set download to use revision

.github/workflows/base.yml

@@ -24,12 +24,6 @@ jobs:
             python_version: "3.11"
             pytorch: 2.3.1
             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: "3.10"
-            pytorch: 2.4.1
-            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
           - cuda: "124"
             cuda_version: 12.4.1
             cudnn_version: ""
@@ -40,25 +34,23 @@ jobs:
             cuda_version: 12.4.1
             cudnn_version: ""
             python_version: "3.11"
-            pytorch: 2.5.1
+            pytorch: 2.4.1
             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

@@ -4,13 +4,11 @@ on:
   push:
     branches:
       - "main"
-    tags:
-      - "v*"
   workflow_dispatch:
 
 jobs:
   build-axolotl:
-    if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
+    if: ${{ ! contains(github.event.commits[0].message, '[skip docker]]') && github.repository_owner == 'axolotl-ai-cloud' }}
     strategy:
       fail-fast: false
       matrix:
@@ -31,11 +29,6 @@ jobs:
             python_version: "3.11"
             pytorch: 2.4.1
             axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.1
-            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -44,12 +37,7 @@ jobs:
         id: metadata
         uses: docker/metadata-action@v5
         with:
-          images: |
-            winglian/axolotl
-            axolotlai/axolotl
-          tags: |
-            type=ref,event=branch
-            type=pep440,pattern={{version}}
+          images: winglian/axolotl
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
       - name: Login to Docker Hub
@@ -63,7 +51,7 @@ jobs:
         with:
           context: .
           build-args: |
-            BASE_TAG=${{ 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 }}
@@ -77,7 +65,7 @@ jobs:
 
   build-axolotl-cloud:
     needs: build-axolotl
-    if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
+    if: ${{ ! contains(github.event.commits[0].message, '[skip docker]]') && github.repository_owner == 'axolotl-ai-cloud' }}
     # this job needs to be run on self-hosted GPU runners...
     strategy:
       matrix:
@@ -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
@@ -111,25 +94,20 @@ jobs:
         id: metadata
         uses: docker/metadata-action@v5
         with:
-          images: |
-            winglian/axolotl-cloud
-            axolotlai/axolotl-cloud
-          tags: |
-            type=ref,event=branch
-            type=pep440,pattern={{version}}
+          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' }}
@@ -140,7 +118,7 @@ jobs:
 
   build-axolotl-cloud-no-tmux:
     needs: build-axolotl
-    if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
+    if: ${{ ! contains(github.event.commits[0].message, '[skip docker]]') && github.repository_owner == 'axolotl-ai-cloud' }}
     # this job needs to be run on self-hosted GPU runners...
     strategy:
       matrix:
@@ -158,25 +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=pep440,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

@@ -8,14 +8,9 @@ on:
   schedule:
     - cron: '0 0 * * 1,4'  # Runs at 00:00 UTC every monday & thursday
 
-# 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:
   test-axolotl-multigpu:
-    if: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' }}
+    if: ${{ ! contains(github.event.commits[0].message, '[skip docker]]') && github.repository_owner == 'axolotl-ai-cloud' }}
     strategy:
       fail-fast: false
       matrix:
@@ -26,17 +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

@@ -7,7 +7,7 @@ on:
 
 jobs:
   build-axolotl:
-    if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
+    if: ${{ ! contains(github.event.commits[0].message, '[skip docker]]') && github.repository_owner == 'axolotl-ai-cloud' }}
     strategy:
       fail-fast: false
       matrix:
@@ -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
@@ -71,7 +64,7 @@ jobs:
 
   build-axolotl-cloud:
     needs: build-axolotl
-    if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
+    if: ${{ ! contains(github.event.commits[0].message, '[skip docker]]') && github.repository_owner == 'axolotl-ai-cloud' }}
     # this job needs to be run on self-hosted GPU runners...
     strategy:
       matrix:
@@ -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

@@ -3,24 +3,12 @@ name: publish pypi
 on:
   push:
     tags:
-      - 'v*'
-  workflow_dispatch:
+      - '*'
 
 jobs:
-  setup_release:
-    name: Create Release
-    runs-on: ubuntu-latest
-    permissions:
-      contents: write
-    steps:
-      - name: Create release
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: gh release create "$GITHUB_REF_NAME" # GITHUB_REF_NAME is the tag name in `on.push.tags` workflows
   pypi-publish:
     name: Upload release to PyPI
     runs-on: ubuntu-latest
-    needs: [setup_release]
     environment:
       name: pypi
       url: https://pypi.org/p/axolotl
@@ -28,10 +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"
 
@@ -39,7 +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
@@ -49,9 +37,9 @@ jobs:
         run: |
           sed -i -E 's/version="([0-9.]+)",/version="${{ steps.tag.outputs.TAG_NAME }}",/g' setup.py
 
-      - name: Build a source dist
+      - name: Build a binary wheel
         run: |
-          python setup.py sdist
+          python setup.py sdist bdist_wheel
 
       - name: Publish package distributions to PyPI
         uses: pypa/gh-action-pypi-publish@release/v1

.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
@@ -47,15 +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
-          sed -i 's#^datasets.*#datasets @ git+https://github.com/huggingface/datasets.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: |
@@ -83,17 +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
 
@@ -41,97 +36,61 @@ 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: |
-          pytest -n8 --ignore=tests/e2e/ tests/
+          pytest --ignore=tests/e2e/ tests/
 
       - name: cleanup pip cache
         run: |
           find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
 
-  pytest-sdist:
-    name: PyTest from Source Dist
-    runs-on: ubuntu-latest
-    strategy:
-      fail-fast: false
-      matrix:
-        python_version: ["3.11"]
-        pytorch_version: ["2.4.1", "2.5.1"]
-    timeout-minutes: 20
-
-    steps:
-      - name: Check out repository code
-        uses: actions/checkout@v4
-
-      - name: Setup Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: ${{ matrix.python_version }}
-          cache: 'pip' # caching pip dependencies
-
-      - name: upgrade pip
-        run: |
-          pip3 install --upgrade pip
-          pip3 install --upgrade packaging setuptools wheel
-
-      - name: Install PyTorch
-        run: |
-          pip3 install torch==${{ matrix.pytorch_version }}
-
-      - name: Install dependencies
-        run: |
-          pip3 show torch
-          python3 setup.py sdist
-          pip3 install dist/axolotl*.tar.gz
-          pip3 install -r requirements-dev.txt -r requirements-tests.txt
-
-      - name: Run tests
-        run: |
-          pytest -n8 --ignore=tests/e2e/ tests/
-
-      - name: cleanup pip cache
-        run: |
-          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
-
-  docker-e2e-tests-1st:
-    if: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' }}
+  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, pytest-sdist]
+    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"
@@ -160,49 +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

.gitignore

@@ -182,6 +182,3 @@ submit.sh
 
 typings/
 out/
-
-# vim
-*.swp

MANIFEST.in

@@ -1,4 +0,0 @@
-include requirements.txt
-include README.md
-include LICENSE
-recursive-include axolotl *.py

README.md

@@ -1,21 +1,8 @@
-<p align="center">
-    <picture>
-        <source media="(prefers-color-scheme: dark)" srcset="image/axolotl_logo_digital_white.svg">
-        <source media="(prefers-color-scheme: light)" srcset="image/axolotl_logo_digital_black.svg">
-        <img alt="Axolotl" src="image/axolotl_logo_digital_black.svg" width="400" height="104" style="max-width: 100%;">
-    </picture>
-</p>
+# Axolotl
 
-<p align="center">
-    <img src="https://img.shields.io/github/license/axolotl-ai-cloud/axolotl.svg?color=blue" alt="GitHub License">
-    <img src="https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/tests.yml/badge.svg" alt="tests">
-    <a href="https://github.com/axolotl-ai-cloud/axolotl/releases"><img src="https://img.shields.io/github/release/axolotl-ai-cloud/axolotl.svg" alt="Releases"></a>
-    <img src="https://img.shields.io/github/stars/axolotl-ai-cloud/axolotl" alt="GitHub Repo stars">
-</p>
-<p align="center">
-    <img src="https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/tests-nightly.yml/badge.svg" alt="tests-nightly">
-    <img src="https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/multi-gpu-e2e.yml/badge.svg" alt="multigpu-semi-weekly tests">
-</p>
+![tests](https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/tests.yml/badge.svg)
+![tests-nightly](https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/tests-nightly.yml/badge.svg)
+![multigpu-semi-weekly tests](https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/multi-gpu-e2e.yml/badge.svg)
 
 Axolotl is a tool designed to streamline the fine-tuning of various AI models, offering support for multiple configurations and architectures.
 
@@ -88,7 +75,7 @@ Features:
 <td>
 
 <div align="center">
-  <img src="image/axolotl_symbol_digital_white.svg" alt="axolotl" width="160">
+  <img src="image/axolotl.png" alt="axolotl" width="160">
   <div>
     <p>
       <b>Axolotl provides a unified repository for fine-tuning <br />a variety of AI models with ease</b>
@@ -134,7 +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
@@ -172,7 +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:
@@ -191,7 +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:
@@ -223,7 +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)
@@ -332,7 +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
@@ -396,10 +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
+      # 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
@@ -574,8 +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 }}"
@@ -23,12 +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; \
-        sed -i 's#^datasets.*#datasets @ git+https://github.com/huggingface/datasets.git@main#' requirements.txt; \
     fi
 
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
@@ -38,7 +37,7 @@ RUN if [ "$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 -n8 --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/multigpu.sh

@@ -2,4 +2,4 @@
 set -e
 
 # only run one test at a time so as not to OOM the GPU
-pytest -v -n2 /workspace/axolotl/tests/e2e/multigpu/
+pytest -n1 /workspace/axolotl/tests/e2e/multigpu/

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
@@ -91,7 +91,15 @@ datasets:
     name: # Optional[str] name of dataset configuration to load
     train_on_split: train # Optional[str] name of dataset split to load from
     revision: # Optional[str] The specific revision of the dataset to use when loading from the Hugging Face Hub. This can be a commit hash, tag, or branch name. If not specified, the latest version will be used. This parameter is ignored for local datasets.
-    trust_remote_code: # Optional[bool] Trust remote code for untrusted source
+
+    # 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
@@ -116,48 +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
@@ -175,19 +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
@@ -309,7 +266,6 @@ wandb_log_model: # "checkpoint" to log model to wandb Artifacts every `save_step
 # mlflow configuration if you're using it
 mlflow_tracking_uri: # URI to mlflow
 mlflow_experiment_name: # Your experiment name
-mlflow_run_name: # Your run name
 hf_mlflow_log_artifacts:  # set to true to copy each saved checkpoint on each save to mlflow artifact registry
 
 # Comet configuration if you're using it
@@ -358,7 +314,7 @@ max_steps:
 
 eval_table_size: # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0
 eval_max_new_tokens: # Total number of tokens generated for predictions sent to wandb. Default is 128
-eval_causal_lm_metrics: # HF evaluate metrics used during evaluation. Default is ["sacrebleu", "comet", "ter", "chrf", "perplexity"]
+eval_causal_lm_metrics: # HF evaluate metrics used during evaluation. Default is ["sacrebleu", "comet", "ter", chrf]
 
 loss_watchdog_threshold: # High loss value, indicating the learning has broken down (a good estimate is ~2 times the loss at the start of training)
 loss_watchdog_patience: # Number of high-loss steps in a row before the trainer aborts (default: 3)
@@ -406,7 +362,6 @@ 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).

docs/unsloth.qmd

@@ -11,10 +11,12 @@ standard industry baselines.
 
 ### Installation
 
-The following will install the correct unsloth and extras from source.
+The following will install unsloth from source and downgrade xformers as unsloth is incompatible with the most up
+to date libraries.
 
 ```bash
-python scripts/unsloth_install.py | sh
+pip install --no-deps "unsloth @ git+https://github.com/unslothai/unsloth.git"
+pip install --no-deps --force-reinstall xformers==0.0.26.post1
 ```
 
