changed yml

.github/workflows/base.yml

@@ -36,29 +36,21 @@ jobs:
             python_version: "3.11"
             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: ""
-            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"
     steps:
       - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
       - name: Docker metadata
         id: metadata
-        uses: docker/metadata-action@v5
+        uses: docker/metadata-action@v3
         with:
-          images: |
-            winglian/axolotl-base
-            axolotlai/axolotl-base
+          images: winglian/axolotl-base
       - name: Login to Docker Hub
         uses: docker/login-action@v2
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3
+        uses: docker/setup-buildx-action@v2
       - name: Build
         uses: docker/build-push-action@v4
         with:

.github/workflows/docs.yml

@@ -17,7 +17,7 @@ jobs:
         - name: Set up Quarto
           uses: quarto-dev/quarto-actions/setup@v2
         - name: Setup Python
-          uses: actions/setup-python@v5
+          uses: actions/setup-python@v3
           with:
             python-version: '3.10'
         - name: install dependencies

.github/workflows/lint.yml

@@ -15,9 +15,9 @@ jobs:
     name: pre-commit
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v4
-      - uses: actions/setup-python@v5
+      - uses: actions/checkout@v3
+      - uses: actions/setup-python@v4
         with:
           python-version: "3.10"
           cache: 'pip' # caching pip dependencies
-      - uses: pre-commit/action@v3.0.1
+      - uses: pre-commit/action@v3.0.0

.github/workflows/main.yml

@@ -4,8 +4,6 @@ on:
   push:
     branches:
       - "main"
-    tags:
-      - "v*"
   workflow_dispatch:
 
 jobs:
@@ -31,11 +29,6 @@ jobs:
             python_version: "3.11"
             pytorch: 2.4.1
             axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.1
-            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -44,12 +37,7 @@ jobs:
         id: metadata
         uses: docker/metadata-action@v5
         with:
-          images: |
-            winglian/axolotl
-            axolotlai/axolotl
-          tags: |
-            type=ref,event=branch
-            type=semver,pattern={{version}}
+          images: winglian/axolotl
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
       - name: Login to Docker Hub
@@ -63,7 +51,7 @@ jobs:
         with:
           context: .
           build-args: |
-            BASE_TAG=${{ github.ref_type == 'tag' && 'main' || github.ref_name }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}
+            BASE_TAG=${{ github.ref_name }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}
             CUDA=${{ matrix.cuda }}
             PYTORCH_VERSION=${{ matrix.pytorch }}
             AXOLOTL_ARGS=${{ matrix.axolotl_args }}
@@ -98,11 +86,6 @@ jobs:
             python_version: "3.11"
             pytorch: 2.4.1
             axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.1
-            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -111,22 +94,20 @@ jobs:
         id: metadata
         uses: docker/metadata-action@v5
         with:
-          images: |
-            winglian/axolotl-cloud
-            axolotlai/axolotl-cloud
+          images: winglian/axolotl-cloud
       - name: Login to Docker Hub
         uses: docker/login-action@v3
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3
+        uses: docker/setup-buildx-action@v2
       - name: Build
         uses: docker/build-push-action@v5
         with:
           context: .
           build-args: |
-            BASE_TAG=${{ github.ref_type == 'tag' && 'main' || github.ref_name }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
+            BASE_TAG=${{ github.ref_name }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
             CUDA=${{ matrix.cuda }}
           file: ./docker/Dockerfile-cloud
           push: ${{ github.event_name != 'pull_request' }}
@@ -155,25 +136,20 @@ jobs:
         id: metadata
         uses: docker/metadata-action@v5
         with:
-          images: |
-            winglian/axolotl-cloud-term
-            axolotlai/axolotl-cloud-term
-          tags: |
-            type=ref,event=branch
-            type=semver,pattern={{version}}
+          images: winglian/axolotl-cloud-term
       - name: Login to Docker Hub
         uses: docker/login-action@v3
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3
+        uses: docker/setup-buildx-action@v2
       - name: Build
         uses: docker/build-push-action@v5
         with:
           context: .
           build-args: |
-            BASE_TAG=${{ github.ref_type == 'tag' && 'main' || github.ref_name }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
+            BASE_TAG=${{ github.ref_name }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
             CUDA=${{ matrix.cuda }}
           file: ./docker/Dockerfile-cloud-no-tmux
           push: ${{ github.event_name != 'pull_request' }}

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

@@ -21,17 +21,10 @@ jobs:
             pytorch: 2.3.1
             axolotl_extras:
             num_gpus: 2
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 121
+            cuda_version: 12.1.1
             python_version: "3.11"
-            pytorch: 2.4.1
-            axolotl_extras:
-            num_gpus: 2
-            nightly_build: "true"
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.1
+            pytorch: 2.3.1
             axolotl_extras:
             num_gpus: 2
             nightly_build: "true"

.github/workflows/nightlies.yml

@@ -28,11 +28,6 @@ jobs:
             python_version: "3.11"
             pytorch: 2.4.1
             axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.1
-            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -41,9 +36,7 @@ jobs:
         id: metadata
         uses: docker/metadata-action@v5
         with:
-          images: |
-            winglian/axolotl
-            axolotlai/axolotl
+          images: winglian/axolotl
           tags: |
             type=raw,value={{ branch }}-{{ date 'YYYYMMDD' }}
       - name: Set up Docker Buildx
@@ -92,11 +85,6 @@ jobs:
             python_version: "3.11"
             pytorch: 2.4.1
             axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.1
-            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -105,9 +93,7 @@ jobs:
         id: metadata
         uses: docker/metadata-action@v5
         with:
-          images: |
-            winglian/axolotl-cloud
-            axolotlai/axolotl-cloud
+          images: winglian/axolotl-cloud
           tags: |
             type=raw,value={{ branch }}-{{ date 'YYYYMMDD' }}
       - name: Login to Docker Hub
@@ -116,7 +102,7 @@ jobs:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3
+        uses: docker/setup-buildx-action@v2
       - name: Build
         uses: docker/build-push-action@v5
         with:

.github/workflows/pypi.yml

@@ -3,31 +3,12 @@ name: publish pypi
 on:
   push:
     tags:
-      - 'v*'
-  workflow_dispatch:
+      - '*'
 
 jobs:
-  setup_release:
-    name: Create Release
-    runs-on: ubuntu-latest
-    steps:
-      - name: Get the tag version
-        id: extract_branch
-        run: echo ::set-output name=branch::${GITHUB_REF#refs/tags/}
-        shell: bash
-
-      - name: Create Release
-        id: create_release
-        uses: actions/create-release@v1
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        with:
-          tag_name: ${{ steps.extract_branch.outputs.branch }}
-          release_name: ${{ steps.extract_branch.outputs.branch }}
   pypi-publish:
     name: Upload release to PyPI
     runs-on: ubuntu-latest
-    needs: [setup_release]
     environment:
       name: pypi
       url: https://pypi.org/p/axolotl
@@ -35,10 +16,10 @@ jobs:
       id-token: write  # IMPORTANT: this permission is mandatory for trusted publishing
     steps:
       - name: Check out repository code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
 
       - name: Setup Python
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v4
         with:
           python-version: "3.10"
 
@@ -46,7 +27,7 @@ jobs:
         run: |
           pip3 install wheel packaging
           pip3 install -e .
-          pip3 install -r requirements-dev.txt -r requirements-tests.txt
+          pip3 install -r requirements-tests.txt
 
       - name: Extract tag name
         id: tag

.github/workflows/tests-nightly.yml

@@ -9,12 +9,12 @@ jobs:
     name: pre-commit
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v4
-      - uses: actions/setup-python@v5
+      - uses: actions/checkout@v3
+      - uses: actions/setup-python@v4
         with:
           python-version: "3.10"
           cache: 'pip' # caching pip dependencies
-      - uses: pre-commit/action@v3.0.1
+      - uses: pre-commit/action@v3.0.0
         env:
           SKIP: no-commit-to-branch
 
@@ -25,15 +25,15 @@ jobs:
       fail-fast: false
       matrix:
         python_version: ["3.10", "3.11"]
-        pytorch_version: ["2.3.1", "2.4.1", "2.5.1"]
+        pytorch_version: ["2.3.1", "2.4.1"]
     timeout-minutes: 20
 
     steps:
       - name: Check out repository code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
 
       - name: Setup Python
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v4
         with:
           python-version: ${{ matrix.python_version }}
           cache: 'pip' # caching pip dependencies
@@ -47,14 +47,13 @@ jobs:
           sed -i 's#^transformers.*#transformers @ git+https://github.com/huggingface/transformers.git@main#' requirements.txt
           sed -i 's#^peft.*#peft @ git+https://github.com/huggingface/peft.git@main#' requirements.txt
           sed -i 's#^accelerate.*#accelerate @ git+https://github.com/huggingface/accelerate.git@main#' requirements.txt
-          sed -i 's#^trl.*#trl @ git+https://github.com/huggingface/trl.git@main#' requirements.txt
 
       - name: Install dependencies
         run: |
           pip3 install --upgrade pip
           pip3 install --upgrade packaging
           pip3 install -U -e .
-          pip3 install -r requirements-dev.txt -r requirements-tests.txt
+          pip3 install -r requirements-tests.txt
 
       - name: Run tests
         run: |
@@ -82,17 +81,17 @@ jobs:
             num_gpus: 1
             axolotl_extras: mamba-ssm
             nightly_build: "true"
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 121
+            cuda_version: 12.1.1
             python_version: "3.11"
-            pytorch: 2.4.1
+            pytorch: 2.3.1
             num_gpus: 1
-            axolotl_extras:
+            axolotl_extras: mamba-ssm
             nightly_build: "true"
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.5.1
+            pytorch: 2.4.1
             num_gpus: 1
             axolotl_extras:
             nightly_build: "true"

.github/workflows/tests.yml

@@ -15,22 +15,17 @@ on:
        - '.github/workflows/*.yml'
   workflow_dispatch:
 
-# Cancel jobs on the same ref if a new one is triggered
-concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
-  cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
-
 jobs:
   pre-commit:
     name: pre-commit
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v4
-      - uses: actions/setup-python@v5
+      - uses: actions/checkout@v3
+      - uses: actions/setup-python@v4
         with:
           python-version: "3.10"
           cache: 'pip' # caching pip dependencies
-      - uses: pre-commit/action@v3.0.1
+      - uses: pre-commit/action@v3.0.0
         env:
           SKIP: no-commit-to-branch
 
@@ -41,33 +36,29 @@ jobs:
       fail-fast: false
       matrix:
         python_version: ["3.10", "3.11"]
-        pytorch_version: ["2.3.1", "2.4.1", "2.5.1"]
+        pytorch_version: ["2.3.1", "2.4.1"]
     timeout-minutes: 20
 
     steps:
       - name: Check out repository code
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
 
       - name: Setup Python
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@v4
         with:
           python-version: ${{ matrix.python_version }}
           cache: 'pip' # caching pip dependencies
 
-      - name: upgrade pip
-        run: |
-          pip3 install --upgrade pip
-          pip3 install --upgrade packaging setuptools wheel
-
       - name: Install PyTorch
         run: |
-          pip3 install torch==${{ matrix.pytorch_version }}
+          pip3 install torch==${{ matrix.pytorch_version }} --index-url https://download.pytorch.org/whl/cpu
 
       - name: Install dependencies
         run: |
-          pip3 show torch
+          pip3 install --upgrade pip
+          pip3 install --upgrade packaging
           pip3 install -U -e .
-          pip3 install -r requirements-dev.txt -r requirements-tests.txt
+          pip3 install -r requirements-tests.txt
 
       - name: Run tests
         run: |
@@ -77,17 +68,29 @@ jobs:
         run: |
           find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
 
-  docker-e2e-tests-1st:
+  docker-e2e-tests:
     if: github.repository_owner == 'axolotl-ai-cloud'
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
-    timeout-minutes: 90
+    timeout-minutes: 60
     needs: [pre-commit, pytest]
 
     strategy:
       fail-fast: false
       matrix:
         include:
+          - cuda: 121
+            cuda_version: 12.1.1
+            python_version: "3.10"
+            pytorch: 2.3.1
+            num_gpus: 1
+            axolotl_extras: mamba-ssm
+          - cuda: 121
+            cuda_version: 12.1.1
+            python_version: "3.11"
+            pytorch: 2.3.1
+            num_gpus: 1
+            axolotl_extras: mamba-ssm
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
@@ -116,49 +119,3 @@ jobs:
       - name: Run tests job on Modal
         run: |
           modal run cicd.tests
-
-  docker-e2e-tests:
-    if: github.repository_owner == 'axolotl-ai-cloud'
-    # this job needs to be run on self-hosted GPU runners...
-    runs-on: [self-hosted, modal]
-    timeout-minutes: 90
-    needs: [pre-commit, pytest, docker-e2e-tests-1st]
-
-    strategy:
-      fail-fast: false
-      matrix:
-        include:
-          - cuda: 121
-            cuda_version: 12.1.1
-            python_version: "3.10"
-            pytorch: 2.3.1
-            num_gpus: 1
-            axolotl_extras: mamba-ssm
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.1
-            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"
-      - name: Install Modal
-        run: |
-          python -m pip install --upgrade pip
-          pip install modal==0.63.64 jinja2
-      - name: Update env vars
-        run: |
-          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
-          echo "PYTORCH_VERSION=${{ matrix.pytorch}}" >> $GITHUB_ENV
-          echo "AXOLOTL_ARGS=${{ matrix.axolotl_args}}" >> $GITHUB_ENV
-          echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV
-          echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
-          echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
-      - name: Run tests job on Modal
-        run: |
-          modal run cicd.tests

README.md

@@ -121,7 +121,7 @@ Features:
 
 Get started with Axolotl in just a few steps! This quickstart guide will walk you through setting up and running a basic fine-tuning task.
 
-**Requirements**: Nvidia GPU (Ampere architecture or newer for `bf16` and Flash Attention), Python >=3.10 and PyTorch >=2.3.1.
+**Requirements**: Python >=3.10 and Pytorch >=2.1.1.
 
 ```bash
 git clone https://github.com/axolotl-ai-cloud/axolotl
@@ -159,7 +159,7 @@ accelerate launch -m axolotl.cli.train https://raw.githubusercontent.com/axolotl
 #### Docker
 
   ```bash
-  docker run --gpus '"all"' --rm -it axolotlai/axolotl:main-latest
+  docker run --gpus '"all"' --rm -it winglian/axolotl:main-latest
   ```
 
   Or run on the current files for development:
@@ -178,7 +178,7 @@ accelerate launch -m axolotl.cli.train https://raw.githubusercontent.com/axolotl
   A more powerful Docker command to run would be this:
 
   ```bash
-docker run --privileged --gpus '"all"' --shm-size 10g --rm -it --name axolotl --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --mount type=bind,src="${PWD}",target=/workspace/axolotl -v ${HOME}/.cache/huggingface:/root/.cache/huggingface axolotlai/axolotl:main-latest
+docker run --privileged --gpus '"all"' --shm-size 10g --rm -it --name axolotl --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --mount type=bind,src="${PWD}",target=/workspace/axolotl -v ${HOME}/.cache/huggingface:/root/.cache/huggingface winglian/axolotl:main-latest
   ```
 
   It additionally:
@@ -210,7 +210,7 @@ docker run --privileged --gpus '"all"' --shm-size 10g --rm -it --name axolotl --
 
 #### Cloud GPU
 
-For cloud GPU providers that support docker images, use [`axolotlai/axolotl-cloud:main-latest`](https://hub.docker.com/r/axolotlai/axolotl-cloud/tags)
+For cloud GPU providers that support docker images, use [`winglian/axolotl-cloud:main-latest`](https://hub.docker.com/r/winglian/axolotl-cloud/tags)
 
 - on Latitude.sh use this [direct link](https://latitude.sh/blueprint/989e0e79-3bf6-41ea-a46b-1f246e309d5c)
 - on JarvisLabs.ai use this [direct link](https://jarvislabs.ai/templates/axolotl)
@@ -319,7 +319,7 @@ Write a job description in YAML as below:
 # dstack.yaml
 type: task
 
-image: axolotlai/axolotl-cloud:main-latest
+image: winglian/axolotl-cloud:main-20240429-py3.11-cu121-2.2.2
 
 env:
   - HUGGING_FACE_HUB_TOKEN
@@ -383,10 +383,11 @@ See [examples](examples) for quick start. It is recommended to duplicate and mod
         - typescript
       type: ... # unimplemented custom format
 
-      # chat_template https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/conversation.html#chat_template
+      # fastchat conversation (deprecation soon, use chat_template)
+      # See 'conversation' options: https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
     - path: ...
-      type: chat_template
-      chat_template: chatml # defaults to tokenizer's chat_template
+      type: sharegpt
+      conversation: chatml # default: vicuna_v1.1
 
       # local
     - path: data.jsonl # or json
@@ -561,8 +562,7 @@ plugins:
   - axolotl.integrations.liger.LigerPlugin
 liger_rope: true
 liger_rms_norm: true
-liger_glu_activation: true
-liger_layer_norm: true
+liger_swiglu: true
 liger_fused_linear_cross_entropy: true
 ```
 

cicd/Dockerfile.jinja

@@ -1,4 +1,4 @@
-FROM axolotlai/axolotl-base:{{ BASE_TAG }}
+FROM winglian/axolotl-base:{{ BASE_TAG }}
 
 ENV TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX"
 ENV AXOLOTL_EXTRAS="{{ AXOLOTL_EXTRAS }}"
@@ -23,11 +23,11 @@ RUN git fetch origin +$GITHUB_REF && \
     git checkout FETCH_HEAD
 
 # If AXOLOTL_EXTRAS is set, append it in brackets
+RUN pip install causal_conv1d
 RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
         sed -i 's#^transformers.*#transformers @ git+https://github.com/huggingface/transformers.git@main#' requirements.txt; \
         sed -i 's#^peft.*#peft @ git+https://github.com/huggingface/peft.git@main#' requirements.txt; \
         sed -i 's#^accelerate.*#accelerate @ git+https://github.com/huggingface/accelerate.git@main#' requirements.txt; \
-        sed -i 's#^trl.*#trl @ git+https://github.com/huggingface/trl.git@main#' requirements.txt; \
     fi
 
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
@@ -37,7 +37,7 @@ RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
     fi
 
 # So we can test the Docker image
-RUN pip install -r requirements-dev.txt -r requirements-tests.txt
+RUN pip install -r requirements-tests.txt
 
 # fix so that git fetch/pull from remote works
 RUN git config remote.origin.fetch "+refs/heads/*:refs/remotes/origin/*" && \

cicd/cicd.sh

@@ -1,6 +1,6 @@
 #!/bin/bash
 set -e
 
-pytest -n4 --ignore=tests/e2e/ /workspace/axolotl/tests/
+pytest --ignore=tests/e2e/ /workspace/axolotl/tests/
 pytest -n1 --dist loadfile -v /workspace/axolotl/tests/e2e/patched/ /workspace/axolotl/tests/e2e/integrations/
 pytest --ignore=tests/e2e/patched/ --ignore=tests/e2e/multigpu/ --ignore=tests/e2e/integrations/ /workspace/axolotl/tests/e2e/

cicd/multigpu.py

@@ -10,7 +10,7 @@ import tempfile
 import jinja2
 import modal
 from jinja2 import select_autoescape
-from modal import App, Image
+from modal import Image, Stub
 
 cicd_path = pathlib.Path(__file__).parent.resolve()
 
@@ -46,7 +46,7 @@ cicd_image = (
     .pip_install("fastapi==0.110.0", "pydantic==2.6.3")
 )
 
-app = App("Axolotl CI/CD", secrets=[])
+stub = Stub("Axolotl CI/CD", secrets=[])
 
 
 N_GPUS = int(os.environ.get("N_GPUS", 2))
@@ -61,10 +61,10 @@ def run_cmd(cmd: str, run_folder: str):
         exit(exit_code)  # pylint: disable=consider-using-sys-exit
 
 
-@app.function(
+@stub.function(
     image=cicd_image,
     gpu=GPU_CONFIG,
-    timeout=60 * 60,
+    timeout=45 * 60,
     cpu=8.0,
     memory=131072 * N_GPUS,
 )
@@ -72,6 +72,6 @@ def cicd_pytest():
     run_cmd("./cicd/multigpu.sh", "/workspace/axolotl")
 
 
-@app.local_entrypoint()
+@stub.local_entrypoint()
 def main():
     cicd_pytest.remote()

cicd/tests.py

@@ -10,7 +10,7 @@ import tempfile
 import jinja2
 import modal
 from jinja2 import select_autoescape
-from modal import App, Image
+from modal import Image, Stub
 
 cicd_path = pathlib.Path(__file__).parent.resolve()
 
@@ -47,7 +47,7 @@ cicd_image = (
     .pip_install("fastapi==0.110.0", "pydantic==2.6.3")
 )
 
-app = App("Axolotl CI/CD", secrets=[])
+stub = Stub("Axolotl CI/CD", secrets=[])
 
 
 N_GPUS = int(os.environ.get("N_GPUS", 1))
@@ -62,10 +62,10 @@ def run_cmd(cmd: str, run_folder: str):
         exit(exit_code)  # pylint: disable=consider-using-sys-exit
 
 
-@app.function(
+@stub.function(
     image=cicd_image,
     gpu=GPU_CONFIG,
-    timeout=60 * 60,
+    timeout=45 * 60,
     cpu=8.0,
     memory=131072,
 )
@@ -73,6 +73,6 @@ def cicd_pytest():
     run_cmd("./cicd/cicd.sh", "/workspace/axolotl")
 
 
-@app.local_entrypoint()
+@stub.local_entrypoint()
 def main():
     cicd_pytest.remote()

deepspeed_configs/zero3_bf16.json

@@ -14,6 +14,15 @@
   "bf16": {
     "enabled": true
   },
+  "fp16": {
+    "enabled": "auto",
+    "auto_cast": false,
+    "loss_scale": 0,
+    "initial_scale_power": 32,
+    "loss_scale_window": 1000,
+    "hysteresis": 2,
+    "min_loss_scale": 1
+  },
   "gradient_accumulation_steps": "auto",
   "gradient_clipping": "auto",
   "train_batch_size": "auto",

deepspeed_configs/zero3_bf16_cpuoffload_all.json

@@ -24,6 +24,15 @@
   "bf16": {
     "enabled": true
   },
+  "fp16": {
+    "enabled": "auto",
+    "auto_cast": false,
+    "loss_scale": 0,
+    "initial_scale_power": 32,
+    "loss_scale_window": 1000,
+    "hysteresis": 2,
+    "min_loss_scale": 1
+  },
   "gradient_accumulation_steps": "auto",
   "gradient_clipping": "auto",
   "train_batch_size": "auto",

deepspeed_configs/zero3_bf16_cpuoffload_params.json

@@ -20,6 +20,15 @@
   "bf16": {
     "enabled": true
   },
+  "fp16": {
+    "enabled": "auto",
+    "auto_cast": false,
+    "loss_scale": 0,
+    "initial_scale_power": 32,
+    "loss_scale_window": 1000,
+    "hysteresis": 2,
+    "min_loss_scale": 1
+  },
   "gradient_accumulation_steps": "auto",
   "gradient_clipping": "auto",
   "train_batch_size": "auto",

devtools/dev_sharegpt.yml

@@ -1,4 +1,4 @@
-# Example config for debugging the chat_template prompt format
+# Example config for debugging the sharegpt prompt format
 base_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
@@ -7,8 +7,8 @@ load_in_8bit: true
 load_in_4bit: false
 
 datasets:
-  - path: fozziethebeat/alpaca_messages_2k_test
-    type: chat_template
+  - path: philschmid/guanaco-sharegpt-style
+    type: sharegpt
     shards: 10
 val_set_size: 0
 output_dir: temp_debug/axolotl_outputs/model

docker/Dockerfile

@@ -1,5 +1,5 @@
 ARG BASE_TAG=main-base
-FROM axolotlai/axolotl-base:$BASE_TAG
+FROM winglian/axolotl-base:$BASE_TAG
 
 ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX"
 ARG AXOLOTL_EXTRAS=""
@@ -20,6 +20,7 @@ RUN git clone --depth=1 https://github.com/axolotl-ai-cloud/axolotl.git
 WORKDIR /workspace/axolotl
 
 # If AXOLOTL_EXTRAS is set, append it in brackets
+RUN pip install causal_conv1d
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
         pip install -e .[deepspeed,flash-attn,optimizers,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \

docker/Dockerfile-cloud

@@ -1,5 +1,5 @@
 ARG BASE_TAG=main
-FROM axolotlai/axolotl:$BASE_TAG
+FROM winglian/axolotl:$BASE_TAG
 
 ENV HF_DATASETS_CACHE="/workspace/data/huggingface-cache/datasets"
 ENV HUGGINGFACE_HUB_CACHE="/workspace/data/huggingface-cache/hub"

docker/Dockerfile-cloud-no-tmux

@@ -1,5 +1,5 @@
 ARG BASE_TAG=main
-FROM axolotlai/axolotl:$BASE_TAG
+FROM winglian/axolotl:$BASE_TAG
 
 ENV HF_DATASETS_CACHE="/workspace/data/huggingface-cache/datasets"
 ENV HUGGINGFACE_HUB_CACHE="/workspace/data/huggingface-cache/hub"

docker/Dockerfile-tests

@@ -1,5 +1,5 @@
 ARG BASE_TAG=main-base
-FROM axolotlai/axolotl-base:$BASE_TAG
+FROM winglian/axolotl-base:$BASE_TAG
 
 ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX"
 ARG AXOLOTL_EXTRAS=""

docs/config.qmd

@@ -83,7 +83,7 @@ lora_on_cpu: true
 datasets:
   # HuggingFace dataset repo | s3://,gs:// path | "json" for local dataset, make sure to fill data_files
   - path: vicgalle/alpaca-gpt4
-    # The type of prompt to use for training. [alpaca, gpteacher, oasst, reflection]
+  # The type of prompt to use for training. [alpaca, sharegpt, gpteacher, oasst, reflection]
     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
@@ -92,6 +92,15 @@ datasets:
     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.
 
+    # Optional[str] fastchat conversation type, only used with type: sharegpt
+    conversation: # Options (see Conversation 'name'): https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
+    field_human: # Optional[str]. Human key to use for conversation.
+    field_model: # Optional[str]. Assistant key to use for conversation.
+    # Add additional keys from your dataset as input or output roles
+    roles:
+      input: # Optional[List[str]]. These will be masked based on train_on_input
+      output: # Optional[List[str]].
+
   # Custom user instruction prompt
   - path: repo
     type:
@@ -115,48 +124,6 @@ datasets:
       # For `completion` datsets only, uses the provided field instead of `text` column
       field:
 
-  # Using chat template
-  - path: ...
-    # Set type to `chat_template` to use this strategy
-    type: chat_template
-    # Specify the name of the chat template to use
-    # The name of the chat template to use for training, following values are supported:
-    # - tokenizer_default: Uses the chat template that is available in the tokenizer_config.json. If the chat template is not available in the tokenizer, it will raise an error. This is the default.
-    # - alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates are available in the axolotl codebase at src/axolotl/utils/chat_templates.py
-    # - tokenizer_default_fallback_*: where * is the name of the chat template to fallback to if the tokenizer does not have a chat template else default to tokenizer. E.g. tokenizer_default_fallback_chatml.
-    # - jinja: Uses a custom jinja template for the chat template. The custom jinja template should be provided in the chat_template_jinja field.
-    chat_template: tokenizer_default
-    # Custom jinja template for chat template. This will be only used if `chat_template` is set to `jinja` or empty (in which case chat_template is automatically set to `jinja`).
-    chat_template_jinja:
-    # The key in the data example that contains the messages. Default is "messages".
-    field_messages: messages
-    # The key in the message turn that contains the role. Default is "role".
-    message_field_role: role
-    # The key in the message turn that contains the content. Default is "content".
-    message_field_content: content
-    # Optional[Dict[str, List]]. Roles mapping for the messages.
-    roles:
-      user: ["human", "user"]
-      assistant: ["gpt", "assistant", "ai"]
-      system: ["system"]
-
-    ## NOTE: Leaving the below empty will default to using the simple legacy tokenization strategy where only last message is trained on.
-
-    # Optional[List[str]]. Roles to train on. The tokens from these roles will be considered for the loss.
-    roles_to_train: ["gpt", "assistant"]
-    # Optional[str]. Which EOS tokens to train on in the conversation. Possible values are:
-    # - all: train on all EOS tokens
-    # - turn: train on the EOS token at the end of each trainable turn
-    # - last: train on the last EOS token in the conversation
-    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
-    # The key in the message turn that contains the training details. Useful to selectively train on certain tokens in a turn.
-    # The value of the key is a List[Dict] containing `begin_offset` (start character index in content), `end_offset` (end character index in content), and `train` (boolean whether to train).
-    # See example at `docs/dataset-formats/conversation.qmd`
-    message_field_training_detail: train_detail
-
-
 # If false, the datasets will not be shuffled and will keep their original order in `datasets`.
 # The same applies to the `test_datasets` option and the `pretraining_dataset` option. Default is true.
 shuffle_merged_datasets: true
@@ -174,19 +141,10 @@ test_datasets:
 
 # use RL training: 'dpo', 'ipo', 'kto'
 rl:
-# whether to perform weighting if doing DPO training. Boolean.
-dpo_use_weighting:
 
-# The name of the chat template to use for training, following values are supported:
-# - tokenizer_default: Uses the chat template that is available in the tokenizer_config.json. If the chat template is not available in the tokenizer, it will raise an error. This is the default value.
-# - alpaca/inst/chatml/gemma/cohere/llama3/phi_3/deepseek_v2/jamba: These chat templates are available in the axolotl codebase at src/axolotl/utils/chat_templates.py
-# - tokenizer_default_fallback_*: where * is the name of the chat template to fallback to. E.g. tokenizer_default_fallback_chatml. This is useful when the chat template is not available in the tokenizer.
-# - jinja: Uses a custom jinja template for the chat template. The custom jinja template should be provided in the chat_template_jinja field.
-# The selected chat template will be saved to the tokenizer_config.json for easier inferencing
-# Note: It is recommended to set train_on_inputs to true when using a chat template that is different from the model's default chat template.
-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
+# Saves the desired chat template to the tokenizer_config.json for easier inferencing
+# Currently supports chatml and inst (mistral/mixtral)
+chat_template: chatml
 # Changes the default system message
 default_system_message: You are a helpful assistant. Please give a long and detailed answer. # Currently only supports chatml.
 # Axolotl attempts to save the dataset as an arrow after packing the data together so
@@ -405,7 +363,6 @@ lr_div_factor: # Learning rate div factor
 # - adamw_torch_fused
 # - adamw_torch_xla
 # - adamw_apex_fused
-# - adopt_adamw (only for torch version >= 2.5.1)
 # - adafactor
 # - adamw_anyprecision
 # - sgd

docs/dataset-formats/conversation.qmd

@@ -6,8 +6,31 @@ order: 3
 
 ## sharegpt
 
-IMPORTANT: ShareGPT is deprecated!. Please see `chat_template` section below.
+conversations where `from` is `human`/`gpt`. (optional: first row with role `system` to override default system prompt)
 
+```{.json filename="data.jsonl"}
+{"conversations": [{"from": "...", "value": "..."}]}
+```
+
+Note: `type: sharegpt` opens special configs:
+- `conversation`: enables conversions to many Conversation types. Refer to the 'name' [here](https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py) for options.
+- `roles`: allows you to specify the roles for input and output. This is useful for datasets with custom roles such as `tool` etc to support masking.
+- `field_human`: specify the key to use instead of `human` in the conversation.
+- `field_model`: specify the key to use instead of `gpt` in the conversation.
+
+```yaml
+datasets:
+    path: ...
+    type: sharegpt
+
+    conversation: # Options (see Conversation 'name'): https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
+    field_human: # Optional[str]. Human key to use for conversation.
+    field_model: # Optional[str]. Assistant key to use for conversation.
+    # Add additional keys from your dataset as input or output roles
+    roles:
+      input: # Optional[List[str]]. These will be masked based on train_on_input
+      output: # Optional[List[str]].
+```
 
