proof of concept for sage attention

* updated colab notebook

* update pip installtation

* cleared cell output

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

Co-authored-by: NanoCode012 <nano@axolotl.ai>

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

Co-authored-by: NanoCode012 <nano@axolotl.ai>

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

Co-authored-by: NanoCode012 <nano@axolotl.ai>

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

Co-authored-by: NanoCode012 <nano@axolotl.ai>

* modified notebook

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

Co-authored-by: NanoCode012 <nano@axolotl.ai>

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

Co-authored-by: NanoCode012 <nano@axolotl.ai>

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

Co-authored-by: NanoCode012 <nano@axolotl.ai>

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

Co-authored-by: NanoCode012 <nano@axolotl.ai>

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

Co-authored-by: NanoCode012 <nano@axolotl.ai>

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

Co-authored-by: NanoCode012 <nano@axolotl.ai>

* cleared cell output

* cleared unnecessary logs

---------

Co-authored-by: NanoCode012 <nano@axolotl.ai>

.github/workflows/base.yml

@@ -27,7 +27,7 @@ jobs:
           - cuda: "124"
             cuda_version: 12.4.1
             cudnn_version: ""
-            python_version: "3.11"
+            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"
@@ -40,23 +40,25 @@ jobs:
             cuda_version: 12.4.1
             cudnn_version: ""
             python_version: "3.11"
-            pytorch: 2.5.0
+            pytorch: 2.5.1
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
       - name: Docker metadata
         id: metadata
-        uses: docker/metadata-action@v3
+        uses: docker/metadata-action@v5
         with:
-          images: winglian/axolotl-base
+          images: |
+            winglian/axolotl-base
+            axolotlai/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@v2
+        uses: docker/setup-buildx-action@v3
       - 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@v3
+          uses: actions/setup-python@v5
           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@v3
-      - uses: actions/setup-python@v4
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
         with:
           python-version: "3.10"
           cache: 'pip' # caching pip dependencies
-      - uses: pre-commit/action@v3.0.0
+      - uses: pre-commit/action@v3.0.1

.github/workflows/main.yml

@@ -4,11 +4,13 @@ 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: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
     strategy:
       fail-fast: false
       matrix:
@@ -32,7 +34,7 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.5.0
+            pytorch: 2.5.1
             axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
@@ -42,7 +44,12 @@ jobs:
         id: metadata
         uses: docker/metadata-action@v5
         with:
-          images: winglian/axolotl
+          images: |
+            winglian/axolotl
+            axolotlai/axolotl
+          tags: |
+            type=ref,event=branch
+            type=pep440,pattern={{version}}
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
       - name: Login to Docker Hub
@@ -56,7 +63,7 @@ jobs:
         with:
           context: .
           build-args: |
-            BASE_TAG=${{ github.ref_name }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}
+            BASE_TAG=${{ github.ref_type == 'tag' && 'main' || 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 }}
@@ -70,7 +77,7 @@ 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:
@@ -94,7 +101,7 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.5.0
+            pytorch: 2.5.1
             axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
@@ -104,20 +111,25 @@ jobs:
         id: metadata
         uses: docker/metadata-action@v5
         with:
-          images: winglian/axolotl-cloud
+          images: |
+            winglian/axolotl-cloud
+            axolotlai/axolotl-cloud
+          tags: |
+            type=ref,event=branch
+            type=pep440,pattern={{version}}
       - name: Login to Docker Hub
         uses: docker/login-action@v3
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v2
+        uses: docker/setup-buildx-action@v3
       - name: Build
         uses: docker/build-push-action@v5
         with:
           context: .
           build-args: |
-            BASE_TAG=${{ github.ref_name }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
+            BASE_TAG=${{ github.ref_type == 'tag' && 'main' || github.ref_name }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
             CUDA=${{ matrix.cuda }}
           file: ./docker/Dockerfile-cloud
           push: ${{ github.event_name != 'pull_request' }}
@@ -128,7 +140,7 @@ 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:
@@ -146,20 +158,25 @@ jobs:
         id: metadata
         uses: docker/metadata-action@v5
         with:
-          images: winglian/axolotl-cloud-term
+          images: |
+            winglian/axolotl-cloud-term
+            axolotlai/axolotl-cloud-term
+          tags: |
+            type=ref,event=branch
+            type=pep440,pattern={{version}}
       - name: Login to Docker Hub
         uses: docker/login-action@v3
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v2
+        uses: docker/setup-buildx-action@v3
       - name: Build
         uses: docker/build-push-action@v5
         with:
           context: .
           build-args: |
-            BASE_TAG=${{ github.ref_name }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
+            BASE_TAG=${{ github.ref_type == 'tag' && 'main' || github.ref_name }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
             CUDA=${{ matrix.cuda }}
           file: ./docker/Dockerfile-cloud-no-tmux
           push: ${{ github.event_name != 'pull_request' }}

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

@@ -8,9 +8,14 @@ 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 docker]]') && github.repository_owner == 'axolotl-ai-cloud' }}
+    if: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' }}
     strategy:
       fail-fast: false
       matrix:
@@ -31,7 +36,7 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.5.0
+            pytorch: 2.5.1
             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:
@@ -31,7 +31,7 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.5.0
+            pytorch: 2.5.1
             axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
@@ -41,7 +41,9 @@ jobs:
         id: metadata
         uses: docker/metadata-action@v5
         with:
-          images: winglian/axolotl
+          images: |
+            winglian/axolotl
+            axolotlai/axolotl
           tags: |
             type=raw,value={{ branch }}-{{ date 'YYYYMMDD' }}
       - name: Set up Docker Buildx
@@ -69,7 +71,7 @@ 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:
@@ -93,7 +95,7 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.5.0
+            pytorch: 2.5.1
             axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
@@ -103,7 +105,9 @@ jobs:
         id: metadata
         uses: docker/metadata-action@v5
         with:
-          images: winglian/axolotl-cloud
+          images: |
+            winglian/axolotl-cloud
+            axolotlai/axolotl-cloud
           tags: |
             type=raw,value={{ branch }}-{{ date 'YYYYMMDD' }}
       - name: Login to Docker Hub
@@ -112,7 +116,7 @@ jobs:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v2
+        uses: docker/setup-buildx-action@v3
       - name: Build
         uses: docker/build-push-action@v5
         with:

.github/workflows/pypi.yml

@@ -3,12 +3,24 @@ 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
@@ -16,10 +28,10 @@ jobs:
       id-token: write  # IMPORTANT: this permission is mandatory for trusted publishing
     steps:
       - name: Check out repository code
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Setup Python
-        uses: actions/setup-python@v4
+        uses: actions/setup-python@v5
         with:
           python-version: "3.10"
 
@@ -37,9 +49,9 @@ jobs:
         run: |
           sed -i -E 's/version="([0-9.]+)",/version="${{ steps.tag.outputs.TAG_NAME }}",/g' setup.py
 
-      - name: Build a binary wheel
+      - name: Build a source dist
         run: |
-          python setup.py sdist bdist_wheel
+          python setup.py sdist
 
       - 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@v3
-      - uses: actions/setup-python@v4
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
         with:
           python-version: "3.10"
           cache: 'pip' # caching pip dependencies
-      - uses: pre-commit/action@v3.0.0
+      - uses: pre-commit/action@v3.0.1
         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", "2.5.0"]
+        pytorch_version: ["2.3.1", "2.4.1", "2.5.1"]
     timeout-minutes: 20
 
     steps:
       - name: Check out repository code
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Setup Python
-        uses: actions/setup-python@v4
+        uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.python_version }}
           cache: 'pip' # caching pip dependencies
@@ -48,6 +48,7 @@ jobs:
           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: |
@@ -82,13 +83,6 @@ jobs:
             num_gpus: 1
             axolotl_extras: mamba-ssm
             nightly_build: "true"
-          - cuda: 121
-            cuda_version: 12.1.1
-            python_version: "3.11"
-            pytorch: 2.3.1
-            num_gpus: 1
-            axolotl_extras: mamba-ssm
-            nightly_build: "true"
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
@@ -99,7 +93,7 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.5.0
+            pytorch: 2.5.1
             num_gpus: 1
             axolotl_extras:
             nightly_build: "true"

.github/workflows/tests.yml

@@ -15,17 +15,22 @@ 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@v3
-      - uses: actions/setup-python@v4
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
         with:
           python-version: "3.10"
           cache: 'pip' # caching pip dependencies
-      - uses: pre-commit/action@v3.0.0
+      - uses: pre-commit/action@v3.0.1
         env:
           SKIP: no-commit-to-branch
 
@@ -36,15 +41,15 @@ jobs:
       fail-fast: false
       matrix:
         python_version: ["3.10", "3.11"]
-        pytorch_version: ["2.3.1", "2.4.1", "2.5.0"]
+        pytorch_version: ["2.3.1", "2.4.1", "2.5.1"]
     timeout-minutes: 20
 
     steps:
       - name: Check out repository code
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Setup Python
-        uses: actions/setup-python@v4
+        uses: actions/setup-python@v5
         with:
           python-version: ${{ matrix.python_version }}
           cache: 'pip' # caching pip dependencies
@@ -66,47 +71,119 @@ jobs:
 
       - name: Run tests
         run: |
-          pytest --ignore=tests/e2e/ tests/
+          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:
-    if: github.repository_owner == 'axolotl-ai-cloud'
+  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' }}
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
     timeout-minutes: 90
-    needs: [pre-commit, pytest]
+    needs: [pre-commit, pytest, pytest-sdist]
 
     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"
             pytorch: 2.4.1
             num_gpus: 1
             axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.5.0
-            num_gpus: 1
-            axolotl_extras:
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Install Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.10"
+      - 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
+
+  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

.gitignore

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

1991.yml

@@ -1,295 +0,0 @@
-base_model: Qwen/Qwen2.5-14B-Instruct
-model_type: AutoModelForCausalLM #nohup accelerate launch -m axolotl.cli.train /home/ubuntu/qwen2.5_14B.yml > training_output.log 2>&1 &
-tokenizer_type: AutoTokenizer
-trust_remote_code: true
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-datasets:
-  - path: tatsu-lab/alpaca
-    type: alpaca
-
-chat_template: chatml
-dataset_prepared_path:
-val_set_size: 0
-output_dir: ./outputs/out
-
-sequence_len: 2048
-sample_packing: true
-eval_sample_packing: true
-pad_to_sequence_len: true
-
-unfrozen_parameters:
-- ^lm_head.weight$
-- ^model.embed_tokens.weight$
-# input_layernorm layers
-- model.layers.0.input_layernorm
-- model.layers.1.input_layernorm
-- model.layers.2.input_layernorm
-- model.layers.3.input_layernorm
-- model.layers.4.input_layernorm
-- model.layers.5.input_layernorm
-- model.layers.6.input_layernorm
-- model.layers.7.input_layernorm
-- model.layers.8.input_layernorm
-- model.layers.9.input_layernorm
-- model.layers.10.input_layernorm
-- model.layers.11.input_layernorm
-- model.layers.12.input_layernorm
-- model.layers.13.input_layernorm
-- model.layers.14.input_layernorm
-- model.layers.15.input_layernorm
-- model.layers.16.input_layernorm
-- model.layers.17.input_layernorm
-- model.layers.18.input_layernorm
-- model.layers.19.input_layernorm
-- model.layers.20.input_layernorm
-- model.layers.21.input_layernorm
-- model.layers.22.input_layernorm
-- model.layers.23.input_layernorm
-# lm_head layers
-# mlp.down_proj layers
-- model.layers.1.mlp.down_proj
-- model.layers.35.mlp.down_proj
-- model.layers.38.mlp.down_proj
-- model.layers.37.mlp.down_proj
-- model.layers.36.mlp.down_proj
-- model.layers.15.mlp.down_proj
-- model.layers.11.mlp.down_proj
-- model.layers.12.mlp.down_proj
-- model.layers.34.mlp.down_proj
-- model.layers.44.mlp.down_proj
-- model.layers.45.mlp.down_proj
-- model.layers.9.mlp.down_proj
-- model.layers.41.mlp.down_proj
-- model.layers.33.mlp.down_proj
-- model.layers.43.mlp.down_proj
-- model.layers.40.mlp.down_proj
-- model.layers.13.mlp.down_proj
-- model.layers.8.mlp.down_proj
-- model.layers.39.mlp.down_proj
-- model.layers.10.mlp.down_proj
-- model.layers.14.mlp.down_proj
-- model.layers.16.mlp.down_proj
-- model.layers.31.mlp.down_proj
-- model.layers.32.mlp.down_proj
-# mlp.gate_proj layers
-- model.layers.1.mlp.gate_proj
-- model.layers.44.mlp.gate_proj
-- model.layers.46.mlp.gate_proj
-- model.layers.45.mlp.gate_proj
-- model.layers.43.mlp.gate_proj
-- model.layers.47.mlp.gate_proj
-- model.layers.42.mlp.gate_proj
-- model.layers.32.mlp.gate_proj
-- model.layers.27.mlp.gate_proj
-- model.layers.33.mlp.gate_proj
-- model.layers.28.mlp.gate_proj
-- model.layers.39.mlp.gate_proj
-- model.layers.41.mlp.gate_proj
-- model.layers.40.mlp.gate_proj
-- model.layers.30.mlp.gate_proj
-- model.layers.29.mlp.gate_proj
-- model.layers.31.mlp.gate_proj
-- model.layers.26.mlp.gate_proj
-- model.layers.37.mlp.gate_proj
-- model.layers.10.mlp.gate_proj
-- model.layers.38.mlp.gate_proj
-- model.layers.12.mlp.gate_proj
-- model.layers.36.mlp.gate_proj
-- model.layers.13.mlp.gate_proj
-# mlp.up_proj layers
-- model.layers.1.mlp.up_proj
-- model.layers.13.mlp.up_proj
-- model.layers.11.mlp.up_proj
-- model.layers.14.mlp.up_proj
-- model.layers.15.mlp.up_proj
-- model.layers.12.mlp.up_proj
-- model.layers.8.mlp.up_proj
-- model.layers.16.mlp.up_proj
-- model.layers.9.mlp.up_proj
-- model.layers.19.mlp.up_proj
-- model.layers.10.mlp.up_proj
-- model.layers.7.mlp.up_proj
-- model.layers.17.mlp.up_proj
-- model.layers.20.mlp.up_proj
-- model.layers.21.mlp.up_proj
-- model.layers.18.mlp.up_proj
-- model.layers.38.mlp.up_proj
-- model.layers.37.mlp.up_proj
-- model.layers.39.mlp.up_proj
-- model.layers.42.mlp.up_proj
-- model.layers.41.mlp.up_proj
-- model.layers.27.mlp.up_proj
-- model.layers.28.mlp.up_proj
-- model.layers.34.mlp.up_proj
-# model.norm layers
-# post_attention_layernorm layers
-- model.layers.0.post_attention_layernorm
-- model.layers.1.post_attention_layernorm
-- model.layers.2.post_attention_layernorm
-- model.layers.3.post_attention_layernorm
-- model.layers.4.post_attention_layernorm
-- model.layers.5.post_attention_layernorm
-- model.layers.6.post_attention_layernorm
-- model.layers.7.post_attention_layernorm
-- model.layers.8.post_attention_layernorm
-- model.layers.9.post_attention_layernorm
-- model.layers.10.post_attention_layernorm
-- model.layers.11.post_attention_layernorm
-- model.layers.12.post_attention_layernorm
-- model.layers.13.post_attention_layernorm
-- model.layers.14.post_attention_layernorm
-- model.layers.15.post_attention_layernorm
-- model.layers.16.post_attention_layernorm
-- model.layers.17.post_attention_layernorm
-- model.layers.18.post_attention_layernorm
-- model.layers.19.post_attention_layernorm
-- model.layers.20.post_attention_layernorm
-- model.layers.21.post_attention_layernorm
-- model.layers.22.post_attention_layernorm
-- model.layers.23.post_attention_layernorm
-# self_attn.k_proj layers
-- model.layers.47.self_attn.k_proj
-- model.layers.39.self_attn.k_proj
-- model.layers.41.self_attn.k_proj
-- model.layers.37.self_attn.k_proj
-- model.layers.35.self_attn.k_proj
-- model.layers.44.self_attn.k_proj
-- model.layers.38.self_attn.k_proj
-- model.layers.14.self_attn.k_proj
-- model.layers.7.self_attn.k_proj
-- model.layers.12.self_attn.k_proj
-- model.layers.11.self_attn.k_proj
-- model.layers.32.self_attn.k_proj
-- model.layers.10.self_attn.k_proj
-- model.layers.8.self_attn.k_proj
-- model.layers.9.self_attn.k_proj
-- model.layers.6.self_attn.k_proj
-- model.layers.45.self_attn.k_proj
-- model.layers.42.self_attn.k_proj
-- model.layers.5.self_attn.k_proj
-- model.layers.40.self_attn.k_proj
-- model.layers.33.self_attn.k_proj
-- model.layers.0.self_attn.k_proj
-- model.layers.34.self_attn.k_proj
-- model.layers.13.self_attn.k_proj
-# self_attn.o_proj layers
-- model.layers.12.self_attn.o_proj
-- model.layers.5.self_attn.o_proj
-- model.layers.14.self_attn.o_proj
-- model.layers.16.self_attn.o_proj
-- model.layers.20.self_attn.o_proj
-- model.layers.13.self_attn.o_proj
-- model.layers.11.self_attn.o_proj
-- model.layers.4.self_attn.o_proj
-- model.layers.6.self_attn.o_proj
-- model.layers.19.self_attn.o_proj
-- model.layers.7.self_attn.o_proj
-- model.layers.18.self_attn.o_proj
-- model.layers.8.self_attn.o_proj
-- model.layers.38.self_attn.o_proj
-- model.layers.15.self_attn.o_proj
-- model.layers.17.self_attn.o_proj
-- model.layers.9.self_attn.o_proj
-- model.layers.10.self_attn.o_proj
-- model.layers.21.self_attn.o_proj
-- model.layers.28.self_attn.o_proj
-- model.layers.32.self_attn.o_proj
-- model.layers.35.self_attn.o_proj
-- model.layers.39.self_attn.o_proj
-- model.layers.3.self_attn.o_proj
-# self_attn.q_proj layers
-- model.layers.1.self_attn.q_proj
-- model.layers.2.self_attn.q_proj
-- model.layers.3.self_attn.q_proj
-- model.layers.44.self_attn.q_proj
-- model.layers.29.self_attn.q_proj
-- model.layers.45.self_attn.q_proj
-- model.layers.43.self_attn.q_proj
-- model.layers.32.self_attn.q_proj
-- model.layers.38.self_attn.q_proj
-- model.layers.19.self_attn.q_proj
-- model.layers.42.self_attn.q_proj
-- model.layers.34.self_attn.q_proj
-- model.layers.36.self_attn.q_proj
-- model.layers.40.self_attn.q_proj
-- model.layers.26.self_attn.q_proj
-- model.layers.20.self_attn.q_proj
-- model.layers.39.self_attn.q_proj
-- model.layers.28.self_attn.q_proj
-- model.layers.35.self_attn.q_proj
-- model.layers.41.self_attn.q_proj
-- model.layers.33.self_attn.q_proj
-- model.layers.25.self_attn.q_proj
-- model.layers.30.self_attn.q_proj
-- model.layers.27.self_attn.q_proj
-# self_attn.v_proj layers
-- model.layers.0.self_attn.v_proj
-- model.layers.7.self_attn.v_proj
-- model.layers.39.self_attn.v_proj
-- model.layers.31.self_attn.v_proj
-- model.layers.15.self_attn.v_proj
-- model.layers.10.self_attn.v_proj
-- model.layers.32.self_attn.v_proj
-- model.layers.41.self_attn.v_proj
-- model.layers.6.self_attn.v_proj
-- model.layers.33.self_attn.v_proj
-- model.layers.42.self_attn.v_proj
-- model.layers.29.self_attn.v_proj
-- model.layers.14.self_attn.v_proj
-- model.layers.9.self_attn.v_proj
-- model.layers.35.self_attn.v_proj
-- model.layers.38.self_attn.v_proj
-- model.layers.13.self_attn.v_proj
-- model.layers.30.self_attn.v_proj
-- model.layers.5.self_attn.v_proj
-- model.layers.34.self_attn.v_proj
-- model.layers.28.self_attn.v_proj
-- model.layers.37.self_attn.v_proj
-- model.layers.27.self_attn.v_proj
-- model.layers.11.self_attn.v_proj
-# model.embed_tokens layers
-
-
-gradient_accumulation_steps: 2
-micro_batch_size: 2
-num_epochs: 3
-optimizer: adamw_torch_fused
-lr_scheduler: linear
-learning_rate: 5e-6
-
-train_on_inputs: false
-group_by_length: false
-bf16: auto
-fp16:
-tf32: false
-
-plugins:
-  - axolotl.integrations.liger.LigerPlugin
-liger_rope: true
-liger_rms_norm: true
-liger_swiglu: true
-liger_fused_linear_cross_entropy: true
-
-gradient_checkpointing: unsloth
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-warmup_steps: 10
-evals_per_epoch: 2
-saves_per_epoch: 1
-save_total_limit: 4
-debug:
-deepspeed: deepspeed_configs/zero3_bf16.json
-weight_decay: 0.05
-special_tokens:
-  eos_token: <|im_end|>