 ### Using unsloth w Axolotl

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

@@ -2,15 +2,19 @@
  "cells": [
   {
    "cell_type": "markdown",
-   "metadata": {},
+   "metadata": {
+    "id": "AKjdG7tbTb-n"
+   },
    "source": [
-    "## Setting up"
+    "# Example notebook for running Axolotl on google colab"
    ]
   },
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {},
+   "metadata": {
+    "id": "RcbNpOgWRcii"
+   },
    "outputs": [],
    "source": [
     "import torch\n",
@@ -18,76 +22,82 @@
     "assert (torch.cuda.is_available()==True)"
    ]
   },
+  {
+   "cell_type": "markdown",
+   "metadata": {
+    "id": "h3nLav8oTRA5"
+   },
+   "source": [
+    "## Install Axolotl and dependencies"
+   ]
+  },
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {},
+   "metadata": {
+    "colab": {
+     "base_uri": "https://localhost:8080/"
+    },
+    "id": "3c3yGAwnOIdi",
+    "outputId": "e3777b5a-40ef-424f-e181-62dfecd1dd01"
+   },
    "outputs": [],
    "source": [
-    "!pip install axolotl[deepspeed]"
+    "!pip install -e git+https://github.com/axolotl-ai-cloud/axolotl#egg=axolotl\n",
+    "!pip install flash-attn==\"2.5.0\"\n",
+    "!pip install deepspeed==\"0.13.1\"!pip install mlflow==\"2.13.0\""
    ]
   },
   {
    "cell_type": "markdown",
-   "metadata": {},
+   "metadata": {
+    "id": "BW2MFr7HTjub"
+   },
    "source": [
-    "## Hugging Face login (optional)"
+    "## Create an yaml config file"
    ]
   },
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "from huggingface_hub import notebook_login\n",
-    "notebook_login()"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## Example configuration"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
+   "metadata": {
+    "id": "9pkF2dSoQEUN"
+   },
    "outputs": [],
    "source": [
     "import yaml\n",
     "\n",
+    "# Your YAML string\n",
     "yaml_string = \"\"\"\n",
-    "base_model: NousResearch/Meta-Llama-3.1-8B\n",
+    "base_model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T\n",
+    "model_type: LlamaForCausalLM\n",
+    "tokenizer_type: LlamaTokenizer\n",
     "\n",
     "load_in_8bit: false\n",
     "load_in_4bit: true\n",
     "strict: false\n",
     "\n",
     "datasets:\n",
-    "  - path: tatsu-lab/alpaca\n",
+    "  - path: mhenrichsen/alpaca_2k_test\n",
     "    type: alpaca\n",
-    "dataset_prepared_path: last_run_prepared\n",
+    "dataset_prepared_path:\n",
     "val_set_size: 0.05\n",
-    "output_dir: ./outputs/lora-out\n",
-    "\n",
-    "sequence_len: 2048\n",
-    "sample_packing: true\n",
-    "eval_sample_packing: true\n",
-    "pad_to_sequence_len: true\n",
+    "output_dir: ./outputs/qlora-out\n",
     "\n",
     "adapter: qlora\n",
     "lora_model_dir:\n",
+    "\n",
+    "sequence_len: 4096\n",
+    "sample_packing: true\n",
+    "eval_sample_packing: false\n",
+    "pad_to_sequence_len: true\n",
+    "\n",
     "lora_r: 32\n",
     "lora_alpha: 16\n",
     "lora_dropout: 0.05\n",
+    "lora_target_modules:\n",
     "lora_target_linear: true\n",
     "lora_fan_in_fan_out:\n",
-    "lora_modules_to_save:\n",
-    "  - embed_tokens\n",
-    "  - lm_head\n",
     "\n",
     "wandb_project:\n",
     "wandb_entity:\n",
@@ -95,12 +105,12 @@
     "wandb_name:\n",
     "wandb_log_model:\n",
     "\n",
-    "gradient_accumulation_steps: 2\n",
-    "micro_batch_size: 1\n",
-    "num_epochs: 1\n",
-    "optimizer: paged_adamw_8bit\n",
+    "gradient_accumulation_steps: 4\n",
+    "micro_batch_size: 2\n",
+    "num_epochs: 4\n",
+    "optimizer: paged_adamw_32bit\n",
     "lr_scheduler: cosine\n",
-    "learning_rate: 2e-5\n",
+    "learning_rate: 0.0002\n",
     "\n",
     "train_on_inputs: false\n",
     "group_by_length: false\n",
@@ -111,15 +121,13 @@
     "gradient_checkpointing: true\n",
     "early_stopping_patience:\n",
     "resume_from_checkpoint:\n",
+    "local_rank:\n",
     "logging_steps: 1\n",
     "xformers_attention:\n",
-    "flash_attention: false\n",
-    "sdp_attention: true\n",
+    "flash_attention: true\n",
     "\n",
-    "warmup_steps: 1\n",
-    "max_steps: 25\n",
-    "evals_per_epoch: 1\n",
-    "eval_table_size:\n",
+    "warmup_steps: 10\n",
+    "evals_per_epoch: 4\n",
     "saves_per_epoch: 1\n",
     "debug:\n",
     "deepspeed:\n",
@@ -127,9 +135,8 @@
     "fsdp:\n",
     "fsdp_config:\n",
     "special_tokens:\n",
-    "  pad_token: <|end_of_text|>\n",
-    "\"\"\"\n",
     "\n",
+    "\"\"\"\n",
     "\n",
     "# Convert the YAML string to a Python dictionary\n",
     "yaml_dict = yaml.safe_load(yaml_string)\n",
@@ -139,124 +146,31 @@
     "\n",
     "# Write the YAML file\n",
     "with open(file_path, 'w') as file:\n",
-    "    yaml.dump(yaml_dict, file)"
+    "    yaml.dump(yaml_dict, file)\n"
    ]
   },
   {
    "cell_type": "markdown",
-   "metadata": {},
+   "metadata": {
+    "id": "bidoj8YLTusD"
+   },
    "source": [
-    "Above we have a configuration file with base LLM model and datasets specified, among many other things. Axolotl can automatically detect whether the specified datasets are on HuggingFace repo or local machine.\n",
-    "\n",
-    "The Axolotl configuration options encompass model and dataset selection, data pre-processing, and training. Let's go through them line by line:\n",
-    "\n",
-    "*   \"base model\": String value, specifies the underlying pre-trained LLM that will be used for finetuning\n",
-    "\n",
-    "Next we have options for model weights quantization. Quantization allows for reduction in occupied memory on GPUs.\n",
-    "\n",
-    "*   \"load_in_8bit\": Boolean value, whether to quantize the model weights into 8-bit integer.\n",
-    "\n",
-    "*   \"load_in_4bit\": Boolean value, whether to quantize the model weights into 4-bit integer.\n",
-    "\n",
-    "*   \"strict\": Boolean value. If false, it allows for overriding established configuration options in the yaml file when executing in command-line interface.\n",
-    "\n",
-    "*   \"datasets\": a list of dicts that contain path and type of data sets as well as other optional configurations where datasets are concerned. Supports multiple datasets.\n",
-    "\n",
-    "*   \"val_set_size\": Either a float value less than one or an integer less than the total size of dataset. Sets the size of validation set from the whole dataset. If float, sets the proportion of the dataset assigned for validation. If integer, sets the direct size of validation set.\n",
-    "\n",
-    "*   \"output_dir\": String value. Path of trained model.\n",
-    "\n",
-    "For data preprocessing:\n",
-    "\n",
-    "*   \"sequence_len\": Integer. Specifies the maximum sequence length of the input. Typically 2048 or less.\n",
-    "\n",
-    "*   \"pad_to_sequence_len\": Boolean. Padding input to maximum sequence length.\n",
-    "\n",
-    "*   \"sample_packing\": Boolean. Specifies whether to use multi-packing with block diagonal attention.\n",
-    "\n",
-    "*   \"special_tokens\": Python dict, optional. Allows users to specify the additional special tokens to be ignored by the tokenizer.\n",
-    "\n",
-    "For LoRA configuration and its hyperparamters:\n",
-    "\n",
-    "*   \"adapter\": String. Either \"lora\" or \"qlora\", depending on user's choice.\n",
-    "\n",
-    "*   \"lora_model_dir\": String, Optional. Path to directory that contains LoRA model, if there is already a trained LoRA model the user would like to use.\n",
-    "\n",
-    "*   \"lora_r\": Integer. Refers to the rank of LoRA decomposition matrices. Higher value will reduce LoRA efficiency. Recommended to be set to 8.\n",
-    "\n",
-    "*   \"lora_alpha\": Integer. Scale the weight matrices by $\\frac{\\text{lora_alpha}}{\\text{lora_r}}$Recommended to be fixed at 16.\n",
-    "\n",
-    "*   \"lora_dropout\": Float that is 1 or less. The dropout probability of a lora layer.\n",
-    "\n",
-    "*   \"lora_target_linear\": Boolean. If true, lora will target all linear modules in the transformers architecture.\n",
-    "\n",
-    "*   \"lora_modules_to_save\": If you added new tokens to the tokenizer, you may need to save some LoRA modules because they need to know the new tokens.\n",
-    "\n",
-    "See [LoRA](https://arxiv.org/abs/2106.09685) for detailed explanation of LoRA implementation.\n",
-    "\n",
-    "For the training configurations:\n",
-    "\n",
-    "*   \"gradient_accumulation_steps\": Integer. The number of steps over which to accumulate gradient for batch training. E.g. if 2, backprop is performed every two steps.\n",
-    "\n",
-    "*   \"micro_batch_size\": Integer. Batch size per gpu / gradient_accumulation_steps\n",
-    "\n",
-    "*   \"num_epochs\": Integer. Number of epochs. One epoch is when training has looped over every batch in the whole data set once.\n",
-    "\n",
-    "*   \"optimizer\": The optimizer to use for the training.\n",
-    "\n",
-    "*   \"learning_rate\": The learning rate.\n",
-    "\n",
-    "*   \"lr_scheduler\": The learning rate scheduler to use for adjusting learning rate during training.\n",
-    "\n",
-    "*   \"train_on_inputs\": Boolean. Whether to ignore or include the user's prompt from the training labels.\n",
-    "\n",
-    "*   \"group_by_length\": Boolean. Whether to group similarly sized data to minimize padding.\n",
-    "\n",
-    "*   \"bf16\": Either \"auto\", \"true\", or \"false\". Whether to use CUDA bf16 floating point format. If set to \"auto\", will automatically apply bf16 should the gpu supports it.\n",
-    "\n",
-    "*   \"fp16\": Optional. Specifies whether to use CUDA fp16. Automatically set to true if \"bf16\" is set to true. Otherwise false.\n",
-    "\n",
-    "*   \"tf32\": Boolean. Whether to use CUDA tf32. Will override bf16.\n",
-    "\n",
-    "*   \"gradient_checkpointing\": Boolean. Whether to use gradient checkpointing https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing\n",
-    "\n",
-    "*   \"gradient_checkpointing_kwargs\": Python Dict. Fed into the trainer.\n",
-    "\n",
-    "*   \"logging_steps\": Integer. Log training information over every specified number of steps.\n",
-    "\n",
-    "*   \"flash_attention\": Boolean. Whether to use the [flash attention](https://github.com/Dao-AILab/flash-attention) mechanism.\n",
-    "\n",
-    "*   \"sdp_attention\": Boolean. Whether to use the Scaled Dot Product attention mechanism (the attention mechanism in the [original implementation](https://arxiv.org/abs/1706.03762) of transformers.)\n",
-    "\n",
-    "*   \"warmup_steps\": Integer. The number of pre-training steps where a very low learning rate is used.\n",
-    "\n",
-    "*   \"evals_per_epoch\": Integer. Number of evaluations to be performed within one training epoch.\n",
-    "\n",
-    "*   \"saves_per_epoch\": Integer. Number of times the model is saved in one training epoch.\n",
-    "\n",
-    "*   \"weight_decay\": Positive Float. Sets the \"strength\" of weight decay (i.e. setting the coefficient of L2 regularization)"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "The above is but a snippet aiming to get users familiarized with the types of streamlined configuration options axolotl provides. For a full list of configuration options, see [here](https://axolotl-ai-cloud.github.io/axolotl/docs/config.html)"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "Train the model"
+    "## Launch the training"
    ]
   },
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {},
+   "metadata": {
+    "colab": {
+     "base_uri": "https://localhost:8080/"
+    },
+    "id": "ydTI2Jk2RStU",
+    "outputId": "d6d0df17-4b53-439c-c802-22c0456d301b"
+   },
    "outputs": [],
    "source": [
+    "# By using the ! the comand will be executed as a bash command\n",
     "!accelerate launch -m axolotl.cli.train /content/test_axolotl.yaml"
    ]
   },
@@ -264,7 +178,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "Predict with trained model"
+    "## Play with inference"
    ]
   },
   {
@@ -273,85 +187,36 @@
    "metadata": {},
    "outputs": [],
    "source": [
+    "# By using the ! the comand will be executed as a bash command\n",
     "!accelerate launch -m axolotl.cli.inference /content/test_axolotl.yaml \\\n",
-    "    --lora_model_dir=\"./outputs/lora-out\" --gradio"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## Deeper Dive"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "It is also helpful to gain some familiarity over some of the core inner workings of axolotl"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## Configuration Normalization"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "Axolotl uses a custom Dict class, called ```DictDefault```\n",
-    "to store configurations specified in the yaml configuration file (into a Python variable named ```cfg```). The definition for this custom Dict can be found in the [utils/dict.py](https://github.com/axolotl-ai-cloud/axolotl/blob/main/src/axolotl/utils/dict.py)\n",
-    "\n",
-    "```DictDefault``` is amended such that calling a missing key from it will result in a ```None``` return type. This is important because if some configuration options aren't specified by the user, the ```None``` type allows Axolotl to perform boolean operations to determine the default settings for missing configurations. For more examples on how this is done, check out [utils/config/__init__.py](https://github.com/axolotl-ai-cloud/axolotl/blob/main/src/axolotl/utils/config/__init__.py)"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## Loading Models, Tokenizers, and Trainer"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "If we inspect [cli.train.py](https://github.com/axolotl-ai-cloud/axolotl/blob/main/src/axolotl/cli/train.py), we will find that most of the heavy lifting were done by the function ```train()``` which is itself imported from [src/axolotl/train.py](https://github.com/axolotl-ai-cloud/axolotl/blob/main/src/axolotl/train.py).\n",
-    "\n",
-    "```train()``` takes care of loading the appropriate tokenizer and pre-trained model through ```load_model()``` and ```load_tokenizer()``` from [src/axolotl/utils/models.py](https://github.com/axolotl-ai-cloud/axolotl/blob/main/src/axolotl/utils/models.py) respectively.\n",
-    "\n",
-    "```load_tokenizer()``` loads in the appropriate tokenizer given the desired model, as well as chat templates.\n",
-    "\n",
-    "```ModelLoader``` class follows after tokenizer has been selected. It will automatically discern the base model type, load in the desired model, as well as applying model-appropriate attention mechanism modifications (e.g. flash attention). Depending on which base model the user chooses in the configuration, ```ModelLoader``` will utilize the corresponding \"attention hijacking\" script. For example, if the user specified the base model to be ```NousResearch/Meta-Llama-3.1-8B```, which is of llama type, and set ```flash_attn``` to ```True```, ```ModelLoader``` will load in [llama_attn_hijack_flash.py](https://github.com/axolotl-ai-cloud/axolotl/blob/main/src/axolotl/monkeypatch/llama_attn_hijack_flash.py). For a list of supported attention hijacking, please refer to the directory [/src/axolotl/monkeypatch/](https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/monkeypatch)\n",
-    "\n",
-    "Another important operation encompassed in ```train()``` is setting up the training that takes into account of user-specified traning configurations (e.g. num_epochs, optimizer) through the use of ```setup_trainer()``` from [/src/axolotl/utils/trainer.py](https://github.com/axolotl-ai-cloud/axolotl/blob/main/src/axolotl/utils/trainer.py), which in turn relies on modules from [/src/axolotl/core/trainer_builder.py](https://github.com/axolotl-ai-cloud/axolotl/blob/main/src/axolotl/core/trainer_builder.py).\n",
-    "```trainer_builder.py``` provides a list of trainer object options bespoke for the task type (Causal or Reinforcement learning ('dpo', 'ipo', 'kto') )"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## Monkey patch\n",
-    "\n",
-    "The [Monkey patch directory](https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/monkeypatch) is where model architecture/optimization patching scripts are stored (these are modifications that are not implemented in the official releases, hence the name monkey patch). It includes attention jacking, ReLoRA, and unsloth optimization."
+    "    --qlora_model_dir=\"./qlora-out\" --gradio"
    ]
   }
  ],
  "metadata": {
+  "accelerator": "GPU",
+  "colab": {
+   "gpuType": "T4",
+   "provenance": []
+  },
   "kernelspec": {
-   "display_name": "Python 3",
+   "display_name": "Python 3 (ipykernel)",
    "language": "python",
    "name": "python3"
   },
   "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
    "name": "python",
-   "version": "3.9.6"
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.12.1"
   }
  },
  "nbformat": 4,
- "nbformat_minor": 2
+ "nbformat_minor": 4
 }

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/gemma2/reward-model.yaml

@@ -1,63 +0,0 @@
-base_model: google/gemma-2-2b
-model_type: AutoModelForSequenceClassification
-tokenizer_type: AutoTokenizer
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-reward_model: true
-chat_template: gemma
-datasets:
-  - path: argilla/distilabel-intel-orca-dpo-pairs
-    type: bradley_terry.chat_template
-val_set_size: 0.0
-output_dir: ./outputs/out
-remove_unused_columns: false
-
-sequence_len: 2048
-sample_packing: false
-eval_sample_packing: false
-pad_to_sequence_len: true
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-
-gradient_accumulation_steps: 4
-micro_batch_size: 2
-num_epochs: 4
-optimizer: adamw_bnb_8bit
-lr_scheduler: cosine
-learning_rate: 0.0002
-
-train_on_inputs: false
-group_by_length: false
-bf16: true
-fp16:
-tf32: true
-
-gradient_checkpointing: true
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-warmup_ratio: 0.1
-evals_per_epoch:
-eval_table_size:
-eval_max_new_tokens: 128
-saves_per_epoch: 1
-debug:
-deepspeed:
-weight_decay: 0.0
-fsdp:
-fsdp_config:
-special_tokens:

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-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/instruct-dpo-lora-8b.yml

@@ -11,6 +11,7 @@ rl: dpo
 datasets:
   - path: fozziethebeat/alpaca_messages_2k_dpo_test
     type: chat_template.default
+    chat_template: llama3
     field_messages: conversation
     field_chosen: chosen
     field_rejected: rejected

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

@@ -10,6 +10,7 @@ chat_template: llama3
 datasets:
   - path: fozziethebeat/alpaca_messages_2k_test
     type: chat_template
+    chat_template: llama3
     field_messages: messages
     message_field_role: role
     message_field_content: content

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

examples/qwen2/dpo.yaml

@@ -1,67 +0,0 @@
-base_model: Qwen/Qwen2.5-0.5B
-
-strict: false
-
-chat_template: qwen_25
-rl: dpo
-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
-
-dataset_prepared_path:
-val_set_size: 0.0
-output_dir: ./outputs/dpo-out
-
-sequence_len: 2048
-sample_packing: false
-pad_to_sequence_len: true
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 4
-micro_batch_size: 2
-num_epochs: 4
-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
-
-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:

image/axolotl_logo_digital_black.svg

@@ -1,19 +0,0 @@
-<?xml version="1.0" encoding="UTF-8"?>
-<svg xmlns="http://www.w3.org/2000/svg" version="1.1" viewBox="0 0 1113 283.5">
-    <path fill="#141310" d="M435,234.3l-12.1-48.8h-54.4l-12.1,48.8h-24.7l48.2-185.1h31.6l47.9,185.1h-24.5ZM417.7,164.9l-13.8-55.6c-2.7-10.7-4.8-19.7-6.3-26.9-.9-4.2-1.5-7.5-2-9.9-.5,2.5-1.2,5.8-2,9.9-1.5,7.1-3.6,16.1-6.3,26.7l-13.8,55.9h44.3Z"/>
-    <path fill="#141310" d="M568.2,234.3l-29.9-45.6c-1.2-1.9-2.4-4.1-3.5-6.5-.8-1.7-1.5-3.3-2.1-4.5-.6,1.3-1.4,2.8-2.3,4.5-1.3,2.4-2.6,4.6-4,6.5l-29.9,45.6h-28.5l49.6-71.9-46.5-67.9h28.5l27.6,43.1c1.2,1.9,2.3,3.9,3.4,6.1.7,1.4,1.4,2.7,1.9,3.8.5-1.1,1.1-2.4,1.8-3.8,1.1-2.2,2.2-4.2,3.4-6.1l27.6-43.1h28.5l-46.5,68.2,49.3,71.7h-28.5Z"/>
-    <path fill="#141310" d="M658.6,236.3c-16.7,0-30.2-5-40.1-14.8-9.9-9.8-14.9-23.7-14.9-41.3v-31.7c0-17.7,5-31.7,14.8-41.4,9.8-9.7,23.4-14.7,40.3-14.7s30.4,4.9,40.3,14.7c9.8,9.7,14.8,23.7,14.8,41.4v31.7c0,17.6-5,31.5-14.9,41.3-9.9,9.8-23.4,14.8-40.1,14.8ZM658.6,114.1c-9.5,0-17.1,2.7-22.6,8.1-5.5,5.4-8.3,13.4-8.3,23.8v36.7c0,10.5,2.8,18.5,8.3,23.8,5.5,5.4,13.1,8.1,22.6,8.1s17.3-2.7,22.7-8.1c5.4-5.4,8.2-13.4,8.2-23.9v-36.7c0-10.5-2.8-18.5-8.2-23.9-5.4-5.4-13.1-8.1-22.7-8.1Z"/>
-    <path fill="#141310" d="M860.6,236.3c-16.7,0-30.2-5-40.1-14.8-9.9-9.8-14.9-23.7-14.9-41.3v-31.7c0-17.7,5-31.7,14.8-41.4,9.8-9.7,23.4-14.7,40.3-14.7s30.4,4.9,40.3,14.7c9.8,9.7,14.8,23.7,14.8,41.4v31.7c0,17.6-5,31.5-14.9,41.3-9.9,9.8-23.4,14.8-40.1,14.8ZM860.6,114.1c-9.5,0-17.1,2.7-22.6,8.1-5.5,5.4-8.3,13.4-8.3,23.8v36.7c0,10.5,2.8,18.5,8.3,23.8,5.5,5.4,13.1,8.1,22.6,8.1s17.3-2.7,22.7-8.1c5.4-5.4,8.2-13.4,8.2-23.9v-36.7c0-10.5-2.8-18.5-8.2-23.9-5.4-5.4-13.1-8.1-22.7-8.1Z"/>
-    <path fill="#141310" d="M773.9,234c-18,0-32.6-14.6-32.6-32.6V48.8h24.1v152.6c0,4.7,3.8,8.5,8.5,8.5h16.8v24.1h-16.8Z"/>
-    <path fill="#141310" d="M1036.2,234.3V81.4c0-4.7-3.8-8.5-8.5-8.5h-16.8v-24.1h16.8c18,0,32.6,14.6,32.6,32.6v152.9h-24.1Z"/>
-    <path fill="#141310" d="M978.6,234.3c-18,0-32.6-14.6-32.6-32.6v-85.1h-20.3v-22.1h20.3v-45.3h24.1v45.3h30.2v22.1h-30.2v85.1c0,4.7,3.8,8.5,8.5,8.5h21.7v24.1h-21.7Z"/>
-    <path fill="#141310" d="M51.5,49h12.2v-20.6h-12.2c-16,0-29,13-29,29v32.8h20.6v-32.8c0-4.7,3.8-8.4,8.4-8.4Z"/>
-    <path fill="#141310" d="M92.8,49h12.2v-20.6h-12.2c-16,0-29,13-29,29v12.2h20.6v-12.2c0-4.7,3.8-8.4,8.4-8.4Z"/>
-    <path fill="#141310" d="M249.3,57.4c0-16-13-29-29-29h-12.2v20.6h12.2c4.7,0,8.4,3.8,8.4,8.4v32.8h20.6v-32.8Z"/>
-    <path fill="#141310" d="M187.4,90.2v-20.6h-103.1v20.6h-41.2v20.6h-20.6v41.2c0,11.4,9.2,20.6,20.6,20.6h185.5c11.4,0,20.6-9.2,20.6-20.6v-41.2h-20.6v-20.6h-41.2ZM166.8,141.7c0-5.7-4.6-10.3-10.3-10.3s-10.3,4.6-10.3,10.3v10.3h-20.6v-20.6c0-11.4,9.2-20.6,20.6-20.6s20.6,9.2,20.6,20.6v10.3ZM228.7,141.7c0-5.7-4.6-10.3-10.3-10.3s-10.3,4.6-10.3,10.3v10.3h-20.6v-20.6c0-11.4,9.2-20.6,20.6-20.6s20.6,9.2,20.6,20.6v10.3Z"/>
-    <path fill="#141310" d="M208,57.4c0-16-13-29-29-29h-12.2v20.6h12.2c4.7,0,8.4,3.8,8.4,8.4v12.2h20.6v-12.2Z"/>
-    <rect fill="#141310" x="22.5" y="234.5" width="41.2" height="20.6"/>
-    <rect fill="#141310" x="84.3" y="234.5" width="164.9" height="20.6"/>
-    <rect fill="#141310" x="208" y="193.3" width="41.2" height="20.6"/>
-    <rect fill="#141310" x="22.5" y="193.3" width="164.9" height="20.6"/>
-</svg>

