add output for train loss in assertian err

* adding codecov reporting

* update codecov-action to v5

* fix

---------

Co-authored-by: Dan Saunders <dan@axolotl.ai>

.coveragerc

@@ -0,0 +1,14 @@
+[run]
+source = axolotl
+omit =
+    */tests/*
+    setup.py
+
+[report]
+exclude_lines =
+    pragma: no cover
+    def __repr__
+    raise NotImplementedError
+    if __name__ == .__main__.:
+    pass
+    raise ImportError

.github/workflows/base.yml

@@ -40,12 +40,24 @@ jobs:
             python_version: "3.11"
             pytorch: 2.6.0
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+          - cuda: "126"
+            cuda_version: 12.6.3
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.6.0
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
           - cuda: "128"
             cuda_version: 12.8.1
             cudnn_version: ""
             python_version: "3.11"
             pytorch: nightly
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+          - cuda: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: next
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -67,7 +79,7 @@ jobs:
         uses: docker/build-push-action@v4
         with:
           context: .
-          file: ${{ matrix.pytorch == 'nightly' && './docker/Dockerfile-base-nightly' || './docker/Dockerfile-base' }}
+          file: ${{ matrix.pytorch == 'nightly' && './docker/Dockerfile-base-nightly' || matrix.pytorch == 'next' && './docker/Dockerfile-base-next' || './docker/Dockerfile-base' }}
           push: ${{ github.event_name != 'pull_request' }}
           tags: ${{ steps.metadata.outputs.tags }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
           labels: ${{ steps.metadata.outputs.labels }}

.github/workflows/docs.yml

@@ -20,9 +20,12 @@ jobs:
           uses: actions/setup-python@v5
           with:
             python-version: '3.11'
-        - name: install dependencies
+        - name: Install dependencies
           run: |
-            python3 -m pip install jupyter
+            python3 -m pip install jupyter quartodoc
+            python3 -m pip install -e . --no-deps
+        - name: Build autodoc
+          run: quartodoc build
         - name: Publish to GitHub Pages (and render)
           uses: quarto-dev/quarto-actions/publish@v2
           with:

.github/workflows/main.yml

@@ -25,12 +25,12 @@ jobs:
             python_version: "3.11"
             pytorch: 2.5.1
             axolotl_extras: vllm
-            is_latest: true
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
             pytorch: 2.6.0
-            axolotl_extras:
+            axolotl_extras: vllm
+            is_latest: true
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -87,12 +87,12 @@ jobs:
             python_version: "3.11"
             pytorch: 2.5.1
             axolotl_extras:
-            is_latest: true
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
             pytorch: 2.6.0
             axolotl_extras:
+            is_latest: true
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout

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

@@ -24,6 +24,13 @@ jobs:
       fail-fast: false
       matrix:
         include:
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.6.0
+            axolotl_extras: vllm
+            num_gpus: 2
+            nightly_build: "true"
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
@@ -38,14 +45,6 @@ jobs:
             axolotl_extras: vllm
             num_gpus: 2
             nightly_build: "true"
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.6.0
-            # awaiting vllm#12721
-            axolotl_extras:
-            num_gpus: 2
-            nightly_build: "true"
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
     steps:

.github/workflows/precommit-autoupdate.yml

@@ -0,0 +1,49 @@
+name: Pre-commit auto-update
+
+on:
+  schedule:
+    - cron: '0 0 * * 0'  # Run weekly
+  workflow_dispatch:  # Manual kickoff
+
+jobs:
+  auto-update:
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+      pull-requests: write
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+
+      - name: Update pre-commit hooks
+        id: update
+        run: |
+          pip install pre-commit
+          pre-commit autoupdate
+          if [[ -n $(git status --porcelain) ]]; then
+            echo "changes=true" >> $GITHUB_OUTPUT
+            git diff .pre-commit-config.yaml > pre-commit-update.diff
+          fi
+
+      - name: Create Pull Request
+        if: steps.update.outputs.changes == 'true'
+        uses: peter-evans/create-pull-request@v6
+        with:
+          token: ${{ secrets.GITHUB_TOKEN }}
+          branch: update/pre-commit-hooks
+          delete-branch: true
+          title: "chore: update pre-commit hooks"
+          commit-message: "chore: update pre-commit hooks"
+          body: |
+            Automated PR to update pre-commit hooks to their latest versions.
+
+            <details>
+            <summary>Changes:</summary>
+
+            ```diff
+            ${{ steps.update.outputs.diff }}
+            ```
+            </details>

.github/workflows/tests-nightly.yml

@@ -33,6 +33,15 @@ jobs:
       - name: Check out repository code
         uses: actions/checkout@v4
 
+      - name: Restore HF cache
+        id: hf-cache-restore
+        uses: actions/cache/restore@v4
+        with:
+          path: |
+            /home/runner/.cache/huggingface/hub/datasets--*
+            /home/runner/.cache/huggingface/hub/models--*
+          key: ${{ runner.os }}-hf-hub-cache-v2
+
       - name: Setup Python
         uses: actions/setup-python@v5
         with:
@@ -46,7 +55,7 @@ jobs:
 
       - name: Install PyTorch
         run: |
-          pip3 install torch==${{ matrix.pytorch_version }} --index-url https://download.pytorch.org/whl/cpu
+          pip3 install torch==${{ matrix.pytorch_version }}
 
       - name: Update requirements.txt
         run: |
@@ -58,8 +67,7 @@ jobs:
 
       - name: Install dependencies
         run: |
-          pip3 install --upgrade pip
+          pip3 show torch
-          pip3 install --upgrade packaging==23.2
           pip3 install --no-build-isolation -U -e .
           python scripts/unsloth_install.py | sh
           python scripts/cutcrossentropy_install.py | sh
@@ -73,10 +81,15 @@ jobs:
         run: |
           axolotl --help
 
+      - name: Pre-Download dataset fixture
+        run: |
+          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
+
       - name: Run tests
         run: |
-          pytest -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ tests/
+          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/
-          pytest tests/patched/
+          pytest -v tests/patched/
+          pytest -v tests/cli/
 
       - name: cleanup pip cache
         run: |
@@ -136,4 +149,4 @@ jobs:
           echo "NIGHTLY_BUILD=${{ matrix.nightly_build }}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         run: |
-          modal run cicd.tests
+          modal run cicd.e2e_tests

.github/workflows/tests.yml

@@ -63,7 +63,7 @@ jobs:
           path: |
             /home/runner/.cache/huggingface/hub/datasets--*
             /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ runner.os }}-hf-hub-cache-${{ hashFiles('**/conftest.py') }}
+          key: ${{ runner.os }}-hf-hub-cache-v2
 
       - name: Setup Python
         uses: actions/setup-python@v5
@@ -96,10 +96,22 @@ jobs:
         run: |
           axolotl --help
 
+      - name: Pre-Download dataset fixture
+        run: |
+          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
+
       - name: Run tests
         run: |
-          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ tests/
+          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/ --cov=axolotl --cov-report=xml
-          pytest -v tests/patched/
+          pytest -v tests/patched/ --cov=axolotl --cov-append --cov-report=xml
+          pytest -v tests/cli/ --cov=axolotl --cov-append --cov-report=xml
+
+      - name: Upload coverage to Codecov
+        uses: codecov/codecov-action@v5
+        with:
+          files: ./coverage.xml
+          flags: unittests,pytorch-${{ matrix.pytorch_version }}
+          fail_ci_if_error: false
 
       - name: cleanup pip cache
         run: |
@@ -136,7 +148,7 @@ jobs:
           path: |
             /home/runner/.cache/huggingface/hub/datasets--*
             /home/runner/.cache/huggingface/hub/models--*
-          key: ${{ runner.os }}-hf-hub-cache-${{ hashFiles('**/conftest.py') }}
+          key: ${{ runner.os }}-hf-hub-cache-v2
 
       - name: Setup Python
         uses: actions/setup-python@v5
@@ -170,10 +182,14 @@ jobs:
         run: |
           axolotl --help
 
+      - name: Show HF cache
+        run: huggingface-cli scan-cache
+
       - name: Run tests
         run: |
-          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ tests/
+          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/
           pytest -v tests/patched/
+          pytest -v tests/cli/
 
       - name: cleanup pip cache
         run: |
@@ -199,6 +215,53 @@ jobs:
       fail-fast: false
       matrix:
         include:
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.6.0
+            num_gpus: 1
+            axolotl_extras: vllm
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Install Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+      - name: Install Modal
+        run: |
+          python -m pip install --upgrade pip
+          pip install modal==0.71.8 jinja2
+      - name: Update env vars
+        run: |
+          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
+          echo "PYTORCH_VERSION=${{ 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 "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
+          echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
+      - name: Run tests job on Modal
+        run: |
+          modal run cicd.e2e_tests
+
+  docker-e2e-tests:
+    if: github.repository_owner == 'axolotl-ai-cloud'
+    # this job needs to be run on self-hosted GPU runners...
+    runs-on: [self-hosted, modal]
+    timeout-minutes: 90
+    needs: [pre-commit, pytest, docker-e2e-tests-1st]
+
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.4.1
+            num_gpus: 1
+            axolotl_extras:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
@@ -227,51 +290,4 @@ jobs:
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         run: |
-          modal run cicd.tests
+          modal run cicd.e2e_tests
-
-  docker-e2e-tests:
-    if: github.repository_owner == 'axolotl-ai-cloud'
-    # this job needs to be run on self-hosted GPU runners...
-    runs-on: [self-hosted, modal]
-    timeout-minutes: 90
-    needs: [pre-commit, pytest, docker-e2e-tests-1st]
-
-    strategy:
-      fail-fast: false
-      matrix:
-        include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.4.1
-            num_gpus: 1
-            axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.6.0
-            num_gpus: 1
-            axolotl_extras:
-    steps:
-      - name: Checkout
-        uses: actions/checkout@v4
-      - name: Install Python
-        uses: actions/setup-python@v5
-        with:
-          python-version: "3.11"
-      - name: Install Modal
-        run: |
-          python -m pip install --upgrade pip
-          pip install modal==0.71.8 jinja2
-      - name: Update env vars
-        run: |
-          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
-          echo "PYTORCH_VERSION=${{ 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 "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
-          echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
-      - name: Run tests job on Modal
-        run: |
-          modal run cicd.tests

.gitignore

@@ -181,6 +181,10 @@ prepared-datasets/
 submit.sh
 *.out*
 
+# Quartodoc generated files
+objects.json
+site_libs/
+
 typings/
 out/
 

.isort.cfg

@@ -1,3 +1,4 @@
 [settings]
 profile=black
 known_third_party=wandb,comet_ml
+known_local_folder=src,tests

.pre-commit-config.yaml

@@ -3,7 +3,7 @@ default_language_version:
 
 repos:
 -   repo: https://github.com/pre-commit/pre-commit-hooks
-    rev: v4.4.0
+    rev: v5.0.0
     hooks:
     -   id: check-yaml
     -   id: end-of-file-fixer
@@ -11,23 +11,23 @@ repos:
     -   id: no-commit-to-branch
         args: ['--branch', 'main']
 -   repo: https://github.com/psf/black
-    rev: 23.3.0
+    rev: 25.1.0
     hooks:
     -   id: black
 -   repo: https://github.com/pycqa/isort
-    rev: 5.12.0
+    rev: 6.0.1
     hooks:
       - id: isort
 -   repo: https://github.com/PyCQA/flake8
-    rev: 6.1.0
+    rev: 7.1.2
     hooks:
     - id: flake8
 -   repo: https://github.com/pylint-dev/pylint
-    rev: c8c96d20cde3552a79858c7456bb1483bf83d633
+    rev: v3.3.6
     hooks:
     - id: pylint
 -   repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.3.0
+    rev: v1.15.0
     hooks:
     - id: mypy
       additional_dependencies:
@@ -36,7 +36,7 @@ repos:
             'pydantic>=2.5.3',
         ]
 -   repo: https://github.com/PyCQA/bandit
-    rev: 1.7.5
+    rev: 1.8.3
     hooks:
     -   id: bandit
         args: [

CNAME

@@ -0,0 +1 @@
+docs.axolotl.ai

README.md

@@ -9,6 +9,7 @@
 <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://codecov.io/gh/axolotl-ai-cloud/axolotl"><img src="https://codecov.io/gh/axolotl-ai-cloud/axolotl/branch/main/graph/badge.svg" alt="codecov"></a>
     <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>
     <br/>
     <a href="https://github.com/axolotl-ai-cloud/axolotl/graphs/contributors"><img src="https://img.shields.io/github/contributors-anon/axolotl-ai-cloud/axolotl?color=yellow&style=flat-square" alt="contributors" style="height: 20px;"></a>
@@ -63,7 +64,7 @@ axolotl fetch examples
 axolotl fetch deepspeed_configs  # OPTIONAL
 ```
 
-Other installation approaches are described [here](https://axolotl-ai-cloud.github.io/axolotl/docs/installation.html).
+Other installation approaches are described [here](https://docs.axolotl.ai/docs/installation.html).
 
 ### Your First Fine-tune
 
@@ -78,7 +79,7 @@ axolotl fetch examples --dest path/to/folder
 axolotl train examples/llama-3/lora-1b.yml
 ```
 
-That's it! Check out our [Getting Started Guide](https://axolotl-ai-cloud.github.io/axolotl/docs/getting-started.html) for a more detailed walkthrough.
+That's it! Check out our [Getting Started Guide](https://docs.axolotl.ai/docs/getting-started.html) for a more detailed walkthrough.
 