MANIFEST.in

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

README.md

@@ -1,8 +1,21 @@
-# Axolotl
+<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>
 
-![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)
+<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>
 
 Axolotl is a tool designed to streamline the fine-tuning of various AI models, offering support for multiple configurations and architectures.
 
@@ -75,7 +88,7 @@ Features:
 <td>
 
 <div align="center">
-  <img src="image/axolotl.png" alt="axolotl" width="160">
+  <img src="image/axolotl_symbol_digital_white.svg" alt="axolotl" width="160">
   <div>
     <p>
       <b>Axolotl provides a unified repository for fine-tuning <br />a variety of AI models with ease</b>
@@ -159,7 +172,7 @@ accelerate launch -m axolotl.cli.train https://raw.githubusercontent.com/axolotl
 #### Docker
 
   ```bash
-  docker run --gpus '"all"' --rm -it winglian/axolotl:main-latest
+  docker run --gpus '"all"' --rm -it axolotlai/axolotl:main-latest
   ```
 
   Or run on the current files for development:
@@ -178,7 +191,7 @@ 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 ${HOME}/.cache/huggingface:/root/.cache/huggingface winglian/axolotl:main-latest
+docker run --privileged --gpus '"all"' --shm-size 10g --rm -it --name axolotl --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --mount type=bind,src="${PWD}",target=/workspace/axolotl -v ${HOME}/.cache/huggingface:/root/.cache/huggingface axolotlai/axolotl:main-latest
   ```
 
   It additionally:
@@ -210,7 +223,7 @@ docker run --privileged --gpus '"all"' --shm-size 10g --rm -it --name axolotl --
 
 #### Cloud GPU
 
-For cloud GPU providers that support docker images, use [`winglian/axolotl-cloud:main-latest`](https://hub.docker.com/r/winglian/axolotl-cloud/tags)
+For cloud GPU providers that support docker images, use [`axolotlai/axolotl-cloud:main-latest`](https://hub.docker.com/r/axolotlai/axolotl-cloud/tags)
 
 - on Latitude.sh use this [direct link](https://latitude.sh/blueprint/989e0e79-3bf6-41ea-a46b-1f246e309d5c)
 - on JarvisLabs.ai use this [direct link](https://jarvislabs.ai/templates/axolotl)
@@ -319,7 +332,7 @@ Write a job description in YAML as below:
 # dstack.yaml
 type: task
 
-image: winglian/axolotl-cloud:main-20240429-py3.11-cu121-2.2.2
+image: axolotlai/axolotl-cloud:main-latest
 
 env:
   - HUGGING_FACE_HUB_TOKEN
@@ -383,11 +396,10 @@ See [examples](examples) for quick start. It is recommended to duplicate and mod
         - typescript
       type: ... # unimplemented custom format
 
-      # fastchat conversation (deprecation soon, use chat_template https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/conversation.html#chat_template)
-      # See 'conversation' options: https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
+      # chat_template https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/conversation.html#chat_template
     - path: ...
-      type: sharegpt
-      conversation: chatml # default: vicuna_v1.1
+      type: chat_template
+      chat_template: chatml # defaults to tokenizer's chat_template
 
       # local
     - path: data.jsonl # or json
@@ -562,7 +574,8 @@ plugins:
   - axolotl.integrations.liger.LigerPlugin
 liger_rope: true
 liger_rms_norm: true
-liger_swiglu: true
+liger_glu_activation: true
+liger_layer_norm: true
 liger_fused_linear_cross_entropy: true
 ```
 

cicd/Dockerfile.jinja

@@ -1,4 +1,4 @@
-FROM winglian/axolotl-base:{{ BASE_TAG }}
+FROM axolotlai/axolotl-base:{{ BASE_TAG }}
 
 ENV TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX"
 ENV AXOLOTL_EXTRAS="{{ AXOLOTL_EXTRAS }}"
@@ -28,6 +28,7 @@ RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
         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 \

cicd/cicd.sh

@@ -1,6 +1,6 @@
 #!/bin/bash
 set -e
 
-pytest -n4 --ignore=tests/e2e/ /workspace/axolotl/tests/
+pytest -n8 --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 Image, Stub
+from modal import App, Image
 
 cicd_path = pathlib.Path(__file__).parent.resolve()
 
@@ -46,7 +46,7 @@ cicd_image = (
     .pip_install("fastapi==0.110.0", "pydantic==2.6.3")
 )
 
-stub = Stub("Axolotl CI/CD", secrets=[])
+app = App("Axolotl CI/CD", secrets=[])
 
 
 N_GPUS = int(os.environ.get("N_GPUS", 2))
@@ -61,7 +61,7 @@ def run_cmd(cmd: str, run_folder: str):
         exit(exit_code)  # pylint: disable=consider-using-sys-exit
 
 
-@stub.function(
+@app.function(
     image=cicd_image,
     gpu=GPU_CONFIG,
     timeout=60 * 60,
@@ -72,6 +72,6 @@ def cicd_pytest():
     run_cmd("./cicd/multigpu.sh", "/workspace/axolotl")
 
 
-@stub.local_entrypoint()
+@app.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 -n1 /workspace/axolotl/tests/e2e/multigpu/
+pytest -v -n2 /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 Image, Stub
+from modal import App, Image
 
 cicd_path = pathlib.Path(__file__).parent.resolve()
 
@@ -47,7 +47,7 @@ cicd_image = (
     .pip_install("fastapi==0.110.0", "pydantic==2.6.3")
 )
 
-stub = Stub("Axolotl CI/CD", secrets=[])
+app = App("Axolotl CI/CD", secrets=[])
 
 
 N_GPUS = int(os.environ.get("N_GPUS", 1))
@@ -62,7 +62,7 @@ def run_cmd(cmd: str, run_folder: str):
         exit(exit_code)  # pylint: disable=consider-using-sys-exit
 
 
-@stub.function(
+@app.function(
     image=cicd_image,
     gpu=GPU_CONFIG,
     timeout=60 * 60,
@@ -73,6 +73,6 @@ def cicd_pytest():
     run_cmd("./cicd/cicd.sh", "/workspace/axolotl")
 
 
-@stub.local_entrypoint()
+@app.local_entrypoint()
 def main():
     cicd_pytest.remote()

devtools/dev_chat_template.yml

@@ -1,4 +1,4 @@
-# Example config for debugging the sharegpt prompt format
+# Example config for debugging the chat_template 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: philschmid/guanaco-sharegpt-style
-    type: sharegpt
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
     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 winglian/axolotl-base:$BASE_TAG
+FROM axolotlai/axolotl-base:$BASE_TAG
 
 ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX"
 ARG AXOLOTL_EXTRAS=""

docker/Dockerfile-cloud

@@ -1,5 +1,5 @@
 ARG BASE_TAG=main
-FROM winglian/axolotl:$BASE_TAG
+FROM axolotlai/axolotl:$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 winglian/axolotl:$BASE_TAG
+FROM axolotlai/axolotl:$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 winglian/axolotl-base:$BASE_TAG
+FROM axolotlai/axolotl-base:$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, sharegpt, gpteacher, oasst, reflection]
+    # The type of prompt to use for training. [alpaca, gpteacher, oasst, reflection]
     type: alpaca # format | format:<prompt_style> (chat/instruct) | <prompt_strategies>.load_<load_fn>
     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
@@ -91,15 +91,7 @@ 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.
-
-    # 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]].
+    trust_remote_code: # Optional[bool] Trust remote code for untrusted source
 
   # Custom user instruction prompt
   - path: repo
@@ -183,6 +175,8 @@ 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.
@@ -412,6 +406,7 @@ 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

@@ -6,33 +6,8 @@ order: 3
 
 ## sharegpt
 
-UPDATE: ShareGPT is being deprecated in the next release. Please see `chat_template` section below.
+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
 
@@ -40,38 +15,6 @@ datasets:
 {"conversations": [{"role": "...", "value": "..."}]}
 ```
 
-## sharegpt.load_role
-
-conversations where `role` is used instead of `from`
-
-```{.json filename="data.jsonl"}
-{"conversations": [{"role": "...", "value": "..."}]}
-```
-
-## sharegpt.load_guanaco
-
-conversations where `from` is `prompter` `assistant` instead of default sharegpt
-
-```{.json filename="data.jsonl"}
-{"conversations": [{"from": "...", "value": "..."}]}
-```
-
-## sharegpt.load_ultrachat
-
-conversations where the turns field is 'messages', human is 'user' and gpt is 'assistant'.
-
-```{.json filename="data.jsonl"}
-{"messages": [{"user": "...", "assistant": "..."}]}
-```
-
-## 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": "..."}]}
-```
-
 
 ## chat_template
 

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 `sharegpt` 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 `chat_template` format.  This is the format used when you have the following in your axolotl config:
 
 ```yaml
 datasets:
-  - path: <path to your sharegpt formatted dataset> # example on HF Hub: philschmid/guanaco-sharegpt-style
-    type: sharegpt
+  - path: <path to your chat_template formatted dataset> # example on HF Hub: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
 ```
 
 >[!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_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.
+For example, to mimic the command `cd devtools && CUDA_VISIBLE_DEVICES=0 accelerate launch -m axolotl.cli.train dev_chat_template.yml`, you would use the below configuration[^1].  Note that we add additional flags that override the axolotl config and incorporate the tips above (see the comments). We also set the working directory to `devtools` and set the `env` variable `HF_HOME` to a temporary folder that is later partially deleted.  This is because we want to delete the HF dataset cache before each run in order to ensure that the data preprocessing code is run from scratch.
 
 ```jsonc
 // .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 - sharegpt",
+            "name": "Debug axolotl prompt - chat_template",
             "type": "python",
             "module": "accelerate.commands.launch",
             "request": "launch",
             "args": [
-                "-m", "axolotl.cli.train", "dev_sharegpt.yml",
+                "-m", "axolotl.cli.train", "dev_chat_template.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/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.
+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.
 
 ### 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 ${HOME}/.cache/huggingface:/root/.cache/huggingface winglian/axolotl:main-py3.10-cu118-2.0.1
+docker run --privileged --gpus '"all"' --shm-size 10g --rm -it --name axolotl --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --mount type=bind,src="${PWD}",target=/workspace/axolotl -v ${HOME}/.cache/huggingface:/root/.cache/huggingface axolotlai/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/winglian/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/axolotlai/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/sharegpt.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/chat_template.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

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

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

@@ -2,19 +2,15 @@
  "cells": [
   {
    "cell_type": "markdown",
-   "metadata": {
-    "id": "AKjdG7tbTb-n"
-   },
+   "metadata": {},
    "source": [
-    "# Example notebook for running Axolotl on google colab"
+    "## Setting up"
    ]
   },
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "id": "RcbNpOgWRcii"
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
     "import torch\n",
@@ -22,82 +18,76 @@
     "assert (torch.cuda.is_available()==True)"
    ]
   },
-  {
-   "cell_type": "markdown",
-   "metadata": {
-    "id": "h3nLav8oTRA5"
-   },
-   "source": [
-    "## Install Axolotl and dependencies"
-   ]
-  },
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "colab": {
-     "base_uri": "https://localhost:8080/"
-    },
-    "id": "3c3yGAwnOIdi",
-    "outputId": "e3777b5a-40ef-424f-e181-62dfecd1dd01"
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
-    "!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\""
+    "!pip install axolotl[deepspeed]"
    ]
   },
   {
    "cell_type": "markdown",
-   "metadata": {
-    "id": "BW2MFr7HTjub"
-   },
+   "metadata": {},
    "source": [
-    "## Create an yaml config file"
+    "## Hugging Face login (optional)"
    ]
   },
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "id": "9pkF2dSoQEUN"
-   },
+   "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": {},
    "outputs": [],
    "source": [
     "import yaml\n",
     "\n",
-    "# Your YAML string\n",
     "yaml_string = \"\"\"\n",
-    "base_model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T\n",
-    "model_type: LlamaForCausalLM\n",
-    "tokenizer_type: LlamaTokenizer\n",
+    "base_model: NousResearch/Meta-Llama-3.1-8B\n",
     "\n",
     "load_in_8bit: false\n",
     "load_in_4bit: true\n",
     "strict: false\n",
     "\n",
     "datasets:\n",
-    "  - path: mhenrichsen/alpaca_2k_test\n",
+    "  - path: tatsu-lab/alpaca\n",
     "    type: alpaca\n",
-    "dataset_prepared_path:\n",
+    "dataset_prepared_path: last_run_prepared\n",
     "val_set_size: 0.05\n",
-    "output_dir: ./outputs/qlora-out\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",
     "\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",
@@ -105,12 +95,12 @@
     "wandb_name:\n",
     "wandb_log_model:\n",
     "\n",
-    "gradient_accumulation_steps: 4\n",
-    "micro_batch_size: 2\n",
-    "num_epochs: 4\n",
-    "optimizer: paged_adamw_32bit\n",
+    "gradient_accumulation_steps: 2\n",
+    "micro_batch_size: 1\n",
+    "num_epochs: 1\n",
+    "optimizer: paged_adamw_8bit\n",
     "lr_scheduler: cosine\n",
-    "learning_rate: 0.0002\n",
+    "learning_rate: 2e-5\n",
     "\n",
     "train_on_inputs: false\n",
     "group_by_length: false\n",
@@ -121,13 +111,15 @@
     "gradient_checkpointing: true\n",
     "early_stopping_patience:\n",
     "resume_from_checkpoint:\n",
-    "local_rank:\n",
     "logging_steps: 1\n",
     "xformers_attention:\n",
-    "flash_attention: true\n",
+    "flash_attention: false\n",
+    "sdp_attention: true\n",
     "\n",
-    "warmup_steps: 10\n",
-    "evals_per_epoch: 4\n",
+    "warmup_steps: 1\n",
+    "max_steps: 25\n",
+    "evals_per_epoch: 1\n",
+    "eval_table_size:\n",
     "saves_per_epoch: 1\n",
     "debug:\n",
     "deepspeed:\n",
@@ -135,9 +127,10 @@
     "fsdp:\n",
     "fsdp_config:\n",
     "special_tokens:\n",
-    "\n",
+    "  pad_token: <|end_of_text|>\n",
     "\"\"\"\n",
     "\n",
+    "\n",
     "# Convert the YAML string to a Python dictionary\n",
     "yaml_dict = yaml.safe_load(yaml_string)\n",
     "\n",
@@ -146,31 +139,124 @@
     "\n",
     "# Write the YAML file\n",
     "with open(file_path, 'w') as file:\n",
-    "    yaml.dump(yaml_dict, file)\n"
+    "    yaml.dump(yaml_dict, file)"
    ]
   },
   {
    "cell_type": "markdown",
-   "metadata": {
-    "id": "bidoj8YLTusD"
-   },
+   "metadata": {},
    "source": [
-    "## Launch the training"
+    "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"
    ]
   },
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "colab": {
-     "base_uri": "https://localhost:8080/"
-    },
-    "id": "ydTI2Jk2RStU",
-    "outputId": "d6d0df17-4b53-439c-c802-22c0456d301b"
-   },
+   "metadata": {},
    "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"
    ]
   },
@@ -178,7 +264,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "## Play with inference"
+    "Predict with trained model"
    ]
   },
   {
@@ -187,36 +273,85 @@
    "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",
-    "    --qlora_model_dir=\"./qlora-out\" --gradio"
+    "    --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."
    ]
   }
  ],
  "metadata": {
-  "accelerator": "GPU",
-  "colab": {
-   "gpuType": "T4",
-   "provenance": []
-  },
   "kernelspec": {
-   "display_name": "Python 3 (ipykernel)",
+   "display_name": "Python 3",
    "language": "python",
    "name": "python3"
   },
   "language_info": {
-   "codemirror_mode": {
-    "name": "ipython",
-    "version": 3
-   },
-   "file_extension": ".py",
-   "mimetype": "text/x-python",
    "name": "python",
-   "nbconvert_exporter": "python",
-   "pygments_lexer": "ipython3",
-   "version": "3.12.1"
+   "version": "3.9.6"
   }
  },
  "nbformat": 4,
- "nbformat_minor": 4
+ "nbformat_minor": 2
 }

