Dropout or have bias terms may need to
+be re-finetuned without these features in order to be useful.
+
+## Implementation details
+
+### Custom autograd functions
+
+The LoRA MLP autograd function optimizes the entire MLP computation path. It fuses the
+LoRA and base weight computations together and provides a single, efficient backward
+pass for the entire MLP block.
+
+For attention components, similar optimizations are provided through a function that
+handles the query, key, and value projections, and a function that handles the output
+projection. They are designed to work with the existing `transformers` attention
+implementation via some monkey-patching logic.
+
+### Triton kernels
+
+Two activation functions (SwiGLU and GeGLU) are implemented with Triton kernels for
+improved speed and memory performance. These kernels handle both the forward and
+backward passes.
+
+### Integration
+
+The custom autograd functions and Triton kernels are designed to work together. The
+autograd function manages the high-level computation flow and gradient tracking, while
+calling the Triton kernels for the activation function computation. During the backward
+pass, the kernel computes both the activation output and the required gradients, which
+the autograd function then uses to compute the final gradients for the entire
+computation path.
+
+## Future Work
+
+- Support for additional model architectures
+- Support for the FSDP setting
+- Support for dropout and bias
+- Additional operator fusions

docs/mac.qmd

@@ -19,4 +19,5 @@ Current support:
 - [ ] DeepSpeed
 
 Untested:
+
 - FSDP

docs/multi-gpu.qmd

@@ -1,5 +1,5 @@
 ---
-title: "Multi-GPU Training Guide"
+title: "Multi-GPU"
 format:
   html:
     toc: true
@@ -35,7 +35,11 @@ deepspeed: deepspeed_configs/zero1.json
 ### Usage {#sec-deepspeed-usage}
 
 ```{.bash}
-accelerate launch -m axolotl.cli.train examples/llama-2/config.yml --deepspeed deepspeed_configs/zero1.json
+# Passing arg via config
+axolotl train config.yml
+
+# Passing arg via cli
+axolotl train config.yml --deepspeed deepspeed_configs/zero1.json
 ```
 
 ### ZeRO Stages {#sec-zero-stages}
@@ -70,25 +74,7 @@ For combining FSDP with QLoRA, see our [dedicated guide](fsdp_qlora.qmd).
 
 ### Liger Kernel Integration {#sec-liger}
 
-::: {.callout-note}
-Liger Kernel provides efficient Triton kernels for LLM training, offering:
-
-- 20% increase in multi-GPU training throughput
-- 60% reduction in memory usage
-- Compatibility with both FSDP and DeepSpeed
-:::
-
-Configuration:
-
-```{.yaml}
-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
-```
+Please see [docs](custom_integrations.qmd#liger) for more info.
 
 ## Troubleshooting {#sec-troubleshooting}
 

docs/multi-node.qmd

@@ -3,6 +3,18 @@ title: Multi Node
 description: How to use Axolotl on multiple machines
 ---
 
+The below are three ways to train multi-node in Axolotl.
+
+::: {.callout-important}
+Each machine needs a copy of Axolotl, we suggest using the same commit to ensure compatibility.
+
+You will also need to have the same configuration file for your model on each machine.
+
+Make sure the main machine is reachable by other machines.
+:::
+
+## Accelerate
+
 You will need to create a configuration for accelerate, either by using `accelerate config` and follow the instructions or you can use one of the preset below:
 
 ~/.cache/huggingface/accelerate/default_config.yaml
@@ -26,7 +38,7 @@ tpu_use_sudo: false
 use_cpu: false
 ```
 
-Configure your model to use FSDP with for example:
+Configure your model to use FSDP in the Axolotl yaml. For example:
 ```yaml
 fsdp:
   - full_shard
@@ -37,12 +49,40 @@ fsdp_config:
   fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
 ```
 
-## Machine configuration
-
-On each machine you need a copy of Axolotl, we suggest using the same commit to ensure compatibility.
-
-You will also need to have the same configuration file for your model on each machine.
-
-On the main machine only, make sure the port you set as `main_process_port` is open in TCP and reachable by other machines.
-
 All you have to do now is launch using accelerate as you would usually do on each machine and voila, the processes will start once you have launched accelerate on every machine.
+
+## Raytrain
+
+Please see ray train doc [here](ray-integration.qmd).
+
+## Torchrun
+
+If you are using Infiniband, we recommend torchrun to utilize the full bandwidth.
+
+Set the following env (change buffersize/socketname depending on your system):
+
+```bash
+export NCCL_IB_DISABLE=0
+export NCCL_SOCKET_IFNAME="eth0,en,eth,em,bond"
+export NCCL_BUFFSIZE=2097152
+```
+
+Run the following on each node:
+
+```bash
+torchrun --nnodes $num_nodes --nproc_per_node $gpu_per_node --rdzv_id $rdzv_id --rdzv_backend c10d --rdzv_endpoint "$head_node_ip:$head_node_port" -m axolotl.cli.train config.yaml
+```
+
+Please make sure to substitute the placeholder variables.
+
+- `num_nodes`: Number of nodes (containing GPUs)
+- `gpu_per_node`: Number of gpus per node
+- `head_node_ip`: IP of the head node (make sure other machines can connect to this)
+- `head_node_port`: Port of the head node (make sure other machines can connect to this. Default 29400)
+- `rdzv_id`: A unique job ID that is used by the job across nodes.
+
+::: {.callout-note}
+You need to call `axolotl.cli.train` instead of `axolotl train` as the latter calls accelerate under the hood
+:::
+
+More info on the available configs can be found on the Pytorch docs [here](https://pytorch.org/docs/stable/elastic/run.html)

docs/nccl.qmd

@@ -13,13 +13,13 @@ Often, this timeout will happen after 30 minutes (the default setting) and is ac
 
 Forcing cross-GPU communication via [NVLink](https://en.wikipedia.org/wiki/NVLink) may help without increasing timeouts. To verify that your configuration is leveraging NVLink run the following command:
 
-```shell
+```bash
 nvidia-smi nvlink --status
 ```
 
 To force NCCL to use NVLink, simply set this in the environment:
 
-```shell
+```bash
 export NCCL_P2P_LEVEL=NVL
 ```
 
@@ -33,13 +33,13 @@ If NVLink is not available in your environment there are other options for ``NCC
 
 To validate that acceptable data transfer speeds exist for your training job, running [NCCL Tests](https://github.com/NVIDIA/nccl-tests/blob/master/README.md) can help pinpoint bottlenecks, for example:
 
-```shell
+```bash
 ./build/all_reduce_perf -b 8 -e 128M -f 2 -g 3
 ```
 
 It can be useful when debugging NCCL communication timeouts to activate additional logging in both PyTorch and NCCL:
 
-```shell
+```bash
 export NCCL_DEBUG=INFO
 export NCCL_DEBUG_SUBSYS=ALL
 export TORCH_DISTRIBUTED_DEBUG=INFO

docs/ray-integration.qmd

@@ -1,5 +1,5 @@
 ---
-title: Ray Train integration
+title: Ray Train
 description: How to use Axolotl with Ray Train
 ---
 
@@ -9,7 +9,7 @@ With the `--use-ray` CLI flag, Axolotl will use Ray Train's [`TorchTrainer`](htt
 
 ## Ray cluster setup
 
-A prerequisite using the Ray Train integration is to setup a Ray cluster on your desired node(s). For a detailed guide on how you can get started with ray clusters, check the official Ray docs here: https://docs.ray.io/en/latest/cluster/getting-started.html
+A prerequisite using the Ray Train integration is to setup a Ray cluster on your desired node(s). For a detailed guide on how you can get started with ray clusters, check the official Ray docs [here](https://docs.ray.io/en/latest/cluster/getting-started.html).
 
 Every Ray cluster has one _head_ node and a set of worker nodes. The head node is just like any other worker node, but it also runs certain special processes related to scheduling and orchestration. Ray-enabled scripts are run on the head node and depending on the resources (number of CPUs, GPUs, etc) they request, will be scheduled to run certain tasks on the worker nodes. For more on key concepts behind a Ray cluster, you can refer this [doc](https://docs.ray.io/en/latest/cluster/key-concepts.html#cluster-key-concepts).
 
@@ -58,13 +58,11 @@ You can find an example configuration at `configs/llama-3/lora-1b-ray.yaml`.
 The key parameters to note here are:
 
 ```yaml
-...
 use_ray: true
 ray_num_workers: 4
 # optional
 resources_per_worker:
     GPU: 1
-...
 ```
 
 - `use_ray`: This is the flag that enables the Ray Train integration. You can either use the corresponding `--use-ray` flag in the CLI or set `use_ray` in the config file.

docs/reward_modelling.qmd

@@ -28,8 +28,23 @@ val_set_size: 0.1
 eval_steps: 100
 ```
 
+Bradley-Terry chat templates expect single-turn conversations in the following format:
+
+```json
+{
+    "system": "...", // optional
+    "input": "...",
+    "chosen": "...",
+    "rejected": "..."
+}
+```
+
 ### Process Reward Models (PRM)
 
+::: {.callout-tip}
+Check out our [PRM blog](https://axolotlai.substack.com/p/process-reward-models).
+:::
+
 Process reward models are trained using data which contains preference annotations for each step in a series of interactions. Typically, PRMs are trained to provide reward signals over each step of a reasoning trace and are used for downstream reinforcement learning.
 ```yaml
 base_model: Qwen/Qwen2.5-3B
@@ -45,3 +60,5 @@ datasets:
 val_set_size: 0.1
 eval_steps: 100
 ```
+
+Please see [stepwise_supervised](dataset-formats/stepwise_supervised.qmd) for more details on the dataset format.

docs/rlhf.qmd

@@ -1,26 +1,40 @@
 ---
 title: "RLHF (Beta)"
 description: "Reinforcement Learning from Human Feedback is a method whereby a language model is optimized from data using human feedback."
+back-to-top-navigation: true
+toc: true
+toc-expand: 2
+toc-depth: 4
 ---
 
-### Overview
+## Overview
 
 Reinforcement Learning from Human Feedback is a method whereby a language model is optimized from data using human
 feedback. Various methods include, but not limited to:
 
+- [Direct Preference Optimization (DPO)](#dpo)
+- [Identity Preference Optimization (IPO)](#ipo)
+- [Kahneman-Tversky Optimization (KTO)](#kto)
+- [Odds Ratio Preference Optimization (ORPO)](#orpo)
 - Proximal Policy Optimization (PPO) (not yet supported in axolotl)
-- Direct Preference Optimization (DPO)
-- Identity Preference Optimization (IPO)
 
 
-### RLHF using Axolotl
+## RLHF using Axolotl
 
->[!IMPORTANT]
->This is a BETA feature and many features are not fully implemented. You are encouraged to open new PRs to improve the integration and functionality.
+::: {.callout-important}
+This is a BETA feature and many features are not fully implemented. You are encouraged to open new PRs to improve the integration and functionality.
+:::
 
-The various RL training methods are implemented in trl and wrapped via axolotl. Below are various examples with how you can use various preference datasets to train models that use ChatML
+We rely on the [TRL](https://github.com/huggingface/trl) library for implementations of various RL training methods, which we wrap around to expose in axolotl. Each method has their own supported ways of loading datasets and prompt formats.
+
+::: {.callout-tip}
+You can find what each method supports by going into `src/axolotl/prompt_strategies/{method}` where `{method}` is one of our supported methods. The `type: ` can be retrieved from `{method}.{function_name}`.
+:::
+
+### DPO
+
+Example config:
 
-#### DPO
 ```yaml
 rl: dpo
 datasets:
@@ -32,12 +46,265 @@ datasets:
     type: chatml
 ```
 
-#### IPO
+DPO supports the following types with the following dataset format:
+
+#### chatml.argilla
+
+```json
+{
+    "system": "...", // optional
+    "instruction": "...",
+    "chosen_response": "...",
+    "rejected_response": "..."
+}
+```
+
+#### chatml.argilla_chat
+
+```json
+{
+    "chosen": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ],
+    "rejected": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ]
+}
+```
+
+#### chatml.icr
+
+```json
+{
+    "system": "...", // optional
+    "input": "...",
+    "chosen": "...",
+    "rejected": "..."
+}
+```
+
+#### chatml.intel
+
+```json
+{
+    "system": "...", // optional
+    "question": "...",
+    "chosen": "...",
+    "rejected": "..."
+}
+```
+
+#### chatml.prompt_pairs
+
+```json
+{
+    "system": "...", // optional
+    "prompt": "...",
+    "chosen": "...",
+    "rejected": "..."
+}
+```
+
+#### chatml.ultra
+
+```json
+{
+    "system": "...", // optional
+    "prompt": "...",
+    "chosen": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ],
+    "rejected": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ]
+}
+```
+
+#### llama3.argilla
+
+```json
+{
+    "system": "...", // optional
+    "instruction": "...",
+    "chosen_response": "...",
+    "rejected_response": "..."
+}
+```
+
+#### llama3.argilla_chat
+
+```json
+{
+    "chosen": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ],
+    "rejected": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ]
+}
+```
+
+#### llama3.icr
+
+```json
+{
+    "system": "...", // optional
+    "input": "...",
+    "chosen": "...",
+    "rejected": "..."
+}
+```
+
+#### llama3.intel
+
+```json
+{
+    "system": "...", // optional
+    "question": "...",
+    "chosen": "...",
+    "rejected": "..."
+}
+```
+
+#### llama3.prompt_pairs
+
+```json
+{
+    "system": "...", // optional
+    "prompt": "...",
+    "chosen": "...",
+    "rejected": "..."
+}
+```
+
+#### llama3.ultra
+
+```json
+{
+    "system": "...", // optional
+    "prompt": "...",
+    "chosen": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ],
+    "rejected": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ]
+}
+```
+
+#### zephyr.nectar
+
+```json
+{
+    "prompt": "...",
+    "answers": [
+        {
+            "answer": "...",
+            "rank": 1
+        },
+        {
+            "answer": "...",
+            "rank": 2
+        }
+        // ... more answers with ranks
+    ]
+}
+```
+
+#### chat_template.default
+
+```yaml
+rl: dpo
+datasets:
+  - path: ...
+    split: train
+    type: chat_template.default
+    field_messages: "messages"
+    field_chosen: "chosen"
+    field_rejected: "rejected"
+    message_property_mappings:
+      role: role
+      content: content
+    roles:
+      user: ["user"]
+      assistant: ["assistant"]
+      system: ["system"]
+```
+
+Sample input format:
+
+```json
+{
+    "messages": [
+        {
+            "role": "system",
+            "content": "..."
+        },
+        {
+            "role": "user",
+            "content": "..."
+        },
+        // ... more messages
+    ],
+    "chosen": {
+        "role": "assistant",
+        "content": "..."
+    },
+    "rejected": {
+        "role": "assistant",
+        "content": "..."
+    }
+}
+```
+
+#### user_defined.default
+
+For custom behaviors,
+
+```yaml
+rl: dpo
+datasets:
+  - path: ...
+    split: train
+    type: user_defined.default
+
+    field_prompt: "prompt"
+    field_system: "system"
+    field_chosen: "chosen"
+    field_rejected: "rejected"
+    prompt_format: "{prompt}"
+    chosen_format: "{chosen}"
+    rejected_format: "{rejected}"
+```
+
+The input format is a simple JSON input with customizable fields based on the above config.
+
+```json
+{
+    "system": "...",  // optional
+    "prompt": "...",
+    "chosen": "...",
+    "rejected": "..."
+}
+```
+
+### IPO
+
+As IPO is just DPO with a different loss function, all supported dataset formats for [DPO](#dpo) are also supported for IPO.
+
 ```yaml
 rl: ipo
 ```
 
-#### ORPO
+### ORPO
 
 Paper: https://arxiv.org/abs/2403.07691
 
@@ -52,13 +319,34 @@ datasets:
     type: chat_template.argilla
 ```
 