 ## ✨ Key Features
 
@@ -91,19 +92,20 @@ That's it! Check out our [Getting Started Guide](https://axolotl-ai-cloud.github
 
 ## 📚 Documentation
 
-- [Installation Options](https://axolotl-ai-cloud.github.io/axolotl/docs/installation.html) - Detailed setup instructions for different environments
+- [Installation Options](https://docs.axolotl.ai/docs/installation.html) - Detailed setup instructions for different environments
-- [Configuration Guide](https://axolotl-ai-cloud.github.io/axolotl/docs/config.html) - Full configuration options and examples
+- [Configuration Guide](https://docs.axolotl.ai/docs/config.html) - Full configuration options and examples
-- [Dataset Guide](https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/) - Supported formats and how to use them
+- [Dataset Guide](https://docs.axolotl.ai/docs/dataset-formats/) - Supported formats and how to use them
-- [Multi-GPU Training](https://axolotl-ai-cloud.github.io/axolotl/docs/multi-gpu.html)
+- [Multi-GPU Training](https://docs.axolotl.ai/docs/multi-gpu.html)
-- [Multi-Node Training](https://axolotl-ai-cloud.github.io/axolotl/docs/multi-node.html)
+- [Multi-Node Training](https://docs.axolotl.ai/docs/multi-node.html)
-- [Multipacking](https://axolotl-ai-cloud.github.io/axolotl/docs/multipack.html)
+- [Multipacking](https://docs.axolotl.ai/docs/multipack.html)
-- [FAQ](https://axolotl-ai-cloud.github.io/axolotl/docs/faq.html) - Frequently asked questions
+- [API Reference](https://docs.axolotl.ai/docs/api/) - Auto-generated code documentation
+- [FAQ](https://docs.axolotl.ai/docs/faq.html) - Frequently asked questions
 
 ## 🤝 Getting Help
 
 - Join our [Discord community](https://discord.gg/HhrNrHJPRb) for support
 - Check out our [Examples](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/) directory
-- Read our [Debugging Guide](https://axolotl-ai-cloud.github.io/axolotl/docs/debugging.html)
+- Read our [Debugging Guide](https://docs.axolotl.ai/docs/debugging.html)
 - Need dedicated support? Please contact [✉️wing@axolotl.ai](mailto:wing@axolotl.ai) for options
 
 ## 🌟 Contributing

_quarto.yml

@@ -1,6 +1,180 @@
 project:
   type: website
 
+quartodoc:
+  dir: docs/api
+  package: axolotl
+  title: API Reference
+  parser: google
+
+  sections:
+    - title: Core
+      desc: Core functionality for training
+      contents:
+        - train
+        - evaluate
+        - datasets
+        - convert
+        - prompt_tokenizers
+        - logging_config
+        - core.trainer_builder
+        - core.training_args
+        - core.chat.messages
+        - core.chat.format.chatml
+        - core.chat.format.llama3x
+        - core.chat.format.shared
+        - core.datasets.chat
+        - core.datasets.transforms.chat_builder
+    - title: CLI
+      desc: Command-line interface
+      contents:
+        - cli.main
+        - cli.train
+        - cli.evaluate
+        - cli.args
+        - cli.checks
+        - cli.config
+        - cli.inference
+        - cli.merge_lora
+        - cli.merge_sharded_fsdp_weights
+        - cli.preprocess
+        - cli.sweeps
+        - cli.utils
+        - cli.vllm_serve
+        - cli.cloud.base
+        - cli.cloud.modal_
+    - title: Trainers
+      desc: Training implementations
+      contents:
+        - core.trainers.base
+        - core.trainers.trl
+        - core.trainers.dpo.trainer
+        - core.trainers.grpo.trainer
+    - title: Prompt Strategies
+      desc: Prompt formatting strategies
+      contents:
+        - prompt_strategies.base
+        - prompt_strategies.chat_template
+        - prompt_strategies.alpaca_chat
+        - prompt_strategies.alpaca_instruct
+        - prompt_strategies.alpaca_w_system
+        - prompt_strategies.user_defined
+        - prompt_strategies.llama2_chat
+        - prompt_strategies.completion
+        - prompt_strategies.input_output
+        - prompt_strategies.stepwise_supervised
+        - prompt_strategies.metharme
+        - prompt_strategies.orcamini
+        - prompt_strategies.pygmalion
+        - prompt_strategies.messages.chat
+        - prompt_strategies.dpo.chat_template
+        - prompt_strategies.dpo.llama3
+        - prompt_strategies.dpo.chatml
+        - prompt_strategies.dpo.zephyr
+        - prompt_strategies.dpo.user_defined
+        - prompt_strategies.dpo.passthrough
+        - prompt_strategies.kto.llama3
+        - prompt_strategies.kto.chatml
+        - prompt_strategies.kto.user_defined
+        - prompt_strategies.orpo.chat_template
+        - prompt_strategies.bradley_terry.llama3
+    - title: Kernels
+      desc: Low-level performance optimizations
+      contents:
+        - kernels.lora
+        - kernels.geglu
+        - kernels.swiglu
+        - kernels.quantize
+        - kernels.utils
+    - title: MonkeyPatches
+      desc: Runtime patches for model optimizations
+      contents:
+        - monkeypatch.llama_attn_hijack_flash
+        - monkeypatch.llama_attn_hijack_xformers
+        - monkeypatch.mistral_attn_hijack_flash
+        - monkeypatch.multipack
+        - monkeypatch.relora
+        - monkeypatch.llama_expand_mask
+        - monkeypatch.lora_kernels
+        - monkeypatch.utils
+        - monkeypatch.btlm_attn_hijack_flash
+        - monkeypatch.llama_patch_multipack
+        - monkeypatch.stablelm_attn_hijack_flash
+        - monkeypatch.trainer_fsdp_optim
+        - monkeypatch.transformers_fa_utils
+        - monkeypatch.unsloth_
+        - monkeypatch.attention.mllama
+        - monkeypatch.data.batch_dataset_fetcher
+        - monkeypatch.mixtral
+    - title: Utils
+      desc: Utility functions
+      contents:
+        - utils.models
+        - utils.tokenization
+        - utils.chat_templates
+        - utils.lora
+        - utils.lora_embeddings
+        - utils.model_shard_quant
+        - utils.bench
+        - utils.freeze
+        - utils.trainer
+        - utils.schedulers
+        - utils.distributed
+        - utils.dict
+        - utils.optimizers.adopt
+        - utils.data.pretraining
+        - utils.data.sft
+        - utils.gradient_checkpointing.unsloth
+    - title: Schemas
+      desc: Pydantic data models for Axolotl config
+      contents:
+        - utils.schemas.config
+        - utils.schemas.model
+        - utils.schemas.training
+        - utils.schemas.datasets
+        - utils.schemas.peft
+        - utils.schemas.trl
+        - utils.schemas.multimodal
+        - utils.schemas.integrations
+        - utils.schemas.enums
+        - utils.schemas.utils
+    - title: Integrations
+      desc: Third-party integrations and extensions
+      contents:
+        - integrations.base
+        - integrations.cut_cross_entropy.args
+        - integrations.grokfast.optimizer
+        - integrations.kd.trainer
+        - integrations.liger.args
+        - integrations.lm_eval.args
+        - integrations.spectrum.args
+    - title: Common
+      desc: Common utilities and shared functionality
+      contents:
+        - common.architectures
+        - common.const
+        - common.datasets
+    - title: Models
+      desc: Custom model implementations
+      contents:
+        - models.mamba.modeling_mamba
+    - title: Data Processing
+      desc: Data processing utilities
+      contents:
+        - utils.collators.core
+        - utils.collators.batching
+        - utils.collators.mamba
+        - utils.collators.mm_chat
+        - utils.samplers.multipack
+    - title: Callbacks
+      desc: Training callbacks
+      contents:
+        - utils.callbacks.perplexity
+        - utils.callbacks.profiler
+        - utils.callbacks.lisa
+        - utils.callbacks.mlflow_
+        - utils.callbacks.comet_
+
 website:
   title: "Axolotl"
   description: "We make fine-tuning accessible, scalable, and fun"
@@ -35,6 +209,8 @@ website:
             - docs/inference.qmd
             - docs/cli.qmd
             - docs/config.qmd
+            - text: "API Reference"
+              href: docs/api
 
         - section: "Dataset Formats"
           contents: docs/dataset-formats/*
@@ -55,6 +231,7 @@ website:
             - docs/reward_modelling.qmd
             - docs/lr_groups.qmd
             - docs/lora_optims.qmd
+            - docs/dataset_loading.qmd
 
         - section: "Core Concepts"
           contents:
@@ -68,6 +245,7 @@ website:
             - docs/unsloth.qmd
             - docs/torchao.qmd
             - docs/custom_integrations.qmd
+            - docs/sequence_parallelism.qmd
 
         - section: "Troubleshooting"
           contents:
@@ -80,3 +258,22 @@ format:
     theme: darkly
     css: styles.css
     toc: true
+    # Enable better handling of line breaks in markdown
+    preserve-tabs: true
+    html-math-method: mathjax
+    # Improved markdown processing options
+    md-extensions:
+      - markdown_it
+      - def_list
+      - attr_list
+      - fenced_divs
+      - tables
+      - html_admonition
+      - lineblocks
+      - fancy_lists
+    # Control whitespace handling
+    whitespace: preserve
+    # Process newlines in paragraphs
+    wrap: preserve
+    # Better line break handling
+    preserve-linebreaks: true

cicd/Dockerfile.jinja

@@ -33,9 +33,9 @@ RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
 
 RUN pip install packaging==23.2 setuptools==75.8.0
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        pip install --no-build-isolation -e .[deepspeed,flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \
-        pip install --no-build-isolation -e .[deepspeed,flash-attn,optimizers,ray] $AXOLOTL_ARGS; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray] $AXOLOTL_ARGS; \
     fi
 
 RUN python scripts/unsloth_install.py | sh

cicd/cicd.sh

@@ -3,9 +3,59 @@ set -e
 
 python -c "import torch; assert '$PYTORCH_VERSION' in torch.__version__"
 