image/axolotl_logo_digital_white.svg

@@ -1,11 +0,0 @@
-<?xml version="1.0" encoding="UTF-8"?>
-<svg xmlns="http://www.w3.org/2000/svg" version="1.1" viewBox="0 0 1113 283.5">
-    <path fill="#fff" d="M462.9,234.2l-12.1-48.8h-54.4l-12.1,48.8h-24.7l48.2-185h31.6l47.9,185h-24.4ZM445.7,164.8l-13.8-55.6c-2.7-10.7-4.8-19.7-6.3-26.9-.9-4.2-1.5-7.5-2-9.9-.5,2.5-1.2,5.8-2,9.9-1.5,7.1-3.6,16.1-6.3,26.7l-13.8,55.9h44.3Z"/>
-    <path fill="#fff" d="M596.1,234.2l-29.9-45.6c-1.2-1.9-2.4-4.1-3.5-6.5-.8-1.7-1.5-3.3-2.1-4.5-.6,1.3-1.4,2.8-2.3,4.5-1.3,2.4-2.6,4.6-4,6.5l-29.9,45.6h-28.5l49.5-71.9-46.5-67.9h28.5l27.6,43.1c1.2,1.9,2.3,3.9,3.4,6.1.7,1.4,1.3,2.7,1.9,3.8.5-1.1,1.1-2.4,1.8-3.8,1.1-2.2,2.2-4.2,3.4-6.1l27.6-43.1h28.5l-46.5,68.1,49.3,71.6h-28.5Z"/>
-    <path fill="#fff" d="M686.4,236.2c-16.7,0-30.2-5-40.1-14.8-9.9-9.8-14.9-23.7-14.9-41.3v-31.7c0-17.7,5-31.6,14.8-41.4,9.8-9.7,23.4-14.7,40.2-14.7s30.4,4.9,40.2,14.7c9.8,9.7,14.8,23.7,14.8,41.4v31.7c0,17.6-5,31.4-14.9,41.3-9.9,9.8-23.4,14.8-40.1,14.8ZM686.4,114.1c-9.5,0-17.1,2.7-22.6,8.1-5.5,5.4-8.3,13.4-8.3,23.8v36.7c0,10.5,2.8,18.5,8.3,23.8,5.5,5.4,13.1,8.1,22.6,8.1s17.3-2.7,22.7-8.1c5.4-5.4,8.2-13.4,8.2-23.8v-36.7c0-10.5-2.8-18.5-8.2-23.8-5.4-5.4-13.1-8.1-22.7-8.1Z"/>
-    <path fill="#fff" d="M888.3,236.2c-16.7,0-30.2-5-40.1-14.8-9.9-9.8-14.9-23.7-14.9-41.3v-31.7c0-17.7,5-31.6,14.8-41.4,9.8-9.7,23.4-14.7,40.2-14.7s30.4,4.9,40.2,14.7c9.8,9.7,14.8,23.7,14.8,41.4v31.7c0,17.6-5,31.4-14.9,41.3-9.9,9.8-23.4,14.8-40.1,14.8ZM888.3,114.1c-9.5,0-17.1,2.7-22.6,8.1-5.5,5.4-8.3,13.4-8.3,23.8v36.7c0,10.5,2.8,18.5,8.3,23.8,5.5,5.4,13.1,8.1,22.6,8.1s17.3-2.7,22.7-8.1c5.4-5.4,8.2-13.4,8.2-23.8v-36.7c0-10.5-2.8-18.5-8.2-23.8-5.4-5.4-13.1-8.1-22.7-8.1Z"/>
-    <path fill="#fff" d="M801.7,234c-18,0-32.6-14.6-32.6-32.6V48.8h24.1v152.5c0,4.7,3.8,8.5,8.5,8.5h16.7v24.1h-16.7Z"/>
-    <path fill="#fff" d="M1063.8,234.2V81.4c0-4.7-3.8-8.5-8.5-8.5h-16.7v-24.1h16.7c18,0,32.6,14.6,32.6,32.6v152.8h-24.1Z"/>
-    <path fill="#fff" d="M1006.2,234.2c-18,0-32.6-14.6-32.6-32.6v-85h-20.3v-22.1h20.3v-45.2h24.1v45.2h30.2v22.1h-30.2v85c0,4.7,3.8,8.5,8.5,8.5h21.7v24.1h-21.7Z"/>
-    <path fill="#fff" d="M160.2,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM222,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM222,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM160.2,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM160.2,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM222,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM222,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM160.2,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM160.2,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM222,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM222,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM160.2,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM277.3,57.4c0-23.8-19.3-43.1-43.1-43.1h-12.2c-3.9,0-7.6,1.6-10.2,4.4-5.9-2.9-12.3-4.4-18.9-4.4h-12.2c-7.7,0-14.1,6.3-14.1,14.1v20.6c0,2.4.6,4.6,1.6,6.6h-37c1-2,1.6-4.2,1.6-6.6v-20.6c0-7.7-6.3-14.1-14.1-14.1h-12.2c-6.5,0-13,1.5-18.9,4.4-2.6-2.8-6.3-4.4-10.2-4.4h-12.2c-23.8,0-43.1,19.3-43.1,43.1v32.8c0,4.1,1.7,7.7,4.5,10.3-2.8,2.6-4.5,6.2-4.5,10.3v41.2c0,11,5.2,20.8,13.2,27.2-7.3.4-13.2,6.6-13.2,14v20.6c0,4.1,1.7,7.7,4.5,10.3-2.8,2.6-4.5,6.2-4.5,10.3v20.6c0,7.7,6.3,14.1,14.1,14.1h41.2c4.1,0,7.7-1.7,10.3-4.5,2.6,2.8,6.2,4.5,10.3,4.5h164.9c7.7,0,14.1-6.3,14.1-14.1v-20.6c0-4.1-1.7-7.7-4.5-10.3,2.8-2.6,4.5-6.2,4.5-10.3v-20.6c0-7.5-5.8-13.6-13.2-14,8-6.4,13.2-16.2,13.2-27.2v-41.2c0-4.1-1.7-7.7-4.5-10.3,2.8-2.6,4.5-6.2,4.5-10.3v-32.8ZM77.8,255.1h-41.2v-20.6h41.2v20.6ZM36.5,213.9v-20.6h164.9v20.6H36.5ZM263.3,255.1H98.4v-20.6h164.9v20.6ZM263.3,213.9h-41.2v-20.6h41.2v20.6ZM263.3,90.2h-20.6v20.6h20.6v41.2c0,11.4-9.2,20.6-20.6,20.6H57.2c-11.4,0-20.6-9.2-20.6-20.6v-41.2h20.6v-20.6h-20.6v-32.8c0-16,13-29,29-29h12.2v20.6h-12.2c-4.7,0-8.4,3.8-8.4,8.4v32.8h41.2v-20.6h-20.6v-12.2c0-16,13-29,29-29h12.2v20.6h-12.2c-4.7,0-8.4,3.8-8.4,8.4v12.2h103.1v-12.2c0-4.7-3.8-8.4-8.4-8.4h-12.2v-20.6h12.2c16,0,29,13,29,29v12.2h-20.6v20.6h41.2v-32.8c0-4.7-3.8-8.4-8.4-8.4h-12.2v-20.6h12.2c16,0,29,13,29,29v32.8ZM201.4,152h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6s-20.6,9.2-20.6,20.6v20.6ZM160.2,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM160.2,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM222,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM222,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM160.2,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM160.2,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM222,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM222,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM160.2,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM160.2,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM222,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM222,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM160.2,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6Z"/>
-</svg>

image/axolotl_symbol_digital_black.svg

@@ -1,26 +0,0 @@
-<?xml version="1.0" encoding="UTF-8"?>
-<svg xmlns="http://www.w3.org/2000/svg" version="1.1" viewBox="0 0 283.5 283.5">
-  <defs>
-    <style>
-      .cls-1 {
-        fill: #141310;
-      }
-    </style>
-  </defs>
-  <!-- Generator: Adobe Illustrator 28.7.1, SVG Export Plug-In . SVG Version: 1.2.0 Build 142)  -->
-  <g>
-    <g id="Layer_1">
-      <g>
-        <path class="cls-1" d="M46.9,37.4h13.7V14.2h-13.7c-18,0-32.7,14.6-32.7,32.7v36.9h23.2v-36.9c0-5.2,4.2-9.5,9.5-9.5Z"/>
-        <path class="cls-1" d="M93.2,37.4h13.7V14.2h-13.7c-18,0-32.7,14.6-32.7,32.7v13.7h23.2v-13.7c0-5.2,4.2-9.5,9.5-9.5Z"/>
-        <path class="cls-1" d="M269.3,46.9c0-18-14.6-32.7-32.7-32.7h-13.7v23.2h13.7c5.2,0,9.5,4.2,9.5,9.5v36.9h23.2v-36.9Z"/>
-        <path class="cls-1" d="M199.7,83.8v-23.2h-116v23.2h-46.4v23.2H14.2v46.4c0,12.8,10.4,23.2,23.2,23.2h208.7c12.8,0,23.2-10.4,23.2-23.2v-46.4h-23.2v-23.2h-46.4ZM176.5,141.7c0-6.4-5.2-11.6-11.6-11.6s-11.6,5.2-11.6,11.6v11.6h-23.2v-23.2c0-12.8,10.4-23.2,23.2-23.2s23.2,10.4,23.2,23.2v11.6ZM246.1,141.7c0-6.4-5.2-11.6-11.6-11.6s-11.6,5.2-11.6,11.6v11.6h-23.2v-23.2c0-12.8,10.4-23.2,23.2-23.2s23.2,10.4,23.2,23.2v11.6Z"/>
-        <path class="cls-1" d="M222.9,46.9c0-18-14.6-32.7-32.7-32.7h-13.7v23.2h13.7c5.2,0,9.5,4.2,9.5,9.5v13.7h23.2v-13.7Z"/>
-        <rect class="cls-1" x="14.2" y="246.1" width="46.4" height="23.2"/>
-        <rect class="cls-1" x="83.8" y="246.1" width="185.5" height="23.2"/>
-        <rect class="cls-1" x="222.9" y="199.7" width="46.4" height="23.2"/>
-        <rect class="cls-1" x="14.2" y="199.7" width="185.5" height="23.2"/>
-      </g>
-    </g>
-  </g>
-</svg>

image/axolotl_symbol_digital_white.svg

@@ -1,16 +0,0 @@
-<?xml version="1.0" encoding="UTF-8"?>
-<svg xmlns="http://www.w3.org/2000/svg" version="1.1" viewBox="0 0 283.5 283.5">
-  <defs>
-    <style>
-      .cls-1 {
-        fill: #fff;
-      }
-    </style>
-  </defs>
-  <!-- Generator: Adobe Illustrator 28.7.1, SVG Export Plug-In . SVG Version: 1.2.0 Build 142)  -->
-  <g>
-    <g id="Layer_1">
-      <path class="cls-1" d="M152,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM214,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM214,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM152,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM152,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM214,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM214,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM152,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM152,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM214,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM214,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM152,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM269.3,57.3c0-23.8-19.4-43.1-43.1-43.1h-12.2c-3.9,0-7.6,1.6-10.2,4.4-5.9-2.9-12.3-4.4-18.9-4.4h-12.2c-7.8,0-14.1,6.3-14.1,14.1v20.6c0,2.4.6,4.6,1.6,6.6h-37c1-2,1.6-4.2,1.6-6.6v-20.6c0-7.8-6.3-14.1-14.1-14.1h-12.2c-6.6,0-13,1.5-18.9,4.4-2.6-2.8-6.3-4.4-10.2-4.4h-12.2c-23.8,0-43.1,19.4-43.1,43.1v32.8c0,4.1,1.7,7.7,4.5,10.3-2.8,2.6-4.5,6.2-4.5,10.3v41.3c0,11,5.2,20.9,13.2,27.2-7.4.4-13.2,6.6-13.2,14v20.6c0,4.1,1.7,7.7,4.5,10.3-2.8,2.6-4.5,6.2-4.5,10.3v20.6c0,7.8,6.3,14.1,14.1,14.1h41.3c4.1,0,7.7-1.7,10.3-4.5,2.6,2.8,6.2,4.5,10.3,4.5h165.1c7.8,0,14.1-6.3,14.1-14.1v-20.6c0-4.1-1.7-7.7-4.5-10.3,2.8-2.6,4.5-6.2,4.5-10.3v-20.6c0-7.5-5.9-13.6-13.2-14,8-6.4,13.2-16.2,13.2-27.2v-41.3c0-4.1-1.7-7.7-4.5-10.3,2.8-2.6,4.5-6.2,4.5-10.3v-32.8ZM69.5,255.2H28.2v-20.6h41.3v20.6ZM28.2,214v-20.6h165.1v20.6H28.2ZM255.2,255.2H90.1v-20.6h165.1v20.6ZM255.2,214h-41.3v-20.6h41.3v20.6ZM255.2,90.1h-20.6v20.6h20.6v41.3c0,11.4-9.2,20.6-20.6,20.6H48.9c-11.4,0-20.6-9.2-20.6-20.6v-41.3h20.6v-20.6h-20.6v-32.8c0-16.1,13-29.1,29.1-29.1h12.2v20.6h-12.2c-4.7,0-8.4,3.8-8.4,8.4v32.8h41.3v-20.6h-20.6v-12.2c0-16.1,13-29.1,29.1-29.1h12.2v20.6h-12.2c-4.7,0-8.4,3.8-8.4,8.4v12.2h103.2v-12.2c0-4.7-3.8-8.4-8.4-8.4h-12.2v-20.6h12.2c16.1,0,29.1,13,29.1,29.1v12.2h-20.6v20.6h41.3v-32.8c0-4.7-3.8-8.4-8.4-8.4h-12.2v-20.6h12.2c16.1,0,29.1,13,29.1,29.1v32.8ZM193.3,152h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6s-20.6,9.2-20.6,20.6v20.6ZM152,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM152,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM214,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM214,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM152,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM152,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM214,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM214,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM152,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM152,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM214,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM214,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6ZM152,110.8c-11.4,0-20.6,9.2-20.6,20.6v20.6h20.6v-10.3c0-5.7,4.6-10.3,10.3-10.3s10.3,4.6,10.3,10.3v-10.3c0-11.4-9.2-20.6-20.6-20.6Z"/>
-    </g>
-  </g>
-</svg>

image/axolotl_wordmark_digital_black.svg

@@ -1,17 +0,0 @@
-<?xml version="1.0" encoding="UTF-8"?>
-<svg xmlns="http://www.w3.org/2000/svg" version="1.1" viewBox="0 0 765.4 212.6">
-  <!-- Generator: Adobe Illustrator 28.7.1, SVG Export Plug-In . SVG Version: 1.2.0 Build 142)  -->
-  <g>
-    <g id="Layer_1">
-      <g>
-        <path d="M121.6,198.1l-12.1-48.8h-54.4l-12.1,48.8h-24.7L66.6,12.9h31.6l47.9,185.1h-24.5ZM104.4,128.6l-13.8-55.6c-2.7-10.7-4.8-19.7-6.3-26.9-.9-4.2-1.5-7.5-2-9.9-.5,2.5-1.2,5.8-2,9.9-1.5,7.1-3.6,16.1-6.3,26.7l-13.8,55.9h44.3Z"/>
-        <path d="M254.9,198.1l-29.9-45.6c-1.2-1.9-2.4-4.1-3.5-6.5-.8-1.7-1.5-3.3-2.1-4.5-.6,1.3-1.4,2.8-2.3,4.5-1.3,2.4-2.6,4.6-4,6.5l-29.9,45.6h-28.5l49.6-71.9-46.5-67.9h28.5l27.6,43.1c1.2,1.9,2.3,3.9,3.4,6.1.7,1.4,1.4,2.7,1.9,3.8.5-1.1,1.1-2.4,1.8-3.8,1.1-2.2,2.2-4.2,3.4-6.1l27.6-43.1h28.5l-46.5,68.2,49.3,71.7h-28.5Z"/>
-        <path d="M345.2,200.1c-16.7,0-30.2-5-40.1-14.8-9.9-9.8-14.9-23.7-14.9-41.3v-31.7c0-17.7,5-31.7,14.8-41.4,9.8-9.7,23.4-14.7,40.3-14.7s30.4,4.9,40.3,14.7c9.8,9.7,14.8,23.7,14.8,41.4v31.7c0,17.6-5,31.5-14.9,41.3-9.9,9.8-23.4,14.8-40.1,14.8ZM345.2,77.8c-9.5,0-17.1,2.7-22.6,8.1-5.5,5.4-8.3,13.4-8.3,23.8v36.7c0,10.5,2.8,18.5,8.3,23.8,5.5,5.4,13.1,8.1,22.6,8.1s17.3-2.7,22.7-8.1c5.4-5.4,8.2-13.4,8.2-23.9v-36.7c0-10.5-2.8-18.5-8.2-23.9-5.4-5.4-13.1-8.1-22.7-8.1Z"/>
-        <path d="M547.3,200.1c-16.7,0-30.2-5-40.1-14.8-9.9-9.8-14.9-23.7-14.9-41.3v-31.7c0-17.7,5-31.7,14.8-41.4,9.8-9.7,23.4-14.7,40.3-14.7s30.4,4.9,40.3,14.7c9.8,9.7,14.8,23.7,14.8,41.4v31.7c0,17.6-5,31.5-14.9,41.3-9.9,9.8-23.4,14.8-40.1,14.8ZM547.3,77.8c-9.5,0-17.1,2.7-22.6,8.1-5.5,5.4-8.3,13.4-8.3,23.8v36.7c0,10.5,2.8,18.5,8.3,23.8,5.5,5.4,13.1,8.1,22.6,8.1s17.3-2.7,22.7-8.1c5.4-5.4,8.2-13.4,8.2-23.9v-36.7c0-10.5-2.8-18.5-8.2-23.9-5.4-5.4-13.1-8.1-22.7-8.1Z"/>
-        <path d="M460.6,197.8c-18,0-32.6-14.6-32.6-32.6V12.5h24.1v152.6c0,4.7,3.8,8.5,8.5,8.5h16.8v24.1h-16.8Z"/>
-        <path d="M722.8,198.1V45.2c0-4.7-3.8-8.5-8.5-8.5h-16.8V12.5h16.8c18,0,32.6,14.6,32.6,32.6v152.9h-24.1Z"/>
-        <path d="M665.2,198.1c-18,0-32.6-14.6-32.6-32.6v-85.1h-20.3v-22.1h20.3V12.9h24.1v45.3h30.2v22.1h-30.2v85.1c0,4.7,3.8,8.5,8.5,8.5h21.7v24.1h-21.7Z"/>
-      </g>
-    </g>
-  </g>
-</svg>

image/axolotl_wordmark_digital_white.svg

@@ -1,24 +0,0 @@
-<?xml version="1.0" encoding="UTF-8"?>
-<svg xmlns="http://www.w3.org/2000/svg" version="1.1" viewBox="0 0 765.4 212.6">
-  <defs>
-    <style>
-      .cls-1 {
-        fill: #fff;
-      }
-    </style>
-  </defs>
-  <!-- Generator: Adobe Illustrator 28.7.1, SVG Export Plug-In . SVG Version: 1.2.0 Build 142)  -->
-  <g>
-    <g id="Layer_1">
-      <g>
-        <path class="cls-1" d="M121.6,198.1l-12.1-48.8h-54.4l-12.1,48.8h-24.7L66.6,12.9h31.6l47.9,185.1h-24.5ZM104.4,128.6l-13.8-55.6c-2.7-10.7-4.8-19.7-6.3-26.9-.9-4.2-1.5-7.5-2-9.9-.5,2.5-1.2,5.8-2,9.9-1.5,7.1-3.6,16.1-6.3,26.7l-13.8,55.9h44.3Z"/>
-        <path class="cls-1" d="M254.9,198.1l-29.9-45.6c-1.2-1.9-2.4-4.1-3.5-6.5-.8-1.7-1.5-3.3-2.1-4.5-.6,1.3-1.4,2.8-2.3,4.5-1.3,2.4-2.6,4.6-4,6.5l-29.9,45.6h-28.5l49.6-71.9-46.5-67.9h28.5l27.6,43.1c1.2,1.9,2.3,3.9,3.4,6.1.7,1.4,1.4,2.7,1.9,3.8.5-1.1,1.1-2.4,1.8-3.8,1.1-2.2,2.2-4.2,3.4-6.1l27.6-43.1h28.5l-46.5,68.2,49.3,71.7h-28.5Z"/>
-        <path class="cls-1" d="M345.2,200.1c-16.7,0-30.2-5-40.1-14.8-9.9-9.8-14.9-23.7-14.9-41.3v-31.7c0-17.7,5-31.7,14.8-41.4,9.8-9.7,23.4-14.7,40.3-14.7s30.4,4.9,40.3,14.7c9.8,9.7,14.8,23.7,14.8,41.4v31.7c0,17.6-5,31.5-14.9,41.3-9.9,9.8-23.4,14.8-40.1,14.8ZM345.2,77.8c-9.5,0-17.1,2.7-22.6,8.1-5.5,5.4-8.3,13.4-8.3,23.8v36.7c0,10.5,2.8,18.5,8.3,23.8,5.5,5.4,13.1,8.1,22.6,8.1s17.3-2.7,22.7-8.1c5.4-5.4,8.2-13.4,8.2-23.9v-36.7c0-10.5-2.8-18.5-8.2-23.9-5.4-5.4-13.1-8.1-22.7-8.1Z"/>
-        <path class="cls-1" d="M547.3,200.1c-16.7,0-30.2-5-40.1-14.8-9.9-9.8-14.9-23.7-14.9-41.3v-31.7c0-17.7,5-31.7,14.8-41.4,9.8-9.7,23.4-14.7,40.3-14.7s30.4,4.9,40.3,14.7c9.8,9.7,14.8,23.7,14.8,41.4v31.7c0,17.6-5,31.5-14.9,41.3-9.9,9.8-23.4,14.8-40.1,14.8ZM547.3,77.8c-9.5,0-17.1,2.7-22.6,8.1-5.5,5.4-8.3,13.4-8.3,23.8v36.7c0,10.5,2.8,18.5,8.3,23.8,5.5,5.4,13.1,8.1,22.6,8.1s17.3-2.7,22.7-8.1c5.4-5.4,8.2-13.4,8.2-23.9v-36.7c0-10.5-2.8-18.5-8.2-23.9-5.4-5.4-13.1-8.1-22.7-8.1Z"/>
-        <path class="cls-1" d="M460.6,197.8c-18,0-32.6-14.6-32.6-32.6V12.5h24.1v152.6c0,4.7,3.8,8.5,8.5,8.5h16.8v24.1h-16.8Z"/>
-        <path class="cls-1" d="M722.8,198.1V45.2c0-4.7-3.8-8.5-8.5-8.5h-16.8V12.5h16.8c18,0,32.6,14.6,32.6,32.6v152.9h-24.1Z"/>
-        <path class="cls-1" d="M665.2,198.1c-18,0-32.6-14.6-32.6-32.6v-85.1h-20.3v-22.1h20.3V12.9h24.1v45.3h30.2v22.1h-30.2v85.1c0,4.7,3.8,8.5,8.5,8.5h21.7v24.1h-21.7Z"/>
-      </g>
-    </g>
-  </g>
-</svg>