+ORPO supports the following types with the following dataset format:
 
-#### KTO
+#### chat_template.argilla
+
+```json
+{
+    "system": "...",  // optional
+    "prompt": "...",  // if available, will be taken as user message for single-turn instead of from list below
+
+    // chosen/rejected should be same till last content and only even-number of alternating user/assistant turns
+    "chosen": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ],
+    "rejected": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ]
+}
+```
+
+### KTO
 
 ```yaml
 rl: kto
-rl_beta: 0.5
-kto_desirable_weight: 0.2
+rl_beta: 0.1  # default
+kto_desirable_weight: 1.0  # default
+kto_undesirable_weight: 1.0  # default
 
 remove_unused_columns: false
 
@@ -72,7 +360,206 @@ gradient_checkpointing_kwargs:
   use_reentrant: true
 ```
 
-#### Using local dataset files
+KTO supports the following types with the following dataset format:
+
+#### chatml.argilla
+
+```json
+{
+    "system": "...", // optional
+    "instruction": "...",
+    "completion": "..."
+}
+```
+
+#### chatml.argilla_chat
+
+```json
+{
+    "chosen": [
+        {"role": "user", "content": "..."}
+    ],
+    "completion": [
+        {"role": "assistant", "content": "..."}
+    ]
+}
+```
+
+#### chatml.intel
+
+```json
+{
+    "system": "...", // optional
+    "question": "...",
+    "completion": "..."
+}
+```
+
+#### chatml.prompt_pairs
+
+```json
+{
+    "system": "...", // optional
+    "prompt": "...",
+    "completion": "..."
+}
+```
+
+#### chatml.ultra
+
+```json
+{
+    "system": "...", // optional
+    "prompt": "...",
+    "completion": "..."
+}
+```
+
+#### llama3.argilla
+
+```json
+{
+    "system": "...", // optional
+    "instruction": "...",
+    "completion": "..."
+}
+```
+
+#### llama3.argilla_chat
+
+```json
+{
+    "completion": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ]
+}
+```
+
+#### llama3.intel
+
+```json
+{
+    "system": "...", // optional
+    "question": "...",
+    "completion": "..."
+}
+```
+
+#### llama3.prompt_pairs
+
+```json
+{
+    "system": "...", // optional
+    "prompt": "...",
+    "completion": "..."
+}
+```
+
+#### llama3.ultra
+
+```json
+{
+    "system": "...", // optional
+    "prompt": "...",
+    "completion": "..."
+}
+```
+
+#### user_defined.default
+
+For custom behaviors,
+
+```yaml
+rl: kto
+datasets:
+  - path: ...
+    split: train
+    type: user_defined.default
+
+    field_prompt: "prompt"
+    field_system: "system"
+    field_completion: "completion"
+    field_label: "label"
+    prompt_format: "{prompt}"
+    completion_format: "{completion}"
+```
+
+The input format is a simple JSON input with customizable fields based on the above config.
+
+```json
+{
+    "system": "...",  // optional
+    "prompt": "...",
+    "completion": "...",
+    "label": "..."
+}
+```
+
+### GRPO
+
+::: {.callout-tip}
+Check out our [GRPO cookbook](https://github.com/axolotl-ai-cloud/axolotl-cookbook/tree/main/grpo#training-an-r1-style-large-language-model-using-grpo).
+:::
+
+GRPO uses custom reward functions and transformations. Please have them ready locally.
+
+For ex, to load OpenAI's GSM8K and use a random reward for completions:
+
+```python
+# rewards.py
+import random
+
+def rand_reward_func(completions, **kwargs) -> list[float]:
+    return [random.uniform(0, 1) for _ in completions]
+
+def oai_gsm8k_transform(cfg, *args, **kwargs):
+    def transform_fn(example, tokenizer=None):
+        label = example["answer"].split("####")[-1].strip().replace(",", "")
+        return {
+            "prompt": [{"role": "user", "content": example["question"]},],
+            "answer": label,
+        }
+    return transform_fn, {"remove_columns": ["question"]}
+```
+
+```yaml
+rl: grpo
+
+trl:
+    beta: 0.001
+    max_completion_length: 256
+    use_vllm: True
+    vllm_device: auto
+    vllm_gpu_memory_utilization: 0.15
+    num_generations: 4
+    reward_funcs: ["rewards.rand_reward_func"]    # format: '{file_name}.{fn_name}'
+    reward_weights: [1.0]
+datasets:
+  - path: openai/gsm8k
+    name: main
+    type: rewards.oai_gsm8k_transform  # format: '{file_name}.{fn_name}'
+```
+
+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 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).
+
+### SimPO
+
+SimPO uses [CPOTrainer](https://huggingface.co/docs/trl/main/en/cpo_trainer) but with alternative loss function.
+
+```yaml
+rl: simpo
+rl_beta: 0.1  # default in CPOTrainer
+cpo_alpha: 1.0  # default in CPOTrainer
+simpo_gamma: 0.5  # default in CPOTrainer
+```
+
+This method uses the same dataset format as [DPO](#dpo).
+
+### Using local dataset files
+
 ```yaml
 datasets:
   - ds_type: json
@@ -82,9 +569,9 @@ datasets:
     type: chatml.intel
 ```
 
-#### Trl autounwrap for peft
+### TRL auto-unwrapping for PEFT
 
-Trl supports autounwrapping peft models, so that a ref model does not need to be additionally loaded, leading to less VRAM needed. This is on by default. To turn it off, pass the following config.
+TRL supports auto-unwrapping PEFT models for RL training paradigms which rely on a reference model. This significantly reduces memory pressure as an additional refreference model does not need to be loaded, and reference model log-probabilities can be obtained by disabling PEFT adapters. This is enabled by default. To turn it off, pass the following config:
 
 ```yaml
 # load ref model when adapter training.

docs/torchao.qmd

@@ -3,6 +3,12 @@ title: "PyTorch ao"
 description: "Custom data types and layouts for training and inference"
 ---
 
+To use experimental optimizers (`AdamWFp8`, `AdamW4bit`, `AdamW8bit`) from Pytorch Ao, please install the package as shown below.
+
+::: {.callout-tip}
+Some experimental optimizers are already present in regular Pytorch, so please re-check if you actually need this package!
+:::
+
 ### Installation
 
 Stable Release from the PyTorch index

docs/unsloth.qmd

@@ -8,6 +8,12 @@ description: "Hyper-optimized QLoRA finetuning for single GPUs"
 Unsloth provides hand-written optimized kernels for LLM finetuning that slightly improve speed and VRAM over
 standard industry baselines.
 
+::: {.callout-important}
+Due to breaking changes in transformers `v4.48.0`, users will need to downgrade to `<=v4.47.1` to use this patch.
+
+This will later be deprecated in favor of [LoRA Optimizations](lora_optims.qmd).
+:::
+
 
 ### Installation
 
@@ -17,7 +23,7 @@ The following will install the correct unsloth and extras from source.
 python scripts/unsloth_install.py | sh
 ```
 
-### Using unsloth w Axolotl
+### Usage
 
 Axolotl exposes a few configuration options to try out unsloth and get most of the performance gains.
 

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

@@ -21,8 +21,9 @@ datasets:
     type: chat_template
     split: train[:20%]
     field_messages: conversations
-    message_field_role: from
-    message_field_content: value
+    message_property_mappings:
+      role: from
+      content: value
 
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0

examples/gemma2/qlora.yml

@@ -16,8 +16,9 @@ datasets:
     type: chat_template
     drop_system_message: true
     field_messages: conversations
-    message_field_role: from
-    message_field_content: value
+    message_property_mappings:
+      role: from
+      content: value
 
 val_set_size: 0.0
 output_dir: ./outputs/out

examples/jamba/qlora_fsdp_large.yaml

@@ -13,8 +13,9 @@ datasets:
     type: chat_template
     drop_system_message: true
     field_messages: conversations
-    message_field_role: from
-    message_field_content: value
+    message_property_mappings:
+      role: from
+      content: value
 
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0

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

@@ -17,8 +17,9 @@ datasets:
     type: chat_template
     split: train[:20%]
     field_messages: conversations
-    message_field_role: from
-    message_field_content: value
+    message_property_mappings:
+      role: from
+      content: value
 
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.02

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

@@ -17,8 +17,9 @@ datasets:
     field_messages: conversation
     field_chosen: chosen
     field_rejected: rejected
-    message_field_role: role
-    message_field_content: content
+    message_property_mappings:
+      role: role
+      content: content
     roles:
       system:
         - system

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

@@ -14,8 +14,9 @@ datasets:
   - path: fozziethebeat/alpaca_messages_2k_test
     type: chat_template
     field_messages: messages
-    message_field_role: role
-    message_field_content: content
+    message_property_mappings:
+      role: role
+      content: content
     roles:
       user:
         - user

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

@@ -17,8 +17,9 @@ datasets:
     field_messages: conversation
     field_chosen: chosen
     field_rejected: rejected
-    message_field_role: role
-    message_field_content: content
+    message_property_mappings:
+      role: role
+      content: content
     roles:
       system:
         - system
@@ -31,8 +32,9 @@ datasets:
     field_messages: conversation
     field_chosen: chosen
     field_rejected: rejected
-    message_field_role: role
-    message_field_content: content
+    message_property_mappings:
+      role: role
+      content: content
     roles:
       system:
         - system

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

@@ -0,0 +1,82 @@
+base_model: NousResearch/Llama-3.2-1B
+# 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
+
+datasets:
+  - path: teknium/GPT4-LLM-Cleaned
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/lora-out
+
+adapter: lora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+pad_to_sequence_len: true
+
+lora_r: 16
+lora_alpha: 32
+# Currently, we don't support dropout with our custom Triton kernels
+# lora_dropout: 0.05
+lora_fan_in_fan_out:
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+# These options enable our custom Triton kernels / autograd
+# functions for MLP and attention calculations
+lora_mlp_kernel: true
+lora_qkv_kernel: true
+lora_o_kernel: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+loss_watchdog_threshold: 5.0
+loss_watchdog_patience: 3
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  pad_token: "<|end_of_text|>"

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

@@ -55,7 +55,7 @@ tf32: true
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
-  use_reentrant: true
+  use_reentrant: false
 early_stopping_patience:
 resume_from_checkpoint:
 local_rank:

examples/mistral/mistral-dpo-qlora.yml

@@ -22,8 +22,9 @@ datasets:
     field_messages: conversation
     field_chosen: chosen
     field_rejected: rejected
-    message_field_role: role
-    message_field_content: content
+    message_property_mappings:
+      role: role
+      content: content
 
 dataset_prepared_path:
 val_set_size: 0.05

examples/phi/lora-3.5.yaml

@@ -14,8 +14,9 @@ datasets:
   - path: fozziethebeat/alpaca_messages_2k_test
     type: chat_template
     field_messages: messages
-    message_field_role: role
-    message_field_content: content
+    message_property_mappings:
+      role: role
+      content: content
     roles:
       user:
         - user

examples/qwen2/dpo.yaml

@@ -12,8 +12,9 @@ datasets:
     field_messages: conversation
     field_chosen: chosen
     field_rejected: rejected
-    message_field_role: role
-    message_field_content: content
+    message_property_mappings:
+      role: role
+      content: content
     roles:
       system:
         - system