 ## pygmalion
 
@@ -15,137 +38,34 @@ IMPORTANT: ShareGPT is deprecated!. Please see `chat_template` section below.
 {"conversations": [{"role": "...", "value": "..."}]}
 ```
 
+## sharegpt.load_role
 
-## chat_template
-
-Chat Template strategy uses a jinja2 template that converts a list of messages into a prompt. Support using tokenizer's template, a supported template, or custom jinja2.
+conversations where `role` is used instead of `from`
 
 ```{.json filename="data.jsonl"}
-{"conversations": [{"role": "...", "content": "..."}]}
+{"conversations": [{"role": "...", "value": "..."}]}
 ```
 
-See `config.qmd` for full configs and supported templates.
+## sharegpt.load_guanaco
 
-### Migrating from sharegpt
-
-Most configs can be adapted as follows:
-
-```yaml
-# old
-chat_template: chatml
-datasets:
-  - path: ...
-    type: sharegpt
-    conversation: chatml
-
-# new (if using tokenizer's chat_template)
-datasets:
-  - path: ...
-    type: chat_template
-
-    field_messages: conversations
-    message_field_role: from
-    message_field_content: value
-
-# new (if setting a new chat_template like chatml, gemma, etc)
-chat_template: chatml
-datasets:
-  - path: ...
-    type: chat_template
-
-    field_messages: conversations
-    message_field_role: from
-    message_field_content: value
-```
-
-We recommend checking the below examples for other usecases.
-
-### Examples
-
-1. Using the default chat template in the tokenizer_config.json on OpenAI messages format, training on only last message.
-
-```yaml
-datasets:
-  - path: ...
-    type: 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
-chat_template: gemma # this overwrites the tokenizer's chat_template
-datasets:
-  - path: ...
-    type: chat_template
-    roles_to_train: ["assistant"]
-```
-
-3. Using the tokenizer_config.json's chat template or `chatml` as fallback if the former's chat template does not exist, on OpenAI messages format, training on all assistant messages.
-
-```yaml
-chat_template: tokenizer_default_fallback_chatml # this overwrites the tokenizer's chat_template
-datasets:
-  - path: ...
-    type: chat_template
-    roles_to_train: ["assistant"]
-```
-
-4. Using a custom jinja template on OpenAI messages format, training on all assistant messages.
-
-```yaml
-# chat_template: jinja # `jinja` will be implied if the `chat_template_jinja` is set and this field is empty
-chat_template_jinja: "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'system') %}{{'<|system|>' + '\n' + message['content'] + '<|end|>' + '\n'}}{% elif (message['role'] == 'user') %}{{'<|user|>' + '\n' + message['content'] + '<|end|>' + '\n' + '<|assistant|>' + '\n'}}{% elif message['role'] == 'assistant' %}{{message['content'] + '<|end|>' + '\n'}}{% endif %}{% endfor %}"
-
-datasets:
-  - path: ...
-    type: chat_template
-    roles_to_train: ["assistant"]
-```
-
-5. (Advanced) Using fine-grained control over tokens and turns to train in a conversation
-
-For a data sample that looks like:
+conversations where `from` is `prompter` `assistant` instead of default sharegpt
 
 ```{.json filename="data.jsonl"}
-{
-  "conversations": [
-    {"from": "system", "value": "You are an AI assistant.", "train": false},
-    {"from": "human", "value": "Hello", "train": false},
-    {"from": "assistant", "value": "Hello", "train": true},
-    {"from": "human", "value": "How are you?", "train": true},
-    {
-      "from": "assistant",
-      "value": "I'm doing very well, thank you!",
-      "train_detail": [
-        {"begin_offset": 0, "end_offset": 8, "train": false},
-        {"begin_offset": 9, "end_offset": 18, "train": true},
-        {"begin_offset": 19, "end_offset": 30, "train": false},
-      ],
-    },
-    {
-        "from": "human",
-        "value": "I'm doing very well, thank you!",
-        "train": true,
-    },
-    {"from": "assistant", "value": "Hi there!", "train": true}
-  ]
-}
+{"conversations": [{"from": "...", "value": "..."}]}
 ```
 
-The configuration would look like:
+## sharegpt.load_ultrachat
 
-```yaml
-datasets:
-  - path: ...
-    type: chat_template
-    chat_template: tokenizer_default
-    field_messages: conversations
-    message_field_role: from
-    message_field_content: value
-    roles_to_train: []
-    train_on_eos: turn
-    message_field_training: train
-    message_field_training_detail: train_detail
+conversations where the turns field is 'messages', human is 'user' and gpt is 'assistant'.
+
+```{.json filename="data.jsonl"}
+{"messages": [{"user": "...", "assistant": "..."}]}
 ```
 
-Tip: It is not necessary to use both `message_field_training` and `message_field_training_detail` at a time.
+## sharegpt_jokes
+
+creates a chat where bot is asked to tell a joke, then explain why the joke is funny
+
+```{.json filename="data.jsonl"}
+{"conversations": [{"title": "...", "text": "...", "explanation": "..."}]}
+```

docs/debugging.qmd

@@ -51,12 +51,12 @@ While debugging it's helpful to simplify your test scenario as much as possible.
 
 ### Background
 
-The below example shows how to configure VSCode to debug data preprocessing of the `chat_template` format.  This is the format used when you have the following in your axolotl config:
+The below example shows how to configure VSCode to debug data preprocessing of the `sharegpt` format.  This is the format used when you have the following in your axolotl config:
 
 ```yaml
 datasets:
-  - path: <path to your chat_template formatted dataset> # example on HF Hub: fozziethebeat/alpaca_messages_2k_test
-    type: chat_template
+  - path: <path to your sharegpt formatted dataset> # example on HF Hub: philschmid/guanaco-sharegpt-style
+    type: sharegpt
 ```
 
 >[!Important]
@@ -83,7 +83,7 @@ If you developing on a remote host, you can easily use VSCode to debug remotely.
 
 The easiest way to get started is to modify the [.vscode/launch.json](../.vscode/launch.json) file in this project.  This is just an example configuration, so you may need to modify or copy it to suit your needs.
 
-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.
+For example, to mimic the command `cd devtools && CUDA_VISIBLE_DEVICES=0 accelerate launch -m axolotl.cli.train dev_sharegpt.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
 // .vscode/launch.json
@@ -91,12 +91,12 @@ For example, to mimic the command `cd devtools && CUDA_VISIBLE_DEVICES=0 acceler
     "version": "0.2.0",
     "configurations": [
         {
-            "name": "Debug axolotl prompt - chat_template",
+            "name": "Debug axolotl prompt - sharegpt",
             "type": "python",
             "module": "accelerate.commands.launch",
             "request": "launch",
             "args": [
-                "-m", "axolotl.cli.train", "dev_chat_template.yml",
+                "-m", "axolotl.cli.train", "dev_sharegpt.yml",
                 // The flags below simplify debugging by overriding the axolotl config
                 // with the debugging tips above.  Modify as needed.
                 "--dataset_processes=1",      // limits data preprocessing to one process
@@ -185,7 +185,7 @@ style="border-radius: 10px; display: block; margin: auto;" width="560" height="3
 
 ## Debugging With Docker
 
-Using [official Axolotl Docker images](https://hub.docker.com/r/axolotlai/axolotl/tags) is a great way to debug your code, and is a very popular way to use Axolotl.  Attaching VSCode to Docker takes a few more steps.
+Using [official Axolotl Docker images](https://hub.docker.com/r/winglian/axolotl/tags) is a great way to debug your code, and is a very popular way to use Axolotl.  Attaching VSCode to Docker takes a few more steps.
 
 ### Setup
 
@@ -202,11 +202,11 @@ cd axolotl
 Next, run the desired docker image and mount the current directory. Below is a docker command you can run to do this:[^2]
 
 ```bash
-docker run --privileged --gpus '"all"' --shm-size 10g --rm -it --name axolotl --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --mount type=bind,src="${PWD}",target=/workspace/axolotl -v ${HOME}/.cache/huggingface:/root/.cache/huggingface axolotlai/axolotl:main-py3.10-cu118-2.0.1
+docker run --privileged --gpus '"all"' --shm-size 10g --rm -it --name axolotl --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --mount type=bind,src="${PWD}",target=/workspace/axolotl -v ${HOME}/.cache/huggingface:/root/.cache/huggingface winglian/axolotl:main-py3.10-cu118-2.0.1
 ```
 
 >[!Tip]
-> To understand which containers are available, see the [Docker section of the README](../README.md#docker) and the [DockerHub repo](https://hub.docker.com/r/axolotlai/axolotl/tags).  For details of how the Docker containers are built, see axolotl's [Docker CI builds](../.github/workflows/main.yml).
+> To understand which containers are available, see the [Docker section of the README](../README.md#docker) and the [DockerHub repo](https://hub.docker.com/r/winglian/axolotl/tags).  For details of how the Docker containers are built, see axolotl's [Docker CI builds](../.github/workflows/main.yml).
 
 You will now be in the container.  Next, perform an editable install of Axolotl:
 
@@ -240,6 +240,6 @@ style="border-radius: 10px; display: block; margin: auto;" width="560" height="3
 </div>
 <br>
 
-[^1]: The config actually mimics the command `CUDA_VISIBLE_DEVICES=0 python -m accelerate.commands.launch -m axolotl.cli.train devtools/chat_template.yml`, but this is the same thing.
+[^1]: The config actually mimics the command `CUDA_VISIBLE_DEVICES=0 python -m accelerate.commands.launch -m axolotl.cli.train devtools/sharegpt.yml`, but this is the same thing.
 
 [^2]: Many of the below flags are recommended best practices by Nvidia when using nvidia-container-toolkit.  You can read more about these flags [here](https://docs.nvidia.com/deeplearning/frameworks/user-guide/index.html).

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

@@ -44,7 +44,7 @@
    "outputs": [],
    "source": [
     "!pip install -e git+https://github.com/axolotl-ai-cloud/axolotl#egg=axolotl\n",
-    "!pip install flash-attn==\"2.7.0.post2\"\n",
+    "!pip install flash-attn==\"2.5.0\"\n",
     "!pip install deepspeed==\"0.13.1\"!pip install mlflow==\"2.13.0\""
    ]
   },

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

@@ -9,17 +9,14 @@ strict: false
 plugins:
   - axolotl.integrations.liger.LigerPlugin
 liger_rms_norm: true
-liger_glu_activation: true
+liger_swiglu: true
 liger_fused_linear_cross_entropy: true
 
 chat_template: deepseek_v2
 datasets:
   - path: mlabonne/FineTome-100k
     type: chat_template
-    split: train[:20%]
-    field_messages: conversations
-    message_field_role: from
-    message_field_content: value
+    split: train
 
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0

examples/gemma2/qlora.yml

@@ -11,11 +11,8 @@ chat_template: gemma
 datasets:
   - path: cgato/SlimOrcaDedupCleaned
     type: chat_template
+    chat_template: gemma
     drop_system_message: true
-    field_messages: conversations
-    message_field_role: from
-    message_field_content: value
-
 val_set_size: 0.0
 output_dir: ./outputs/out
 

examples/jamba/qlora_fsdp_large.yaml

@@ -4,15 +4,11 @@ tokenizer_type: AutoTokenizer
 load_in_4bit: true
 strict: false
 use_tensorboard: true
-chat_template: jamba
 datasets:
   - path: cgato/SlimOrcaDedupCleaned
     type: chat_template
+    chat_template: jamba
     drop_system_message: true
-    field_messages: conversations
-    message_field_role: from
-    message_field_content: value
-
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0
 output_dir: jamba-large-fsdp-qlora-ft

examples/llama-3/fft-8b-fsdp.yml

@@ -0,0 +1,50 @@
+base_model: meta-llama/Llama-3.1-8B-Instruct
+
+save_safetensors: true
+
+datasets:
+  - path: teknium/GPT4-LLM-Cleaned
+    type: alpaca
+
+dataset_prepared_path: ./last_run_prepared
+
+output_dir: ./outputs/fft-out
+sequence_len: 8192
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch
+learning_rate: 2e-5
+
+bf16: auto
+fp16:
+tf32: false
+
+logging_steps: 2
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 2
+evals_per_epoch: 2
+save_steps: 2
+max_steps: 2
+weight_decay: 0.0
+
+fsdp:
+  - full_shard
+  - auto_wrap
+fsdp_config:
+  fsdp_limit_all_gathers: true
+  fsdp_sync_module_states: true
+  fsdp_offload_params: false
+  fsdp_use_orig_params: true
+  fsdp_cpu_ram_efficient_loading: false
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_backward_prefetch: BACKWARD_PRE
+
+special_tokens:
+  pad_token: "<|end_of_text|>"

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

@@ -4,7 +4,7 @@ plugins:
   - axolotl.integrations.liger.LigerPlugin
 liger_rope: true
 liger_rms_norm: true
-liger_glu_activation: true
+liger_swiglu: true
 liger_fused_linear_cross_entropy: true
 
 strict: false
@@ -14,10 +14,6 @@ datasets:
   - path: mlabonne/FineTome-100k
     type: chat_template
     split: train[:20%]
-    field_messages: conversations
-    message_field_role: from
-    message_field_content: value
-
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.02
 output_dir: ./outputs/out

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

@@ -1,77 +0,0 @@
-base_model: meta-llama/Llama-3.2-1B
-
-load_in_8bit: false
-load_in_4bit: true
-strict: false
-
-datasets:
-  - path: teknium/GPT4-LLM-Cleaned
-    type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.1
-output_dir: ./outputs/qlora-out
-
-adapter: qlora
-lora_model_dir:
-
-sequence_len: 2048
-sample_packing: true
-eval_sample_packing: true
-pad_to_sequence_len: true
-
-lora_r: 32
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_linear: true
-lora_fan_in_fan_out:
-lora_target_modules:
-  - gate_proj
-  - down_proj
-  - up_proj
-  - q_proj
-  - v_proj
-  - k_proj
-  - o_proj
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 4
-micro_batch_size: 2
-num_epochs: 1
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-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
-eval_table_size:
-eval_max_new_tokens: 128
-saves_per_epoch: 1
-debug:
-deepspeed:
-weight_decay: 0.0
-fsdp:
-fsdp_config:
-special_tokens:
-  pad_token: "<|end_of_text|>"

examples/mistral/mistral-dpo-qlora.yml

@@ -1,93 +0,0 @@
-#Note that we are switching from the regular chat template to chatml.
-#If you experience problems with the special tokens, training for more epochs can help.
-#After training, merge the model before inference otherwise you might
-#face problems with the special tokens.
-
-base_model: mistralai/Mistral-7B-Instruct-v0.2
-model_type: MistralForCausalLM
-tokenizer_type: LlamaTokenizer
-
-load_in_8bit: false
-load_in_4bit: true
-strict: false
-
-chat_template: chatml
-rl: dpo
-datasets:
-  - path: olivermolenschot/alpaca_messages_dpo_test
-    type: chat_template.default
-    field_messages: conversation
-    field_chosen: chosen
-    field_rejected: rejected
-    message_field_role: role
-    message_field_content: content
-
-dataset_prepared_path:
-val_set_size: 0.05
-output_dir: ./outputs/dpo-qlora
-
-sequence_len: 2048
-sample_packing: false
-pad_to_sequence_len: true
-
-adapter: qlora
-lora_model_dir:
-lora_r: 8
-lora_alpha: 16
-lora_dropout: 0.2
-lora_target_linear: true
-lora_fan_in_fan_out:
-
-lora_target_modules:
-  - gate_proj
-  - down_proj
-  - up_proj
-  - q_proj
-  - v_proj
-  - k_proj
-  - o_proj
-lora_modules_to_save:
- - embed_tokens
- - lm_head
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 4
-micro_batch_size: 16
-num_epochs: 6
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0001
-
-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: false
-s2_attention:
-
-warmup_steps: 10
-evals_per_epoch: 4
-eval_table_size:
-eval_max_new_tokens: 128
-saves_per_epoch: 1
-debug:
-deepspeed:
-weight_decay: 0.0
-fsdp:
-fsdp_config:
-special_tokens:
-  bos_token: "<|im_start|>"
-  eos_token: "<|im_end|>"

examples/phi/lora-3.5.yaml

@@ -10,6 +10,7 @@ chat_template: phi_3
 datasets:
   - path: fozziethebeat/alpaca_messages_2k_test
     type: chat_template
+    chat_template: phi_3
     field_messages: messages
     message_field_role: role
     message_field_content: content

requirements-dev.txt

@@ -2,4 +2,3 @@ pre-commit
 black
 mypy
 types-requests
-tbparse

requirements-tests.txt

@@ -1,3 +1,2 @@
 pytest
 pytest-xdist
-pytest-retry

requirements.txt

@@ -1,22 +1,22 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
 packaging==23.2
 peft==0.13.2
-transformers==4.46.1
+transformers==4.45.2
 tokenizers>=0.20.1
 bitsandbytes==0.44.1
-accelerate==1.1.0
+accelerate==1.0.1
 datasets==3.0.1
-deepspeed==0.15.3
+deepspeed==0.14.4
 pydantic==2.6.3
 addict
 fire
 PyYAML>=6.0
 requests
-flash-attn==2.7.0.post2
+flash-attn==2.6.3
 sentencepiece
 wandb
 einops
-xformers>=0.0.23.post1
+xformers==0.0.28.post1
 optimum==1.16.2
 hf_transfer
 colorama
@@ -28,12 +28,13 @@ scipy
 scikit-learn==1.4.2
 pynvml
 art
+fschat @ git+https://github.com/lm-sys/FastChat.git@27a05b04a35510afb1d767ae7e5990cbd278f8fe
 gradio==3.50.2
 tensorboard
 python-dotenv==1.0.1
 autoawq>=0.2.5
 triton>=2.3.0
-liger-kernel==0.4.1
+liger-kernel==0.3.0
 
 mamba-ssm==1.2.0.post1
 
@@ -42,7 +43,7 @@ s3fs>=2024.5.0
 gcsfs>=2024.5.0
 # adlfs
 
-trl==0.12.0
+trl==0.9.6
 zstandard==0.22.0
 fastcore
 

setup.py

@@ -31,18 +31,13 @@ def parse_requirements():
     try:
         xformers_version = [req for req in _install_requires if "xformers" in req][0]
         torchao_version = [req for req in _install_requires if "torchao" in req][0]
-        autoawq_version = [req for req in _install_requires if "autoawq" in req][0]
-
         if "Darwin" in platform.system():
             # don't install xformers on MacOS
             _install_requires.pop(_install_requires.index(xformers_version))
         else:
             # detect the version of torch already installed
             # and set it so dependencies don't clobber the torch version
-            try:
-                torch_version = version("torch")
-            except PackageNotFoundError:
-                torch_version = "2.5.1"
+            torch_version = version("torch")
             _install_requires.append(f"torch=={torch_version}")
 
             version_match = re.match(r"^(\d+)\.(\d+)(?:\.(\d+))?", torch_version)
@@ -55,20 +50,10 @@ def parse_requirements():
             else:
                 raise ValueError("Invalid version format")
 
-            if (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.pop(_install_requires.index(autoawq_version))
-            elif (major, minor) >= (2, 4):
+            if (major, minor) >= (2, 4):
                 if patch == 0:
                     _install_requires.pop(_install_requires.index(xformers_version))
                     _install_requires.append("xformers>=0.0.27")
-                else:
-                    _install_requires.pop(_install_requires.index(xformers_version))
-                    _install_requires.append("xformers==0.0.28.post1")
             elif (major, minor) >= (2, 3):
                 _install_requires.pop(_install_requires.index(torchao_version))
                 if patch == 0:
@@ -88,6 +73,7 @@ def parse_requirements():
 
     except PackageNotFoundError:
         pass
+
     return _install_requires, _dependency_links
 
 
@@ -96,7 +82,7 @@ install_requires, dependency_links = parse_requirements()
 
 setup(
     name="axolotl",
-    version="0.5.0",
+    version="0.4.1",
     description="LLM Trainer",
     long_description="Axolotl is a tool designed to streamline the fine-tuning of various AI models, offering support for multiple configurations and architectures.",
     package_dir={"": "src"},
@@ -105,7 +91,10 @@ setup(
     dependency_links=dependency_links,
     extras_require={
         "flash-attn": [
-            "flash-attn==2.7.0.post2",
+            "flash-attn==2.6.3",
+        ],
+        "fused-dense-lib": [
+            "fused-dense-lib  @ git+https://github.com/Dao-AILab/flash-attention@v2.6.2#subdirectory=csrc/fused_dense_lib",
         ],
         "deepspeed": [
             "deepspeed==0.14.4",
@@ -113,7 +102,6 @@ setup(
         ],
         "mamba-ssm": [
             "mamba-ssm==1.2.0.post1",
-            "causal_conv1d",
         ],
         "auto-gptq": [
             "auto-gptq==0.5.1",

src/axolotl/cli/__init__.py

@@ -30,7 +30,7 @@ from axolotl.common.cli import TrainerCliArgs, load_model_and_tokenizer
 from axolotl.integrations.base import PluginManager
 from axolotl.logging_config import configure_logging
 from axolotl.train import TrainDatasetMeta
-from axolotl.utils.chat_templates import get_chat_template
+from axolotl.utils.chat_templates import chat_templates
 from axolotl.utils.comet_ import setup_comet_env_vars
 from axolotl.utils.config import (
     normalize_cfg_datasets,
@@ -190,15 +190,18 @@ def do_inference(
 ):
     model, tokenizer = load_model_and_tokenizer(cfg=cfg, cli_args=cli_args)
     prompter = cli_args.prompter
+    default_tokens = {"unk_token": "<unk>", "bos_token": "<s>", "eos_token": "</s>"}
+
+    for token, symbol in default_tokens.items():
+        # If the token isn't already specified in the config, add it
+        if not (cfg.special_tokens and token in cfg.special_tokens):
+            tokenizer.add_special_tokens({token: symbol})
 
     prompter_module = None
-    chat_template_str = None
     if prompter:
         prompter_module = getattr(
             importlib.import_module("axolotl.prompters"), prompter
         )
-    elif cfg.chat_template:
-        chat_template_str = get_chat_template(cfg.chat_template)
 
     model = model.to(cfg.device, dtype=cfg.torch_dtype)
 
@@ -208,31 +211,13 @@ def do_inference(
         instruction = get_multi_line_input()
         if not instruction:
             return
-
         if prompter_module:
             prompt: str = next(
                 prompter_module().build_prompt(instruction=instruction.strip("\n"))
             )
         else:
             prompt = instruction.strip()
-
-        if chat_template_str:
-            batch = tokenizer.apply_chat_template(
-                [
-                    {
-                        "role": "user",
-                        "content": prompt,
-                    }
-                ],
-                return_tensors="pt",
-                add_special_tokens=True,
-                add_generation_prompt=True,
-                chat_template=chat_template_str,
-                tokenize=True,
-                return_dict=True,
-            )
-        else:
-            batch = tokenizer(prompt, return_tensors="pt", add_special_tokens=True)
+        batch = tokenizer(prompt, return_tensors="pt", add_special_tokens=True)
 
         print("=" * 40)
         model.eval()
@@ -272,6 +257,13 @@ def do_inference_gradio(
 
     model, tokenizer = load_model_and_tokenizer(cfg=cfg, cli_args=cli_args)
     prompter = cli_args.prompter
+    # default_tokens = {"unk_token": "<unk>", "bos_token": "<s>", "eos_token": "</s>"}
+    default_tokens: Dict[str, str] = {}
+
+    for token, symbol in default_tokens.items():
+        # If the token isn't already specified in the config, add it
+        if not (cfg.special_tokens and token in cfg.special_tokens):
+            tokenizer.add_special_tokens({token: symbol})
 
     prompter_module = None
     chat_template_str = None
@@ -280,7 +272,7 @@ def do_inference_gradio(
             importlib.import_module("axolotl.prompters"), prompter
         )
     elif cfg.chat_template:
-        chat_template_str = get_chat_template(cfg.chat_template, tokenizer=tokenizer)
+        chat_template_str = chat_templates(cfg.chat_template)
 
     model = model.to(cfg.device, dtype=cfg.torch_dtype)
 
@@ -470,12 +462,7 @@ def load_datasets(
         processor=processor,
     )
 
-    if (
-        cli_args.debug
-        or cfg.debug
-        or cli_args.debug_text_only
-        or int(cli_args.debug_num_examples) > 0
-    ):
+    if cli_args.debug or cfg.debug:
         LOG.info("check_dataset_labels...")
         check_dataset_labels(
             train_dataset.select(

src/axolotl/cli/preprocess.py

@@ -23,6 +23,10 @@ from axolotl.cli import (
 )
 from axolotl.common.cli import PreprocessCliArgs
 from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
+from axolotl.prompt_strategies.sharegpt import (
+    register_chatml_template,
+    register_llama3_template,
+)
 from axolotl.utils.trainer import disable_datasets_caching
 
 LOG = logging.getLogger("axolotl.cli.preprocess")
@@ -40,6 +44,23 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
         return_remaining_strings=True
     )
 
+    if parsed_cfg.chat_template == "chatml":
+        if parsed_cfg.default_system_message:
+            LOG.info(
+                f"ChatML set. Adding default system message: {parsed_cfg.default_system_message}"
+            )
+            register_chatml_template(parsed_cfg.default_system_message)
+        else:
+            register_chatml_template()
+    elif parsed_cfg.chat_template == "llama3":
+        if parsed_cfg.default_system_message:
+            LOG.info(
+                f"LLaMA-3 set. Adding default system message: {parsed_cfg.default_system_message}"
+            )
+            register_llama3_template(parsed_cfg.default_system_message)
+        else:
+            register_llama3_template()
+
     if not parsed_cfg.dataset_prepared_path:
         msg = (
             Fore.RED

src/axolotl/cli/train.py

@@ -19,6 +19,10 @@ from axolotl.cli import (
 )
 from axolotl.common.cli import TrainerCliArgs
 from axolotl.integrations.base import PluginManager
+from axolotl.prompt_strategies.sharegpt import (
+    register_chatml_template,
+    register_llama3_template,
+)
 from axolotl.train import train
 
 LOG = logging.getLogger("axolotl.cli.train")
@@ -38,6 +42,21 @@ def do_train(cfg, cli_args) -> None:
     print_axolotl_text_art()
     check_accelerate_default_config()
     check_user_token()
+    if cfg.chat_template == "chatml" and cfg.default_system_message:
+        LOG.info(
+            f"ChatML set. Adding default system message: {cfg.default_system_message}"
+        )
+        register_chatml_template(cfg.default_system_message)
+    else:
+        register_chatml_template()
+
+    if cfg.chat_template == "llama3" and cfg.default_system_message:
+        LOG.info(
+            f"LLaMA-3 set. Adding default system message: {cfg.default_system_message}"
+        )
+        register_llama3_template(cfg.default_system_message)
+    else:
+        register_llama3_template()
 
     if cfg.rl:  # and cfg.rl != "orpo":
         dataset_meta = load_rl_datasets(cfg=cfg, cli_args=cli_args)

src/axolotl/common/cli.py

@@ -23,7 +23,7 @@ class TrainerCliArgs:
 
     debug: bool = field(default=False)
     debug_text_only: bool = field(default=False)
-    debug_num_examples: int = field(default=0)
+    debug_num_examples: int = field(default=5)
     inference: bool = field(default=False)
     merge_lora: bool = field(default=False)
     prompter: Optional[str] = field(default=None)

src/axolotl/core/trainer_builder.py

@@ -7,7 +7,6 @@ import abc
 import gc
 import importlib
 import importlib.util
-import inspect
 import logging
 import math
 import os
@@ -28,6 +27,7 @@ from torch.optim.lr_scheduler import OneCycleLR
 from torch.utils.data import BatchSampler, DataLoader, RandomSampler, SequentialSampler
 from transformers import (
     EarlyStoppingCallback,
+    PreTrainedModel,
     Trainer,
     TrainerCallback,
     TrainingArguments,
@@ -48,7 +48,6 @@ from trl import (
 )
 from trl.trainer.utils import RewardDataCollatorWithPadding, pad_to_length
 
-from axolotl.integrations.base import PluginManager
 from axolotl.monkeypatch.multipack import SUPPORTED_MULTIPACK_MODEL_TYPES
 from axolotl.monkeypatch.relora import ReLoRACallback, ReLoRAScheduler
 from axolotl.utils import is_comet_available, is_mlflow_available
@@ -64,7 +63,7 @@ from axolotl.utils.callbacks import (
     log_prediction_callback_factory,
 )
 from axolotl.utils.callbacks.lisa import lisa_callback_factory
-from axolotl.utils.chat_templates import get_chat_template
+from axolotl.utils.chat_templates import chat_templates
 from axolotl.utils.collators import (
     BatchSamplerDataCollatorForSeq2Seq,
     DataCollatorForSeq2Seq,
@@ -436,13 +435,7 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
         if (
             self.args.loraplus_lr_ratio is None
             and self.args.alternate_optimizer
-            not in [
-                "optimi_adamw",
-                "ao_adamw_8bit",
-                "ao_adamw_4bit",
-                "ao_adamw_fp8",
-                "adopt_adamw",
-            ]
+            not in ["optimi_adamw", "ao_adamw_8bit", "ao_adamw_4bit", "ao_adamw_fp8"]
         ):
             return super().create_optimizer()
 
@@ -511,14 +504,6 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
                 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
-
-                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
-                    ADOPT(
-                        optimizer_grouped_parameters, decoupled=True, **optimizer_kwargs
-                    )
-                )
 
         if is_sagemaker_mp_enabled():
             self.optimizer = smp.DistributedOptimizer(  # pylint: disable=attribute-defined-outside-init
@@ -681,9 +666,7 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
         return DataLoader(bench_dataset, **dataloader_params)
         # return self.accelerator.prepare(DataLoader(bench_dataset, **dataloader_params))
 
-    def compute_loss(
-        self, model, inputs, return_outputs=False, num_items_in_batch=None
-    ):
+    def compute_loss(self, model, inputs, return_outputs=False):
         # use one's weighted cross entropy loss calc
         # if self.args.sample_packing:
         #     labels = inputs.pop("labels")
@@ -691,18 +674,8 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
         #     loss = trainer_weighted_loss(outputs, labels, shift_labels=True)
         #     return (loss, outputs) if return_outputs else loss
         if self.args.orpo_alpha:
-            return self.orpo_compute_loss(
-                model,
-                inputs,
-                return_outputs=return_outputs,
-                num_items_in_batch=num_items_in_batch,
-            )
-        return super().compute_loss(
-            model,
-            inputs,
-            return_outputs=return_outputs,
-            num_items_in_batch=num_items_in_batch,
-        )
+            return self.orpo_compute_loss(model, inputs, return_outputs=return_outputs)
+        return super().compute_loss(model, inputs, return_outputs=return_outputs)
 
     @staticmethod
     def orpo_concatenate_inputs(inputs, label_pad_token=-100, pad_token=0, device=None):
@@ -798,13 +771,7 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
         ).squeeze(2)
         return torch.mul(per_token_logps, mask).sum(dim=1) / mask.sum(dim=1)
 
-    def orpo_compute_loss(
-        self,
-        model,
-        inputs,
-        return_outputs=False,
-        num_items_in_batch=None,  # pylint: disable=unused-argument
-    ):
+    def orpo_compute_loss(self, model, inputs, return_outputs=False):
         concat_inputs = AxolotlTrainer.orpo_concatenate_inputs(
             inputs,
             label_pad_token=-100,
@@ -910,13 +877,13 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
         for key, value in metrics.items():
             self._stored_metrics[train_eval][key].append(value)
 