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

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

examples/gemma2/qlora.yml

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

examples/jamba/qlora_fsdp_large.yaml

@@ -4,11 +4,15 @@ 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_swiglu: true
+liger_glu_activation: true
 liger_fused_linear_cross_entropy: true
 
 strict: false
@@ -14,6 +14,10 @@ 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/mistral/mistral-dpo-qlora.yml

@@ -0,0 +1,93 @@
+#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

@@ -10,7 +10,6 @@ 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

@@ -0,0 +1,67 @@
+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

@@ -0,0 +1,19 @@
+<?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

@@ -0,0 +1,11 @@
+<?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

@@ -0,0 +1,26 @@
+<?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

@@ -0,0 +1,16 @@
+<?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

@@ -0,0 +1,17 @@
+<?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

@@ -0,0 +1,24 @@
+<?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-tests.txt

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

requirements.txt

@@ -1,18 +1,18 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
 packaging==23.2
 peft==0.13.2
-transformers==4.46.0
+transformers==4.46.3
 tokenizers>=0.20.1
 bitsandbytes==0.44.1
-accelerate==1.0.1
-datasets==3.0.1
-deepspeed==0.15.3
+accelerate==1.1.0
+datasets==3.1.0
+deepspeed==0.15.4
 pydantic==2.6.3
 addict
 fire
 PyYAML>=6.0
 requests
-flash-attn==2.6.3
+flash-attn==2.7.0.post2
 sentencepiece
 wandb
 einops
@@ -28,13 +28,12 @@ scipy
 scikit-learn==1.4.2
 pynvml
 art
-fschat @ git+https://github.com/lm-sys/FastChat.git@27a05b04a35510afb1d767ae7e5990cbd278f8fe
 gradio==3.50.2
 tensorboard
 python-dotenv==1.0.1
-autoawq>=0.2.5
+autoawq==0.2.7.post2
 triton>=2.3.0
-liger-kernel==0.3.0
+liger-kernel==0.4.2
 
 mamba-ssm==1.2.0.post1
 
@@ -43,7 +42,7 @@ s3fs>=2024.5.0
 gcsfs>=2024.5.0
 # adlfs
 
-trl @ git+https://github.com/huggingface/trl.git@31d02cfb795284591a084416b9dcb7bef5d08924
+trl==0.12.0
 zstandard==0.22.0
 fastcore
 
@@ -54,3 +53,4 @@ immutabledict==4.2.0
 antlr4-python3-runtime==4.13.2
 
 torchao==0.5.0
+schedulefree==1.3.0

scripts/cloud-entrypoint.sh

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

scripts/unsloth_install.py

@@ -0,0 +1,33 @@
+# 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

@@ -39,7 +39,10 @@ def parse_requirements():
         else:
             # detect the version of torch already installed
             # and set it so dependencies don't clobber the torch version
-            torch_version = version("torch")
+            try:
+                torch_version = version("torch")
+            except PackageNotFoundError:
+                torch_version = "2.5.1"
             _install_requires.append(f"torch=={torch_version}")
 
             version_match = re.match(r"^(\d+)\.(\d+)(?:\.(\d+))?", torch_version)
@@ -54,6 +57,10 @@ def parse_requirements():
 
             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 patch == 0:
@@ -89,22 +96,19 @@ install_requires, dependency_links = parse_requirements()
 
 setup(
     name="axolotl",
-    version="0.4.1",
+    version="0.5.2",
     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(),
+    packages=find_packages("src"),
     install_requires=install_requires,
     dependency_links=dependency_links,
     extras_require={
         "flash-attn": [
-            "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",
+            "flash-attn==2.7.0.post2",
         ],
         "deepspeed": [
-            "deepspeed==0.14.4",
+            "deepspeed==0.15.4",
             "deepspeed-kernels",
         ],
         "mamba-ssm": [

src/axolotl/cli/__init__.py

@@ -30,7 +30,10 @@ from axolotl.common.cli import TrainerCliArgs, load_model_and_tokenizer
 from axolotl.integrations.base import PluginManager
 from axolotl.logging_config import configure_logging
 from axolotl.train import TrainDatasetMeta
-from axolotl.utils.chat_templates import get_chat_template
+from axolotl.utils.chat_templates import (
+    get_chat_template,
+    get_chat_template_from_config,
+)
 from axolotl.utils.comet_ import setup_comet_env_vars
 from axolotl.utils.config import (
     normalize_cfg_datasets,
@@ -190,18 +193,19 @@ 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)
 
@@ -211,13 +215,31 @@ 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()
-        batch = tokenizer(prompt, return_tensors="pt", add_special_tokens=True)
+
+        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)
 
         print("=" * 40)
         model.eval()
@@ -257,13 +279,6 @@ 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
@@ -272,7 +287,7 @@ def do_inference_gradio(
             importlib.import_module("axolotl.prompters"), prompter
         )
     elif cfg.chat_template:
-        chat_template_str = get_chat_template(cfg.chat_template)
+        chat_template_str = get_chat_template(cfg.chat_template, tokenizer=tokenizer)
 
     model = model.to(cfg.device, dtype=cfg.torch_dtype)
 

src/axolotl/cli/preprocess.py

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

src/axolotl/cli/train.py

@@ -19,10 +19,6 @@ 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")
@@ -42,21 +38,6 @@ 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/core/trainer_builder.py

@@ -48,6 +48,7 @@ from trl import (
 )
 from trl.trainer.utils import RewardDataCollatorWithPadding, pad_to_length
 
+from axolotl.integrations.base import PluginManager
 from axolotl.monkeypatch.multipack import SUPPORTED_MULTIPACK_MODEL_TYPES
 from axolotl.monkeypatch.relora import ReLoRACallback, ReLoRAScheduler
 from axolotl.utils import is_comet_available, is_mlflow_available
@@ -435,7 +436,13 @@ 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"]
+            not in [
+                "optimi_adamw",
+                "ao_adamw_8bit",
+                "ao_adamw_4bit",
+                "ao_adamw_fp8",
+                "adopt_adamw",
+            ]
         ):
             return super().create_optimizer()
 
@@ -504,6 +511,14 @@ 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
@@ -895,13 +910,13 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
         for key, value in metrics.items():
             self._stored_metrics[train_eval][key].append(value)
 
-    def _save_checkpoint(self, model, trial):
+    def _save_checkpoint(self, model, trial, **kwargs):
         # 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)
+        return super()._save_checkpoint(model, trial, **kwargs)
 
 
 class AxolotlMambaTrainer(AxolotlTrainer):
@@ -1023,24 +1038,37 @@ class AxolotlDPOTrainer(SchedulerMixin, DPOTrainer):
 
         return super().push_to_hub(*args, **kwargs)
 
+    @staticmethod
     def tokenize_row(
-        self,
         features,
         processing_class,
         max_prompt_length,
         max_completion_length,
         add_special_tokens,
     ) -> Dict:
-        res = super().tokenize_row(
+        res = DPOTrainer.tokenize_row(
             features,
             processing_class,
             max_prompt_length,
             max_completion_length,
             add_special_tokens,
         )
-        if processing_class.bos_token_id is None and res["prompt_input_ids"][0] is None:
+        # fix when the tokenizer doesn't have a bos_token_id, e.g. Qwen
+        if processing_class.bos_token is None and res["prompt_input_ids"][0] is None:
             for key in res.keys():
                 res[key] = res[key][1:]
+
+        if processing_class.bos_token and processing_class.bos_token_id is not None:
+            # dpo trainer may incorrectly prepend the bos_token_id to the dpo outputs
+            if res["chosen_input_ids"][0] == processing_class.bos_token_id:
+                res["chosen_input_ids"] = res["chosen_input_ids"][1:]
+                res["chosen_labels"] = res["chosen_labels"][1:]
+                res["chosen_attention_mask"] = res["chosen_attention_mask"][1:]
+            if res["rejected_input_ids"][0] == processing_class.bos_token_id:
+                res["rejected_input_ids"] = res["rejected_input_ids"][1:]
+                res["rejected_labels"] = res["rejected_labels"][1:]
+                res["rejected_attention_mask"] = res["rejected_attention_mask"][1:]
+
         return res
 
     def training_step(
@@ -1147,6 +1175,12 @@ 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)
@@ -1173,11 +1207,23 @@ 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
@@ -1260,6 +1306,8 @@ 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):
@@ -1377,17 +1425,15 @@ 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["evaluation_strategy"] = "no"
+            training_arguments_kwargs["eval_strategy"] = "no"
         elif self.cfg.eval_steps:
-            training_arguments_kwargs["evaluation_strategy"] = "steps"
+            training_arguments_kwargs["eval_strategy"] = "steps"
             training_arguments_kwargs["eval_steps"] = self.cfg.eval_steps
-        elif self.cfg.evaluation_strategy:
-            training_arguments_kwargs[
-                "evaluation_strategy"
-            ] = self.cfg.evaluation_strategy
+        elif self.cfg.eval_strategy:
+            training_arguments_kwargs["eval_strategy"] = self.cfg.eval_strategy
         else:
             # we have an eval set, but no steps defined, default to use epoch
-            training_arguments_kwargs["evaluation_strategy"] = "epoch"
+            training_arguments_kwargs["eval_strategy"] = "epoch"
 
         if self.cfg.save_steps:
             training_arguments_kwargs["save_strategy"] = "steps"
@@ -1595,7 +1641,8 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         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
+                self.cfg.chat_template,
+                tokenizer=self.tokenizer,
             )
 
         if self.cfg.rl == "orpo":
@@ -1611,11 +1658,13 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         if self.cfg.reward_model:
             trainer_kwargs["max_length"] = self.cfg.sequence_len
 
+        # pylint: disable=duplicate-code
         if self.cfg.optimizer in [
             "optimi_adamw",
             "ao_adamw_4bit",
             "ao_adamw_8bit",
             "ao_adamw_fp8",
+            "adopt_adamw",
         ]:
             # Set default so transformers doesn't throw
             training_arguments_kwargs["optim"] = "adamw_hf"
@@ -1790,7 +1839,7 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
-        callbacks = []
+        callbacks = super().get_post_trainer_create_callbacks(trainer=trainer)
         return callbacks
 
     def build_training_arguments(self, total_num_steps):
@@ -1818,10 +1867,10 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             training_args_kwargs["save_safetensors"] = self.cfg.save_safetensors
 
         if self.eval_dataset:
-            training_args_kwargs["evaluation_strategy"] = "steps"
+            training_args_kwargs["eval_strategy"] = "steps"
             training_args_kwargs["eval_steps"] = self.cfg.eval_steps
         else:
-            training_args_kwargs["evaluation_strategy"] = "no"
+            training_args_kwargs["eval_strategy"] = "no"
 
         if self.cfg.bf16 or self.cfg.bfloat16:
             training_args_kwargs["bf16"] = True
@@ -1876,17 +1925,18 @@ 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"
@@ -1895,13 +1945,13 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             if self.cfg.cpo_alpha is not None:
                 training_args_kwargs["cpo_alpha"] = self.cfg.cpo_alpha
 
-        if self.cfg.rl == "orpo":
+        elif 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
 
-        if self.cfg.rl == "kto":
+        elif self.cfg.rl == "kto":
             training_args_cls = AxolotlKTOConfig
 
             training_args_kwargs["desirable_weight"] = (
@@ -1916,6 +1966,17 @@ 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,
@@ -1936,7 +1997,6 @@ 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:
@@ -1950,12 +2010,6 @@ 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"] = self.cfg.use_wandb
         elif self.cfg.rl == "orpo":
             trainer_cls = AxolotlORPOTrainer
             trainer_cls_args = [self.model]
@@ -1999,11 +2053,11 @@ class HFPPOTrainerBuilder(TrainerBuilderBase):
     """
 
     def get_callbacks(self):
-        callbacks = []
+        callbacks = super().get_callbacks()
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
-        callbacks = []
+        callbacks = super().get_post_trainer_create_callbacks(trainer=trainer)
         return callbacks
 
     def build(self, total_num_steps):

src/axolotl/integrations/base.py

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

src/axolotl/integrations/grokfast/LICENSE

@@ -0,0 +1,21 @@
+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

@@ -0,0 +1,13 @@
+# 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

@@ -0,0 +1,50 @@
+"""
+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

@@ -0,0 +1,15 @@
+"""
+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

@@ -0,0 +1,63 @@
+# 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

@@ -18,20 +18,24 @@ 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.geglu import LigerGEGLUMLP
+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.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):
     """
@@ -42,59 +46,31 @@ class LigerPlugin(BasePlugin):
         return "axolotl.integrations.liger.LigerArgs"
 
     def pre_model_load(self, cfg):
-        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"
+        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.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
-
+            apply_liger_fn(**kwargs)
         elif cfg.model_config_type == "jamba":
             from transformers.models.jamba import modeling_jamba
 
@@ -104,30 +80,14 @@ 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_swiglu:
+            if cfg.liger_glu_activation:
                 modeling_jamba.JambaMLP = LigerSwiGLUMLP
             if cfg.liger_cross_entropy:
-                modeling_jamba.CrossEntropyLoss = LigerCrossEntropyLoss
+                from transformers.loss.loss_utils import nn
+
+                nn.functional.cross_entropy = liger_cross_entropy
             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
@@ -146,44 +106,11 @@ 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_swiglu:
+            if cfg.liger_glu_activation:
                 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

@@ -15,9 +15,12 @@
 """
 Module for handling LIGER input arguments.
 """
+import logging
 from typing import Optional
 
-from pydantic import BaseModel
+from pydantic import BaseModel, model_validator
+
+LOG = logging.getLogger("axolotl.integrations.liger.args")
 
 
 class LigerArgs(BaseModel):
@@ -27,6 +30,24 @@ 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

@@ -0,0 +1,361 @@
+"""
+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

@@ -0,0 +1,622 @@
+"""
+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

@@ -0,0 +1,158 @@
+"""
+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

@@ -1,231 +0,0 @@
-"""
-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/multipack.py

@@ -1,4 +1,5 @@
 """multipack patching for v2 of sample packing"""
+
 import importlib
 
 import transformers
@@ -27,74 +28,28 @@ SUPPORTED_MULTIPACK_MODEL_TYPES = [
 ]
 
 
-# def patch_for_multipack(model_type, model_name=None, is_remote_code=False):
 def patch_for_multipack(model_type, model_name=None, has_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:
+    if has_remote_code:
+        patch_remote(model_name)
     elif hasattr(transformers, "modeling_flash_attention_utils"):
-        if not has_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
-
-    # 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
+        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()
 