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==4.46.3
-tokenizers>=0.20.1
-bitsandbytes==0.44.1
-accelerate==1.1.0
-datasets==3.1.0
-deepspeed==0.15.4
+peft==0.13.0
+transformers==4.45.1
+tokenizers>=0.19.1
+bitsandbytes==0.44.0
+accelerate==0.34.2
+datasets==2.21.0
+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==0.2.7.post2
+autoawq>=0.2.5
 triton>=2.3.0
-liger-kernel==0.4.2
+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
 
@@ -51,6 +52,3 @@ lm_eval==0.4.4
 langdetect==1.0.9
 immutabledict==4.2.0
 antlr4-python3-runtime==4.13.2
-
-torchao==0.5.0
-schedulefree==1.3.0

requirements_env.txt

@@ -1,315 +0,0 @@
-accelerate==0.34.1
-addict==2.4.0
-aiofiles==23.2.1
-aiohttp==3.9.0
-aiosignal==1.3.1
-aiostream==0.5.2
-alembic==1.13.1
-annotated-types==0.6.0
-annoy==1.17.3
-ansible==6.7.0
-ansible-core==2.13.13
-ansible-vault==2.1.0
-anyio==3.7.1
-appdirs==1.4.4
-art==6.0
-asgiref==3.7.2
-async-timeout==4.0.2
-attrdict==2.0.1
-attrs==22.2.0
-awscli==1.32.75
--e git+ssh://git@github.com/OpenAccess-AI-Collective/axolotl.git@6e354682e3c1735d3f7fb9e362280c38e922260f#egg=axolotl
-backoff==2.2.1
-base58==2.1.1
-beartype==0.17.2
-bitnet==0.2.1
-bitsandbytes==0.42.0
-bittensor==6.7.0
-black==23.7.0
-blinker==1.7.0
-boto3==1.34.75
-botocore==1.34.75
-cachetools==5.3.3
-cachy==0.1.1
-certifi==2023.7.22
-cffi==1.16.0
-cfgv==3.3.1
-chai-guanaco==1.2.4
-charset-normalizer==3.2.0
-cleo==0.6.8
-click==8.1.7
-cloudpickle==2.0.0
-cohere==4.11.2
-colorama==0.4.4
-coloredlogs==15.0.1
-CoLT5-attention==0.10.20
-contextlib2==21.6.0
-contourpy==1.2.0
-cryptography==41.0.3
-cycler==0.12.1
-cytoolz==0.12.3
-databricks-cli==0.18.0
-dataclasses-json==0.5.7
-datasets==2.11.0
-ddt==1.6.0
-decorator==5.1.1
-deepspeed==0.15.0
-# Editable Git install with no remote (dialogpt==0.1)
--e /Users/wing/Projects/ml/dialogpt/src
-dill==0.3.6
-distlib==0.3.6
-docker==7.0.0
-docker-pycreds==0.4.0
-docstring-parser==0.15
-docutils==0.16
-ecdsa==0.18.0
-einops==0.7.0
-einops-exts==0.0.4
-einx==0.1.3
-entrypoints==0.4
-eth-hash==0.6.0
-eth-keys==0.5.0
-eth-typing==4.0.0
-eth-utils==2.3.1
-evaluate==0.4.0
-exceptiongroup==1.1.1
-fastapi==0.109.2
-fastcore==1.5.29
-ffmpy==0.4.0
-filelock==3.12.2
--e git+https://github.com/NousResearch/finetuning-subnet.git@24e9407d6b4430a7ca39d344692f89ce5a97d27e#egg=finetuning_subnet
-fire==0.5.0
-first==2.0.2
-flake8==7.0.0
-Flask==3.0.1
-fonttools==4.47.2
-frozendict==2.4.1
-frozenlist==1.3.3
-fschat @ git+https://github.com/lm-sys/FastChat.git@27a05b04a35510afb1d767ae7e5990cbd278f8fe
-fsspec==2023.6.0
-fuzzywuzzy==0.18.0
-gitdb==4.0.10
-GitPython==3.1.31
-google-pasta==0.2.0
-gradio==4.42.0
-gradio_client==1.3.0
-greenlet==2.0.2
-grpclib==0.4.7
-gunicorn==21.2.0
-h11==0.14.0
-h2==4.1.0
-hpack==4.0.0
-httpcore==0.17.3
-httpx==0.24.1
-huggingface-hub==0.23.4
-humanfriendly==10.0
-hyperframe==6.0.1
-identify==2.5.24
-idna==3.4
-immutables==0.20
-importlib-metadata==6.7.0
-importlib-resources==6.1.1
-inflection==0.5.1
-iniconfig==2.0.0
-itsdangerous==2.1.2
-Jinja2==3.1.2
-jmespath==1.0.1
-joblib==1.3.2
-jsonlines==3.1.0
-jsonschema==2.6.0
-kiwisolver==1.4.5
-langchain==0.0.144
-Levenshtein==0.24.0
-libcst==1.1.0
-liger-kernel==0.0.0
-lion-pytorch==0.1.2
-llama-cpp-python==0.1.36
-llvmlite==0.40.1
-local-attention==1.9.0
-loguru==0.7.0
-Mako==1.3.2
-Markdown==3.5.2
-markdown-it-py==3.0.0
-markdown2==2.4.10
-MarkupSafe==2.1.2
-marshmallow==3.19.0
-marshmallow-enum==1.5.1
-matplotlib==3.8.2
-mccabe==0.7.0
-mdurl==0.1.2
-MEGABYTE-pytorch==0.0.7
--e git+https://github.com/cg123/mergekit.git@53c5f414774a0558b8d84858fb6374bc93a8f1c1#egg=mergekit
-mlflow==2.10.0
-modal==0.62.77
-more-itertools==10.2.0
-mpmath==1.2.1
-msgpack==1.0.7
-msgpack-numpy-opentensor==0.5.0
-multidict==6.0.4
-multiprocess==0.70.14
-munch==2.5.0
-mypy==1.3.0
-mypy-extensions==1.0.0
-nest-asyncio==1.6.0
-netaddr==0.10.1
-networkx==3.0rc1
-nh3==0.2.14
-nodeenv==1.8.0
-nomic==2.0.2
-numba==0.57.1
-numexpr==2.8.4
-numpy==1.24.4
-oauthlib==3.2.2
-openai==0.27.4
-openapi==1.1.0
-openapi-schema-pydantic==1.2.4
-optimum==1.8.6
-orjson==3.10.7
-packaging==23.1
-pandas==2.0.0
-parameterized==0.9.0
-password-strength==0.0.3.post2
-pastel==0.1.1
-pathos==0.3.0
-pathspec==0.11.1
-pathtools==0.1.2
-peft==0.11.1
-pendulum==3.0.0
-Pillow==9.5.0
-pip-tools==1.11.0
-platformdirs==3.2.0
-pluggy==1.4.0
-poetry==0.7.1
-pox==0.3.2
-ppft==1.7.6.6
-pre-commit==3.3.2
-prettytable==3.10.0
-prompt-toolkit==3.0.39
-protobuf==3.20.2
-protobuf3-to-dict==0.1.5
-psutil==5.9.5
-psycopg==3.1.18
-PuLP==2.8.0
-py==1.11.0
-py-bip39-bindings==0.1.11
-py-cpuinfo==9.0.0
-py-ed25519-zebra-bindings==1.0.1
-py-sr25519-bindings==0.2.0
-pyarrow==11.0.0
-pyasn1==0.6.0
-pycodestyle==2.11.1
-pycparser==2.21
-pycryptodome==3.20.0
-pydantic==2.5.3
-pydantic_core==2.14.6
-pydub==0.25.1
-pyfiglet==0.8.post1
-pyflakes==3.2.0
-Pygments==2.15.1
-PyJWT==2.8.0
-pylev==1.4.0
-PyNaCl==1.5.0
-pynvml==11.5.0
-pyparsing==2.4.7
-pyrsistent==0.14.11
-pytest==8.0.2
-pytest-asyncio==0.23.4
-python-dateutil==2.8.2
-python-dotenv==1.0.1
-python-Levenshtein==0.24.0
-python-multipart==0.0.9
-pytz==2023.3
-PyYAML==6.0.1
-querystring-parser==1.2.4
-rapidfuzz==3.6.1
-regex==2023.6.3
-requests==2.31.0
-requests-toolbelt==0.8.0
-resolvelib==0.8.1
-responses==0.18.0
-retry==0.9.2
-rich==13.7.0
-rsa==4.7.2
-ruff==0.6.3
-s3transfer==0.10.1
-safetensors==0.4.5
-sagemaker==2.148.0
-scalecodec==1.2.7
-schedulefree==1.2.1
-schema==0.7.5
-scikit-learn==1.4.0
-scipy==1.9.3
-seaborn==0.13.2
-semantic-version==2.10.0
-sentencepiece==0.2.0
-sentry-sdk==1.19.1
-setproctitle==1.3.2
-shellingham==1.5.4
-shortuuid==1.0.11
-shtab==1.6.5
-sigtools==4.0.1
-six==1.16.0
-skypilot==0.4.1
-smdebug-rulesconfig==1.0.1
-smmap==5.0.0
-sniffio==1.3.0
-SQLAlchemy==1.4.47
-sqlparse==0.4.4
-starlette==0.36.3
-substrate-interface==1.5.2
-svgwrite==1.4.3
-sympy==1.11.1
-synchronicity==0.6.7
-tabulate==0.9.0
-tblib==1.7.0
-tenacity==8.2.2
-tensor-parallel==2.0.0
-termcolor==2.2.0
-text2art==0.2.0
-threadpoolctl==3.2.0
-tiktoken==0.6.0
-time-machine==2.14.1
-timm==0.9.16
-tokenizers==0.19.1
-tokenmonster==1.1.12
-toml==0.9.6
-tomli==2.0.1
-tomlkit==0.12.0
-toolz==0.12.1
-torch==2.2.0
-torchdata==0.6.1
-torchdiffeq==0.2.3
-TorchFix==0.4.0
-torchtext==0.15.2
-torchvision==0.17.0
-tqdm==4.66.2
-transformers==4.44.2
-trl==0.9.6
-typer==0.12.5
-types-certifi==2021.10.8.3
-types-requests==2.31.0.20240125
-types-setuptools==69.0.0.20240125
-types-toml==0.10.8.7
-typing==3.7.4.3
-typing-inspect==0.8.0
-typing_extensions==4.9.0
-tyro==0.5.18
-tzdata==2023.3
-unique-names-generator==1.0.2
-urllib3==2.2.2
-uvicorn==0.22.0
-vector_quantize_pytorch==1.14.1
-virtualenv==20.23.0
-voyager==2.0.2
-wandb==0.16.2
-watchfiles==0.21.0
-wavedrom==2.0.3.post3
-wcwidth==0.2.6
-websocket-client==1.7.0
-websockets==12.0
-Werkzeug==3.0.1
-wonderwords==2.2.0
-xxhash==3.2.0
-yarl==1.8.2
-zetascale==2.2.7
-zipp==3.15.0

scripts/chat_datasets.py

@@ -1,60 +0,0 @@
-"""
-helper script to parse chat datasets into a usable yaml
-"""
-import click
-import yaml
-from datasets import load_dataset
-
-
-@click.command()
-@click.argument("dataset", type=str)
-@click.option("--split", type=str, default="train")
-def parse_dataset(dataset=None, split="train"):
-    ds_cfg = {}
-    ds_cfg["path"] = dataset
-    ds_cfg["split"] = split
-    ds_cfg["type"] = "chat_template"
-    ds_cfg["chat_template"] = "<<<Replace based on your model>>>"
-
-    dataset = load_dataset(dataset, split=split)
-    features = dataset.features
-    feature_keys = features.keys()
-    field_messages = None
-    for key in ["conversation", "conversations", "messages"]:
-        if key in feature_keys:
-            field_messages = key
-            break
-    if not field_messages:
-        raise ValueError(
-            f'No conversation field found in dataset: {", ".join(feature_keys)}'
-        )
-    ds_cfg["field_messages"] = field_messages
-
-    message_fields = features["conversations"][0].keys()
-    message_field_role = None
-    for key in ["from", "role"]:
-        if key in message_fields:
-            message_field_role = key
-            break
-    if not message_field_role:
-        raise ValueError(
-            f'No role field found in messages: {", ".join(message_fields)}'
-        )
-    ds_cfg["message_field_role"] = message_field_role
-
-    message_field_content = None
-    for key in ["content", "text", "value"]:
-        if key in message_fields:
-            message_field_content = key
-            break
-    if not message_field_content:
-        raise ValueError(
-            f'No content field found in messages: {", ".join(message_fields)}'
-        )
-    ds_cfg["message_field_content"] = message_field_content
-
-    print(yaml.dump({"datasets": [ds_cfg]}))
-
-
-if __name__ == "__main__":
-    parse_dataset()

scripts/cloud-entrypoint.sh

@@ -2,7 +2,7 @@
 
 # Export specific ENV variables to /etc/rp_environment
 echo "Exporting environment variables..."
-printenv | grep -E '^HF_|^BNB_|^CUDA_|^NCCL_|^NV|^RUNPOD_|^PATH=|^_=' | sed 's/^\([^=]*\)=\(.*\)$/export \1="\2"/' | grep -v 'printenv' >> /etc/rp_environment
+printenv | grep -E '^RUNPOD_|^PATH=|^_=' | sed 's/^\(.*\)=\(.*\)$/export \1="\2"/' >> /etc/rp_environment
 echo 'source /etc/rp_environment' >> ~/.bashrc
 
 add_keys_to_authorized() {

scripts/unsloth_install.py

@@ -1,33 +0,0 @@
-# noqa
-# pylint: skip-file
-try:
-    import torch
-except ImportError:
-    raise ImportError("Install torch via `pip install torch`")
-from packaging.version import Version as V
-
-v = V(torch.__version__)
-cuda = str(torch.version.cuda)
-is_ampere = torch.cuda.get_device_capability()[0] >= 8
-if cuda != "12.1" and cuda != "11.8" and cuda != "12.4":
-    raise RuntimeError(f"CUDA = {cuda} not supported!")
-if v <= V("2.1.0"):
-    raise RuntimeError(f"Torch = {v} too old!")
-elif v <= V("2.1.1"):
-    x = "cu{}{}-torch211"
-elif v <= V("2.1.2"):
-    x = "cu{}{}-torch212"
-elif v < V("2.3.0"):
-    x = "cu{}{}-torch220"
-elif v < V("2.4.0"):
-    x = "cu{}{}-torch230"
-elif v < V("2.5.0"):
-    x = "cu{}{}-torch240"
-elif v < V("2.6.0"):
-    x = "cu{}{}-torch250"
-else:
-    raise RuntimeError(f"Torch = {v} too new!")
-x = x.format(cuda.replace(".", ""), "-ampere" if is_ampere else "")
-print(
-    f'pip install unsloth-zoo && pip install --no-deps "unsloth[{x}] @ git+https://github.com/unslothai/unsloth.git"'
-)

setup.py

@@ -30,19 +30,13 @@ 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)
@@ -55,22 +49,11 @@ 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 (major, minor) >= (2, 3):
                 if patch == 0:
                     _install_requires.pop(_install_requires.index(xformers_version))
                     _install_requires.append("xformers>=0.0.26.post1")
@@ -78,16 +61,15 @@ def parse_requirements():
                     _install_requires.pop(_install_requires.index(xformers_version))
                     _install_requires.append("xformers>=0.0.27")
             elif (major, minor) >= (2, 2):
-                _install_requires.pop(_install_requires.index(torchao_version))
                 _install_requires.pop(_install_requires.index(xformers_version))
                 _install_requires.append("xformers>=0.0.25.post1")
             else:
-                _install_requires.pop(_install_requires.index(torchao_version))
                 _install_requires.pop(_install_requires.index(xformers_version))
                 _install_requires.append("xformers>=0.0.23.post1")
 
     except PackageNotFoundError:
         pass
+
     return _install_requires, _dependency_links
 
 
@@ -96,24 +78,26 @@ install_requires, dependency_links = parse_requirements()
 
 setup(
     name="axolotl",
-    version="0.5.2",
+    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"},
-    packages=find_packages("src"),
+    packages=find_packages(),
     install_requires=install_requires,
     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==0.15.4",
+            "deepspeed==0.14.4",
             "deepspeed-kernels",
         ],
         "mamba-ssm": [
             "mamba-ssm==1.2.0.post1",
-            "causal_conv1d",
         ],
         "auto-gptq": [
             "auto-gptq==0.5.1",

src/axolotl/cli/__init__.py

@@ -30,10 +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,
-    get_chat_template_from_config,
-)
+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,
@@ -193,19 +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)
-    elif cfg.datasets[0].type == "chat_template":
-        chat_template_str = get_chat_template_from_config(
-            cfg=cfg, ds_cfg=cfg.datasets[0], tokenizer=tokenizer
-        )
 
     model = model.to(cfg.device, dtype=cfg.torch_dtype)
 
@@ -215,31 +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()
@@ -279,6 +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
@@ -287,7 +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)
 