-    def _save_checkpoint(self, model, trial, **kwargs):
+    def _save_checkpoint(self, model, trial, metrics=None):
         # make sure the checkpoint dir exists, since trainer is flakey
         checkpoint_folder = f"{PREFIX_CHECKPOINT_DIR}-{self.state.global_step}"
         run_dir = self._get_output_dir(trial=trial)
         output_dir = os.path.join(run_dir, checkpoint_folder)
         os.makedirs(output_dir, exist_ok=True)
-        return super()._save_checkpoint(model, trial, **kwargs)
+        return super()._save_checkpoint(model, trial, metrics=metrics)
 
 
 class AxolotlMambaTrainer(AxolotlTrainer):
@@ -931,7 +898,6 @@ class AxolotlMambaTrainer(AxolotlTrainer):
         model,
         inputs,
         return_outputs=False,  # pylint: disable=unused-argument
-        num_items_in_batch=None,  # pylint: disable=unused-argument
     ):
         input_ids = inputs.pop("input_ids")
         lm_logits = model(input_ids).logits
@@ -1039,32 +1005,18 @@ class AxolotlDPOTrainer(SchedulerMixin, DPOTrainer):
         return super().push_to_hub(*args, **kwargs)
 
     def tokenize_row(
-        self,
-        features,
-        processing_class,
-        max_prompt_length,
-        max_completion_length,
-        add_special_tokens,
+        self, feature, model: Optional[Union[PreTrainedModel, torch.nn.Module]] = None
     ) -> Dict:
-        res = super().tokenize_row(
-            features,
-            processing_class,
-            max_prompt_length,
-            max_completion_length,
-            add_special_tokens,
-        )
-        if processing_class.bos_token_id is None and res["prompt_input_ids"][0] is None:
+        res = super().tokenize_row(feature, model=model)
+        if self.tokenizer.bos_token_id is None and res["prompt_input_ids"][0] is None:
             for key in res.keys():
                 res[key] = res[key][1:]
         return res
 
     def training_step(
-        self,
-        model: nn.Module,
-        inputs: Dict[str, Union[torch.Tensor, Any]],
-        num_items_in_batch=None,
+        self, model: nn.Module, inputs: Dict[str, Union[torch.Tensor, Any]]
     ) -> torch.Tensor:
-        loss: torch.Tensor = super().training_step(model, inputs, num_items_in_batch)
+        loss: torch.Tensor = super().training_step(model, inputs)
         gc.collect()
         torch.cuda.empty_cache()
         return loss
@@ -1162,28 +1114,17 @@ class TrainerBuilderBase(abc.ABC):
 
     def get_callbacks(self) -> List[TrainerCallback]:
         callbacks = []
-
-        plugin_manager = PluginManager.get_instance()
-        callbacks.extend(
-            plugin_manager.add_callbacks_pre_trainer(cfg=self.cfg, model=self.model)
-        )
-
         if self.cfg.use_wandb:
             callbacks.append(
                 SaveAxolotlConfigtoWandBCallback(self.cfg.axolotl_config_path)
             )
         if self.cfg.use_mlflow and is_mlflow_available():
-            from transformers.integrations.integration_utils import MLflowCallback
-
             from axolotl.utils.callbacks.mlflow_ import (
                 SaveAxolotlConfigtoMlflowCallback,
             )
 
-            callbacks.extend(
-                [
-                    SaveAxolotlConfigtoMlflowCallback(self.cfg.axolotl_config_path),
-                    MLflowCallback,
-                ]
+            callbacks.append(
+                SaveAxolotlConfigtoMlflowCallback(self.cfg.axolotl_config_path)
             )
         if self.cfg.use_comet and is_comet_available():
             from axolotl.utils.callbacks.comet_ import SaveAxolotlConfigtoCometCallback
@@ -1194,17 +1135,11 @@ class TrainerBuilderBase(abc.ABC):
 
         return callbacks
 
+    @abstractmethod
     def get_post_trainer_create_callbacks(self, trainer):
         """
         Callbacks added after the trainer is created, usually b/c these need access to the trainer
         """
-        callbacks = []
-
-        plugin_manager = PluginManager.get_instance()
-        callbacks.extend(
-            plugin_manager.add_callbacks_post_trainer(cfg=self.cfg, trainer=trainer)
-        )
-        return callbacks
 
     def hook_pre_create_training_args(self, training_arguments_kwargs):
         # TODO
@@ -1250,7 +1185,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
-        callbacks = super().get_post_trainer_create_callbacks(trainer=trainer)
+        callbacks = []
         if self.cfg.use_wandb and self.cfg.eval_table_size > 0:
             LogPredictionCallback = log_prediction_callback_factory(
                 trainer, self.tokenizer, "wandb"
@@ -1287,18 +1222,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
 
         if self.cfg.lisa_step_interval and self.cfg.lisa_n_layers:
             callbacks.append(lisa_callback_factory(trainer))
-
-        if self.cfg.plugins:
-            plugin_manager = PluginManager.get_instance()
-            callbacks.extend(
-                [
-                    cb
-                    for cb in plugin_manager.add_callbacks_post_trainer(
-                        self.cfg, trainer
-                    )
-                    if cb
-                ]
-            )
         return callbacks
 
     def _get_trainer_cls(self):
@@ -1633,9 +1556,8 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs["model_type"] = self.cfg.model_config_type
         training_arguments_kwargs["pretraining"] = bool(self.cfg.pretraining_dataset)
         if self.cfg.chat_template:
-            training_arguments_kwargs["chat_template"] = get_chat_template(
-                self.cfg.chat_template,
-                tokenizer=self.tokenizer,
+            training_arguments_kwargs["chat_template"] = chat_templates(
+                self.cfg.chat_template
             )
 
         if self.cfg.rl == "orpo":
@@ -1651,13 +1573,11 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         if self.cfg.reward_model:
             trainer_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"
@@ -1742,17 +1662,12 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 return_tensors="pt",
                 **data_collator_kwargs,
             )
-        sig = inspect.signature(trainer_cls)
-        if "processing_class" in sig.parameters.keys():
-            trainer_kwargs["processing_class"] = self.tokenizer
-        else:
-            trainer_kwargs["tokenizer"] = self.tokenizer
-
         trainer = trainer_cls(
             model=self.model,
             train_dataset=self.train_dataset,
             eval_dataset=self.eval_dataset,
             args=training_args,
+            tokenizer=self.tokenizer,
             data_collator=self.build_collator(training_args, **data_collator_kwargs),
             callbacks=self.get_callbacks(),
             **trainer_kwargs,
@@ -1793,8 +1708,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         ]
         if self.cfg.reward_model:
             collator = RewardDataCollatorWithPadding
-            if "max_length" in kwargs:
-                kwargs.pop("max_length")
         elif use_batch_sampler_collator:
             if self.cfg.model_config_type in SUPPORTED_MULTIPACK_MODEL_TYPES:
                 collator = V2BatchSamplerDataCollatorForSeq2Seq
@@ -1832,7 +1745,7 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
-        callbacks = super().get_post_trainer_create_callbacks(trainer=trainer)
+        callbacks = []
         return callbacks
 
     def build_training_arguments(self, total_num_steps):
@@ -1918,18 +1831,17 @@ 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.rl_beta:
             training_args_kwargs["beta"] = 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
 
+        training_args_kwargs["dataset_num_proc"] = self.cfg.dataset_processes
+        training_args_cls = AxolotlDPOConfig
         if self.cfg.rpo_alpha is not None:
             training_args_kwargs["rpo_alpha"] = self.cfg.rpo_alpha
 
-        training_args_cls = None
         if self.cfg.rl == "simpo":
             training_args_cls = AxolotlCPOConfig
             training_args_kwargs["loss_type"] = "simpo"
@@ -1938,13 +1850,13 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             if self.cfg.cpo_alpha is not None:
                 training_args_kwargs["cpo_alpha"] = self.cfg.cpo_alpha
 
-        elif self.cfg.rl == "orpo":
+        if self.cfg.rl == "orpo":
             training_args_cls = AxolotlORPOConfig
             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 == "kto":
+        if self.cfg.rl == "kto":
             training_args_cls = AxolotlKTOConfig
 
             training_args_kwargs["desirable_weight"] = (
@@ -1959,17 +1871,6 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             if self.cfg.max_prompt_len:
                 training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
 
-        else:
-            training_args_cls = AxolotlDPOConfig
-            if self.cfg.rl == "ipo":
-                training_args_kwargs["loss_type"] = "ipo"
-            training_args_kwargs["max_length"] = self.cfg.sequence_len
-            training_args_kwargs["max_completion_length"] = None
-            training_args_kwargs["max_prompt_length"] = self.cfg.sequence_len
-            training_args_kwargs["generate_during_eval"] = self.cfg.use_wandb
-            if self.cfg.dpo_use_weighting is not None:
-                training_args_kwargs["use_weighting"] = self.cfg.dpo_use_weighting
-
         training_args = training_args_cls(  # pylint: disable=unexpected-keyword-arg
             output_dir=self.cfg.output_dir,
             per_device_train_batch_size=self.cfg.micro_batch_size,
@@ -1990,6 +1891,7 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         training_args = self.build_training_arguments(total_num_steps)
         dpo_trainer_kwargs = {}
         if self.cfg.rl == "ipo":
+            dpo_trainer_kwargs["loss_type"] = "ipo"
             if self.cfg.dpo_label_smoothing:
                 dpo_trainer_kwargs["label_smoothing"] = self.cfg.dpo_label_smoothing
         if self.eval_dataset:
@@ -2003,6 +1905,12 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         if self.cfg.rl in ["dpo", "ipo"]:
             trainer_cls = AxolotlDPOTrainer
             trainer_cls_args = [self.model, self.model_ref]
+
+            # these aren't used for the ORPO trainer
+            dpo_trainer_kwargs["max_length"] = self.cfg.sequence_len
+            dpo_trainer_kwargs["max_target_length"] = None
+            dpo_trainer_kwargs["max_prompt_length"] = self.cfg.sequence_len
+            dpo_trainer_kwargs["generate_during_eval"] = True
         elif self.cfg.rl == "orpo":
             trainer_cls = AxolotlORPOTrainer
             trainer_cls_args = [self.model]
@@ -2014,17 +1922,11 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             trainer_cls_args = [self.model]
         else:
             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:
-            dpo_trainer_kwargs["tokenizer"] = self.tokenizer
-
         dpo_trainer = trainer_cls(
             *trainer_cls_args,
             args=training_args,
             train_dataset=self.train_dataset,
+            tokenizer=self.tokenizer,
             callbacks=self.get_callbacks(),
             **dpo_trainer_kwargs,
         )
@@ -2046,11 +1948,11 @@ class HFPPOTrainerBuilder(TrainerBuilderBase):
     """
 
     def get_callbacks(self):
-        callbacks = super().get_callbacks()
+        callbacks = []
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
-        callbacks = super().get_post_trainer_create_callbacks(trainer=trainer)
+        callbacks = []
         return callbacks
 
     def build(self, total_num_steps):

src/axolotl/integrations/base.py

@@ -18,10 +18,9 @@ Plugins can be used to integrate third-party models, modify the training process
 
 To create a new plugin, you need to inherit from the BasePlugin class and implement the required methods.
 """
-import collections
 import importlib
 import logging
-from typing import OrderedDict
+from typing import List
 
 
 class BasePlugin:
@@ -48,7 +47,7 @@ class BasePlugin:
         Initializes the BasePlugin.
         """
 
-    def register(self, cfg):  # pylint: disable=unused-argument
+    def register(self, cfg):
         """
         Registers the plugin with the given configuration.
 
@@ -64,7 +63,7 @@ class BasePlugin:
         Returns a pydantic model for the plugin's input arguments.
         """
 
-    def pre_model_load(self, cfg):  # pylint: disable=unused-argument
+    def pre_model_load(self, cfg):
         """
         Performs actions before the model is loaded.
 
@@ -75,7 +74,7 @@ class BasePlugin:
         None
         """
 
-    def post_model_load(self, cfg, model):  # pylint: disable=unused-argument
+    def post_model_load(self, cfg, model):
         """
         Performs actions after the model is loaded.
 
@@ -87,7 +86,7 @@ class BasePlugin:
         None
         """
 
-    def pre_lora_load(self, cfg, model):  # pylint: disable=unused-argument
+    def pre_lora_load(self, cfg, model):
         """
         Performs actions before LoRA weights are loaded.
 
@@ -99,7 +98,7 @@ class BasePlugin:
         None
         """
 
-    def post_lora_load(self, cfg, model):  # pylint: disable=unused-argument
+    def post_lora_load(self, cfg, model):
         """
         Performs actions after LoRA weights are loaded.
 
@@ -111,7 +110,7 @@ class BasePlugin:
         None
         """
 
-    def create_optimizer(self, cfg, trainer):  # pylint: disable=unused-argument
+    def create_optimizer(self, cfg, trainer):
         """
         Creates and returns an optimizer for training.
 
@@ -123,9 +122,7 @@ class BasePlugin:
         object: The created optimizer.
         """
 
-    def create_lr_scheduler(
-        self, cfg, trainer, optimizer
-    ):  # pylint: disable=unused-argument
+    def create_lr_scheduler(self, cfg, trainer, optimizer):
         """
         Creates and returns a learning rate scheduler.
 
@@ -138,9 +135,9 @@ class BasePlugin:
         object: The created learning rate scheduler.
         """
 
-    def add_callbacks_pre_trainer(self, cfg, model):  # pylint: disable=unused-argument
+    def add_callbacks_pre_trainer(self, cfg, model):
         """
-        setup callbacks before creating the trainer.
+        Adds callbacks to the trainer before training.
 
         Parameters:
         cfg (dict): The configuration for the plugin.
@@ -149,25 +146,20 @@ class BasePlugin:
         Returns:
         List[callable]: A list of callback functions to be added to the TrainingArgs
         """
-        return []
 
-    def add_callbacks_post_trainer(
-        self, cfg, trainer
-    ):  # pylint: disable=unused-argument
+    def add_callbacks_post_trainer(self, cfg, trainer):
         """
-        Adds callbacks to the trainer after creating the trainer.
-        This is useful for callbacks that require access to the model or trainer.
+        Adds callbacks to the trainer after training.
 
         Parameters:
         cfg (dict): The configuration for the plugin.
         trainer (object): The trainer object for training.
 
         Returns:
-        List[callable]: A list of callback functions to be added
+        List[callable]: A list of callback functions to be added to the TrainingArgs
         """
-        return []
 
-    def post_train(self, cfg, model):  # pylint: disable=unused-argument
+    def post_train(self, cfg, model):
         """
         Performs actions after training is complete.
 
@@ -179,7 +171,7 @@ class BasePlugin:
         None
         """
 
-    def post_train_unload(self, cfg):  # pylint: disable=unused-argument
+    def post_train_unload(self, cfg):
         """
         Performs actions after training is complete and the model is unloaded.
 
@@ -235,7 +227,7 @@ class PluginManager:
     pre_model_load(cfg): Calls the pre_model_load method of all registered plugins.
     """
 
-    plugins: OrderedDict[str, BasePlugin] = collections.OrderedDict()
+    plugins: List[BasePlugin] = []
 
     _instance = None
 
@@ -245,7 +237,7 @@ class PluginManager:
         """
         if cls._instance is None:
             cls._instance = super(PluginManager, cls).__new__(cls)
-            cls._instance.plugins = collections.OrderedDict()
+            cls._instance.plugins: List[BasePlugin] = []
         return cls._instance
 
     @staticmethod
@@ -273,7 +265,7 @@ class PluginManager:
         """
         try:
             plugin = load_plugin(plugin_name)
-            self.plugins[plugin_name] = plugin
+            self.plugins.append(plugin)
         except ImportError:
             logging.error(f"Failed to load plugin: {plugin_name}")
 
@@ -285,7 +277,7 @@ class PluginManager:
         list[str]: A list of Pydantic classes for all registered plugins' input arguments.'
         """
         input_args = []
-        for plugin in self.plugins.values():
+        for plugin in self.plugins:
             input_args_from_plugin = plugin.get_input_args()
             if input_args_from_plugin is not None:
                 input_args.append(input_args_from_plugin)
@@ -301,7 +293,7 @@ class PluginManager:
         Returns:
         None
         """
-        for plugin in self.plugins.values():
+        for plugin in self.plugins:
             plugin.pre_model_load(cfg)
 
     def post_model_load(self, cfg, model):
@@ -315,7 +307,7 @@ class PluginManager:
         Returns:
         None
         """
-        for plugin in self.plugins.values():
+        for plugin in self.plugins:
             plugin.post_model_load(cfg, model)
 
     def pre_lora_load(self, cfg, model):
@@ -329,7 +321,7 @@ class PluginManager:
         Returns:
         None
         """
-        for plugin in self.plugins.values():
+        for plugin in self.plugins:
             plugin.pre_lora_load(cfg, model)
 
     def post_lora_load(self, cfg, model):
@@ -343,7 +335,7 @@ class PluginManager:
         Returns:
         None
         """
-        for plugin in self.plugins.values():
+        for plugin in self.plugins:
             plugin.post_lora_load(cfg, model)
 
     def create_optimizer(self, cfg, trainer):
@@ -357,7 +349,7 @@ class PluginManager:
         Returns:
         object: The created optimizer, or None if none was found.
         """
-        for plugin in self.plugins.values():
+        for plugin in self.plugins:
             optimizer = plugin.create_optimizer(cfg, trainer)
             if optimizer is not None:
                 return optimizer
@@ -375,7 +367,7 @@ class PluginManager:
         Returns:
         object: The created learning rate scheduler, or None if none was found.
         """
-        for plugin in self.plugins.values():
+        for plugin in self.plugins:
             scheduler = plugin.create_lr_scheduler(cfg, trainer, optimizer)
             if scheduler is not None:
                 return scheduler
@@ -393,10 +385,8 @@ class PluginManager:
         List[callable]: A list of callback functions to be added to the TrainingArgs.
         """
         callbacks = []
-        for plugin in self.plugins.values():
-            plugin_callbacks = plugin.add_callbacks_pre_trainer(cfg, model)
-            if plugin_callbacks:  # if the plugin returned a list of callbacks
-                callbacks.extend(plugin_callbacks)
+        for plugin in self.plugins:
+            callbacks.extend(plugin.add_callbacks_pre_trainer(cfg, model))
         return callbacks
 
     def add_callbacks_post_trainer(self, cfg, trainer):
@@ -411,10 +401,8 @@ class PluginManager:
         List[callable]: A list of callback functions to be added to the TrainingArgs.
         """
         callbacks = []
-        for plugin in self.plugins.values():
-            plugin_callbacks = plugin.add_callbacks_post_trainer(cfg, trainer)
-            if plugin_callbacks:
-                callbacks.extend(plugin_callbacks)
+        for plugin in self.plugins:
+            callbacks.extend(plugin.add_callbacks_post_trainer(cfg, trainer))
         return callbacks
 
     def post_train_unload(self, cfg):
@@ -428,5 +416,5 @@ class PluginManager:
         Returns:
         None
         """
-        for plugin in self.plugins.values():
+        for plugin in self.plugins:
             plugin.post_train_unload(cfg)

src/axolotl/integrations/grokfast/LICENSE

@@ -1,21 +0,0 @@
-MIT License
-
-Copyright (c) 2024 Jaerin Lee, Bong Gyun Kang, Kihoon Kim, Kyoung Mu Lee
-
-Permission is hereby granted, free of charge, to any person obtaining a copy
-of this software and associated documentation files (the "Software"), to deal
-in the Software without restriction, including without limitation the rights
-to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-copies of the Software, and to permit persons to whom the Software is
-furnished to do so, subject to the following conditions:
-
-The above copyright notice and this permission notice shall be included in all
-copies or substantial portions of the Software.
-
-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 NONINFRINGEMENT. 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.

src/axolotl/integrations/grokfast/README.md

@@ -1,13 +0,0 @@
-# Grokfast Optimizer
-
-See https://github.com/ironjr/grokfast
-
-### Usage
-
-```yaml
-plugins:
-  - axolotl.integrations.grokfast.GrokfastPlugin
-
-grokfast_alpha: 2.0
-grokfast_lamb: 0.98
-```

src/axolotl/integrations/grokfast/__init__.py

@@ -1,50 +0,0 @@
-"""
-Grokfast plugin for Axolotl
-"""
-import logging
-
-from transformers.trainer_callback import TrainerCallback
-
-from ..base import BasePlugin
-from .args import GrokfastArgs  # pylint: disable=unused-import. # noqa: F401
-from .optimizer import gradfilter_ema
-
-LOG = logging.getLogger("axolotl.integrations.grokfast")
-
-
-class GrokfastCallbackHandler(TrainerCallback):
-    """
-    Transformer trainer callbacks for Grokfast
-    """
-
-    def __init__(self, *args_, alpha=0.98, lamb=2.0, **kwargs):
-        super().__init__(*args_, **kwargs)
-        self.grads = None
-        self.alpha = alpha
-        self.lamb = lamb
-
-    def on_train_begin(self, *args_, **kwargs):  # pylint: disable=unused-argument
-        self.grads = None
-
-    def on_pre_optimizer_step(
-        self, args_, state, control, **kwargs
-    ):  # pylint: disable=unused-argument
-        model = kwargs.pop("model")
-        self.grads = gradfilter_ema(model, self.grads, alpha=self.alpha, lamb=self.lamb)
-        return control
-
-
-class GrokfastPlugin(BasePlugin):
-    """
-    Plugin for Grokfast optimizer integraton with Axolotl.
-    """
-
-    def get_input_args(self):
-        return "axolotl.integrations.grokfast.GrokfastArgs"
-
-    def add_callbacks_post_trainer(self, cfg, trainer):
-        LOG.info("Adding Grokfast callback to the trainer")
-        callback = GrokfastCallbackHandler(
-            alpha=cfg.grokfast_alpha, lamb=cfg.grokfast_lamb
-        )
-        return [callback]

src/axolotl/integrations/grokfast/args.py

@@ -1,15 +0,0 @@
-"""
-config args for grokfast plugin
-"""
-from typing import Optional
-
-from pydantic import BaseModel
-
-
-class GrokfastArgs(BaseModel):
-    """
-    Input args for Grokfast optimizer.
-    """
-
-    grokfast_alpha: Optional[float] = 0.98
-    grokfast_lamb: Optional[float] = 2.0

src/axolotl/integrations/grokfast/optimizer.py

@@ -1,63 +0,0 @@
-# Copyright: MIT License (c) 2024 Jaerin Lee, Bong Gyun Kang, Kihoon Kim, Kyoung Mu Lee
-# Reference: https://github.com/ironjr/grokfast
-
-# pylint: skip-file
-from collections import deque
-from typing import Dict, Literal, Optional
-
-import torch
-import torch.nn as nn
-
-
-def gradfilter_ma(
-    m: nn.Module,
-    grads: Optional[Dict[str, deque]] = None,
-    window_size: int = 100,
-    lamb: float = 5.0,
-    filter_type: Literal["mean", "sum"] = "mean",
-    warmup: bool = True,
-    trigger: bool = False,  # For ablation study.
-) -> Dict[str, deque]:
-    if grads is None:
-        grads = {
-            n: deque(maxlen=window_size)
-            for n, p in m.named_parameters()
-            if p.requires_grad and p.grad is not None
-        }
-
-    for n, p in m.named_parameters():
-        if p.requires_grad and p.grad is not None:
-            grads[n].append(p.grad.data.detach())  # .cpu())
-
-            # Modify the gradients.
-            if not warmup or len(grads[n]) == window_size and not trigger:
-                if filter_type == "mean":
-                    avg = sum(grads[n]) / len(grads[n])
-                elif filter_type == "sum":
-                    avg = sum(grads[n])
-                else:
-                    raise ValueError(f"Unrecognized filter_type {filter_type}")
-                p.grad.data = p.grad.data + avg * lamb
-
-    return grads
-
-
-def gradfilter_ema(
-    m: nn.Module,
-    grads: Optional[Dict[str, torch.Tensor]] = None,
-    alpha: float = 0.98,
-    lamb: float = 2.0,
-) -> Dict[str, torch.Tensor]:
-    if grads is None:
-        grads = {
-            n: p.grad.data.detach()
-            for n, p in m.named_parameters()
-            if p.requires_grad and p.grad is not None
-        }
-
-    for n, p in m.named_parameters():
-        if p.requires_grad and p.grad is not None:
-            grads[n] = grads[n] * alpha + p.grad.data.detach() * (1 - alpha)
-            p.grad.data = p.grad.data + grads[n] * lamb
-
-    return grads

src/axolotl/integrations/liger/__init__.py

@@ -18,24 +18,20 @@ Module for the Plugin for LIGER integraton with Axolotl.
 Liger Kernel is the collection of Triton-native kernels for LLM Training.
 It is designed to be performant, correct, and light-weight.
 """
-import inspect
 import logging
 import sys
+from functools import partial
 
 from liger_kernel.transformers.cross_entropy import LigerCrossEntropyLoss
-from liger_kernel.transformers.functional import liger_cross_entropy
-from liger_kernel.transformers.monkey_patch import MODEL_TYPE_TO_APPLY_LIGER_FN
+from liger_kernel.transformers.geglu import LigerGEGLUMLP
 from liger_kernel.transformers.rms_norm import LigerRMSNorm
 from liger_kernel.transformers.rope import liger_rotary_pos_emb
 from liger_kernel.transformers.swiglu import LigerSwiGLUMLP
 
 from axolotl.integrations.base import BasePlugin
 
-from ...utils.distributed import zero_only
 from .args import LigerArgs  # pylint: disable=unused-import. # noqa: F401
 
-LOG = logging.getLogger("axolotl.integrations.liger")
-
 
 class LigerPlugin(BasePlugin):
     """
@@ -46,31 +42,59 @@ class LigerPlugin(BasePlugin):
         return "axolotl.integrations.liger.LigerArgs"
 
     def pre_model_load(self, cfg):
-        if cfg.model_config_type in MODEL_TYPE_TO_APPLY_LIGER_FN:
-            apply_liger_fn = MODEL_TYPE_TO_APPLY_LIGER_FN[cfg.model_config_type]
-            liger_fn_sig = inspect.signature(apply_liger_fn)
-            kwargs = {}
-            if "rope" in liger_fn_sig.parameters:
-                kwargs["rope"] = cfg.liger_rope
-            if "cross_entropy" in liger_fn_sig.parameters:
-                kwargs["cross_entropy"] = cfg.liger_cross_entropy
-            if "fused_linear_cross_entropy" in liger_fn_sig.parameters:
-                kwargs[
-                    "fused_linear_cross_entropy"
-                ] = cfg.liger_fused_linear_cross_entropy
-            if "rms_norm" in liger_fn_sig.parameters:
-                kwargs["rms_norm"] = cfg.liger_rms_norm
-            if "layer_norm" in liger_fn_sig.parameters:
-                kwargs["layer_norm"] = cfg.liger_layer_norm
-            if "geglu" in liger_fn_sig.parameters:
-                kwargs["geglu"] = cfg.liger_glu_activation
-            elif "swiglu" in liger_fn_sig.parameters:
-                kwargs["swiglu"] = cfg.liger_glu_activation
-            with zero_only():
-                LOG.info(
-                    f"Applying LIGER to {cfg.model_config_type} with kwargs: {kwargs}"
+        if cfg.model_config_type == "llama":
+            from liger_kernel.transformers.model.llama import (
+                lce_forward as llama_lce_forward,
+            )
+            from transformers.models.llama import modeling_llama
+
+            if cfg.liger_rope:
+                modeling_llama.apply_rotary_pos_emb = liger_rotary_pos_emb
+            if cfg.liger_rms_norm:
+                modeling_llama.LlamaRMSNorm = LigerRMSNorm
+            if cfg.liger_swiglu:
+                modeling_llama.LlamaMLP = LigerSwiGLUMLP
+            if cfg.liger_cross_entropy:
+                modeling_llama.CrossEntropyLoss = LigerCrossEntropyLoss
+            elif cfg.liger_fused_linear_cross_entropy:
+                modeling_llama.LlamaForCausalLM.forward = llama_lce_forward
+
+        elif cfg.model_config_type == "mistral":
+            from liger_kernel.transformers.model.mistral import (
+                lce_forward as mistral_lce_forward,
+            )
+            from transformers.models.mistral import modeling_mistral
+
+            if cfg.liger_rope:
+                modeling_mistral.apply_rotary_pos_emb = liger_rotary_pos_emb
+            if cfg.liger_rms_norm:
+                modeling_mistral.MistralRMSNorm = LigerRMSNorm
+            if cfg.liger_swiglu:
+                modeling_mistral.MistralMLP = LigerSwiGLUMLP
+            if cfg.liger_cross_entropy:
+                modeling_mistral.CrossEntropyLoss = LigerCrossEntropyLoss
+            if cfg.liger_fused_linear_cross_entropy:
+                modeling_mistral.MistralForCausalLM.forward = mistral_lce_forward
+
+        elif cfg.model_config_type == "gemma":
+            from liger_kernel.transformers.model.gemma import (
+                lce_forward as gemma_lce_forward,
+            )
+            from transformers.models.gemma import modeling_gemma
+
+            if cfg.liger_rope:
+                modeling_gemma.apply_rotary_pos_emb = liger_rotary_pos_emb
+            if cfg.liger_rms_norm:
+                modeling_gemma.GemmaRMSNorm = partial(
+                    LigerRMSNorm, offset=1.0, init_fn="zeros", casting_mode="gemma"
                 )
-            apply_liger_fn(**kwargs)
+            if cfg.liger_swiglu:
+                modeling_gemma.GemmaMLP = LigerGEGLUMLP
+            if cfg.liger_cross_entropy:
+                modeling_gemma.CrossEntropyLoss = LigerCrossEntropyLoss
+            if cfg.liger_fused_linear_cross_entropy:
+                modeling_gemma.GemmaForCausalLM.forward = gemma_lce_forward
+
         elif cfg.model_config_type == "jamba":
             from transformers.models.jamba import modeling_jamba
 