-pytest -v --durations=10 -n8 --ignore=tests/e2e/ --ignore=tests/patched/ /workspace/axolotl/tests/
+# Run unit tests with initial coverage report
-pytest -v --durations=10 /workspace/axolotl/tests/e2e/patched/lora_kernels  # running these with the other patches causes a failure
+pytest -v --durations=10 -n8 \
-pytest -v --durations=10 --ignore=tests/e2e/patched/lora_kernels /workspace/axolotl/tests/e2e/patched
+  --ignore=tests/e2e/ \
-pytest -v --durations=10 -n1 /workspace/axolotl/tests/e2e/solo/
+  --ignore=tests/patched/ \
-pytest -v --durations=10 /workspace/axolotl/tests/e2e/integrations/
+  --ignore=tests/cli \
-pytest -v --durations=10 --ignore=tests/e2e/solo/ --ignore=tests/e2e/patched/ --ignore=tests/e2e/multigpu/ --ignore=tests/e2e/integrations/ /workspace/axolotl/tests/e2e/
+  /workspace/axolotl/tests/ \
+  --cov=axolotl \
+  --cov-report=xml:coverage.xml
+
+# Run lora kernels tests with coverage append
+pytest -v --durations=10 \
+  /workspace/axolotl/tests/e2e/patched/lora_kernels \
+  --cov=axolotl \
+  --cov-append
+
+# Run patched tests excluding lora kernels with coverage append
+pytest -v --durations=10 \
+  --ignore=tests/e2e/patched/lora_kernels \
+  /workspace/axolotl/tests/e2e/patched \
+  --cov=axolotl \
+  --cov-append
+
+# Run solo tests with coverage append
+pytest -v --durations=10 -n1 \
+  /workspace/axolotl/tests/e2e/solo/ \
+  --cov=axolotl \
+  --cov-append
+
+# Run integration tests with coverage append
+pytest -v --durations=10 \
+  /workspace/axolotl/tests/e2e/integrations/ \
+  --cov=axolotl \
+  --cov-append
+
+pytest -v --durations=10 /workspace/axolotl/tests/cli \
+  --cov=axolotl \
+  --cov-append
+
+# Run remaining e2e tests with coverage append and final report
+pytest -v --durations=10 \
+  --ignore=tests/e2e/solo/ \
+  --ignore=tests/e2e/patched/ \
+  --ignore=tests/e2e/multigpu/ \
+  --ignore=tests/e2e/integrations/ \
+  --ignore=tests/cli \
+  /workspace/axolotl/tests/e2e/ \
+  --cov=axolotl \
+  --cov-append \
+  --cov-report=xml:coverage.xml
+
+# Upload coverage to Codecov
+if [ -f e2e-coverage.xml ]; then
+  codecov -f e2e-coverage.xml -F e2e,pytorch-${PYTORCH_VERSION}
+else
+  echo "Coverage file not found. Coverage report may have failed."
+fi

cicd/e2e_tests.py

@@ -1,4 +1,5 @@
 """Modal app to run axolotl GPU tests"""
+
 # pylint: disable=duplicate-code
 
 import os

cicd/multigpu.py

@@ -1,6 +1,7 @@
 """
- modal application to run axolotl gpu tests in Modal
+modal application to run axolotl gpu tests in Modal
- """
+"""
+
 # pylint: disable=duplicate-code
 
 import os
@@ -67,7 +68,7 @@ def run_cmd(cmd: str, run_folder: str):
 @app.function(
     image=cicd_image,
     gpu=GPU_CONFIG,
-    timeout=60 * 60,
+    timeout=90 * 60,
     cpu=8.0,
     memory=131072 * N_GPUS,
     volumes=VOLUME_CONFIG,

cicd/multigpu.sh

@@ -2,4 +2,24 @@
 set -e
 
 # only run one test at a time so as not to OOM the GPU
-pytest -v -n2 /workspace/axolotl/tests/e2e/multigpu/
+pytest -v  --durations=10 -n2 /workspace/axolotl/tests/e2e/multigpu/ --ignore=/workspace/axolotl/tests/e2e/multigpu/solo/
+pytest -v  --durations=10 -n1 /workspace/axolotl/tests/e2e/multigpu/solo/
+
+# Only run two tests at a time to avoid OOM on GPU (with coverage collection)
+pytest -v -n2 \
+  --ignore=/workspace/axolotl/tests/e2e/multigpu/solo/
+  /workspace/axolotl/tests/e2e/multigpu/ \
+  --cov=axolotl \
+  --cov-report=xml:multigpu-coverage.xml
+
+pytest -v  --durations=10 -n1 /workspace/axolotl/tests/e2e/multigpu/solo/ \
+  --cov=axolotl \
+  --cov-append \
+  --cov-report=xml:multigpu-coverage.xml
+
+# Upload coverage to Codecov
+if [ -f multigpu-coverage.xml ]; then
+  codecov -f multigpu-coverage.xml -F multigpu,docker-tests,pytorch-${PYTORCH_VERSION}
+else
+  echo "Coverage file not found. Coverage report may have failed."
+fi

codecov.yml

@@ -0,0 +1,51 @@
+codecov:
+  require_ci_to_pass: yes
+
+coverage:
+  precision: 2
+  round: down
+  range: "70...100"
+  status:
+    project:
+      default:
+        # basic
+        target: auto
+        threshold: 0%
+        base: auto
+        # advanced
+        branches: null
+        if_no_uploads: error
+        if_not_found: success
+        if_ci_failed: error
+        only_pulls: false
+        flags: null
+        paths: null
+    patch:
+      default:
+        # basic
+        target: auto
+        threshold: 0%
+        base: auto
+        # advanced
+        branches: null
+        if_no_uploads: error
+        if_not_found: success
+        if_ci_failed: error
+        only_pulls: false
+        flags: null
+        paths: null
+
+parsers:
+  gcov:
+    branch_detection:
+      conditional: yes
+      loop: yes
+      method: no
+      macro: no
+
+comment:
+  layout: "reach,diff,flags,files,footer"
+  behavior: default
+  require_changes: no
+  require_base: no
+  require_head: yes

docker/Dockerfile

@@ -20,9 +20,9 @@ WORKDIR /workspace/axolotl
 
 # If AXOLOTL_EXTRAS is set, append it in brackets
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        pip install --no-build-isolation -e .[deepspeed,flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \
-        pip install --no-build-isolation -e .[deepspeed,flash-attn,optimizers,ray] $AXOLOTL_ARGS; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray] $AXOLOTL_ARGS; \
     fi
 
 RUN python scripts/unsloth_install.py | sh

docker/Dockerfile-base

@@ -29,7 +29,7 @@ ENV PATH="/root/miniconda3/envs/py${PYTHON_VERSION}/bin:${PATH}"
 WORKDIR /workspace
 
 RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==23.2 setuptools==75.8.0 wheel && \
-    python3 -m pip install --no-cache-dir -U torch==${PYTORCH_VERSION}+cu${CUDA} --extra-index-url https://download.pytorch.org/whl/cu$CUDA && \
+    python3 -m pip install --no-cache-dir -U torch==${PYTORCH_VERSION}+cu${CUDA} torchvision --extra-index-url https://download.pytorch.org/whl/cu$CUDA && \
     python3 -m pip install --no-cache-dir "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main" && \
     python3 -m pip install --no-cache-dir "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main"
 

docker/Dockerfile-base-next

@@ -0,0 +1,38 @@
+ARG CUDA_VERSION="12.8.1"
+ARG CUDNN_VERSION="8"
+ARG UBUNTU_VERSION="22.04"
+ARG MAX_JOBS=4
+
+FROM nvidia/cuda:$CUDA_VERSION-cudnn$CUDNN_VERSION-devel-ubuntu$UBUNTU_VERSION AS base-builder
+
+ENV PATH="/root/miniconda3/bin:${PATH}"
+
+ARG PYTHON_VERSION="3.11"
+ARG PYTORCH_VERSION="next"
+ARG CUDA="128"
+ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 9.0+PTX"
+
+ENV PYTHON_VERSION=$PYTHON_VERSION
+ENV TORCH_CUDA_ARCH_LIST=$TORCH_CUDA_ARCH_LIST
+
+RUN apt-get update \
+    && apt-get install -y wget git build-essential ninja-build git-lfs libaio-dev pkg-config && rm -rf /var/lib/apt/lists/* \
+    && wget \
+    https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh \
+    && mkdir /root/.conda \
+    && bash Miniconda3-latest-Linux-x86_64.sh -b \
+    && rm -f Miniconda3-latest-Linux-x86_64.sh \
+    && conda create -n "py${PYTHON_VERSION}" python="${PYTHON_VERSION}"
+
+ENV PATH="/root/miniconda3/envs/py${PYTHON_VERSION}/bin:${PATH}"
+
+WORKDIR /workspace
+
+RUN python3 -m pip install --upgrade pip && pip3 install packaging && \
+    python3 -m pip install --no-cache-dir -U torch==2.7.0 --extra-index-url https://download.pytorch.org/whl/test/cu$CUDA && \
+    python3 -m pip install --no-cache-dir "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main" && \
+    python3 -m pip install --no-cache-dir "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main"
+
+RUN git lfs install --skip-repo && \
+    pip3 install awscli && \
+    pip3 install -U --no-cache-dir pydantic==2.10.6

docs/.gitignore

@@ -1,2 +1,4 @@
 /.quarto/
 _site/
+/api/*.qmd
+/api/*.html

docs/cli.qmd

@@ -1,5 +1,5 @@
 ---
-title: "CLI Reference"
+title: "Command Line Interface (CLI)"
 format:
   html:
     toc: true
@@ -170,7 +170,7 @@ axolotl merge-sharded-fsdp-weights config.yml
 
 ### evaluate
 
-Evaluates a model's performance using metrics specified in the config.
+Evaluates a model's performance (loss etc) on the train and eval datasets.
 
 ```bash
 # Basic evaluation
@@ -197,6 +197,8 @@ lm_eval_batch_size: # Batch size for evaluation
 output_dir: # Directory to save evaluation results
 ```
 
+See [LM Eval Harness](https://github.com/EleutherAI/lm-evaluation-harness) for more details.
+
 ## Legacy CLI Usage
 
 While the new Click-based CLI is preferred, Axolotl still supports the legacy module-based CLI:
@@ -235,7 +237,7 @@ Create a cloud config YAML with your Modal settings:
 ```yaml
 # cloud_config.yml
 provider: modal
-gpu: a100  # Supported: l40s, a100-40gb, a100-80gb, a10g, h100, t4, l4
+gpu: a100       # Supported: l40s, a100-40gb, a100-80gb, a10g, h100, t4, l4
 gpu_count: 1    # Number of GPUs to use
 timeout: 86400  # Maximum runtime in seconds (24 hours)
 branch: main    # Git branch to use (optional)
@@ -248,7 +250,7 @@ volumes:        # Persistent storage volumes
   - name: axolotl-artifacts
     mount: /workspace/artifacts
 
-env:            # Environment variables
+secrets:        # Secrets to inject
   - WANDB_API_KEY
   - HF_TOKEN
 ```
@@ -274,15 +276,27 @@ axolotl lm-eval config.yml --cloud cloud_config.yml
 ### Cloud Configuration Options
 
 ```yaml
-provider: # compute provider, currently only `modal` is supported
+provider:    # compute provider, currently only `modal` is supported
-gpu: # GPU type to use
+gpu:         # GPU type to use
-gpu_count: # Number of GPUs (default: 1)
+gpu_count:   # Number of GPUs (default: 1)
-memory: # RAM in GB (default: 128)
+memory:      # RAM in GB (default: 128)
-timeout: # Maximum runtime in seconds
+timeout:     # Maximum runtime in seconds
 timeout_preprocess: # Preprocessing timeout
-branch: # Git branch to use
+branch:      # Git branch to use
-docker_tag: # Custom Docker image tag
+docker_tag:  # Custom Docker image tag
-volumes: # List of persistent storage volumes
+volumes:     # List of persistent storage volumes
-env: # Environment variables to pass
+
-secrets: # Secrets to inject
+# Environment variables to pass. Can be specified in two ways:
+# 1. As a string: Will load the value from the host computer's environment variables
+# 2. As a key-value pair: Will use the specified value directly
+# Example:
+# env:
+#   - CUSTOM_VAR  # Loads from host's $CUSTOM_VAR
+#   - {CUSTOM_VAR: "value"}  # Uses "value" directly
+env:
+
+# Secrets to inject. Same input format as `env` but for sensitive data.
+secrets:
+  # - HF_TOKEN
+  # - WANDB_API_KEY
 ```

docs/config.qmd

@@ -32,6 +32,9 @@ tokenizer_legacy:
 resize_token_embeddings_to_32x:
 # Optional[bool] Whether to shrink the embeddings to len(tokenizer). By default, we won't shrink.
 shrink_embeddings:
+# Whether to load the model with randomly initialized weights. Useful for
+# pre-training a model from scratch or debugging purposes.
+random_init_weights:
 
 # (Internal use only)
 # Used to identify which the model is based on
@@ -87,7 +90,7 @@ lora_on_cpu: true
 
 # List[str]. Add plugins to extend the pipeline.
 # See `src/axolotl/integrations` for the available plugins or doc below for more details.
-# https://axolotl-ai-cloud.github.io/axolotl/docs/custom_integrations.html
+# https://docs.axolotl.ai/docs/custom_integrations.html
 plugins:
   # - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
 
@@ -106,7 +109,7 @@ datasets:
     preprocess_shards: # Optional[int] process dataset in N sequential chunks for memory efficiency (exclusive with `shards`)
 
     name: # Optional[str] name of dataset configuration to load
-    train_on_split: train # Optional[str] name of dataset split to load from
+    split: train # Optional[str] name of dataset split to load from
     revision: # Optional[str] The specific revision of the dataset to use when loading from the Hugging Face Hub. This can be a commit hash, tag, or branch name. If not specified, the latest version will be used. This parameter is ignored for local datasets.
     trust_remote_code: # Optional[bool] Trust remote code for untrusted source
 
@@ -162,7 +165,9 @@ datasets:
       content: value
       # ...
 