-def patch_remote(model_name, config_name, modeling_name):
+
+def patch_remote(model_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)
-    module_name = model_config.__class__.__module__.replace(config_name, modeling_name)
+    parts = model_config.__class__.__module__.split(".")
+    parts[-1] = parts[-1].replace("configuration_", "modeling_", 1)
+    module_name = ".".join(parts)
     modeling_arch = importlib.import_module(module_name)
-    modeling_arch._get_unpad_data = get_unpad_data  # pylint: disable=protected-access
+    if hasattr(modeling_arch, "_get_unpad_data"):
+        modeling_arch._get_unpad_data = (  # pylint: disable=protected-access
+            get_unpad_data
+        )

src/axolotl/monkeypatch/unsloth_.py

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

src/axolotl/prompt_strategies/instruct.py

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

src/axolotl/prompt_strategies/llama2_chat.py

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

src/axolotl/prompt_strategies/sharegpt.py

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

src/axolotl/prompt_strategies/sharegpt_jokes.py

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

src/axolotl/prompt_tokenizers.py

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

src/axolotl/prompters.py

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

src/axolotl/utils/bench.py

@@ -4,6 +4,9 @@ import functools
 import pynvml
 import torch
 from pynvml.nvml import NVMLError
+from transformers.utils.import_utils import is_torch_npu_available
+
+from axolotl.utils.distributed import get_device_type
 
 
 def check_cuda_device(default_value):
@@ -53,6 +56,12 @@ def mps_memory_usage_all():
     return usage, reserved - usage, 0
 
 
+def npu_memory_usage_all(device=0):
+    usage = torch.npu.memory_allocated(device) / 1024.0**3
+    reserved = torch.npu.memory_reserved(device) / 1024.0**3
+    return usage, reserved - usage, 0
+
+
 @check_cuda_device(0.0)
 def gpu_memory_usage_smi(device=0):
     if isinstance(device, torch.device):
@@ -69,8 +78,11 @@ def gpu_memory_usage_smi(device=0):
 
 
 def log_gpu_memory_usage(log, msg, device):
+    cur_device = get_device_type()
     if torch.backends.mps.is_available():
         usage, cache, misc = mps_memory_usage_all()
+    elif "npu" in str(cur_device) and is_torch_npu_available():
+        usage, cache, misc = npu_memory_usage_all(device)
     else:
         usage, cache, misc = gpu_memory_usage_all(device)
     extras = []
@@ -79,6 +91,7 @@ def log_gpu_memory_usage(log, msg, device):
     if misc > 0:
         extras.append(f"+{misc:.03f}GB misc")
     log.info(
-        f"GPU memory usage {msg}: {usage:.03f}GB ({', '.join(extras)})", stacklevel=2
+        f"{str(cur_device)} memory usage {msg}: {usage:.03f}GB ({', '.join(extras)})",
+        stacklevel=2,
     )
     return usage, cache, misc

src/axolotl/utils/callbacks/__init__.py

@@ -64,10 +64,7 @@ class EvalFirstStepCallback(
         control: TrainerControl,
         **kwargs,
     ):
-        if (
-            args.evaluation_strategy == IntervalStrategy.STEPS
-            and state.global_step == 1
-        ):
+        if args.eval_strategy == IntervalStrategy.STEPS and state.global_step == 1:
             control.should_evaluate = True
         return control
 

src/axolotl/utils/config/__init__.py

@@ -1,16 +1,14 @@
 """Module for working with config dicts"""
-import json
 import logging
 import os
-from pathlib import Path
 from typing import Optional
 
 import torch
 from transformers.utils import is_torch_bf16_gpu_available
+from transformers.utils.import_utils import is_torch_npu_available
 
 from axolotl.integrations.config import merge_input_args
 from axolotl.utils.bench import log_gpu_memory_usage
-from axolotl.utils.config.models.input.v0_4_1 import SUPPORTED_METRICS
 from axolotl.utils.config.models.input.v0_4_1 import (
     AxolotlConfigWCapabilities as AxolotlConfigWCapabilitiesBase,
 )
@@ -32,7 +30,10 @@ def choose_device(cfg):
             if torch.backends.mps.is_available():
                 return "mps"
 
-            raise SystemError("No CUDA/mps device found")
+            if is_torch_npu_available():
+                return f"npu:{cfg.local_rank}"
+
+            raise SystemError("No CUDA/mps/npu device found")
         except Exception:  # pylint: disable=broad-exception-caught
             return "cpu"
 
@@ -42,6 +43,8 @@ def choose_device(cfg):
     else:
         if cfg.device.startswith("cuda"):
             cfg.device_map = {"": torch.cuda.current_device()}
+        elif cfg.device.startswith("npu"):
+            cfg.device_map = {"npu": torch.npu.current_device()}
         else:
             cfg.device_map = {"": cfg.device}
 
@@ -215,11 +218,6 @@ def normalize_cfg_datasets(cfg):
     if cfg.chat_template:
         if cfg.datasets:
             for idx, ds_cfg in enumerate(cfg.datasets):
-                if ds_cfg.type == "sharegpt" and not ds_cfg.conversation:
-                    LOG.info(
-                        f"updating dataset {ds_cfg.path} with `conversation: {cfg.chat_template}` to match your chat_template"
-                    )
-                    cfg.datasets[idx].conversation = cfg.chat_template
                 if (
                     ds_cfg.type in ["orpo.chat_template", "chat_template"]
                     and not ds_cfg.chat_template
@@ -252,391 +250,3 @@ def validate_config(cfg: DictDefault, capabilities: Optional[dict] = None):
     return DictDefault(
         dict(AxolotlInputConfig(**cfg.to_dict()).model_dump(exclude_none=True))
     )
-
-
-def legacy_validate_config(cfg):
-    """
-    This is a "pre-validation" step that handles the yaml configuration before we have any
-    information about the model architecture
-    """
-    if is_torch_bf16_gpu_available():
-        if not cfg.bf16 and not cfg.bfloat16:
-            LOG.info("bf16 support detected, but not enabled for this configuration.")
-    else:
-        if (
-            not cfg.merge_lora
-            and not cfg.is_preprocess
-            and (cfg.bf16 is True or cfg.bfloat16 is True)
-        ):
-            raise ValueError(
-                "bf16 requested, but AMP is not supported on this GPU. Requires Ampere series or above."
-            )
-    if (
-        # pylint: disable=too-many-boolean-expressions
-        not (cfg.bf16 or cfg.bfloat16)
-        and (cfg.fp16 or cfg.float16)
-        and not cfg.adapter
-        and not cfg.flash_attention
-        and cfg.sample_packing
-    ):
-        LOG.warning(
-            "Full fine tune w/o FA2 w/ sample packing and fp16/float16 is likely to raise errors. Try LoRA."
-        )
-        # ValueError: Attempting to unscale FP16 gradients.
-        # OR
-        # RuntimeError: expected mat1 and mat2 to have the same dtype, but got: float != c10::Half
-    if cfg.max_packed_sequence_len:
-        raise DeprecationWarning("`max_packed_sequence_len` is no longer supported")
-
-    if cfg.sample_packing and cfg.rl:
-        raise ValueError("`sample_packing: true` does not work with RLHF training")
-
-    if cfg.sample_packing and not cfg.pad_to_sequence_len:
-        LOG.warning(
-            "`pad_to_sequence_len: true` is recommended when using sample_packing"
-        )
-
-    if cfg.gradient_accumulation_steps and cfg.batch_size:
-        raise ValueError(
-            "please set only one of gradient_accumulation_steps or batch_size"
-        )
-    if cfg.batch_size:
-        LOG.warning(
-            "%s\n%s",
-            "batch_size is not recommended. Please use gradient_accumulation_steps instead.",
-            "To calculate the equivalent gradient_accumulation_steps, divide batch_size / micro_batch_size / number of gpus.",
-        )
-    if (
-        cfg.eval_batch_size
-        and cfg.micro_batch_size
-        and cfg.eval_batch_size != cfg.micro_batch_size
-    ):
-        LOG.warning(
-            "eval_batch_size != micro_batch_size. This can lead to VRAM instability."
-        )
-
-    if cfg.adapter == "qlora":
-        if cfg.merge_lora:
-            # can't merge qlora if loaded in 8bit or 4bit
-            if cfg.load_in_8bit:
-                raise ValueError("Can't merge qlora if loaded in 8bit")
-
-            if cfg.gptq:
-                raise ValueError("Can't merge qlora if gptq")
-
-            if cfg.load_in_4bit:
-                raise ValueError("Can't merge qlora if loaded in 4bit")
-
-        else:
-            if cfg.load_in_8bit:
-                raise ValueError("Can't load qlora in 8bit")
-
-            if cfg.gptq:
-                raise ValueError("Can't load qlora if gptq")
-
-            if not cfg.load_in_4bit:
-                raise ValueError("Require cfg.load_in_4bit to be True for qlora")
-
-        if cfg.flash_attn_fuse_qkv or cfg.flash_attn_fuse_mlp:
-            raise ValueError("Fused modules are not supported with QLoRA")
-
-    loftq = cfg.peft and cfg.peft.loftq_config and cfg.peft.loftq_config.loftq_bits
-    if not cfg.load_in_8bit and cfg.adapter == "lora" and not loftq:
-        LOG.warning("We recommend setting `load_in_8bit: true` for LORA finetuning")
-
-    if cfg.adapter == "lora" and (cfg.flash_attn_fuse_qkv or cfg.flash_attn_fuse_mlp):
-        raise ValueError("Fused modules are not supported with LoRA")
-
-    if cfg.adapter and cfg.peft_layers_to_transform and cfg.unfrozen_parameters:
-        raise ValueError(
-            "`unfrozen_parameters` used with `peft_layers_to_transform` can have unexpected behavior."
-        )
-
-    if cfg.relora_steps:
-        if cfg.adapter not in ("lora", "qlora"):
-            raise ValueError("cfg.adapter must be lora or qlora to use ReLoRA")
-
-        if cfg.fsdp:
-            raise ValueError("fsdp not supported with ReLoRA")
-
-        if cfg.deepspeed:
-            raise ValueError("deepspeed not supported with ReLoRA")
-
-        if cfg.lr_scheduler == "one_cycle":
-            raise ValueError("ReLoRA is not compatible with the one_cycle scheduler")
-
-        if cfg.flash_attn_fuse_qkv or cfg.flash_attn_fuse_mlp:
-            raise ValueError("Fused modules are not supported with ReLoRA")
-
-    if cfg.trust_remote_code:
-        LOG.warning(
-            "`trust_remote_code` is set to true. Please make sure that you reviewed the remote code/model."
-        )
-
-    if cfg.push_dataset_to_hub and cfg.hf_use_auth_token is not True:
-        raise ValueError(
-            "Require cfg.hf_use_auth_token to be True for push_dataset_to_hub"
-        )
-
-    if (cfg.base_model and "falcon" in cfg.base_model.lower()) and cfg.fsdp:
-        raise ValueError("FSDP is not supported for falcon models")
-
-    if (
-        cfg.base_model and "mpt" in cfg.base_model.lower()
-    ) and cfg.gradient_checkpointing:
-        raise ValueError("gradient_checkpointing is not supported for MPT models")
-
-    if cfg.flash_optimum is True:
-        if cfg.adapter:
-            LOG.warning("BetterTransformers probably doesn't work with PEFT adapters")
-        if cfg.fp16 or cfg.bf16:
-            raise ValueError("AMP is not supported with BetterTransformer")
-        if cfg.float16 is not True and cfg.bfloat16 is not True:
-            LOG.warning(
-                "You should probably set bfloat16 or float16 to true to "
-                "load the model in float16 for BetterTransformers"
-            )
-        if int(torch.__version__.split(".", maxsplit=1)[0]) < 2:
-            LOG.warning("torch>=2.0.0 required")
-            raise ValueError(
-                f"flash_optimum for BetterTransformers may not be used with {torch.__version__}"
-            )
-
-    if cfg.pretraining_dataset and cfg.group_by_length:
-        LOG.warning(
-            "You probably want to disable group_by_length as it will force a streamed dataset to download completely."
-        )
-    if cfg.pretraining_dataset and not cfg.max_steps:
-        raise ValueError(
-            "max_steps must be set when using iterable pretraining_dataset, Trainer can't infer length and schedule optimizer/learning rate without it!"
-        )
-
-    if any([cfg.adam_beta1, cfg.adam_beta2, cfg.adam_epsilon]) and (
-        not cfg.optimizer or "adamw" not in cfg.optimizer
-    ):
-        LOG.warning("adamw hyperparameters found, but no adamw optimizer set")
-
-    if cfg.push_to_hub_model_id:
-        raise ValueError(
-            "push_to_hub_model_id is deprecated. Please use hub_model_id instead."
-        )
-
-    if cfg.hub_model_id and cfg.save_strategy not in ["steps", "epoch", None]:
-        LOG.warning(
-            "hub_model_id is set without any models being saved. To save a model, set save_strategy to steps, epochs or leave empty."
-        )
-
-    if cfg.gptq and cfg.revision_of_model:
-        raise ValueError(
-            "revision_of_model is not supported for GPTQ models. "
-            + "Please download the model from HuggingFace Hub manually for correct branch, "
-            + "point to its path, and remove revision_of_model from the config."
-        )
-
-    # if cfg.sample_packing and cfg.sdp_attention:
-    #     # incompatible due to bug w/ accelerate causing 0.0 loss when using llama2
-    #     raise ValueError(
-    #         "sample_packing not compatible with sdp_attention. Use flash_attention"
-    #     )
-
-    if cfg.sample_packing and cfg.xformers_attention:
-        raise ValueError(
-            "sample_packing not compatible with xformers_attention. Use flash_attention"
-        )
-
-    if cfg.sample_packing and cfg.sdp_attention and (cfg.bfloat16 or cfg.bf16):
-        # https://github.com/pytorch/pytorch/blob/1b03423526536b5f3d35bdfa95ccc6197556cf9b/test/test_transformers.py#L2440-L2450
-        LOG.warning(
-            "sample_packing & torch sdpa with bf16 is unsupported may results in 0.0 loss. "
-            "This may work on H100s."
-        )
-
-    if cfg.early_stopping_patience:
-        if not cfg.save_steps or not cfg.eval_steps:
-            raise ValueError(
-                "`early_stopping_patience` requires save_steps and eval_steps to be set. eval_steps should evenly divide save_steps."
-            )
-        if cfg.save_steps % cfg.eval_steps != 0:
-            raise ValueError(
-                "`early_stopping_patience` requires that eval_steps should evenly divide save_steps."
-            )
-
-    if cfg.datasets:
-        for idx, ds_cfg in enumerate(cfg.datasets):
-            if not ds_cfg.type:
-                continue
-            if ds_cfg.type == "sharegpt:chat":
-                LOG.warning(
-                    PendingDeprecationWarning(
-                        "`type: sharegpt:chat` will soon be deprecated. simply use `type: sharegpt` instead."
-                    )
-                )
-                cfg.datasets[idx].type = "sharegpt"
-            if "sharegpt_simple" in ds_cfg.type:
-                LOG.warning(
-                    PendingDeprecationWarning(
-                        "`type: sharegpt_simple` will soon be deprecated. simply use `type: sharegpt` instead."
-                    )
-                )
-                cfg.datasets[idx].type = cfg.datasets[idx].type.replace(
-                    "sharegpt_simple", "sharegpt"
-                )
-
-    if cfg.saves_per_epoch and cfg.save_steps:
-        raise ValueError(
-            "save_steps and saves_per_epoch are mutually exclusive and cannot be used together."
-        )
-    if cfg.save_strategy and cfg.saves_per_epoch and cfg.save_strategy != "steps":
-        raise ValueError(
-            "save_strategy must be empty or set to `steps` when used with saves_per_epoch."
-        )
-    if cfg.save_strategy and cfg.save_steps and cfg.save_strategy != "steps":
-        raise ValueError(
-            "save_strategy and save_steps mismatch. Please set save_strategy to 'steps' or remove save_steps."
-        )
-    if cfg.evals_per_epoch and cfg.eval_steps:
-        raise ValueError(
-            "eval_steps and evals_per_epoch are mutually exclusive and cannot be used together."
-        )
-    if (
-        cfg.evals_per_epoch
-        and cfg.evaluation_strategy
-        and cfg.evaluation_strategy != "steps"
-    ):
-        raise ValueError(
-            "evaluation_strategy must be empty or set to `steps` when used with evals_per_epoch."
-        )
-    if (
-        cfg.evaluation_strategy
-        and cfg.eval_steps
-        and cfg.evaluation_strategy != "steps"
-    ):
-        raise ValueError(
-            "evaluation_strategy and eval_steps mismatch. Please set evaluation_strategy to 'steps' or remove eval_steps."
-        )
-
-    if (
-        cfg.val_set_size == 0
-        and (cfg.eval_steps or cfg.evaluation_strategy)
-        and not cfg.test_datasets
-    ):
-        raise ValueError(
-            "eval_steps and evaluation_strategy are not supported with val_set_size == 0"
-        )
-
-    if (
-        cfg.sample_packing
-        and cfg.eval_table_size
-        and cfg.eval_sample_packing is not False
-    ):
-        raise ValueError(
-            "eval_table_size and eval_sample_packing are not supported together with sample_packing. Please set 'eval_sample_packing' to false."
-        )
-
-    if not cfg.adapter and (cfg.load_in_8bit or cfg.load_in_4bit):
-        raise ValueError(
-            "load_in_8bit and load_in_4bit are not supported without setting an adapter."
-            "If you want to full finetune, please turn off load_in_8bit and load_in_4bit."
-        )
-
-    if cfg.rope_scaling:
-        LOG.warning("`rope_scaling` should now be be a key under `model_config`")
-
-    if cfg.wandb_run_id and not cfg.wandb_name:
-        cfg.wandb_name = cfg.wandb_run_id
-
-        LOG.warning(
-            "wandb_run_id sets the ID of the run. If you would like to set the name, please use wandb_name instead."
-        )
-
-    if cfg.noisy_embedding_alpha is not None:
-        # Deprecated, use neftune_noise_alpha
-        LOG.warning("noisy_embedding_alpha is deprecated, use neftune_noise_alpha")
-        if cfg.neftune_noise_alpha is None:
-            cfg.neftune_noise_alpha = cfg.noisy_embedding_alpha
-        else:
-            # User is providing both; bail and have them sort out their settings
-            raise ValueError(
-                "noisy_embedding_alpha is deprecated, use neftune_noise_alpha; both are set, please remove the deprecated noisy_embedding_alpha setting"
-            )
-
-    if cfg.neftune_noise_alpha is not None and cfg.neftune_noise_alpha <= 0.0:
-        raise ValueError("neftune_noise_alpha must be > 0.0")
-
-    if cfg.max_memory is not None and cfg.gpu_memory_limit is not None:
-        raise ValueError(
-            "max_memory and gpu_memory_limit are mutually exclusive and cannot be used together."
-        )
-
-    if (
-        cfg.unfrozen_parameters
-        and cfg.gradient_checkpointing_kwargs
-        and cfg.gradient_checkpointing_kwargs.use_reentrant is True
-    ):
-        # https://github.com/huggingface/transformers/issues/21381
-        raise ValueError(
-            "`use_reentrant` must be false when used with partially frozen model."
-        )
-
-    if cfg.deepspeed and Path(cfg.deepspeed).is_file():
-        with open(cfg.deepspeed, encoding="utf-8") as file:
-            contents = file.read()
-            deepspeed_cfg: DictDefault = DictDefault(json.loads(contents))
-            if cfg.flash_attention:
-                if (
-                    deepspeed_cfg.zero_optimization
-                    and deepspeed_cfg.zero_optimization.stage == 3
-                ):
-                    if not (
-                        (
-                            deepspeed_cfg.bf16
-                            and deepspeed_cfg.bf16.enabled  # pylint: disable=no-member
-                            is True
-                        )
-                        or (
-                            deepspeed_cfg.fp16
-                            and deepspeed_cfg.fp16.enabled  # pylint: disable=no-member
-                            is True
-                        )
-                    ):
-                        raise ValueError(
-                            "bf16.enabled or fp16.enabled must be set to true when using ZeRO-3 with flash-attention"
-                        )
-            if "8bit" in cfg.optimizer and deepspeed_cfg.optimizer:
-                LOG.warning(
-                    f"conflicting optimizer: {cfg.optimizer} used alongside deepspeed optimizer."
-                )
-
-    if cfg.test_datasets and cfg.val_set_size:
-        raise ValueError(
-            "non-zero val_set_size should not be used with test_datasets configuration"
-        )
-
-    if cfg.fsdp and "bnb" in cfg.optimizer:
-        raise ValueError(f"FSDP not compatible with {cfg.optimizer}")
-
-    if cfg.do_causal_lm_eval and cfg.eval_sample_packing:
-        raise ValueError(
-            "do_causal_lm_eval is enabled, eval_sample_packing must be set to False"
-        )
-
-    if cfg.eval_causal_lm_metrics:
-        if not isinstance(cfg.eval_causal_lm_metrics, list):
-            raise ValueError("eval_causal_lm_metrics must be a list")
-        # only ["sacrebleu", "comet", "ter", "chrf"] supported
-        if set(cfg.eval_causal_lm_metrics) - SUPPORTED_METRICS:
-            raise ValueError(
-                f"eval_causal_lm_metrics must be one of {SUPPORTED_METRICS}"
-            )
-
-    # TODO
-    # MPT 7b
-    # https://github.com/facebookresearch/bitsandbytes/issues/25
-    # no 8bit adaAmw w bf16
-
-    # GPT-NeoX
-    # evals broken when extending context len
-    # File "/root/miniconda3/envs/py3.9/lib/python3.9/site-packages/transformers/models/gpt_neox/modeling_gpt_neox.py", line 162, in forward                        attn_output, attn_weights = self._attn(query, key, value, attention_mask, head_mask)
-    # File "/root/miniconda3/envs/py3.9/lib/python3.9/site-packages/optimum/bettertransformer/models/attention.py", line 74, in gpt2_wrapped_scaled_dot_product
-    # attention_mask = causal_mask + attention_mask
-    # RuntimeError: The size of tensor a (2048) must match the size of tensor b (8132) at non-singleton dimension 3