@@ -80,14 +104,30 @@ class LigerPlugin(BasePlugin):
                 modeling_jamba.apply_rotary_pos_emb = liger_rotary_pos_emb
             if cfg.liger_rms_norm:
                 modeling_jamba.JambaRMSNorm = LigerRMSNorm
-            if cfg.liger_glu_activation:
+            if cfg.liger_swiglu:
                 modeling_jamba.JambaMLP = LigerSwiGLUMLP
             if cfg.liger_cross_entropy:
-                from transformers.loss.loss_utils import nn
-
-                nn.functional.cross_entropy = liger_cross_entropy
+                modeling_jamba.CrossEntropyLoss = LigerCrossEntropyLoss
             if cfg.liger_fused_linear_cross_entropy:
                 modeling_jamba.JambaForCausalLM.forward = jamba_lce_forward
+
+        elif cfg.model_config_type == "qwen2":
+            from liger_kernel.transformers.model.qwen2 import (
+                lce_forward as qwen2_lce_forward,
+            )
+            from transformers.models.qwen2 import modeling_qwen2
+
+            if cfg.liger_rope:
+                modeling_qwen2.apply_rotary_pos_emb = liger_rotary_pos_emb
+            if cfg.liger_rms_norm:
+                modeling_qwen2.Qwen2RMSNorm = LigerRMSNorm
+            if cfg.liger_swiglu:
+                modeling_qwen2.Qwen2MLP = LigerSwiGLUMLP
+            if cfg.liger_cross_entropy:
+                modeling_qwen2.CrossEntropyLoss = LigerCrossEntropyLoss
+            if cfg.liger_fused_linear_cross_entropy:
+                modeling_qwen2.Qwen2ForCausalLM.forward = qwen2_lce_forward
+
         elif cfg.model_config_type == "deepseek_v2":
             from accelerate import init_empty_weights
             from transformers import AutoModelForCausalLM
@@ -106,11 +146,44 @@ class LigerPlugin(BasePlugin):
                 logging.warning("Fused liger_rope is not supported for DeepseekV2.")
             if cfg.liger_rms_norm:
                 modeling_mod.DeepseekV2RMSNorm = LigerRMSNorm
-            if cfg.liger_glu_activation:
+            if cfg.liger_swiglu:
                 modeling_mod.DeepseekV2MLP.forward = LigerSwiGLUMLP.forward
             if cfg.liger_cross_entropy:
-                # We do not patch `nn.functional.cross_entropy` for DeepseekV2 as it still uses
-                # nn.CrossEntropyLoss in the forward method.
                 modeling_mod.CrossEntropyLoss = LigerCrossEntropyLoss
             if cfg.liger_fused_linear_cross_entropy:
                 modeling_mod.DeepseekV2ForCausalLM.forward = deepseekv2_lce_forward
+
+        elif cfg.model_config_type == "gemma2":
+            from transformers.models.gemma2 import modeling_gemma2
+
+            if cfg.liger_rope:
+                modeling_gemma2.apply_rotary_pos_emb = liger_rotary_pos_emb
+            if cfg.liger_rms_norm:
+                modeling_gemma2.Gemma2RMSNorm = partial(
+                    LigerRMSNorm, offset=1.0, init_fn="zeros", casting_mode="gemma"
+                )
+            if cfg.liger_swiglu:
+                modeling_gemma2.Gemma2MLP = LigerGEGLUMLP
+            if cfg.liger_cross_entropy:
+                modeling_gemma2.CrossEntropyLoss = LigerCrossEntropyLoss
+            if cfg.liger_fused_linear_cross_entropy:
+                logging.warning(
+                    "Fused linear cross entropy is not supported for Gemma 2."
+                )
+
+        elif cfg.model_config_type == "phi3":
+            from liger_kernel.transformers.model.phi3 import (
+                lce_forward as phi3_lce_forward,
+            )
+            from transformers.models.phi3 import modeling_phi3
+
+            if cfg.liger_rope:
+                modeling_phi3.apply_rotary_pos_emb = liger_rotary_pos_emb
+            if cfg.liger_rms_norm:
+                modeling_phi3.Phi3RMSNorm = LigerRMSNorm
+            if cfg.liger_swiglu:
+                modeling_phi3.Phi3MLP = LigerSwiGLUMLP
+            if cfg.liger_cross_entropy:
+                modeling_phi3.CrossEntropyLoss = LigerCrossEntropyLoss
+            if cfg.liger_fused_linear_cross_entropy:
+                modeling_phi3.Phi3ForCausalLM.forward = phi3_lce_forward

src/axolotl/integrations/liger/args.py

@@ -15,12 +15,9 @@
 """
 Module for handling LIGER input arguments.
 """
-import logging
 from typing import Optional
 
-from pydantic import BaseModel, model_validator
-
-LOG = logging.getLogger("axolotl.integrations.liger.args")
+from pydantic import BaseModel
 
 
 class LigerArgs(BaseModel):
@@ -30,24 +27,6 @@ class LigerArgs(BaseModel):
 
     liger_rope: Optional[bool] = None
     liger_rms_norm: Optional[bool] = None
-    liger_layer_norm: Optional[bool] = None
     liger_swiglu: Optional[bool] = None
-    liger_glu_activation: Optional[bool] = None
     liger_cross_entropy: Optional[bool] = None
     liger_fused_linear_cross_entropy: Optional[bool] = None
-
-    @model_validator(mode="before")
-    @classmethod
-    def check_deprecated_swiglu(cls, data):
-        if data.get("liger_swiglu") is not None:
-            if data.get("liger_glu_activation") is not None:
-                raise ValueError(
-                    "You cannot have both `liger_swiglu` and `liger_glu_activation` set."
-                )
-
-            LOG.warning(
-                "The 'liger_swiglu' argument is deprecated and will be removed in a future release. "
-                "Please use 'liger_glu_activation' instead."
-            )
-            data["liger_glu_activation"] = data.pop("liger_swiglu")
-        return data

src/axolotl/monkeypatch/fastchat_conversation_turns.py

@@ -0,0 +1,231 @@
+"""
+monkeypatch to add a get_turns method
+"""
+
+import logging
+from typing import Generator, Tuple
+
+from fastchat.conversation import SeparatorStyle
+
+LOG = logging.getLogger("axolotl.monkeypatch.fastchat_conversation_turns")
+
+
+def get_prompt(self) -> str:
+    ret = ""
+    for role, msg in self.get_turns():
+        ret += role + msg
+    return ret
+
+
+def get_turns(  # pylint: disable=too-many-return-statements
+    self,
+) -> Generator[Tuple[str, str], None, None]:
+    """Get the prompt for generation."""
+    system_prompt = self.system_template.format(system_message=self.system_message)
+    if self.sep_style == SeparatorStyle.ADD_COLON_SINGLE:
+        yield "", system_prompt + self.sep
+        for role, message in self.messages:
+            if message:
+                yield role + ": ", message + self.sep
+            else:
+                yield role + ":", ""
+        return
+    if self.sep_style == SeparatorStyle.ADD_COLON_TWO:
+        seps = [self.sep, self.sep2]
+        yield "", system_prompt + seps[0]
+        for i, (role, message) in enumerate(self.messages):
+            if message:
+                yield role + ": ", message + seps[i % 2]
+            else:
+                yield role + ":", ""
+        return
+    if self.sep_style == SeparatorStyle.ADD_COLON_SPACE_SINGLE:
+        yield "", system_prompt + self.sep
+        for role, message in self.messages:
+            if message:
+                yield role + ": ", message + self.sep
+            else:
+                yield role + ": ", ""  # must be end with a space
+        return
+    if self.sep_style == SeparatorStyle.ADD_NEW_LINE_SINGLE:
+        yield "", "" if system_prompt == "" else system_prompt + self.sep
+        for role, message in self.messages:
+            if message:
+                yield role + "\n", message + self.sep
+            else:
+                yield role + "\n", ""
+        return
+    if self.sep_style == SeparatorStyle.NO_COLON_SINGLE:
+        yield "", system_prompt
+        for role, message in self.messages:
+            if message:
+                yield role, message + self.sep
+            else:
+                yield role, ""
+        return
+    if self.sep_style == SeparatorStyle.NO_COLON_TWO:
+        seps = [self.sep, self.sep2]
+        yield "", system_prompt
+        for i, (role, message) in enumerate(self.messages):
+            if message:
+                yield role, message + seps[i % 2]
+            else:
+                yield role, ""
+        return
+    if self.sep_style == SeparatorStyle.RWKV:
+        yield "", system_prompt
+        for i, (role, message) in enumerate(self.messages):
+            if message:
+                yield role + ": ", message.replace("\r\n", "\n").replace(
+                    "\n\n", "\n"
+                ) + "\n\n"
+            else:
+                yield role + ":", ""
+        return
+    if self.sep_style == SeparatorStyle.LLAMA2 and self.name != "mistral":
+        if self.system_message:
+            if self.messages:
+                # For llama, the system message is incorporated into the first human instruction
+                first_role, first_msg = self.messages[0]
+                if first_role == self.roles[0]:
+                    system_prompt += first_msg
+                    self.messages.pop(0)
+            yield "", system_prompt
+        for i, (role, message) in enumerate(self.messages):
+            if message:
+                if (i % 2 == 0 and not self.system_message) or (
+                    i % 2 != 0 and self.system_message
+                ):
+                    role = "<s> " + role
+                yield role + " ", message
+            else:
+                yield role, ""
+        return
+    if self.sep_style == SeparatorStyle.LLAMA2 and self.name == "mistral":
+        contains_sys_msg = False
+        if self.system_message:
+            contains_sys_msg = True
+            if self.messages:
+                # There is no clear guidance on how to handle system messages in Mistral so we just prepend it to the first human instruction separated by a newline
+                first_role, first_msg = self.messages[0]
+                if first_role == self.roles[0]:
+                    system_prompt = self.system_template.format(
+                        system_message=" " + self.system_message
+                    )
+                    system_prompt += first_msg
+                    self.messages.pop(0)
+            yield "", system_prompt
+        for i, (role, message) in enumerate(self.messages):
+            if message and i == 0 and not contains_sys_msg:
+                yield "", system_prompt.strip() + " " + message  # if there is no system message, we need to make sure there is the a `<s> [INST]` at the beginning of the first instruction.
+            elif message:
+                yield role + " ", message
+            else:
+                yield role, ""
+        return
+    if self.sep_style == SeparatorStyle.LLAMA3:
+        if self.system_message:
+            # For llama3, the system message is NOT incorporated into the first human instruction
+            # All messages follow <|start_header_id|>' + role + '<|end_header_id|>\n\n'+ message + '<|eot_id|>
+            yield "", system_prompt
+        for i, (role, message) in enumerate(self.messages):
+            if message:
+                yield f"<|start_header_id|>{role}<|end_header_id|>\n\n", f"{message.strip()}<|eot_id|>"
+            else:
+                yield f"<|start_header_id|>{role}<|end_header_id|>\n\n", ""
+        return
+    if self.sep_style == SeparatorStyle.GEMMA:
+        if self.system_message:
+            raise ValueError("Gemma chat template does not support system messages")
+        for i, (role, message) in enumerate(self.messages):
+            prefix = "<bos>" if i == 0 else ""
+            message_str = message if message else ""
+            yield prefix + "<start_of_turn>" + role + "\n", message_str + "<end_of_turn>\n"
+        return
+    if self.sep_style == SeparatorStyle.CHATGLM:
+        # source: https://huggingface.co/THUDM/chatglm-6b/blob/1d240ba371910e9282298d4592532d7f0f3e9f3e/modeling_chatglm.py#L1302-L1308
+        # source2: https://huggingface.co/THUDM/chatglm2-6b/blob/e186c891cf64310ac66ef10a87e6635fa6c2a579/modeling_chatglm.py#L926
+        round_add_n = 1 if self.name == "chatglm2" else 0
+        if system_prompt:
+            yield "", system_prompt + self.sep
+
+        for i, (role, message) in enumerate(self.messages):
+            if i % 2 == 0:
+                yield "", f"[Round {i//2 + round_add_n}]{self.sep}"
+
+            if message:
+                yield f"{role}：", f"{message}{self.sep}"
+            else:
+                yield f"{role}：", ""
+        return
+    if self.sep_style == SeparatorStyle.CHATML:
+        yield "", "" if system_prompt == "" else system_prompt + self.sep + "\n"
+        for role, message in self.messages:
+            if message:
+                yield role + "\n", message + self.sep + "\n"
+            else:
+                yield role + "\n", ""
+        return
+    if self.sep_style == SeparatorStyle.CHATGLM3:
+        if self.system_message:
+            yield "", system_prompt
+        for role, message in self.messages:
+            if message:
+                yield role + "\n", " " + message
+            else:
+                yield role
+        return
+    if self.sep_style == SeparatorStyle.CHATINTERN:
+        # source: https://huggingface.co/internlm/internlm-chat-7b-8k/blob/bd546fa984b4b0b86958f56bf37f94aa75ab8831/modeling_internlm.py#L771
+        seps = [self.sep, self.sep2]
+        yield "", system_prompt
+        for i, (role, message) in enumerate(self.messages):
+            prefix = "<s>" if i % 2 == 0 else ""
+            if message:
+                yield prefix + role + ":", message + seps[i % 2] + "\n"
+            else:
+                yield role + ":", ""
+        return
+    if self.sep_style == SeparatorStyle.DOLLY:
+        seps = [self.sep, self.sep2]
+        yield "", system_prompt
+        for i, (role, message) in enumerate(self.messages):
+            if message:
+                suffix = "\n\n" if i % 2 == 1 else ""
+                yield role + ":\n", message + seps[i % 2] + suffix
+            else:
+                yield role + ":\n", ""
+        return
+    if self.sep_style == SeparatorStyle.PHOENIX:
+        yield "", system_prompt
+        for role, message in self.messages:
+            if message:
+                yield role + ": ", "<s>" + message + "</s>"
+            else:
+                yield role + ": " + "<s>", ""
+        return
+    if self.sep_style == SeparatorStyle.ROBIN:
+        yield "", system_prompt + self.sep
+        for role, message in self.messages:
+            if message:
+                yield role + ":\n", message + self.sep
+            else:
+                yield role + ":\n", ""
+        return
+    if self.sep_style == SeparatorStyle.FALCON_CHAT:
+        if self.system_message:
+            yield "", system_prompt + self.sep
+        for role, message in self.messages:
+            if message:
+                yield role + ": ", message + self.sep
+            else:
+                yield role + ":", ""
+    else:
+        raise ValueError(f"Invalid style: {self.sep_style}")
+
+
+def add_get_turns_to_conversation():
+    import fastchat.conversation
+
+    fastchat.conversation.Conversation.get_turns = get_turns
+    fastchat.conversation.Conversation.get_prompt = get_prompt

src/axolotl/monkeypatch/llama_attn_hijack_flash.py

@@ -22,6 +22,7 @@ from transformers.models.llama.modeling_llama import (
     apply_rotary_pos_emb,
     repeat_kv,
 )
+from xformers.ops import SwiGLU
 
 from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids, set_module_name
 
@@ -43,19 +44,7 @@ except ImportError:
 LOG = logging.getLogger("axolotl")
 
 
-def is_xformers_available() -> bool:
-    try:
-        import xformers  # pylint: disable=unused-import  # noqa: F401
-
-        return True
-    except ImportError:
-        return False
-
-
 def is_xformers_swiglu_available() -> bool:
-    if not is_xformers_available():
-        return False
-
     from xformers.ops.common import get_xformers_operator
 
     try:
@@ -68,11 +57,6 @@ def is_xformers_swiglu_available() -> bool:
 
 
 def replace_llama_mlp_with_swiglu(model):
-    if is_xformers_swiglu_available():
-        from axolotl.monkeypatch.xformers_ import FusedMLP
-    else:
-        raise RuntimeError("xformers SwiGLU not available for this environment")
-
     for name, module in model.named_modules():
         if isinstance(module, LlamaMLP):
             mlp = FusedMLP(
@@ -197,6 +181,49 @@ class FusedAttention(LlamaAttention):
         set_module_name(model, name, new_attn)
 
 
+class FusedMLP(torch.nn.Module):
+    """
+    Fused MLP layer for incrementally improved training efficiency
+    """
+
+    def __init__(
+        self,
+        config,
+        gate_proj: torch.nn.Linear,
+        up_proj: torch.nn.Linear,
+        down_proj: torch.nn.Linear,
+    ):
+        super().__init__()
+        self.config = config
+        self.swiglu = SwiGLU(
+            in_features=config.hidden_size,
+            hidden_features=config.intermediate_size,
+            bias=False,
+            _pack_weights=True,
+        )
+        # overwrite initialized weights with pretrained weights
+        self.swiglu.w12.weight.data = torch.cat(
+            (gate_proj.weight.data, up_proj.weight.data), dim=0
+        )
+        self.swiglu.w3.weight.data = down_proj.weight.data
+
+    def _post_training(self, model, name):
+        w1, w2 = torch.split(  # pylint: disable=invalid-name
+            self.swiglu.w12.weight.data, self.config.intermediate_size, dim=0
+        )
+
+        # Assign the split weights back to the original layers
+        new_mlp = LlamaMLP(self.config)
+        new_mlp.gate_proj.weight.data = w1
+        new_mlp.up_proj.weight.data = w2
+        new_mlp.down_proj.weight.data = self.swiglu.w3.weight.data
+
+        set_module_name(model, name, new_mlp)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:  # pylint: disable=invalid-name
+        return self.swiglu(x)
+
+
 # Disable the transformation of the attention mask in LlamaModel as the flash attention
 # requires the attention mask to be the same as the key_padding_mask
 def _prepare_decoder_attention_mask(

src/axolotl/monkeypatch/unsloth_.py

@@ -16,6 +16,26 @@ from transformers.models.llama.modeling_llama import (
 
 LOG = get_logger("axolotl.monkeypatch.unsloth")
 
+ORIGINAL_CEL_CODE = """# Shift so that tokens < n predict n
+        shift_logits = logits[..., :-1, :].contiguous()
+        shift_labels = labels[..., 1:].contiguous()
+        # Flatten the tokens
+        loss_fct = CrossEntropyLoss()
+        shift_logits = shift_logits.view(-1, self.config.vocab_size)
+        shift_labels = shift_labels.view(-1)
+        # Enable model parallelism
+        shift_labels = shift_labels.to(shift_logits.device)
+        loss = loss_fct(shift_logits, shift_labels)
+"""
+
+PATCHED_CEL_CODE = """shift_logits = logits[..., :-1, :].contiguous()
+        shift_labels = labels[..., 1:].contiguous()
+        loss = fast_cross_entropy_loss(
+            logits = shift_logits,
+            labels = shift_labels,
+        )
+"""
+
 ORIGINAL_QKV_CODE = """
     query_states = self.q_proj(hidden_states)
     key_states = self.k_proj(hidden_states)
@@ -60,6 +80,12 @@ def get_forward_code() -> str:
     return forward
 
 
+def check_cel_is_patchable() -> bool:
+    forward = get_forward_code()
+    forward, _ = detab_code(forward)
+    return ORIGINAL_CEL_CODE in forward
+
+
 def get_self_attn_code() -> str:
     forward = inspect.getsource(LlamaFlashAttention2.forward)
     return forward
@@ -72,31 +98,48 @@ def check_self_attn_is_patchable() -> bool:
 
 
 def integrate_cross_entropy_loss_patch(model_type: str = "llama") -> None:
-    from unsloth.kernels.cross_entropy_loss import fast_cross_entropy_loss
-
-    def UnslothForCausalLMLoss(  # pylint: disable=invalid-name
-        logits,
-        labels,
-        vocab_size: int,  # pylint: disable=unused-argument
-        num_items_in_batch: int = None,
-        ignore_index: int = -100,  # pylint: disable=unused-argument
-        **kwargs,  # pylint: disable=unused-argument
-    ):
-        # Upcast to float if we need to compute the loss to avoid potential precision issues
-        logits = logits.float()
-        # Shift so that tokens < n predict n
-        shift_logits = logits[..., :-1, :].contiguous()
-        shift_labels = labels[..., 1:].contiguous()
-
-        loss = fast_cross_entropy_loss(
-            logits=shift_logits, labels=shift_labels, n_items=num_items_in_batch
-        )
-        return loss
-
     if model_type == "llama":
-        from transformers.loss import loss_utils
+        forward = get_forward_code()
+        LlamaForCausalLM._original_forward = forward  # pylint: disable=protected-access
+        forward, _ = detab_code(forward)
+        assert ORIGINAL_CEL_CODE in forward, "Original forward code not found"
 
-        loss_utils.ForCausalLMLoss = UnslothForCausalLMLoss  # type: ignore[assignment]
+        forward = forward.replace(
+            "@add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)", ""
+        )
+        forward = forward.replace(
+            "@replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)",
+            "",
+        )
+        forward = forward.replace(ORIGINAL_CEL_CODE, PATCHED_CEL_CODE)
+        forward = forward.replace(
+            "def forward(",
+            "def fast_cross_entropy_loss_forward(",
+            1,
+        )
+
+        # load imports necessary
+        import transformers.models.llama.modeling_llama
+
+        items_to_import = []
+        for item in dir(transformers.models.llama.modeling_llama):
+            if item in forward:
+                items_to_import.append(item)
+
+        exec(  # pylint: disable=exec-used  # nosec B102
+            "from unsloth.kernels.cross_entropy_loss import fast_cross_entropy_loss",
+            globals(),
+        )
+
+        exec(  # pylint: disable=exec-used  # nosec B102
+            "from transformers.models.llama.modeling_llama import ("
+            + ", ".join(x for x in items_to_import)
+            + ")",
+            globals(),
+        )
+        exec(forward, globals())  # pylint: disable=exec-used  # nosec B102
+        LOG.info("patching unsloth fast_cross_entropy_loss", main_process_only=True)
+        LlamaForCausalLM.forward = fast_cross_entropy_loss_forward  # pylint: disable=undefined-variable  # noqa: F821
     else:
         raise ValueError("Unsupported model type")
 

src/axolotl/monkeypatch/xformers_/__init__.py

@@ -1,51 +0,0 @@
-"""
-Fused MLP layer for incrementally improved training efficiency
-"""
-import torch
-from transformers.models.llama.modeling_llama import LlamaMLP
-from xformers.ops import SwiGLU
-
-from axolotl.monkeypatch.utils import set_module_name
-
-
-class FusedMLP(torch.nn.Module):
-    """
-    Fused MLP layer for incrementally improved training efficiency
-    """
-
-    def __init__(
-        self,
-        config,
-        gate_proj: torch.nn.Linear,
-        up_proj: torch.nn.Linear,
-        down_proj: torch.nn.Linear,
-    ):
-        super().__init__()
-        self.config = config
-        self.swiglu = SwiGLU(
-            in_features=config.hidden_size,
-            hidden_features=config.intermediate_size,
-            bias=False,
-            _pack_weights=True,
-        )
-        # overwrite initialized weights with pretrained weights
-        self.swiglu.w12.weight.data = torch.cat(
-            (gate_proj.weight.data, up_proj.weight.data), dim=0
-        )
-        self.swiglu.w3.weight.data = down_proj.weight.data
-
-    def _post_training(self, model, name):
-        w1, w2 = torch.split(  # pylint: disable=invalid-name
-            self.swiglu.w12.weight.data, self.config.intermediate_size, dim=0
-        )
-
-        # Assign the split weights back to the original layers
-        new_mlp = LlamaMLP(self.config)
-        new_mlp.gate_proj.weight.data = w1
-        new_mlp.up_proj.weight.data = w2
-        new_mlp.down_proj.weight.data = self.swiglu.w3.weight.data
-
-        set_module_name(model, name, new_mlp)
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:  # pylint: disable=invalid-name
-        return self.swiglu(x)

src/axolotl/prompt_strategies/bradley_terry/__init__.py

@@ -6,7 +6,7 @@ import logging
 
 from axolotl.prompt_strategies.user_defined import UserDefinedDatasetConfig
 
-LOG = logging.getLogger("axolotl.prompt_strategies.bradley_terry")
+LOG = logging.getLogger("axolotl.prompt_strategies")
 
 
 def load(strategy, tokenizer, cfg, ds_cfg):

src/axolotl/prompt_strategies/bradley_terry/chat_template.py

@@ -2,18 +2,13 @@
 Bradley-Terry model with chat template prompt strategy.
 """
 
-import logging
 from typing import Any, Dict, Optional
 
 from axolotl.prompt_strategies.chat_template import (
     ChatTemplatePrompter,
     ChatTemplateStrategy,
 )
-from axolotl.utils.chat_templates import get_chat_template_from_config
-
-# Configure the logger
-LOG = logging.getLogger("axolotl.prompt_strategies.bradley_terry.chat_template")
-LOG.setLevel(logging.INFO)
+from axolotl.utils.chat_templates import chat_templates
 
 
 class BTChatTemplateStrategy(ChatTemplateStrategy):
@@ -32,24 +27,18 @@ class BTChatTemplateStrategy(ChatTemplateStrategy):
         # pylint: disable=duplicate-code
         prompt[self.messages] = []
         if prompt["system"]:
-            prompt[self.messages].append(
-                {"role": "system", "content": prompt["system"]}
-            )
-        prompt[self.messages].append({"role": "user", "content": prompt["input"]})
-        prompt[self.messages].append({"role": "assistant", "content": prompt["chosen"]})
+            prompt[self.messages].append({"from": "system", "value": prompt["system"]})
+        prompt[self.messages].append({"from": "user", "value": prompt["input"]})
+        prompt[self.messages].append({"from": "assistant", "value": prompt["chosen"]})
         chosen_tokenized = super().tokenize_prompt(prompt)
 
         self.messages = "rejected_messages"
         # pylint: disable=duplicate-code
         prompt[self.messages] = []
         if prompt["system"]:
-            prompt[self.messages].append(
-                {"role": "system", "content": prompt["system"]}
-            )
-        prompt[self.messages].append({"role": "user", "content": prompt["input"]})
-        prompt[self.messages].append(
-            {"role": "assistant", "content": prompt["rejected"]}
-        )
+            prompt[self.messages].append({"from": "system", "value": prompt["system"]})
+        prompt[self.messages].append({"from": "user", "value": prompt["input"]})
+        prompt[self.messages].append({"from": "assistant", "value": prompt["rejected"]})
         rejected_tokenized = super().tokenize_prompt(prompt)
 
         return {
@@ -64,18 +53,15 @@ class BTChatTemplateStrategy(ChatTemplateStrategy):
 
 def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
     ds_cfg = ds_cfg or {}
-    chat_template_string = get_chat_template_from_config(
-        cfg=cfg, ds_cfg=ds_cfg, tokenizer=tokenizer
-    )
 
     prompter_params = {
         "tokenizer": tokenizer,
-        "chat_template": chat_template_string,
-        "message_field_role": ds_cfg.get("message_field_role", "role"),
-        "message_field_content": ds_cfg.get("message_field_content", "content"),
-        "message_field_training": ds_cfg.get("message_field_training", None),
+        "chat_template": chat_templates(ds_cfg.get("chat_template", "chatml")),
+        "message_field_role": ds_cfg.get("message_field_role", "from"),
+        "message_field_content": ds_cfg.get("message_field_content", "value"),
+        "message_field_training": ds_cfg.get("message_field_training", "training"),
         "message_field_training_detail": ds_cfg.get(
-            "message_field_training_detail", None
+            "message_field_training_detail", "train_detail"
         ),
         "roles": ds_cfg.get("roles"),
         "drop_system_message": ds_cfg.get("drop_system_message", False),
@@ -88,8 +74,8 @@ def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
     strategy_params = {
         "train_on_inputs": cfg.train_on_inputs,
         "sequence_len": cfg.sequence_len,
-        "roles_to_train": ds_cfg.get("roles_to_train", []),
-        "train_on_eos": ds_cfg.get("train_on_eos", None),
+        "roles_to_train": ds_cfg.get("roles_to_train", ["gpt", "assistant"]),
+        "train_on_eos": ds_cfg.get("train_on_eos", "turn"),
     }
 
     strategy = BTChatTemplateStrategy(

src/axolotl/prompt_strategies/chat_template.py

@@ -9,7 +9,7 @@ from transformers import ProcessorMixin
 
 from axolotl.prompt_tokenizers import PromptTokenizingStrategy
 from axolotl.prompters import IGNORE_TOKEN_ID, Prompter
-from axolotl.utils.chat_templates import get_chat_template_from_config
+from axolotl.utils.chat_templates import chat_templates
 
 # Configure the logger
 LOG = logging.getLogger("axolotl")
@@ -405,14 +405,10 @@ class ChatTemplateStrategy(PromptTokenizingStrategy):
 def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None, processor=None):
     # pylint: disable=duplicate-code
     ds_cfg = ds_cfg or {}
-    chat_template_string = get_chat_template_from_config(
-        cfg=cfg, ds_cfg=ds_cfg, tokenizer=tokenizer
-    )
-    LOG.info(f"Using chat template:\n---\n{chat_template_string!s}\n---")
 
     prompter_params = {
         "tokenizer": tokenizer,
-        "chat_template": chat_template_string,
+        "chat_template": chat_templates(ds_cfg.get("chat_template", "chatml")),
         "message_field_role": ds_cfg.get("message_field_role", "role"),
         "message_field_content": ds_cfg.get("message_field_content", "content"),
         "message_field_training": ds_cfg.get("message_field_training", None),

src/axolotl/prompt_strategies/dpo/chat_template.py

@@ -2,16 +2,15 @@
 DPO prompt strategies for using tokenizer chat templates.
 """
 
-from axolotl.utils.chat_templates import extract_chat_template_args, get_chat_template
+from axolotl.utils.chat_templates import chat_templates
 
 
 def default(
     cfg, dataset_idx=0, **kwargs
 ):  # pylint: disable=possibly-unused-variable,unused-argument
     ds_cfg = cfg["datasets"][dataset_idx]
-    chat_template_choice, chat_template_jinja = extract_chat_template_args(
-        cfg=cfg, ds_cfg=ds_cfg
-    )
+    chat_template_str = chat_templates(cfg.chat_template)
+
     field_messages = ds_cfg.get("field_messages", "messages")
     field_chosen = ds_cfg.get("field_chosen", "chosen")
     field_rejected = ds_cfg.get("field_rejected", "rejected")
@@ -31,12 +30,6 @@ def default(
             role_map[source] = target
 
     def transform_fn(sample, tokenizer=None):
-        chat_template_string = get_chat_template(
-            user_choice=chat_template_choice,
-            jinja_template=chat_template_jinja,
-            tokenizer=tokenizer,
-        )
-
         messages = sample[field_messages]
         messages = [
             {
@@ -53,29 +46,28 @@ def default(
             "role": role_map[sample[field_rejected][field_message_role]],
             "content": sample[field_rejected][field_message_content],
         }
-        dummy_user_message = {"role": "user", "content": "[[dummy_message]]"}
 
         result = {}
         result["prompt"] = tokenizer.apply_chat_template(
             messages,
             add_generation_prompt=True,
-            chat_template=chat_template_string,
+            chat_template=chat_template_str,
             tokenize=False,
         )
 
         result["chosen"] = tokenizer.apply_chat_template(
-            [dummy_user_message, chosen],
+            [chosen],
             add_generation_prompt=False,
-            chat_template=chat_template_string,
+            chat_template=chat_template_str,
             tokenize=False,
         )
         chosen_strip_index = result["chosen"].find(chosen["content"])
         result["chosen"] = result["chosen"][chosen_strip_index:].rstrip()
 
         result["rejected"] = tokenizer.apply_chat_template(
-            [dummy_user_message, rejected],
+            [rejected],
             add_generation_prompt=False,
-            chat_template=chat_template_string,
+            chat_template=chat_template_str,
             tokenize=False,
         )
         rejected_strip_index = result["rejected"].find(rejected["content"])

src/axolotl/prompt_strategies/instruct.py

@@ -0,0 +1,33 @@
+"""Module containing the InstructShareGPTPromptTokenizingStrategy class"""
+from typing import Any, Dict, Optional
+
+from axolotl.prompt_tokenizers import ShareGPTPromptTokenizingStrategy
+from axolotl.prompters import ShareGPTPrompterV2
+
+
+def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
+    conversation = (
+        ds_cfg["conversation"] if ds_cfg and "conversation" in ds_cfg else None
+    )
+    strategy = InstructShareGPTPromptTokenizingStrategy(
+        # pylint: disable=duplicate-code
+        ShareGPTPrompterV2(
+            conversation=conversation,
+        ),
+        tokenizer,
+        cfg.train_on_inputs,
+        cfg.sequence_len,
+    )
+    return strategy
+
+
+class InstructShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
+    """
+    basic sharegpt strategy to grab conversations from the sample row
+    """
+
+    def get_conversation_thread(self, prompt):
+        return [
+            {"from": "human", "value": prompt["instruction"]},
+            {"from": "gpt", "value": prompt["output"]},
+        ]

src/axolotl/prompt_strategies/llama2_chat.py

@@ -29,7 +29,7 @@ from dataclasses import dataclass, field
 from typing import Generator, List, Sequence
 
 from axolotl.prompt_tokenizers import PromptTokenizingStrategy
-from axolotl.prompters import ALTERNATING_ASSERTION_FAILED_ROLE, IGNORE_TOKEN_ID
+from axolotl.prompters import IGNORE_TOKEN_ID, SHAREGPT_ASSERTION_FAILED_ROLE
 
 
 @dataclass
@@ -75,7 +75,7 @@ class Llama2ChatConversation:
 
 class LLama2ChatTokenizingStrategy(PromptTokenizingStrategy):
     """
-    Tokenizing strategy for Llama2 prompts.
+    Tokenizing strategy for ShareGPT prompts.
     adapted from https://github.com/lm-sys/FastChat/blob/main/fastchat/train/train.py
     """
 