-    # Optional[Dict[str, List]]. Roles mapping in the messages. The default is:
+    # Optional[Dict[str, List]]. Roles mapping in the messages.
+    # The format is {target_role: [source_roles]}. All source roles will be mapped to the target role.
+    # The default is:
     roles:
       user: ["human", "user"]
       assistant: ["gpt", "assistant"]
@@ -235,10 +240,10 @@ simpo_gamma: 0.5  # Target reward margin for the SimPO loss
 # grpo
 trl:
   use_vllm: # Optional[bool]. Whether to use VLLM for RL training.
-  vllm_device: # Optional[str]. Device to use for VLLM.
+  vllm_server_host: # Optional[str]. Host of the vLLM server to connect to.
-  vllm_gpu_memory_utilization: # Optional[float]. GPU memory utilization for VLLM.
+  vllm_server_port: # Optional[int]. Port of the vLLM server to connect to.
-  vllm_max_model_len: # Optional[int]. Maximum length of the model for VLLM.
+  vllm_server_timeout: # Optional[int]. Total timeout (in seconds) to wait for the vLLM server to respond.
-  vllm_dtype: # Optional[str]. Data type for VLLM.
+  vllm_guided_decoding_regex: # Optional[str]. Regex for vLLM guided decoding.
 
   beta: # Optional[float]. Beta parameter for the RL training. Same as `rl_beta`. Use
   max_completion_length: # Optional[int]. Maximum length of the completion for RL training.
@@ -317,9 +322,13 @@ total_num_tokens:
 sample_packing_group_size: 100000
 # The number of samples which can be packed into one sequence. Increase if using a large sequence_len with many short samples.
 sample_packing_bin_size: 200
+sample_pack_sequentially: # Optional[bool]. Whether to pack samples sequentially.
+
 # whether to concatenate samples during pretraining
 pretraining_sample_concatenation:
 
+curriculum_sampling: # Optional[bool]. Whether to use sequential sampling for curriculum learning
+
 # Use batch flattening for speedups when not using sample_packing
 batch_flattening:
 
@@ -351,7 +360,27 @@ lora_target_modules:
 #  - down_proj
 #  - up_proj
 lora_target_linear: # If true, will target all linear modules
-peft_layers_to_transform: # The layer indices to transform, otherwise, apply to all layers
+
+# List[int] | int. # The layer indices to transform, otherwise, apply to all layers
+# https://huggingface.co/docs/peft/v0.15.0/en/package_reference/lora#peft.LoraConfig.layers_to_transform
+peft_layers_to_transform:
+
+# Optional[bool]. Whether to use DoRA.
+# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#weight-decomposed-low-rank-adaptation-dora
+peft_use_dora:
+
+# Optional[bool]. Whether to use RSLoRA.
+# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#rank-stabilized-lora
+peft_use_rslora:
+
+# Optional[list[tuple[int, int]]]. List of layer indices to replicate.
+# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#memory-efficient-layer-replication-with-lora
+peft_layer_replication:
+
+# bool | Literal["gaussian", "eva", "olora", "pissa", "pissa_niter_[number of iters]", "corda", "loftq"]
+# How to initialize LoRA weights. Default to True which is MS original implementation.
+# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#initialization
+peft_init_lora_weights:
 
 # If you added new tokens to the tokenizer, you may need to save some LoRA modules because they need to know the new tokens.
 # For LLaMA and Mistral, you need to save `embed_tokens` and `lm_head`. It may vary for other models.
@@ -365,7 +394,7 @@ lora_fan_in_fan_out: false
 
 # Apply custom LoRA autograd functions and activation function Triton kernels for
 # speed and memory savings
-# See: https://axolotl-ai-cloud.github.io/axolotl/docs/lora_optims.html
+# See: https://docs.axolotl.ai/docs/lora_optims.html
 lora_mlp_kernel: true
 lora_qkv_kernel: true
 lora_o_kernel: true
@@ -463,6 +492,7 @@ auto_find_batch_size: # Optional[bool]
 
 eval_table_size: # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0
 eval_max_new_tokens: # Total number of tokens generated for predictions sent to wandb. Default is 128
+do_causal_lm_eval: # Whether to run causal language model evaluation for metrics in `eval_causal_lm_metrics`.
 eval_causal_lm_metrics: # HF evaluate metrics used during evaluation. Default is ["sacrebleu", "comet", "ter", "chrf", "perplexity"]
 
 profiler_steps: # enable the pytorch profiler to capture the first N steps of training to the output_dir.
@@ -482,7 +512,8 @@ train_on_inputs: false
 # Note that training loss may have an oscillating pattern with this enabled.
 group_by_length: false
 
-# Whether to use gradient checkpointing https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing
+# Whether to use gradient checkpointing. Available options are: true, false, "offload".
+# https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing
 gradient_checkpointing: false
 # additional kwargs to pass to the trainer for gradient checkpointing
 # gradient_checkpointing_kwargs:
@@ -503,36 +534,58 @@ lr_div_factor: # Learning rate div factor
 
 # Specify optimizer
 # Valid values are driven by the Transformers OptimizerNames class, see:
-# https://github.com/huggingface/transformers/blob/95b374952dc27d8511541d6f5a4e22c9ec11fb24/src/transformers/training_args.py#L134
+# https://github.com/huggingface/transformers/blob/cbf924b76c03828101a34069a96d209314114fd5/src/transformers/training_args.py#L144-L189
 #
 # Note that not all optimizers may be available in your environment, ex: 'adamw_anyprecision' is part of
 # torchdistx, 'adamw_bnb_8bit' is part of bnb.optim.Adam8bit, etc. When in doubt, it is recommended to start with the optimizer used
 # in the examples/ for your model and fine-tuning use case.
 #
 # Valid values for 'optimizer' include:
-# - adamw_hf
 # - adamw_torch
 # - adamw_torch_fused
 # - adamw_torch_xla
+# - adamw_torch_npu_fused
 # - adamw_apex_fused
-# - adopt_adamw (an EXPERIMENTAL optimizer, only for torch version >= 2.5.1)
+# - adopt_adamw  (an EXPERIMENTAL optimizer, only for torch version >= 2.5.1)
 # - adafactor
 # - adamw_anyprecision
+# - adamw_torch_4bit
+# - ademamix
 # - sgd
 # - adagrad
 # - adamw_bnb_8bit
+# - adamw_8bit   # alias for adamw_bnb_8bit
+# - ademamix_8bit
 # - lion_8bit
 # - lion_32bit
 # - paged_adamw_32bit
 # - paged_adamw_8bit
+# - paged_ademamix_32bit
+# - paged_ademamix_8bit
 # - paged_lion_32bit
 # - paged_lion_8bit
+# - rmsprop
+# - rmsprop_bnb
+# - rmsprop_bnb_8bit
+# - rmsprop_bnb_32bit
 # - galore_adamw
 # - galore_adamw_8bit
 # - galore_adafactor
 # - galore_adamw_layerwise
 # - galore_adamw_8bit_layerwise
 # - galore_adafactor_layerwise
+# - lomo
+# - adalomo
+# - grokadamw
+# - schedule_free_adamw
+# - schedule_free_sgd
+# - apollo_adamw
+# - apollo_adamw_layerwise
+#
+# Additional custom optimizers include:
+# - optimi_adamw
+# - ao_adamw_8bit
+# - ao_adamw_fp8
 optimizer:
 # Dictionary of arguments to pass to the optimizer
 optim_args:
@@ -561,29 +614,42 @@ max_grad_norm:
 # currently only supported on Llama and Mistral
 neftune_noise_alpha:
 
-# Whether to bettertransformers
+# Optional[bool]. Whether to bettertransformers
 flash_optimum:
-# Whether to use xformers attention patch https://github.com/facebookresearch/xformers:
+
+# Note: Only one of the following attention patches can be used at a time.
+# For example, if you set `xformers_attention` to `true`, do not set `flash_attention` to `true`.
+
+# Optional[bool]. Whether to use xformers attention patch https://github.com/facebookresearch/xformers:
 xformers_attention:
-# Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:
+# Optional[bool]. Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:
 flash_attention:
-flash_attn_cross_entropy:  # Whether to use flash-attention cross entropy implementation - advanced use only
+flash_attn_cross_entropy:  # Optional[bool]. Whether to use flash-attention cross entropy implementation - advanced use only
-flash_attn_rms_norm:  # Whether to use flash-attention rms norm implementation - advanced use only
+flash_attn_rms_norm:  # Optional[bool]. Whether to use flash-attention rms norm implementation - advanced use only
-flash_attn_fuse_qkv: # Whether to fuse QKV into a single operation
+flash_attn_fuse_qkv: # Optional[bool]. Whether to fuse QKV into a single operation
-flash_attn_fuse_mlp: # Whether to fuse part of the MLP into a single operation
+flash_attn_fuse_mlp: # Optional[bool]. Whether to fuse part of the MLP into a single operation
-# Whether to use scaled-dot-product attention
+# Optional[bool]. Whether to use scaled-dot-product attention
 # https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html
 sdp_attention:
-# Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf
+# Optional[bool]. Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf
 s2_attention:
+
 # Optional[bool]. Whether to use low_cpu_mem_usage
 low_cpu_mem_usage:
-# Resume from a specific checkpoint dir
+# Optional[str]. Resume from a specific checkpoint dir
 resume_from_checkpoint:
-# If resume_from_checkpoint isn't set and you simply want it to start where it left off.
+# Optional[bool]. If resume_from_checkpoint isn't set and you simply want it to start where it left off.
 # Be careful with this being turned on between different models.
 auto_resume_from_checkpoints: false
 
+## Multimodal section
+# int | tuple[int, int] | None . Size to resize images to, width x height.
+# Will read from model/processor config if not set.
+image_size:
+# str. Algorithm to use for image resizing. "bilinear", "bicubic", "lanczos". Default is "bilinear".
+image_resize_algorithm: 'bilinear'
+## End of multimodal section
+
 # Don't mess with this, it's here for accelerate and torchrun
 local_rank:
 
@@ -617,6 +683,20 @@ ddp_timeout:
 ddp_bucket_cap_mb:
 ddp_broadcast_buffers:
 
+# Sequence parallelism
+# Set to a divisor of the number of GPUs available to split sequences into chunks of equal size.
+# Use in long context training to prevent OOM when sequences cannot fit into a single GPU's VRAM.
+# E.g., if 4 GPUs are available, set this value to 2 to split each sequence into two equal-sized
+# subsequences, or set to 4 to split into four equal-sized subsequences.
+# See https://docs.axolotl.ai/docs/sequence_parallelism.html for more details.
+sequence_parallel_degree:
+# Optional; strides across the key dimension. Larger values use more memory but should make training faster.
+# Must evenly divide the number of KV heads in your model.
+heads_k_stride: 1
+# One of "varlen_llama3", "batch_ring", "batch_zigzag", "batch_stripe". Defaults to "varlen_llama3"
+# in the sample packing case, and "batch_ring" in the non-sample packing case.
+ring_attn_func:
+
 # Path to torch distx for optim 'adamw_anyprecision'
 torchdistx_path:
 

docs/dataset-formats/index.qmd

@@ -13,6 +13,13 @@ As there are a lot of available options in Axolotl, this guide aims to provide a
 
 Axolotl supports 3 kinds of training methods: pre-training, supervised fine-tuning, and preference-based post-training (e.g. DPO, ORPO, PRMs). Each method has their own dataset format which are described below.
 