index.qmd

@@ -1,7 +1,7 @@
 ---
-toc-location: right-body
-toc-title: Table Of Contents
-toc-expand: 2
+# toc-location: right-body
+# toc-title: Table Of Contents
+# toc-expand: 2
 ---
 
 ```{python}

pyproject.toml

@@ -1,5 +1,5 @@
 [build-system]
-requires = ["setuptools>=64", "wheel", "setuptools_scm>=8"]
+requires = ["setuptools>=64", "wheel", "setuptools_scm>=8", "packaging==23.2"]
 build-backend = "setuptools.build_meta"
 
 [project]
@@ -8,6 +8,7 @@ dynamic = ["version", "dependencies", "optional-dependencies"]
 description = "LLM Trainer"
 readme = "README.md"
 requires-python = ">=3.10"
+# license = "Apache-2.0"
 
 [project.scripts]
 axolotl = "axolotl.cli.main:main"

requirements.txt

@@ -1,24 +1,24 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
 
 # START section of dependencies that don't install on Darwin/MacOS
-bitsandbytes==0.45.1
+bitsandbytes==0.45.3
 triton>=3.0.0
 mamba-ssm==1.2.0.post1
-flash-attn==2.7.0.post2
+flash-attn==2.7.4.post1
 xformers>=0.0.23.post1
 autoawq==0.2.7.post3
-liger-kernel==0.5.2
+liger-kernel==0.5.3
 # END section
 
 packaging==23.2
 
-peft==0.14.0
-transformers==4.48.1
-tokenizers>=0.21.0
-accelerate==1.3.0
-datasets==3.2.0
-deepspeed==0.16.1
-trl==0.13.0
+peft==0.15.0
+transformers==4.49.0
+tokenizers>=0.21.1
+accelerate==1.5.2
+datasets==3.4.1
+deepspeed==0.16.4
+trl==0.15.1
 
 optimum==1.16.2
 hf_transfer
@@ -26,7 +26,7 @@ sentencepiece
 gradio==3.50.2
 
 modal==0.70.5
-pydantic==2.6.3
+pydantic==2.10.6
 addict
 fire
 PyYAML>=6.0
@@ -62,4 +62,5 @@ antlr4-python3-runtime==4.13.2
 torchao==0.7.0
 schedulefree==1.3.0
 
-axolotl-contribs-lgpl==0.0.3
+axolotl-contribs-lgpl==0.0.6
+axolotl-contribs-mit==0.0.3

scripts/chat_datasets.py

@@ -1,6 +1,7 @@
 """
 helper script to parse chat datasets into a usable yaml
 """
+
 import click
 import yaml
 from datasets import load_dataset
@@ -31,27 +32,26 @@ def parse_dataset(dataset=None, split="train"):
     ds_cfg["field_messages"] = field_messages
 
     message_fields = features[field_messages][0].keys()
-    message_field_role = None
+
+    message_property_mappings = {"role": None, "content": None}
     for key in ["from", "role"]:
         if key in message_fields:
-            message_field_role = key
+            message_property_mappings["role"] = key
             break
-    if not message_field_role:
+    if not message_property_mappings["role"]:
         raise ValueError(
             f'No role field found in messages: {", ".join(message_fields)}'
         )
-    ds_cfg["message_field_role"] = message_field_role
 
-    message_field_content = None
     for key in ["content", "text", "value"]:
         if key in message_fields:
-            message_field_content = key
+            message_property_mappings["content"] = key
             break
-    if not message_field_content:
+    if not message_property_mappings["content"]:
         raise ValueError(
             f'No content field found in messages: {", ".join(message_fields)}'
         )
-    ds_cfg["message_field_content"] = message_field_content
+    ds_cfg["message_property_mappings"] = message_property_mappings
 
     print(yaml.dump({"datasets": [ds_cfg]}))
 

scripts/cutcrossentropy_install.py

@@ -1,4 +1,5 @@
 """Script to output the correct installation command for cut-cross-entropy."""
+
 import importlib.util
 import sys
 
@@ -24,5 +25,5 @@ if cce_spec:
 
 print(
     UNINSTALL_PREFIX
-    + 'pip install "cut-cross-entropy @ git+https://github.com/apple/ml-cross-entropy.git@9c297c905f55b73594b5d650722d1e78183b77bd"'
+    + 'pip install "cut-cross-entropy[transformers] @ git+https://github.com/apple/ml-cross-entropy.git@24fbe4b5dab9a6c250a014573613c1890190536c"'
 )

setup.py

@@ -71,12 +71,15 @@ def parse_requirements():
             else:
                 raise ValueError("Invalid version format")
 
-            if (major, minor) >= (2, 5):
+            if (major, minor) >= (2, 6):
+                _install_requires.pop(_install_requires.index(xformers_version))
+                _install_requires.append("xformers==0.0.29.post2")
+            elif (major, minor) >= (2, 5):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 if patch == 0:
                     _install_requires.append("xformers==0.0.28.post2")
                 else:
-                    _install_requires.append("xformers==0.0.28.post3")
+                    _install_requires.append("xformers>=0.0.28.post3")
                 _install_requires.pop(_install_requires.index(autoawq_version))
             elif (major, minor) >= (2, 4):
                 if patch == 0:
@@ -122,10 +125,10 @@ setup(
     },
     extras_require={
         "flash-attn": [
-            "flash-attn==2.7.0.post2",
+            "flash-attn==2.7.4.post1",
         ],
         "deepspeed": [
-            "deepspeed==0.16.1",
+            "deepspeed==0.16.4",
             "deepspeed-kernels",
         ],
         "mamba-ssm": [
@@ -153,5 +156,8 @@ setup(
         "ray": [
             "ray[train]",
         ],
+        "vllm": [
+            "vllm==0.7.2",
+        ],
     },
 )

src/axolotl/__init__.py

@@ -4,4 +4,4 @@ import pkgutil
 
 __path__ = pkgutil.extend_path(__path__, __name__)  # Make this a namespace package
 
-__version__ = "0.6.0"
+__version__ = "0.8.0.dev0"

src/axolotl/cli/cloud/__init__.py

@@ -1,6 +1,7 @@
 """
 launch axolotl in supported cloud platforms
 """
+
 from pathlib import Path
 from typing import Union
 
@@ -35,13 +36,18 @@ def do_cli_train(
     cloud_config: Union[Path, str],
     config: Union[Path, str],
     accelerate: bool = True,
+    cwd=None,
+    **kwargs,
 ) -> None:
     print_axolotl_text_art()
     cloud_cfg = load_cloud_cfg(cloud_config)
     cloud = ModalCloud(cloud_cfg)
     with open(config, "r", encoding="utf-8") as file:
         config_yaml = file.read()
-    cloud.train(config_yaml, accelerate=accelerate)
+    local_dirs = {}
+    if cwd and not Path(cwd).joinpath("src", "axolotl").exists():
+        local_dirs = {"/workspace/mounts": cwd}
+    cloud.train(config_yaml, accelerate=accelerate, local_dirs=local_dirs, **kwargs)
 
 
 def do_cli_lm_eval(

src/axolotl/cli/cloud/base.py

@@ -1,6 +1,7 @@
 """
 base class for cloud platforms from cli
 """
+
 from abc import ABC, abstractmethod
 
 

src/axolotl/cli/cloud/modal_.py

@@ -1,12 +1,14 @@
 """
 Modal Cloud support from CLI
 """
+
 import copy
 import json
 import os
 import subprocess  # nosec B404
 from pathlib import Path
 from random import randint
+from typing import Optional
 
 import modal
 
@@ -22,8 +24,18 @@ def run_cmd(cmd: str, run_folder: str, volumes=None):
 
     # modal workaround so it doesn't use the automounted axolotl
     new_env = copy.deepcopy(os.environ)
+
     if "PYTHONPATH" in new_env:
-        del new_env["PYTHONPATH"]
+        paths = ["/workspace/mounts"]
+        for sub_python_path_str in new_env["PYTHONPATH"].split(":"):
+            sub_python_path = Path(sub_python_path_str)
+            if not sub_python_path.joinpath("src", "axolotl").exists():
+                # we don't want to use the automounted axolotl or unexpected behavior happens
+                paths.append(str(sub_python_path))
+        if paths:
+            new_env["PYTHONPATH"] = ":".join(paths)
+        else:
+            del new_env["PYTHONPATH"]
 
     # Propagate errors from subprocess.
     if exit_code := subprocess.call(  # nosec B603
@@ -102,7 +114,7 @@ class ModalCloud(Cloud):
                 [
                     # Random id for cache busting of branch commits
                     f"RUN echo '{str(randint(0, 1000000))}'",  # nosec B311
-                    f"RUN cd /workspace/axolotl && git fetch && git checkout {self.config.branch}",
+                    f"RUN cd /workspace/axolotl && git fetch && git checkout {self.config.branch} && git pull",
                 ]
             )
 
@@ -112,8 +124,6 @@ class ModalCloud(Cloud):
         if env := self.get_env():
             image = image.env(env)
 
-        image = image.pip_install("fastapi==0.110.0", "pydantic==2.6.3")
-
         return image
 
     def get_secrets(self):
@@ -203,9 +213,12 @@ class ModalCloud(Cloud):
             memory = int(self.config.memory)
         return 1024 * memory
 
-    def get_train_env(self):
+    def get_train_env(self, local_dirs=None):
+        image = self.get_image()
+        for mount, local_dir in (local_dirs or {}).items():
+            image = image.add_local_dir(local_dir, mount)
         return self.app.function(
-            image=self.get_image(),
+            image=image,
             volumes={k: v[0] for k, v in self.volumes.items()},
             cpu=16.0,
             gpu=self.get_train_gpu(),
@@ -214,14 +227,21 @@ class ModalCloud(Cloud):
             secrets=self.get_secrets(),
         )
 
-    def train(self, config_yaml: str, accelerate: bool = True):
-        modal_fn = self.get_train_env()(_train)
+    def train(
+        self,
+        config_yaml: str,
+        accelerate: bool = True,
+        local_dirs: Optional[dict[str, str]] = None,
+        **kwargs,
+    ):
+        modal_fn = self.get_train_env(local_dirs)(_train)
         with modal.enable_output():
             with self.app.run(detach=True):
                 modal_fn.remote(
                     config_yaml,
                     accelerate=accelerate,
                     volumes={k: v[0] for k, v in self.volumes.items()},
+                    **kwargs,
                 )
 
     def lm_eval(self, config_yaml: str):
@@ -239,44 +259,43 @@ class ModalCloud(Cloud):
 
 
 def _preprocess(config_yaml: str, volumes=None):
-    Path("/workspace/artifacts/axolotl").mkdir(parents=True, exist_ok=True)
-    with open(
-        "/workspace/artifacts/axolotl/config.yaml", "w", encoding="utf-8"
-    ) as f_out:
+    Path("/workspace/mounts").mkdir(parents=True, exist_ok=True)
+    with open("/workspace/mounts/config.yaml", "w", encoding="utf-8") as f_out:
         f_out.write(config_yaml)
-    run_folder = "/workspace/artifacts/axolotl"
+    run_folder = "/workspace/mounts"
     run_cmd(
-        "axolotl preprocess /workspace/artifacts/axolotl/config.yaml --dataset-processes=8",
+        "axolotl preprocess /workspace/mounts/config.yaml --dataset-processes=8",
         run_folder,
         volumes,
     )
 
 
-def _train(config_yaml: str, accelerate: bool = True, volumes=None):
-    with open(
-        "/workspace/artifacts/axolotl/config.yaml", "w", encoding="utf-8"
-    ) as f_out:
+def _train(config_yaml: str, accelerate: bool = True, volumes=None, **kwargs):
+    Path("/workspace/mounts").mkdir(parents=True, exist_ok=True)
+    with open("/workspace/mounts/config.yaml", "w", encoding="utf-8") as f_out:
         f_out.write(config_yaml)
-    run_folder = "/workspace/artifacts/axolotl"
+    run_folder = "/workspace/mounts"
     if accelerate:
         accelerate_args = "--accelerate"
     else:
         accelerate_args = "--no-accelerate"
+    num_processes_args = ""
+    if num_processes := kwargs.pop("num_processes", None):
+        num_processes_args = f"--num-processes {num_processes}"
     run_cmd(
-        f"axolotl train {accelerate_args} /workspace/artifacts/axolotl/config.yaml",
+        f"axolotl train {accelerate_args} {num_processes_args} /workspace/mounts/config.yaml",
         run_folder,
         volumes,
     )
 
 
 def _lm_eval(config_yaml: str, volumes=None):
-    with open(
-        "/workspace/artifacts/axolotl/config.yaml", "w", encoding="utf-8"
-    ) as f_out:
+    Path("/workspace/mounts").mkdir(parents=True, exist_ok=True)
+    with open("/workspace/mounts/config.yaml", "w", encoding="utf-8") as f_out:
         f_out.write(config_yaml)
-    run_folder = "/workspace/artifacts/axolotl"
+    run_folder = "/workspace/mounts"
     run_cmd(
-        "axolotl lm-eval /workspace/artifacts/axolotl/config.yaml",
+        "axolotl lm-eval /workspace/mounts/config.yaml",
         run_folder,
         volumes,
     )

src/axolotl/cli/convert_linear_attention.py

@@ -1,135 +0,0 @@
-"""CLI to run training on a model."""
-
-import logging
-import os
-from pathlib import Path
-from typing import Union
-
-import fire
-from dotenv import load_dotenv
-from transformers.hf_argparser import HfArgumentParser
-
-from axolotl.cli.args import TrainerCliArgs
-from axolotl.cli.art import print_axolotl_text_art
-from axolotl.cli.checks import check_accelerate_default_config, check_user_token
-from axolotl.cli.config import load_cfg
-from axolotl.cli.utils import load_model_and_tokenizer
-from axolotl.common.datasets import load_datasets
-from axolotl.integrations.base import PluginManager
-from axolotl.integrations.lolcats.linear_llama.configuration_linear_llama import (
-    LinearLlamaConfig,
-)
-from axolotl.integrations.lolcats.linear_llama.modeling_linear_llama import (
-    LinearLlamaForCausalLM,
-)
-from axolotl.utils.dict import DictDefault
-from axolotl.utils.models import load_model_config
-from axolotl.utils.trainer import setup_trainer
-
-LOG = logging.getLogger(__name__)
-
-
-def do_linearize(cfg: DictDefault, cli_args: TrainerCliArgs) -> None:
-    """
-    Convert attention to linear attention and perform attention transfer via distillation.
-    """
-    print_axolotl_text_art()
-    check_accelerate_default_config()
-    check_user_token()
-
-    # ensure quantization and peft are turned off (due to how we need to re-apply peft later)
-    cfg.load_in_8bit = False
-    cfg.load_in_4bit = False
-    cfg.adapter = None
-
-    # load model
-    model, tokenizer = load_model_and_tokenizer(cfg=cfg)
-
-    # freeze model
-    for p in model.parameters():
-        p.requires_grad = False
-
-    # convert to linear llama
-    linear_llama_config = LinearLlamaConfig.from_llama(
-        model.config, cfg.attention_config
-    )
-    model = LinearLlamaForCausalLM.from_llama(
-        model, config=linear_llama_config, train_attention=True
-    )
-
-    # set save_path, save tokenizer and model config.
-    save_path = str(os.path.join(cfg.output_dir, "distilled"))
-    tokenizer.save_pretrained(save_path)
-    if hasattr(model, "config"):
-        model.config.save_pretrained(save_path)
-
-    # Get datasets
-    dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
-    train_dataset = dataset_meta.train_dataset
-    eval_dataset = dataset_meta.eval_dataset
-    total_num_steps = dataset_meta.total_num_steps
-
-    # toggle attention to be trainable
-    model.toggle_attention(train=True)
-
-    # Setup trainer
-    trainer = setup_trainer(
-        cfg=cfg,
-        train_dataset=train_dataset,
-        eval_dataset=eval_dataset,
-        model=(model, None, None),
-        tokenizer=tokenizer,
-        processor=None,
-        total_num_steps=total_num_steps,
-    )
-
-    # train
-    trainer.train(resume_from_checkpoint=cfg.resume_from_checkpoint)
-
-    # drop base_attention + remove training attn
-    model.toggle_attention(train=False)
-    model.remove_base_attention()
-
-    # NOTE: If in peft mode, consider whether to auto-merge
-
-    # save model
-    safe_serialization = cfg.save_safetensors is True
-    # NOTE: may need to consider other ways of saving due to multi-gpu etc
-    model.save_pretrained(save_path, safe_serialization=safe_serialization)
-
-    # cleanup
-    plugin_manager = PluginManager.get_instance()
-
-    del model
-    del tokenizer
-
-    plugin_manager.post_train_unload(cfg)
-
-
-def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs) -> None:
-    """
-    Parses `axolotl` config, CLI args, and calls `do_train`.
-
-    Args:
-        config: Path to `axolotl` config YAML file.
-        kwargs: Additional keyword arguments to override config file values.
-    """
-    # load cfg, force linearize and add plugin to linearize
-    parsed_cfg = load_cfg(
-        config,
-        linearize=True,
-        plugins=["axolotl.integrations.lolcats.LinearizePlugin"],
-        **kwargs,
-    )
-
-    parser = HfArgumentParser(TrainerCliArgs)
-    parsed_cli_args, _ = parser.parse_args_into_dataclasses(
-        return_remaining_strings=True
-    )
-
-    do_linearize(parsed_cfg, parsed_cli_args)
-
-
-if __name__ == "__main__":
-    load_dotenv()
-    fire.Fire(do_cli)

src/axolotl/cli/main.py

@@ -1,20 +1,21 @@
 """Click CLI definitions for various axolotl commands."""
+
 # pylint: disable=redefined-outer-name
 
 import logging
-import random
+import os
 import subprocess  # nosec B404
 import tempfile
-from copy import deepcopy
-from itertools import product
 from pathlib import Path
 from typing import Optional
 
 import click
 import yaml
+from dotenv import load_dotenv
 
 import axolotl
 from axolotl.cli.args import EvaluateCliArgs, PreprocessCliArgs, TrainerCliArgs
+from axolotl.cli.sweeps import generate_sweep_configs
 from axolotl.cli.utils import (
     add_options_from_config,
     add_options_from_dataclass,
@@ -27,76 +28,6 @@ from axolotl.utils import set_pytorch_cuda_alloc_conf
 from axolotl.utils.config.models.input.v0_4_1 import AxolotlInputConfig
 
 
-def generate_sweep_configs(base_config, sweeps_config):
-    """
-    Recursively generates all possible configurations by applying sweeps to the base config.
-
-    Args:
-        base_config (dict): The original configuration dictionary
-        sweeps_config (dict): Dictionary where keys are parameters and values are either:
-            - lists of values to sweep independently
-            - or for paired values, a list of dicts under the '_' key
-
-    Returns:
-        list: List of all possible configuration dictionaries
-
-    Example:
-        sweeps_config = {
-            'learning_rate': [0.1, 0.01],
-            '_': [
-                {'load_in_8bit': True, 'adapter': 'lora'},
-                {'load_in_4bit': True, 'adapter': 'qlora'}
-            ]
-        }
-    """
-    # Separate paired values from regular sweeps
-    paired_values = sweeps_config.get("_", [])
-    regular_sweeps = {k: v for k, v in sweeps_config.items() if k != "_"}
-
-    # Process regular sweeps
-    param_names = list(regular_sweeps.keys())
-    param_values = list(regular_sweeps.values())
-
-    # Generate combinations for regular sweeps
-    regular_combinations = list(product(*param_values)) if param_values else [()]
-
-    # Combine regular sweeps with paired values
-    all_combinations = []
-    for reg_combo in regular_combinations:
-        if paired_values:
-            for paired_set in paired_values:
-                new_config = {}
-                # new_config = deepcopy(base_config)
-                # Combine regular parameters with paired parameters
-                full_combo = {**dict(zip(param_names, reg_combo)), **paired_set}
-                for param_name, param_value in full_combo.items():
-                    new_config[param_name] = param_value
-                print(new_config)
-                all_combinations.append(new_config)
-        else:
-            # If no paired values, just use regular combinations
-            # new_config = deepcopy(base_config)
-            new_config = {}
-            for param_name, param_value in zip(param_names, reg_combo):
-                new_config[param_name] = param_value
-            print(new_config)
-            all_combinations.append(new_config)
-
-    # randomize the order of trials
-    random.seed(42)
-    random.shuffle(all_combinations)
-
-    # Generate a new config for each combination
-    result_configs = []
-    for combination in all_combinations:
-        new_config = deepcopy(base_config)
-        for param_name, param_value in combination.items():
-            new_config[param_name] = param_value
-        result_configs.append(new_config)
-
-    return result_configs
-
-
 @click.group()
 @click.version_option(version=axolotl.__version__, prog_name="axolotl")
 def cli():
@@ -165,7 +96,6 @@ def train(
     """
     # Enable expandable segments for cuda allocation to improve VRAM usage
     set_pytorch_cuda_alloc_conf()
-    from axolotl.cli.cloud import do_cli_train
 
     if "use_ray" in kwargs and kwargs["use_ray"]:
         accelerate = False
@@ -199,7 +129,16 @@ def train(
         try:
             if accelerate:
                 if cloud:
-                    do_cli_train(cloud_config=cloud, config=config, accelerate=True)
+                    from axolotl.cli.cloud import do_cli_train
+
+                    cwd = os.getcwd()
+                    do_cli_train(
+                        cloud_config=cloud,
+                        config=config,
+                        accelerate=True,
+                        cwd=cwd,
+                        **kwargs,
+                    )
                 else:
                     accelerate_args = []
                     if "main_process_port" in kwargs:
@@ -208,7 +147,7 @@ def train(
                         accelerate_args.append(str(main_process_port))
                     if "num_processes" in kwargs:
                         num_processes = kwargs.pop("num_processes", None)