@@ -191,7 +191,7 @@ class Llama2ChatPrompter:  # pylint: disable=too-few-public-methods
         conv.messages = []  # pylint: disable=R0801
         for j, sentence in enumerate(source):
             role = roles[sentence["from"]]
-            assert role == conv.roles[j % 2], ALTERNATING_ASSERTION_FAILED_ROLE
+            assert role == conv.roles[j % 2], SHAREGPT_ASSERTION_FAILED_ROLE
             if sentence["value"]:
                 conv.append_message(role, sentence["value"])
         yield conv

src/axolotl/prompt_strategies/orpo/chat_template.py

@@ -5,7 +5,7 @@ from pydantic import BaseModel
 
 from axolotl.prompt_tokenizers import IGNORE_INDEX, PromptTokenizingStrategy
 from axolotl.prompters import Prompter
-from axolotl.utils.chat_templates import get_chat_template_from_config
+from axolotl.utils.chat_templates import chat_templates
 
 
 class Message(BaseModel):
@@ -28,13 +28,18 @@ def load(
     """
     chatml transforms for datasets with system, input, chosen, rejected
     """
-    chat_template_string = get_chat_template_from_config(
-        cfg=cfg, ds_cfg=ds_cfg, tokenizer=tokenizer
-    )
-    tokenizer.chat_template = chat_template_string
+
+    chat_template = chat_templates("chatml")
+    if ds_cfg and "chat_template" in ds_cfg:
+        chat_template = ds_cfg["chat_template"]
+        try:
+            chat_template = chat_templates(chat_template)
+        except ValueError:
+            pass
+    tokenizer.chat_template = chat_template
 
     return ORPOTokenizingStrategy(
-        ORPOPrompter(chat_template_string, tokenizer),
+        ORPOPrompter(chat_template, tokenizer),
         tokenizer,
         cfg.train_on_inputs,
         cfg.sequence_len,
@@ -243,30 +248,28 @@ class ORPOPrompter(Prompter):
 def argilla(cfg, **kwargs):  # pylint: disable=possibly-unused-variable,unused-argument
     dataset_parser = ORPODatasetParsingStrategy()
 
+    chat_template_str = chat_templates(cfg.chat_template)
+
     def transform_fn(sample, tokenizer=None):
         res = {}
 
-        chat_template_string = get_chat_template_from_config(
-            cfg=cfg, tokenizer=tokenizer
-        )
-
         res["prompt"] = tokenizer.apply_chat_template(
             [msg.model_dump() for msg in dataset_parser.get_prompt(sample).messages],
             add_generation_prompt=True,
-            chat_template=chat_template_string,
+            chat_template=chat_template_str,
             tokenize=False,
         )
         prompt_str_len = len(res["prompt"])
         res["chosen"] = tokenizer.apply_chat_template(
             [msg.model_dump() for msg in dataset_parser.get_chosen(sample).messages],
             add_generation_prompt=False,
-            chat_template=chat_template_string,
+            chat_template=chat_template_str,
             tokenize=False,
         )[prompt_str_len:]
         res["rejected"] = tokenizer.apply_chat_template(
             [msg.model_dump() for msg in dataset_parser.get_rejected(sample).messages],
             add_generation_prompt=False,
-            chat_template=chat_template_string,
+            chat_template=chat_template_str,
             tokenize=False,
         )[prompt_str_len:]
 

src/axolotl/prompt_strategies/sharegpt.py

@@ -0,0 +1,223 @@
+"""Module containing the SimpleShareGPTPromptTokenizingStrategy class"""
+
+import logging
+from typing import Any, Dict, Optional, Type
+
+from fastchat.conversation import Conversation, SeparatorStyle, register_conv_template
+
+from axolotl.prompt_tokenizers import ShareGPTPromptTokenizingStrategy
+from axolotl.prompters import ShareGPTPrompterV2
+from axolotl.utils.tokenization import (
+    chatml_to_conversation,
+    merge_consecutive_messages,
+)
+
+LOG = logging.getLogger("axolotl")
+
+
+def register_chatml_template(system_message=None):
+    system_message = system_message or "You are a helpful assistant."
+    register_conv_template(
+        Conversation(
+            name="chatml",
+            system_template="<|im_start|>system\n{system_message}",
+            system_message=system_message,
+            roles=("<|im_start|>user", "<|im_start|>assistant"),
+            sep_style=SeparatorStyle.CHATML,
+            sep="<|im_end|>",
+        )
+    )
+    register_conv_template(
+        Conversation(
+            name="chatml_glaive",
+            system_template="<|im_start|>system\n{system_message}",
+            system_message=system_message,
+            roles=("<|im_start|>user", "<|im_start|>assistant", "<|im_start|>tool"),
+            sep_style=SeparatorStyle.CHATML,
+            sep="<|im_end|>",
+        )
+    )
+
+
+def register_llama3_template(system_message=None):
+    system_message = system_message or "You are a helpful assistant."
+    register_conv_template(
+        Conversation(
+            name="llama3",
+            system_template="<|start_header_id|>system<|end_header_id|>\n\n{system_message}<|eot_id|>",
+            system_message=system_message,
+            roles=("user", "assistant"),
+            sep_style=SeparatorStyle.LLAMA3,
+            sep="",
+            stop_str="<|eot_id|>",
+            stop_token_ids=[128001, 128009],
+        )
+    )
+
+
+def build_loader(
+    tokenization_strategy_cls: Type["ShareGPTPromptTokenizingStrategy"],
+    prompter_cls: Type["ShareGPTPrompterV2"],
+    default_conversation: Optional[str] = None,
+):
+    def _load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
+        LOG.warning(
+            "sharegpt type support will be deprecated in the next release of Axolotl. Please use chat_template instead.",
+        )
+        conversation = (
+            ds_cfg["conversation"]
+            if ds_cfg and "conversation" in ds_cfg
+            else default_conversation
+        )
+        field_human = (
+            ds_cfg["field_human"] if ds_cfg and "field_human" in ds_cfg else None
+        )
+        field_model = (
+            ds_cfg["field_model"] if ds_cfg and "field_model" in ds_cfg else None
+        )
+        roles = ds_cfg["roles"].to_dict() if ds_cfg and "roles" in ds_cfg else None
+        strategy = tokenization_strategy_cls(
+            prompter_cls(
+                conversation=conversation,
+                role_key_model=field_model,
+                role_key_human=field_human,
+                roles=roles,
+            ),
+            tokenizer,
+            cfg.train_on_inputs,
+            cfg.sequence_len,
+        )
+        if ds_cfg and "strict" in ds_cfg and hasattr(strategy, "strict"):
+            strategy.strict = ds_cfg["strict"]
+        if ds_cfg and "field_messages" in ds_cfg and hasattr(strategy, "messages"):
+            strategy.messages = ds_cfg["field_messages"]
+        return strategy
+
+    return _load
+
+
+class SimpleShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
+    """
+    basic sharegpt strategy to grab conversations from the sample row
+    """
+
+    _strict = False
+    _messages = "conversations"
+
+    @property
+    def strict(self):
+        return self._strict
+
+    @strict.setter
+    def strict(self, strict):
+        self._strict = strict
+
+    @property
+    def messages(self):
+        return self._messages
+
+    @messages.setter
+    def messages(self, messages):
+        self._messages = messages
+
+    def get_conversation_thread(self, prompt):
+        conversations = prompt[self.messages]
+        if self.strict:
+            return conversations
+        role_key = "from"
+        if "role" in conversations[0].keys():
+            role_key = "role"
+        value_key = "value"
+        if "text" in conversations[0].keys():
+            value_key = "text"
+        elif "content" in conversations[0].keys():
+            value_key = "content"
+        # remap roles - allow for assistant turn"
+        role_map = {
+            "user": "human",
+            "human": "human",
+            "assistant": "gpt",
+            "gpt": "gpt",
+            "system": "system",
+        }
+        turns = [
+            {
+                "from": (
+                    role_map[t[role_key]] if t[role_key] in role_map else t[role_key]
+                ),
+                "value": t[value_key],
+                "weight": 1
+                if "weight" not in t or t["weight"] is None
+                else t["weight"],
+            }
+            for t in conversations
+        ]
+        return turns
+
+
+class SimpleRoleShareGPTPromptTokenizingStrategy(
+    SimpleShareGPTPromptTokenizingStrategy
+):
+    """
+    basic sharegpt strategy to grab conversations from the sample row, but uses role instead of from
+    """
+
+    def get_conversation_thread(self, prompt):
+        conversations = prompt["conversations"]
+        # remap role: prompter/assistant, text: ... => from: human/gpt, value: ...
+        turns = [{"from": t["role"], "value": t["value"]} for t in conversations]
+        return turns
+
+
+class GuanacoShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
+    """
+    sharegpt strategy that remaps oasst data to sharegpt format
+    """
+
+    def get_conversation_thread(self, prompt):
+        conversations = prompt["conversations"]
+        # remap role: prompter/assistant, text: ... => from: human/gpt, value: ...
+        role_map = {"prompter": "human", "assistant": "gpt"}
+        turns = [
+            {"from": role_map[t["role"]], "value": t["text"]} for t in conversations
+        ]
+        return turns
+
+
+class UltrachatShareGPTPromptTokenizingStrategy(SimpleShareGPTPromptTokenizingStrategy):
+    """
+    sharegpt strategy that remaps ultrachat data to sharegpt format
+    """
+
+    def get_conversation_thread(self, prompt):
+        conversations = prompt["messages"]
+        role_map = {"user": "human", "assistant": "gpt"}
+        turns = [
+            {"from": role_map[t["role"]], "value": t["content"]} for t in conversations
+        ]
+        return turns
+
+
+class GlaiveShareGPTPromptTokenizingStrategy(SimpleShareGPTPromptTokenizingStrategy):
+    """
+    sharegpt strategy that remaps glaive data to sharegpt format
+    """
+
+    def get_conversation_thread(self, prompt):
+        conversation = chatml_to_conversation(prompt)
+        conversation = merge_consecutive_messages(conversation)
+
+        return conversation
+
+
+load = build_loader(SimpleShareGPTPromptTokenizingStrategy, ShareGPTPrompterV2)
+load_role = build_loader(SimpleRoleShareGPTPromptTokenizingStrategy, ShareGPTPrompterV2)
+load_ultrachat = build_loader(
+    UltrachatShareGPTPromptTokenizingStrategy, ShareGPTPrompterV2
+)
+load_guanaco = build_loader(GuanacoShareGPTPromptTokenizingStrategy, ShareGPTPrompterV2)
+load_glaive = build_loader(
+    GlaiveShareGPTPromptTokenizingStrategy,
+    ShareGPTPrompterV2,
+    default_conversation="chatml_glaive",
+)

src/axolotl/prompt_strategies/sharegpt_jokes.py

@@ -0,0 +1,28 @@
+"""Module for Jokes prompts using sharegpt style """
+from axolotl.prompt_tokenizers import ShareGPTPromptTokenizingStrategy
+from axolotl.prompters import ShareGPTPrompterV2
+
+
+def load(tokenizer, cfg):
+    return SimpleJokesShareGPTPromptTokenizingStrategy(
+        ShareGPTPrompterV2(),
+        tokenizer,
+        cfg.train_on_inputs,
+        cfg.sequence_len,
+    )
+
+
+class SimpleJokesShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
+    """
+    Tokenization strategy for asking bot to tell a joke and then explain why its funny
+    """
+
+    # title, text, explanation
+    def get_conversation_thread(self, prompt):
+        title = "" if not prompt["title"] else prompt["title"] + " "
+        return [
+            {"from": "human", "value": "Tell me a joke."},
+            {"from": "gpt", "value": title + prompt["text"]},
+            {"from": "human", "value": "Why is that joke funny?"},
+            {"from": "gpt", "value": prompt["explanation"]},
+        ]

src/axolotl/prompt_tokenizers.py

@@ -1,12 +1,17 @@
 """Module containing PromptTokenizingStrategy and Prompter classes"""
 
 import abc
+import copy
 import logging
 from typing import Dict, List, Tuple, Union
 
+from fastchat.conversation import Conversation
 from transformers import BatchEncoding, PreTrainedTokenizer
 
-from axolotl.prompters import Prompter
+from axolotl.monkeypatch.fastchat_conversation_turns import (
+    add_get_turns_to_conversation,
+)
+from axolotl.prompters import IGNORE_TOKEN_ID, Prompter
 
 LOG = logging.getLogger("axolotl")
 
@@ -16,6 +21,8 @@ LLAMA_DEFAULT_EOS_TOKEN = "</s>"  # nosec
 LLAMA_DEFAULT_BOS_TOKEN = "<s>"  # nosec
 LLAMA_DEFAULT_UNK_TOKEN = "<unk>"  # nosec
 
+add_get_turns_to_conversation()
+
 
 class InvalidDataException(Exception):
     """
@@ -324,6 +331,154 @@ class AlpacaReflectionPTStrategy(ReflectionPromptTokenizingStrategy):
         )
 
 
+class ShareGPTPromptTokenizingStrategy(PromptTokenizingStrategy):
+    """
+    Tokenizing strategy for ShareGPT prompts.
+    """
+
+    def get_conversation_thread(self, prompt):
+        return prompt["conversations"]
+
+    def tokenize_prompt(self, prompt):
+        # Initial values. We will append to these as we go through the conversation.
+        result, current_len = tokenize_prompt_default()
+        conversation: Conversation = (
+            self.prompter._conversation.copy()  # pylint: disable=protected-access
+        )
+
+        input_roles = {conversation.roles[0]}
+        output_roles = {conversation.roles[1]}
+
+        if len(conversation.roles) == 3:
+            tool_role_label = conversation.roles[2]
+            input_roles.add(tool_role_label)
+
+        # Add roles from the config
+        if self.prompter.roles:
+            if "input" in self.prompter.roles and self.prompter.roles["input"]:
+                for role in self.prompter.roles["input"]:
+                    input_roles.add(role)
+
+            if "output" in self.prompter.roles and self.prompter.roles["output"]:
+                for role in self.prompter.roles["output"]:
+                    output_roles.add(role)
+
+        # support for custom roles from the dataset, only useful for vicuna style prompts/roles
+        role_remap = []
+        if (
+            conversation.name == "vicuna_v1.1"
+            and "roles" in prompt
+            and len(prompt["roles"]) >= 2
+        ):
+            role_remap = [
+                {"from": conversation.roles[0], "to": prompt["roles"][0]},
+                {"from": conversation.roles[1], "to": prompt["roles"][1]},
+            ]
+
+        try:
+            for _, part in enumerate(
+                self.prompter.build_prompt(self.get_conversation_thread(prompt))
+            ):
+                if not isinstance(part, tuple):
+                    LOG.warning(f"expected tuple, got {part}")
+                    continue
+
+                if len(part) <= 2:
+                    role, content = part
+                    weight = 1
+                else:
+                    role, content, weight = part
+
+                # Uses "in" because role contains extra characters
+                input_turn = any(r.lower() in role.lower() for r in input_roles)
+                output_turn = any(r.lower() in role.lower() for r in output_roles)
+                empty_role = role.strip() == ""
+
+                if not any([input_turn, output_turn, empty_role]):
+                    LOG.warning(f"unhandled role: {role}")
+                    continue
+
+                if input_turn:
+                    role = (
+                        role.replace(role_remap[0]["from"], role_remap[0]["to"])
+                        if role_remap
+                        else role
+                    )
+                    turn = role + content
+                    # this is still the user query, we should
+                    if not content.strip():
+                        LOG.warning(f"user turn has empty text: {prompt}")
+                    res = self._tokenize(
+                        turn,
+                        add_eos_token=False,
+                        strip_bos_token=True,
+                    )
+                    if self.train_on_inputs and weight == 1:
+                        labels = copy.deepcopy(res["input_ids"])
+                    else:
+                        # everything from this is masked out from the labels
+                        labels = [IGNORE_TOKEN_ID] * len(res["input_ids"])
+                elif output_turn:
+                    role = (
+                        role.replace(role_remap[1]["from"], role_remap[1]["to"])
+                        if role_remap
+                        else role
+                    )
+                    turn = role + content
+                    # this should be the assistant response, should end with an eos token
+                    if not content.strip():
+                        LOG.warning(f"assistant turn has empty text: {prompt}")
+                    add_eos_token = not (
+                        conversation.name == "chatml"
+                        and conversation.sep == self.tokenizer.eos_token
+                    )
+                    res = self._tokenize(
+                        turn,
+                        add_eos_token=add_eos_token,
+                        strip_bos_token=True,
+                    )
+                    role_res = self._tokenize(
+                        role.rstrip(),
+                        add_eos_token=False,
+                        strip_bos_token=True,
+                    )
+                    labels = copy.deepcopy(res["input_ids"])
+                    if not self.train_on_inputs:
+                        # mask out role tokens from the labels
+                        len_role = len(role_res["input_ids"])
+                        labels[:len_role] = [IGNORE_TOKEN_ID] * min(
+                            len_role, len(labels)
+                        )
+                    if weight == 0:
+                        # everything from this is masked out from the labels
+                        # (role is masked out too because it makes no sense if contents is masked out)
+                        labels = [IGNORE_TOKEN_ID] * len(res["input_ids"])
+
+                elif empty_role:
+                    turn = content
+                    # this is only ever the first part, should include the bos token and the user query
+                    res = self._tokenize(
+                        turn, add_eos_token=False, strip_bos_token=False
+                    )
+                    if self.train_on_inputs and weight == 1:
+                        labels = copy.deepcopy(res["input_ids"])
+                    else:
+                        # everything from this is masked out from the labels
+                        labels = [IGNORE_TOKEN_ID] * len(res["input_ids"])
+
+                # pylint: disable=duplicate-code
+                result, current_len = parse_tokenized_to_result(
+                    result,
+                    current_len,
+                    res,
+                    labels,
+                    pad_token_id=self.tokenizer.pad_token_id,
+                )
+            return result
+        except (KeyError, AssertionError, IndexError) as err:
+            raise InvalidDataException(str(err)) from err
+
+
 def tokenize_prompt_default() -> Tuple[Dict[str, List[int]], int]:
     """
     Returns the default values for the tokenize prompt function

src/axolotl/prompters.py

@@ -5,6 +5,7 @@ from enum import Enum
 from typing import Generator, Optional, Union
 
 from colorama import Fore
+from fastchat.conversation import Conversation, get_conv_template
 
 LOG = logging.getLogger("axolotl")
 IGNORE_TOKEN_ID = -100
@@ -261,10 +262,166 @@ class ReflectAlpacaPrompter(Prompter):
         )
 
 
-ALTERNATING_ASSERTION_FAILED_ROLE = (
+SHAREGPT_ASSERTION_FAILED_ROLE = (
     "Role did not alternate between turns (gpt and human). Please check your data."
 )
 
+CONVERSATION_ROLE_FORMAT = {
+    "chatml": "<|im_start|>{ROLE}",
+    "zephyr": "<|{ROLE}|>",
+    "vicuna_v1.1": "{ROLE}",
+    "llama3": "<|start_header_id|>{ROLE}<|end_header_id|>",
+}
+
+
+class ShareGPTPrompter(Prompter):  # pylint: disable=too-few-public-methods
+    """
+    A prompter that generates prompts for the ShareGPT
+    """
+
+    role_key_human = "human"
+    role_key_model = "gpt"
+    # Optional, only used for tool usage datasets.
+    role_key_tool: Optional[str] = None
+    # Optional, role input/output mapping
+    roles: Optional[dict] = None
+
+    def __init__(
+        self,
+        prompt_style=None,  # pylint: disable=unused-argument
+        conversation: Optional[Union[str, Conversation]] = None,
+        role_key_human: Optional[str] = None,
+        role_key_model: Optional[str] = None,
+        role_key_tool: Optional[str] = None,
+        roles: Optional[dict] = None,
+    ):
+        if conversation:
+            if isinstance(conversation, Conversation):
+                self._conversation = conversation
+            else:
+                self._conversation = get_conv_template(conversation)
+        else:
+            self._conversation = get_conv_template("vicuna_v1.1")
+        if role_key_human:
+            self.role_key_human = role_key_human
+        if role_key_model:
+            self.role_key_model = role_key_model
+        if role_key_tool:
+            self.role_key_tool = role_key_tool
+        if roles:
+            self.roles = roles
+
+    def _build_result(self, source):
+        if len(source) < 2:
+            # If there isn't a back and forth conversation, ignore it
+            # also happens on the data splitting leaving empty conversations
+            raise IndexError(
+                f"A conversation entry has less than 2 messages :\n{source}"
+            )
+
+        conv = self._conversation.copy()
+
+        original_source = source.copy()
+        # Add the conversation system prompt if provided, otherwise use the default one
+        if source[0]["from"] == "system":
+            conv.set_system_message(source[0]["value"])
+            source.pop(0)
+
+        roles = {self.role_key_human: conv.roles[0], self.role_key_model: conv.roles[1]}
+        if self.role_key_tool:
+            roles[self.role_key_tool] = conv.roles[2]
+
+        try:
+            # Apply prompt templates
+            if source[0]["from"] not in roles:
+                # Skip the first one if it is not from human
+                source = source[1:]
+        except IndexError as err:
+            # sometimes there is a bing or system chat
+            raise err
+
+        conv.messages = []
+        for _, sentence in enumerate(source):
+            from_role = sentence["from"]
+            if from_role in roles:
+                role = roles[from_role]
+            else:
+                if self._conversation.name not in CONVERSATION_ROLE_FORMAT:
+                    raise NotImplementedError(
+                        f"Role ({role}) not in default roles, and {self._conversation.name} does not support role remapping yet."
+                        "Please help us by creating an Issue to add support for this conversation type."
+                    )
+
+                if self._conversation.name in ["llama3"]:
+                    role = from_role
+                else:
+                    role = CONVERSATION_ROLE_FORMAT[self._conversation.name].format(
+                        ROLE=from_role
+                    )
+
+            if len(conv.messages) > 0 and ((role == conv.messages[-1][0])):
+                if (
+                    role != "assistant"
+                ):  # back to back assistant calls may be okay for tool calls
+                    LOG.warning(f"{SHAREGPT_ASSERTION_FAILED_ROLE}: {sentence}")
+
+            conv.append_message(role, sentence["value"])
+        turns = list(conv.get_turns())
+        original_source_length = len(original_source)
+        assert len(turns) in [
+            original_source_length - 1,
+            original_source_length,
+            original_source_length + 1,
+        ]
+        if len(turns) == original_source_length + 1:
+            original_source = [{"weight": None}] + original_source
+        elif len(turns) == original_source_length - 1:
+            original_source = original_source[1:]
+        return [
+            (*turn, weight)
+            for turn, weight in zip(
+                turns,
+                [
+                    1 if "weight" not in e or e["weight"] is None else e["weight"]
+                    for e in original_source
+                ],
+            )
+        ]
+
+    def build_prompt(self, source) -> Generator[str, None, None]:
+        turns = self._build_result(source)
+
+        for part in turns:
+            if part[0] and not part[1]:
+                LOG.warning(f"role with empty message: {part[0]}")
+            yield part
+
+    def __repr__(self) -> str:
+        turns = self._build_result([{"from": "{from}", "value": "{value}"}])
+        return "\n".join([REPR_TEMPLATE.format(full_prompt=part) for part in turns])
+
+
+class ShareGPTPrompterV2(ShareGPTPrompter):
+    """
+    A V2 prompter that generates prompts for the ShareGPT
+    """
+
+    def __init__(
+        self,
+        conversation: Optional[Union[str, Conversation]] = None,
+        role_key_human: Optional[str] = None,
+        role_key_model: Optional[str] = None,
+        role_key_tool: Optional[str] = None,
+        roles: Optional[dict] = None,
+    ):
+        super().__init__(
+            conversation=conversation,
+            role_key_human=role_key_human,
+            role_key_model=role_key_model,
+            role_key_tool=role_key_tool,
+            roles=roles,
+        )
+
 
 class UnsupportedPrompter(Prompter):
     """

src/axolotl/train.py

@@ -260,10 +260,8 @@ def train(
 
     if not cfg.hub_model_id:
         try:
-            trainer.create_model_card(
-                model_name=cfg.output_dir.lstrip("./").encode("utf-8").decode("utf-8")
-            )
-        except (AttributeError, UnicodeDecodeError):
+            trainer.create_model_card(model_name=cfg.output_dir.lstrip("./"))
+        except AttributeError:
             pass
     elif cfg.hub_model_id:
         # defensively push to the hub to ensure the model card is updated

src/axolotl/utils/config/__init__.py

@@ -215,6 +215,11 @@ def normalize_cfg_datasets(cfg):
     if cfg.chat_template:
         if cfg.datasets:
             for idx, ds_cfg in enumerate(cfg.datasets):
+                if ds_cfg.type == "sharegpt" and not ds_cfg.conversation:
+                    LOG.info(
+                        f"updating dataset {ds_cfg.path} with `conversation: {cfg.chat_template}` to match your chat_template"
+                    )
+                    cfg.datasets[idx].conversation = cfg.chat_template
                 if (
                     ds_cfg.type in ["orpo.chat_template", "chat_template"]
                     and not ds_cfg.chat_template
@@ -223,7 +228,6 @@ def normalize_cfg_datasets(cfg):
                         f"updating dataset {ds_cfg.path} with `chat_template: {cfg.chat_template}` to match your chat_template"
                     )
                     cfg.datasets[idx].chat_template = cfg.chat_template
-                    cfg.datasets[idx].chat_template_jinja = cfg.chat_template_jinja
 
 
 def validate_config(cfg: DictDefault, capabilities: Optional[dict] = None):
@@ -456,6 +460,27 @@ def legacy_validate_config(cfg):
                 "`early_stopping_patience` requires that eval_steps should evenly divide save_steps."
             )
 
+    if cfg.datasets:
+        for idx, ds_cfg in enumerate(cfg.datasets):
+            if not ds_cfg.type:
+                continue
+            if ds_cfg.type == "sharegpt:chat":
+                LOG.warning(
+                    PendingDeprecationWarning(
+                        "`type: sharegpt:chat` will soon be deprecated. simply use `type: sharegpt` instead."
+                    )
+                )
+                cfg.datasets[idx].type = "sharegpt"
+            if "sharegpt_simple" in ds_cfg.type:
+                LOG.warning(
+                    PendingDeprecationWarning(
+                        "`type: sharegpt_simple` will soon be deprecated. simply use `type: sharegpt` instead."
+                    )
+                )
+                cfg.datasets[idx].type = cfg.datasets[idx].type.replace(
+                    "sharegpt_simple", "sharegpt"
+                )
+
     if cfg.saves_per_epoch and cfg.save_steps:
         raise ValueError(
             "save_steps and saves_per_epoch are mutually exclusive and cannot be used together."

src/axolotl/utils/config/models/input/v0_4_1/__init__.py

@@ -8,16 +8,9 @@ import logging
 import os
 from enum import Enum
 from importlib.metadata import version
-from typing import Annotated, Any, Dict, List, Literal, Optional, Tuple, Union
+from typing import Any, Dict, List, Literal, Optional, Tuple, Union
 
-from pydantic import (
-    BaseModel,
-    Field,
-    StringConstraints,
-    conlist,
-    field_validator,
-    model_validator,
-)
+from pydantic import BaseModel, Field, conlist, field_validator, model_validator
 from transformers import SchedulerType
 from transformers.training_args import OptimizerNames
 
@@ -28,39 +21,6 @@ LOG = logging.getLogger("axolotl.utils.config.models.input")
 SUPPORTED_METRICS = {"sacrebleu", "comet", "ter", "chrf", "perplexity"}
 
 
-class RLType(str, Enum):
-    """RL trainer type configuration subset"""
-
-    dpo = "dpo"  # pylint: disable=invalid-name
-    ipo = "ipo"  # pylint: disable=invalid-name
-    orpo = "orpo"  # pylint: disable=invalid-name
-    kto = "kto"  # pylint: disable=invalid-name
-    simpo = "simpo"  # pylint: disable=invalid-name
-
-
-class ChatTemplate(str, Enum):
-    """Chat templates configuration subset"""
-
-    alpaca = "alpaca"  # pylint: disable=invalid-name
-    chatml = "chatml"  # pylint: disable=invalid-name
-    mistral_v1 = "mistral_v1"  # pylint: disable=invalid-name
-    mistral_v2v3 = "mistral_v2v3"  # pylint: disable=invalid-name
-    mistral_v3_tekken = "mistral_v3_tekken"  # pylint: disable=invalid-name
-    gemma = "gemma"  # pylint: disable=invalid-name
-    cohere = "cohere"  # pylint: disable=invalid-name
-    llama3 = "llama3"  # pylint: disable=invalid-name
-    llama3_2_vision = "llama3_2_vision"  # pylint: disable=invalid-name
-    phi_3 = "phi_3"  # pylint: disable=invalid-name
-    phi_35 = "phi_35"  # pylint: disable=invalid-name
-    deepseek_v2 = "deepseek_v2"  # pylint: disable=invalid-name
-    jamba = "jamba"  # pylint: disable=invalid-name
-    jinja = "jinja"  # pylint: disable=invalid-name
-    qwen_25 = "qwen_25"  # pylint: disable=invalid-name
-    tokenizer_default = "tokenizer_default"  # pylint: disable=invalid-name
-    exaone = "exaone"  # pylint: disable=invalid-name
-    metharme = "metharme"  # pylint: disable=invalid-name
-
-
 class DeprecatedParameters(BaseModel):
     """configurations that are deprecated"""
 
@@ -145,19 +105,13 @@ class SFTDataset(BaseModel):
     input_transform: Optional[str] = None
     shards: Optional[int] = None
     conversation: Optional[str] = None
-    # Do not make this too strict or it will break the validator to choose different dataset class
-    chat_template: Optional[
-        Union[
-            ChatTemplate,
-            str,
-        ]
-    ] = None
-    chat_template_jinja: Optional[str] = None
+    chat_template: Optional[str] = None
     data_files: Optional[Union[str, List[str]]] = None
     input_format: Optional[str] = None
     name: Optional[str] = None
     ds_type: Optional[str] = None
     train_on_split: Optional[str] = None
+
     field: Optional[str] = None
     field_human: Optional[str] = None
     field_model: Optional[str] = None
@@ -168,32 +122,13 @@ class SFTDataset(BaseModel):
     message_field_training_detail: Optional[str] = None
     roles_to_train: Optional[List[str]] = None
     train_on_eos: Optional[str] = None
+
     roles: Optional[Dict[str, List[str]]] = None
     drop_system_message: Optional[bool] = None
+
     trust_remote_code: Optional[bool] = False
     revision: Optional[str] = None
 
-    @model_validator(mode="before")
-    @classmethod
-    def check_chat_template_config(cls, data):
-        # Set chat_template to tokenizer_default if not set
-        if data.get("type") == "chat_template" and not data.get("chat_template"):
-            data["chat_template"] = ChatTemplate.tokenizer_default
-
-        # if chat_template is set to jinja, chat_template_jinja is required
-        if data.get("chat_template") == ChatTemplate.jinja and not data.get(
-            "chat_template_jinja"
-        ):
-            raise ValueError(
-                "chat_template_jinja is required when chat_template is set to jinja"
-            )
-
-        # If chat_template_jinja is set, set chat_template to jinja
-        if data.get("chat_template_jinja") and not data.get("chat_template"):
-            data["chat_template"] = ChatTemplate.jinja
-
-        return data
-
 
 class UserDefinedDPOType(BaseModel):
     """User defined typing for DPO"""
@@ -239,6 +174,35 @@ class KTODataset(BaseModel):
     revision: Optional[str] = None
 