+::: {.callout-tip}
+
+This guide will mainly use JSONL as an introduction. Please refer to the [dataset loading docs](../dataset_loading.qmd) to understand how to load datasets from other sources.
+
+For `pretraining_dataset:` specifically, please refer to the [Pre-training section](#pre-training).
+:::
+
 ## Pre-training
 
 When aiming to train on large corpora of text datasets, pre-training is your go-to choice. Due to the size of these datasets, downloading the entire-datasets before beginning training would be prohibitively time-consuming. Axolotl supports [streaming](https://huggingface.co/docs/datasets/en/stream) to only load batches into memory at a time.
@@ -450,10 +457,7 @@ datasets:
     type: alpaca
 ```
 
-Axolotl supports many kinds of instruction dataset. All of them can be found here (https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/inst_tune.html) with their respective type and sample row format.
+Axolotl supports many kinds of instruction dataset. All of them can be found in the [Instruction Dataset Documentation](inst_tune.qmd) with their respective type and sample row format.
-
-
-Reference: [Instruction Dataset Documentation](inst_tune.qmd).
 
 #### Custom Instruct Prompt Format
 

docs/dataset_loading.qmd

@@ -0,0 +1,276 @@
+---
+title: Dataset Loading
+description: Understanding how to load datasets from different sources
+back-to-top-navigation: true
+toc: true
+toc-depth: 5
+---
+
+## Overview
+
+Datasets can be loaded in a number of different ways depending on the how it is saved (the extension of the file) and where it is stored.
+
+## Loading Datasets
+
+We use the `datasets` library to load datasets and a mix of `load_dataset` and `load_from_disk` to load them.
+
+You may recognize the similar named configs between `load_dataset` and the `datasets` section of the config file.
+
+```yaml
+datasets:
+  - path:
+    name:
+    data_files:
+    split:
+    revision:
+    trust_remote_code:
+```
+
+::: {.callout-tip}
+
+Do not feel overwhelmed by the number of options here. A lot of them are optional. In fact, the most common config to use would be `path` and sometimes `data_files`.
+
+:::
+
+This matches the API of [`datasets.load_dataset`](https://github.com/huggingface/datasets/blob/0b5998ac62f08e358f8dcc17ec6e2f2a5e9450b6/src/datasets/load.py#L1838-L1858), so if you're familiar with that, you will feel right at home.
+
+For HuggingFace's guide to load different dataset types, see [here](https://huggingface.co/docs/datasets/loading).
+
+For full details on the config, see [config.qmd](config.qmd).
+
+::: {.callout-note}
+
+You can set multiple datasets in the config file by more than one entry under `datasets`.
+
+```yaml
+datasets:
+  - path: /path/to/your/dataset
+  - path: /path/to/your/other/dataset
+```
+
+:::
+
+### Local dataset
+
+#### Files
+
+Usually, to load a JSON file, you would do something like this:
+
+```python
+from datasets import load_dataset
+
+dataset = load_dataset("json", data_files="data.json")
+```
+
+Which translates to the following config:
+
+```yaml
+datasets:
+  - path: json
+    data_files: /path/to/your/file.jsonl
+```
+
+However, to make things easier, we have added a few shortcuts for loading local dataset files.
+
+You can just point the `path` to the file or directory along with the `ds_type` to load the dataset. The below example shows for a JSON file:
+
+```yaml
+datasets:
+  - path: /path/to/your/file.jsonl
+    ds_type: json
+```
+
+This works for CSV, JSON, Parquet, and Arrow files.
+
+::: {.callout-tip}
+
+If `path` points to a file and `ds_type` is not specified, we will automatically infer the dataset type from the file extension, so you could omit `ds_type` if you'd like.
+
+:::
+
+#### Directory
+
+If you're loading a directory, you can point the `path` to the directory.
+
+Then, you have two options:
+
+##### Loading entire directory
+
+You do not need any additional configs.
+
+We will attempt to load in the following order:
+- datasets saved with `datasets.save_to_disk`
+- loading entire directory of files (such as with parquet/arrow files)
+
+```yaml
+datasets:
+  - path: /path/to/your/directory
+```
+
+##### Loading specific files in directory
+
+Provide `data_files` with a list of files to load.
+
+```yaml
+datasets:
+    # single file
+  - path: /path/to/your/directory
+    ds_type: csv
+    data_files: file1.csv
+
+    # multiple files
+  - path: /path/to/your/directory
+    ds_type: json
+    data_files:
+      - file1.jsonl
+      - file2.jsonl
+
+    # multiple files for parquet
+  - path: /path/to/your/directory
+    ds_type: parquet
+    data_files:
+      - file1.parquet
+      - file2.parquet
+
+```
+
+### HuggingFace Hub
+
+The method you use to load the dataset depends on how the dataset was created, whether a folder was uploaded directly or a HuggingFace Dataset was pushed.
+
+::: {.callout-note}
+
+If you're using a private dataset, you will need to enable the `hf_use_auth_token` flag in the root-level of the config file.
+
+:::
+
+#### Folder uploaded
+
+This would mean that the dataset is a single file or file(s) uploaded to the Hub.
+
+```yaml
+datasets:
+  - path: org/dataset-name
+    data_files:
+      - file1.jsonl
+      - file2.jsonl
+```
+
+#### HuggingFace Dataset
+
+This means that the dataset is created as a HuggingFace Dataset and pushed to the Hub via `datasets.push_to_hub`.
+
+```yaml
+datasets:
+  - path: org/dataset-name
+```
+
+::: {.callout-note}
+
+There are some other configs which may be required like `name`, `split`, `revision`, `trust_remote_code`, etc depending on the dataset.
+
+:::
+
+### Remote Filesystems
+
+Via the `storage_options` config under `load_dataset`, you can load datasets from remote filesystems like S3, GCS, Azure, and OCI.
+
+::: {.callout-warning}
+
+This is currently experimental. Please let us know if you run into any issues!
+
+:::
+
+The only difference between the providers is that you need to prepend the path with the respective protocols.
+
+```yaml
+datasets:
+    # Single file
+  - path: s3://bucket-name/path/to/your/file.jsonl
+
+    # Directory
+  - path: s3://bucket-name/path/to/your/directory
+```
+
+For directory, we load via `load_from_disk`.
+
+#### S3
+
+Prepend the path with `s3://`.
+
+The credentials are pulled in the following order:
+
+- `AWS_ACCESS_KEY_ID`, `AWS_SECRET_ACCESS_KEY`, and `AWS_SESSION_TOKEN` environment variables
+- from the `~/.aws/credentials` file
+- for nodes on EC2, the IAM metadata provider
+
+::: {.callout-note}
+
+We assume you have credentials setup and not using anonymous access. If you want to use anonymous access, let us know! We may have to open a config option for this.
+
+:::
+
+Other environment variables that can be set can be found in [boto3 docs](https://boto3.amazonaws.com/v1/documentation/api/latest/guide/configuration.html#using-environment-variables)
+
+#### GCS
+
+Prepend the path with `gs://` or `gcs://`.
+
+The credentials are loaded in the following order:
+
+- gcloud credentials
+- for nodes on GCP, the google metadata service
+- anonymous access
+
+#### Azure
+
+##### Gen 1
+
+Prepend the path with `adl://`.
+
+Ensure you have the following environment variables set:
+
+- `AZURE_STORAGE_TENANT_ID`
+- `AZURE_STORAGE_CLIENT_ID`
+- `AZURE_STORAGE_CLIENT_SECRET`
+
+##### Gen 2
+
+Prepend the path with `abfs://` or `az://`.
+
+Ensure you have the following environment variables set:
+
+- `AZURE_STORAGE_ACCOUNT_NAME`
+- `AZURE_STORAGE_ACCOUNT_KEY`
+
+Other environment variables that can be set can be found in [adlfs docs](https://github.com/fsspec/adlfs?tab=readme-ov-file#setting-credentials)
+
+#### OCI
+
+Prepend the path with `oci://`.
+
+It would attempt to read in the following order:
+
+- `OCIFS_IAM_TYPE`, `OCIFS_CONFIG_LOCATION`, and `OCIFS_CONFIG_PROFILE` environment variables
+- when on OCI resource, resource principal
+
+Other environment variables:
+
+- `OCI_REGION_METADATA`
+
+Please see the [ocifs docs](https://ocifs.readthedocs.io/en/latest/getting-connected.html#Using-Environment-Variables).
+
+### HTTPS
+
+The path should start with `https://`.
+
+```yaml
+datasets:
+  - path: https://path/to/your/dataset/file.jsonl
+```
+
+This must be publically accessible.
+
+## Next steps
+
+Now that you know how to load datasets, you can learn more on how to load your specific dataset format into your target output format [dataset formats docs](dataset-formats).

docs/dataset_preprocessing.qmd

@@ -6,7 +6,7 @@ description: How datasets are processed
 ## Overview
 
 Dataset pre-processing is the step where Axolotl takes each dataset you've configured alongside
-the [dataset format](docs/dataset-formats) and prompt strategies to:
+the [dataset format](dataset-formats) and prompt strategies to:
 
  - parse the dataset based on the *dataset format*
  - transform the dataset to how you would interact with the model based on the *prompt strategy*

docs/docker.qmd

@@ -103,8 +103,7 @@ This uses the same tags as the [`main` image](#sec-main-tags).
 
 - `JUPYTER_DISABLE`: Disable Jupyter lab.
 - `JUPYTER_PASSWORD`: Set a password for the Jupyter lab.
-- `PUBLIC_KEY`: Add a public key for the SSH service.
+- `PUBLIC_KEY` / `SSH_KEY`: Add a public key for the SSH service.
-- `SSH_KEY`: Add a private key for the SSH service.
 
 #### Volume mounts
 

docs/faq.qmd

@@ -35,7 +35,21 @@ description: Frequently asked questions
 
 **Q: How to call Axolotl via custom python scripts?**
 
-> A: Yes, since Axolotl is just Python, please see `src/axolotl/cli/main.py` on how each command is called.
+> A: Since Axolotl is just Python, please see `src/axolotl/cli/main.py` on how each command is called.
+
+**Q: How to know the value to use for `fsdp_transformer_layer_cls_to_wrap`?**
+
+> A: This is the class name of the transformer layer to wrap with FSDP. For example, for `LlamaForCausalLM`, the value is `LlamaDecoderLayer`. To find this for a specific model, check the model's `PreTrainedModel` definition and look for `_no_split_modules` variable in the `modeling_<model_name>.py` file within `transformers` library.
+
+**Q: ValueError: Asking to pad but the tokenizer does not have a padding token. Please select a token to use as pad_token**
+
+> A: This is because the tokenizer does not have a padding token. Please add a padding token to the tokenizer via:
+
+> ```yaml
+> special_tokens:
+>   # str. If you're not sure, set to same as `eos_token`.
+>   pad_token: "..."
+> ```
 
 ### Chat templates
 

docs/lora_optims.qmd

@@ -17,6 +17,7 @@ We currently support several common model architectures, including (but not limi
 - `qwen2`
 - `gemma`
 - `gemma2`
+- `gemma3`
 
 <details>
 

docs/multi-gpu.qmd

@@ -18,6 +18,7 @@ Axolotl supports several methods for multi-GPU training:
 
 - DeepSpeed (recommended)
 - FSDP (Fully Sharded Data Parallel)
+- Sequence parallelism
 - FSDP + QLoRA
 
 ## DeepSpeed {#sec-deepspeed}
@@ -35,6 +36,9 @@ deepspeed: deepspeed_configs/zero1.json
 ### Usage {#sec-deepspeed-usage}
 
 ```{.bash}
+# Fetch deepspeed configs (if not already present)
+axolotl fetch deepspeed_configs
+
 # Passing arg via config
 axolotl train config.yml
 
@@ -47,10 +51,20 @@ axolotl train config.yml --deepspeed deepspeed_configs/zero1.json
 We provide default configurations for:
 
 - ZeRO Stage 1 (`zero1.json`)
+- ZeRO Stage 1 with torch compile (`zero1_torch_compile.json`)
 - ZeRO Stage 2 (`zero2.json`)
 - ZeRO Stage 3 (`zero3.json`)
+- ZeRO Stage 3 with bf16 (`zero3_bf16.json`)
+- ZeRO Stage 3 with bf16 and CPU offload params(`zero3_bf16_cpuoffload_params.json`)
+- ZeRO Stage 3 with bf16 and CPU offload params and optimizer (`zero3_bf16_cpuoffload_all.json`)
 
-Choose based on your memory requirements and performance needs.
+::: {.callout-tip}
+
+Choose the configuration that offloads the least amount to memory while still being able to fit on VRAM for best performance.
+
+Start from Stage 1 -> Stage 2 -> Stage 3.
+
+:::
 
 ## FSDP {#sec-fsdp}
 
@@ -66,6 +80,28 @@ fsdp_config:
   fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
 ```
 
+## Sequence parallelism {#sec-sequence-parallelism}
+
+We support sequence parallelism (SP) via the
+[ring-flash-attention](https://github.com/zhuzilin/ring-flash-attention) project. This
+allows one to split up sequences across GPUs, which is useful in the event that a
+single sequence causes OOM errors during model training.
+
+First, install `ring-flash-attn`, recommended via `pip install axolotl[ring-flash-attn]`,
+or from source with `pip install .[ring-flash-attn]`.
+
+Your Axolotl YAML config should contain the following lines:
+
+```{.yaml}
+sequence_parallel_degree: 4  # Split each sequence into 4 parts, one per GPU
+flash_attention: true  # Required with sequence parallelism
+
+# Optional; strides across the key dimension. Larger values use more memory but will make training faster.
+heads_k_stride: 1
+```
+
+See our [dedicated guide](sequence_parallelism.qmd) for more details.
+
 ### FSDP + QLoRA {#sec-fsdp-qlora}
 
 For combining FSDP with QLoRA, see our [dedicated guide](fsdp_qlora.qmd).

docs/multimodal.qmd

@@ -1,28 +1,180 @@
-# MultiModal / Vision Language Models (BETA)
+---
+title: MultiModal / Vision Language Models (BETA)
+format:
+  html:
+    toc: true
+    toc-depth: 3
+---
 
-### Supported Models
+## Supported Models
 
-- Mllama, i.e. llama with vision models
+- [Mllama](#sec-mllama)
+- [Llama4](#sec-llama4)
+- [Pixtral](#sec-pixtral)
+- [Llava-1.5](#sec-llava-15)
+- [Mistral-Small-3.1](#sec-mistral-small-31)
+- [Gemma-3](#sec-gemma-3)
+- [Qwen2-VL](#sec-qwen2-vl)
+- [Qwen2.5-VL](#sec-qwen25-vl)
 
-### Usage
+## Usage
 
-Currently multimodal support is limited and doesn't have full feature parity. To finetune a multimodal Llama w/ LoRA,
+Multimodal support is limited and doesn't have full feature parity.
-you'll need to use the following in YAML in combination with the rest of the required hyperparams.
+
+Here are the hyperparams you'll need to use to finetune a multimodal model.
 