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

@@ -7,7 +7,6 @@ Module for pydantic models for configuration
 import logging
 import os
 from enum import Enum
-from importlib.metadata import version
 from typing import Annotated, Any, Dict, List, Literal, Optional, Tuple, Union
 
 from pydantic import (
@@ -20,6 +19,7 @@ from pydantic import (
 )
 from transformers import SchedulerType
 from transformers.training_args import OptimizerNames
+from transformers.utils.import_utils import is_torch_npu_available
 
 from axolotl.utils.config.models.internals import GPUCapabilities
 
@@ -57,6 +57,8 @@ class ChatTemplate(str, Enum):
     jinja = "jinja"  # pylint: disable=invalid-name
     qwen_25 = "qwen_25"  # pylint: disable=invalid-name
     tokenizer_default = "tokenizer_default"  # pylint: disable=invalid-name
+    exaone = "exaone"  # pylint: disable=invalid-name
+    metharme = "metharme"  # pylint: disable=invalid-name
 
 
 class DeprecatedParameters(BaseModel):
@@ -66,6 +68,7 @@ class DeprecatedParameters(BaseModel):
     rope_scaling: Optional[Any] = None
     noisy_embedding_alpha: Optional[float] = None
     dpo_beta: Optional[float] = None
+    evaluation_strategy: Optional[str] = None
 
     @field_validator("max_packed_sequence_len")
     @classmethod
@@ -97,6 +100,13 @@ class DeprecatedParameters(BaseModel):
             LOG.warning("dpo_beta is deprecated, use rl_beta instead")
         return dpo_beta
 
+    @field_validator("evaluation_strategy")
+    @classmethod
+    def validate_evaluation_strategy(cls, evaluation_strategy):
+        if evaluation_strategy is not None:
+            LOG.warning("evaluation_strategy is deprecated, use eval_strategy instead")
+        return evaluation_strategy
+
 
 class RemappedParameters(BaseModel):
     """parameters that have been remapped to other names"""
@@ -240,8 +250,10 @@ class KTODataset(BaseModel):
 class LoftQConfig(BaseModel):
     """LoftQ configuration subset"""
 
-    loftq_bits: int = Field(default=4, metadata={"help": "Quantization bits for LoftQ"})
-    # loftq_iter: int = Field(default=1, metadata={"help": "Alternating iterations for LoftQ"})
+    loftq_bits: int = Field(
+        default=4, json_schema_extra={"description": "Quantization bits for LoftQ"}
+    )
+    # loftq_iter: int = Field(default=1, json_schema_extra={"description": "Alternating iterations for LoftQ"})
 
 
 class PeftConfig(BaseModel):
@@ -284,8 +296,8 @@ class LoraConfig(BaseModel):
 
     qlora_sharded_model_loading: Optional[bool] = Field(
         default=False,
-        metadata={
-            "help": "load qlora model in sharded format for FSDP using answer.ai technique."
+        json_schema_extra={
+            "description": "load qlora model in sharded format for FSDP using answer.ai technique."
         },
     )
     lora_on_cpu: Optional[bool] = None
@@ -294,13 +306,15 @@ class LoraConfig(BaseModel):
 
     loraplus_lr_ratio: Optional[float] = Field(
         default=None,
-        metadata={
-            "help": "loraplus learning rate ratio lr_B / lr_A. Recommended value is 2^4."
+        json_schema_extra={
+            "description": "loraplus learning rate ratio lr_B / lr_A. Recommended value is 2^4."
         },
     )
     loraplus_lr_embedding: Optional[float] = Field(
         default=1e-6,
-        metadata={"help": "loraplus learning rate for lora embedding layers."},
+        json_schema_extra={
+            "description": "loraplus learning rate for lora embedding layers."
+        },
     )
 
     merge_lora: Optional[bool] = None
@@ -370,10 +384,10 @@ class ModelInputConfig(BaseModel):
     tokenizer_use_fast: Optional[bool] = None
     tokenizer_legacy: Optional[bool] = None
     tokenizer_type: Optional[str] = Field(
-        default=None, metadata={"help": "transformers tokenizer class"}
+        default=None, json_schema_extra={"description": "transformers tokenizer class"}
     )
     processor_type: Optional[str] = Field(
-        default=None, metadata={"help": "transformers processor class"}
+        default=None, json_schema_extra={"description": "transformers processor class"}
     )
     trust_remote_code: Optional[bool] = None
 
@@ -395,18 +409,18 @@ class HyperparametersConfig(BaseModel):
     gradient_accumulation_steps: Optional[int] = Field(default=1)
     micro_batch_size: Optional[int] = Field(
         default=1,
-        metadata={"help": "per gpu micro batch size for training"},
+        json_schema_extra={"description": "per gpu micro batch size for training"},
     )
     batch_size: Optional[int] = Field(
         default=None,
-        metadata={
-            "help": "Total batch size, we do not recommended setting this manually"
+        json_schema_extra={
+            "description": "Total batch size, we do not recommended setting this manually"
         },
     )
     eval_batch_size: Optional[int] = Field(
         default=None,
-        metadata={
-            "help": "per gpu micro batch size for evals, defaults to value of micro_batch_size"
+        json_schema_extra={
+            "description": "per gpu micro batch size for evals, defaults to value of micro_batch_size"
         },
     )
 
@@ -426,16 +440,18 @@ class HyperparametersConfig(BaseModel):
                 "ao_adamw_4bit",
                 "ao_adamw_8bit",
                 "ao_adamw_fp8",
+                "adopt_adamw",
             ],
         ]
     ] = OptimizerNames.ADAMW_HF.value
     optim_args: Optional[Union[str, Dict[str, Any]]] = Field(
-        default=None, metadata={"help": "Optional arguments to supply to optimizer."}
+        default=None,
+        json_schema_extra={"description": "Optional arguments to supply to optimizer."},
     )
     optim_target_modules: Optional[Union[List[str], Literal["all_linear"]]] = Field(
         default=None,
-        metadata={
-            "help": "The target modules to optimize, i.e. the module names that you would like to train."
+        json_schema_extra={
+            "description": "The target modules to optimize, i.e. the module names that you would like to train."
         },
     )
     torchdistx_path: Optional[str] = None
@@ -495,15 +511,15 @@ class LISAConfig(BaseModel):
 
     lisa_n_layers: Optional[int] = Field(
         default=None,
-        metadata={"help": "the number of activate layers in LISA"},
+        json_schema_extra={"description": "the number of activate layers in LISA"},
     )
     lisa_step_interval: Optional[int] = Field(
         default=None,
-        metadata={"help": "how often to switch layers in LISA"},
+        json_schema_extra={"description": "how often to switch layers in LISA"},
     )
     lisa_layers_attribute: Optional[str] = Field(
         default="model.layers",
-        metadata={"help": "path under the model to access the layers"},
+        json_schema_extra={"description": "path under the model to access the layers"},
     )
 
 
@@ -587,6 +603,9 @@ class AxolotlInputConfig(
 
     rl: Optional[RLType] = None
     reward_model: Optional[bool] = None
+    dpo_use_weighting: Optional[
+        bool
+    ] = None  # whether to use weighting in DPO trainer. If none, default is false in the trainer.
 
     datasets: Optional[conlist(Union[SFTDataset, DPODataset, KTODataset], min_length=1)] = None  # type: ignore
     test_datasets: Optional[conlist(Union[SFTDataset, DPODataset, KTODataset], min_length=1)] = None  # type: ignore
@@ -599,7 +618,8 @@ class AxolotlInputConfig(
     pretraining_dataset: Optional[  # type: ignore
         conlist(Union[PretrainingDataset, SFTDataset], min_length=1)
     ] = Field(
-        default=None, metadata={"help": {"streaming dataset to use for pretraining"}}
+        default=None,
+        json_schema_extra={"description": "streaming dataset to use for pretraining"},
     )
     dataset_processes: Optional[int] = Field(default=os.cpu_count())
     dataset_keep_in_memory: Optional[bool] = None
@@ -659,7 +679,8 @@ class AxolotlInputConfig(
     sequence_len: int = Field(default=512)
     min_sample_len: Optional[int] = None
     max_prompt_len: int = Field(
-        default=512, metadata={"help": "maximum prompt length for RL training"}
+        default=512,
+        json_schema_extra={"description": "maximum prompt length for RL training"},
     )
     sample_packing: Optional[bool] = None
     sample_packing_group_size: Optional[int] = 100_000
@@ -678,8 +699,8 @@ class AxolotlInputConfig(
     pretrain_multipack_buffer_size: Optional[int] = 10_000
     pretrain_multipack_attn: Optional[bool] = Field(
         default=True,
-        metadata={
-            "help": "whether to prevent cross attention for packed sequences during pretraining",
+        json_schema_extra={
+            "description": "whether to prevent cross attention for packed sequences during pretraining",
         },
     )
 
@@ -725,7 +746,7 @@ class AxolotlInputConfig(
     warmup_ratio: Optional[float] = None
     eval_steps: Optional[Union[int, float]] = None
     evals_per_epoch: Optional[Union[int]] = None
-    evaluation_strategy: Optional[str] = None
+    eval_strategy: Optional[str] = None
     save_steps: Optional[Union[int, float]] = None
     saves_per_epoch: Optional[int] = None
     save_strategy: Optional[str] = None
@@ -777,28 +798,25 @@ class AxolotlInputConfig(
     is_mistral_derived_model: Optional[bool] = Field(default=None)
     is_qwen_derived_model: Optional[bool] = Field(default=None)
 
+    plugins: Optional[List[str]] = Field(default=None)
+
     @field_validator("datasets", mode="before")
     @classmethod
-    def fix_sharegpt_datasets(cls, datasets):
-        for idx, ds_cfg in enumerate(datasets):
-            if not ds_cfg["type"]:
+    def deprecate_sharegpt_datasets(cls, datasets):
+        for _, ds_cfg in enumerate(datasets):
+            if not ds_cfg.get("type"):
                 continue
-            if ds_cfg["type"] == "sharegpt:chat":
-                LOG.warning(
-                    PendingDeprecationWarning(
-                        "`type: sharegpt:chat` will soon be deprecated. simply use `type: sharegpt` instead."
-                    )
-                )
-                datasets[idx]["type"] = "sharegpt"
-            if "sharegpt_simple" in ds_cfg["type"]:
-                LOG.warning(
-                    PendingDeprecationWarning(
-                        "`type: sharegpt_simple` will soon be deprecated. simply use `type: sharegpt` instead."
-                    )
-                )
-                datasets[idx]["type"] = datasets[idx]["type"].replace(
-                    "sharegpt_simple", "sharegpt"
+
+            ds_type = ds_cfg["type"]
+            # skip if it's a dict (for custom user instruction prompt)
+            if isinstance(ds_type, dict):
+                continue
+
+            if isinstance(ds_type, str) and ds_type.startswith("sharegpt"):
+                raise ValueError(
+                    "`type: sharegpt.*` is deprecated. Please use `type: chat_template` instead."
                 )
+
         return datasets
 
     @model_validator(mode="before")
@@ -1030,21 +1048,21 @@ class AxolotlInputConfig(
     @classmethod
     def check_evals(cls, data):
         if (
-            data.get("evaluation_strategy")
+            data.get("eval_strategy")
             and data.get("eval_steps")
-            and data.get("evaluation_strategy") != "steps"
+            and data.get("eval_strategy") != "steps"
         ):
             raise ValueError(
-                "evaluation_strategy and eval_steps mismatch. Please set evaluation_strategy to 'steps' or remove eval_steps."
+                "eval_strategy and eval_steps mismatch. Please set eval_strategy to 'steps' or remove eval_steps."
             )
 
         if (
             data.get("val_set_size") == 0
-            and (data.get("eval_steps") or data.get("evaluation_strategy"))
+            and (data.get("eval_steps") or data.get("eval_strategy"))
             and not data.get("test_datasets")
         ):
             raise ValueError(
-                "eval_steps and evaluation_strategy are not supported with val_set_size == 0"
+                "eval_steps and eval_strategy are not supported with val_set_size == 0"
             )
         if data.get("evals_per_epoch") and data.get("eval_steps"):
             raise ValueError(
@@ -1052,11 +1070,11 @@ class AxolotlInputConfig(
             )
         if (
             data.get("evals_per_epoch")
-            and data.get("evaluation_strategy")
-            and data.get("evaluation_strategy") != "steps"
+            and data.get("eval_strategy")
+            and data.get("eval_strategy") != "steps"
         ):
             raise ValueError(
-                "evaluation_strategy must be empty or set to `steps` when used with evals_per_epoch."
+                "eval_strategy must be empty or set to `steps` when used with evals_per_epoch."
             )
 
         if data.get("do_bench_eval") and not (
@@ -1288,6 +1306,26 @@ class AxolotlInputConfig(
             )
         return data
 
+    @model_validator(mode="before")
+    @classmethod
+    def warn_qlora_zero3_w_use_reentrant(cls, data):
+        if (
+            data.get("adapter") == "qlora"
+            and data.get("gradient_checkpointing_kwargs", {})
+            and data.get("gradient_checkpointing_kwargs", {}).get("use_reentrant")
+            is False
+            and data.get("deepspeed", "") is not None
+            and "zero3" in data.get("deepspeed", "")
+        ):
+            # may result in:
+            # torch.utils.checkpoint.CheckpointError: torch.utils.checkpoint:
+            # Recomputed values for the following tensors have different metadata
+            # than during the forward pass.
+            LOG.warning(
+                "qlora + zero3 with use_reentrant: false may result in a CheckpointError about recomputed values"
+            )
+        return data
+
     @model_validator(mode="before")
     @classmethod
     def check_val_w_test_datasets(cls, data):
@@ -1297,6 +1335,19 @@ class AxolotlInputConfig(
             )
         return data
 