-                        accelerate_args.append("--num-processes")
+                        accelerate_args.append("--num_processes")
                         accelerate_args.append(str(num_processes))
 
                     base_cmd = ["accelerate", "launch"]
@@ -220,7 +159,11 @@ def train(
                     subprocess.run(cmd, check=True)  # nosec B603
             else:
                 if cloud:
-                    do_cli_train(cloud_config=cloud, config=config, accelerate=False)
+                    from axolotl.cli.cloud import do_cli_train
+
+                    do_cli_train(
+                        cloud_config=cloud, config=config, accelerate=False, **kwargs
+                    )
                 else:
                     from axolotl.cli.train import do_cli
 
@@ -381,4 +324,5 @@ def main():
 
 
 if __name__ == "__main__":
+    load_dotenv()
     main()

src/axolotl/cli/sweeps.py

@@ -0,0 +1,77 @@
+"""Utilities for handling sweeps over configs for axolotl train CLI command"""
+
+import random
+from copy import deepcopy
+from itertools import product
+
+
+def generate_sweep_configs(
+    base_config: dict[str, list], sweeps_config: dict[str, list]
+) -> list[dict[str, list]]:
+    """
+    Recursively generates all possible configurations by applying sweeps to the base config.
+
+    Args:
+        base_config (dict): The original configuration dictionary
+        sweeps_config (dict): Dictionary where keys are parameters and values are either:
+            - lists of values to sweep independently
+            - or for paired values, a list of dicts under the '_' key
+
+    Returns:
+        list: List of all possible configuration dictionaries
+
+    Example:
+        sweeps_config = {
+            'learning_rate': [0.1, 0.01],
+            '_': [
+                {'load_in_8bit': True, 'adapter': 'lora'},
+                {'load_in_4bit': True, 'adapter': 'qlora'}
+            ]
+        }
+    """
+    # Separate paired values from regular sweeps
+    paired_values = sweeps_config.get("_", [])
+    regular_sweeps = {k: v for k, v in sweeps_config.items() if k != "_"}
+
+    # Process regular sweeps
+    param_names = list(regular_sweeps.keys())
+    param_values = list(regular_sweeps.values())
+
+    # Generate combinations for regular sweeps
+    regular_combinations = list(product(*param_values)) if param_values else [()]
+
+    # Combine regular sweeps with paired values
+    all_combinations = []
+    for reg_combo in regular_combinations:
+        if paired_values:
+            for paired_set in paired_values:
+                new_config = {}
+                # new_config = deepcopy(base_config)
+                # Combine regular parameters with paired parameters
+                full_combo = {**dict(zip(param_names, reg_combo)), **paired_set}
+                for param_name, param_value in full_combo.items():
+                    new_config[param_name] = param_value
+                print(new_config)
+                all_combinations.append(new_config)
+        else:
+            # If no paired values, just use regular combinations
+            # new_config = deepcopy(base_config)
+            new_config = {}
+            for param_name, param_value in zip(param_names, reg_combo):
+                new_config[param_name] = param_value
+            print(new_config)
+            all_combinations.append(new_config)
+
+    # randomize the order of trials
+    random.seed(42)
+    random.shuffle(all_combinations)
+
+    # Generate a new config for each combination
+    result_configs = []
+    for combination in all_combinations:
+        new_config = deepcopy(base_config)
+        for param_name, param_value in combination.items():
+            new_config[param_name] = param_value
+        result_configs.append(new_config)
+
+    return result_configs

src/axolotl/cli/train.py

@@ -1,6 +1,7 @@
 """CLI to run training on a model."""
 
 import logging
+import os
 from pathlib import Path
 from typing import Union
 
@@ -34,18 +35,20 @@ def do_train(cfg: DictDefault, cli_args: TrainerCliArgs) -> None:
     """
     print_axolotl_text_art()
     check_accelerate_default_config()
-    check_user_token()
+    if int(os.getenv("LOCAL_RANK", "0")) == 0:
+        check_user_token()
 
     if cfg.rl:
         dataset_meta = load_preference_datasets(cfg=cfg, cli_args=cli_args)
     else:
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-    model, tokenizer = train(cfg=cfg, dataset_meta=dataset_meta)
+    model, tokenizer, trainer = train(cfg=cfg, dataset_meta=dataset_meta)
     plugin_manager = PluginManager.get_instance()
 
     del model
     del tokenizer
+    del trainer
 
     plugin_manager.post_train_unload(cfg)
 

src/axolotl/cli/utils.py

@@ -5,7 +5,6 @@ import dataclasses
 import hashlib
 import json
 import logging
-import typing
 from functools import wraps
 from pathlib import Path
 from types import NoneType
@@ -24,7 +23,7 @@ configure_logging()
 LOG = logging.getLogger(__name__)
 
 
-def strip_optional_type(field_type: type | typing._SpecialForm | None):
+def strip_optional_type(field_type: type | str | None):
     """
     Extracts the non-`None` type from an `Optional` / `Union` type.
 

src/axolotl/common/datasets.py

@@ -24,8 +24,8 @@ class TrainDatasetMeta:
     """Dataclass with fields for training and validation datasets and metadata."""
 
     train_dataset: Dataset
-    eval_dataset: Optional[Dataset] = None
-    total_num_steps: Optional[int] = None
+    eval_dataset: Dataset | None = None
+    total_num_steps: int | None = None
 
 
 def sample_dataset(dataset: Dataset, num_samples: int) -> Dataset:
@@ -122,9 +122,11 @@ def load_preference_datasets(
         `total_num_steps`.
     """
     train_dataset, eval_dataset = load_prepare_preference_datasets(cfg)
-    total_num_steps = int(
+    total_num_steps: Optional[int] = int(
         math.ceil(len(train_dataset) * cfg.num_epochs / cfg.batch_size)
     )
+    if cfg.rl == "grpo":
+        total_num_steps = None
 
     if cli_args.debug or cfg.debug:
         LOG.info("check_dataset_labels...")

src/axolotl/convert.py

@@ -1,6 +1,5 @@
 """Module containing File Reader, File Writer, Json Parser, and Jsonl Serializer classes"""
 
-
 import json
 import sys
 

src/axolotl/core/chat/format/chatml.py

@@ -1,6 +1,7 @@
 """
 ChatML transformation functions for MessageContents
 """
+
 from typing import Optional
 
 from ..messages import MessageContents, Messages

src/axolotl/core/chat/format/llama3x.py

@@ -1,6 +1,7 @@
 """
 Llama 3.x chat formatting functions for MessageContents
 """
+
 from typing import Optional
 
 from ..messages import MessageContents, Messages

src/axolotl/core/chat/format/shared.py

@@ -1,6 +1,7 @@
 """
 shared functions for format transforms
 """
+
 from axolotl.core.chat.messages import MessageContents, Messages
 
 

src/axolotl/core/chat/messages.py

@@ -1,6 +1,7 @@
 """
 internal message representations of chat messages
 """
+
 import json
 from enum import Enum
 from typing import Any, Callable, List, Optional, Union

src/axolotl/core/datasets/chat.py

@@ -1,6 +1,7 @@
 """
 chat dataset module
 """
+
 import os
 from typing import Callable, Optional, Union
 
@@ -43,7 +44,7 @@ class TokenizedChatDataset(Dataset):
         process_or_cpu_count: int = (
             process_count or os.cpu_count()  # type: ignore[assignment]
         )
-        num_proc = min(64, process_or_cpu_count)
+        num_proc = min(32, process_or_cpu_count)
         features = data.features.keys()
         tokenized_data = data.map(
             map_fn,

src/axolotl/core/datasets/transforms/chat_builder.py

@@ -1,6 +1,7 @@
 """
 This module contains a function that builds a transform that takes a row from the dataset and converts it to a Chat.
 """
+
 from typing import Any, Mapping, Union
 
 

src/axolotl/core/trainer_builder.py

@@ -35,11 +35,11 @@ from transformers import (
     EarlyStoppingCallback,
     TrainerCallback,
 )
+from transformers.training_args import OptimizerNames
 from trl.trainer.utils import RewardDataCollatorWithPadding
 
 from axolotl.core.trainers.base import (
     AxolotlCPOTrainer,
-    AxolotlDPOTrainer,
     AxolotlKTOTrainer,
     AxolotlMambaTrainer,
     AxolotlORPOTrainer,
@@ -48,9 +48,11 @@ from axolotl.core.trainers.base import (
     AxolotlTrainer,
     ReLoRATrainer,
 )
+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,
-    AxolotlDPOConfig,
     AxolotlKTOConfig,
     AxolotlORPOConfig,
     AxolotlPRMConfig,
@@ -83,6 +85,7 @@ from axolotl.utils.collators import (
     V2BatchSamplerDataCollatorForSeq2Seq,
 )
 from axolotl.utils.collators.mm_chat import MultiModalChatDataCollator
+from axolotl.utils.config.models.input.v0_4_1 import CustomSupportedOptimizers
 from axolotl.utils.models import ensure_dtype
 
 try:
@@ -90,13 +93,11 @@ try:
 except ImportError:
     pass
 
-LOG = logging.getLogger("axolotl.core.trainer_builder")
+LOG = logging.getLogger(__name__)
 
 
 class TrainerBuilderBase(abc.ABC):
-    """
-    Base class for trainer builder
-    """
+    """Base class for trainer builder."""
 
     _train_dataset = None
     _eval_dataset = None
@@ -109,9 +110,9 @@ class TrainerBuilderBase(abc.ABC):
         self.tokenizer = tokenizer
         self.processor = processor
 
-        # in case the model supports tagging, add the axolotl tag.
+        # 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 instad of  trainer.push_to_hub.
+        # model.push_to_hub instead of trainer.push_to_hub.
         if hasattr(model, "add_model_tags"):
             model.add_model_tags(["axolotl"])
 
@@ -226,8 +227,8 @@ class TrainerBuilderBase(abc.ABC):
 
 class HFCausalTrainerBuilder(TrainerBuilderBase):
     """
-    Build the HuggingFace training args/trainer for causal models
-    and reward modelling using TRL.
+    Build the HuggingFace training args/trainer for causal models and reward modeling
+    using TRL.
     """
 
     def get_callbacks(self):
@@ -329,6 +330,12 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         )
 
         training_arguments_kwargs = {}
+
+        if self.cfg.include_tokens_per_second is not None:
+            training_arguments_kwargs["include_tokens_per_second"] = (
+                self.cfg.include_tokens_per_second
+            )
+
         if self.cfg.bf16 == "full":
             training_arguments_kwargs["bf16_full_eval"] = True
         else:
@@ -344,13 +351,13 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             training_arguments_kwargs["seed"] = self.cfg.seed
 
         if self.cfg.gradient_checkpointing:
-            training_arguments_kwargs[
-                "gradient_checkpointing"
-            ] = self.cfg.gradient_checkpointing
+            training_arguments_kwargs["gradient_checkpointing"] = (
+                self.cfg.gradient_checkpointing
+            )
             if self.cfg.gradient_checkpointing_kwargs is not None:
-                training_arguments_kwargs[
-                    "gradient_checkpointing_kwargs"
-                ] = self.cfg.gradient_checkpointing_kwargs
+                training_arguments_kwargs["gradient_checkpointing_kwargs"] = (
+                    self.cfg.gradient_checkpointing_kwargs
+                )
         if self.cfg.fsdp:
             training_arguments_kwargs["fsdp"] = self.cfg.fsdp
             if self.cfg.fsdp_config:
@@ -366,9 +373,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             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
+            training_arguments_kwargs["lr_quadratic_warmup"] = (
+                self.cfg.lr_quadratic_warmup
+            )
 
         if self.cfg.adam_beta1:
             training_arguments_kwargs["adam_beta1"] = self.cfg.adam_beta1
@@ -392,28 +399,28 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             training_arguments_kwargs["save_safetensors"] = self.cfg.save_safetensors
 
         if self.cfg.dataloader_pin_memory is not None:
-            training_arguments_kwargs[
-                "dataloader_pin_memory"
-            ] = self.cfg.dataloader_pin_memory
+            training_arguments_kwargs["dataloader_pin_memory"] = (
+                self.cfg.dataloader_pin_memory
+            )
         if self.cfg.dataloader_num_workers is not None:
-            training_arguments_kwargs[
-                "dataloader_num_workers"
-            ] = self.cfg.dataloader_num_workers
+            training_arguments_kwargs["dataloader_num_workers"] = (
+                self.cfg.dataloader_num_workers
+            )
         if self.cfg.dataloader_prefetch_factor is not None:
-            training_arguments_kwargs[
-                "dataloader_prefetch_factor"
-            ] = self.cfg.dataloader_prefetch_factor
+            training_arguments_kwargs["dataloader_prefetch_factor"] = (
+                self.cfg.dataloader_prefetch_factor
+            )
         if self.cfg.dataloader_drop_last is not None:
-            training_arguments_kwargs[
-                "dataloader_drop_last"
-            ] = self.cfg.dataloader_drop_last
+            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
+            training_arguments_kwargs["remove_unused_columns"] = (
+                self.cfg.remove_unused_columns
+            )
 
         if not self.cfg.test_datasets and self.cfg.val_set_size == 0:
             # no eval set, so don't eval
@@ -445,9 +452,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         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
+            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
 
@@ -460,13 +467,13 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 )
                 training_arguments_kwargs["torch_compile"] = self.cfg.torch_compile
                 if self.cfg.torch_compile_backend:
-                    training_arguments_kwargs[
-                        "torch_compile_backend"
-                    ] = self.cfg.torch_compile_backend
+                    training_arguments_kwargs["torch_compile_backend"] = (
+                        self.cfg.torch_compile_backend
+                    )
                 if self.cfg.torch_compile_mode:
-                    training_arguments_kwargs[
-                        "torch_compile_mode"
-                    ] = self.cfg.torch_compile_mode
+                    training_arguments_kwargs["torch_compile_mode"] = (
+                        self.cfg.torch_compile_mode
+                    )
 
         # DDP Config
         if self.cfg.ddp_timeout:
@@ -475,32 +482,32 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         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
+            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_steps"] = (
             total_num_steps if self.cfg.max_steps else -1
         )
         training_arguments_kwargs["max_seq_length"] = self.cfg.sequence_len
-        training_arguments_kwargs[
-            "per_device_train_batch_size"
-        ] = self.cfg.micro_batch_size
+        training_arguments_kwargs["per_device_train_batch_size"] = (
+            self.cfg.micro_batch_size
+        )
         if self.cfg.eval_batch_size:
-            training_arguments_kwargs[
-                "per_device_eval_batch_size"
-            ] = self.cfg.eval_batch_size
+            training_arguments_kwargs["per_device_eval_batch_size"] = (
+                self.cfg.eval_batch_size
+            )
         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[
-            "gradient_accumulation_steps"
-        ] = self.cfg.gradient_accumulation_steps
-        training_arguments_kwargs[
-            "eval_accumulation_steps"
-        ] = self.cfg.gradient_accumulation_steps
+            training_arguments_kwargs["auto_find_batch_size"] = (
+                self.cfg.auto_find_batch_size
+            )
+        training_arguments_kwargs["gradient_accumulation_steps"] = (
+            self.cfg.gradient_accumulation_steps
+        )
+        training_arguments_kwargs["eval_accumulation_steps"] = (
+            self.cfg.gradient_accumulation_steps
+        )
         training_arguments_kwargs["num_train_epochs"] = self.cfg.num_epochs
         training_arguments_kwargs["learning_rate"] = self.cfg.learning_rate
         training_arguments_kwargs["output_dir"] = self.cfg.output_dir
@@ -544,34 +551,12 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             training_arguments_kwargs["run_name"] = self.cfg.mlflow_run_name
         else:
             training_arguments_kwargs["run_name"] = None
-        training_arguments_kwargs["optim"] = (
-            self.cfg.optimizer if self.cfg.optimizer else "adamw_hf"
-        )
-        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_arguments_kwargs["optim_args"] = optim_args
-        if self.cfg.optim_target_modules:
-            training_arguments_kwargs[
-                "optim_target_modules"
-            ] = self.cfg.optim_target_modules
-        training_arguments_kwargs["loraplus_lr_ratio"] = self.cfg.loraplus_lr_ratio
-        training_arguments_kwargs[
-            "loraplus_lr_embedding"
-        ] = self.cfg.loraplus_lr_embedding
-        training_arguments_kwargs["embedding_lr"] = self.cfg.embedding_lr
-        training_arguments_kwargs["embedding_lr_scale"] = self.cfg.embedding_lr_scale
-        training_arguments_kwargs["lr_groups"] = self.cfg.lr_groups
 
-        if self.cfg.lr_scheduler in ["one_cycle", "log_sweep"]:
+        if self.cfg.lr_scheduler in ["one_cycle", "rex", "log_sweep"]:
             training_arguments_kwargs["lr_scheduler_type"] = "cosine"
-            training_arguments_kwargs[
-                "alternate_lr_scheduler_type"
-            ] = self.cfg.lr_scheduler
+            training_arguments_kwargs["alternate_lr_scheduler_type"] = (
+                self.cfg.lr_scheduler
+            )
         else:
             training_arguments_kwargs["lr_scheduler_type"] = (
                 self.cfg.lr_scheduler if self.cfg.lr_scheduler else "cosine"
@@ -580,9 +565,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             self.cfg.lr_scheduler_kwargs if self.cfg.lr_scheduler_kwargs else {}
         )
         training_arguments_kwargs["cosine_min_lr_ratio"] = self.cfg.cosine_min_lr_ratio
-        training_arguments_kwargs[
-            "cosine_constant_lr_ratio"
-        ] = self.cfg.cosine_constant_lr_ratio
+        training_arguments_kwargs["cosine_constant_lr_ratio"] = (
+            self.cfg.cosine_constant_lr_ratio
+        )
         training_arguments_kwargs["weight_decay"] = (
             self.cfg.weight_decay if self.cfg.weight_decay is not None else 0.0
         )
@@ -595,40 +580,40 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             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
+            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
+            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
+            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
+            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
+                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
+                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["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
@@ -641,63 +626,128 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 tokenizer=self.tokenizer,
             )
 
-        if self.cfg.rl == "orpo":
-            training_arguments_kwargs["orpo_alpha"] = self.cfg.orpo_alpha
-
         if self.cfg.neftune_noise_alpha is not None:
-            training_arguments_kwargs[
-                "neftune_noise_alpha"
-            ] = self.cfg.neftune_noise_alpha
+            training_arguments_kwargs["neftune_noise_alpha"] = (
+                self.cfg.neftune_noise_alpha
+            )
 
         trainer_kwargs = {}
 
         if self.cfg.reward_model:
-            trainer_kwargs["max_length"] = self.cfg.sequence_len
+            training_arguments_kwargs["max_length"] = self.cfg.sequence_len
 
-        # pylint: disable=duplicate-code
-        if self.cfg.optimizer in [
-            "optimi_adamw",
-            "ao_adamw_4bit",
-            "ao_adamw_8bit",
-            "ao_adamw_fp8",
-            "adopt_adamw",
-        ]:
-            # Set default so transformers doesn't throw
-            training_arguments_kwargs["optim"] = "adamw_hf"
-            training_arguments_kwargs["alternate_optimizer"] = self.cfg.optimizer
+        # Handle custom optimizer
+        custom_supported_optimizers = [opt.value for opt in CustomSupportedOptimizers]
+        if self.cfg.optimizer in custom_supported_optimizers:
+            # Common optimizer kwargs
+            optimizer_kwargs = {
+                "lr": training_arguments_kwargs.get("learning_rate"),
+                "weight_decay": training_arguments_kwargs.get("weight_decay"),
+            }
 
-        if self.cfg.optimizer == "lion_pytorch":
-            from lion_pytorch import Lion
+            # Adam-specific kwargs
+            adam_kwargs = {}