 ```yaml
-base_model: alpindale/Llama-3.2-11B-Vision-Instruct
 processor_type: AutoProcessor
-skip_prepare_dataset: true
 
-chat_template: llama3_2_vision
+skip_prepare_dataset: true
+remove_unused_columns: false  # leave columns in place as they are needed to handle image embeddings during training
+sample_packing: false  # not yet supported with multimodal
+
+chat_template:  # see in next section
+
+# example dataset
 datasets:
   - path: HuggingFaceH4/llava-instruct-mix-vsft
     type: chat_template
     split: train[:1%]
     field_messages: messages
-remove_unused_columns: false
-sample_packing: false
 
-# only finetune the Language model, leave the vision model and vision tower frozen
+# (optional) if doing lora, only finetune the Language model,
+# leave the vision model and vision tower frozen
+# load_in_8bit: true
+adapter: lora
 lora_target_modules: 'language_model.model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+
+# (optional) if you want to resize images to a set size
+image_size: 512
+image_resize_algorithm: bilinear
+```
+
+Please see [examples](https://github.com/axolotl-ai/axolotl/tree/main/examples) folder for full configs.
+
+::: {.callout-warning}
+Some of our chat_templates have been extended to support broader dataset types. This should not break any existing configs.
+:::
+
+### Mllama {#sec-mllama}
+
+```yaml
+base_model: meta-llama/Llama-3.2-11B-Vision-Instruct
+
+chat_template: llama3_2_vision
+```
+
+### Llama4 {#sec-llama4}
+
+```yaml
+base_model: meta-llama/Llama-4-Scout-17B-16E-Instruct
+
+chat_template: llama4
+```
+
+### Pixtral {#sec-pixtral}
+
+```yaml
+base_model: mistralai/Pixtral-12B-2409
+
+chat_template: pixtral
+```
+
+### Llava-1.5 {#sec-llava-15}
+
+```yaml
+base_model: llava-hf/llava-1.5-7b-hf
+
+chat_template: llava
+```
+
+### Mistral-Small-3.1 {#sec-mistral-small-31}
+
+```yaml
+base_model: mistralai/Mistral-Small-3.1-24B-Instruct-2503
+
+chat_template: mistral_v7_tekken
+```
+
+### Gemma-3 {#sec-gemma-3}
+
+::: {.callout-tip}
+The Gemma3-1B model is a text-only model, so please train as regular text model.
+:::
+
+For multi-modal 4B/12B/27B models, use the following config:
+
+```yaml
+base_model: google/gemma-3-4b-it
+
+chat_template: gemma3
+```
+
+### Qwen2-VL {#sec-qwen2-vl}
+
+```yaml
+base_model: Qwen/Qwen2-VL-7B-Instruct
+
+chat_template: qwen2_vl
+```
+
+### Qwen2.5-VL {#sec-qwen25-vl}
+
+```yaml
+base_model: Qwen/Qwen2.5-VL-7B-Instruct
+
+chat_template: qwen2_vl  # same as qwen2-vl
+```
+
+## Dataset Format
+
+For multi-modal datasets, we adopt an extended `chat_template` format similar to OpenAI's Message format.
+
+- A message is a list of `role` and `content`.
+- `role` can be `system`, `user`, `assistant`, etc.
+- `content` is a list of `type` and (`text` or `image` or `path` or `url` or `base64`).
+
+::: {.callout-note}
+For backwards compatibility:
+
+- If the dataset has a `images` or `image` column of `list[Image]`, it will be appended to the first `content` list as `{"type": "image", "image": ...}`. However, if the content already has a `{"type": "image"}` but no `image` key, it will be set the `image` key.
+- If `content` is a string, it will be converted to a list with `type` as `text`.
+:::
+
+::: {.callout-tip}
+For image loading, you can use the following keys within `content` alongside `"type": "image"`:
+
+- `"path": "/path/to/image.jpg"`
+- `"url": "https://example.com/image.jpg"`
+- `"base64": "..."`
+- `"image": PIL.Image`
+:::
+
+Here is an example of a multi-modal dataset:
+```json
+[
+  {
+    "messages": [
+        {
+            "role": "system",
+            "content": [
+              {"type": "text", "text": "You are a helpful assistant."}
+              ]
+        },
+        {
+            "role": "user",
+            "content": [
+                {"type": "image", "image": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"},
+                {"type": "text", "text": "Describe this image in detail."}
+            ]
+        },
+        {
+            "role": "assistant",
+            "content": [
+              {"type": "text", "text": "The image is a bee."}
+            ]
+        }
+    ]
+  }
+]
 ```

docs/rlhf.qmd

@@ -502,9 +502,48 @@ The input format is a simple JSON input with customizable fields based on the ab
 Check out our [GRPO cookbook](https://github.com/axolotl-ai-cloud/axolotl-cookbook/tree/main/grpo#training-an-r1-style-large-language-model-using-grpo).
 :::
 
+If you have multiple GPUs available, we reccomend using `vLLM` with the `GRPOTrainer` to significantly speedup trajectory generation during training.
+First, launch a `vLLM` server using `trl vllm-serve` - you may use a config file or CLI overrides to configure your vLLM server. In this example, we're
+using 4 GPUs - 2 for training, and 2 for vLLM:
+
+::: {.callout-important}
+Make sure you've installed the correct version of vLLM by including it as an extra when installing axolotl, e.g. `pip install axolotl[vllm]`.
+:::
+
+```yaml
+base_model: Qwen/Qwen2.5-1.5B-Instruct
+
+vllm:
+    host: 0.0.0.0
+    port: 8000
+    tensor_parallel_size: 2
+    gpu_memory_utilization: 0.85
+    dtype: auto
+    # max_model_len: # you may find it useful to set the vLLM model context length if you know this beforehand
+
+rl: grpo
+trl:
+    use_vllm: true
+    vllm_server_host: 0.0.0.0
+    vllm_server_port: 8000
+    vllm_server_timeout: 300
+```
+
+```bash
+CUDA_VISIBLE_DEVICES=2,3 axolotl vllm-serve grpo.yaml
+```
+
+Your `vLLM` instance will now attempt to spin up, and it's time to kick off training utilizing our remaining two GPUs. In another terminal, execute:
+
+```bash
+CUDA_VISIBLE_DEVICES=0,1 axolotl train grpo.yaml --num-processes 2
+```
+
+#### Reward functions
+
 GRPO uses custom reward functions and transformations. Please have them ready locally.
 
-For ex, to load OpenAI's GSM8K and use a random reward for completions:
+For example, to load OpenAI's GSM8K and use a random reward for completions:
 
 ```python
 # rewards.py
@@ -530,8 +569,6 @@ trl:
     beta: 0.001
     max_completion_length: 256
     use_vllm: True
-    vllm_device: auto
-    vllm_gpu_memory_utilization: 0.15
     num_generations: 4
     reward_funcs: ["rewards.rand_reward_func"]    # format: '{file_name}.{fn_name}'
     reward_weights: [1.0]

docs/sequence_parallelism.qmd

@@ -0,0 +1,100 @@
+---
+title: Sequence Parallelism
+description: Train with long sequences split across multiple GPUs.
+---
+
+# Sequence Parallelism
+
+Sequence parallelism is a technique that splits sequences across multiple GPUs,
+allowing you to train with very long sequences that wouldn't fit on a single GPU. Each
+GPU processes a different portion of the sequence, and the results are aggregated
+through a ring communication pattern.
+
+## When to Use Sequence Parallelism
+
+Use sequence parallelism when:
+
+- You need to train with sequence lengths that don't fit into a single GPU's memory
+- You have multiple GPUs available
+- You're experiencing OOM (Out Of Memory) errors with long sequences
+
+## Configuration
+
+To enable sequence parallelism, add the following to your configuration file:
+
+```yaml
+# Set to a divisor (> 1) of the number of GPUs available
+sequence_parallel_degree: 4  # Split sequences across 4 GPUs
+# Optional; strides across the key dimension. Larger values use more memory but should make training faster.
+heads_k_stride: 1
+# Optional; one of "varlen_llama3", "batch_ring", "batch_zigzag", "batch_stripe". Defaults to
+# "varlen_llama3" when `sample_packing: true`, and "batch_ring" otherwise.
+ring_attn_func:
+```
+
+The `sequence_parallel_degree` should be a divisor of the total number of GPUs. For example:
+
+- With 8 GPUs, valid values would be 2, 4, or 8
+- With 4 GPUs, valid values would be 2 or 4
+
+## Implementation Details
+
+When sequence parallelism is enabled:
+
+1. Each sequence is divided into equal chunks across the GPUs in a sequence parallel group
+2. The data collator handles the chunking of input_ids, attention_mask, labels, and position_ids
+3. Position IDs are adjusted to maintain proper relative positions, especially for packed sequences
+4. The trainer uses special ring communication patterns for attention operations
+
+## Requirements
+
+To use sequence parallelism, you need:
+
+- Multiple GPUs (at least 2)
+- The `ring-flash-attn` package. Install with:
+  - `pip install axolotl[ring-flash-attn]` (preferred)
+  - `pip install ring-flash-attn>=0.1.4`
+
+## Limitations
+
+- Flash attention must be enabled for this to work (`flash_attention: true` in config YAML)
+- May have a small performance overhead due to communication between GPUs
+
+## Example
+
+```yaml
+base_model: meta-llama/Llama-3-8B-Instruct
+sequence_len: 8192
+
+...
+
+sequence_parallel_degree: 4  # Split each sequence into 4 parts, one per GPU
+flash_attention: true  # Required with sequence parallelism
+# Optional; strides across the key dimension. Larger values use more memory but should make training faster.
+heads_k_stride: 1
+
+...
+```
+
+This will train the Llama 3 8B model with 8K context length, with each sequence split
+into 2 subsequences of length 4096 across 2 GPUs.
+
+## Sample Packing with Sequence Parallelism
+
+Sequence parallelism is compatible with Axolotl's sample packing functionality. When using both features together:
+
+1. Samples are first packed together
+2. The packed sequences are then divided across GPUs in the sequence parallel group
+3. Position IDs are automatically adjusted to maintain proper relative positions
+
+## Effect on Batch Size
+
+When using sequence parallelism, your effective global batch size is **divided** by the `sequence_parallel_degree`. This happens because:
+
+- Each group of `sequence_parallel_degree` GPUs works on the same batch (just different parts of each sequence)
+- The number of batches processed per step decreases
+
+For example:
+- With 8 GPUs and no sequence parallelism: 8 different batches processed per step
+- With 8 GPUs and `sequence_parallel_degree=4`: Only 2 different batches processed per step (each split across 4 GPUs)
+- If your per-GPU `micro_batch_size` is 2, the global batch size decreases from 16 to 4

examples/cerebras/btlm-ft.yml

@@ -8,10 +8,6 @@ tokenizer_type: GPT2Tokenizer
 trust_remote_code: true
 tokenizer_use_fast: true
 tokenizer_legacy: true
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
 push_dataset_to_hub:
 hf_use_auth_token: true
 datasets:
@@ -34,7 +30,6 @@ lora_alpha:
 lora_dropout:
 lora_target_modules:
 lora_target_linear:
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -58,16 +53,12 @@ learning_rate: 0.000085
 train_on_inputs: true
 group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 
 gradient_checkpointing: false
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
 
-xformers_attention:
 flash_attention: true
 sdp_attention:
 flash_optimum:
@@ -80,8 +71,6 @@ evals_per_epoch: 4
 saves_per_epoch: 1
 save_total_limit:
 