+    @model_validator(mode="before")
+    @classmethod
+    def check_eval_strategy(cls, data):
+        if (
+            data.get("evaluation_strategy") is not None
+            and data.get("eval_strategy") is None
+        ):
+            LOG.info(
+                "explicitly setting `eval_strategy` from the `evaluation_strategy`"
+            )
+            data["eval_strategy"] = data.get("evaluation_strategy")
+        return data
+
     @model_validator(mode="before")
     @classmethod
     def check_fsdp_offload_w_8bit_optimizer(cls, data):
@@ -1375,21 +1426,6 @@ class AxolotlInputConfig(
                 )
         return data
 
-    @model_validator(mode="before")
-    @classmethod
-    def check_unsloth_xformers_version(cls, data):
-        if (
-            data.get("unsloth_lora_mlp")
-            or data.get("unsloth_lora_qkv")
-            or data.get("unsloth_lora_o")
-        ):
-            xformers_version = version("xformers")
-            if xformers_version == "0.0.27":
-                raise ValueError(
-                    "xformers version 0.0.27 is not supported with unsloth. Please downgrade to 0.0.26.post1"
-                )
-        return data
-
     @model_validator(mode="before")
     @classmethod
     def check_torch_compile_deepspeed(cls, data):
@@ -1399,6 +1435,40 @@ class AxolotlInputConfig(
             )
         return data
 
+    @model_validator(mode="before")
+    @classmethod
+    def check_npu_config(cls, data):
+        if is_torch_npu_available():
+            # check attention config
+            attn_list = ["flash_attention", "sdp_attention", "s2_attention"]
+            for attn in attn_list:
+                if data.get(attn):
+                    raise NotImplementedError(
+                        f"{attn} is currently not supported in Ascend npu, please disable this configuration."
+                    )
+
+            # check quant config
+            if data.get("optimizer") is not None and "bit" in data.get("optimizer"):
+                optimizer = data.get("optimizer")
+                raise NotImplementedError(
+                    f"{optimizer} is currently not supported in Ascend npu, choose another one please."
+                )
+
+            quant_list = ["load_in_8bit", "load_in_4bit"]
+            for quant in quant_list:
+                if data.get(quant):
+                    raise NotImplementedError(
+                        f"Quantification is currently not supported in Ascend npu, please disable {quant}."
+                    )
+
+            # check dtype config
+            if data.get("tf32"):
+                raise NotImplementedError(
+                    "tf32 dtype is currently not supported in Ascend npu, please disable this configuration"
+                )
+
+        return data
+
 
 class AxolotlConfigWCapabilities(AxolotlInputConfig):
     """wrapper to valdiate gpu capabilities with the configured options"""

src/axolotl/utils/data/rl.py

@@ -64,15 +64,57 @@ def map_dataset(cfg, data_set, ds_transform_fn, tokenizer):
             tokenizer = load_tokenizer(cfg)
         ds_transform_fn = partial(ds_transform_fn, tokenizer=tokenizer)
 
+    if isinstance(data_set, DatasetDict):
+        data_set = data_set["train"]
+
     data_set = data_set.map(
         ds_transform_fn,
         desc="Mapping RL Dataset",
     )
-    if isinstance(data_set, DatasetDict):
-        data_set = data_set["train"]
+
     return data_set
 
 
+def drop_long_rl_seq(
+    sample, rl, tokenizer, sequence_len  # pylint: disable=invalid-name
+):
+    if rl in ("dpo", "ipo", "orpo", "simpo"):
+        if not (
+            sample.get("prompt") and sample.get("chosen") and sample.get("rejected")
+        ):
+            raise ValueError(
+                "Prompt, chosen and rejected keys are required for DPO/ORPO datasets"
+            )
+
+        prompt = sample["prompt"]
+        chosen = sample["chosen"]
+        rejected = sample["rejected"]
+
+        len_prompt = len(tokenizer(prompt, add_special_tokens=False)["input_ids"])
+        len_chosen = len(tokenizer(chosen, add_special_tokens=False)["input_ids"])
+        len_rejected = len(tokenizer(rejected, add_special_tokens=False)["input_ids"])
+
+        return (len_prompt + len_chosen) <= sequence_len and (
+            len_prompt + len_rejected
+        ) <= sequence_len
+
+    if rl == "kto":
+        if not (sample.get("prompt") and sample.get("completion")):
+            raise ValueError("Prompt and completion keys are required for KTO datasets")
+
+        prompt = sample["prompt"]
+        completion = sample["completion"]
+
+        len_prompt = len(tokenizer(prompt, add_special_tokens=False)["input_ids"])
+        len_completion = len(
+            tokenizer(completion, add_special_tokens=False)["input_ids"]
+        )
+
+        return (len_prompt + len_completion) <= sequence_len
+
+    raise ValueError("Unknown RL type")
+
+
 def load_prepare_dpo_datasets(cfg):
     def load_split(dataset_cfgs, _cfg):
         split_datasets: List[Any] = []
@@ -94,7 +136,7 @@ def load_prepare_dpo_datasets(cfg):
                 )
                 split_datasets.insert(i, ds)
 
-        tokenizer = None
+        tokenizer = load_tokenizer(cfg)
 
         for i, data_set in enumerate(split_datasets):
             _type = dataset_cfgs[i]["type"]
@@ -121,7 +163,28 @@ def load_prepare_dpo_datasets(cfg):
                 # "prompt", "chosen" and "rejected" already preprocessed
                 split_datasets[i] = data_set
 
-        return concatenate_datasets(split_datasets)
+            drop_long = partial(
+                drop_long_rl_seq,
+                rl=_cfg.rl,
+                tokenizer=tokenizer,
+                sequence_len=cfg.sequence_len,
+            )
+
+            prior_len = len(split_datasets[i])
+            split_datasets[i] = split_datasets[i].filter(
+                drop_long,
+                num_proc=cfg.dataset_processes,
+                load_from_cache_file=not cfg.is_preprocess,
+                desc="Dropping Long Sequences",
+            )
+            dropped = prior_len - len(split_datasets[i])
+            if dropped:
+                LOG.warning(f"Dropped {dropped} long samples from dataset index {i}")
+
+        combined_datasets = concatenate_datasets(split_datasets)
+        combined_datasets = combined_datasets.shuffle(seed=cfg.seed)
+
+        return combined_datasets
 
     with zero_first(is_main_process()):
         train_is_preprocessed = False

src/axolotl/utils/data/sft.py

@@ -2,9 +2,11 @@
 
 import functools
 import logging
+import time
 from pathlib import Path
 from typing import List, Optional, Tuple, Union
 
+import requests
 from datasets import (
     Dataset,
     DatasetDict,
@@ -53,6 +55,28 @@ from axolotl.utils.trainer import (
 LOG = logging.getLogger("axolotl")
 
 
+def retry_on_request_exceptions(max_retries=3, delay=1):
+    def decorator(func):
+        @functools.wraps(func)
+        def wrapper(*args, **kwargs):  # pylint: disable=inconsistent-return-statements
+            for attempt in range(max_retries):
+                try:
+                    return func(*args, **kwargs)
+                except (
+                    requests.exceptions.ReadTimeout,
+                    requests.exceptions.ConnectionError,
+                ) as exc:
+                    if attempt < max_retries - 1:
+                        time.sleep(delay)
+                    else:
+                        raise exc
+
+        return wrapper
+
+    return decorator
+
+
+@retry_on_request_exceptions(max_retries=3, delay=5)
 def prepare_dataset(cfg, tokenizer, processor=None):
     prompters = []
     if not cfg.pretraining_dataset:
@@ -236,6 +260,7 @@ def load_tokenized_prepared_datasets(
         for config_dataset in for_d_in_datasets(cfg_datasets):
             ds: Optional[Union[Dataset, DatasetDict]] = None
             ds_from_hub = False
+            ds_trust_remote_code = config_dataset.trust_remote_code
             try:
                 # this is just a basic check to see if the path is a
                 # valid HF dataset that's loadable
@@ -245,6 +270,7 @@ def load_tokenized_prepared_datasets(
                     streaming=True,
                     token=use_auth_token,
                     revision=config_dataset.revision,
+                    trust_remote_code=ds_trust_remote_code,
                 )
                 ds_from_hub = True
             except (FileNotFoundError, ConnectionError, HFValidationError, ValueError):
@@ -324,7 +350,15 @@ def load_tokenized_prepared_datasets(
                             split=None,
                         )
                     else:
-                        ds = load_from_disk(config_dataset.path)
+                        try:
+                            ds = load_from_disk(config_dataset.path)
+                        except FileNotFoundError:
+                            ds = load_dataset(
+                                config_dataset.path,
+                                name=config_dataset.name,
+                                streaming=False,
+                                split=None,
+                            )
                 elif local_path.is_file():
                     ds_type = get_ds_type(config_dataset)
 
@@ -342,7 +376,7 @@ def load_tokenized_prepared_datasets(
             elif ds_from_hub:
                 load_ds_kwargs = {}
                 if config_dataset.split:
-                    load_ds_kwargs = {"split": config_dataset.split}
+                    load_ds_kwargs["split"] = config_dataset.split
                 ds = load_dataset(
                     config_dataset.path,
                     name=config_dataset.name,
@@ -350,6 +384,7 @@ def load_tokenized_prepared_datasets(
                     data_files=config_dataset.data_files,
                     token=use_auth_token,
                     revision=config_dataset.revision,
+                    trust_remote_code=config_dataset.trust_remote_code,
                     **load_ds_kwargs,
                 )
             elif ds_from_cloud and remote_file_system:
@@ -367,6 +402,7 @@ def load_tokenized_prepared_datasets(
                         streaming=False,
                         split=None,
                         storage_options=storage_options,
+                        trust_remote_code=config_dataset.trust_remote_code,
                     )
             elif config_dataset.path.startswith("https://"):
                 ds_type = get_ds_type(config_dataset)
@@ -377,6 +413,7 @@ def load_tokenized_prepared_datasets(
                     streaming=False,
                     split=None,
                     storage_options=storage_options,
+                    trust_remote_code=config_dataset.trust_remote_code,
                 )
             else:
                 if isinstance(config_dataset.data_files, str):

src/axolotl/utils/distributed.py

@@ -9,10 +9,44 @@ from datetime import timedelta
 import torch
 import torch.distributed as dist
 from accelerate import PartialState
+from transformers.utils.import_utils import (
+    is_torch_cuda_available,
+    is_torch_mps_available,
+    is_torch_npu_available,
+)
 
 distributed_state = None  # pylint: disable=invalid-name
 
 
+def get_device_type():
+    device = torch.device("cpu")
+    if is_torch_cuda_available():
+        device = torch.device("cuda")
+    elif is_torch_mps_available():
+        device = torch.device("mps")
+    elif is_torch_npu_available():
+        device = torch.device("npu")
+    return device
+
+
+def get_device_count():
+    cur_device = get_device_type()
+    if "cuda" in str(cur_device):
+        return torch.cuda.device_count()
+    if "npu" in str(cur_device):
+        return torch.npu.device_count()
+    return 1
+
+
+def get_current_device():
+    cur_device = get_device_type()
+    if "cuda" in str(cur_device):
+        return torch.cuda.current_device()
+    if "npu" in str(cur_device):
+        return torch.npu.current_device()
+    return 0
+
+
 def is_distributed():
     """
     Check if distributed training is initialized.
@@ -91,7 +125,7 @@ def gather_scalar_from_all_ranks(fn, world_size=1):  # pylint: disable=invalid-n
     if not is_distributed():
         return [value_scalar]
     value_tensor = torch.tensor(
-        value_scalar, device=torch.cuda.current_device()
+        value_scalar, device=f"{get_device_type()}:{get_current_device()}"
     ).float()
 
     if not is_main_process():
@@ -115,13 +149,14 @@ def broadcast_dict(vals: dict):
     if not is_distributed():
         return vals
 
+    cur_device = get_device_type()
     if is_main_process():
         data_byte = pickle.dumps(vals)
-        data_tensor = torch.ByteTensor(list(data_byte)).to("cuda")
-        data_size = torch.IntTensor([len(data_byte)]).to("cuda")
+        data_tensor = torch.ByteTensor(list(data_byte)).to(cur_device)
+        data_size = torch.IntTensor([len(data_byte)]).to(cur_device)
     else:
-        data_tensor = torch.empty([1024], dtype=torch.uint8, device="cuda")
-        data_size = torch.IntTensor([0]).to("cuda")
+        data_tensor = torch.empty([1024], dtype=torch.uint8, device=cur_device)
+        data_size = torch.IntTensor([0]).to(cur_device)
 
     dist.broadcast(data_size, 0)
     if not is_main_process():
@@ -150,14 +185,15 @@ def compute_and_broadcast(fn):  # pylint: disable=invalid-name
     Returns:
     - The computed value (int or float).
     """
+    cur_device = f"{get_device_type()}:{get_current_device()}"
     if is_main_process():
         value_scalar = fn()
         value_tensor = torch.tensor(
-            value_scalar, device=torch.cuda.current_device(), dtype=torch.float32
+            value_scalar, device=cur_device, dtype=torch.float32
         )
     else:
         value_tensor = torch.tensor(
-            0.0, device=torch.cuda.current_device(), dtype=torch.float32
+            0.0, device=cur_device, dtype=torch.float32
         )  # Placeholder tensor
 
     # Broadcast the tensor to all processes.
@@ -184,7 +220,7 @@ def gather_from_all_ranks(fn, world_size=1):  # pylint: disable=invalid-name
     """
     value_scalar = fn()
     value_tensor = torch.tensor(
-        value_scalar, device=torch.cuda.current_device()
+        value_scalar, device=f"{get_device_type()}:{get_current_device()}"
     ).float()
 
     # Placeholder tensor for gathering results

src/axolotl/utils/environment.py

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

src/axolotl/utils/gradient_checkpointing/unsloth.py

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

src/axolotl/utils/models.py

@@ -46,6 +46,7 @@ from transformers.integrations.deepspeed import (
 )
 
 from axolotl.common.architectures import MOE_ARCH_BLOCK
+from axolotl.integrations.sageattention.lib.core import monkeypatch_sdp_w_sage_attention
 from axolotl.models.mamba import fix_mamba_attn_for_loss
 from axolotl.monkeypatch.multipack import (
     SUPPORTED_MULTIPACK_MODEL_TYPES,
@@ -55,7 +56,7 @@ from axolotl.prompt_tokenizers import LLAMA_DEFAULT_EOS_TOKEN
 from axolotl.utils.bench import log_gpu_memory_usage
 from axolotl.utils.chat_templates import get_chat_template_from_config
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.distributed import zero_only
+from axolotl.utils.distributed import get_device_count, get_device_type, zero_only
 from axolotl.utils.gradient_checkpointing import hf_grad_checkpoint_unsloth_wrapper
 from axolotl.utils.lora_embeddings import get_linear_embedding_layers
 from axolotl.utils.model_shard_quant import load_sharded_model, load_sharded_model_quant
@@ -238,6 +239,7 @@ def load_tokenizer(cfg):
                         x in cfg.lora_modules_to_save for x in lora_modules_to_save
                     )
                 )
+                and k != "pad_token"
             ):
                 lora_modules_to_save = ", ".join(
                     [f"`{x}`" for x in lora_modules_to_save]
@@ -396,9 +398,11 @@ class ModelLoader:
         ):
             has_remote_code = (
                 "auto_map" in self.model_config
-                and self.model_type in self.model_config["auto_map"]
+                and "AutoModelForCausalLM" in self.model_config["auto_map"]
             )
-
+            if has_remote_code and self.cfg.trust_remote_code is False:
+                # if explicitly set in the YAML, we should prefer that, for example if explicitly disabled
+                has_remote_code = self.cfg.trust_remote_code
             patch_for_multipack(
                 self.cfg.model_config_type,
                 model_name=self.cfg.base_model,
@@ -567,7 +571,8 @@ class ModelLoader:
             )
 
             max_memory = {}
-            for i in range(torch.cuda.device_count()):
+            num_device = get_device_count()
+            for i in range(num_device):
                 max_memory[i] = gpu_memory_limit
             max_memory["cpu"] = "256GiB"  # something sufficiently large to fit anything
 
@@ -592,8 +597,11 @@ class ModelLoader:
         self.model_kwargs["device_map"] = device_map
         self.model_kwargs["torch_dtype"] = self.cfg.torch_dtype
 
-        if torch.backends.mps.is_available():
+        cur_device = get_device_type()
+        if "mps" in str(cur_device):
             self.model_kwargs["device_map"] = "mps:0"
+        elif "npu" in str(cur_device):
+            self.model_kwargs["device_map"] = "npu:0"
 
         # TODO can we put the reference model on it's own gpu? I think we have to move logits around to calculate loss
         # if cfg.rl:
@@ -645,9 +653,7 @@ class ModelLoader:
                 self.model_kwargs["quantization_config"] = BitsAndBytesConfig(
                     **self.model_config.quantization_config
                 )
-        elif self.cfg.adapter == "qlora" and (
-            "load_in_4bit" in self.model_kwargs and self.model_kwargs["load_in_4bit"]
-        ):
+        elif self.cfg.adapter == "qlora" and self.model_kwargs["load_in_4bit"]:
             bnb_config = {
                 "load_in_4bit": True,
                 "llm_int8_threshold": 6.0,
@@ -670,9 +676,7 @@ class ModelLoader:
             self.model_kwargs["quantization_config"] = BitsAndBytesConfig(
                 **bnb_config,
             )
-        elif self.cfg.adapter == "lora" and (
-            "load_in_8bit" in self.model_kwargs and self.model_kwargs["load_in_8bit"]
-        ):
+        elif self.cfg.adapter == "lora" and self.model_kwargs["load_in_8bit"]:
             bnb_config = {
                 "load_in_8bit": True,
             }
@@ -685,10 +689,8 @@ class ModelLoader:
 