 
+class RLType(str, Enum):
+    """RL trainer type configuration subset"""
+
+    dpo = "dpo"  # pylint: disable=invalid-name
+    ipo = "ipo"  # pylint: disable=invalid-name
+    orpo = "orpo"  # pylint: disable=invalid-name
+    kto = "kto"  # pylint: disable=invalid-name
+    simpo = "simpo"  # pylint: disable=invalid-name
+
+
+class ChatTemplate(str, Enum):
+    """Chat templates configuration subset"""
+
+    alpaca = "alpaca"  # pylint: disable=invalid-name
+    chatml = "chatml"  # pylint: disable=invalid-name
+    mistral_v1 = "mistral_v1"  # pylint: disable=invalid-name
+    mistral_v2v3 = "mistral_v2v3"  # pylint: disable=invalid-name
+    mistral_v3_tekken = "mistral_v3_tekken"  # pylint: disable=invalid-name
+    gemma = "gemma"  # pylint: disable=invalid-name
+    cohere = "cohere"  # pylint: disable=invalid-name
+    llama3 = "llama3"  # pylint: disable=invalid-name
+    llama3_2_vision = "llama3_2_vision"  # pylint: disable=invalid-name
+    phi_3 = "phi_3"  # pylint: disable=invalid-name
+    phi_35 = "phi_35"  # pylint: disable=invalid-name
+    deepseek_v2 = "deepseek_v2"  # pylint: disable=invalid-name
+    jamba = "jamba"  # pylint: disable=invalid-name
+    qwen_25 = "qwen_25"  # pylint: disable=invalid-name
+
+
 class LoftQConfig(BaseModel):
     """LoftQ configuration subset"""
 
@@ -428,7 +392,6 @@ class HyperparametersConfig(BaseModel):
                 "ao_adamw_4bit",
                 "ao_adamw_8bit",
                 "ao_adamw_fp8",
-                "adopt_adamw",
             ],
         ]
     ] = OptimizerNames.ADAMW_HF.value
@@ -586,13 +549,9 @@ class AxolotlInputConfig(
     resume_from_checkpoint: Optional[str] = None
     auto_resume_from_checkpoints: Optional[bool] = None
     resize_token_embeddings_to_32x: Optional[bool] = None
-    mean_resizing_embeddings: Optional[bool] = False
 
     rl: Optional[RLType] = None
     reward_model: Optional[bool] = None
-    dpo_use_weighting: Optional[
-        bool
-    ] = None  # whether to use weighting in DPO trainer. If none, default is false in the trainer.
 
     datasets: Optional[conlist(Union[SFTDataset, DPODataset, KTODataset], min_length=1)] = None  # type: ignore
     test_datasets: Optional[conlist(Union[SFTDataset, DPODataset, KTODataset], min_length=1)] = None  # type: ignore
@@ -759,13 +718,7 @@ class AxolotlInputConfig(
     gpu_memory_limit: Optional[Union[int, str]] = None
     low_cpu_mem_usage: Optional[bool] = None
 
-    chat_template: Optional[
-        Union[
-            ChatTemplate,
-            Annotated[str, StringConstraints(pattern="^tokenizer_default_fallback_")],
-        ]
-    ] = None
-    chat_template_jinja: Optional[str] = None
+    chat_template: Optional[ChatTemplate] = None
     default_system_message: Optional[str] = None
 
     fix_untrained_tokens: Optional[bool] = None
@@ -783,25 +736,28 @@ class AxolotlInputConfig(
     is_mistral_derived_model: Optional[bool] = Field(default=None)
     is_qwen_derived_model: Optional[bool] = Field(default=None)
 
-    plugins: Optional[List[str]] = Field(default=None)
-
     @field_validator("datasets", mode="before")
     @classmethod
-    def deprecate_sharegpt_datasets(cls, datasets):
-        for _, ds_cfg in enumerate(datasets):
-            if not ds_cfg.get("type"):
+    def fix_sharegpt_datasets(cls, datasets):
+        for idx, ds_cfg in enumerate(datasets):
+            if not ds_cfg["type"]:
                 continue
-
-            ds_type = ds_cfg["type"]
-            # skip if it's a dict (for custom user instruction prompt)
-            if isinstance(ds_type, dict):
-                continue
-
-            if isinstance(ds_type, str) and ds_type.startswith("sharegpt"):
-                raise ValueError(
-                    "`type: sharegpt.*` is deprecated. Please use `type: chat_template` instead."
+            if ds_cfg["type"] == "sharegpt:chat":
+                LOG.warning(
+                    PendingDeprecationWarning(
+                        "`type: sharegpt:chat` will soon be deprecated. simply use `type: sharegpt` instead."
+                    )
+                )
+                datasets[idx]["type"] = "sharegpt"
+            if "sharegpt_simple" in ds_cfg["type"]:
+                LOG.warning(
+                    PendingDeprecationWarning(
+                        "`type: sharegpt_simple` will soon be deprecated. simply use `type: sharegpt` instead."
+                    )
+                )
+                datasets[idx]["type"] = datasets[idx]["type"].replace(
+                    "sharegpt_simple", "sharegpt"
                 )
-
         return datasets
 
     @model_validator(mode="before")
@@ -871,23 +827,6 @@ class AxolotlInputConfig(
 
         return data
 
-    @model_validator(mode="before")
-    @classmethod
-    def check_chat_template_config(cls, data):
-        # if chat_template is set to jinja, chat_template_jinja is required
-        if data.get("chat_template") == ChatTemplate.jinja and not data.get(
-            "chat_template_jinja"
-        ):
-            raise ValueError(
-                "chat_template_jinja is required when chat_template is set to jinja"
-            )
-
-        # If chat_template_jinja is set, set chat_template to jinja
-        if data.get("chat_template_jinja") and not data.get("chat_template"):
-            data["chat_template"] = ChatTemplate.jinja
-
-        return data
-
     @model_validator(mode="before")
     @classmethod
     def check_sample_packing_wo_flash(cls, data):

src/axolotl/utils/data/sft.py

@@ -2,11 +2,9 @@
 
 import functools
 import logging
-import time
 from pathlib import Path
 from typing import List, Optional, Tuple, Union
 
-import requests
 from datasets import (
     Dataset,
     DatasetDict,
@@ -55,28 +53,6 @@ from axolotl.utils.trainer import (
 LOG = logging.getLogger("axolotl")
 
 
-def retry_on_request_exceptions(max_retries=3, delay=1):
-    def decorator(func):
-        @functools.wraps(func)
-        def wrapper(*args, **kwargs):  # pylint: disable=inconsistent-return-statements
-            for attempt in range(max_retries):
-                try:
-                    return func(*args, **kwargs)
-                except (
-                    requests.exceptions.ReadTimeout,
-                    requests.exceptions.ConnectionError,
-                ) as exc:
-                    if attempt < max_retries - 1:
-                        time.sleep(delay)
-                    else:
-                        raise exc
-
-        return wrapper
-
-    return decorator
-
-
-@retry_on_request_exceptions(max_retries=3, delay=5)
 def prepare_dataset(cfg, tokenizer, processor=None):
     prompters = []
     if not cfg.pretraining_dataset:

src/axolotl/utils/environment.py

@@ -1,25 +0,0 @@
-"""
-utils to get GPU info for the current environment
-"""
-from accelerate.utils.environment import (
-    check_cuda_p2p_ib_support as accelerate_check_cuda_p2p_ib_support,
-)
-from accelerate.utils.environment import get_gpu_info
-
-
-def check_cuda_p2p_ib_support():
-    if not accelerate_check_cuda_p2p_ib_support():
-        return False
-    unsupported_devices = {"RTX 6000 Ada"}
-    try:
-        device_names, device_count = get_gpu_info()
-        if 1 < device_count < 8:
-            if any(
-                unsupported_device in device_name
-                for device_name in device_names
-                for unsupported_device in unsupported_devices
-            ):
-                return False
-    except Exception:  # pylint: disable=broad-except # nosec
-        pass
-    return True

src/axolotl/utils/gradient_checkpointing/unsloth.py

@@ -14,16 +14,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 import torch
-from packaging import version
-
-torch_version = version.parse(torch.__version__)
-
-if torch_version < version.parse("2.4.0"):
-    torch_cuda_amp_custom_fwd = torch.cuda.amp.custom_fwd
-    torch_cuda_amp_custom_bwd = torch.cuda.amp.custom_bwd
-else:
-    torch_cuda_amp_custom_fwd = torch.amp.custom_fwd(device_type="cuda")
-    torch_cuda_amp_custom_bwd = torch.amp.custom_bwd(device_type="cuda")
 
 
 class Unsloth_Offloaded_Gradient_Checkpointer(  # pylint: disable=invalid-name
@@ -35,7 +25,7 @@ class Unsloth_Offloaded_Gradient_Checkpointer(  # pylint: disable=invalid-name
     """
 
     @staticmethod
-    @torch_cuda_amp_custom_fwd
+    @torch.cuda.amp.custom_fwd
     def forward(ctx, forward_function, hidden_states, *args):
         saved_hidden_states = hidden_states.to("cpu", non_blocking=True)
         with torch.no_grad():
@@ -46,7 +36,7 @@ class Unsloth_Offloaded_Gradient_Checkpointer(  # pylint: disable=invalid-name
         return output
 
     @staticmethod
-    @torch_cuda_amp_custom_bwd
+    @torch.cuda.amp.custom_bwd
     def backward(ctx, dY):
         (hidden_states,) = ctx.saved_tensors
         hidden_states = hidden_states.to("cuda", non_blocking=True).detach()

src/axolotl/utils/mlflow_.py

@@ -16,7 +16,3 @@ def setup_mlflow_env_vars(cfg: DictDefault):
     # Enable mlflow if experiment name is present
     if cfg.mlflow_experiment_name and len(cfg.mlflow_experiment_name) > 0:
         cfg.use_mlflow = True
-
-    # Enable logging hf artifacts in mlflow if value is truthy
-    if cfg.hf_mlflow_log_artifacts is True:
-        os.environ["HF_MLFLOW_LOG_ARTIFACTS"] = "true"

src/axolotl/utils/optimizers/adopt.py

@@ -1,508 +0,0 @@
-"""
-Copied from https://github.com/iShohei220/adopt
-
-ADOPT: Modified Adam Can Converge with Any β2 with the Optimal Rate (2024)
-Taniguchi, Shohei and Harada, Keno and Minegishi, Gouki and Oshima, Yuta and Jeong, Seong Cheol and Nagahara, Go and Iiyama, Tomoshi and Suzuki, Masahiro and Iwasawa, Yusuke and Matsuo, Yutaka
-"""
-# mypy: ignore-errors
-# pylint: skip-file
-# mypy: allow-untyped-decorators
-# mypy: allow-untyped-defs
-from typing import List, Optional, Tuple, Union, cast
-
-import torch
-from torch import Tensor
-from torch.optim.optimizer import (
-    Optimizer,
-    ParamsT,
-    _default_to_fused_or_foreach,
-    _device_dtype_check_for_fused,
-    _disable_dynamo_if_unsupported,
-    _get_capturable_supported_devices,
-    _get_scalar_dtype,
-    _get_value,
-    _use_grad_for_differentiable,
-    _view_as_real,
-)
-
-__all__ = ["ADOPT", "adopt"]
-
-
-class ADOPT(Optimizer):
-    def __init__(
-        self,
-        params: ParamsT,
-        lr: Union[float, Tensor] = 1e-3,
-        betas: Tuple[float, float] = (0.9, 0.9999),
-        eps: float = 1e-6,
-        weight_decay: float = 0.0,
-        decoupled: bool = False,
-        *,
-        foreach: Optional[bool] = None,
-        maximize: bool = False,
-        capturable: bool = False,
-        differentiable: bool = False,
-        fused: Optional[bool] = None,
-    ):
-        if isinstance(lr, Tensor):
-            if foreach and not capturable:
-                raise ValueError(
-                    "lr as a Tensor is not supported for capturable=False and foreach=True"
-                )
-            if lr.numel() != 1:
-                raise ValueError("Tensor lr must be 1-element")
-        if not 0.0 <= lr:
-            raise ValueError(f"Invalid learning rate: {lr}")
-        if not 0.0 <= eps:
-            raise ValueError(f"Invalid epsilon value: {eps}")
-        if not 0.0 <= betas[0] < 1.0:
-            raise ValueError(f"Invalid beta parameter at index 0: {betas[0]}")
-        if not 0.0 <= betas[1] < 1.0:
-            raise ValueError(f"Invalid beta parameter at index 1: {betas[1]}")
-        if not 0.0 <= weight_decay:
-            raise ValueError(f"Invalid weight_decay value: {weight_decay}")
-
-        defaults = dict(
-            lr=lr,
-            betas=betas,
-            eps=eps,
-            weight_decay=weight_decay,
-            decoupled=decoupled,
-            maximize=maximize,
-            foreach=foreach,
-            capturable=capturable,
-            differentiable=differentiable,
-            fused=fused,
-        )
-        super().__init__(params, defaults)
-
-        if fused:
-            # TODO: support fused
-            raise RuntimeError("`fused` is not currently supported")
-
-            if differentiable:
-                raise RuntimeError("`fused` does not support `differentiable`")
-            self._step_supports_amp_scaling = True
-            # TODO(crcrpar): [low prec params & their higher prec copy]
-            # Support AMP with FP16/BF16 model params which would need
-            # higher prec copy of params to do update math in higher prec to
-            # alleviate the loss of information.
-            if foreach:
-                raise RuntimeError("`fused` and `foreach` cannot be `True` together.")
-
-    def __setstate__(self, state):
-        super().__setstate__(state)
-        for group in self.param_groups:
-            group.setdefault("maximize", False)
-            group.setdefault("foreach", None)
-            group.setdefault("capturable", False)
-            group.setdefault("differentiable", False)
-            fused = group.setdefault("fused", None)
-            for p in group["params"]:
-                p_state = self.state.get(p, [])
-                if len(p_state) != 0 and not torch.is_tensor(p_state["step"]):
-                    step_val = float(p_state["step"])
-                    p_state["step"] = (
-                        torch.tensor(
-                            step_val,
-                            dtype=_get_scalar_dtype(is_fused=fused),
-                            device=p.device,
-                        )
-                        if group["capturable"] or group["fused"]
-                        else torch.tensor(step_val, dtype=_get_scalar_dtype())
-                    )
-
-    def _init_group(
-        self,
-        group,
-        params_with_grad,
-        grads,
-        exp_avgs,
-        exp_avg_sqs,
-        state_steps,
-    ):
-        has_complex = False
-        for p in group["params"]:
-            if p.grad is not None:
-                has_complex |= torch.is_complex(p)
-                params_with_grad.append(p)
-                if p.grad.is_sparse:
-                    raise RuntimeError("ADOPT does not support sparse gradients")
-                grads.append(p.grad)
-
-                state = self.state[p]
-                # Lazy state initialization
-                if len(state) == 0:
-                    if group["fused"]:
-                        _device_dtype_check_for_fused(p)
-                    # note(crcrpar): [special device hosting for step]
-                    # Deliberately host `step` on CPU if both capturable and fused are off.
-                    # This is because kernel launches are costly on CUDA and XLA.
-                    state["step"] = (
-                        torch.zeros(
-                            (),
-                            dtype=_get_scalar_dtype(is_fused=group["fused"]),
-                            device=p.device,
-                        )
-                        if group["capturable"] or group["fused"]
-                        else torch.tensor(0.0, dtype=_get_scalar_dtype())
-                    )
-                    # Exponential moving average of gradient values
-                    state["exp_avg"] = torch.zeros_like(
-                        p, memory_format=torch.preserve_format
-                    )
-                    # Exponential moving average of squared gradient values
-                    state["exp_avg_sq"] = torch.zeros_like(
-                        p, memory_format=torch.preserve_format
-                    )
-
-                exp_avgs.append(state["exp_avg"])
-                exp_avg_sqs.append(state["exp_avg_sq"])
-
-                if group["differentiable"] and state["step"].requires_grad:
-                    raise RuntimeError(
-                        "`requires_grad` is not supported for `step` in differentiable mode"
-                    )
-
-                # Foreach without capturable does not support a tensor lr
-                if (
-                    group["foreach"]
-                    and torch.is_tensor(group["lr"])
-                    and not group["capturable"]
-                ):
-                    raise RuntimeError(
-                        "lr as a Tensor is not supported for capturable=False and foreach=True"
-                    )
-
-                state_steps.append(state["step"])
-        return has_complex
-
-    @_use_grad_for_differentiable
-    def step(self, closure=None):
-        """Perform a single optimization step.
-
-        Args:
-            closure (Callable, optional): A closure that reevaluates the model
-                and returns the loss.
-        """
-        self._cuda_graph_capture_health_check()
-
-        loss = None
-        if closure is not None:
-            with torch.enable_grad():
-                loss = closure()
-
-        for group in self.param_groups:
-            params_with_grad: List[Tensor] = []
-            grads: List[Tensor] = []
-            exp_avgs: List[Tensor] = []
-            exp_avg_sqs: List[Tensor] = []
-            state_steps: List[Tensor] = []
-            beta1, beta2 = group["betas"]
-
-            has_complex = self._init_group(
-                group,
-                params_with_grad,
-                grads,
-                exp_avgs,
-                exp_avg_sqs,
-                state_steps,
-            )
-
-            adopt(
-                params_with_grad,
-                grads,
-                exp_avgs,
-                exp_avg_sqs,
-                state_steps,
-                has_complex=has_complex,
-                beta1=beta1,
-                beta2=beta2,
-                lr=group["lr"],
-                weight_decay=group["weight_decay"],
-                decoupled=group["decoupled"],
-                eps=group["eps"],
-                maximize=group["maximize"],
-                foreach=group["foreach"],
-                capturable=group["capturable"],
-                differentiable=group["differentiable"],
-                fused=group["fused"],
-                grad_scale=getattr(self, "grad_scale", None),
-                found_inf=getattr(self, "found_inf", None),
-            )
-
-        return loss
-
-
-def _single_tensor_adopt(
-    params: List[Tensor],
-    grads: List[Tensor],
-    exp_avgs: List[Tensor],
-    exp_avg_sqs: List[Tensor],
-    state_steps: List[Tensor],
-    grad_scale: Optional[Tensor],
-    found_inf: Optional[Tensor],
-    *,
-    has_complex: bool,
-    beta1: float,
-    beta2: float,
-    lr: Union[float, Tensor],
-    weight_decay: float,
-    decoupled: bool,
-    eps: float,
-    maximize: bool,
-    capturable: bool,
-    differentiable: bool,
-):
-    assert grad_scale is None and found_inf is None
-
-    if torch.jit.is_scripting():
-        # this assert is due to JIT being dumb and not realizing that the ops below
-        # have overloads to handle both float and Tensor lrs, so we just assert it's
-        # a float since most people using JIT are using floats
-        assert isinstance(lr, float)
-
-    for i, param in enumerate(params):
-        grad = grads[i] if not maximize else -grads[i]
-        exp_avg = exp_avgs[i]
-        exp_avg_sq = exp_avg_sqs[i]
-        step_t = state_steps[i]
-
-        # If compiling, the compiler will handle cudagraph checks, see note [torch.compile x capturable]
-        if not torch._utils.is_compiling() and capturable:
-            capturable_supported_devices = _get_capturable_supported_devices()
-            assert (
-                param.device.type == step_t.device.type
-                and param.device.type in capturable_supported_devices
-            ), f"If capturable=True, params and state_steps must be on supported devices: {capturable_supported_devices}."
-
-        # update step
-        step_t += 1
-
-        if weight_decay != 0:
-            if decoupled:
-                param.add_(param, alpha=-lr * weight_decay)
-            else:
-                grad = grad.add(param, alpha=weight_decay)
-
-        if torch.is_complex(param):
-            grad = torch.view_as_real(grad)
-            if exp_avg is not None:
-                exp_avg = torch.view_as_real(exp_avg)
-            if exp_avg_sq is not None:
-                exp_avg_sq = torch.view_as_real(exp_avg_sq)
-            param = torch.view_as_real(param)
-
-        step = step_t if capturable or differentiable else _get_value(step_t)
-        if step == 1:
-            exp_avg_sq.addcmul_(grad, grad.conj())
-            continue
-
-        denom = torch.clamp(exp_avg_sq.sqrt(), eps)
-        if step == 2:
-            exp_avg.addcdiv_(grad, denom)
-        else:
-            exp_avg.mul_(beta1).addcdiv_(grad, denom, value=1 - beta1)
-
-        param.add_(exp_avg, alpha=-lr)
-        exp_avg_sq.mul_(beta2).addcmul_(grad, grad.conj(), value=1 - beta2)
-
-
-def _multi_tensor_adopt(
-    params: List[Tensor],
-    grads: List[Tensor],
-    exp_avgs: List[Tensor],
-    exp_avg_sqs: List[Tensor],
-    state_steps: List[Tensor],
-    grad_scale: Optional[Tensor],
-    found_inf: Optional[Tensor],
-    *,
-    has_complex: bool,
-    beta1: float,
-    beta2: float,
-    lr: Union[float, Tensor],
-    weight_decay: float,
-    decoupled: bool,
-    eps: float,
-    maximize: bool,
-    capturable: bool,
-    differentiable: bool,
-):
-    if len(params) == 0:
-        return
-
-    if isinstance(lr, Tensor) and not capturable:
-        raise RuntimeError(
-            "lr as a Tensor is not supported for capturable=False and foreach=True"
-        )
-
-    # If compiling, the compiler will handle cudagraph checks, see note [torch.compile x capturable]
-    if not torch._utils.is_compiling() and capturable:
-        capturable_supported_devices = _get_capturable_supported_devices(
-            supports_xla=False
-        )
-        assert all(
-            p.device.type == step.device.type
-            and p.device.type in capturable_supported_devices
-            for p, step in zip(params, state_steps)
-        ), f"If capturable=True, params and state_steps must be on supported devices: {capturable_supported_devices}."
-
-    assert grad_scale is None and found_inf is None
-
-    assert not differentiable, "_foreach ops don't support autograd"
-
-    grouped_tensors = Optimizer._group_tensors_by_device_and_dtype(
-        [params, grads, exp_avgs, exp_avg_sqs, state_steps]  # type: ignore[list-item]
-    )
-    for (
-        device_params_,
-        device_grads_,
-        device_exp_avgs_,
-        device_exp_avg_sqs_,
-        device_state_steps_,
-    ), _ in grouped_tensors.values():
-        device_params = cast(List[Tensor], device_params_)
-        device_grads = cast(List[Tensor], device_grads_)
-        device_exp_avgs = cast(List[Tensor], device_exp_avgs_)
-        device_exp_avg_sqs = cast(List[Tensor], device_exp_avg_sqs_)
-        device_state_steps = cast(List[Tensor], device_state_steps_)
-
-        # Handle complex parameters
-        if has_complex:
-            _view_as_real(
-                device_params, device_grads, device_exp_avgs, device_exp_avg_sqs
-            )
-
-        if maximize:
-            device_grads = torch._foreach_neg(device_grads)  # type: ignore[assignment]
-
-        # Update steps
-        # If steps are on CPU, foreach will fall back to the slow path, which is a for-loop calling t.add(1) over
-        # and over. 1 will then be wrapped into a Tensor over and over again, which is slower than if we just
-        # wrapped it once now. The alpha is required to assure we go to the right overload.
-        if not torch._utils.is_compiling() and device_state_steps[0].is_cpu:
-            torch._foreach_add_(
-                device_state_steps, torch.tensor(1.0, device="cpu"), alpha=1.0
-            )
-        else:
-            torch._foreach_add_(device_state_steps, 1)
-
-        if weight_decay != 0:
-            if decoupled:
-                torch._foreach_add_(
-                    device_params, device_params, alpha=-lr * weight_decay
-                )
-            else:
-                # Re-use the intermediate memory (device_grads) already allocated for maximize
-                if maximize:
-                    torch._foreach_add_(device_grads, device_params, alpha=weight_decay)
-                else:
-                    device_grads = torch._foreach_add(  # type: ignore[assignment]
-                        device_grads, device_params, alpha=weight_decay
-                    )
-
-        if device_state_steps[0] == 1:
-            torch._foreach_addcmul_(device_exp_avg_sqs, device_grads, device_grads)
-            continue
-
-        exp_avg_sq_sqrt = torch._foreach_sqrt(device_exp_avg_sqs)
-        exp_avg_sq_sqrt = torch._foreach_maximum(exp_avg_sq_sqrt, eps)
-
-        if device_state_steps[0] == 2:
-            torch._foreach_addcdiv_(device_exp_avgs, device_grads, exp_avg_sq_sqrt)
-        else:
-            torch._foreach_mul_(device_exp_avgs, beta1)
-            torch._foreach_addcdiv_(
-                device_exp_avgs, device_grads, exp_avg_sq_sqrt, value=1 - beta1
-            )
-
-        torch._foreach_add_(device_params, device_exp_avgs, alpha=-lr)
-        torch._foreach_mul_(device_exp_avg_sqs, beta2)
-        torch._foreach_addcmul_(
-            device_exp_avg_sqs, device_grads, device_grads, value=1 - beta2
-        )
-
-
-@_disable_dynamo_if_unsupported(single_tensor_fn=_single_tensor_adopt)
-def adopt(
-    params: List[Tensor],
-    grads: List[Tensor],
-    exp_avgs: List[Tensor],
-    exp_avg_sqs: List[Tensor],
-    state_steps: List[Tensor],
-    # kwonly args with defaults are not supported by functions compiled with torchscript issue #70627
-    # setting this as kwarg for now as functional API is compiled by torch/distributed/optim
-    foreach: Optional[bool] = None,
-    capturable: bool = False,
-    differentiable: bool = False,
-    fused: Optional[bool] = None,
-    grad_scale: Optional[Tensor] = None,
-    found_inf: Optional[Tensor] = None,
-    has_complex: bool = False,
-    *,
-    beta1: float,
-    beta2: float,
-    lr: Union[float, Tensor],
-    weight_decay: float,
-    decoupled: bool,
-    eps: float,
-    maximize: bool,
-):
-    r"""Functional API that performs ADOPT algorithm computation."""
-    # Respect when the user inputs False/True for foreach or fused. We only want to change
-    # the default when neither have been user-specified. Note that we default to foreach
-    # and pass False to use_fused. This is not a mistake--we want to give the fused impl
-    # bake-in time before making it the default, even if it is typically faster.
-    if fused is None and foreach is None:
-        _, foreach = _default_to_fused_or_foreach(
-            params, differentiable, use_fused=False
-        )
-        # Do not flip on foreach for the unsupported case where lr is a Tensor and capturable=False.
-        if foreach and isinstance(lr, Tensor) and not capturable:
-            foreach = False
-    if fused is None:
-        fused = False
-    if foreach is None:
-        foreach = False
-
-    # this check is slow during compilation, so we skip it
-    # if it's strictly needed we can add this check back in dynamo
-    if not torch._utils.is_compiling() and not all(
-        isinstance(t, torch.Tensor) for t in state_steps
-    ):
-        raise RuntimeError(
-            "API has changed, `state_steps` argument must contain a list of singleton tensors"
-        )
-
-    if foreach and torch.jit.is_scripting():
-        raise RuntimeError("torch.jit.script not supported with foreach optimizers")
-    if fused and torch.jit.is_scripting():
-        raise RuntimeError("torch.jit.script not supported with fused optimizers")
-
-    # if fused and not torch.jit.is_scripting():
-    #     func = _fused_adopt
-    # elif foreach and not torch.jit.is_scripting():
-    if foreach and not torch.jit.is_scripting():
-        func = _multi_tensor_adopt
-    else:
-        func = _single_tensor_adopt
-
-    func(
-        params,
-        grads,
-        exp_avgs,
-        exp_avg_sqs,
-        state_steps,
-        has_complex=has_complex,
-        beta1=beta1,
-        beta2=beta2,
-        lr=lr,
-        weight_decay=weight_decay,
-        decoupled=decoupled,
-        eps=eps,
-        maximize=maximize,
-        capturable=capturable,
-        differentiable=differentiable,
-        grad_scale=grad_scale,
-        found_inf=found_inf,
-    )

src/axolotl/utils/samplers/multipack.py

@@ -133,8 +133,6 @@ class MultipackBatchSampler(BatchSampler):
         self.eff_total_used = 0
         self.eff_total_slots = 0
 
-        self.len_across_ranks = None
-
     def set_epoch(self, epoch: int):
         self.epoch = epoch
 
@@ -197,14 +195,15 @@ class MultipackBatchSampler(BatchSampler):
             LOG.info(f"gather_len_batches: {repr(estimates)}")
             return math.floor(0.998 * min(estimates))
 
-        min_len_batches = reduce_and_broadcast(lambda: num, calc_min_len)
+        min_len_batches = reduce_and_broadcast(
+            lambda: num,
+            calc_min_len,
+        )
         return min_len_batches
 
     def __len__(self):
-        if not self.len_across_ranks:
-            len_batches = self.num_batches()
-            self.len_across_ranks = self.gather_len_batches(len_batches)
-        return self.len_across_ranks
+        len_batches = self.num_batches()
+        return self.gather_len_batches(len_batches)
 
     def _len_est(self):
         efficiency = (

src/axolotl/utils/tokenization.py

@@ -1,6 +1,8 @@
 """Module for tokenization utilities"""
 
 import logging
+import re
+from typing import Dict, List
 
 from termcolor import colored
 
@@ -91,3 +93,65 @@ def check_rl_example_labels(example, tokenizer, text_only=False):
     LOG.info(f"REJECTED RESPONSE: {delimiter.join(colored_rejecteds)}\n\n\n")
 
     return delimiter.join(colored_tokens)
+
+
+GLAIVE_ROLES = ["USER", "ASSISTANT", "FUNCTION RESPONSE"]
+GLAIVE_TO_SHAREGPT_ROLE = {
+    "SYSTEM": "system",
+    "USER": "human",
+    "ASSISTANT": "gpt",
+    "FUNCTION RESPONSE": "tool",
+}
+
+GLAIVE_MSG_REGEX = re.compile(rf"({'|'.join(GLAIVE_ROLES)}): ")
+
+
+def chatml_to_conversation(row: Dict[str, str]) -> List[Dict[str, str]]:
+    """
+    Converts a ChatML formatted row to a list of messages in ShareGPT format.
+    Initially based off https://github.com/lilacai/lilac/blob/main/notebooks/GlaiveToShareGPT.ipynb.
+    """
+
+    system_prompt = row.get("system")
+    if system_prompt:
+        system_prompt = system_prompt.removeprefix("SYSTEM: ")
+
+    chat_str = row["chat"]
+    chat_msgs = [s.strip() for s in GLAIVE_MSG_REGEX.split(chat_str) if s]
+
+    chat_msg_dicts = [
+        {"from": GLAIVE_TO_SHAREGPT_ROLE[role], "value": value}
+        for role, value in zip(chat_msgs[::2], chat_msgs[1::2])
+    ]
+
+    if system_prompt:
+        chat_msg_dicts = [
+            {"from": GLAIVE_TO_SHAREGPT_ROLE["SYSTEM"], "value": system_prompt}
+        ] + chat_msg_dicts
+
+    return chat_msg_dicts
+
+
+def merge_consecutive_messages(messages):
+    """
+    Merge consecutive messages from the same sender into a single message.
+    This can be useful with datasets that contain multiple consecutive tool calls.
+    """
+
+    merged_messages = []
+    current_from = None
+    current_message = ""
+
+    for msg in messages:
+        if current_from == msg["from"]:
+            current_message += msg["value"]
+        else:
+            if current_from is not None:
+                merged_messages.append({"from": current_from, "value": current_message})
+            current_from = msg["from"]
+            current_message = msg["value"]
+
+    if current_from is not None:
+        merged_messages.append({"from": current_from, "value": current_message})
+
+    return merged_messages

src/axolotl/utils/trainer.py

@@ -17,7 +17,6 @@ from transformers.utils import is_torch_bf16_gpu_available
 
 from axolotl.core.trainer_builder import HFCausalTrainerBuilder, HFRLTrainerBuilder
 from axolotl.utils.distributed import reduce_and_broadcast
-from axolotl.utils.environment import check_cuda_p2p_ib_support
 from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
 