+            if training_arguments_kwargs.get(
+                "adam_beta1"
+            ) and training_arguments_kwargs.get("adam_beta2"):
+                adam_kwargs["betas"] = (
+                    training_arguments_kwargs.get("adam_beta1"),
+                    training_arguments_kwargs.get("adam_beta2"),
+                )
+            if training_arguments_kwargs.get("adam_epsilon"):
+                adam_kwargs["eps"] = training_arguments_kwargs.get("adam_epsilon")
 
-            lion_kwargs = {"lr": training_arguments_kwargs["learning_rate"]}
-            if "weight_decay" in training_arguments_kwargs:
-                lion_kwargs["weight_decay"] = training_arguments_kwargs["weight_decay"]
-
-            if (
-                "adam_beta1" in training_arguments_kwargs
-                and "adam_beta2" in training_arguments_kwargs
-            ):
-                lion_kwargs["betas"] = (
-                    training_arguments_kwargs["adam_beta1"],
-                    training_arguments_kwargs["adam_beta2"],
+            if self.cfg.optimizer == "muon":
+                from axolotl.contribs.mit.muon import (  # pylint: disable=no-name-in-module
+                    MuonOptimizerFactory,
                 )
 
-            trainer_kwargs["optimizers"] = (
-                Lion(params=self.model.parameters(), **lion_kwargs),
-                None,
+                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)
+
+            # 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
+
+            trainer_kwargs["optimizer_cls_and_kwargs"] = (
+                optimizer_cls,
+                optimizer_kwargs,
             )
-            # Set default so transformers doesn't throw
-            training_arguments_kwargs["optim"] = "adamw_hf"
+        else:
+            # Use transformers' optimizer
+            training_arguments_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_arguments_kwargs["optim_args"] = optim_args
 
         if self.cfg.optimizer == "adamw_anyprecision":
             if Path(self.cfg.torchdistx_path).exists():
                 sys.path.append(self.cfg.torchdistx_path)
                 importlib.import_module("torchdistx")
 
+        if self.cfg.optim_target_modules:
+            training_arguments_kwargs["optim_target_modules"] = (
+                self.cfg.optim_target_modules
+            )
+
+        training_arguments_kwargs["embedding_lr"] = self.cfg.embedding_lr
+        training_arguments_kwargs["embedding_lr_scale"] = self.cfg.embedding_lr_scale
+
+        training_arguments_kwargs["loraplus_lr_ratio"] = self.cfg.loraplus_lr_ratio
+        training_arguments_kwargs["loraplus_lr_embedding"] = (
+            self.cfg.loraplus_lr_embedding
+        )
+        training_arguments_kwargs["lr_groups"] = self.cfg.lr_groups
+
         if self.cfg.accelerator_config:
-            training_arguments_kwargs[
-                "accelerator_config"
-            ] = self.cfg.accelerator_config
+            training_arguments_kwargs["accelerator_config"] = (
+                self.cfg.accelerator_config
+            )
 
         if self.cfg.kd_ce_alpha is not None:
             training_arguments_kwargs["kd_ce_alpha"] = self.cfg.kd_ce_alpha
@@ -706,13 +756,13 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         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
+            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
+            training_arguments_kwargs["kd_top_k_before_softmax"] = (
+                self.cfg.kd_top_k_before_softmax
+            )
 
         if self.cfg.reward_model:
             training_args_cls = AxolotlRewardConfig
@@ -868,9 +918,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
 
 
 class HFRLTrainerBuilder(TrainerBuilderBase):
-    """
-    Trainer factory class for TRL-based RLHF trainers (e.g. DPO)
-    """
+    """Trainer factory class for TRL-based RLHF trainers (e.g. DPO)"""
 
     def get_callbacks(self):
         callbacks = super().get_callbacks()
@@ -924,32 +972,32 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             self.cfg.lr_scheduler_kwargs if self.cfg.lr_scheduler_kwargs else {}
         )
         if self.cfg.remove_unused_columns is not None:
-            training_args_kwargs[
-                "remove_unused_columns"
-            ] = self.cfg.remove_unused_columns
+            training_args_kwargs["remove_unused_columns"] = (
+                self.cfg.remove_unused_columns
+            )
         else:
             training_args_kwargs["remove_unused_columns"] = False
 
         if self.cfg.dataloader_pin_memory is not None:
-            training_args_kwargs[
-                "dataloader_pin_memory"
-            ] = self.cfg.dataloader_pin_memory
+            training_args_kwargs["dataloader_pin_memory"] = (
+                self.cfg.dataloader_pin_memory
+            )
         if self.cfg.dataloader_num_workers is not None:
-            training_args_kwargs[
-                "dataloader_num_workers"
-            ] = self.cfg.dataloader_num_workers
+            training_args_kwargs["dataloader_num_workers"] = (
+                self.cfg.dataloader_num_workers
+            )
         if self.cfg.dataloader_prefetch_factor is not None:
-            training_args_kwargs[
-                "dataloader_prefetch_factor"
-            ] = self.cfg.dataloader_prefetch_factor
+            training_args_kwargs["dataloader_prefetch_factor"] = (
+                self.cfg.dataloader_prefetch_factor
+            )
         if self.cfg.gradient_checkpointing:
-            training_args_kwargs[
-                "gradient_checkpointing"
-            ] = 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
+                training_args_kwargs["gradient_checkpointing_kwargs"] = (
+                    self.cfg.gradient_checkpointing_kwargs
+                )
             else:
                 training_args_kwargs["gradient_checkpointing_kwargs"] = {
                     "use_reentrant": False
@@ -965,10 +1013,11 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             # default to saving each epoch if not defined
             training_args_kwargs["save_strategy"] = "epoch"
 
-        training_args_kwargs["dataset_num_proc"] = self.cfg.dataset_processes
+        if self.cfg.dataset_processes:
+            training_args_kwargs["dataset_num_proc"] = self.cfg.dataset_processes
 
-        if self.cfg.rl_beta:
-            training_args_kwargs["beta"] = self.cfg.rl_beta
+        if (self.cfg.trl and self.cfg.trl.beta) or self.cfg.rl_beta:
+            training_args_kwargs["beta"] = self.cfg.trl.beta or self.cfg.rl_beta
         if self.cfg.orpo_alpha:
             # trl does some odd mapping of alpha to beta to reuse the beta parameter ???
             training_args_kwargs["beta"] = self.cfg.orpo_alpha
@@ -977,6 +1026,7 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             training_args_kwargs["rpo_alpha"] = self.cfg.rpo_alpha
 
         training_args_cls = None
+        blocklist_args_kwargs = []
         if self.cfg.rl == "simpo":
             training_args_cls = AxolotlCPOConfig
             training_args_kwargs["loss_type"] = "simpo"
@@ -1001,11 +1051,15 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
                 self.cfg.kto_undesirable_weight or 1.0
             )
 
-            training_args_kwargs["dataset_num_proc"] = self.cfg.dataset_processes
             training_args_kwargs["max_length"] = self.cfg.sequence_len
             if self.cfg.max_prompt_len:
                 training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
 
+        elif self.cfg.rl == "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()
+
         else:
             training_args_cls = AxolotlDPOConfig
             if self.cfg.rl == "ipo":
@@ -1016,11 +1070,21 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             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
+                )
 
+        for blocklist_key in blocklist_args_kwargs:
+            if blocklist_key in training_args_kwargs:
+                del training_args_kwargs[blocklist_key]
+
+        max_steps = self.cfg.max_steps or total_num_steps or -1
+        training_args_kwargs["num_train_epochs"] = self.cfg.num_epochs
         training_args = training_args_cls(  # pylint: disable=unexpected-keyword-arg
-            output_dir=self.cfg.output_dir,
+            self.cfg.output_dir,
             per_device_train_batch_size=self.cfg.micro_batch_size,
-            max_steps=self.cfg.max_steps or total_num_steps,
+            max_steps=max_steps,
             gradient_accumulation_steps=self.cfg.gradient_accumulation_steps,
             learning_rate=self.cfg.learning_rate,
             warmup_steps=self.cfg.warmup_steps,
@@ -1044,11 +1108,16 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         if self.cfg.adapter and self.peft_config:
             dpo_trainer_kwargs["peft_config"] = self.peft_config
         if self.cfg.precompute_ref_log_probs is not None:
-            dpo_trainer_kwargs[
-                "precompute_ref_log_probs"
-            ] = self.cfg.precompute_ref_log_probs
-        if self.cfg.rl in ["dpo", "ipo"]:
-            trainer_cls = AxolotlDPOTrainer
+            dpo_trainer_kwargs["precompute_ref_log_probs"] = (
+                self.cfg.precompute_ref_log_probs
+            )
+        if self.cfg.rl == "grpo":
+            trainer_cls = GRPOStrategy.get_trainer_class()
+            trainer_cls_args = [self.model]
+            trainer_cls_args.extend(GRPOStrategy.set_trainer_args(self.cfg))
+            dpo_trainer_kwargs.update(GRPOStrategy.set_trainer_kwargs(self.cfg))
+        elif self.cfg.rl in ["dpo", "ipo"]:
+            trainer_cls = DPOStrategy.get_trainer_class()
             trainer_cls_args = [self.model, self.model_ref]
         elif self.cfg.rl == "orpo":
             trainer_cls = AxolotlORPOTrainer
@@ -1063,12 +1132,14 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             raise ValueError(f"Unsupported RL: {self.cfg.rl}")
 
         sig = inspect.signature(trainer_cls)
-        if "processing_class" in sig.parameters.keys():
-            dpo_trainer_kwargs["processing_class"] = self.tokenizer
-        else:
+        if "tokenizer" in sig.parameters.keys():
             dpo_trainer_kwargs["tokenizer"] = self.tokenizer
+        else:
+            dpo_trainer_kwargs["processing_class"] = self.tokenizer
 
-        if self.cfg.datasets is not None and (trainer_cls is AxolotlDPOTrainer):
+        if self.cfg.datasets is not None and (
+            trainer_cls is DPOStrategy.get_trainer_class()
+        ):
             dpo_trainer_kwargs["dataset_tags"] = [
                 d["path"] for d in self.cfg.datasets if not Path(d["path"]).is_dir()
             ]

src/axolotl/core/trainers/base.py

@@ -5,12 +5,11 @@ module for customized trainers
 from __future__ import annotations
 
 # pylint: disable=too-many-lines
-import gc
 import logging
 import os
 from collections import defaultdict
 from functools import wraps
-from typing import Any, Dict, Literal, Optional, Union
+from typing import Dict, Literal, Optional
 
 import torch
 from datasets import Dataset
@@ -21,19 +20,14 @@ from torch.utils.data import BatchSampler, DataLoader, RandomSampler, Sequential
 from transformers import Trainer
 from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR, seed_worker
 from transformers.utils import is_sagemaker_mp_enabled
-from trl import (
-    CPOTrainer,
-    DPOTrainer,
-    KTOTrainer,
-    ORPOTrainer,
-    PRMTrainer,
-    RewardTrainer,
-)
+from trl import CPOTrainer, KTOTrainer, ORPOTrainer, PRMTrainer, RewardTrainer
 from trl.trainer.utils import pad_to_length
 
+from axolotl.integrations.base import BaseOptimizerFactory
 from axolotl.monkeypatch.relora import ReLoRAScheduler
 from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
 from axolotl.utils.schedulers import (
+    RexLR,
     get_cosine_schedule_with_min_lr,
     get_cosine_schedule_with_quadratic_warmup,
     get_cosine_schedule_with_warmup_decay_constant,
@@ -124,6 +118,17 @@ class SchedulerMixin(Trainer):
                     **extra_lr_kwargs,
                     **self.args.lr_scheduler_kwargs,
                 )
+            elif self.args.alternate_lr_scheduler_type == "rex":
+                if use_cosine_min_lr:
+                    assert 0 <= self.args.cosine_min_lr_ratio <= 1.0, "cosine_min_lr_ratio must be between 0.0 and 1.0"
+
+                self.lr_scheduler = RexLR(
+                    optimizer=optimizer,
+                    max_lr=self.args.learning_rate,
+                    min_lr=0 if not use_cosine_min_lr else (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("Both cosine quadratic warmup and min lr detected. Using quadratic warmup.")
@@ -163,47 +168,18 @@ class SchedulerMixin(Trainer):
         return self.lr_scheduler
 
 
-class AxolotlTrainer(SchedulerMixin, Trainer):
+class OptimizerMixin(Trainer):
     """
-    Extend the base Trainer for axolotl helpers
+    Mixin class for shared handling of building custom optimizers
     """
 
     args = None  # type: "AxolotlTrainingArguments"  # type: ignore[name-defined]
-    tag_names = ["axolotl"]
 
-    def __init__(
-        self,
-        *_args,
-        bench_data_collator=None,
-        eval_data_collator=None,
-        dataset_tags=None,
-        **kwargs,
-    ):
-        self.bench_data_collator = bench_data_collator
-        self.eval_data_collator = eval_data_collator
-        self.dataset_tags = dataset_tags
-        self._signature_columns = None  # workaround for pylint
-        super().__init__(*_args, **kwargs)
-        self.train_data_collator = self.data_collator
-        self._stored_metrics = defaultdict(lambda: defaultdict(list))
-        if self.args.orpo_alpha:
-            self.loss_fct = torch.nn.CrossEntropyLoss(reduction="none")
-
-    def _wrap_model(self, model, training=True, dataloader=None):
-        if self.args.torch_compile:
-            torch._dynamo.config.accumulated_cache_size_limit = (  # pylint: disable=protected-access
-                256
-            )
-            model = torch.compile(
-                model,
-                backend=self.args.torch_compile_backend,
-                mode=self.args.torch_compile_mode,
-            )
-        return super()._wrap_model(model, training=training, dataloader=dataloader)
-
-    def create_optimizer_grouped_parameters(self, opt_model, optimizer_kwargs):
+    def create_optimizer_grouped_parameters(
+        self, opt_model, optimizer_kwargs
+    ) -> list[dict]:
         decay_parameters = self.get_decay_parameter_names(opt_model)
-        params = {
+        params: dict = {
             "to_weight_decay": {},  # LayerNorm and bias
             "embeddings": {},  # lm_head, embed_tokens,
             "no_weight_decay": {},
@@ -290,23 +266,30 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
             and self.args.embedding_lr_scale is None
             and self.args.embedding_lr is None
             and self.args.lr_groups is None
-            and self.args.alternate_optimizer
-            not in [
-                "optimi_adamw",
-                "ao_adamw_8bit",
-                "ao_adamw_4bit",
-                "ao_adamw_fp8",
-                "adopt_adamw",
-            ]
+            and self.optimizer_cls_and_kwargs 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,
+
+        if (
+            not self.optimizer
+            and self.optimizer_cls_and_kwargs is not None
+            and issubclass(self.optimizer_cls_and_kwargs[0], BaseOptimizerFactory)
+        ):
+            optimizer_factory_cls, optimizer_kwargs = self.optimizer_cls_and_kwargs
+            self.optimizer = optimizer_factory_cls()(
+                opt_model, self.args, **optimizer_kwargs
             )
+
+        if not self.optimizer:
+            if self.optimizer_cls_and_kwargs is not None:
+                optimizer_cls, optimizer_kwargs = self.optimizer_cls_and_kwargs
+            else:
+                optimizer_cls, optimizer_kwargs = self.get_optimizer_cls_and_kwargs(
+                    self.args, opt_model
+                )
+
             optimizer_grouped_parameters = self.create_optimizer_grouped_parameters(
                 opt_model, optimizer_kwargs
             )
@@ -323,50 +306,47 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
                     loraplus_lr_embedding=loraplus_lr_embedding,
                     **optimizer_kwargs,
                 )
-            elif (
-                self.args.embedding_lr_scale is not None
-                or self.args.embedding_lr is not None
-                or self.args.lr_groups is not None
-            ):
-                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
-                    optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs)
-                )
-            elif self.args.alternate_optimizer == "optimi_adamw":
-                from optimi import AdamW
+            else:
+                # Overwrite `params` in case it's created by `get_optimizer_cls_and_kwargs`
+                # e.g. for GaLore optimizer.
+                if "params" in optimizer_kwargs:
+                    optimizer_grouped_parameters = optimizer_kwargs.pop("params")
 
-                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
-                    AdamW(
-                        optimizer_grouped_parameters, foreach=False, **optimizer_kwargs
+                # Overwrite `model` in case it's created by `get_optimizer_cls_and_kwargs`
+                # e.g. for LOMO optimizer.
+                if "model" in optimizer_kwargs:
+                    optimizer_grouped_parameters = optimizer_kwargs.pop("model")
+
+                # For layer-wise dummy optimizers we overwrite optimizer_grouped_parameters with `optimizer_dict`
+                # to avoid arguments conflicts.
+                if "optimizer_dict" in optimizer_kwargs:
+                    optimizer_grouped_parameters = optimizer_kwargs.pop(
+                        "optimizer_dict"
                     )
-                )
-            elif self.args.alternate_optimizer == "ao_adamw_4bit":
-                from torchao.prototype.low_bit_optim import AdamW4bit
 
-                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
-                    AdamW4bit(optimizer_grouped_parameters, **optimizer_kwargs)
+                self.optimizer = optimizer_cls(
+                    optimizer_grouped_parameters, **optimizer_kwargs
                 )
-            elif self.args.alternate_optimizer == "ao_adamw_8bit":
-                from torchao.prototype.low_bit_optim import AdamW8bit
 
-                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
-                    AdamW8bit(optimizer_grouped_parameters, **optimizer_kwargs)
-                )
-            elif self.args.alternate_optimizer == "ao_adamw_fp8":
-                from torchao.prototype.low_bit_optim import AdamWFp8
+            if optimizer_cls.__name__ == "Adam8bit":
+                import bitsandbytes
 
-                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
-                    AdamWFp8(optimizer_grouped_parameters, **optimizer_kwargs)
-                )
-            elif self.args.alternate_optimizer == "adopt_adamw":
-                from axolotl.utils.optimizers.adopt import ADOPT
+                manager = bitsandbytes.optim.GlobalOptimManager.get_instance()
 
-                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
-                    ADOPT(
-                        optimizer_grouped_parameters,
-                        decouple=True,
-                        **optimizer_kwargs,
-                    )
-                )
+                skipped = 0
+                for module in opt_model.modules():
+                    if isinstance(module, nn.Embedding):
+                        skipped += sum(
+                            {
+                                p.data_ptr(): p.numel() for p in module.parameters()
+                            }.values()
+                        )
+                        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")
 
         if is_sagemaker_mp_enabled():
             self.optimizer = smp.DistributedOptimizer(  # pylint: disable=attribute-defined-outside-init
@@ -375,6 +355,45 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
 
         return self.optimizer
 
+
+class AxolotlTrainer(SchedulerMixin, OptimizerMixin, Trainer):
+    """
+    Extend the base Trainer for axolotl helpers
+    """
+
+    args = None  # type: "AxolotlTrainingArguments"  # type: ignore[name-defined]
+    tag_names = ["axolotl"]
+
+    def __init__(
+        self,
+        *_args,
+        bench_data_collator=None,
+        eval_data_collator=None,
+        dataset_tags=None,
+        **kwargs,
+    ):
+        self.bench_data_collator = bench_data_collator