-debug:
-deepspeed:
 weight_decay: 0.1
 special_tokens:
   pad_token: "<|endoftext|>"

examples/cerebras/qlora.yml

@@ -4,7 +4,6 @@ base_model: cerebras/Cerebras-GPT-1.3B
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 push_dataset_to_hub:
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
@@ -22,7 +21,6 @@ lora_target_modules:
   - c_attn
   - c_proj
 lora_target_linear:
-lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
@@ -36,15 +34,10 @@ optimizer: paged_adamw_8bit
 torchdistx_path:
 lr_scheduler: cosine
 learning_rate: 0.0002
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
 xformers_attention: true
 flash_attention:
@@ -53,10 +46,6 @@ gptq_model_v1:
 warmup_steps: 10
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.1
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: "<|endoftext|>"

examples/code-llama/13b/lora.yml

@@ -7,7 +7,6 @@ tokenizer_type: CodeLlamaTokenizer
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test
@@ -26,7 +25,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -41,29 +39,18 @@ 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
-s2_attention:
 
 warmup_steps: 10
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   bos_token: "<s>"
   eos_token: "</s>"

examples/code-llama/13b/qlora.yml

@@ -7,7 +7,6 @@ tokenizer_type: CodeLlamaTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test
@@ -26,9 +25,7 @@ pad_to_sequence_len: true
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -43,28 +40,18 @@ optimizer: paged_adamw_32bit
 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
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   bos_token: "<s>"
   eos_token: "</s>"

examples/code-llama/34b/lora.yml

@@ -7,7 +7,6 @@ tokenizer_type: CodeLlamaTokenizer
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test
@@ -26,7 +25,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -41,29 +39,18 @@ 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
-s2_attention:
 
 warmup_steps: 10
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   bos_token: "<s>"
   eos_token: "</s>"

examples/code-llama/34b/qlora.yml

@@ -7,7 +7,6 @@ tokenizer_type: CodeLlamaTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test
@@ -26,9 +25,7 @@ pad_to_sequence_len: true
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -43,28 +40,18 @@ optimizer: paged_adamw_32bit
 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
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   bos_token: "<s>"
   eos_token: "</s>"

examples/code-llama/7b/lora.yml

@@ -7,7 +7,6 @@ tokenizer_type: CodeLlamaTokenizer
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test
@@ -26,7 +25,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -41,29 +39,18 @@ 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
-s2_attention:
 
 warmup_steps: 10
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   bos_token: "<s>"
   eos_token: "</s>"

examples/code-llama/7b/qlora.yml

@@ -7,7 +7,6 @@ tokenizer_type: CodeLlamaTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test
@@ -26,9 +25,7 @@ pad_to_sequence_len: true
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -43,28 +40,18 @@ optimizer: paged_adamw_32bit
 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
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   bos_token: "<s>"
   eos_token: "</s>"

examples/cohere/command-r-7b-qlora.yml

@@ -0,0 +1,58 @@
+base_model: CohereForAI/c4ai-command-r7b-12-2024
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+
+load_in_8bit: false
+load_in_4bit: true
+
+# huggingface repo
+chat_template: cohere
+datasets:
+  - path: cgato/SlimOrcaDedupCleaned
+    type: chat_template
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+adapter: qlora
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+sequence_len: 2048
+sample_packing: true
+eval_sample_packing: false
+pad_to_sequence_len: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch:
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/dbrx/16bit-lora.yaml

@@ -4,10 +4,6 @@ base_model: LnL-AI/dbrx-base-converted-v2
 
 trust_remote_code: true
 
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
 datasets:
   - path: tatsu-lab/alpaca
     type: alpaca
@@ -48,26 +44,20 @@ optimizer: paged_adamw_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: false  # don't use with fsdp_activation_checkpointing
 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:
 saves_per_epoch: 1
-debug:
+
 weight_decay: 0.0
 fsdp:
   - full_shard

examples/dbrx/8bit-lora.yaml

@@ -6,7 +6,6 @@ trust_remote_code: true
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 datasets:
   - path: tatsu-lab/alpaca
@@ -48,26 +47,20 @@ optimizer: paged_adamw_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: false  # don't use with fsdp_activation_checkpointing
 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:
 saves_per_epoch: 1
-debug:
+
 weight_decay: 0.0
 fsdp:
   - full_shard

examples/dbrx/fft-ds-zero3.yaml

@@ -4,10 +4,6 @@ base_model: LnL-AI/dbrx-base-converted-v2
 
 trust_remote_code: true
 
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
 datasets:
   - path: tatsu-lab/alpaca
     type: alpaca
@@ -35,25 +31,19 @@ optimizer: paged_adamw_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: true
 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:
 saves_per_epoch: 1
-debug:
+
 weight_decay: 0.0
 deepspeed: deepspeed_configs/zero3_bf16.json

examples/deepcoder/deepcoder-14B-preview-lora.yml

@@ -0,0 +1,58 @@
+base_model: agentica-org/DeepCoder-14B-Preview
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+    field_messages: messages
+    message_property_mappings:
+      role: role
+      content: content
+
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/lora-out
+
+sequence_len: 4096
+sample_packing: true
+eval_sample_packing: false
+pad_to_sequence_len: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 2
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/deepcogito/cogito-v1-preview-llama-3B-lora.yml

@@ -0,0 +1,58 @@
+base_model: deepcogito/cogito-v1-preview-llama-3B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+    field_messages: messages
+    message_property_mappings:
+      role: role
+      content: content
+
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/lora-out
+
+sequence_len: 4096
+sample_packing: true
+eval_sample_packing: false
+pad_to_sequence_len: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/deepcogito/cogito-v1-preview-qwen-14B-lora.yml

@@ -0,0 +1,58 @@
+base_model: deepcogito/cogito-v1-preview-qwen-14B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+    field_messages: messages
+    message_property_mappings:
+      role: role
+      content: content
+
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/lora-out
+
+sequence_len: 4096
+sample_packing: true
+eval_sample_packing: false
+pad_to_sequence_len: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/deepseek-v2/fft-fsdp-16b.yaml

@@ -3,10 +3,6 @@ base_model: deepseek-ai/DeepSeek-V2-Lite
 # hub_model_id: username/custom_model_name
 trust_remote_code: true
 
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
 datasets:
   - path: tatsu-lab/alpaca
     type: alpaca
@@ -31,27 +27,19 @@ optimizer: adamw_torch_fused
 lr_scheduler: cosine
 learning_rate: 2e-5
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
-early_stopping_patience:
 resume_from_checkpoint:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 100
 evals_per_epoch: 2
-eval_table_size:
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
 special_tokens:
 fsdp:

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

@@ -6,7 +6,6 @@ trust_remote_code: true
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 
 plugins:
@@ -52,27 +51,19 @@ optimizer: adamw_torch_fused
 lr_scheduler: cosine
 learning_rate: 2e-5
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
-early_stopping_patience:
 resume_from_checkpoint:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 100
 evals_per_epoch: 2
-eval_table_size:
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
 special_tokens:
 fsdp:

examples/falcon/config-7b-lora.yml

@@ -11,7 +11,6 @@ trust_remote_code: true
 load_in_8bit: true
 load_in_4bit: false
 gptq: false
-strict: false
 push_dataset_to_hub:
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
@@ -25,9 +24,7 @@ max_packed_sequence_len:
 lora_r: 16
 lora_alpha: 32
 lora_dropout: 0.0
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
@@ -41,15 +38,10 @@ optimizer: adamw_bnb_8bit
 torchdistx_path:
 lr_scheduler: cosine
 learning_rate: 0.00003
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
 xformers_attention: true
 flash_attention:
@@ -58,11 +50,7 @@ gptq_model_v1:
 warmup_steps: 40
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: "<|endoftext|>"
   bos_token: "<|endoftext|>"

examples/falcon/config-7b-qlora.yml

@@ -15,7 +15,6 @@ load_in_8bit: false
 # enable 4bit for QLoRA
 load_in_4bit: true
 gptq: false
-strict: false
 push_dataset_to_hub:
 datasets:
   - path: QingyiSi/Alpaca-CoT
@@ -38,9 +37,7 @@ lora_alpha: 16
 # 0.05 for 33B and 65B models
 lora_dropout: 0.05
 # add LoRA modules on all linear layers of the base model
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -67,10 +64,7 @@ lr_scheduler: cosine
 # - 2e-4 for 7b & 13b
 # - 1e-4 for 33b & 64b
 learning_rate: 0.0002
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 gradient_checkpointing: true
 # stop training after this many evaluation losses have increased in a row
@@ -78,7 +72,6 @@ gradient_checkpointing: true
 early_stopping_patience: 3
 resume_from_checkpoint:
 auto_resume_from_checkpoints: true
-local_rank:
 logging_steps: 1
 xformers_attention: true
 flash_attention:
@@ -87,11 +80,7 @@ gptq_model_v1:
 warmup_steps: 10
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.000001
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: "<|endoftext|>"
   bos_token: "<|endoftext|>"

examples/falcon/config-7b.yml

@@ -7,11 +7,7 @@ tokenizer_type: AutoTokenizer
 
 # required by falcon custom model code: https://huggingface.co/tiiuae/falcon-7b/tree/main
 trust_remote_code: true
-
-load_in_8bit: false
-load_in_4bit: false
 gptq: false
-strict: false
 push_dataset_to_hub:
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
@@ -25,9 +21,7 @@ max_packed_sequence_len:
 lora_r: 64
 lora_alpha: 32
 lora_dropout: 0.0
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
@@ -41,15 +35,10 @@ optimizer: adamw_bnb_8bit
 torchdistx_path:
 lr_scheduler: cosine
 learning_rate: 0.00003
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
 xformers_attention: true
 flash_attention:
@@ -58,11 +47,7 @@ gptq_model_v1:
 warmup_steps: 40
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: "<|endoftext|>"
   bos_token: "<|endoftext|>"

examples/gemma/qlora.yml

@@ -8,7 +8,6 @@ tokenizer_type: AutoTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 # huggingface repo
 datasets:
@@ -42,28 +41,16 @@ 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_ratio: 0.1
 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:

examples/gemma2/qlora.yml

@@ -7,7 +7,6 @@ tokenizer_type: AutoTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 # huggingface repo
 chat_template: gemma
@@ -48,28 +47,16 @@ optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_ratio: 0.1
 evals_per_epoch:
-eval_table_size:
-eval_max_new_tokens: 128
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:

examples/gemma2/reward-model.yaml

@@ -6,10 +6,6 @@ tokenizer_type: AutoTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
 reward_model: true
 chat_template: gemma
 datasets:
@@ -38,8 +34,6 @@ optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bf16: true
 fp16:
 tf32: true
@@ -47,21 +41,12 @@ tf32: true
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_ratio: 0.1
 evals_per_epoch:
-eval_table_size:
-eval_max_new_tokens: 128
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:

examples/gemma3/gemma-3-1b-qlora.yml

@@ -0,0 +1,66 @@
+base_model: google/gemma-3-1b-it
+# optionally might have model_type or tokenizer_type
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+# gemma3 doesn't seem to play nice with ddp
+ddp_find_unused_parameters: true
+
+load_in_8bit: false
+load_in_4bit: true
+
+# huggingface repo
+chat_template: gemma3
+datasets:
+  - path: cgato/SlimOrcaDedupCleaned
+    type: chat_template
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+adapter: qlora
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+sequence_len: 2048
+sample_packing: true
+eval_sample_packing: false
+pad_to_sequence_len: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch:
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/gemma3/gemma-3-4b-qlora.yml