 LOG = get_logger("axolotl")
@@ -185,10 +184,11 @@ def process_datasets_for_packing(cfg, train_dataset, eval_dataset):
         min_sequence_len=cfg.min_sample_len or 2,
     )
 
-    min_input_len = np.min(get_dataset_lengths(train_dataset))
-    LOG.debug(f"min_input_len: {min_input_len}", main_process_only=True)
-    max_input_len = np.max(get_dataset_lengths(train_dataset))
-    LOG.debug(f"max_input_len: {max_input_len}", main_process_only=True)
+    if cfg.is_preprocess:
+        min_input_len = np.min(get_dataset_lengths(train_dataset))
+        LOG.debug(f"min_input_len: {min_input_len}", main_process_only=True)
+        max_input_len = np.max(get_dataset_lengths(train_dataset))
+        LOG.debug(f"max_input_len: {max_input_len}", main_process_only=True)
 
     if cfg.model_config_type == "mamba":
         LOG.info("dropping attention_mask column")
@@ -461,9 +461,6 @@ def setup_fsdp_envs(cfg):
 
 
 def prepare_optim_env(cfg):
-    if not check_cuda_p2p_ib_support():
-        if os.getenv("NCCL_P2P_DISABLE") is None:
-            os.environ["NCCL_P2P_DISABLE"] = "1"
     if cfg.fsdp:
         setup_fsdp_envs(cfg)
     elif cfg.deepspeed:

tests/e2e/integrations/liger.py

@@ -1,6 +1,7 @@
 """
 Simple end-to-end test for Liger integration
 """
+
 import unittest
 from pathlib import Path
 

tests/e2e/multigpu/test_eval.py

@@ -1,155 +0,0 @@
-"""
-E2E tests for multigpu eval
-"""
-import logging
-import os
-import unittest
-from pathlib import Path
-
-import yaml
-from accelerate.test_utils import execute_subprocess_async
-
-from axolotl.utils.dict import DictDefault
-
-from ..utils import with_temp_dir
-
-LOG = logging.getLogger("axolotl.tests.e2e.multigpu")
-os.environ["WANDB_DISABLED"] = "true"
-
-AXOLOTL_ROOT = Path(__file__).parent.parent.parent.parent
-
-
-class TestMultiGPUEval(unittest.TestCase):
-    """
-    Test case for MultiGPU Eval Sample Packing
-    """
-
-    @with_temp_dir
-    def test_eval_sample_packing(self, temp_dir):
-        # pylint: disable=duplicate-code
-        cfg = DictDefault(
-            {
-                "base_model": "JackFram/llama-68m",
-                "load_in_8bit": False,
-                "load_in_4bit": True,
-                "strict": False,
-                "sequence_len": 2048,
-                "adapter": "qlora",
-                "sample_packing": True,
-                "eval_sample_packing": True,
-                "pad_to_sequence_len": True,
-                "lora_r": 8,
-                "lora_alpha": 16,
-                "lora_dropout": 0.05,
-                "lora_target_linear": True,
-                "lora_modules_to_save": ["embed_tokens", "lm_head"],
-                "val_set_size": 0.1,
-                "special_tokens": {"pad_token": "<|end_of_text|>"},
-                "datasets": [
-                    {
-                        "path": "teknium/GPT4-LLM-Cleaned",
-                        "type": "alpaca",
-                    },
-                ],
-                "num_epochs": 1,
-                "max_steps": 5,
-                "micro_batch_size": 2,
-                "gradient_accumulation_steps": 4,
-                "output_dir": temp_dir,
-                "learning_rate": 0.00001,
-                "optimizer": "adamw_8bit",
-                "lr_scheduler": "cosine",
-                "flash_attention": True,
-                "loss_watchdog_threshold": 5.0,
-                "loss_watchdog_patience": 3,
-                "bf16": "auto",
-                "warmup_steps": 1,
-                "evals_per_epoch": 2,
-                "eval_max_new_tokens": 128,
-                "saves_per_epoch": 1,
-                "logging_steps": 1,
-                "weight_decay": 0.0,
-            }
-        )
-
-        # write cfg to yaml file
-        Path(temp_dir).mkdir(parents=True, exist_ok=True)
-        with open(Path(temp_dir) / "config.yaml", "w", encoding="utf-8") as fout:
-            fout.write(yaml.dump(cfg.to_dict(), Dumper=yaml.Dumper))
-
-        execute_subprocess_async(
-            [
-                "accelerate",
-                "launch",
-                "--num-processes",
-                "2",
-                "-m",
-                "axolotl.cli.train",
-                str(Path(temp_dir) / "config.yaml"),
-            ]
-        )
-
-    @with_temp_dir
-    def test_eval(self, temp_dir):
-        # pylint: disable=duplicate-code
-        cfg = DictDefault(
-            {
-                "base_model": "JackFram/llama-68m",
-                "load_in_8bit": False,
-                "load_in_4bit": True,
-                "strict": False,
-                "sequence_len": 2048,
-                "adapter": "qlora",
-                "sample_packing": True,
-                "eval_sample_packing": False,
-                "pad_to_sequence_len": True,
-                "lora_r": 8,
-                "lora_alpha": 16,
-                "lora_dropout": 0.05,
-                "lora_target_linear": True,
-                "lora_modules_to_save": ["embed_tokens", "lm_head"],
-                "val_set_size": 0.1,
-                "special_tokens": {"pad_token": "<|end_of_text|>"},
-                "datasets": [
-                    {
-                        "path": "teknium/GPT4-LLM-Cleaned",
-                        "type": "alpaca",
-                    },
-                ],
-                "num_epochs": 1,
-                "max_steps": 5,
-                "micro_batch_size": 2,
-                "gradient_accumulation_steps": 4,
-                "output_dir": temp_dir,
-                "learning_rate": 0.00001,
-                "optimizer": "adamw_8bit",
-                "lr_scheduler": "cosine",
-                "flash_attention": True,
-                "loss_watchdog_threshold": 5.0,
-                "loss_watchdog_patience": 3,
-                "bf16": "auto",
-                "warmup_steps": 1,
-                "evals_per_epoch": 2,
-                "eval_max_new_tokens": 128,
-                "saves_per_epoch": 1,
-                "logging_steps": 1,
-                "weight_decay": 0.0,
-            }
-        )
-
-        # write cfg to yaml file
-        Path(temp_dir).mkdir(parents=True, exist_ok=True)
-        with open(Path(temp_dir) / "config.yaml", "w", encoding="utf-8") as fout:
-            fout.write(yaml.dump(cfg.to_dict(), Dumper=yaml.Dumper))
-
-        execute_subprocess_async(
-            [
-                "accelerate",
-                "launch",
-                "--num-processes",
-                "2",
-                "-m",
-                "axolotl.cli.train",
-                str(Path(temp_dir) / "config.yaml"),
-            ]
-        )

tests/e2e/multigpu/test_llama.py

@@ -14,7 +14,7 @@ from huggingface_hub import snapshot_download
 
 from axolotl.utils.dict import DictDefault
 
-from ..utils import is_hopper, with_temp_dir
+from ..utils import with_temp_dir
 
 LOG = logging.getLogger("axolotl.tests.e2e.multigpu")
 os.environ["WANDB_DISABLED"] = "true"
@@ -59,7 +59,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 15,
+                "max_steps": 100,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -116,7 +116,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 15,
+                "max_steps": 50,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -144,146 +144,6 @@ class TestMultiGPULlama(unittest.TestCase):
             ]
         )
 
-    @pytest.mark.skipif(is_hopper(), reason="h100 doesn't support 8-bit lora")
-    @with_temp_dir
-    def test_dpo_lora_ddp(self, temp_dir):
-        # pylint: disable=duplicate-code
-        cfg = DictDefault(
-            {
-                "base_model": "TinyLlama/TinyLlama_v1.1",
-                "tokenizer_type": "LlamaTokenizer",
-                "sequence_len": 2048,
-                "sample_packing": False,
-                "eval_sample_packing": False,
-                "pad_to_sequence_len": True,
-                "load_in_8bit": True,
-                "adapter": "lora",
-                "lora_r": 8,
-                "lora_alpha": 16,
-                "lora_dropout": 0.05,
-                "lora_target_linear": True,
-                "val_set_size": 0.05,
-                "special_tokens": {
-                    "unk_token": "<unk>",
-                    "bos_token": "<s>",
-                    "eos_token": "</s>",
-                },
-                "rl": "dpo",
-                "chat_template": "llama3",
-                "datasets": [
-                    {
-                        "path": "fozziethebeat/alpaca_messages_2k_dpo_test",
-                        "type": "chat_template.default",
-                        "field_messages": "conversation",
-                        "field_chosen": "chosen",
-                        "field_rejected": "rejected",
-                        "message_field_role": "role",
-                        "message_field_content": "content",
-                        "roles": {
-                            "system": ["system"],
-                            "user": ["user"],
-                            "assistant": ["assistant"],
-                        },
-                    },
-                ],
-                "num_epochs": 1,
-                "max_steps": 15,
-                "micro_batch_size": 4,
-                "gradient_accumulation_steps": 4,
-                "output_dir": temp_dir,
-                "warmup_steps": 0,
-                "learning_rate": 0.00001,
-                "optimizer": "adamw_8bit",
-                "lr_scheduler": "cosine",
-                "flash_attention": True,
-            }
-        )
-
-        # write cfg to yaml file
-        Path(temp_dir).mkdir(parents=True, exist_ok=True)
-        with open(Path(temp_dir) / "config.yaml", "w", encoding="utf-8") as fout:
-            fout.write(yaml.dump(cfg.to_dict(), Dumper=yaml.Dumper))
-
-        execute_subprocess_async(
-            [
-                "accelerate",
-                "launch",
-                "--num-processes",
-                "2",
-                "-m",
-                "axolotl.cli.train",
-                str(Path(temp_dir) / "config.yaml"),
-            ]
-        )
-
-    @with_temp_dir
-    def test_dpo_qlora_ddp(self, temp_dir):
-        # pylint: disable=duplicate-code
-        cfg = DictDefault(
-            {
-                "base_model": "HuggingFaceTB/SmolLM-135M",
-                "sequence_len": 2048,
-                "sample_packing": False,
-                "eval_sample_packing": False,
-                "pad_to_sequence_len": True,
-                "load_in_4bit": True,
-                "adapter": "qlora",
-                "lora_r": 8,
-                "lora_alpha": 16,
-                "lora_dropout": 0.05,
-                "lora_target_linear": True,
-                "val_set_size": 0.05,
-                "special_tokens": {
-                    "pad_token": "<|endoftext|>",
-                },
-                "rl": "dpo",
-                "chat_template": "chatml",
-                "datasets": [
-                    {
-                        "path": "fozziethebeat/alpaca_messages_2k_dpo_test",
-                        "type": "chat_template.default",
-                        "field_messages": "conversation",
-                        "field_chosen": "chosen",
-                        "field_rejected": "rejected",
-                        "message_field_role": "role",
-                        "message_field_content": "content",
-                        "roles": {
-                            "system": ["system"],
-                            "user": ["user"],
-                            "assistant": ["assistant"],
-                        },
-                    },
-                ],
-                "num_epochs": 1,
-                "max_steps": 15,
-                "micro_batch_size": 4,
-                "gradient_accumulation_steps": 4,
-                "output_dir": temp_dir,
-                "warmup_steps": 0,
-                "learning_rate": 0.00001,
-                "optimizer": "adamw_8bit",
-                "lr_scheduler": "cosine",
-                "flash_attention": True,
-            }
-        )
-
-        # write cfg to yaml file
-        Path(temp_dir).mkdir(parents=True, exist_ok=True)
-        with open(Path(temp_dir) / "config.yaml", "w", encoding="utf-8") as fout:
-            fout.write(yaml.dump(cfg.to_dict(), Dumper=yaml.Dumper))
-
-        execute_subprocess_async(
-            [
-                "accelerate",
-                "launch",
-                "--num-processes",
-                "2",
-                "-m",
-                "axolotl.cli.train",
-                str(Path(temp_dir) / "config.yaml"),
-            ]
-        )
-
     @with_temp_dir
     def test_fsdp(self, temp_dir):
         # pylint: disable=duplicate-code
@@ -305,7 +165,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 15,
+                "max_steps": 100,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -371,7 +231,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 15,
+                "max_steps": 100,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -413,6 +273,7 @@ class TestMultiGPULlama(unittest.TestCase):
             ]
         )
 
+    @pytest.mark.skip("disabled due to upstream issue")
     @with_temp_dir
     def test_fsdp_qlora_prequant_packed(self, temp_dir):
         # pylint: disable=duplicate-code
@@ -421,7 +282,6 @@ class TestMultiGPULlama(unittest.TestCase):
                 "base_model": "axolotl-ai-co/TinyLlama_v1.1-bnb-nf4-bf16",
                 "tokenizer_type": "AutoTokenizer",
                 "adapter": "qlora",
-                "mean_resizing_embeddings": True,
                 "load_in_4bit": True,
                 "lora_r": 8,
                 "lora_alpha": 16,
@@ -437,7 +297,7 @@ class TestMultiGPULlama(unittest.TestCase):
                 "sequence_len": 2048,
                 "val_set_size": 0.05,
                 "special_tokens": {
-                    "pad_token": "</s>",
+                    "pad_token": "<|end_of_text|>",
                 },
                 "datasets": [
                     {
@@ -447,7 +307,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 15,
+                "max_steps": 100,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -513,7 +373,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 15,
+                "max_steps": 100,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -572,7 +432,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 15,
+                "max_steps": 100,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,

tests/e2e/multigpu/test_qwen2.py

@@ -47,7 +47,7 @@ class TestMultiGPUQwen2(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 15,
+                "max_steps": 100,
                 "warmup_steps": 20,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 2,

tests/e2e/patched/test_4d_multipack_llama.py

@@ -13,7 +13,7 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import require_torch_2_3_1, with_temp_dir
+from ..utils import require_torch_2_1_1, with_temp_dir
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
@@ -24,7 +24,7 @@ class Test4dMultipackLlama(unittest.TestCase):
     Test case for Llama models using 4d attention with multipack
     """
 
-    @require_torch_2_3_1
+    @require_torch_2_1_1
     @with_temp_dir
     def test_sdp_lora_packing(self, temp_dir):
         # pylint: disable=duplicate-code

tests/e2e/patched/test_unsloth_integration.py

@@ -1,12 +1,22 @@
 """Test module for checking whether the integration of Unsloth with Hugging Face Transformers is working as expected."""
 import unittest
 
-from axolotl.monkeypatch.unsloth_ import check_self_attn_is_patchable
+from axolotl.monkeypatch.unsloth_ import (
+    check_cel_is_patchable,
+    check_self_attn_is_patchable,
+)
 
 
 class TestUnslothIntegration(unittest.TestCase):
     """Unsloth monkeypatch integration tests."""
 
+    def test_is_cel_patchable(self):
+        # ensures the current version of transformers has loss code that matches our patching code
+        self.assertTrue(
+            check_cel_is_patchable(),
+            "HF transformers loss code has changed and isn't patchable",
+        )
+
     def test_is_self_attn_patchable(self):
         # ensures the current version of transformers has loss code that matches our patching code
         self.assertTrue(

tests/e2e/test_dpo.py

@@ -115,51 +115,6 @@ class TestDPOLlamaLora(unittest.TestCase):
         train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
         assert (Path(temp_dir) / "checkpoint-20/adapter_model.safetensors").exists()
 
-    @with_temp_dir
-    def test_dpo_use_weighting(self, temp_dir):
-        # pylint: disable=duplicate-code
-        cfg = DictDefault(
-            {
-                "base_model": "JackFram/llama-68m",
-                "tokenizer_type": "LlamaTokenizer",
-                "sequence_len": 1024,
-                "load_in_8bit": True,
-                "adapter": "lora",
-                "lora_r": 64,
-                "lora_alpha": 32,
-                "lora_dropout": 0.1,
-                "lora_target_linear": True,
-                "special_tokens": {},
-                "rl": "dpo",
-                "dpo_use_weighting": True,
-                "datasets": [
-                    {
-                        "path": "arcee-ai/distilabel-intel-orca-dpo-pairs-binarized",
-                        "type": "chatml.ultra",
-                        "split": "train",
-                    },
-                ],
-                "num_epochs": 1,
-                "micro_batch_size": 4,
-                "gradient_accumulation_steps": 1,
-                "output_dir": temp_dir,
-                "learning_rate": 0.00001,
-                "optimizer": "paged_adamw_8bit",
-                "lr_scheduler": "cosine",
-                "max_steps": 20,
-                "save_steps": 10,
-                "warmup_steps": 5,
-                "gradient_checkpointing": True,
-                "gradient_checkpointing_kwargs": {"use_reentrant": True},
-            }
-        )
-        normalize_config(cfg)
-        cli_args = TrainerCliArgs()
-        dataset_meta = load_rl_datasets(cfg=cfg, cli_args=cli_args)
-
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "checkpoint-20/adapter_model.safetensors").exists()
-
     @pytest.mark.skip("kto_pair no longer supported in trl")
     @with_temp_dir
     def test_kto_pair_lora(self, temp_dir):

tests/e2e/test_load_model.py

@@ -1,95 +0,0 @@
-"""Module for testing ModelLoader."""
-
-import shutil
-import tempfile
-
-import pytest
-import torch
-
-from axolotl.utils.dict import DictDefault
-from axolotl.utils.models import ModelLoader, load_model, load_tokenizer
-
-
-@pytest.fixture(name="temp_dir")
-def fixture_temp_dir():
-    temp_dir = tempfile.mkdtemp()
-    yield temp_dir
-    shutil.rmtree(temp_dir)
-
-
-class TestLoadModelUtils:
-    """
-    Testing module testing ModelLoader.
-    """
-
-    def setup_method(self):
-        # load config
-        self.cfg = DictDefault(
-            {
-                "base_model": "JackFram/llama-68m",
-                "tokenizer_type": "LlamaTokenizer",
-                "tokenizer_config": "JackFram/llama-68m",
-                "sequence_len": 1024,
-                "load_in_8bit": False,
-                "adapter": "lora",
-                "lora_r": 8,
-                "lora_alpha": 16,
-                "lora_dropout": 0.05,
-                "lora_target_linear": True,
-                "val_set_size": 0.1,
-                "special_tokens": {
-                    "unk_token": "<unk>",
-                    "bos_token": "<s>",
-                    "eos_token": "</s>",
-                },
-                "datasets": [
-                    {
-                        "path": "mhenrichsen/alpaca_2k_test",
-                        "type": "alpaca",
-                    },
-                ],
-                "num_epochs": 1,
-                "micro_batch_size": 8,
-                "gradient_accumulation_steps": 1,
-                "learning_rate": 0.00001,
-                "optimizer": "adamw_torch",
-                "lr_scheduler": "cosine",
-            }
-        )
-        self.model_loader = (  # pylint: disable=attribute-defined-outside-init
-            ModelLoader(
-                cfg=self.cfg,
-                tokenizer="",
-            )
-        )
-
-    @pytest.mark.parametrize("embedding_modules", ["embed_tokens", "lm_head"])
-    @pytest.mark.parametrize(
-        "dist_dtype", [torch.bfloat16, torch.float16, torch.float32]
-    )
-    @pytest.mark.parametrize("before_kbit_train_or_finetune", [True, False])
-    def test_convert_embedding_modules_dtype(
-        self, temp_dir, embedding_modules, dist_dtype, before_kbit_train_or_finetune
-    ):
-        self.cfg.output_dir = temp_dir
-        self.model_loader.tokenizer = load_tokenizer(self.cfg)  # pylint: disable=all
-        self.model_loader.model, _ = load_model(
-            self.cfg,
-            self.model_loader.tokenizer,
-            inference=False,
-            reference_model=True,
-        )
-        self.model_loader.convert_embedding_modules_dtype(
-            embedding_modules, dist_dtype, before_kbit_train_or_finetune
-        )
-        for name, module in self.model_loader.model.named_modules():
-            if (
-                "norm" in name
-                or (before_kbit_train_or_finetune and name.endswith(".gate"))
-                or (
-                    any(m in name for m in embedding_modules)
-                    and hasattr(module, "weight")
-                )
-            ):
-                for _, param in module.named_parameters():
-                    assert param.dtype == dist_dtype

tests/e2e/test_optimizers.py

@@ -13,7 +13,7 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 
-from .utils import require_torch_2_5_1, with_temp_dir
+from .utils import with_temp_dir
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
@@ -65,46 +65,3 @@ class TestCustomOptimizers(unittest.TestCase):
 
         train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
         assert (Path(temp_dir) / "adapter_model.bin").exists()
-
-    @with_temp_dir
-    @require_torch_2_5_1
-    def test_adopt_adamw(self, temp_dir):
-        # pylint: disable=duplicate-code
-        cfg = DictDefault(
-            {
-                "base_model": "JackFram/llama-68m",
-                "tokenizer_type": "LlamaTokenizer",
-                "sequence_len": 1024,
-                "load_in_8bit": True,
-                "adapter": "lora",
-                "lora_r": 8,
-                "lora_alpha": 16,
-                "lora_dropout": 0.05,
-                "lora_target_linear": True,
-                "val_set_size": 0.1,
-                "special_tokens": {
-                    "unk_token": "<unk>",
-                    "bos_token": "<s>",
-                    "eos_token": "</s>",
-                },
-                "datasets": [
-                    {
-                        "path": "mhenrichsen/alpaca_2k_test",
-                        "type": "alpaca",
-                    },
-                ],
-                "num_epochs": 1,
-                "micro_batch_size": 8,
-                "gradient_accumulation_steps": 1,
-                "output_dir": temp_dir,
-                "learning_rate": 0.00001,
-                "optimizer": "adopt_adamw",
-                "lr_scheduler": "cosine",
-            }
-        )
-        normalize_config(cfg)
-        cli_args = TrainerCliArgs()
-        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
-
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "adapter_model.bin").exists()

tests/e2e/test_packing_loss.py

@@ -1,74 +0,0 @@
-"""
-E2E tests for packed training
-"""
-
-import logging
-import os
-import unittest
-
-from tbparse import SummaryReader
-from transformers.utils import is_torch_bf16_gpu_available
-
-from axolotl.cli import load_datasets
-from axolotl.common.cli import TrainerCliArgs
-from axolotl.train import train
-from axolotl.utils.config import normalize_config
-from axolotl.utils.dict import DictDefault
-
-from .utils import most_recent_subdir, with_temp_dir
-
-LOG = logging.getLogger("axolotl.tests.e2e")
-os.environ["WANDB_DISABLED"] = "true"
-
-
-class TestPackedLlama(unittest.TestCase):
-    """
-    Test case for Packed training of llama models
-    """
-
-    @with_temp_dir
-    def test_loss_packed(self, temp_dir):
-        # pylint: disable=duplicate-code
-        cfg = DictDefault(
-            {
-                "base_model": "HuggingFaceTB/SmolLM-135M",
-                "sequence_len": 1024,
-                "sample_packing": True,
-                "flash_attention": True,
-                "val_set_size": 0.0,
-                "special_tokens": {
-                    "pad_token": "<|endoftext|>",
-                },
-                "datasets": [
-                    {
-                        "path": "vicgalle/alpaca-gpt4",
-                        "type": "alpaca",
-                    },
-                ],
-                "num_epochs": 1,
-                "micro_batch_size": 2,
-                "gradient_accumulation_steps": 4,
-                "output_dir": temp_dir,
-                "learning_rate": 0.00001,
-                "optimizer": "adamw_torch",
-                "lr_scheduler": "cosine",
-                "max_steps": 5,
-                "use_tensorboard": True,
-            }
-        )
-        if is_torch_bf16_gpu_available():
-            cfg.bf16 = True
-        else:
-            cfg.fp16 = True
-        normalize_config(cfg)
-        cli_args = TrainerCliArgs()
-        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
-
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-
-        tb_log_path = most_recent_subdir(temp_dir + "/runs")
-        event_file = os.path.join(tb_log_path, sorted(os.listdir(tb_log_path))[0])
-        reader = SummaryReader(event_file)
-        df = reader.scalars  # pylint: disable=invalid-name
-        df = df[(df.tag == "train/train_loss")]  # pylint: disable=invalid-name
-        assert df.value.values[-1] < 2.0, "Loss is too high"

tests/e2e/utils.py

@@ -6,13 +6,9 @@ import shutil
 import tempfile
 import unittest
 from functools import wraps
+from importlib.metadata import version
 from pathlib import Path
 
-import torch
-
-# from importlib.metadata import version
-from packaging import version
-
 
 def with_temp_dir(test_func):
     @wraps(test_func)
@@ -39,30 +35,13 @@ def most_recent_subdir(path):
     return subdir
 
 
-def require_torch_2_3_1(test_case):
+def require_torch_2_1_1(test_case):
     """
-    Decorator marking a test that requires torch >= 2.3.1
+    Decorator marking a test that requires torch >= 2.1.1
     """
 
-    def is_min_2_3_1():
-        torch_version = version.parse(torch.__version__)
-        return torch_version >= version.parse("2.3.1")
+    def is_min_2_1_1():
+        torch_version = version("torch")
+        return torch_version >= "2.1.1"
 
-    return unittest.skipUnless(is_min_2_3_1(), "test torch 2.3.1")(test_case)
-
-
-def require_torch_2_5_1(test_case):
-    """
-    Decorator marking a test that requires torch >= 2.3.1
-    """
-
-    def is_min_2_5_1():
-        torch_version = version.parse(torch.__version__)
-        return torch_version >= version.parse("2.5.1")
-
-    return unittest.skipUnless(is_min_2_5_1(), "test torch 2.5.1")(test_case)
-
-
-def is_hopper():
-    compute_capability = torch.cuda.get_device_capability()
-    return compute_capability == (9, 0)
+    return unittest.skipUnless(is_min_2_1_1(), "test torch 2.1.1")(test_case)

tests/integrations/liger.py

@@ -1,80 +0,0 @@
-"""
-config validation tests for swiglu args
-"""
-# pylint: disable=duplicate-code
-import logging
-from typing import Optional
-
-import pytest
-
-from axolotl.utils.config import validate_config
-from axolotl.utils.dict import DictDefault
-
-
-@pytest.fixture(name="minimal_base_cfg")
-def fixture_cfg():
-    return DictDefault(
-        {
-            "base_model": "TinyLlama/TinyLlama-1.1B-Chat-v0.6",
-            "learning_rate": 0.000001,
-            "datasets": [
-                {
-                    "path": "mhenrichsen/alpaca_2k_test",
-                    "type": "alpaca",
-                }
-            ],
-            "micro_batch_size": 1,
-            "gradient_accumulation_steps": 1,
-        }
-    )
-
-
-class BaseValidation:
-    """
-    Base validation module to setup the log capture
-    """
-
-    _caplog: Optional[pytest.LogCaptureFixture] = None
-
-    @pytest.fixture(autouse=True)
-    def inject_fixtures(self, caplog):
-        self._caplog = caplog
-
-
-# pylint: disable=too-many-public-methods
-class TestValidation(BaseValidation):
-    """
-    Test the validation module for liger
-    """
-
-    def test_deprecated_swiglu(self, minimal_cfg):
-        test_cfg = DictDefault(
-            {
-                "liger_swiglu": False,
-            }
-            | minimal_cfg
-        )
-
-        with self._caplog.at_level(logging.WARNING):
-            updated_cfg = validate_config(test_cfg)
-            assert (
-                "The 'liger_swiglu' argument is deprecated"
-                in self._caplog.records[0].message
-            )
-            assert updated_cfg.liger_swiglu is None
-            assert updated_cfg.liger_glu_activations is False
-
-    def test_conflict_swiglu_ligergluactivation(self, minimal_cfg):
-        test_cfg = DictDefault(
-            {
-                "liger_swiglu": False,
-                "liger_glu_activations": True,
-            }
-            | minimal_cfg
-        )
-
-        with pytest.raises(
-            ValueError,
-            match=r".*You cannot have both `liger_swiglu` and `liger_glu_activation` set.*",
-        ):
-            validate_config(test_cfg)

tests/prompt_strategies/test_chat_template_utils.py

@@ -1,125 +0,0 @@
-"""
-Tests for utils in axolotl.utils.chat_templates
-"""
-import unittest
-
-import pytest
-from transformers import AutoTokenizer
-
-from axolotl.utils.chat_templates import (
-    _CHAT_TEMPLATES,
-    extract_chat_template_args,
-    get_chat_template,
-)
-
-
-@pytest.fixture(name="llama3_tokenizer")
-def fixture_llama3_tokenizer():
-    tokenizer = AutoTokenizer.from_pretrained("NousResearch/Meta-Llama-3-8B")
-
-    return tokenizer
-
-
-class TestGetChatTemplateUtils:
-    """
-    Tests the get_chat_template function.
-    """
-
-    def test_known_chat_template(self):
-        chat_template_str = get_chat_template("llama3")
-        assert chat_template_str == _CHAT_TEMPLATES["llama3"]
-
-    def test_invalid_chat_template(self):
-        with pytest.raises(ValueError) as exc:
-            get_chat_template("invalid_template")
-            assert str(exc) == "Template 'invalid_template' not found."
-
-    def test_tokenizer_default_no_tokenizer(self):
-        with pytest.raises(ValueError):
-            get_chat_template("tokenizer_default", tokenizer=None)
-
-    def test_tokenizer_default_no_chat_template_on_tokenizer(self, llama3_tokenizer):
-        with pytest.raises(ValueError):
-            get_chat_template("tokenizer_default", tokenizer=llama3_tokenizer)
-
-    def test_tokenizer_default_with_chat_template_on_tokenizer(self, llama3_tokenizer):
-        llama3_tokenizer.chat_template = "test_template"
-        chat_template_str = get_chat_template(
-            "tokenizer_default", tokenizer=llama3_tokenizer
-        )
-        assert chat_template_str == "test_template"
-
-    def test_tokenizer_default_fallback_no_tokenizer(self):
-        with pytest.raises(ValueError):
-            get_chat_template("tokenizer_default_fallback_test", tokenizer=None)
-
-    def test_tokenizer_default_fallback_no_chat_template_on_tokenizer(
-        self, llama3_tokenizer
-    ):
-        chat_template_str = get_chat_template(
-            "tokenizer_default_fallback_chatml", tokenizer=llama3_tokenizer
-        )
-        assert chat_template_str == get_chat_template("chatml")
-
-    def test_tokenizer_default_fallback_with_chat_template_on_tokenizer(
-        self, llama3_tokenizer
-    ):
-        llama3_tokenizer.chat_template = "test_template"
-        chat_template_str = get_chat_template(
-            "tokenizer_default_fallback_chatml", tokenizer=llama3_tokenizer
-        )
-        assert chat_template_str == "test_template"
-
-    def test_jinja_template_mode(self):
-        jinja_template = "example_jinja_template"
-        chat_template_str = get_chat_template("jinja", jinja_template=jinja_template)
-        assert chat_template_str == jinja_template
-
-    def test_jinja_template_mode_no_jinja_template(self):
-        with pytest.raises(ValueError):
-            get_chat_template("jinja", jinja_template=None)
-
-    def test_extract_chat_template_args(self):
-        # No ds_cfg
-        chat_template_choice, chat_template_jinja = extract_chat_template_args(
-            cfg={"chat_template": "chatml"},
-        )
-        assert chat_template_choice == "chatml"
-        assert chat_template_jinja is None
-
-        # ds_cfg provided
-        chat_template_choice, chat_template_jinja = extract_chat_template_args(
-            cfg={
-                "chat_template": "jinja",
-                "chat_template_jinja": "global_jinja_template",
-            },
-            ds_cfg={"chat_template": "llama3", "chat_template_jinja": None},
-        )
-        assert chat_template_choice == "llama3"
-        assert chat_template_jinja is None
-
-        # ds_cfg provided with jinja template
-        chat_template_choice, chat_template_jinja = extract_chat_template_args(
-            cfg={"chat_template": "chatml", "chat_template_jinja": None},
-            ds_cfg={
-                "chat_template": "jinja",
-                "chat_template_jinja": "ds_jinja_template",
-            },
-        )
-        assert chat_template_choice == "jinja"
-        assert chat_template_jinja == "ds_jinja_template"
-
-        # ds_cfg provided with no chat_template
-        chat_template_choice, chat_template_jinja = extract_chat_template_args(
-            cfg={
-                "chat_template": "jinja",
-                "chat_template_jinja": "global_jinja_template",
-            },
-            ds_cfg={"chat_template": None, "chat_template_jinja": "ds_jinja_template"},
-        )
-        assert chat_template_choice == "jinja"
-        assert chat_template_jinja == "global_jinja_template"
-
-
-if __name__ == "__main__":
-    unittest.main()