+        self.eval_data_collator = eval_data_collator
+        self.dataset_tags = dataset_tags
+        self._signature_columns = None  # workaround for pylint
+        super().__init__(*_args, **kwargs)
+        self.train_data_collator = self.data_collator
+        self._stored_metrics = defaultdict(lambda: defaultdict(list))
+        if self.args.orpo_alpha:
+            self.loss_fct = torch.nn.CrossEntropyLoss(reduction="none")
+
+    def _wrap_model(self, model, training=True, dataloader=None):
+        if self.args.torch_compile:
+            torch._dynamo.config.accumulated_cache_size_limit = (  # pylint: disable=protected-access
+                256
+            )
+            model = torch.compile(
+                model,
+                backend=self.args.torch_compile_backend,
+                mode=self.args.torch_compile_mode,
+            )
+        return super()._wrap_model(model, training=training, dataloader=dataloader)
+
     def _get_train_sampler(self) -> Optional[torch.utils.data.Sampler]:
         if self.args.sample_packing and not self.args.pretraining:
             if self.args.multipack_real_batches:
@@ -443,9 +462,9 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
                 "pin_memory": self.args.dataloader_pin_memory,
             }
             if self.args.dataloader_prefetch_factor:
-                dataloader_params[
-                    "prefetch_factor"
-                ] = self.args.dataloader_prefetch_factor
+                dataloader_params["prefetch_factor"] = (
+                    self.args.dataloader_prefetch_factor
+                )
 
             sampler = self._get_train_sampler()
             if isinstance(sampler, BatchSampler):
@@ -490,9 +509,9 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
                 "pin_memory": self.args.dataloader_pin_memory,
             }
             if self.args.dataloader_prefetch_factor:
-                dataloader_params[
-                    "prefetch_factor"
-                ] = self.args.dataloader_prefetch_factor
+                dataloader_params["prefetch_factor"] = (
+                    self.args.dataloader_prefetch_factor
+                )
 
             if isinstance(eval_sampler, BatchSampler):
                 dataloader_params["batch_sampler"] = eval_sampler
@@ -847,107 +866,6 @@ class ReLoRATrainer(AxolotlTrainer):
         return self.lr_scheduler
 
 
-class AxolotlDPOTrainer(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):
-        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)
-            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:
-        """
-        Overwrite the `push_to_hub` method in order to force-add the tags when pushing the
-        model on the Hub. Please refer to `~transformers.Trainer.push_to_hub` for more details.
-        """
-        kwargs = _sanitize_kwargs_for_ds_tagging(
-            dataset_tags=self.dataset_tags, kwargs=kwargs
-        )
-        kwargs = _sanitize_kwargs_for_tagging(tag_names=self.tag_names, kwargs=kwargs)
-
-        return super().push_to_hub(*args, **kwargs)
-
-    @staticmethod
-    def tokenize_row(
-        features,
-        processing_class,
-        max_prompt_length,
-        max_completion_length,
-        add_special_tokens,
-    ) -> Dict:
-        res = DPOTrainer.tokenize_row(
-            features,
-            processing_class,
-            max_prompt_length,
-            max_completion_length,
-            add_special_tokens,
-        )
-        # fix when the tokenizer doesn't have a bos_token_id, e.g. Qwen
-        if processing_class.bos_token is None and res["prompt_input_ids"][0] is None:
-            for key in res.keys():
-                res[key] = res[key][1:]
-
-        if processing_class.bos_token and processing_class.bos_token_id is not None:
-            # dpo trainer may incorrectly prepend the bos_token_id to the dpo outputs
-            if res["chosen_input_ids"][0] == processing_class.bos_token_id:
-                res["chosen_input_ids"] = res["chosen_input_ids"][1:]
-                res["chosen_labels"] = res["chosen_labels"][1:]
-                res["chosen_attention_mask"] = res["chosen_attention_mask"][1:]
-            if res["rejected_input_ids"][0] == processing_class.bos_token_id:
-                res["rejected_input_ids"] = res["rejected_input_ids"][1:]
-                res["rejected_labels"] = res["rejected_labels"][1:]
-                res["rejected_attention_mask"] = res["rejected_attention_mask"][1:]
-
-        return res
-
-    def training_step(
-        self,
-        model: nn.Module,
-        inputs: Dict[str, Union[torch.Tensor, Any]],
-        num_items_in_batch=None,
-    ) -> torch.Tensor:
-        loss: torch.Tensor = super().training_step(model, inputs, num_items_in_batch)
-        gc.collect()
-        torch.cuda.empty_cache()
-        return loss
-
-
 class AxolotlORPOTrainer(SchedulerMixin, ORPOTrainer):
     """
     Extend the base ORPOTrainer for axolotl helpers

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

@@ -0,0 +1,34 @@
+"""
+DPO Specific Strategy for training
+"""
+
+from axolotl.core.trainers.dpo.trainer import AxolotlDPOTrainer
+
+
+class DPOStrategy:
+    """
+    Strategy for DPO training
+    """
+
+    @classmethod
+    def get_trainer_class(cls):
+        return AxolotlDPOTrainer
+
+    @classmethod
+    def get_training_args_class(cls):
+        from axolotl.core.trainers.dpo.args import AxolotlDPOConfig
+
+        return AxolotlDPOConfig
+
+    @classmethod
+    def set_training_args_kwargs(cls, cfg):
+        training_args_kwargs = {}
+        if cfg.rl == "ipo":
+            training_args_kwargs["loss_type"] = "ipo"
+        training_args_kwargs["max_length"] = cfg.sequence_len
+        training_args_kwargs["max_completion_length"] = None
+        training_args_kwargs["max_prompt_length"] = cfg.sequence_len
+        training_args_kwargs["generate_during_eval"] = cfg.use_wandb
+        if cfg.dpo_use_weighting is not None:
+            training_args_kwargs["use_weighting"] = cfg.dpo_use_weighting
+        return training_args_kwargs

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

@@ -0,0 +1,16 @@
+"""
+Axolotl specific DPO args
+"""
+
+from dataclasses import dataclass
+
+from trl import DPOConfig
+
+from axolotl.core.training_args import AxolotlTrainingMixins
+
+
+@dataclass
+class AxolotlDPOConfig(AxolotlTrainingMixins, DPOConfig):
+    """
+    DPO config for DPO training
+    """

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

@@ -0,0 +1,126 @@
+"""
+DPO trainer for axolotl
+"""
+
+import gc
+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.base import (
+    SchedulerMixin,
+    _sanitize_kwargs_for_ds_tagging,
+    _sanitize_kwargs_for_tagging,
+)
+
+if is_sagemaker_mp_enabled():
+    import smdistributed.modelparallel.torch as smp
+
+
+class AxolotlDPOTrainer(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:
+        """
+        Overwrite the `push_to_hub` method in order to force-add the tags when pushing the
+        model on the Hub. Please refer to `~transformers.Trainer.push_to_hub` for more details.
+        """
+        kwargs = _sanitize_kwargs_for_ds_tagging(
+            dataset_tags=self.dataset_tags, kwargs=kwargs
+        )
+        kwargs = _sanitize_kwargs_for_tagging(tag_names=self.tag_names, kwargs=kwargs)
+
+        return super().push_to_hub(*args, **kwargs)
+
+    @staticmethod
+    def tokenize_row(
+        features,
+        processing_class,
+        max_prompt_length,
+        max_completion_length,
+        add_special_tokens,
+    ) -> Dict:
+        res = DPOTrainer.tokenize_row(
+            features,
+            processing_class,
+            max_prompt_length,
+            max_completion_length,
+            add_special_tokens,
+        )
+        # fix when the tokenizer doesn't have a bos_token_id, e.g. Qwen
+        if processing_class.bos_token is None and res["prompt_input_ids"][0] is None:
+            for key in res.keys():
+                res[key] = res[key][1:]
+
+        if processing_class.bos_token and processing_class.bos_token_id is not None:
+            # dpo trainer may incorrectly prepend the bos_token_id to the dpo outputs
+            if res["chosen_input_ids"][0] == processing_class.bos_token_id:
+                res["chosen_input_ids"] = res["chosen_input_ids"][1:]
+                res["chosen_labels"] = res["chosen_labels"][1:]
+                res["chosen_attention_mask"] = res["chosen_attention_mask"][1:]
+            if res["rejected_input_ids"][0] == processing_class.bos_token_id:
+                res["rejected_input_ids"] = res["rejected_input_ids"][1:]
+                res["rejected_labels"] = res["rejected_labels"][1:]
+                res["rejected_attention_mask"] = res["rejected_attention_mask"][1:]
+
+        return res
+
+    def training_step(
+        self,
+        model: nn.Module,
+        inputs: Dict[str, Union[torch.Tensor, Any]],
+        num_items_in_batch=None,
+    ) -> torch.Tensor:
+        loss: torch.Tensor = super().training_step(model, inputs, num_items_in_batch)
+        gc.collect()
+        torch.cuda.empty_cache()
+        return loss

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

@@ -0,0 +1,134 @@
+"""
+GRPO Specific Strategy for training
+"""
+
+import importlib
+import inspect
+import logging
+
+from trl.trainer.grpo_trainer import RewardFunc
+
+from axolotl.core.trainers.grpo.trainer import AxolotlGRPOTrainer
+from axolotl.utils.config.models.input.v0_4_1.trl import TRLConfig
+
+LOG = logging.getLogger("axolotl")
+
+
+class GRPOStrategy:
+    """
+    Strategy for GRPO training
+    """
+
+    @classmethod
+    def get_trainer_class(cls):
+        return AxolotlGRPOTrainer
+
+    @classmethod
+    def get_training_args_class(cls):
+        from axolotl.core.trainers.grpo.args import AxolotlGRPOConfig
+
+        return AxolotlGRPOConfig
+
+    @classmethod
+    def set_training_args_kwargs(cls, cfg):
+        grpo_args_kwargs = {}
+
+        if not hasattr(cfg, "trl") or not cfg.trl:
+            return grpo_args_kwargs
+
+        trl: TRLConfig = cfg.trl  # type: ignore
+
+        if trl.use_vllm:
+            grpo_args_kwargs["use_vllm"] = trl.use_vllm
+            grpo_args_kwargs["vllm_device"] = (
+                trl.vllm_device if trl.vllm_device else "auto"
+            )
+
+            if trl.vllm_gpu_memory_utilization:
+                grpo_args_kwargs["vllm_gpu_memory_utilization"] = (
+                    trl.vllm_gpu_memory_utilization
+                )
+
+            if trl.vllm_max_model_len:
+                grpo_args_kwargs["vllm_max_model_len"] = trl.vllm_max_model_len
+
+        if trl.num_generations:
+            grpo_args_kwargs["num_generations"] = trl.num_generations
+
+        if trl.sync_ref_model:
+            grpo_args_kwargs["sync_ref_model"] = trl.sync_ref_model
+
+            if trl.ref_model_mixup_alpha:
+                grpo_args_kwargs["ref_model_mixup_alpha"] = trl.ref_model_mixup_alpha
+
+            if trl.ref_model_sync_steps:
+                grpo_args_kwargs["ref_model_sync_steps"] = trl.ref_model_sync_steps
+
+        grpo_args_kwargs["max_completion_length"] = trl.max_completion_length
+        grpo_args_kwargs["log_completions"] = trl.log_completions
+
+        if trl.reward_weights:
+            grpo_args_kwargs["reward_weights"] = trl.reward_weights
+
+        return grpo_args_kwargs
+
+    @classmethod
+    def set_trainer_args(cls, cfg):
+        trainer_args = []
+        if cfg.trl and cfg.trl.reward_funcs:
+            reward_funcs = []
+            for reward_func_fqn in cfg.trl.reward_funcs:
+                reward_funcs.append(cls.get_reward_func(reward_func_fqn))
+            trainer_args.append(reward_funcs)
+        return trainer_args
+
+    @classmethod
+    def set_trainer_kwargs(cls, cfg):
+        trainer_kwargs = {}
+        if cfg.trl and cfg.trl.reward_processing_classes:
+            trainer_kwargs["reward_processing_classes"] = (
+                cfg.trl.reward_processing_classes
+            )
+        return trainer_kwargs
+
+    @classmethod
+    def get_collator(cls, *args, **kwargs):  # pylint: disable=unused-argument
+        # No data collation is needed in GRPO, handled by trl's trainer __init__
+        return None
+
+    @classmethod
+    def get_blocklist_args_kwargs(cls):
+        return ["dataset_num_proc"]
+
+    @classmethod
+    def get_reward_func(cls, reward_func_fqn: str) -> RewardFunc:
+        """
+        Returns the reward function from the given fully qualified name, or the path to the reward function model.
+
+        Args:
+            reward_func_fqn (str): Fully qualified name of the reward function (e.g. r1_grpo.gsm8k_transform),
+                or a HF hub path to the reward model.
+        Raises:
+            ValueError: If the reward function does not accept at least two arguments.
+
+        Returns:
+            RewardFunc: A callable that accepts prompts and completions and returns rewards,
+                or a path to a reward model.
+
+        """
+        try:
+            # use importlib to dynamically load the reward function from the module
+            reward_func_module_name = reward_func_fqn.split(".")[-1]
+            reward_func_module = importlib.import_module(reward_func_fqn.split(".")[-2])
+            reward_func = getattr(reward_func_module, reward_func_module_name)
+            if not len(inspect.signature(reward_func).parameters) >= 2:
+                raise ValueError(
+                    "Reward function must accept at least two arguments: prompts: list and completions: list"
+                )
+            return reward_func
+        except ModuleNotFoundError:
+            # the user has passed a string (ideally indicating the path of a reward model)
+            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

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

@@ -0,0 +1,16 @@
+"""
+Axolotl Specific Training Args
+"""
+
+from dataclasses import dataclass
+
+from trl import GRPOConfig
+
+from axolotl.core.training_args import AxolotlTrainingMixins
+
+
+@dataclass
+class AxolotlGRPOConfig(AxolotlTrainingMixins, GRPOConfig):
+    """
+    Axolotl GRPO Config for GRPO training
+    """

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

@@ -0,0 +1,109 @@
+"""
+Axolotl GRPO trainer
+"""
+
+from accelerate.utils import is_peft_model
+from accelerate.utils.other import is_compiled_module
+from transformers import PreTrainedModel
+from trl import GRPOConfig, GRPOTrainer
+from trl.models import unwrap_model_for_generation
+
+from axolotl.core.trainers.base import SchedulerMixin
+
+
+# mypy: ignore-errors
+class AxolotlGRPOTrainer(SchedulerMixin, GRPOTrainer):
+    """
+    Extend the base GRPOTrainer for axolotl helpers
+    """
+
+    _tag_names = ["trl", "grpo", "axolotl"]
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+        # pylint: disable=access-member-before-definition
+        # Enable gradient checkpointing if requested
+        if kwargs["args"].gradient_checkpointing:
+            # Ensure use_cache is disabled
+            if hasattr(self.model, "config"):
+                self.model.config.use_cache = False
+
+            # Enable gradient checkpointing on the base model for PEFT
+            if is_peft_model(self.model) and hasattr(
+                self.model.base_model, "gradient_checkpointing_enable"
+            ):
+                self.model.base_model.gradient_checkpointing_enable()
+            # Enable gradient checkpointing for non-PEFT models
+            elif hasattr(self.model, "gradient_checkpointing_enable"):
+                self.model.gradient_checkpointing_enable()
+            self.model = self._enable_gradient_checkpointing(self.model, kwargs["args"])
+        # pylint: enable=access-member-before-definition
+
+    def _enable_gradient_checkpointing(
+        self, model: PreTrainedModel, args: GRPOConfig
+    ) -> PreTrainedModel:
+        """Enables gradient checkpointing for the model."""
+        # pylint: disable=unused-argument,redefined-builtin
+        gradient_checkpointing_kwargs = args.gradient_checkpointing_kwargs or {}
+        use_reentrant = (
+            "use_reentrant" not in gradient_checkpointing_kwargs
+            or gradient_checkpointing_kwargs["use_reentrant"]
+        )
+
+        if use_reentrant:
+            if hasattr(model, "enable_input_require_grads"):
+                model.enable_input_require_grads()
+            else:
+
+                def make_inputs_require_grad(module, input, output):
+                    output.requires_grad_(True)
+
+                model.get_input_embeddings().register_forward_hook(
+                    make_inputs_require_grad
+                )
+
+        return model
+        # pylint: enable=unused-argument,redefined-builtin
+
+    def _move_model_to_vllm(self):
+        with unwrap_model_for_generation(
+            self.model,
+            self.accelerator,
+            gather_deepspeed3_params=self.args.ds3_gather_for_generation,
+        ) as unwrapped_model:
+            if is_compiled_module(unwrapped_model):
+                unwrapped_model = (
+                    unwrapped_model._orig_mod  # pylint: disable=protected-access
+                )
+            if is_peft_model(unwrapped_model):
+                unwrapped_model.merge_adapter()
+                state_dict = unwrapped_model.state_dict()
+                # Remove base_model and base_layer prefixes
+                state_dict = {
+                    k.removeprefix("base_model.model.")
+                    .removeprefix("base_model.model.")
+                    .replace(".base_layer", ""): v
+                    for k, v in state_dict.items()
+                }
+                # Remove values with adapter prefix (example: "_lora")
+                state_dict = {
+                    k: v
+                    for k, v in state_dict.items()
+                    if unwrapped_model.prefix not in k
+                }
+                # When module to save, remove its prefix and discard the original module
+                state_dict = {
+                    k.replace("modules_to_save.default.", ""): v
+                    for k, v in state_dict.items()
+                    if "original_module" not in k
+                }
+            else:
+                state_dict = unwrapped_model.state_dict()
+            if self.accelerator.is_main_process:
+                llm_model = (
+                    self.llm.llm_engine.model_executor.driver_worker.model_runner.model
+                )
+                llm_model.load_weights(state_dict.items())
+            if is_peft_model(unwrapped_model):
+                unwrapped_model.unmerge_adapter()

src/axolotl/core/trainers/trl.py

@@ -1,6 +1,7 @@
 """
 module for TRL PPO training
 """
+
 import torch
 from tqdm import tqdm
 from trl import PPOTrainer

src/axolotl/core/training_args.py

@@ -1,11 +1,12 @@
 """
 extra axolotl specific training args
 """
+
 from dataclasses import dataclass, field
 from typing import Optional
 
 from transformers import TrainingArguments
-from trl import CPOConfig, DPOConfig, KTOConfig, ORPOConfig, PRMConfig, RewardConfig
+from trl import CPOConfig, KTOConfig, ORPOConfig, PRMConfig, RewardConfig
 
 
 @dataclass
@@ -217,13 +218,6 @@ class AxolotlTrainingArguments(AxolotlTrainingMixins, TrainingArguments):
     """
 
 
-@dataclass
-class AxolotlDPOConfig(AxolotlTrainingMixins, DPOConfig):
-    """
-    DPO config for DPO training
-    """
-
-
 @dataclass
 class AxolotlORPOConfig(AxolotlTrainingMixins, ORPOConfig):
     """

src/axolotl/integrations/base.py

@@ -23,6 +23,8 @@ import importlib
 import logging
 from typing import OrderedDict
 
+import torch
+
 
 class BasePlugin:
     """
@@ -469,3 +471,14 @@ class PluginManager:
         """
         for plugin in self.plugins.values():
             plugin.post_train_unload(cfg)
+
+
+class BaseOptimizerFactory:
+    """
+    Base class for factories to create custom optimizers
+    """
+
+    def __call__(
+        self, opt_model, training_args, **optimizer_kwargs
+    ) -> "torch.optim.Optimizer":
+        pass

src/axolotl/integrations/cut_cross_entropy/README.md

@@ -1,6 +1,26 @@
 # Cut Cross Entropy
 
-### Usage
+Cut Cross Entropy reduces VRAM usage through optimization on the cross-entropy operation during loss calculation.
+
+See https://github.com/apple/ml-cross-entropy
+
+## Requirements
+
+- PyTorch 2.4.0 or higher
+
+## Installation
+
+Run the following command to install `cut_cross_entropy[transformers]` if you don't have it already.
+
+```bash
+# if you are in dev environment
+python scripts/cutcrossentropy_install.py | sh
+
+# if you are not in dev environment
+pip3 uninstall -y cut-cross-entropy && pip3 install "cut-cross-entropy[transformers] @ git+https://github.com/apple/ml-cross-entropy.git@24fbe4b5dab9a6c250a014573613c1890190536c"
+```
+
+## Usage
 