         # no longer needed per https://github.com/huggingface/transformers/pull/26610
         if "quantization_config" in self.model_kwargs or self.cfg.gptq:
-            if "load_in_8bit" in self.model_kwargs:
-                del self.model_kwargs["load_in_8bit"]
-            if "load_in_4bit" in self.model_kwargs:
-                del self.model_kwargs["load_in_4bit"]
+            self.model_kwargs.pop("load_in_8bit", None)
+            self.model_kwargs.pop("load_in_4bit", None)
 
     def set_attention_config(self) -> None:
         """
@@ -706,6 +708,7 @@ class ModelLoader:
             self.model_config._attn_implementation = (  # pylint: disable=protected-access
                 "sdpa"
             )
+            monkeypatch_sdp_w_sage_attention()
         elif self.cfg.eager_attention:
             self.model_kwargs["attn_implementation"] = "eager"
             self.model_config._attn_implementation = (  # pylint: disable=protected-access
@@ -973,17 +976,10 @@ class ModelLoader:
         if is_deepspeed_zero3_enabled():
             skip_prepare_model_for_kbit_training = True
 
-        is_load_in_8bit = (
-            "load_in_8bit" in self.model_kwargs and self.model_kwargs["load_in_8bit"]
-        )
-        is_load_in_4bit = (
-            "load_in_4bit" in self.model_kwargs and self.model_kwargs["load_in_4bit"]
-        )
-
         if (
             not skip_prepare_model_for_kbit_training
             and self.cfg.adapter in ["lora", "qlora"]
-            and (is_load_in_8bit or is_load_in_4bit)
+            and (self.cfg.load_in_8bit or self.cfg.load_in_4bit)
         ):
             LOG.info("converting PEFT model w/ prepare_model_for_kbit_training")
             self.model = prepare_model_for_kbit_training(
@@ -1060,7 +1056,11 @@ class ModelLoader:
         self.ajust_model_config()
 
         # log device memory usage
-        if hasattr(self.model, "device") and self.model.device.type in ("cuda", "mps"):
+        if hasattr(self.model, "device") and self.model.device.type in (
+            "cuda",
+            "mps",
+            "npu",
+        ):
             log_gpu_memory_usage(LOG, "after model load", self.model.device)
 
         # make sure these are fp32 per Ramesh et al. (2021)
@@ -1121,22 +1121,16 @@ class ModelLoader:
         # ---------------------------------------------------------
         #  put model to accelerator
         # ---------------------------------------------------------
-        is_load_in_8bit = (
-            "load_in_8bit" in self.model_kwargs and self.model_kwargs["load_in_8bit"]
-        )
-        is_load_in_4bit = (
-            "load_in_4bit" in self.model_kwargs and self.model_kwargs["load_in_4bit"]
-        )
         if (
             self.cfg.ddp
-            and not is_load_in_8bit
-            and not (self.cfg.rl and is_load_in_4bit)
+            and not self.cfg.load_in_8bit
+            and not (self.cfg.rl and self.cfg.load_in_4bit)
             and not skip_move_to_device
         ):
             # TODO revaldate this conditional
-            self.model.to(f"cuda:{self.cfg.local_rank}")
+            self.model.to(f"{str(get_device_type())}:{self.cfg.local_rank}")
 
-        if torch.cuda.device_count() > 1 and int(os.getenv("WORLD_SIZE", "1")) == 1:
+        if get_device_count() > 1 and int(os.getenv("WORLD_SIZE", "1")) == 1:
             setattr(self.model, "is_parallelizable", True)
             setattr(self.model, "model_parallel", True)
 

src/axolotl/utils/optimizers/adopt.py

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

src/axolotl/utils/tokenization.py

@@ -1,8 +1,6 @@
 """Module for tokenization utilities"""
 
 import logging
-import re
-from typing import Dict, List
 
 from termcolor import colored
 
@@ -68,90 +66,47 @@ def process_tokens_for_rl_debug(tokens, color, tokenizer, text_only):
 
 
 def check_rl_example_labels(example, tokenizer, text_only=False):
-    field_prompt, field_chosen, field_rejected = "prompt", "chosen", "rejected"
+    field_prompt, field_chosen, field_rejected, field_completion = (
+        "prompt",
+        "chosen",
+        "rejected",
+        "completion",
+    )
 
     input_tokens = example[field_prompt]
-    labels_chosen, labels_rejected = example[field_chosen], example[field_rejected]
+
+    labels_chosen = example.get(field_chosen)
+    labels_rejected = example.get(field_rejected)
+    labels_completion = example.get(field_completion)
+
+    # Create a delimiter based on text_only flag
+    delimiter = "" if text_only else " "
 
     # Process and color each type of token
     colored_tokens = process_tokens_for_rl_debug(
         input_tokens, "yellow", tokenizer, text_only
     )
-    colored_chosens = process_tokens_for_rl_debug(
-        labels_chosen, "green", tokenizer, text_only
-    )
-    colored_rejecteds = process_tokens_for_rl_debug(
-        labels_rejected, "red", tokenizer, text_only
-    )
 
-    # Create a delimiter based on text_only flag
-    delimiter = "" if text_only else " "
+    # Process tokens
+    if labels_completion is None:
+        colored_chosens = process_tokens_for_rl_debug(
+            labels_chosen, "green", tokenizer, text_only
+        )
+        colored_rejecteds = process_tokens_for_rl_debug(
+            labels_rejected, "red", tokenizer, text_only
+        )
+    else:
+        colored_completion = process_tokens_for_rl_debug(
+            labels_completion, "green", tokenizer, text_only
+        )
 
     # Logging information
     LOG.info(f"INPUT PROMPT: {delimiter.join(colored_tokens)}\n\n")
-    LOG.info(f"CHOSEN RESPONSE: {delimiter.join(colored_chosens)}\n\n")
-    LOG.info(f"REJECTED RESPONSE: {delimiter.join(colored_rejecteds)}\n\n\n")
+
+    if labels_completion is None:
+        LOG.info(f"CHOSEN RESPONSE: {delimiter.join(colored_chosens)}\n\n")
+        LOG.info(f"REJECTED RESPONSE: {delimiter.join(colored_rejecteds)}\n\n\n")
+    else:
+        LOG.info(f"COMPLETION RESPONSE: {delimiter.join(colored_completion)}\n\n\n")
 
     return delimiter.join(colored_tokens)
-
-
-GLAIVE_ROLES = ["USER", "ASSISTANT", "FUNCTION RESPONSE"]
-GLAIVE_TO_SHAREGPT_ROLE = {
-    "SYSTEM": "system",
-    "USER": "human",
-    "ASSISTANT": "gpt",
-    "FUNCTION RESPONSE": "tool",
-}
-
-GLAIVE_MSG_REGEX = re.compile(rf"({'|'.join(GLAIVE_ROLES)}): ")
-
-
-def chatml_to_conversation(row: Dict[str, str]) -> List[Dict[str, str]]:
-    """
-    Converts a ChatML formatted row to a list of messages in ShareGPT format.
-    Initially based off https://github.com/lilacai/lilac/blob/main/notebooks/GlaiveToShareGPT.ipynb.
-    """
-
-    system_prompt = row.get("system")
-    if system_prompt:
-        system_prompt = system_prompt.removeprefix("SYSTEM: ")
-
-    chat_str = row["chat"]
-    chat_msgs = [s.strip() for s in GLAIVE_MSG_REGEX.split(chat_str) if s]
-
-    chat_msg_dicts = [
-        {"from": GLAIVE_TO_SHAREGPT_ROLE[role], "value": value}
-        for role, value in zip(chat_msgs[::2], chat_msgs[1::2])
-    ]
-
-    if system_prompt:
-        chat_msg_dicts = [
-            {"from": GLAIVE_TO_SHAREGPT_ROLE["SYSTEM"], "value": system_prompt}
-        ] + chat_msg_dicts
-
-    return chat_msg_dicts
-
-
-def merge_consecutive_messages(messages):
-    """
-    Merge consecutive messages from the same sender into a single message.
-    This can be useful with datasets that contain multiple consecutive tool calls.
-    """
-
-    merged_messages = []
-    current_from = None
-    current_message = ""
-
-    for msg in messages:
-        if current_from == msg["from"]:
-            current_message += msg["value"]
-        else:
-            if current_from is not None:
-                merged_messages.append({"from": current_from, "value": current_message})
-            current_from = msg["from"]
-            current_message = msg["value"]
-
-    if current_from is not None:
-        merged_messages.append({"from": current_from, "value": current_message})
-
-    return merged_messages

src/axolotl/utils/trainer.py

@@ -17,6 +17,7 @@ from transformers.utils import is_torch_bf16_gpu_available
 
 from axolotl.core.trainer_builder import HFCausalTrainerBuilder, HFRLTrainerBuilder
 from axolotl.utils.distributed import reduce_and_broadcast
+from axolotl.utils.environment import check_cuda_p2p_ib_support
 from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
 
 LOG = get_logger("axolotl")
@@ -184,11 +185,10 @@ def process_datasets_for_packing(cfg, train_dataset, eval_dataset):
         min_sequence_len=cfg.min_sample_len or 2,
     )
 
-    if cfg.is_preprocess:
-        min_input_len = np.min(get_dataset_lengths(train_dataset))
-        LOG.debug(f"min_input_len: {min_input_len}", main_process_only=True)
-        max_input_len = np.max(get_dataset_lengths(train_dataset))
-        LOG.debug(f"max_input_len: {max_input_len}", main_process_only=True)
+    min_input_len = np.min(get_dataset_lengths(train_dataset))
+    LOG.debug(f"min_input_len: {min_input_len}", main_process_only=True)
+    max_input_len = np.max(get_dataset_lengths(train_dataset))
+    LOG.debug(f"max_input_len: {max_input_len}", main_process_only=True)
 
     if cfg.model_config_type == "mamba":
         LOG.info("dropping attention_mask column")
@@ -203,37 +203,59 @@ def process_datasets_for_packing(cfg, train_dataset, eval_dataset):
         if eval_dataset and "token_type_ids" in eval_dataset.column_names:
             eval_dataset = eval_dataset.remove_columns("token_type_ids")
 
+    prior_len = len(train_dataset)
     train_dataset = train_dataset.filter(
         drop_long,
         num_proc=cfg.dataset_processes,
         load_from_cache_file=not cfg.is_preprocess,
         desc="Dropping Long Sequences",
     )
+    dropped = prior_len - len(train_dataset)
+    if dropped:
+        LOG.warning(f"Dropped {dropped} long samples from train dataset")
+
     if eval_dataset:
+        prior_len = len(eval_dataset)
         eval_dataset = eval_dataset.filter(
             drop_long,
             num_proc=cfg.dataset_processes,
             load_from_cache_file=not cfg.is_preprocess,
             desc="Dropping Long Sequences",
         )
+        dropped = prior_len - len(eval_dataset)
+        if dropped:
+            LOG.warning(f"Dropped {dropped} long samples from eval dataset")
 
     # drop samples with where the number of elements with labels not equal to -100 is zero
     def drop_no_trainable_tokens(sample):
         return np.sum(np.array(sample["labels"]) != -100) > 0
 
+    prior_len = len(train_dataset)
     train_dataset = train_dataset.filter(
         drop_no_trainable_tokens,
         num_proc=cfg.dataset_processes,
         load_from_cache_file=not cfg.is_preprocess,
         desc="Drop Samples with Zero Trainable Tokens",
     )
+    dropped = prior_len - len(train_dataset)
+    if dropped:
+        LOG.warning(
+            f"Dropped {dropped} samples with no trainable tokens from train dataset"
+        )
+
     if eval_dataset:
+        prior_len = len(eval_dataset)
         eval_dataset = eval_dataset.filter(
             drop_no_trainable_tokens,
             num_proc=cfg.dataset_processes,
             load_from_cache_file=not cfg.is_preprocess,
             desc="Drop Samples with Zero Trainable Tokens",
         )
+        dropped = prior_len - len(eval_dataset)
+        if dropped:
+            LOG.warning(
+                f"Dropped {dropped} samples with no trainable tokens from eval dataset"
+            )
 
     if cfg.group_by_length:
         train_dataset = train_dataset.map(
@@ -461,6 +483,9 @@ def setup_fsdp_envs(cfg):
 
 
 def prepare_optim_env(cfg):
+    if not check_cuda_p2p_ib_support():
+        if os.getenv("NCCL_P2P_DISABLE") is None:
+            os.environ["NCCL_P2P_DISABLE"] = "1"
     if cfg.fsdp:
         setup_fsdp_envs(cfg)
     elif cfg.deepspeed:
@@ -490,7 +515,7 @@ def prepare_opinionated_env(cfg):
 def setup_trainer(
     cfg, train_dataset, eval_dataset, model, tokenizer, processor, total_num_steps
 ):
-    if cfg.rl in ["dpo", "ipo", "orpo", "kto", "simpo"]:
+    if cfg.rl in ("dpo", "ipo", "orpo", "kto", "simpo"):
         trainer_builder = HFRLTrainerBuilder(cfg, model[0], tokenizer, processor)
         trainer_builder.model_ref = model[1]
         trainer_builder.peft_config = model[2]

tests/e2e/conftest.py

@@ -0,0 +1,35 @@
+"""
+shared pytest fixtures
+"""
+import shutil
+import tempfile
+
+import pytest
+from huggingface_hub import snapshot_download
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_smollm2_135m_model():
+    # download the model
+    snapshot_download("HuggingFaceTB/SmolLM2-135M")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_tatsu_lab_alpaca_dataset():
+    # download the model
+    snapshot_download("tatsu-lab/alpaca", repo_type="dataset")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_mhenrichsen_alpaca_2k_dataset():
+    # download the model
+    snapshot_download("mhenrichsen/alpaca_2k_test", repo_type="dataset")
+
+
+@pytest.fixture
+def temp_dir():
+    # Create a temporary directory
+    _temp_dir = tempfile.mkdtemp()
+    yield _temp_dir
+    # Clean up the directory after the test
+    shutil.rmtree(_temp_dir)

tests/e2e/integrations/liger.py

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

tests/e2e/multigpu/test_eval.py

@@ -3,28 +3,25 @@ E2E tests for multigpu eval
 """
 import logging
 import os
-import unittest
 from pathlib import Path
 
 import yaml
 from accelerate.test_utils import execute_subprocess_async
+from transformers.testing_utils import get_torch_dist_unique_port
 
 from axolotl.utils.dict import DictDefault
 
-from ..utils import with_temp_dir
-
 LOG = logging.getLogger("axolotl.tests.e2e.multigpu")
 os.environ["WANDB_DISABLED"] = "true"
 
 AXOLOTL_ROOT = Path(__file__).parent.parent.parent.parent
 
 
-class TestMultiGPUEval(unittest.TestCase):
+class TestMultiGPUEval:
     """
     Test case for MultiGPU Eval Sample Packing
     """
 
-    @with_temp_dir
     def test_eval_sample_packing(self, temp_dir):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
@@ -83,13 +80,14 @@ class TestMultiGPUEval(unittest.TestCase):
                 "launch",
                 "--num-processes",
                 "2",
+                "--main_process_port",
+                f"{get_torch_dist_unique_port()}",
                 "-m",
                 "axolotl.cli.train",
                 str(Path(temp_dir) / "config.yaml"),
             ]
         )
 
-    @with_temp_dir
     def test_eval(self, temp_dir):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
@@ -148,6 +146,8 @@ class TestMultiGPUEval(unittest.TestCase):
                 "launch",
                 "--num-processes",
                 "2",
+                "--main_process_port",
+                f"{get_torch_dist_unique_port()}",
                 "-m",
                 "axolotl.cli.train",
                 str(Path(temp_dir) / "config.yaml"),

tests/e2e/multigpu/test_llama.py

@@ -4,17 +4,17 @@ E2E tests for multigpu lora tinyllama
 
 import logging
 import os
-import unittest
 from pathlib import Path
 
 import pytest
 import yaml
 from accelerate.test_utils import execute_subprocess_async
 from huggingface_hub import snapshot_download
+from transformers.testing_utils import get_torch_dist_unique_port
 
 from axolotl.utils.dict import DictDefault
 
-from ..utils import with_temp_dir
+from ..utils import is_hopper
 
 LOG = logging.getLogger("axolotl.tests.e2e.multigpu")
 os.environ["WANDB_DISABLED"] = "true"
@@ -25,21 +25,19 @@ AXOLOTL_ROOT = Path(__file__).parent.parent.parent.parent
 @pytest.fixture(scope="session", autouse=True)
 def download_model():
     # download the model
-    snapshot_download("TinyLlama/TinyLlama_v1.1")
+    snapshot_download("HuggingFaceTB/SmolLM2-135M")
 
 
-class TestMultiGPULlama(unittest.TestCase):
+class TestMultiGPULlama:
     """
     Test case for Llama models using LoRA
     """
 