tests/prompt_strategies/test_chat_templates.py

@@ -11,7 +11,7 @@ from axolotl.prompt_strategies.chat_template import (
     load,
 )
 from axolotl.prompters import IGNORE_TOKEN_ID
-from axolotl.utils.chat_templates import get_chat_template
+from axolotl.utils.chat_templates import chat_templates
 from axolotl.utils.dict import DictDefault
 
 logging.basicConfig(level=logging.DEBUG)
@@ -73,7 +73,7 @@ class TestAssistantChatTemplateLlama3:
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
                 llama3_tokenizer,
-                chat_template=get_chat_template("llama3"),
+                chat_template=chat_templates("llama3"),
                 message_field_role="role",
                 message_field_content="content",
                 roles={
@@ -113,7 +113,7 @@ class TestAssistantChatTemplateLlama3:
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
                 phi35_tokenizer,
-                chat_template=get_chat_template("phi_35"),
+                chat_template=chat_templates("phi_35"),
                 message_field_role="role",
                 message_field_content="content",
                 roles={
@@ -171,7 +171,7 @@ class TestAssistantChatTemplateLlama3:
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
                 llama3_tokenizer,
-                chat_template=get_chat_template("llama3"),
+                chat_template=chat_templates("llama3"),
                 message_field_role="role",
                 message_field_content="content",
                 message_field_training="training",
@@ -230,7 +230,7 @@ class TestSharegptChatTemplateLlama3:
         # pylint: disable=duplicate-code
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=False,
@@ -283,7 +283,7 @@ class TestSharegptChatTemplateLlama3:
         # pylint: disable=duplicate-code
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=False,
@@ -336,7 +336,7 @@ class TestSharegptChatTemplateLlama3:
         # pylint: disable=duplicate-code
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=False,

tests/prompt_strategies/test_chat_templates_advanced.py

@@ -12,7 +12,7 @@ from axolotl.prompt_strategies.chat_template import (
     ChatTemplateStrategy,
 )
 from axolotl.prompters import IGNORE_TOKEN_ID
-from axolotl.utils.chat_templates import get_chat_template
+from axolotl.utils.chat_templates import chat_templates
 
 logging.basicConfig(level=logging.DEBUG)
 LOG = logging.getLogger("axolotl")
@@ -35,7 +35,7 @@ class TestChatTemplateConfigurations:
         LOG.info("Testing with train_on_inputs=True")
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=True,
@@ -80,7 +80,7 @@ class TestChatTemplateConfigurations:
         LOG.info("Testing with train_on_inputs=False")
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=False,
@@ -123,7 +123,7 @@ class TestChatTemplateConfigurations:
         LOG.info("Testing roles_to_train with assistant only")
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=False,
@@ -151,7 +151,7 @@ class TestChatTemplateConfigurations:
         LOG.info("Testing roles_to_train with all roles")
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=True,
@@ -184,7 +184,7 @@ class TestChatTemplateConfigurations:
         LOG.info("Testing with empty roles_to_train")
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=False,
@@ -205,7 +205,7 @@ class TestChatTemplateConfigurations:
         LOG.info("Testing with train_on_eos='all'")
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=False,
@@ -232,7 +232,7 @@ class TestChatTemplateConfigurations:
         LOG.info("Testing with train_on_eos='turn'")
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=False,
@@ -282,7 +282,7 @@ class TestChatTemplateConfigurations:
         LOG.info("Testing with train_on_eos='last'")
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=False,
@@ -315,7 +315,7 @@ class TestChatTemplateConfigurations:
         LOG.info("Testing with train_on_eos='none'")
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
-                llama3_tokenizer, chat_template=get_chat_template("llama3")
+                llama3_tokenizer, chat_template=chat_templates("llama3")
             ),
             tokenizer=llama3_tokenizer,
             train_on_inputs=False,
@@ -343,7 +343,7 @@ class TestChatTemplateConfigurations:
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
                 llama3_tokenizer,
-                chat_template=get_chat_template("llama3"),
+                chat_template=chat_templates("llama3"),
                 drop_system_message=True,
             ),
             tokenizer=llama3_tokenizer,
@@ -371,7 +371,7 @@ class TestChatTemplateConfigurations:
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
                 llama3_tokenizer,
-                chat_template=get_chat_template("llama3"),
+                chat_template=chat_templates("llama3"),
                 roles=custom_roles,
             ),
             tokenizer=llama3_tokenizer,
@@ -424,7 +424,7 @@ class TestChatTemplateConfigurations:
         strategy = ChatTemplateStrategy(
             ChatTemplatePrompter(
                 llama3_tokenizer,
-                chat_template=get_chat_template("llama3"),
+                chat_template=chat_templates("llama3"),
                 message_field_training="train",
                 message_field_training_detail="train_detail",
             ),

tests/prompt_strategies/test_dpo_chat_templates.py

@@ -86,20 +86,6 @@ def fixture_llama3_tokenizer():
     return tokenizer
 
 
-@pytest.fixture(name="phi3_tokenizer")
-def fixture_phi3_tokenizer():
-    tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3-medium-128k-instruct")
-
-    return tokenizer
-
-
-@pytest.fixture(name="gemma_tokenizer")
-def fixture_gemma_tokenizer():
-    tokenizer = AutoTokenizer.from_pretrained("unsloth/gemma-2b-it", revision="703fb4a")
-
-    return tokenizer
-
-
 class TestAssistantDPOChatTemplateLlama3:
     """
     Test class for assistant style datasets with llama-3 prompts using the chat_template strategy.
@@ -113,7 +99,7 @@ class TestAssistantDPOChatTemplateLlama3:
                     "chat_template": "llama3",
                     "datasets": [
                         {
-                            "type": "chat_template",
+                            "chat_template": "llama3",
                         }
                     ],
                 }
@@ -138,7 +124,7 @@ class TestAssistantDPOChatTemplateLlama3:
                     "chat_template": "llama3",
                     "datasets": [
                         {
-                            "type": "chat_template",
+                            "chat_template": "llama3",
                             "field_messages": "conversation",
                             "field_chosen": "better",
                             "field_rejected": "worse",
@@ -166,65 +152,5 @@ class TestAssistantDPOChatTemplateLlama3:
         assert result["rejected"] == "party on<|eot_id|>"
 
 
-class TestAssistantDPOChatTemplatePhi3:
-    """
-    Test class for assistant style datasets with phi-3 prompts using the tokenizer's chat_template strategy.
-    """
-
-    def test_phi3_defaults(self, phi3_tokenizer, assistant_dataset):
-        # pylint: disable=duplicate-code
-        transform_fn = default(
-            DictDefault(
-                {
-                    "chat_template": "tokenizer_default",
-                    "datasets": [
-                        {
-                            "type": "chat_template",
-                        }
-                    ],
-                }
-            )
-        )
-        result = transform_fn(assistant_dataset[0], tokenizer=phi3_tokenizer)
-        assert result["prompt"] == (
-            "<|user|>\nhello<|end|>\n"
-            + "<|assistant|>\nhello<|end|>\n"
-            + "<|user|>\ngoodbye<|end|>\n"
-            + "<|assistant|>\n"
-        )
-        assert result["chosen"] == "goodbye<|end|>"
-        assert result["rejected"] == "party on<|end|>"
-
-
-class TestAssistantDPOChatTemplateGemma:
-    """
-    Test class for assistant style datasets with gemma prompts using the tokenizer's chat_template strategy.
-    """
-
-    def test_gemma_defaults(self, gemma_tokenizer, assistant_dataset):
-        # pylint: disable=duplicate-code
-        transform_fn = default(
-            DictDefault(
-                {
-                    "chat_template": "tokenizer_default",
-                    "datasets": [
-                        {
-                            "type": "chat_template",
-                        }
-                    ],
-                }
-            )
-        )
-        result = transform_fn(assistant_dataset[0], tokenizer=gemma_tokenizer)
-        assert result["prompt"] == (
-            "<bos><start_of_turn>user\nhello<end_of_turn>\n"
-            + "<start_of_turn>model\nhello<end_of_turn>\n"
-            + "<start_of_turn>user\ngoodbye<end_of_turn>\n"
-            + "<start_of_turn>model\n"
-        )
-        assert result["chosen"] == "goodbye<end_of_turn>"
-        assert result["rejected"] == "party on<end_of_turn>"
-
-
 if __name__ == "__main__":
     unittest.main()

tests/prompt_strategies/test_sharegpt.py

@@ -0,0 +1,500 @@
+"""
+Test module for sharegpt integration w chatml
+"""
+
+import pytest
+from datasets import Dataset
+from tokenizers import AddedToken
+from transformers import AutoTokenizer
+
+from axolotl.datasets import TokenizedPromptDataset
+from axolotl.prompt_strategies.sharegpt import (
+    GlaiveShareGPTPromptTokenizingStrategy,
+    SimpleShareGPTPromptTokenizingStrategy,
+    register_chatml_template,
+    register_llama3_template,
+)
+from axolotl.prompters import ShareGPTPrompterV2
+
+register_chatml_template()
+register_llama3_template()
+
+
+@pytest.fixture(name="sharegpt_dataset")
+def fixture_sharegpt_dataset():
+    return Dataset.from_list(
+        [
+            {
+                "conversations": [
+                    {
+                        "from": "system",
+                        "value": "repeat",
+                    },
+                    {
+                        "from": "human",
+                        "value": "hello",
+                    },
+                    {
+                        "from": "gpt",
+                        "value": "hello",
+                    },
+                    {
+                        "from": "human",
+                        "value": "goodbye",
+                    },
+                    {
+                        "from": "gpt",
+                        "value": "goodbye",
+                    },
+                ]
+            }
+        ]
+    )
+
+
+@pytest.fixture(name="sharegpt_dataset_with_weights")
+def fixture_sharegpt_dataset_with_weights():
+    return Dataset.from_list(
+        [
+            {
+                "conversations": [
+                    {
+                        "from": "system",
+                        "value": "repeat",
+                    },
+                    {
+                        "from": "human",
+                        "value": "hello",
+                        "weight": 1,
+                    },
+                    {
+                        "from": "gpt",
+                        "value": "hello",
+                        "weight": 0,
+                    },
+                    {
+                        "from": "human",
+                        "value": "rehello",
+                        "weight": 0,
+                    },
+                    {
+                        "from": "gpt",
+                        "value": "rehello",
+                        "weight": 1,
+                    },
+                    {
+                        "from": "human",
+                        "value": "goodbye",
+                    },
+                    {
+                        "from": "gpt",
+                        "value": "goodbye",
+                        "weight": 0,
+                    },
+                ]
+            }
+        ]
+    )
+
+
+@pytest.fixture(name="glaive_dataset")
+def fixture_sharegpt_glaive_dataset():
+    return Dataset.from_list(
+        [
+            {
+                "system": "SYSTEM: This is a system prompt",
+                "chat": "USER: Can you book a flight for me from New York to London? ASSISTANT: I'm sorry, but I don't have the capability to book flights.  <|endoftext|>",
+            }
+        ]
+    )
+
+
+@pytest.fixture(name="multi_role_dataset")
+def fixture_multi_role_dataset():
+    return Dataset.from_list(
+        [
+            {
+                "conversations": [
+                    {
+                        "from": "system",
+                        "value": "use get_weather(city) to get the weather for a city",
+                    },
+                    {
+                        "from": "human",
+                        "value": "hello, what's the weather in New York?",
+                    },
+                    {
+                        "from": "gpt",
+                        "value": "let me get that for you",
+                    },
+                    {
+                        "from": "tool",
+                        "value": "get_weather(New York)",
+                    },
+                    {
+                        "from": "gpt",
+                        "value": "the weather in New York is 70 degrees and sunny",
+                    },
+                ]
+            }
+        ]
+    )
+
+
+@pytest.fixture(name="tokenizer")
+def fixture_tokenizer():
+    tokenizer = AutoTokenizer.from_pretrained(
+        "casperhansen/mistral-7b-instruct-v0.1-awq"
+    )
+    tokenizer.add_special_tokens(
+        {
+            "eos_token": AddedToken(
+                "<|im_end|>", rstrip=False, lstrip=False, normalized=False
+            )
+        }
+    )
+    tokenizer.add_tokens(
+        [
+            AddedToken("<|im_start|>", rstrip=False, lstrip=False, normalized=False),
+        ]
+    )
+
+    return tokenizer
+
+
+@pytest.fixture(name="llama3_tokenizer")
+def fixture_llama3_tokenizer():
+    tokenizer = AutoTokenizer.from_pretrained("NousResearch/Meta-Llama-3-8B")
+    tokenizer.eos_token = "<|eot_id|>"
+
+    return tokenizer
+
+
+class TestSharegptLlama3:
+    """Test class for ShareGPT style datasets with llama-3 prompts"""
+
+    def test_tokenization(self, sharegpt_dataset, llama3_tokenizer):
+        strategy = SimpleShareGPTPromptTokenizingStrategy(
+            ShareGPTPrompterV2(
+                conversation="llama3",
+                role_key_model=None,
+                role_key_human=None,
+            ),
+            llama3_tokenizer,
+            False,  # train_on_inputs
+            2048,  # sequence_len
+        )
+
+        dataset_wrapper = TokenizedPromptDataset(
+            strategy, sharegpt_dataset, process_count=1
+        )
+
+        input_ids = dataset_wrapper[0]["input_ids"]
+
+        # fmt: off
+        # pylint: disable=duplicate-code
+        assert input_ids == [
+            128000,  # bos
+            128006, 9125, 128007,  # system header
+            271, 31724, 128009,  # sys prompt, eot
+            128006, 882, 128007,  # user header
+            271, 15339, 128009,  # user prompt eot
+            128006, 78191, 128007,  # assistant header
+            271, 15339, 128009,   # assistant response eot
+            128006, 882, 128007,
+            271, 19045, 29474, 128009,
+            128006, 78191, 128007,
+            271, 19045, 29474, 128009,
+        ]
+        # fmt: on
+
+    def test_tokenization_with_weights(
+        self, sharegpt_dataset_with_weights, llama3_tokenizer
+    ):
+        strategy = SimpleShareGPTPromptTokenizingStrategy(
+            ShareGPTPrompterV2(
+                conversation="llama3",
+                role_key_model=None,
+                role_key_human=None,
+            ),
+            llama3_tokenizer,
+            False,  # train_on_inputs
+            2048,  # sequence_len
+        )
+
+        dataset_wrapper = TokenizedPromptDataset(
+            strategy, sharegpt_dataset_with_weights, process_count=1
+        )
+
+        input_ids = dataset_wrapper[0]["input_ids"]
+
+        # fmt: off
+        # pylint: disable=duplicate-code
+        assert input_ids == [
+            128000,  # bos
+            128006, 9125, 128007,  # system header
+            271, 31724, 128009,  # sys prompt, eot
+            128006, 882, 128007,  # user header
+            271, 15339, 128009,  # user prompt eot
+            128006, 78191, 128007,  # assistant header
+            271, 15339, 128009,   # assistant response eot
+            128006, 882, 128007,
+            271, 11310, 4896, 128009,
+            128006, 78191, 128007,
+            271, 11310, 4896, 128009,
+            128006, 882, 128007,
+            271, 19045, 29474, 128009,
+            128006, 78191, 128007,
+            271, 19045, 29474, 128009,
+        ]
+        # fmt: on
+
+
+class TestSharegptChatML:
+    """
+    Test class for sharegpt prompter
+    """
+
+    def test_no_double_im_end(self, sharegpt_dataset, tokenizer):
+        strategy = SimpleShareGPTPromptTokenizingStrategy(
+            ShareGPTPrompterV2(
+                conversation="chatml",
+                role_key_model=None,
+                role_key_human=None,
+            ),
+            tokenizer,
+            False,  # train_on_inputs
+            2048,  # sequence_len
+        )
+
+        dataset_wrapper = TokenizedPromptDataset(
+            strategy, sharegpt_dataset, process_count=1
+        )
+
+        input_ids = dataset_wrapper[0]["input_ids"]
+        # fmt: off
+        assert input_ids == [
+            #  28705, 13, is " \n"
+            1,   # bos
+            32001, 1587, 13, 25997, 32000, 28705, 13,  # system
+            32001, 2188, 13, 21558, 32000, 28705, 13,  # human
+            32001, 13892, 13, 21558, 32000, 28705, 13,  # gpt
+            32001, 2188, 13, 12684, 17664, 32000, 28705, 13,   # human
+            32001, 13892, 13, 12684, 17664, 32000, 28705, 13,  # gpt
+        ]
+        # fmt: on
+
+    def test_no_double_im_end_with_weights(
+        self, sharegpt_dataset_with_weights, tokenizer
+    ):
+        strategy = SimpleShareGPTPromptTokenizingStrategy(
+            ShareGPTPrompterV2(
+                conversation="chatml",
+                role_key_model=None,
+                role_key_human=None,
+            ),
+            tokenizer,
+            False,  # train_on_inputs
+            2048,  # sequence_len
+        )
+
+        dataset_wrapper = TokenizedPromptDataset(
+            strategy, sharegpt_dataset_with_weights, process_count=1
+        )
+
+        input_ids = dataset_wrapper[0]["input_ids"]
+        # fmt: off
+        assert input_ids == [
+            #  28705, 13, is " \n"
+            1,   # bos
+            32001, 1587, 13, 25997, 32000, 28705, 13,  # system
+            32001, 2188, 13, 21558, 32000, 28705, 13,  # human
+            32001, 13892, 13, 21558, 32000, 28705, 13,  # gpt
+            32001, 2188, 13, 267, 21558, 32000, 28705, 13,  # human
+            32001, 13892, 13, 267, 21558, 32000, 28705, 13,  # gpt
+            32001, 2188, 13, 12684, 17664, 32000, 28705, 13,   # human
+            32001, 13892, 13, 12684, 17664, 32000, 28705, 13,  # gpt
+        ]
+        # fmt: on
+
+    def test_no_train_on_input(self, sharegpt_dataset, tokenizer):
+        strategy = SimpleShareGPTPromptTokenizingStrategy(
+            ShareGPTPrompterV2(
+                conversation="chatml",
+                role_key_model=None,
+                role_key_human=None,
+            ),
+            tokenizer,
+            False,  # train_on_inputs
+            2048,  # sequence_len
+        )
+
+        dataset_wrapper = TokenizedPromptDataset(
+            strategy, sharegpt_dataset, process_count=1
+        )
+
+        labels = dataset_wrapper[0]["labels"]
+        # fmt: off
+        assert labels == [
+            -100,   # bos
+            -100, -100, -100, -100, -100, -100, -100,  # system
+            -100, -100, -100, -100, -100, -100, -100,  # human
+            -100, -100, 13, 21558, 32000, 28705, 13,  # gpt
+            -100, -100, -100, -100, -100, -100, -100, -100,   # human
+            -100, -100, 13, 12684, 17664, 32000, 28705, 13,  # gpt
+        ]
+        # fmt: on
+
+    def test_no_train_on_input_with_weights(
+        self, sharegpt_dataset_with_weights, tokenizer
+    ):
+        strategy = SimpleShareGPTPromptTokenizingStrategy(
+            ShareGPTPrompterV2(
+                conversation="chatml",
+                role_key_model=None,
+                role_key_human=None,
+            ),
+            tokenizer,
+            False,  # train_on_inputs
+            2048,  # sequence_len
+        )
+
+        dataset_wrapper = TokenizedPromptDataset(
+            strategy, sharegpt_dataset_with_weights, process_count=1
+        )
+
+        labels = dataset_wrapper[0]["labels"]
+        # fmt: off
+        assert labels == [
+            -100,   # bos
+            -100, -100, -100, -100, -100, -100, -100,  # system
+            -100, -100, -100, -100, -100, -100, -100,  # human
+            -100, -100, -100, -100, -100, -100, -100,  # gpt with weight zero
+            -100, -100, -100, -100, -100, -100, -100, -100,  # human
+            -100, -100, 13, 267, 21558, 32000, 28705, 13,  # gpt
+            -100, -100, -100, -100, -100, -100, -100, -100,   # human
+            -100, -100, -100, -100, -100, -100, -100, -100  # gpt with weight zero
+        ]
+        # fmt: on
+
+    def test_w_train_on_input(self, sharegpt_dataset, tokenizer):
+        strategy = SimpleShareGPTPromptTokenizingStrategy(
+            ShareGPTPrompterV2(
+                conversation="chatml",
+                role_key_model=None,
+                role_key_human=None,
+            ),
+            tokenizer,
+            True,  # train_on_inputs
+            2048,  # sequence_len
+        )
+
+        dataset_wrapper = TokenizedPromptDataset(
+            strategy, sharegpt_dataset, process_count=1
+        )
+
+        labels = dataset_wrapper[0]["labels"]
+        # fmt: off
+        assert labels == [
+            1,   # bos
+            32001, 1587, 13, 25997, 32000, 28705, 13,  # system
+            32001, 2188, 13, 21558, 32000, 28705, 13,  # human
+            32001, 13892, 13, 21558, 32000, 28705, 13,  # gpt
+            32001, 2188, 13, 12684, 17664, 32000, 28705, 13,   # human
+            32001, 13892, 13, 12684, 17664, 32000, 28705, 13,  # gpt
+        ]
+        # fmt: on
+
+    def test_w_train_on_input_with_weights(
+        self, sharegpt_dataset_with_weights, tokenizer
+    ):
+        strategy = SimpleShareGPTPromptTokenizingStrategy(
+            ShareGPTPrompterV2(
+                conversation="chatml",
+                role_key_model=None,
+                role_key_human=None,
+            ),
+            tokenizer,
+            True,  # train_on_inputs
+            2048,  # sequence_len
+        )
+
+        dataset_wrapper = TokenizedPromptDataset(
+            strategy, sharegpt_dataset_with_weights, process_count=1
+        )
+
+        labels = dataset_wrapper[0]["labels"]
+        # fmt: off
+        assert labels == [
+            1,   # bos
+            32001, 1587, 13, 25997, 32000, 28705, 13,  # system
+            32001, 2188, 13, 21558, 32000, 28705, 13,  # human
+            -100, -100, -100, -100, -100, -100, -100,  # gpt with weight 0
+            -100, -100, -100, -100, -100, -100, -100, -100,  # human with weight 0
+            32001, 13892, 13, 267, 21558, 32000, 28705, 13,  # gpt
+            32001, 2188, 13, 12684, 17664, 32000, 28705, 13,  # human
+            -100, -100, -100, -100, -100, -100, -100, -100  # gpt with weight 0
+        ]
+        # fmt: on
+
+    def test_chatml_glaive(self, glaive_dataset, tokenizer):
+        strategy = GlaiveShareGPTPromptTokenizingStrategy(
+            ShareGPTPrompterV2(
+                conversation="chatml",
+                role_key_model=None,
+                role_key_human=None,
+            ),
+            tokenizer,
+            True,  # train_on_inputs
+            2048,  # sequence_len
+        )
+
+        dataset_wrapper = TokenizedPromptDataset(
+            strategy, glaive_dataset, process_count=1
+        )
+
+        labels = dataset_wrapper[0]["labels"]
+        # fmt: off
+        assert labels == [
+            1,  # bos
+            32001, 1587, 13, 3260, 349, 264, 1587, 11510, 32000, 28705, 13,  # system
+            32001, 2188, 13, 6325, 368, 1820, 264, 9314, 354, 528, 477, 1450, 2726, 298, 4222, 28804, 32000, 28705, 13,  # human
+            32001, 13892, 13, 28737, 28742, 28719, 7371, 28725, 562, 315, 949, 28742, 28707, 506, 272, 21368, 298, 1820, 22447, 28723, 28705, 523, 28766, 416, 1009, 772, 28766, 28767, 32000, 28705, 13  # gpt
+        ]
+        # fmt: on
+
+    def test_multi_role_dataset(self, multi_role_dataset, tokenizer):
+        strategy = SimpleShareGPTPromptTokenizingStrategy(
+            ShareGPTPrompterV2(conversation="chatml", roles={"input": ["tool"]}),
+            tokenizer,
+            False,  # train_on_inputs
+            2048,  # sequence_len
+        )
+
+        dataset_wrapper = TokenizedPromptDataset(
+            strategy, multi_role_dataset, process_count=1
+        )
+
+        input_ids = dataset_wrapper[0]["input_ids"]
+        # fmt: off
+        assert input_ids == [
+            1,   # bos
+            32001, 1587, 13, 1730, 625, 28730, 769, 1223, 28732, 18373, 28731, 298, 625, 272, 8086, 354, 264, 2990, 32000, 28705, 13,  # system
+            32001, 2188, 13, 21558, 28725, 767, 28742, 28713, 272, 8086, 297, 1450, 2726, 28804, 32000, 28705, 13,  # human
+            32001, 13892, 13, 895, 528, 625, 369, 354, 368, 32000, 28705, 13,  # gpt
+            32001, 3921, 13, 527, 28730, 769, 1223, 28732, 2972, 2726, 28731, 32000, 28705, 13,  # tool
+            32001, 13892, 13, 1237, 8086, 297, 1450, 2726, 349, 28705, 28787, 28734, 11182, 304, 4376, 1780, 32000, 28705, 13  # gpt
+        ]
+        # fmt: on
+
+        labels = dataset_wrapper[0]["labels"]
+        # fmt: off
+        assert labels == [
+            -100,  # bos
+            -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,  # system
+            -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,  # human
+            -100, -100, 13, 895, 528, 625, 369, 354, 368, 32000, 28705, 13,  # gpt
+            -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100,  # tool
+            -100, -100, 13, 1237, 8086, 297, 1450, 2726, 349, 28705, 28787, 28734, 11182, 304, 4376, 1780, 32000, 28705, 13  # gpt
+        ]
+        # fmt: on

tests/test_datasets.py

@@ -306,10 +306,6 @@ class TestDatasetPreparation(unittest.TestCase):
         """Verify that processing data from the hub works with a specific revision"""
         with tempfile.TemporaryDirectory() as tmp_dir:
             prepared_path = Path(tmp_dir) / "prepared"
-
-            # make sure prepared_path is empty
-            shutil.rmtree(prepared_path, ignore_errors=True)
-
             cfg = DictDefault(
                 {
                     "tokenizer_config": "huggyllama/llama-7b",
@@ -371,44 +367,43 @@ class TestDatasetPreparation(unittest.TestCase):
     def test_load_local_hub_with_revision(self):
         """Verify that a local copy of a hub dataset can be loaded with a specific revision"""
         with tempfile.TemporaryDirectory() as tmp_dir:
-            with tempfile.TemporaryDirectory() as tmp_dir2:
-                tmp_ds_path = Path(tmp_dir2) / "mhenrichsen/alpaca_2k_test"
-                tmp_ds_path.mkdir(parents=True, exist_ok=True)
-                snapshot_download(
-                    repo_id="mhenrichsen/alpaca_2k_test",
-                    repo_type="dataset",
-                    local_dir=tmp_ds_path,
-                    revision="d05c1cb",
-                )
+            tmp_ds_path = Path("mhenrichsen/alpaca_2k_test")
+            tmp_ds_path.mkdir(parents=True, exist_ok=True)
+            snapshot_download(
+                repo_id="mhenrichsen/alpaca_2k_test",
+                repo_type="dataset",
+                local_dir=tmp_ds_path,
+                revision="d05c1cb",
+            )
 
-                prepared_path = Path(tmp_dir) / "prepared"
-                cfg = DictDefault(
-                    {
-                        "tokenizer_config": "huggyllama/llama-7b",
-                        "sequence_len": 1024,
-                        "datasets": [
-                            {
-                                "path": "mhenrichsen/alpaca_2k_test",
-                                "ds_type": "parquet",
-                                "type": "alpaca",
-                                "data_files": [
-                                    f"{tmp_ds_path}/alpaca_2000.parquet",
-                                ],
-                                "revision": "d05c1cb",
-                            },
-                        ],
-                    }
-                )
+            prepared_path = Path(tmp_dir) / "prepared"
+            cfg = DictDefault(
+                {
+                    "tokenizer_config": "huggyllama/llama-7b",
+                    "sequence_len": 1024,
+                    "datasets": [
+                        {
+                            "path": "mhenrichsen/alpaca_2k_test",
+                            "ds_type": "parquet",
+                            "type": "alpaca",
+                            "data_files": [
+                                "mhenrichsen/alpaca_2k_test/alpaca_2000.parquet",
+                            ],
+                            "revision": "d05c1cb",
+                        },
+                    ],
+                }
+            )
 
-                dataset, _ = load_tokenized_prepared_datasets(
-                    self.tokenizer, cfg, prepared_path
-                )
+            dataset, _ = load_tokenized_prepared_datasets(
+                self.tokenizer, cfg, prepared_path
+            )
 
-                assert len(dataset) == 2000
-                assert "input_ids" in dataset.features
-                assert "attention_mask" in dataset.features
-                assert "labels" in dataset.features
-                shutil.rmtree(tmp_ds_path)
+            assert len(dataset) == 2000
+            assert "input_ids" in dataset.features
+            assert "attention_mask" in dataset.features
+            assert "labels" in dataset.features
+            shutil.rmtree(tmp_ds_path)
 
 
 if __name__ == "__main__":

tests/test_normalize_config.py

@@ -39,12 +39,12 @@ class NormalizeConfigTestCase(unittest.TestCase):
                 "datasets": [
                     {
                         "path": "lorem/ipsum",
-                        "type": "chat_template",
-                        "chat_template": "gemma",
+                        "type": "sharegpt",
+                        "conversation": "vicuna_v1.1",
                     },
                     {
                         "path": "sit/amet",
-                        "type": "chat_template",
+                        "type": "sharegpt",
                     },
                 ],
             }
@@ -52,8 +52,8 @@ class NormalizeConfigTestCase(unittest.TestCase):
 
         normalize_cfg_datasets(cfg)
 
-        assert cfg.datasets[0].chat_template == "gemma"
-        assert cfg.datasets[1].chat_template == "chatml"
+        assert cfg.datasets[0].conversation == "vicuna_v1.1"
+        assert cfg.datasets[1].conversation == "chatml"
 
     @patch("axolotl.utils.config.is_torch_bf16_gpu_available")
     def test_bf16_auto_setter_available(self, mock_bf16_avail):

tests/test_packed_pretraining.py

@@ -2,7 +2,6 @@
 import functools
 import unittest
 
-import pytest
 import torch
 from datasets import load_dataset
 from torch.utils.data import DataLoader
@@ -22,7 +21,6 @@ class TestPretrainingPacking(unittest.TestCase):
         self.tokenizer = AutoTokenizer.from_pretrained("huggyllama/llama-7b")
         self.tokenizer.pad_token = "</s>"
 
-    @pytest.mark.flaky(retries=3, delay=5)
     def test_packing_stream_dataset(self):
         # pylint: disable=duplicate-code
         dataset = load_dataset(