 ```yaml
 plugins:
@@ -8,3 +28,19 @@ plugins:
 
 cut_cross_entropy: true
 ```
+
+## Citation
+
+```bib
+@article{wijmans2024cut,
+  author       = {Erik Wijmans and
+                  Brody Huval and
+                  Alexander Hertzberg and
+                  Vladlen Koltun and
+                  Philipp Kr\"ahenb\"uhl},
+  title        = {Cut Your Losses in Large-Vocabulary Language Models},
+  journal      = {arXiv},
+  year         = {2024},
+  url          = {https://arxiv.org/abs/2411.09009},
+}
+```

src/axolotl/integrations/cut_cross_entropy/__init__.py

@@ -33,7 +33,7 @@ LOG = logging.getLogger("axolotl.integrations.cut_cross_entropy")
 
 _CCE_INSTALL_MESSAGE = (
     "Please install cut_cross_entropy with transformers support using "
-    '`pip install "cut-cross-entropy[transformers]==24.11.4"`'
+    '`pip install "cut-cross-entropy[transformers] @ git+https://github.com/apple/ml-cross-entropy.git@24fbe4b5dab9a6c250a014573613c1890190536c"`'
 )
 
 

src/axolotl/integrations/grokfast/README.md

@@ -2,7 +2,7 @@
 
 See https://github.com/ironjr/grokfast
 
-### Usage
+## Usage
 