-    @with_temp_dir
     def test_lora_ddp(self, temp_dir):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "TinyLlama/TinyLlama_v1.1",
-                "tokenizer_type": "LlamaTokenizer",
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
                 "sequence_len": 2048,
                 "adapter": "lora",
                 "lora_r": 8,
@@ -48,9 +46,7 @@ class TestMultiGPULlama(unittest.TestCase):
                 "lora_target_linear": True,
                 "val_set_size": 0.05,
                 "special_tokens": {
-                    "unk_token": "<unk>",
-                    "bos_token": "<s>",
-                    "eos_token": "</s>",
+                    "pad_token": "<|endoftext|>",
                 },
                 "datasets": [
                     {
@@ -59,7 +55,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 100,
+                "max_steps": 15,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -81,19 +77,23 @@ class TestMultiGPULlama(unittest.TestCase):
                 "launch",
                 "--num-processes",
                 "2",
+                "--main_process_port",
+                f"{get_torch_dist_unique_port()}",
                 "-m",
                 "axolotl.cli.train",
                 str(Path(temp_dir) / "config.yaml"),
             ]
         )
 
-    @with_temp_dir
-    def test_lora_ddp_packed(self, temp_dir):
+    @pytest.mark.parametrize(
+        "gradient_accumulation_steps",
+        [1, 4],
+    )
+    def test_lora_ddp_packed(self, temp_dir, gradient_accumulation_steps):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "TinyLlama/TinyLlama_v1.1",
-                "tokenizer_type": "LlamaTokenizer",
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
                 "sequence_len": 2048,
                 "sample_packing": True,
                 "eval_sample_packing": False,
@@ -105,9 +105,7 @@ class TestMultiGPULlama(unittest.TestCase):
                 "lora_target_linear": True,
                 "val_set_size": 0.05,
                 "special_tokens": {
-                    "unk_token": "<unk>",
-                    "bos_token": "<s>",
-                    "eos_token": "</s>",
+                    "pad_token": "<|endoftext|>",
                 },
                 "datasets": [
                     {
@@ -116,9 +114,9 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 50,
+                "max_steps": 15,
                 "micro_batch_size": 4,
-                "gradient_accumulation_steps": 4,
+                "gradient_accumulation_steps": gradient_accumulation_steps,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_8bit",
@@ -138,25 +136,166 @@ class TestMultiGPULlama(unittest.TestCase):
                 "launch",
                 "--num-processes",
                 "2",
+                "--main_process_port",
+                f"{get_torch_dist_unique_port()}",
                 "-m",
                 "axolotl.cli.train",
                 str(Path(temp_dir) / "config.yaml"),
             ]
         )
 
-    @with_temp_dir
-    def test_fsdp(self, temp_dir):
+    @pytest.mark.skipif(is_hopper(), reason="h100 doesn't support 8-bit lora")
+    def test_dpo_lora_ddp(self, temp_dir):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "TinyLlama/TinyLlama_v1.1",
-                "tokenizer_type": "LlamaTokenizer",
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
                 "sequence_len": 2048,
+                "sample_packing": False,
+                "eval_sample_packing": False,
+                "pad_to_sequence_len": True,
+                "load_in_8bit": True,
+                "adapter": "lora",
+                "lora_r": 8,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
                 "val_set_size": 0.05,
                 "special_tokens": {
-                    "unk_token": "<unk>",
-                    "bos_token": "<s>",
-                    "eos_token": "</s>",
+                    "pad_token": "<|endoftext|>",
+                },
+                "rl": "dpo",
+                "chat_template": "chatml",
+                "datasets": [
+                    {
+                        "path": "fozziethebeat/alpaca_messages_2k_dpo_test",
+                        "type": "chat_template.default",
+                        "field_messages": "conversation",
+                        "field_chosen": "chosen",
+                        "field_rejected": "rejected",
+                        "message_field_role": "role",
+                        "message_field_content": "content",
+                        "roles": {
+                            "system": ["system"],
+                            "user": ["user"],
+                            "assistant": ["assistant"],
+                        },
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 15,
+                "micro_batch_size": 4,
+                "gradient_accumulation_steps": 4,
+                "output_dir": temp_dir,
+                "warmup_steps": 0,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_8bit",
+                "lr_scheduler": "cosine",
+                "flash_attention": True,
+            }
+        )
+
+        # write cfg to yaml file
+        Path(temp_dir).mkdir(parents=True, exist_ok=True)
+        with open(Path(temp_dir) / "config.yaml", "w", encoding="utf-8") as fout:
+            fout.write(yaml.dump(cfg.to_dict(), Dumper=yaml.Dumper))
+
+        execute_subprocess_async(
+            [
+                "accelerate",
+                "launch",
+                "--num-processes",
+                "2",
+                "--main_process_port",
+                f"{get_torch_dist_unique_port()}",
+                "-m",
+                "axolotl.cli.train",
+                str(Path(temp_dir) / "config.yaml"),
+            ]
+        )
+
+    def test_dpo_qlora_ddp(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "sequence_len": 2048,
+                "sample_packing": False,
+                "eval_sample_packing": False,
+                "pad_to_sequence_len": True,
+                "load_in_4bit": True,
+                "adapter": "qlora",
+                "lora_r": 8,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "val_set_size": 0.05,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                },
+                "rl": "dpo",
+                "chat_template": "chatml",
+                "datasets": [
+                    {
+                        "path": "fozziethebeat/alpaca_messages_2k_dpo_test",
+                        "type": "chat_template.default",
+                        "field_messages": "conversation",
+                        "field_chosen": "chosen",
+                        "field_rejected": "rejected",
+                        "message_field_role": "role",
+                        "message_field_content": "content",
+                        "roles": {
+                            "system": ["system"],
+                            "user": ["user"],
+                            "assistant": ["assistant"],
+                        },
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 15,
+                "micro_batch_size": 4,
+                "gradient_accumulation_steps": 4,
+                "output_dir": temp_dir,
+                "warmup_steps": 0,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_8bit",
+                "lr_scheduler": "cosine",
+                "flash_attention": True,
+            }
+        )
+
+        # write cfg to yaml file
+        Path(temp_dir).mkdir(parents=True, exist_ok=True)
+        with open(Path(temp_dir) / "config.yaml", "w", encoding="utf-8") as fout:
+            fout.write(yaml.dump(cfg.to_dict(), Dumper=yaml.Dumper))
+
+        execute_subprocess_async(
+            [
+                "accelerate",
+                "launch",
+                "--num-processes",
+                "2",
+                "--main_process_port",
+                f"{get_torch_dist_unique_port()}",
+                "-m",
+                "axolotl.cli.train",
+                str(Path(temp_dir) / "config.yaml"),
+            ]
+        )
+
+    @pytest.mark.parametrize(
+        "gradient_accumulation_steps",
+        [1, 4],
+    )
+    def test_fsdp(self, temp_dir, gradient_accumulation_steps):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "sequence_len": 2048,
+                "val_set_size": 0.01,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
                 },
                 "datasets": [
                     {
@@ -165,9 +304,9 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 100,
+                "max_steps": 10,
                 "micro_batch_size": 4,
-                "gradient_accumulation_steps": 4,
+                "gradient_accumulation_steps": gradient_accumulation_steps,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch",
@@ -184,7 +323,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     "fsdp_use_orig_params": False,
                     "fsdp_cpu_ram_efficient_loading": False,
                     "fsdp_transformer_layer_cls_to_wrap": "LlamaDecoderLayer",
-                    "fsdp_state_dict_type": "SHARDED_STATE_DICT",
+                    "fsdp_state_dict_type": "FULL_STATE_DICT",
                     "fsdp_auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
                 },
             }
@@ -201,28 +340,29 @@ class TestMultiGPULlama(unittest.TestCase):
                 "launch",
                 "--num-processes",
                 "2",
+                "--main_process_port",
+                f"{get_torch_dist_unique_port()}",
                 "-m",
                 "axolotl.cli.train",
                 str(Path(temp_dir) / "config.yaml"),
             ]
         )
 
-    @with_temp_dir
-    def test_fsdp_packed(self, temp_dir):
+    @pytest.mark.parametrize(
+        "fsdp_state_dict_type",
+        ["FULL_STATE_DICT", "SHARDED_STATE_DICT"],
+    )
+    def test_fsdp_packed(self, temp_dir, fsdp_state_dict_type):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "TinyLlama/TinyLlama_v1.1",
-                "tokenizer_type": "LlamaTokenizer",
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
                 "sample_packing": True,
-                "eval_sample_packing": False,
                 "pad_to_sequence_len": True,
                 "sequence_len": 2048,
                 "val_set_size": 0.05,
                 "special_tokens": {
-                    "unk_token": "<unk>",
-                    "bos_token": "<s>",
-                    "eos_token": "</s>",
+                    "pad_token": "<|endoftext|>",
                 },
                 "datasets": [
                     {
@@ -231,7 +371,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 100,
+                "max_steps": 15,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -250,7 +390,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     "fsdp_use_orig_params": False,
                     "fsdp_cpu_ram_efficient_loading": False,
                     "fsdp_transformer_layer_cls_to_wrap": "LlamaDecoderLayer",
-                    "fsdp_state_dict_type": "SHARDED_STATE_DICT",
+                    "fsdp_state_dict_type": fsdp_state_dict_type,
                     "fsdp_auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
                 },
             }
@@ -267,37 +407,37 @@ class TestMultiGPULlama(unittest.TestCase):
                 "launch",
                 "--num-processes",
                 "2",
+                "--main_process_port",
+                f"{get_torch_dist_unique_port()}",
                 "-m",
                 "axolotl.cli.train",
                 str(Path(temp_dir) / "config.yaml"),
             ]
         )
 
-    @pytest.mark.skip("disabled due to upstream issue")
-    @with_temp_dir
     def test_fsdp_qlora_prequant_packed(self, temp_dir):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "axolotl-ai-co/TinyLlama_v1.1-bnb-nf4-bf16",
-                "tokenizer_type": "AutoTokenizer",
+                "base_model": "axolotl-ai-co/SmolLM2-135M-bnb-nf4-bf16",
                 "adapter": "qlora",
+                "mean_resizing_embeddings": True,
                 "load_in_4bit": True,
                 "lora_r": 8,
                 "lora_alpha": 16,
                 "lora_dropout": 0.05,
                 "lora_target_linear": True,
-                "lora_modules_to_save": [
-                    "embed_tokens",
-                    "lm_head",
-                ],
+                # "lora_modules_to_save": [
+                #     "embed_tokens",
+                #     "lm_head",
+                # ],
                 "sample_packing": True,
                 "eval_sample_packing": False,
                 "pad_to_sequence_len": True,
                 "sequence_len": 2048,
                 "val_set_size": 0.05,
                 "special_tokens": {
-                    "pad_token": "<|end_of_text|>",
+                    "pad_token": "<|endoftext|>",
                 },
                 "datasets": [
                     {
@@ -307,7 +447,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 100,
+                "max_steps": 15,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -343,28 +483,29 @@ class TestMultiGPULlama(unittest.TestCase):
                 "launch",
                 "--num-processes",
                 "2",
+                "--main_process_port",
+                f"{get_torch_dist_unique_port()}",
                 "-m",
                 "axolotl.cli.train",
                 str(Path(temp_dir) / "config.yaml"),
             ]
         )
 
-    @with_temp_dir
-    def test_ds_zero3_packed(self, temp_dir):
+    @pytest.mark.parametrize(
+        "gradient_accumulation_steps",
+        [1, 4],
+    )
+    def test_ds_zero3_packed(self, temp_dir, gradient_accumulation_steps):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "TinyLlama/TinyLlama_v1.1",
-                "tokenizer_type": "LlamaTokenizer",
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
                 "sample_packing": True,
-                "eval_sample_packing": False,
                 "pad_to_sequence_len": True,
                 "sequence_len": 2048,
                 "val_set_size": 0.05,
                 "special_tokens": {
-                    "unk_token": "<unk>",
-                    "bos_token": "<s>",
-                    "eos_token": "</s>",
+                    "pad_token": "<|endoftext|>",
                 },
                 "datasets": [
                     {
@@ -373,9 +514,9 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 100,
+                "max_steps": 15,
                 "micro_batch_size": 4,
-                "gradient_accumulation_steps": 4,
+                "gradient_accumulation_steps": gradient_accumulation_steps,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch",
@@ -396,19 +537,19 @@ class TestMultiGPULlama(unittest.TestCase):
                 "launch",
                 "--num-processes",
                 "2",
+                "--main_process_port",
+                f"{get_torch_dist_unique_port()}",
                 "-m",
                 "axolotl.cli.train",
                 str(Path(temp_dir) / "config.yaml"),
             ]
         )
 
-    @with_temp_dir
     def test_ds_zero3_qlora_packed(self, temp_dir):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "TinyLlama/TinyLlama_v1.1",
-                "tokenizer_type": "LlamaTokenizer",
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
                 "load_in_4bit": True,
                 "adapter": "qlora",
                 "lora_r": 8,
@@ -421,9 +562,7 @@ class TestMultiGPULlama(unittest.TestCase):
                 "sequence_len": 2048,
                 "val_set_size": 0.05,
                 "special_tokens": {
-                    "unk_token": "<unk>",
-                    "bos_token": "<s>",
-                    "eos_token": "</s>",
+                    "pad_token": "<|endoftext|>",
                 },
                 "datasets": [
                     {
@@ -432,7 +571,7 @@ class TestMultiGPULlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 100,
+                "max_steps": 15,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -455,6 +594,8 @@ class TestMultiGPULlama(unittest.TestCase):
                 "launch",
                 "--num-processes",
                 "2",
+                "--main_process_port",
+                f"{get_torch_dist_unique_port()}",
                 "-m",
                 "axolotl.cli.train",
                 str(Path(temp_dir) / "config.yaml"),

tests/e2e/multigpu/test_qwen2.py

@@ -4,31 +4,30 @@ E2E tests for multigpu qwen2
 
 import logging
 import os
-import unittest
 from pathlib import Path
 
+import pytest
 import yaml
 from accelerate.test_utils import execute_subprocess_async
+from transformers.testing_utils import get_torch_dist_unique_port
 
 from axolotl.utils.dict import DictDefault
 
-from ..utils import with_temp_dir
-
 LOG = logging.getLogger("axolotl.tests.e2e.multigpu")
 os.environ["WANDB_DISABLED"] = "true"
 
 
-class TestMultiGPUQwen2(unittest.TestCase):
+class TestMultiGPUQwen2:
     """
     Test case for Llama models using LoRA
     """
 
-    @with_temp_dir
-    def test_qlora_fsdp_dpo(self, temp_dir):
+    @pytest.mark.parametrize("base_model", ["Qwen/Qwen2-0.5B", "Qwen/Qwen2.5-0.5B"])
+    def test_qlora_fsdp_dpo(self, base_model, temp_dir):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2-1.5B",
+                "base_model": base_model,
                 "load_in_4bit": True,
                 "rl": "dpo",
                 "chat_template": "chatml",
@@ -47,9 +46,9 @@ class TestMultiGPUQwen2(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 100,
+                "max_steps": 5,
                 "warmup_steps": 20,
-                "micro_batch_size": 4,
+                "micro_batch_size": 2,
                 "gradient_accumulation_steps": 2,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
@@ -91,6 +90,8 @@ class TestMultiGPUQwen2(unittest.TestCase):
                 "launch",
                 "--num-processes",
                 "2",
+                "--main_process_port",
+                f"{get_torch_dist_unique_port()}",
                 "-m",
                 "axolotl.cli.train",
                 str(Path(temp_dir) / "config.yaml"),

tests/e2e/patched/test_4d_multipack_llama.py

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

tests/e2e/patched/test_fa_xentropy.py

@@ -66,6 +66,8 @@ class TestFAXentropyLlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
+                "max_steps": 10,
+                "save_steps": 10,
                 "micro_batch_size": 8,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,

tests/e2e/patched/test_lora_llama_multipack.py

@@ -56,6 +56,8 @@ class TestLoraLlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
+                "max_steps": 20,
+                "save_steps": 10,
                 "micro_batch_size": 8,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
@@ -109,6 +111,7 @@ class TestLoraLlama(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
+                "max_steps": 20,
                 "save_steps": 0.5,
                 "micro_batch_size": 8,
                 "gradient_accumulation_steps": 1,

tests/e2e/test_dpo.py

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

tests/e2e/test_llama.py

@@ -0,0 +1,66 @@
+"""
+E2E tests for llama
+"""
+
+import logging
+import os
+import unittest
+from pathlib import Path
+
+from axolotl.cli import load_datasets
+from axolotl.common.cli import TrainerCliArgs
+from axolotl.train import train
+from axolotl.utils.config import normalize_config
+from axolotl.utils.dict import DictDefault
+
+from .utils import with_temp_dir
+
+LOG = logging.getLogger("axolotl.tests.e2e")
+os.environ["WANDB_DISABLED"] = "true"
+
+
+class TestLlama(unittest.TestCase):
+    """
+    Test case for Llama models
+    """
+
+    @with_temp_dir
+    def test_fft_trust_remote_code(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "JackFram/llama-68m",
+                "tokenizer_type": "LlamaTokenizer",
+                "trust_remote_code": True,
+                "sequence_len": 512,
+                "val_set_size": 0.1,
+                "special_tokens": {
+                    "unk_token": "<unk>",
+                    "bos_token": "<s>",
+                    "eos_token": "</s>",
+                },
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "micro_batch_size": 8,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_bnb_8bit",
+                "lr_scheduler": "cosine",
+                "flash_attention": True,
+                "sample_packing": True,
+                "bf16": True,
+                "save_safetensors": True,
+            }
+        )
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+        assert (Path(temp_dir) / "model.safetensors").exists()