@@ -477,12 +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

@@ -23,7 +23,10 @@ from axolotl.cli import (
 )
 from axolotl.common.cli import PreprocessCliArgs
 from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
-from axolotl.utils.trainer import disable_datasets_caching
+from axolotl.prompt_strategies.sharegpt import (
+    register_chatml_template,
+    register_llama3_template,
+)
 
 LOG = logging.getLogger("axolotl.cli.preprocess")
 
@@ -40,6 +43,23 @@ 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
@@ -50,11 +70,10 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
         LOG.warning(msg)
         parsed_cfg.dataset_prepared_path = DEFAULT_DATASET_PREPARED_PATH
 
-    with disable_datasets_caching():
-        if parsed_cfg.rl:  # and parsed_cfg.rl != "orpo":
-            load_rl_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
-        else:
-            load_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
+    if parsed_cfg.rl:  # and parsed_cfg.rl != "orpo":
+        load_rl_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
+    else:
+        load_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
 
     if parsed_cli_args.download:
         model_name = parsed_cfg.base_model

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/chat/format/chatml.py

@@ -1,34 +0,0 @@
-"""
-ChatML transformation functions for MessageContents
-"""
-from typing import Optional
-
-from ..messages import MessageContents, Messages
-from .shared import wrap_tools
-
-
-def format_message(
-    message: Messages,
-    message_index: Optional[int] = None,  # pylint: disable=unused-argument
-) -> Messages:
-    if message.is_chat_formatted:
-        return message
-
-    # prepend the role prefix within a MessageContents to message.content
-    message.content.insert(
-        0,
-        MessageContents(
-            type="text",
-            value=f"<|im_start|>{message.role}\n",
-            weight=0,
-        ),
-    )
-    message.content.append(
-        MessageContents(type="text", value="<|im_end|>", weight=message.weight)
-    )
-    message.content.append(MessageContents(type="text", value="\n", weight=0))
-
-    message = wrap_tools(message)
-
-    message.is_chat_formatted = True
-    return message

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

@@ -1,45 +0,0 @@
-"""
-Llama 3.x chat formatting functions for MessageContents
-"""
-from typing import Optional
-
-from ..messages import MessageContents, Messages
-from .shared import wrap_tools
-
-
-def format_message(message: Messages, message_index: Optional[int] = None) -> Messages:
-    if message.is_chat_formatted:
-        return message
-
-    message_role = message.role
-    if message.role == "tool":
-        message_role = "ipython"
-
-    # prepend the role prefix within a MessageContents to message.content
-    message.content.insert(
-        0,
-        MessageContents(
-            type="text",
-            value=f"<|start_header_id|>{message_role}<|end_header_id|>\n\n",
-            weight=0,
-        ),
-    )
-
-    message.content.append(
-        MessageContents(type="text", value="<|eot_id|>", weight=message.weight)
-    )
-
-    message = wrap_tools(message)
-
-    if message_index == 0:
-        message.content.insert(
-            0,
-            MessageContents(
-                type="text",
-                value="<|begin_of_text|>",
-                weight=0,
-            ),
-        )
-
-    message.is_chat_formatted = True
-    return message

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

@@ -1,47 +0,0 @@
-"""
-shared functions for format transforms
-"""
-from axolotl.core.chat.messages import MessageContents, Messages
-
-
-def wrap_tools(message: Messages):
-    # loop over message.content by index to find tool calls, we need to wrap each with tags,
-    # so be wary of indexing issues when changing the list while iterating.
-    # iterate over the range in reverse order to avoid index shifting
-    for i in range(len(message.content) - 1, -1, -1):
-        if message.content[i].type == "tool_call":
-            # append a </tool_call> MessageContents text tag after
-            message.content.insert(
-                i + 1,
-                MessageContents(
-                    type="text", value="</tool_call>\n", weight=message.weight
-                ),
-            )
-            # make sure the actual tool call content ends with a newline
-            message.content[i].has_newline = True
-            # prepend a <tool_call> MessageContents text tag before
-            message.content.insert(
-                i,
-                MessageContents(
-                    type="text", value="<tool_call>\n", weight=message.weight
-                ),
-            )
-        elif message.content[i].type == "tool_response":
-            # append a </tool_call> MessageContents text tag after
-            message.content.insert(
-                i + 1,
-                MessageContents(
-                    type="text", value="</tool_response>\n", weight=message.weight
-                ),
-            )
-            # make sure the actual tool response content ends with a newline
-            message.content[i].has_newline = True
-            # prepend a <tool_call> MessageContents text tag before
-            message.content.insert(
-                i,
-                MessageContents(
-                    type="text", value="<tool_response>\n", weight=message.weight
-                ),
-            )
-
-    return message

src/axolotl/core/chat/messages.py

@@ -1,230 +0,0 @@
-"""
-internal message representations of chat messages
-"""
-import json
-from enum import Enum
-from typing import Any, Callable, List, Optional, Union
-
-from pydantic import BaseModel
-from transformers import PreTrainedTokenizer
-
-
-class MessageRoles(str, Enum):
-    """
-    Message roles for the system, user, assistant, and tools
-    """
-
-    system = "system"  # pylint: disable=invalid-name
-    user = "user"  # pylint: disable=invalid-name
-    assistant = "assistant"  # pylint: disable=invalid-name
-    tool = "tool"  # pylint: disable=invalid-name
-    ipython = (  # pylint: disable=invalid-name
-        # for responses from builtin tools
-        "ipython"
-    )
-
-
-class MessageContentTypes(str, Enum):
-    """
-    Message content types for text, image, audio, tool calls, and tool responses
-    """
-
-    special_token = "special_token"  # pylint: disable=invalid-name  # nosec B105
-    text = "text"  # pylint: disable=invalid-name
-    image = "image"  # pylint: disable=invalid-name
-    audio = "audio"  # pylint: disable=invalid-name
-    tool_call = "tool_call"  # pylint: disable=invalid-name  # to differentiate regular responses from tool calls from the assistant
-    tool_response = "tool_response"  # pylint: disable=invalid-name
-
-
-class SpecialToken(str, Enum):
-    """
-    Special tokens for beginning of string and end of string
-    """
-
-    bos_token = "bos_token"  # pylint: disable=invalid-name  # nosec B105
-    eos_token = "eos_token"  # pylint: disable=invalid-name  # nosec B105
-
-
-class ToolCallFunction(BaseModel):
-    """
-    Tool call function with name and arguments
-    """
-
-    name: str
-    arguments: dict[str, str]
-
-
-class Tool(BaseModel):
-    """
-    Tool with description, function, and parameters
-    """
-
-    description: str
-    function: ToolCallFunction
-    parameters: dict[str, str]  # .properties
-
-
-class ToolCallContents(BaseModel):
-    """
-    Tool call contents with name, arguments, and optional id
-    """
-
-    name: str
-    arguments: dict[str, Union[str, int]]
-    id: Optional[str] = None  # pylint: disable=invalid-name
-
-    def __str__(self) -> str:
-        data = {"name": self.name, "arguments": self.arguments}
-        if self.id is not None:
-            data["id"] = self.id
-        return json.dumps(data)
-
-
-class ToolResponseContents(BaseModel):
-    """
-    Tool response contents with name, content, and optional id
-    """
-
-    name: str
-    content: Union[str, dict[str, Union[str, int, float]]]
-    id: Optional[str] = None  # pylint: disable=invalid-name
-
-    def __str__(self) -> str:
-        data = {"name": self.name, "content": self.content}
-        if self.id is not None:
-            data["id"] = self.id
-        return json.dumps(data)
-
-
-class MessageContents(BaseModel):
-    """
-    Message contents with type, value, metadata, weight, newline, and end of contents
-    """
-
-    type: Union[str, MessageContentTypes]
-    value: Union[str, ToolCallContents, ToolResponseContents, SpecialToken]
-    meta: Optional[dict[str, Any]] = None  # support additional arbitrary metadata
-    weight: Optional[Union[int, float]] = None
-    has_newline: bool = False
-    eoc: bool = False  # end of contents
-
-    def __str__(self) -> str:
-        str_val = str(self.value)
-        if self.has_newline and not str_val.endswith("\n"):
-            str_val += "\n"
-        return str_val
-
-
-class Messages(BaseModel):
-    """
-    Messages with role, content, metadata, weight, and chat formatting
-    """
-
-    role: Union[MessageRoles, str]  # allows for arbitrary roles
-    content: List["MessageContents"]
-    meta: Optional[dict[str, Any]] = None  # support additional arbitrary metadata
-    weight: Optional[Union[int, float]] = None
-    is_chat_formatted: bool = False
-
-    def __str__(self) -> str:
-        return "".join(str(c) for c in self.content)
-
-    def tokenized(
-        self, tokenizer: PreTrainedTokenizer, ignore_index=-100
-    ) -> dict[str, List[int]]:
-        # iterate over the contents, tokenizing the concatenated string values up to the current MessageContents
-        # returns a dictionary mapping w input_ids, attention_mask, and labels
-        input_ids: List[int] = []
-        labels: List[int] = []
-        pending_input_ids: List[int] = []
-        pending_weight = self.weight
-        running_content = ""
-        for _, msg_content in enumerate(self.content):
-            # TODO also handle non-text content types
-            if msg_content.type in [
-                MessageContentTypes.text.value,
-                MessageContentTypes.tool_call.value,
-                MessageContentTypes.tool_response.value,
-            ]:
-                running_content += str(msg_content)
-                tok_results = tokenizer(running_content, add_special_tokens=False)
-                tok_input_ids = tok_results["input_ids"]
-                if pending_input_ids:
-                    new_pending_inputs = tok_input_ids[
-                        len(input_ids) : len(input_ids) + len(pending_input_ids)
-                    ]
-                    if new_pending_inputs != pending_input_ids:
-                        # logging.warning("tokenization mismatch from concatenation.")
-                        pending_input_ids = new_pending_inputs
-                    input_ids.extend(pending_input_ids)
-                    if pending_weight:
-                        labels.extend(pending_input_ids)
-                    else:
-                        labels.extend([ignore_index] * len(pending_input_ids))
-                pending_input_ids = tok_results["input_ids"][len(input_ids) :]
-                pending_weight = self.weight and msg_content.weight not in [0, 0.0]
-        input_ids.extend(pending_input_ids)
-        if pending_weight:
-            labels.extend(pending_input_ids)
-        else:
-            labels.extend([ignore_index] * len(pending_input_ids))
-        attention_mask = [1] * len(input_ids)
-        return {
-            "input_ids": input_ids,
-            "attention_mask": attention_mask,
-            "labels": labels,
-        }
-
-
-class Chats(BaseModel):
-    """
-    top level data structure for chat conversations
-    """
-
-    conversation: List[Messages]
-
-    def __str__(self) -> str:
-        return "".join(str(c) for c in self.conversation)
-
-    def tokenized(
-        self, tokenizer: Callable[[str], dict[str, List[int]]], ignore_index=-100
-    ) -> dict[str, List[int]]:
-        input_ids = []
-        attention_mask = []
-        labels = []
-        for msg in self.conversation:
-            msg_results = msg.tokenized(tokenizer, ignore_index)
-            input_ids.extend(msg_results["input_ids"])
-            attention_mask.extend(msg_results["attention_mask"])
-            labels.extend(msg_results["labels"])
-        return {
-            "input_ids": input_ids,
-            "attention_mask": attention_mask,
-            "labels": labels,
-        }
-
-
-class ChatFormattedChats(Chats):
-    """
-    Chat formatted chats with formatter and optional train on inputs
-    """
-
-    formatter: Callable  # [[Union[dict, Chats]], Chats]
-    train_on_inputs: bool = False
-
-    def model_post_init(self, __context):
-        for i, msg in enumerate(self.conversation):
-            self.conversation[i] = self.formatter(msg, message_index=i)
-            if self.train_on_inputs:
-                self.conversation[i].weight = 1
-
-
-class PreferenceChats(BaseModel):
-    """
-    representation for preference data for chat
-    """
-
-    prompt: List[Messages]
-    chosen: Messages
-    rejected: Messages

src/axolotl/core/datasets/chat.py

@@ -1,55 +0,0 @@
-"""
-chat dataset module
-"""
-import os
-from typing import Callable, Optional, Union
-
-from datasets import Dataset
-from transformers import PreTrainedTokenizer
-
-from axolotl.core.chat.messages import ChatFormattedChats
-
-
-class TokenizedChatDataset(Dataset):
-    """
-    Tokenized chat dataset
-    """
-
-    def __init__(
-        self,
-        data: Dataset,
-        model_transform: Union[PreTrainedTokenizer, Callable],
-        *args,
-        message_transform: Optional[Callable] = None,
-        formatter=None,
-        process_count: Optional[int] = None,
-        keep_in_memory: Optional[bool] = False,
-        **kwargs,
-    ):
-        def map_fn(ex):
-            if message_transform is not None:
-                ex = message_transform(ex)
-            if formatter is not None:
-                ex = ChatFormattedChats(
-                    formatter=formatter,
-                    **ex,
-                )
-            else:
-                ex = ChatFormattedChats(
-                    **ex,
-                )
-            return ex.tokenized(model_transform)
-
-        process_or_cpu_count: int = (
-            process_count or os.cpu_count()  # type: ignore[assignment]
-        )
-        num_proc = min(64, process_or_cpu_count)
-        features = data.features.keys()
-        tokenized_data = data.map(
-            map_fn,
-            num_proc=num_proc,
-            keep_in_memory=keep_in_memory,
-            remove_columns=features,
-            desc="Tokenizing Chats",
-        )
-        super().__init__(tokenized_data.data, *args, **kwargs)

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

@@ -1,150 +0,0 @@
-"""
-This module contains a function that builds a transform that takes a row from the dataset and converts it to a Chat.
-"""
-from typing import Any, Mapping, Union
-
-
-def chat_message_transform_builder(  # pylint: disable=dangerous-default-value
-    train_on_inputs=False,
-    conversations_field: str = "conversations",
-    message_field_role: Union[str, list[str]] = ["role", "from"],  # commonly "role"
-    message_field_content: Union[str, list[str]] = [
-        "value",
-        "text",
-        "content",
-    ],  # commonly "content"
-    message_field_training: Union[str, list[str]] = [
-        "train",
-        "weight",
-    ],  # commonly "weight"
-):
-    """Builds a transform that takes a row from the dataset and converts it to a Chat
-
-    Args:
-        train_on_inputs (bool, optional):
-            If True, the transform will train on the inputs. If False, the transform will train on the targets.
-            Defaults to False.
-        conversations_field (str, optional):
-            The field name of the conversations. Defaults to "conversations".
-        message_field_role (str | list[str], optional):
-            The field name of the role. Defaults to "role".
-        message_field_content (str | list[str], optional):
-            The field name of the message content. Defaults to "content".
-        message_field_training (str | list[str], optional):
-            The field name of the train/weight. Defaults to "weight".
-
-    Returns:
-        Callable:
-            A function that takes a list of conversations and returns a list of messages.
-    """
-
-    message_field_role = (
-        [message_field_role]
-        if isinstance(message_field_role, str)
-        else message_field_role
-    )
-    message_field_content = (
-        [message_field_content]
-        if isinstance(message_field_content, str)
-        else message_field_content
-    )
-    message_weight_fields = (
-        [message_field_training]
-        if isinstance(message_field_training, str)
-        else message_field_training
-    )
-
-    role_value_mappings = {
-        "system": "system",
-        "user": "user",
-        "human": "user",
-        "assistant": "assistant",
-        "gpt": "assistant",
-        "tool": "tool",
-        "ipython": "ipython",
-    }
-    if train_on_inputs:
-        role_default_weights_mappings = {
-            "system": 1,
-            "user": 1,
-            "assistant": 1,
-            "tool": 1,
-            "ipython": 1,
-        }
-    else:
-        role_default_weights_mappings = {
-            "system": 0,
-            "user": 0,
-            "assistant": 1,
-            "tool": 0,
-            "ipython": 0,
-        }
-
-    def transform_builder(sample: Mapping[str, Any]):
-        if conversations_field not in sample:
-            raise ValueError(f"Field '{conversations_field}' not found in sample.")
-        # if none of the role fields are in the message, raise an error
-        if not any(
-            role in sample[conversations_field][0] for role in message_field_role
-        ):
-            raise ValueError("No role field found in message.")
-        role_field = next(
-            role
-            for role in message_field_role
-            if role in sample[conversations_field][0]
-        )
-        if not any(
-            field in sample[conversations_field][0] for field in message_field_content
-        ):
-            raise ValueError("No message_content field found in message.")
-        message_content_field = next(
-            field
-            for field in message_field_content
-            if field in sample[conversations_field][0]
-        )
-        if not any(
-            field in sample[conversations_field][0] for field in message_field_training
-        ):
-            message_weight_field = None
-        else:
-            message_weight_field = next(
-                field
-                for field in message_weight_fields
-                if field in sample[conversations_field][0]
-            )
-
-        messages = []
-        for message in sample[conversations_field]:
-            role = role_value_mappings[message[role_field]]
-            weight = (
-                int(message[message_weight_field])
-                if message_weight_field
-                else role_default_weights_mappings[role]
-            )
-
-            # TODO if "tool_calls" in message[message_content_field]: then convert tool call to ToolCallContents
-            if isinstance(message[message_content_field], str):
-                messages.append(
-                    {
-                        "role": role,
-                        "content": [
-                            {
-                                "type": "text",
-                                "value": message[message_content_field],
-                            }
-                        ],
-                        "weight": weight,
-                    }
-                )
-            else:
-                messages.append(
-                    {
-                        "role": role,
-                        "content": message[message_content_field],
-                        "weight": weight,
-                    }
-                )
-
-        return {"conversation": messages}
-
-    return transform_builder

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,
@@ -43,12 +43,9 @@ from trl import (
     KTOTrainer,
     ORPOConfig,
     ORPOTrainer,
-    RewardConfig,
-    RewardTrainer,
 )
-from trl.trainer.utils import RewardDataCollatorWithPadding, pad_to_length
+from trl.trainer.utils import 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 +61,7 @@ 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,
@@ -304,13 +301,6 @@ class AxolotlCPOConfig(AxolotlTrainingMixins, CPOConfig):
     )
 
 
-@dataclass
-class AxolotlRewardConfig(AxolotlTrainingMixins, RewardConfig):
-    """
-    Reward config for Reward training
-    """
-
-
 class SchedulerMixin(Trainer):
     """
     Mixin class for scheduler setup in CausalTrainer.
@@ -408,10 +398,12 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
     def __init__(
         self,
         *_args,
+        num_epochs=1,
         bench_data_collator=None,
         eval_data_collator=None,
         **kwargs,
     ):
+        self.num_epochs = num_epochs
         self.bench_data_collator = bench_data_collator
         self.eval_data_collator = eval_data_collator
         super().__init__(*_args, **kwargs)
@@ -436,13 +428,7 @@ 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 +497,6 @@ 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 +659,7 @@ 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 +667,8 @@ 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 +764,7 @@ 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 +870,13 @@ 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 +891,6 @@ 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
@@ -1038,46 +997,19 @@ class AxolotlDPOTrainer(SchedulerMixin, DPOTrainer):
 
         return super().push_to_hub(*args, **kwargs)
 