 ```yaml
 plugins:
@@ -11,3 +11,14 @@ plugins:
 grokfast_alpha: 2.0
 grokfast_lamb: 0.98
 ```
+
+## Citation
+
+```bib
+@article{lee2024grokfast,
+    title={{Grokfast}: Accelerated Grokking by Amplifying Slow Gradients},
+    author={Lee, Jaerin and Kang, Bong Gyun and Kim, Kihoon and Lee, Kyoung Mu},
+    journal={arXiv preprint arXiv:2405.20233},
+    year={2024}
+}
+```

src/axolotl/integrations/grokfast/__init__.py

@@ -1,6 +1,7 @@
 """
 Grokfast plugin for Axolotl
 """
+
 import logging
 
 from transformers.trainer_callback import TrainerCallback

src/axolotl/integrations/grokfast/args.py

@@ -1,6 +1,7 @@
 """
 config args for grokfast plugin
 """
+
 from typing import Optional
 
 from pydantic import BaseModel

src/axolotl/integrations/kd/README.md

@@ -0,0 +1,23 @@
+# Knowledge Distillation
+
+## Usage
+
+```yaml
+plugins:
+  - "axolotl.integrations.kd.KDPlugin"
+
+kd_trainer: True
+kd_ce_alpha: 0.1
+kd_alpha: 0.9
+kd_temperature: 1.0
+
+torch_compile: True  # torch>=2.5.1, recommended to reduce vram
+
+datasets:
+  - path: ...
+    type: "axolotl.integrations.kd.chat_template"
+    field_messages: "messages_combined"
+    logprobs_field: "llm_text_generation_vllm_logprobs"  # for kd only, field of logprobs
+```
+
+An example dataset can be found at [`axolotl-ai-co/evolkit-logprobs-pipeline-75k-v2-sample`](https://huggingface.co/datasets/axolotl-ai-co/evolkit-logprobs-pipeline-75k-v2-sample)

src/axolotl/integrations/kd/args.py

@@ -26,12 +26,12 @@ class KDArgs(BaseModel):
     """
 
     kd_trainer: Optional[bool] = None  # whether to use KD trainer
-    kd_ce_alpha: Optional[
-        float
-    ] = None  # loss coefficient for cross-entropy loss during KD
+    kd_ce_alpha: Optional[float] = (
+        None  # loss coefficient for cross-entropy loss during KD
+    )
     kd_alpha: Optional[float] = None  # loss coefficient for KD loss
     kd_temperature: Optional[float] = None  # temperature for sampling during KD
     kd_zscore_base_temp: Optional[float] = None  # base temperature for zscore scaling
-    kd_top_k_before_softmax: Optional[
-        bool
-    ] = None  # whether to sample top k before softmax during KD
+    kd_top_k_before_softmax: Optional[bool] = (
+        None  # whether to sample top k before softmax during KD
+    )

src/axolotl/integrations/kd/topk_logprob/LICENSE.md

@@ -1,58 +0,0 @@
-### AXOLOTL COMMUNITY LICENSE AGREEMENT
-
-This Axolotl Community License Agreement (“Agreement”) is entered into by and between Axolotl AI Corp. (“Axolotl”) and
-any individual or entity (“Licensee”) who wishes to use the Software (as defined below) in accordance with the terms
-and conditions set forth in this Agreement.
-
-1.  Definitions
-    1.1 “Licensee” refers to any individual or entity who has obtained a copy of the Software under this Agreement.
-    1.2 “Plugin Integration” means independent integration software modules which may or may not be offered by Axolotl,
-        which may be licensed separately by their respective  authors and/or licensors.
-    1.3 “Software” refers to the specific sub-directory of the Axolotl, Inc. software located at
-        https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/integrations and its subdirectories which
-        permits Plugin Integrations to integrate with the Axolotl service.
-2.  Grant of License
-    2.1	Axolotl hereby grants Licensee a worldwide, non-exclusive, royalty-free, license to use, copy, modify, merge,
-        publish, distribute, sublicense, and/or otherwise exploit the Software, subject to the following conditions:
-        - Licensee must comply with all the terms and conditions of this Agreement.
-        - Licensee must include the original copyright notice and disclaimer of warranty in all copies or substantial
-          portions of the Software.
-    2.2 Licensee may use the Software for any lawful purpose, except as restricted in Section 3.
-3.  Restrictions
-    3.1 Licensee shall not use the Software for any activity that constitutes a commercial activity of offering for
-        free or for sale any services, platform, or equivalent  to third parties for the purposes of allowing such
-        third parties to fine-tune artificial intelligence models.
-    3.2 Licensee shall not:
-        - Use the Software for any illegal or unauthorized purpose.
-        - Reverse engineer, decompile, or disassemble the Software.
-        - Remove or modify any copyright, trademark, or other proprietary notices contained in the Software.
-        - Use the Software in a way that could damage, disable, overburden, or impair the functionality of the
-          Software or interfere with any third-party use of the Software.
-    3.3 Axolotl reserves the right to restrict certain Plugin Integrations for use with the Software. To the extent Licensee integrates a permitted, applicable Plugin Integration with the Software, Licensee shall comply with any additional terms and conditions imposed by the licensors of such Plugin Integration for use of such Plugin Integrations. Licensee shall contact Axolotl if it has questions about whether its use of the Software falls beyond the scope of this Agreement.
-4.  Intellectual Property Rights
-    4.1 Axolotl and its contributors retain all intellectual property rights in and to the Software. Licensee
-        acknowledges that this Agreement does not transfer any ownership rights or intellectual property rights to
-        Licensee.
-5.  Disclaimer of Warranty
-    5.1 THE SOFTWARE IS PROVIDED “AS IS,” WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED
-        TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, AND NON-INFRINGEMENT. IN NO EVENT SHALL
-        THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES, OR OTHER LIABILITY, WHETHER IN AN ACTION OF
-        CONTRACT, TORT, OR OTHERWISE, ARISING FROM, OUT OF, OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
-        DEALINGS IN THE SOFTWARE.
-6.  Termination
-    6.1 Axolotl may terminate this Agreement at any time if Licensee fails to comply with any of the terms and
-        conditions set forth herein. Upon termination, Licensee shall cease all use of the Software and destroy any
-        copies in its possession.
-7.  Governing Law
-    7.1 This Agreement shall be governed by and construed in accordance with the laws of the State of California,
-        without regards to conflicts of laws provisions thereof.
-8.  Entire Agreement
-    8.1 This Agreement constitutes the entire agreement between Axolotl and Licensee with respect to the subject matter
-        hereof and supersedes all prior or contemporaneous understandings or agreements between the parties concerning
-        the Software, whether written or oral. Axolotl may update the terms of this Agreement from time to time, and
-        Licensee’s continued use of the Software after any such updates shall constitute acceptance of updated terms
-        on a go-forward basis.  Axolotl will use commercially reasonable efforts to provide Licensee notice of any
-        material updates. By using the Software, Licensee acknowledges that it has read, understood, and agrees to be
-        bound by the terms and conditions of this Agreement.
-
-This Agreement was last updated on August 23, 2024.