-    @staticmethod
     def tokenize_row(
-        features,
-        processing_class,
-        max_prompt_length,
-        max_completion_length,
-        add_special_tokens,
+        self, feature, model: Optional[Union[PreTrainedModel, torch.nn.Module]] = None
     ) -> Dict:
-        res = DPOTrainer.tokenize_row(
-            features,
-            processing_class,
-            max_prompt_length,
-            max_completion_length,
-            add_special_tokens,
-        )
-        # fix when the tokenizer doesn't have a bos_token_id, e.g. Qwen
-        if processing_class.bos_token is None and res["prompt_input_ids"][0] is None:
+        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:]
-
-        if processing_class.bos_token and processing_class.bos_token_id is not None:
-            # dpo trainer may incorrectly prepend the bos_token_id to the dpo outputs
-            if res["chosen_input_ids"][0] == processing_class.bos_token_id:
-                res["chosen_input_ids"] = res["chosen_input_ids"][1:]
-                res["chosen_labels"] = res["chosen_labels"][1:]
-                res["chosen_attention_mask"] = res["chosen_attention_mask"][1:]
-            if res["rejected_input_ids"][0] == processing_class.bos_token_id:
-                res["rejected_input_ids"] = res["rejected_input_ids"][1:]
-                res["rejected_labels"] = res["rejected_labels"][1:]
-                res["rejected_attention_mask"] = res["rejected_attention_mask"][1:]
-
         return res
 
     def training_step(
-        self,
-        model: nn.Module,
-        inputs: Dict[str, Union[torch.Tensor, Any]],
-        num_items_in_batch=None,
+        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
@@ -1107,14 +1039,6 @@ class AxolotlCPOTrainer(SchedulerMixin, CPOTrainer):
     tag_names = ["axolotl", "cpo"]
 
 
-class AxolotlRewardTrainer(SchedulerMixin, RewardTrainer):
-    """
-    Extend the base RewardTrainer for axolotl helpers
-    """
-
-    tag_names = ["axolotl", "reward"]
-
-
 class TrainerBuilderBase(abc.ABC):
     """
     Base class for trainer builder
@@ -1175,28 +1099,17 @@ 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
@@ -1207,23 +1120,11 @@ 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 = []
-        if self.cfg.plugins:
-            plugin_manager = PluginManager.get_instance()
-            callbacks.extend(
-                [
-                    cb
-                    for cb in plugin_manager.add_callbacks_post_trainer(
-                        self.cfg, trainer
-                    )
-                    if cb
-                ]
-            )
-        return callbacks
 
     def hook_pre_create_training_args(self, training_arguments_kwargs):
         # TODO
@@ -1306,8 +1207,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
 
         if self.cfg.lisa_step_interval and self.cfg.lisa_n_layers:
             callbacks.append(lisa_callback_factory(trainer))
-
-        callbacks.extend(super().get_post_trainer_create_callbacks(trainer=trainer))
         return callbacks
 
     def _get_trainer_cls(self):
@@ -1315,8 +1214,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             return ReLoRATrainer
         if self.cfg.model_config_type == "mamba":
             return AxolotlMambaTrainer
-        if self.cfg.reward_model:
-            return AxolotlRewardTrainer
         return AxolotlTrainer
 
     def build(self, total_num_steps):
@@ -1425,15 +1322,17 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
 
         if not self.cfg.test_datasets and self.cfg.val_set_size == 0:
             # no eval set, so don't eval
-            training_arguments_kwargs["eval_strategy"] = "no"
+            training_arguments_kwargs["evaluation_strategy"] = "no"
         elif self.cfg.eval_steps:
-            training_arguments_kwargs["eval_strategy"] = "steps"
+            training_arguments_kwargs["evaluation_strategy"] = "steps"
             training_arguments_kwargs["eval_steps"] = self.cfg.eval_steps
-        elif self.cfg.eval_strategy:
-            training_arguments_kwargs["eval_strategy"] = self.cfg.eval_strategy
+        elif self.cfg.evaluation_strategy:
+            training_arguments_kwargs[
+                "evaluation_strategy"
+            ] = self.cfg.evaluation_strategy
         else:
             # we have an eval set, but no steps defined, default to use epoch
-            training_arguments_kwargs["eval_strategy"] = "epoch"
+            training_arguments_kwargs["evaluation_strategy"] = "epoch"
 
         if self.cfg.save_steps:
             training_arguments_kwargs["save_strategy"] = "steps"
@@ -1546,12 +1445,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             report_to.append("comet_ml")
 
         training_arguments_kwargs["report_to"] = report_to
-        if self.cfg.use_wandb:
-            training_arguments_kwargs["run_name"] = self.cfg.wandb_name
-        elif self.cfg.use_mlflow:
-            training_arguments_kwargs["run_name"] = self.cfg.mlflow_run_name
-        else:
-            training_arguments_kwargs["run_name"] = None
+        training_arguments_kwargs["run_name"] = (
+            self.cfg.wandb_name if self.cfg.use_wandb else None
+        )
         training_arguments_kwargs["optim"] = (
             self.cfg.optimizer if self.cfg.optimizer else "adamw_hf"
         )
@@ -1640,9 +1536,8 @@ 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":
@@ -1655,16 +1550,11 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
 
         trainer_kwargs = {}
 
-        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"
@@ -1703,13 +1593,10 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 "accelerator_config"
             ] = self.cfg.accelerator_config
 
-        training_args_cls = (
-            AxolotlTrainingArguments
-            if not self.cfg.reward_model
-            else AxolotlRewardConfig
-        )
-        training_args = training_args_cls(  # pylint: disable=unexpected-keyword-arg
-            **training_arguments_kwargs,
+        training_args = (
+            AxolotlTrainingArguments(  # pylint: disable=unexpected-keyword-arg
+                **training_arguments_kwargs,
+            )
         )
         training_args = self.hook_post_create_training_args(training_args)
 
@@ -1731,37 +1618,27 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             # https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html
             data_collator_kwargs["pad_to_multiple_of"] = 64
 
-        if self.cfg.reward_model:
-            data_collator_kwargs["max_length"] = self.cfg.sequence_len
-
         trainer_cls = self._get_trainer_cls()
         trainer_kwargs, trainer_cls = self.hook_pre_create_trainer(
             trainer_kwargs, trainer_cls
         )
-        if eval_data_collator := self.build_collator(
-            training_args, is_eval=True, **data_collator_kwargs
-        ):
-            if not self.cfg.reward_model:
-                trainer_kwargs["eval_data_collator"] = eval_data_collator
-        if not self.cfg.reward_model:
-            trainer_kwargs["bench_data_collator"] = transformers.DataCollatorForSeq2Seq(
-                self.tokenizer,
-                return_tensors="pt",
-                **data_collator_kwargs,
-            )
-        sig = inspect.signature(trainer_cls)
-        if "processing_class" in sig.parameters.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),
+            eval_data_collator=self.build_collator(
+                training_args, is_eval=True, **data_collator_kwargs
+            ),
+            bench_data_collator=transformers.DataCollatorForSeq2Seq(
+                self.tokenizer,
+                return_tensors="pt",
+                **data_collator_kwargs,
+            ),
             callbacks=self.get_callbacks(),
+            num_epochs=self.cfg.num_epochs,
             **trainer_kwargs,
         )
         trainer = self.hook_post_create_trainer(trainer)
@@ -1795,14 +1672,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 V2BatchSamplerDataCollatorForSeq2Seq,
                 BatchSamplerDataCollatorForSeq2Seq,
                 DataCollatorForSeq2Seq,
-                RewardDataCollatorWithPadding,
             ]
         ]
-        if self.cfg.reward_model:
-            collator = RewardDataCollatorWithPadding
-            if "max_length" in kwargs:
-                kwargs.pop("max_length")
-        elif use_batch_sampler_collator:
+        if use_batch_sampler_collator:
             if self.cfg.model_config_type in SUPPORTED_MULTIPACK_MODEL_TYPES:
                 collator = V2BatchSamplerDataCollatorForSeq2Seq
             elif (
@@ -1839,7 +1711,7 @@ 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):
@@ -1867,10 +1739,10 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             training_args_kwargs["save_safetensors"] = self.cfg.save_safetensors
 
         if self.eval_dataset:
-            training_args_kwargs["eval_strategy"] = "steps"
+            training_args_kwargs["evaluation_strategy"] = "steps"
             training_args_kwargs["eval_steps"] = self.cfg.eval_steps
         else:
-            training_args_kwargs["eval_strategy"] = "no"
+            training_args_kwargs["evaluation_strategy"] = "no"
 
         if self.cfg.bf16 or self.cfg.bfloat16:
             training_args_kwargs["bf16"] = True
@@ -1925,18 +1797,17 @@ 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"
@@ -1945,13 +1816,13 @@ 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"] = (
@@ -1966,17 +1837,6 @@ 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,
@@ -1997,6 +1857,7 @@ 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:
@@ -2010,6 +1871,12 @@ 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]
@@ -2021,17 +1888,11 @@ 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,
         )
@@ -2053,11 +1914,11 @@ 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/integrations/sageattention/lib/core.py

@@ -1,361 +0,0 @@
-"""
-Copyright (c) 2024 by SageAttention team.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-"""
-
-from typing import Any, Optional
-
-import torch
-from torch.autograd import Function
-
-from .triton.attn_qk_int8_per_block_causal_varlen import (
-    backward as sageattn_varlen_backward,
-)
-from .triton.attn_qk_int8_per_block_causal_varlen import forward as attn_true_varlen
-from .triton.quant_per_block_varlen import (
-    per_block_int8 as per_block_int8_varlen_triton,
-)
-
-
-def get_cuda_arch_versions():
-    cuda_archs = []
-    for i in range(torch.cuda.device_count()):
-        major, minor = torch.cuda.get_device_capability(i)
-        cuda_archs.append(f"sm{major}{minor}")
-    return cuda_archs
-
-
-def sageattn_varlen(
-    q: torch.Tensor,
-    k: torch.Tensor,
-    v: torch.Tensor,
-    cu_seqlens_q: torch.Tensor,
-    cu_seqlens_k: torch.Tensor,
-    max_seqlen_q: int,
-    max_seqlen_k: int,
-    sm_scale: Optional[float] = None,
-    smooth_k: bool = True,
-    **kwargs: Any,
-) -> torch.Tensor:
-    """
-
-    Parameters
-    ----------
-    q : torch.Tensor
-        The query tensor, shape: ``[cu_seqlens_q[-1], num_qo_heads, head_dim]``.
-
-    k : torch.Tensor
-        The key tensor, shape: ``[cu_seqlens_k[-1], num_kv_heads, head_dim]``.
-
-    v : torch.Tensor
-        The value tensor, shape: ``[cu_seqlens_k[-1], num_kv_heads, head_dim]``.
-
-    cu_seqlens_q : torch.Tensor
-        The cumulative sequence lengths for the query sequences in the batch, used to index into `q`.
-        Shape: ``[batch_size + 1]``, where each entry represents the cumulative length of sequences up to that batch index.
-
-    cu_seqlens_k : torch.Tensor
-        The cumulative sequence lengths for the key and value sequences in the batch, used to index into `k` and `v`.
-        Shape: ``[batch_size + 1]``, where each entry represents the cumulative length of sequences up to that batch index.
-
-    max_seqlen_q : int
-        The maximum sequence length for the query tensor in the batch.
-
-    max_seqlen_k : int
-        The maximum sequence length for the key and value tensors in the batch.
-
-    is_causal : bool
-        Whether to apply causal mask to the attention matrix. Only applicable when qo_len == kv_len for each sequence.
-        Default: False.
-
-    sm_scale : Optional[float]
-        The scale used in softmax, if not provided, will be set to ``1.0 / sqrt(head_dim)``.
-
-    smooth_k : bool
-        Whether to smooth the key tensor by subtracting the mean along the sequence dimension.
-        Default: True.
-
-    Returns
-    -------
-    torch.Tensor
-        The output tensor, shape: ``[cu_seqlens_q[-1], num_qo_heads, head_dim]``.
-
-    Note
-    ----
-    - ``num_qo_heads`` must be divisible by ``num_kv_heads``.
-    - The tensors `q`, `k`, and `v` must have the dtype ``torch.float16``, ``torch.bfloat16`` or ``torch.float32``.
-    - The tensors `cu_seqlens_q` and `cu_seqlens_k` must have the dtype ``torch.int32`` or ``torch.int64``.
-    - All tensors must be on the same cuda device.
-    - `smooth_k` will introduce slight overhead but will improve the accuracy under most circumstances.
-    """
-
-    dtype = q.dtype
-    assert q.is_cuda, "Input tensors must be on cuda."
-    assert dtype in [
-        torch.float16,
-        torch.bfloat16,
-    ], "Input tensors must be in dtype of torch.float16 or torch.bfloat16"
-    assert q.device == k.device == v.device, "All tensors must be on the same device."
-    assert q.dtype == k.dtype == v.dtype, "All tensors must have the same dtype."
-
-    head_dim = q.size(-1)
-    assert head_dim in [64, 128], "varlen only support head_dim [64, 128]."
-
-    assert (
-        q.stride(-1) == 1 and k.stride(-1) == 1 and v.stride(-1) == 1
-    ), "Last dim of qkv must be contiguous."
-    assert (
-        cu_seqlens_q.is_contiguous() and cu_seqlens_k.is_contiguous()
-    ), "cu_seqlens_q and cu_seqlens_k must be contiguous."
-
-    if dtype == torch.bfloat16 or dtype == torch.float32:
-        v = v.to(torch.float16)
-
-    if smooth_k:
-        km = k.mean(
-            dim=0, keepdim=True
-        )  # ! km is calculated on the all the batches. Calculate over each individual sequence requires dedicated kernel.
-        k -= km
-
-    (
-        q_int8,
-        q_scale,
-        k_int8,
-        k_scale,
-        cu_seqlens_q_scale,
-        cu_seqlens_k_scale,
-    ) = per_block_int8_varlen_triton(
-        q, k, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k, sm_scale=sm_scale
-    )
-
-    o = attn_true_varlen(
-        q_int8,
-        k_int8,
-        v,
-        cu_seqlens_q,
-        cu_seqlens_k,
-        max_seqlen_q,
-        q_scale,
-        k_scale,
-        cu_seqlens_q_scale,
-        cu_seqlens_k_scale,
-        output_dtype=dtype,
-    )
-
-    return o
-
-
-class SageAttentionFunction(Function):
-    @staticmethod
-    def forward(
-        ctx,
-        query,
-        key,
-        value,
-        attn_mask=None,
-        dropout_p=0.0,
-        is_causal=False,
-        scale=None,
-    ):
-        """
-        query: Tensor of shape [batch_size, num_heads, seq_len_q, head_dim]
-        key: Tensor of shape [batch_size, num_heads, seq_len_k, head_dim]
-        value: Tensor of shape [batch_size, num_heads, seq_len_k, head_dim]
-        attn_mask: Optional[Tensor], mask tensor
-        dropout_p: float, dropout probability
-        is_causal: bool, whether to apply causal masking
-        scale: Optional[float], scaling factor for attention scores
-        """
-        # Ensure inputs are contiguous
-        query = query.contiguous()
-        key = key.contiguous()
-        value = value.contiguous()
-
-        # Handle default scale
-        if scale is None:
-            scale = 1.0 / (query.size(-1) ** 0.5)
-
-        # Save parameters needed for backward
-        ctx.scale = scale
-        ctx.is_causal = is_causal
-        ctx.dropout_p = dropout_p
-        ctx.attn_mask = attn_mask
-
-        # Prepare cumulative sequence lengths and max sequence lengths
-        # Assuming batch sizes are consistent across query, key, and value
-        batch_size, num_heads, seq_len_q, head_dim = query.shape
-        seq_len_k = key.shape[2]
-
-        # Flatten batch and head dimensions
-        q = query.view(
-            -1, seq_len_q, head_dim
-        )  # [batch_size * num_heads, seq_len_q, head_dim]
-        k = key.view(-1, seq_len_k, head_dim)
-        v = value.view(-1, seq_len_k, head_dim)
-
-        # Create cumulative sequence lengths
-        cu_seqlens_q = torch.arange(
-            0,
-            (batch_size * num_heads + 1) * seq_len_q,
-            seq_len_q,
-            dtype=torch.int32,
-            device=query.device,
-        )
-        cu_seqlens_k = torch.arange(
-            0,
-            (batch_size * num_heads + 1) * seq_len_k,
-            seq_len_k,
-            dtype=torch.int32,
-            device=key.device,
-        )
-        max_seqlen_q = seq_len_q
-        max_seqlen_k = seq_len_k
-
-        # Call your custom per-block int8 quantization function
-        (
-            q_int8,
-            q_scale,
-            k_int8,
-            k_scale,
-            cu_seqlens_q_scale,
-            cu_seqlens_k_scale,
-        ) = per_block_int8_varlen_triton(
-            q, k, cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k, sm_scale=scale
-        )
-
-        # Call your custom attention function
-        if is_causal:
-            output = attn_true_varlen(
-                q_int8,
-                k_int8,
-                v,
-                cu_seqlens_q,
-                cu_seqlens_k,
-                max_seqlen_q,
-                q_scale,
-                k_scale,
-                cu_seqlens_q_scale,
-                cu_seqlens_k_scale,
-                output_dtype=query.dtype,
-            )
-        else:
-            raise NotImplementedError("Non-causal attention is not implemented yet.")
-
-        # Reshape output to match the expected shape
-        output = output.view(batch_size, num_heads, seq_len_q, head_dim)
-
-        # Save tensors for backward
-        ctx.save_for_backward(
-            query,
-            key,
-            value,
-            q_int8,
-            k_int8,
-            q_scale,
-            k_scale,
-            cu_seqlens_q,
-            cu_seqlens_k,
-            cu_seqlens_q_scale,
-            cu_seqlens_k_scale,
-            output,
-        )
-
-        return output
-
-    @staticmethod
-    def backward(ctx, grad_output):
-        (
-            query,
-            key,
-            value,
-            q_int8,
-            k_int8,
-            q_scale,
-            k_scale,
-            cu_seqlens_q,
-            cu_seqlens_k,
-            cu_seqlens_q_scale,
-            cu_seqlens_k_scale,
-            output,
-        ) = ctx.saved_tensors
-
-        scale = ctx.scale
-        is_causal = ctx.is_causal
-        dropout_p = ctx.dropout_p
-        attn_mask = ctx.attn_mask
-
-        # Flatten batch and head dimensions
-        batch_size, num_heads, seq_len_q, head_dim = query.shape
-        seq_len_k = key.shape[2]
-        grad_output = grad_output.contiguous()
-        do = grad_output.view(-1, seq_len_q, head_dim)
-
-        # Compute gradients w.r.t. q, k, v
-        dq, dk, dv = sageattn_varlen_backward(
-            do,
-            query.view(-1, seq_len_q, head_dim),
-            key.view(-1, seq_len_k, head_dim),
-            value.view(-1, seq_len_k, head_dim),
-            cu_seqlens_q,
-            cu_seqlens_k,
-            seq_len_q,
-            seq_len_k,
-            q_int8,
-            k_int8,
-            q_scale,
-            k_scale,
-            cu_seqlens_q_scale,
-            cu_seqlens_k_scale,
-            scale,
-            is_causal,
-        )
-
-        # Reshape gradients to match the input shapes
-        dq = dq.view(batch_size, num_heads, seq_len_q, head_dim)
-        dk = dk.view(batch_size, num_heads, seq_len_k, head_dim)
-        dv = dv.view(batch_size, num_heads, seq_len_k, head_dim)
-
-        # Handle optional arguments
-        d_attn_mask = None  # Assuming attn_mask does not require gradients
-        d_dropout_p = (
-            None  # Dropout probability is a hyperparameter, typically not optimized
-        )
-        d_is_causal = None  # Not differentiable
-        d_scale = None  # If scale is a tensor and requires grad, compute its gradient
-
-        return dq, dk, dv, d_attn_mask, d_dropout_p, d_is_causal, d_scale
-
-
-def scaled_dot_product_attention(
-    query,
-    key,
-    value,
-    attn_mask=None,
-    dropout_p=0.0,
-    is_causal=False,
-    scale=None,
-):
-    """
-    Custom scaled dot product attention using SageAttentionFunction.
-    """
-    return SageAttentionFunction.apply(
-        query, key, value, attn_mask, dropout_p, is_causal, scale
-    )
-
-
-def monkeypatch_sdp_w_sage_attention():
-    """
-    Replace torch.nn.functional.scaled_dot_product_attention with custom scaled dot product attention using SageAttentionFunction.
-    """
-    torch.nn.functional.scaled_dot_product_attention = scaled_dot_product_attention

src/axolotl/integrations/sageattention/lib/triton/attn_qk_int8_per_block_causal_varlen.py

@@ -1,622 +0,0 @@
-"""
-Copyright (c) 2024 by SageAttention team.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-"""
-
-import math
-
-import torch
-import triton
-import triton.language as tl
-
-
-@triton.jit
-def _attn_fwd_inner(
-    acc,
-    l_i,
-    m_i,
-    q,
-    q_scale,
-    kv_len,
-    K_ptrs,
-    K_scale_ptr,
-    V_ptrs,
-    stride_kn,
-    stride_vn,
-    start_m,
-    H: tl.constexpr,
-    BLOCK_M: tl.constexpr,
-    HEAD_DIM: tl.constexpr,
-    BLOCK_N: tl.constexpr,
-    STAGE: tl.constexpr,
-    offs_m: tl.constexpr,
-    offs_n: tl.constexpr,
-):
-    if STAGE == 1:
-        lo, hi = 0, start_m * BLOCK_M
-    elif STAGE == 2:
-        lo, hi = start_m * BLOCK_M, (start_m + 1) * BLOCK_M
-        lo = tl.multiple_of(lo, BLOCK_M)
-        K_scale_ptr += (lo // BLOCK_N) * H
-        K_ptrs += stride_kn * lo
-        V_ptrs += stride_vn * lo
-    for start_n in range(lo, hi, BLOCK_N):
-        start_n = tl.multiple_of(start_n, BLOCK_N)
-        k_mask = offs_n[None, :] < (kv_len - start_n)
-        k = tl.load(K_ptrs, mask=k_mask)
-        k_scale = tl.load(K_scale_ptr)
-        qk = tl.dot(q, k).to(tl.float32) * q_scale * k_scale
-
-        if STAGE == 2:
-            mask = offs_m[:, None] >= (start_n + offs_n[None, :])
-            qk = qk + tl.where(mask, 0, -1.0e6)
-            m_ij = tl.maximum(m_i, tl.max(qk, 1))
-            qk -= m_ij[:, None]
-        else:
-            m_ij = tl.maximum(m_i, tl.max(qk, 1))
-            qk = qk - m_ij[:, None]
-
-        p = tl.math.exp2(qk)
-        l_ij = tl.sum(p, 1)
-
-        alpha = tl.math.exp2(m_i - m_ij)
-        l_i = l_i * alpha + l_ij
-
-        acc = acc * alpha[:, None]
-
-        v = tl.load(V_ptrs, mask=offs_n[:, None] < (kv_len - start_n))
-        p = p.to(tl.float16)
-
-        acc += tl.dot(p, v, out_dtype=tl.float16)
-        m_i = m_ij
-        K_ptrs += BLOCK_N * stride_kn
-        K_scale_ptr += H
-        V_ptrs += BLOCK_N * stride_vn
-    return acc, l_i, m_i
-
-
-@triton.jit
-def _attn_fwd(
-    Q,
-    K,
-    V,
-    cu_seqlens_q,
-    cu_seqlens_k,
-    Q_scale,
-    K_scale,
-    cu_seqlens_q_scale,
-    cu_seqlens_k_scale,
-    Out,
-    stride_qh,
-    stride_qn,
-    stride_kh,
-    stride_kn,
-    stride_vh,
-    stride_vn,
-    stride_oh,
-    stride_on,
-    H: tl.constexpr,
-    num_kv_groups: tl.constexpr,
-    HEAD_DIM: tl.constexpr,
-    BLOCK_M: tl.constexpr,
-    BLOCK_N: tl.constexpr,
-    STAGE: tl.constexpr,
-):
-    start_m = tl.program_id(0)
-
-    off_z = tl.program_id(2).to(tl.int64)
-    off_h = tl.program_id(1).to(tl.int64)
-
-    cu_seqlens_q_start = tl.load(cu_seqlens_q + off_z)
-    cu_seqlens_q_end = tl.load(cu_seqlens_q + off_z + 1)
-
-    qo_len = cu_seqlens_q_end - cu_seqlens_q_start
-
-    if (start_m * BLOCK_M) >= qo_len:
-        return
-
-    cu_seq_lens_q_scale_start = tl.load(cu_seqlens_q_scale + off_z)
-    cu_seq_lens_k_scale_start = tl.load(cu_seqlens_k_scale + off_z)
-
-    q_scale_offset = cu_seq_lens_q_scale_start * H + off_h + start_m * H
-    k_scale_offset = (
-        cu_seq_lens_k_scale_start * (H // num_kv_groups) + off_h // num_kv_groups
-    )
-
-    cu_seqlens_k_start = tl.load(cu_seqlens_k + off_z)
-    cu_seqlens_k_end = tl.load(cu_seqlens_k + off_z + 1)
-
-    kv_len = cu_seqlens_k_end - cu_seqlens_k_start
-
-    offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
-    offs_n = tl.arange(0, BLOCK_N)
-    offs_k = tl.arange(0, HEAD_DIM)
-    Q_ptrs = (
-        Q
-        + (cu_seqlens_q_start * stride_qn + off_h * stride_qh)
-        + offs_m[:, None] * stride_qn
-        + offs_k[None, :]
-    )
-    Q_scale_ptr = Q_scale + q_scale_offset
-    K_ptrs = (
-        K
-        + (cu_seqlens_k_start * stride_kn + (off_h // num_kv_groups) * stride_kh)
-        + offs_n[None, :] * stride_kn
-        + offs_k[:, None]
-    )
-    K_scale_ptr = K_scale + k_scale_offset
-    V_ptrs = (
-        V
-        + (cu_seqlens_k_start * stride_vn + (off_h // num_kv_groups) * stride_vh)
-        + offs_n[:, None] * stride_vn
-        + offs_k[None, :]
-    )
-    O_block_ptr = (
-        Out
-        + (cu_seqlens_q_start * stride_on + off_h * stride_oh)
-        + offs_m[:, None] * stride_on
-        + offs_k[None, :]
-    )
-
-    m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")
-    l_i = tl.zeros([BLOCK_M], dtype=tl.float32) + 1.0
-    acc = tl.zeros([BLOCK_M, HEAD_DIM], dtype=tl.float32)
-
-    q = tl.load(Q_ptrs, mask=offs_m[:, None] < qo_len)
-    q_scale = tl.load(Q_scale_ptr)
-    acc, l_i, m_i = _attn_fwd_inner(
-        acc,
-        l_i,
-        m_i,
-        q,
-        q_scale,
-        kv_len,
-        K_ptrs,
-        K_scale_ptr,
-        V_ptrs,
-        stride_kn,
-        stride_vn,
-        start_m,
-        H // num_kv_groups,
-        BLOCK_M,
-        HEAD_DIM,
-        BLOCK_N,
-        4 - STAGE,
-        offs_m,
-        offs_n,
-    )
-
-    acc, l_i, _ = _attn_fwd_inner(
-        acc,
-        l_i,
-        m_i,
-        q,
-        q_scale,
-        kv_len,
-        K_ptrs,
-        K_scale_ptr,
-        V_ptrs,
-        stride_kn,
-        stride_vn,
-        start_m,
-        H // num_kv_groups,
-        BLOCK_M,
-        HEAD_DIM,
-        BLOCK_N,
-        2,
-        offs_m,
-        offs_n,
-    )
-    acc = acc / l_i[:, None]
-    tl.store(O_block_ptr, acc.to(Out.type.element_ty), mask=(offs_m[:, None] < qo_len))
-
-
-@triton.jit
-def _attn_bwd_inner(
-    dq_acc,
-    dk_acc,
-    dv_acc,
-    l_i,
-    m_i,
-    q,
-    k,
-    v,
-    do,
-    q_scale,
-    k_scale,
-    kv_len,
-    stride_kn,
-    stride_vn,
-    start_m,
-    H,
-    BLOCK_M: tl.constexpr,
-    HEAD_DIM: tl.constexpr,
-    BLOCK_N: tl.constexpr,
-    STAGE: tl.constexpr,
-    offs_m: tl.constexpr,
-    offs_n: tl.constexpr,
-):
-    if STAGE == 1:
-        lo, hi = 0, start_m * BLOCK_M
-    elif STAGE == 2:
-        lo, hi = start_m * BLOCK_M, (start_m + 1) * BLOCK_M
-        lo = tl.multiple_of(lo, BLOCK_M)
-        k += stride_kn * lo
-        v += stride_vn * lo
-
-    for start_n in range(lo, hi, BLOCK_N):
-        start_n = tl.multiple_of(start_n, BLOCK_N)
-        k_mask = offs_n[None, :] < (kv_len - start_n)
-        k_curr = tl.load(k, mask=k_mask)
-        v_curr = tl.load(v, mask=k_mask)
-        k_scale_curr = tl.load(k_scale)
-        s = tl.dot(q, k_curr, trans_b=True).to(tl.float32) * q_scale * k_scale_curr
-
-        if STAGE == 2:
-            mask = offs_m[:, None] >= (start_n + offs_n[None, :])
-            s = s + tl.where(mask, 0.0, -float("inf"))
-            m_ij = tl.maximum(m_i, tl.max(s, 1))
-            s = s - m_ij[:, None]
-        else:
-            m_ij = tl.maximum(m_i, tl.max(s, 1))
-            s = s - m_ij[:, None]
-
-        p = tl.math.exp2(s)
-        l_ij = tl.sum(p, 1)
-        alpha = tl.math.exp2(m_i - m_ij)
-        l_i = l_i * alpha + l_ij
-        m_i = m_ij
-
-        p = p / l_i[:, None]  # Normalize probabilities
-
-        # Compute gradients
-        # Compute softmax gradient
-        do_scaled = do / l_i[:, None]
-        dv_contrib = tl.dot(p.to(tl.float16).T, do_scaled.to(tl.float16))
-        dv_acc += dv_contrib
-
-        dp = tl.dot(do_scaled.to(tl.float16), v_curr.to(tl.float16).T)
-
-        # Compute ds (gradient w.r.t. logits s)
-        p_dp = p * dp
-        sum_p_dp = tl.sum(p_dp, axis=1)
-        ds = (p_dp - p * sum_p_dp[:, None]) * tl.math.log(2.0)  # Adjust for exp2
-
-        # Compute gradients w.r.t q and k
-        dq_contrib = tl.dot(ds.to(tl.float16), k_curr.to(tl.float16))
-        dk_contrib = tl.dot(ds.to(tl.float16).T, q.to(tl.float16))
-
-        dq_acc += dq_contrib * (q_scale * k_scale_curr)
-        dk_acc += dk_contrib * (q_scale * k_scale_curr)
-
-        k += BLOCK_N * stride_kn
-        k_scale += H
-        v += BLOCK_N * stride_vn
-
-    return dq_acc, dk_acc, dv_acc, l_i, m_i
-
-
-@triton.jit
-def _attn_bwd(
-    DO,
-    Q,
-    K,
-    V,
-    cu_seqlens_q,
-    cu_seqlens_k,
-    Q_scale,
-    K_scale,
-    cu_seqlens_q_scale,
-    cu_seqlens_k_scale,
-    L,
-    M,
-    DQ,
-    DK,
-    DV,
-    stride_qh,
-    stride_qn,
-    stride_kh,
-    stride_kn,
-    stride_vh,
-    stride_vn,
-    H: tl.constexpr,
-    num_kv_groups: tl.constexpr,
-    HEAD_DIM: tl.constexpr,
-    BLOCK_M: tl.constexpr,
-    BLOCK_N: tl.constexpr,
-    STAGE: tl.constexpr,
-):
-    start_m = tl.program_id(0)
-    off_z = tl.program_id(2).to(tl.int64)
-    off_h = tl.program_id(1).to(tl.int64)
-
-    cu_seqlens_q_start = tl.load(cu_seqlens_q + off_z)
-    cu_seqlens_q_end = tl.load(cu_seqlens_q + off_z + 1)
-    qo_len = cu_seqlens_q_end - cu_seqlens_q_start
-
-    if (start_m * BLOCK_M) >= qo_len:
-        return
-
-    cu_seq_lens_q_scale_start = tl.load(cu_seqlens_q_scale + off_z)
-    cu_seq_lens_k_scale_start = tl.load(cu_seqlens_k_scale + off_z)
-
-    q_scale_offset = cu_seq_lens_q_scale_start * H + off_h + start_m * H
-    k_scale_offset = (
-        cu_seq_lens_k_scale_start * (H // num_kv_groups) + off_h // num_kv_groups
-    )
-
-    cu_seqlens_k_start = tl.load(cu_seqlens_k + off_z)
-    cu_seqlens_k_end = tl.load(cu_seqlens_k + off_z + 1)
-    kv_len = cu_seqlens_k_end - cu_seqlens_k_start
-
-    offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
-    offs_n = tl.arange(0, BLOCK_N)
-    offs_k = tl.arange(0, HEAD_DIM)
-    Q_ptrs = (
-        Q
-        + (cu_seqlens_q_start * stride_qn + off_h * stride_qh)
-        + offs_m[:, None] * stride_qn
-        + offs_k[None, :]
-    )
-    DO_ptrs = (
-        DO
-        + (cu_seqlens_q_start * stride_qn + off_h * stride_qh)
-        + offs_m[:, None] * stride_qn
-        + offs_k[None, :]
-    )
-    Q_scale_ptr = Q_scale + q_scale_offset
-    K_ptrs = (
-        K
-        + (cu_seqlens_k_start * stride_kn + (off_h // num_kv_groups) * stride_kh)
-        + offs_n[None, :] * stride_kn
-        + offs_k[:, None]
-    )
-    K_scale_ptr = K_scale + k_scale_offset
-    V_ptrs = (
-        V
-        + (cu_seqlens_k_start * stride_vn + (off_h // num_kv_groups) * stride_vh)
-        + offs_n[:, None] * stride_vn
-        + offs_k[None, :]
-    )
-    DQ_ptrs = (
-        DQ
-        + (cu_seqlens_q_start * stride_qn + off_h * stride_qh)
-        + offs_m[:, None] * stride_qn
-        + offs_k[None, :]
-    )
-    DK_ptrs = (
-        DK
-        + (cu_seqlens_k_start * stride_kn + (off_h // num_kv_groups) * stride_kh)
-        + offs_n[None, :] * stride_kn
-        + offs_k[:, None]
-    )
-    DV_ptrs = (
-        DV
-        + (cu_seqlens_k_start * stride_vn + (off_h // num_kv_groups) * stride_vh)
-        + offs_n[:, None] * stride_vn
-        + offs_k[None, :]
-    )
-    L_ptrs = L + (cu_seqlens_q_start + offs_m)
-    M_ptrs = M + (cu_seqlens_q_start + offs_m)
-
-    m_i = tl.load(M_ptrs, mask=offs_m < qo_len, other=float("-inf"))
-    l_i = tl.load(L_ptrs, mask=offs_m < qo_len, other=1.0)
-
-    dq_acc = tl.zeros([BLOCK_M, HEAD_DIM], dtype=tl.float32)
-    dk_acc = tl.zeros([BLOCK_N, HEAD_DIM], dtype=tl.float32)
-    dv_acc = tl.zeros([BLOCK_N, HEAD_DIM], dtype=tl.float32)
-
-    q = tl.load(Q_ptrs, mask=offs_m[:, None] < qo_len)
-    do = tl.load(DO_ptrs, mask=offs_m[:, None] < qo_len)
-    q_scale = tl.load(Q_scale_ptr)
-
-    dq_acc, dk_acc, dv_acc, l_i, m_i = _attn_bwd_inner(
-        dq_acc,
-        dk_acc,
-        dv_acc,
-        l_i,
-        m_i,
-        q,
-        K_ptrs,
-        V_ptrs,
-        do,
-        q_scale,
-        K_scale_ptr,
-        kv_len,
-        stride_kn,
-        stride_vn,
-        start_m,
-        H // num_kv_groups,
-        BLOCK_M,
-        HEAD_DIM,
-        BLOCK_N,
-        4 - STAGE,
-        offs_m,
-        offs_n,
-    )
-
-    dq_acc, dk_acc, dv_acc, l_i, m_i = _attn_bwd_inner(
-        dq_acc,
-        dk_acc,
-        dv_acc,
-        l_i,
-        m_i,
-        q,
-        K_ptrs,
-        V_ptrs,
-        do,
-        q_scale,
-        K_scale_ptr,
-        kv_len,
-        stride_kn,
-        stride_vn,
-        start_m,
-        H // num_kv_groups,
-        BLOCK_M,
-        HEAD_DIM,
-        BLOCK_N,
-        2,
-        offs_m,
-        offs_n,
-    )
-
-    tl.store(DQ_ptrs, dq_acc.to(DQ.dtype.element_ty), mask=offs_m[:, None] < qo_len)
-    tl.store(DK_ptrs, dk_acc.to(DK.dtype.element_ty), mask=offs_n[None, :] < kv_len)
-    tl.store(DV_ptrs, dv_acc.to(DV.dtype.element_ty), mask=offs_n[:, None] < kv_len)
-
-
-def forward(
-    q,
-    k,
-    v,
-    cu_seqlens_q,
-    cu_seqlens_k,
-    max_seqlen_q,
-    q_scale,
-    k_scale,
-    cu_seqlens_q_scale,
-    cu_seqlens_k_scale,
-    output_dtype=torch.float16,
-):
-    BLOCK_M = 128
-    BLOCK_N = 64
-    stage = 3
-
-    o = torch.empty(q.shape, dtype=output_dtype, device=q.device)
-
-    b = cu_seqlens_q.shape[0] - 1
-    _, h_qo, head_dim = q.shape
-    _, h_kv, _ = k.shape
-
-    HEAD_DIM_K = head_dim
-    num_kv_groups = h_qo // h_kv
-
-    grid = (triton.cdiv(max_seqlen_q, BLOCK_M), h_qo, b)
-    _attn_fwd[grid](
-        q,
-        k,
-        v,
-        cu_seqlens_q,
-        cu_seqlens_k,
-        q_scale,
-        k_scale,
-        cu_seqlens_q_scale,
-        cu_seqlens_k_scale,
-        o,
-        q.stride(1),
-        q.stride(0),
-        k.stride(1),
-        k.stride(0),
-        v.stride(1),
-        v.stride(0),
-        o.stride(1),
-        o.stride(0),
-        h_qo,
-        num_kv_groups,
-        BLOCK_M=BLOCK_M,
-        BLOCK_N=BLOCK_N,
-        HEAD_DIM=HEAD_DIM_K,
-        STAGE=stage,
-        num_warps=4 if head_dim == 64 else 8,
-        num_stages=4,
-    )
-    return o
-
-
-def backward(
-    do,
-    q,
-    k,
-    v,
-    cu_seqlens_q,
-    cu_seqlens_k,
-    max_seqlen_q,
-    q_scale,
-    k_scale,
-    cu_seqlens_q_scale,
-    cu_seqlens_k_scale,
-    l,
-    m,
-    output_dtype=torch.float16,
-):
-    BLOCK_M = 128
-    BLOCK_N = 64
-    stage = 3
-
-    device = q.device
-    dtype = q.dtype
-    b = cu_seqlens_q.shape[0] - 1
-    _, h_qo, head_dim = q.shape
-    _, h_kv, _ = k.shape
-    num_kv_groups = h_qo // h_kv
-
-    dq = torch.zeros_like(q, dtype=output_dtype)
-    dk = torch.zeros_like(k, dtype=output_dtype)
-    dv = torch.zeros_like(v, dtype=output_dtype)
-
-    grid = (triton.cdiv(max_seqlen_q, BLOCK_M), h_qo, b)
-    _attn_bwd[grid](
-        do,
-        q,
-        k,
-        v,
-        cu_seqlens_q,
-        cu_seqlens_k,
-        q_scale,
-        k_scale,
-        cu_seqlens_q_scale,
-        cu_seqlens_k_scale,
-        l,
-        m,
-        dq,
-        dk,
-        dv,
-        q.stride(1),
-        q.stride(0),
-        k.stride(1),
-        k.stride(0),
-        v.stride(1),
-        v.stride(0),
-        h_qo,
-        num_kv_groups,
-        HEAD_DIM=head_dim,
-        BLOCK_M=BLOCK_M,
-        BLOCK_N=BLOCK_N,
-        STAGE=stage,
-        num_warps=4 if head_dim == 64 else 8,
-        num_stages=4,
-    )
-    return dq, dk, dv
-
-
-# class TritonAttentionFunction(torch.autograd.Function):
-#     @staticmethod
-#     def forward(ctx, q, k, v, cu_seqlens_q, cu_seqlens_k, q_scale, k_scale, cu_seqlens_q_scale, cu_seqlens_k_scale):
-#         l = torch.zeros(q.shape[0], device=q.device, dtype=torch.float32)
-#         m = torch.zeros(q.shape[0], device=q.device, dtype=torch.float32)
-#         output = forward(q, k, v, cu_seqlens_q, cu_seqlens_k, q.shape[0], q_scale, k_scale, cu_seqlens_q_scale, cu_seqlens_k_scale, l, m)
-#         ctx.save_for_backward(q, k, v, cu_seqlens_q, cu_seqlens_k, q_scale, k_scale, cu_seqlens_q_scale, cu_seqlens_k_scale, l, m)
-#         return output
-#
-#     @staticmethod
-#     def backward(ctx, do):
-#         q, k, v, cu_seqlens_q, cu_seqlens_k, q_scale, k_scale, cu_seqlens_q_scale, cu_seqlens_k_scale, l, m = ctx.saved_tensors
-#         dq, dk, dv = backward(
-#             do, q, k, v,
-#             cu_seqlens_q, cu_seqlens_k,
-#             q.shape[0], q_scale, k_scale,
-#             cu_seqlens_q_scale, cu_seqlens_k_scale,
-#             l, m,
-#         )
-#         return dq, dk, dv, None, None, None, None, None, None

src/axolotl/integrations/sageattention/lib/triton/quant_per_block_varlen.py

@@ -1,158 +0,0 @@
-"""
-Copyright (c) 2024 by SageAttention team.
-
-Licensed under the Apache License, Version 2.0 (the "License");
-you may not use this file except in compliance with the License.
-You may obtain a copy of the License at
-
-    http://www.apache.org/licenses/LICENSE-2.0
-
-Unless required by applicable law or agreed to in writing, software
-distributed under the License is distributed on an "AS IS" BASIS,
-WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-See the License for the specific language governing permissions and
-limitations under the License.
-"""
-
-import torch
-import triton
-import triton.language as tl
-
-
-@triton.jit
-def quant_per_block_int8_kernel(
-    Input,
-    Output,
-    Scale,
-    cu_seqlens_input,
-    cu_seqlens_scale,
-    stride_ih,
-    stride_in,
-    stride_oh,
-    stride_on,
-    sm_scale,
-    H: tl.constexpr,
-    C: tl.constexpr,
-    BLK: tl.constexpr,
-):
-    off_blk = tl.program_id(0)
-    off_h = tl.program_id(1)
-    off_b = tl.program_id(2)
-
-    cu_seqlens_input_start = tl.load(cu_seqlens_input + off_b)
-    cu_seqlens_input_end = tl.load(cu_seqlens_input + off_b + 1)
-
-    L = cu_seqlens_input_end - cu_seqlens_input_start
-
-    if (off_blk * BLK) >= L:
-        return
-
-    cu_seqlens_scale_start = tl.load(cu_seqlens_scale + off_b)
-
-    offs_n = off_blk * BLK + tl.arange(0, BLK)
-    offs_k = tl.arange(0, C)
-
-    input_ptrs = (
-        Input
-        + cu_seqlens_input_start * stride_in
-        + off_h * stride_ih
-        + offs_n[:, None] * stride_in
-        + offs_k[None, :]
-    )
-    output_ptrs = (
-        Output
-        + cu_seqlens_input_start * stride_on
-        + off_h * stride_oh
-        + offs_n[:, None] * stride_on
-        + offs_k[None, :]
-    )
-    scale_ptrs = Scale + cu_seqlens_scale_start * H + off_h + off_blk * H
-
-    x = tl.load(input_ptrs, mask=offs_n[:, None] < L)
-    x = x.to(tl.float32)
-    x *= sm_scale
-    scale = tl.max(tl.abs(x)) / 127.0
-    x_int8 = x / scale
-    x_int8 += 0.5 * tl.where(x_int8 >= 0, 1, -1)
-    x_int8 = x_int8.to(tl.int8)
-    tl.store(output_ptrs, x_int8, mask=offs_n[:, None] < L)
-    tl.store(scale_ptrs, scale)
-
-
-def per_block_int8(
-    q,
-    k,
-    cu_seqlens_q,
-    cu_seqlens_k,
-    max_seqlen_q,
-    max_seqlen_k,
-    BLKQ=128,
-    BLKK=64,
-    sm_scale=None,
-):
-    q_int8 = torch.empty(q.shape, dtype=torch.int8, device=q.device)
-    k_int8 = torch.empty(k.shape, dtype=torch.int8, device=k.device)
-
-    h_qo = q.shape[1]
-    h_kv = k.shape[1]
-    head_dim = q.shape[-1]
-
-    b = cu_seqlens_q.shape[0] - 1
-    q_batch_len = cu_seqlens_q[1:] - cu_seqlens_q[:-1]
-    k_batch_len = cu_seqlens_k[1:] - cu_seqlens_k[:-1]
-
-    q_scale_len = (q_batch_len + BLKQ - 1) // BLKQ
-    k_scale_len = (k_batch_len + BLKK - 1) // BLKK
-
-    cu_seqlens_q_scale = torch.nn.functional.pad(
-        torch.cumsum(q_scale_len, dim=0), (1, 0), value=0
-    )
-    cu_seqlens_k_scale = torch.nn.functional.pad(
-        torch.cumsum(k_scale_len, dim=0), (1, 0), value=0
-    )
-
-    q_scale = torch.empty(
-        (cu_seqlens_q_scale[-1], h_qo), device=q.device, dtype=torch.float32
-    )
-    k_scale = torch.empty(
-        (cu_seqlens_k_scale[-1], h_kv), device=k.device, dtype=torch.float32
-    )
-
-    if sm_scale is None:
-        sm_scale = head_dim**-0.5
-
-    grid = ((max_seqlen_q + BLKQ - 1) // BLKQ, h_qo, b)
-    quant_per_block_int8_kernel[grid](
-        q,
-        q_int8,
-        q_scale,
-        cu_seqlens_q,
-        cu_seqlens_q_scale,
-        q.stride(1),
-        q.stride(0),
-        q_int8.stride(1),
-        q_int8.stride(0),
-        sm_scale=(sm_scale * 1.44269504),
-        H=h_qo,
-        C=head_dim,
-        BLK=BLKQ,
-    )
-
-    grid = ((max_seqlen_k + BLKK - 1) // BLKK, h_kv, b)
-    quant_per_block_int8_kernel[grid](
-        k,
-        k_int8,
-        k_scale,
-        cu_seqlens_k,
-        cu_seqlens_k_scale,
-        k.stride(1),
-        k.stride(0),
-        k_int8.stride(1),
-        k_int8.stride(0),
-        sm_scale=1.0,
-        H=h_kv,
-        C=head_dim,
-        BLK=BLKK,
-    )
-
-    return q_int8, q_scale, k_int8, k_scale, cu_seqlens_q_scale, cu_seqlens_k_scale

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/multipack.py

@@ -1,5 +1,4 @@
 """multipack patching for v2 of sample packing"""
-
 import importlib
 
 import transformers
@@ -28,28 +27,71 @@ SUPPORTED_MULTIPACK_MODEL_TYPES = [
 ]
 
 
-def patch_for_multipack(model_type, model_name=None, has_remote_code=False):
-    if has_remote_code:
-        patch_remote(model_name)
-    elif hasattr(transformers, "modeling_flash_attention_utils"):
+def patch_for_multipack(model_type, model_name=None, is_remote_code=False):
+    if model_type == "gemmoe":
+        patch_remote(model_name, ".configuration_gemmoe", ".modeling_gemmoe")
+    elif model_type == "deepseek_v2":
+        patch_remote(model_name, ".configuration_deepseek", ".modeling_deepseek")
+    elif hasattr(transformers, "modeling_flash_attention_utils") and not is_remote_code:
         transformers.modeling_flash_attention_utils._get_unpad_data = (  # pylint: disable=protected-access
             get_unpad_data
         )
+        if model_type == "mixtral" and is_deepspeed_zero3_enabled():
+            patch_mixtral_moe_forward_zero3()
+        return
 
-    if model_type == "mixtral" and is_deepspeed_zero3_enabled():
-        patch_mixtral_moe_forward_zero3()
+    # retain for legacy
+    if model_type == "mixtral":
+        transformers.models.mixtral.modeling_mixtral._get_unpad_data = (  # pylint: disable=protected-access
+            get_unpad_data
+        )
+        if is_deepspeed_zero3_enabled():
+            patch_mixtral_moe_forward_zero3()
+    elif model_type == "llama":
+        if hasattr(transformers.models.llama.modeling_llama, "_get_unpad_data"):
+            transformers.models.llama.modeling_llama._get_unpad_data = (  # pylint: disable=protected-access
+                get_unpad_data
+            )
+    elif model_type == "mistral":
+        if hasattr(transformers.models.mistral.modeling_mistral, "_get_unpad_data"):
+            transformers.models.llama.modeling_llama._get_unpad_data = (  # pylint: disable=protected-access
+                get_unpad_data
+            )
+    elif model_type == "qwen2":
+        transformers.models.qwen2.modeling_qwen2._get_unpad_data = (  # pylint: disable=protected-access
+            get_unpad_data
+        )
+    elif model_type == "qwen2_moe":
+        transformers.models.qwen2_moe.modeling_qwen2_moe._get_unpad_data = (  # pylint: disable=protected-access
+            get_unpad_data
+        )
+    elif model_type == "falcon":
+        transformers.models.falcon.modeling_falcon._get_unpad_data = (  # pylint: disable=protected-access
+            get_unpad_data
+        )
+    elif model_type == "phi":
+        transformers.models.phi.modeling_phi._get_unpad_data = (  # pylint: disable=protected-access
+            get_unpad_data
+        )
+    elif model_type == "gemma":
+        transformers.models.gemma.modeling_gemma._get_unpad_data = (  # pylint: disable=protected-access
+            get_unpad_data
+        )
+    elif model_type == "gemma2":
+        transformers.models.gemma2.modeling_gemma2._get_unpad_data = (  # pylint: disable=protected-access
+            get_unpad_data
+        )
+    elif model_type == "starcoder2":
+        transformers.models.starcoder2.modeling_starcoder2._get_unpad_data = (  # pylint: disable=protected-access
+            get_unpad_data
+        )
 
 
-def patch_remote(model_name):
+def patch_remote(model_name, config_name, modeling_name):
     model_config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
     # we need to load the model here in order for modeling_* to be available
     with init_empty_weights():
         AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)
-    parts = model_config.__class__.__module__.split(".")
-    parts[-1] = parts[-1].replace("configuration_", "modeling_", 1)
-    module_name = ".".join(parts)
+    module_name = model_config.__class__.__module__.replace(config_name, modeling_name)
     modeling_arch = importlib.import_module(module_name)
-    if hasattr(modeling_arch, "_get_unpad_data"):
-        modeling_arch._get_unpad_data = (  # pylint: disable=protected-access
-            get_unpad_data
-        )
+    modeling_arch._get_unpad_data = get_unpad_data  # pylint: disable=protected-access

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")
 
@@ -188,7 +231,7 @@ def integrate_lora_mlp_patch(peft_model: PeftModelForCausalLM):
             for module in layer_modules
         )
         mlp_not_dora = all(
-            len(getattr(module, "lora_magnitude_vector", []) or []) == 0
+            getattr(module, "lora_magnitude_vector", None) is None
             for module in layer_modules
         )
 
@@ -213,7 +256,7 @@ def integrate_lora_patch(peft_model: PeftModelForCausalLM, cfg):
                 for module in layer_modules
             )
             qkv_not_dora = all(
-                len(getattr(module, "lora_magnitude_vector", []) or []) == 0
+                getattr(module, "lora_magnitude_vector", None) is None
                 for module in layer_modules
             )
 
@@ -232,7 +275,7 @@ def integrate_lora_patch(peft_model: PeftModelForCausalLM, cfg):
                 for module in layer_modules
             )
             o_not_dora = all(
-                len(getattr(module, "lora_magnitude_vector", []) or []) == 0
+                getattr(module, "lora_magnitude_vector", None) is None
                 for module in layer_modules
             )
 

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/__init__.py

@@ -11,10 +11,6 @@ LOG = logging.getLogger("axolotl.prompt_strategies")
 
 def load(strategy, tokenizer, cfg, ds_cfg, processor=None):
     try:
-        if strategy == "messages":
-            from .messages import load as messages_load
-
-            return messages_load(tokenizer, cfg, ds_cfg, processor=processor)
         load_fn = "load"
         if strategy.split(".")[-1].startswith("load_"):
             load_fn = strategy.split(".")[-1]
@@ -35,5 +31,4 @@ def load(strategy, tokenizer, cfg, ds_cfg, processor=None):
         return None
     except Exception as exc:  # pylint: disable=broad-exception-caught
         LOG.error(f"Failed to load prompt strategy `{strategy}`: {str(exc)}")
-        raise exc
-    return None
+        return None

src/axolotl/prompt_strategies/bradley_terry/README.md

@@ -1,10 +0,0 @@
-### example yaml
-
-```yaml
-chat_template: gemma
-datasets:
-  - path: argilla/distilabel-intel-orca-dpo-pairs
-    type: bradley_terry.chat_template
-val_set_size: 0.0
-output_dir: ./outputs/out
-```

src/axolotl/prompt_strategies/bradley_terry/__init__.py

@@ -1,35 +0,0 @@
-"""Module to load prompt strategies."""
-
-import importlib
-import inspect
-import logging
-
-from axolotl.prompt_strategies.user_defined import UserDefinedDatasetConfig
-
-LOG = logging.getLogger("axolotl.prompt_strategies.bradley_terry")
-
-
-def load(strategy, tokenizer, cfg, ds_cfg):
-    # pylint: disable=duplicate-code
-    try:
-        load_fn = "load"
-        if strategy.split(".")[-1].startswith("load_"):
-            load_fn = strategy.split(".")[-1]
-            strategy = ".".join(strategy.split(".")[:-1])
-        mod = importlib.import_module(
-            f".{strategy}", "axolotl.prompt_strategies.bradley_terry"
-        )
-        func = getattr(mod, load_fn)
-        load_kwargs = {}
-        if strategy == "user_defined":
-            load_kwargs["ds_cfg"] = UserDefinedDatasetConfig(**ds_cfg)
-        else:
-            sig = inspect.signature(func)
-            if "ds_cfg" in sig.parameters:
-                load_kwargs["ds_cfg"] = ds_cfg
-        return func(tokenizer, cfg, **load_kwargs)
-    except ModuleNotFoundError:
-        return None
-    except Exception as exc:  # pylint: disable=broad-exception-caught
-        LOG.error(f"Failed to load prompt strategy `{strategy}`: {str(exc)}")
-        return None

src/axolotl/prompt_strategies/bradley_terry/chat_template.py

@@ -1,102 +0,0 @@
-"""
-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)
-
-
-class BTChatTemplateStrategy(ChatTemplateStrategy):
-    """
-    Bradley-Terry reward model pairwise chat template prompt strategy.
-    """
-
-    def tokenize_prompt(self, prompt):
-        """
-
-        :param prompt: the actual row of data from the underlying dataset
-        :return:
-        """
-
-        self.messages = "chosen_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["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"]}
-        )
-        rejected_tokenized = super().tokenize_prompt(prompt)
-
-        return {
-            "input_ids_chosen": chosen_tokenized["input_ids"],
-            "attention_mask_chosen": chosen_tokenized["attention_mask"],
-            "labels_chosen": 1.0,
-            "input_ids_rejected": rejected_tokenized["input_ids"],
-            "attention_mask_rejected": rejected_tokenized["attention_mask"],
-            "labels_rejected": 0.0,
-        }
-
-
-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),
-        "message_field_training_detail": ds_cfg.get(
-            "message_field_training_detail", None
-        ),
-        "roles": ds_cfg.get("roles"),
-        "drop_system_message": ds_cfg.get("drop_system_message", False),
-        # we need to add one for detecting sequences with exceeding the `sequence_len` limit.
-        "max_length": cfg.sequence_len + 1
-        if not cfg.reward_model
-        else cfg.sequence_len,
-    }
-
-    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),
-    }
-
-    strategy = BTChatTemplateStrategy(
-        ChatTemplatePrompter(**prompter_params), tokenizer=tokenizer, **strategy_params
-    )
-
-    if "field_messages" in ds_cfg and hasattr(strategy, "messages"):
-        strategy.messages = ds_cfg["field_messages"]
-
-    return strategy

src/axolotl/prompt_strategies/bradley_terry/llama3.py

@@ -1,27 +0,0 @@
-"""
-chatml transforms for datasets with system, input, chosen, rejected to match llama3 chat template
-"""
-
-
-def icr(
-    cfg,
-    **kwargs,
-):  # pylint: disable=possibly-unused-variable,unused-argument
-    """
-    chatml transforms for datasets with system, input, chosen, rejected
-    ex. https://huggingface.co/datasets/argilla/distilabel-intel-orca-dpo-pairs
-    """
-
-    def transform_fn(sample):
-        if "system" in sample and sample["system"]:
-            prompt = (
-                f"<|start_header_id|>system<|end_header_id|>\n\n{sample['system']}<|eot_id|>"
-                f"<|start_header_id|>user<|end_header_id|>\n\n{sample['input']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
-            )
-        else:
-            prompt = f"<|start_header_id|>user<|end_header_id|>\n\n{sample['input']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
-        sample["chosen"] = prompt + f"{sample['chosen']}<|eot_id|>"
-        sample["rejected"] = prompt + f"{sample['rejected']}<|eot_id|>"
-        return sample
-
-    return transform_fn

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")
@@ -403,16 +403,11 @@ 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"]},
+        ]