@@ -0,0 +1,60 @@
+base_model: google/gemma-3-4b-it
+
+load_in_4bit: true
+
+# gemma3 doesn't seem to play nice with ddp
+ddp_find_unused_parameters: true
+
+chat_template: gemma3
+datasets:
+  - path: cgato/SlimOrcaDedupCleaned
+    type: chat_template
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./outputs/out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+pad_to_sequence_len: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules: 'language_model.model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: true
+fp16:
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+logging_steps: 1
+flash_attention: true
+eager_attention:
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0

examples/gemma3/gemma-3-4b-vision-qlora.yml

@@ -0,0 +1,62 @@
+base_model: google/gemma-3-4b-it
+processor_type: AutoProcessor
+
+load_in_4bit: true
+
+# these 3 lines are needed for now to handle vision chat templates w images
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+# gemma3 doesn't seem to play nice with ddp
+ddp_find_unused_parameters: true
+
+chat_template: gemma3
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+    field_messages: messages
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./outputs/out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+pad_to_sequence_len: false
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules: 'language_model.model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: true
+fp16:
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+logging_steps: 1
+flash_attention: true
+eager_attention:
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0

examples/gptj/qlora.yml

@@ -4,7 +4,6 @@ base_model: EleutherAI/gpt-j-6b
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 push_dataset_to_hub:
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
@@ -18,9 +17,7 @@ max_packed_sequence_len:
 lora_r: 8
 lora_alpha: 32
 lora_dropout: 0.05
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
@@ -34,15 +31,10 @@ optimizer: paged_adamw_8bit
 torchdistx_path:
 lr_scheduler: cosine
 learning_rate: 0.0001
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
 xformers_attention: true
 flash_attention:
@@ -51,10 +43,6 @@ gptq_model_v1:
 warmup_steps: 10
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.1
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: "<|endoftext|>"

examples/jamba/qlora.yaml

@@ -6,7 +6,6 @@ trust_remote_code: true
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test
@@ -40,26 +39,18 @@ optimizer: paged_adamw_8bit
 lr_scheduler: cosine
 learning_rate: 0.00001
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
 evals_per_epoch:
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
 special_tokens:

examples/jamba/qlora_deepspeed.yaml

@@ -5,7 +5,6 @@ trust_remote_code: true
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test
@@ -39,26 +38,20 @@ optimizer: paged_adamw_8bit
 lr_scheduler: cosine
 learning_rate: 0.00001
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
 evals_per_epoch:
 saves_per_epoch: 1
-debug:
+
 deepspeed: deepspeed_configs/zero2.json
 weight_decay: 0.0
 special_tokens:

examples/jamba/qlora_fsdp_large.yaml

@@ -5,7 +5,6 @@ tokenizer_type: AutoTokenizer
 # hub_model_id: username/custom_model_name
 
 load_in_4bit: true
-strict: false
 use_tensorboard: true
 chat_template: jamba
 datasets:
@@ -39,8 +38,6 @@ optimizer: adamw_torch_fused
 lr_scheduler: cosine
 learning_rate: 0.00001
 
-train_on_inputs: false
-group_by_length: false
 bf16: true
 tf32: true
 

examples/jeopardy-bot/config.yml

@@ -33,13 +33,9 @@ optimizer: adamw_bnb_8bit
 torchdistx_path:
 lr_scheduler: cosine
 learning_rate: 0.00003
-train_on_inputs: false
-group_by_length: false
 bf16: auto
 tf32: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 5
 xformers_attention: true
 flash_attention:
@@ -48,11 +44,7 @@ gptq_model_v1:
 warmup_steps: 20
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.1
-fsdp:
-fsdp_config:
 tokens:
   bos_token: "<s>"
   eos_token: "</s>"

examples/llama-2/fft_optimized.yml

@@ -5,10 +5,6 @@ tokenizer_type: LlamaTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
 datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
@@ -26,7 +22,6 @@ lora_r:
 lora_alpha:
 lora_dropout:
 lora_target_linear:
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -41,18 +36,12 @@ 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
 flash_attn_cross_entropy: false
 flash_attn_rms_norm: true
@@ -61,11 +50,8 @@ flash_attn_fuse_mlp: true
 
 warmup_steps: 100
 evals_per_epoch: 4
-eval_table_size:
 saves_per_epoch: 1
-debug:
+
 deepspeed: #deepspeed_configs/zero2.json # multi-gpu only
 weight_decay: 0.1
-fsdp:
-fsdp_config:
 special_tokens:

examples/llama-2/gptq-lora.yml

@@ -10,9 +10,6 @@ gptq_disable_exllama: true
 
 tokenizer_use_fast: true
 tokenizer_legacy: true
-load_in_8bit: false
-load_in_4bit: false
-strict: false
 push_dataset_to_hub:
 hf_use_auth_token: true
 datasets:
@@ -33,7 +30,6 @@ lora_target_modules:
   - q_proj
   - v_proj
 lora_target_linear:
-lora_fan_in_fan_out:
 wandb_project:
 wandb_watch:
 wandb_name:
@@ -50,26 +46,19 @@ torchdistx_path:
 lr_scheduler: cosine
 lr_quadratic_warmup: true
 learning_rate: 0.000017
-train_on_inputs: false
-group_by_length: false
 bf16: false
 fp16: false
 float16: true
 tf32: true
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention:
 sdp_attention:
 flash_optimum:
 warmup_steps: 100
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.1
 special_tokens:
   bos_token: "<s>"

examples/llama-2/lisa.yml

@@ -5,10 +5,6 @@ tokenizer_type: LlamaTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
     type: alpaca
@@ -26,7 +22,6 @@ lora_r:
 lora_alpha:
 lora_dropout:
 lora_target_linear:
-lora_fan_in_fan_out:
 
 lisa_n_layers: 4
 lisa_step_interval: 20
@@ -45,18 +40,12 @@ optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 5e-5 # recommendation from lisa paper for 7b
 
-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
 flash_attn_cross_entropy: false
 flash_attn_rms_norm: true
@@ -65,13 +54,8 @@ flash_attn_fuse_mlp: true
 
 warmup_steps: 100
 evals_per_epoch: 4
-eval_table_size:
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.1
-fsdp:
-fsdp_config:
 special_tokens:
   bos_token: "<s>"
   eos_token: "</s>"

examples/llama-2/loftq.yml

@@ -5,10 +5,6 @@ tokenizer_type: LlamaTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
 datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
@@ -26,7 +22,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 peft:
   loftq_config:
     loftq_bits: 4
@@ -44,29 +39,16 @@ 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
-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:

examples/llama-2/lora.yml

@@ -7,7 +7,6 @@ tokenizer_type: LlamaTokenizer
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test
@@ -26,7 +25,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -41,29 +39,16 @@ 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
-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:

examples/llama-2/qlora-fsdp.yml

@@ -7,7 +7,6 @@ tokenizer_type: LlamaTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: yahma/alpaca-cleaned
@@ -26,9 +25,7 @@ pad_to_sequence_len: true
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -43,28 +40,19 @@ optimizer: adamw_torch_fused
 lr_scheduler: cosine
 learning_rate: 0.00001
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: 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:
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
 fsdp:
   - full_shard

examples/llama-2/qlora.yml

@@ -7,7 +7,6 @@ tokenizer_type: LlamaTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test
@@ -26,9 +25,7 @@ pad_to_sequence_len: true
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -43,27 +40,16 @@ optimizer: paged_adamw_32bit
 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:
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:

examples/llama-2/relora.yml

@@ -5,7 +5,6 @@ tokenizer_type: LlamaTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
@@ -24,9 +23,7 @@ pad_to_sequence_len: true
 lora_r: 8
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 relora_steps: 150
 relora_warmup_steps: 10
@@ -45,28 +42,18 @@ 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
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   bos_token: "<s>"
   eos_token: "</s>"

examples/llama-3-vision/lora-11b.yaml

@@ -4,7 +4,6 @@ processor_type: AutoProcessor
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 
-strict: false
 
 # these 3 lines are needed for now to handle vision chat templates w images
 skip_prepare_dataset: true
@@ -45,14 +44,11 @@ optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bf16: true
 fp16:
 tf32: true
 
 gradient_checkpointing: true
-local_rank:
 logging_steps: 1
 flash_attention: true
 eager_attention:
@@ -60,8 +56,4 @@ eager_attention:
 warmup_ratio: 0.1
 evals_per_epoch: 1
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:

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

@@ -9,7 +9,6 @@ liger_rms_norm: true
 liger_glu_activation: true
 liger_fused_linear_cross_entropy: true
 
-strict: false
 
 chat_template: llama3
 datasets:
@@ -42,27 +41,19 @@ optimizer: adamw_torch_fused
 lr_scheduler: cosine
 learning_rate: 2e-5
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
-early_stopping_patience:
 resume_from_checkpoint:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 100
 evals_per_epoch: 2
-eval_table_size:
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
 fsdp:
   - full_shard

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

@@ -2,10 +2,6 @@ base_model: NousResearch/Meta-Llama-3.1-8B
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
 datasets:
   - path: tatsu-lab/alpaca
     type: alpaca
@@ -30,29 +26,19 @@ optimizer: paged_adamw_8bit
 lr_scheduler: cosine
 learning_rate: 2e-5
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
-early_stopping_patience:
 resume_from_checkpoint:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 100
 evals_per_epoch: 2
-eval_table_size:
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: <|end_of_text|>

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

@@ -7,7 +7,6 @@ tokenizer_type: AutoTokenizer
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 chat_template: llama3
 rl: dpo
@@ -42,7 +41,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -57,28 +55,15 @@ 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
-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:

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

@@ -7,7 +7,6 @@ tokenizer_type: AutoTokenizer
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 chat_template: llama3
 datasets:
@@ -37,7 +36,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -52,30 +50,17 @@ 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
-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:
    pad_token: <|end_of_text|>

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

@@ -7,7 +7,6 @@ tokenizer_type: AutoTokenizer
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 chat_template: llama3
 rl: dpo
@@ -58,7 +57,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -73,28 +71,15 @@ 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
-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:

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

@@ -7,7 +7,6 @@ tokenizer_type: AutoTokenizer
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test
@@ -19,7 +18,6 @@ val_set_size: 0.0
 output_dir: ./outputs/lora-out
 
 dataset_exact_deduplication: true
-test_value: true
 
 sequence_len: 4096
 sample_packing: true
@@ -32,7 +30,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 lora_modules_to_save:
   - embed_tokens
   - lm_head
@@ -50,30 +47,17 @@ 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
-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:
    pad_token: <|end_of_text|>

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

@@ -2,10 +2,6 @@ base_model: NousResearch/Llama-3.2-1B
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
     type: alpaca
@@ -24,7 +20,6 @@ lora_r: 16
 lora_alpha: 32
 # Currently, we don't support dropout with our custom Triton kernels
 # lora_dropout: 0.05
-lora_fan_in_fan_out:
 lora_target_modules:
   - gate_proj
   - down_proj
@@ -53,18 +48,12 @@ optimizer: adamw_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 loss_watchdog_threshold: 5.0
@@ -73,10 +62,6 @@ loss_watchdog_patience: 3
 warmup_steps: 10
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: "<|end_of_text|>"

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

@@ -2,10 +2,6 @@ base_model: NousResearch/Llama-3.2-1B
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
     type: alpaca
@@ -24,7 +20,6 @@ pad_to_sequence_len: true
 lora_r: 16
 lora_alpha: 32
 lora_dropout: 0.05
-lora_fan_in_fan_out:
 lora_target_modules:
   - gate_proj
   - down_proj
@@ -47,18 +42,12 @@ optimizer: adamw_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 loss_watchdog_threshold: 5.0
@@ -67,11 +56,9 @@ loss_watchdog_patience: 3
 warmup_steps: 10
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
+
 deepspeed: deepspeed_configs/zero3.json
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: "<|end_of_text|>"
 

examples/llama-3/lora-1b-sample-packing-sequentially.yml

@@ -0,0 +1,65 @@
+base_model: meta-llama/Llama-3.2-1B
+# optionally might have model_type or tokenizer_type
+model_type: LlamaForCausalLM
+tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: true
+load_in_4bit: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.0
+output_dir: ./outputs/lora-out
+
+test_value: true
+
+sequence_len: 4096
+sample_packing: true
+sample_packing_sequentially: true
+curriculum_sampling: true
+eval_sample_packing: false
+pad_to_sequence_len: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+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: 2
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+  pad_token: <|end_of_text|>

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

@@ -2,10 +2,6 @@ base_model: NousResearch/Llama-3.2-1B
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
     type: alpaca
@@ -24,7 +20,6 @@ pad_to_sequence_len: true
 lora_r: 16
 lora_alpha: 32
 lora_dropout: 0.05
-lora_fan_in_fan_out:
 lora_target_modules:
   - gate_proj
   - down_proj
@@ -47,18 +42,12 @@ optimizer: adamw_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 loss_watchdog_threshold: 5.0
@@ -67,10 +56,6 @@ loss_watchdog_patience: 3
 warmup_steps: 10
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: "<|end_of_text|>"

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

@@ -7,7 +7,6 @@ tokenizer_type: AutoTokenizer
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test
@@ -27,7 +26,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 lora_modules_to_save:
   - embed_tokens
   - lm_head
@@ -45,30 +43,17 @@ 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
-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:
    pad_token: <|end_of_text|>

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

@@ -4,7 +4,6 @@ base_model: meta-llama/Llama-3.2-1B
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 rl: kto
 rl_beta: 0.5
@@ -32,7 +31,6 @@ lora_r: 32
 lora_alpha: 64
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -47,31 +45,19 @@ optimizer: adamw_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 20
 evals_per_epoch: 4
-eval_table_size:
-eval_max_new_tokens: 128
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: "<|end_of_text|>"

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

@@ -4,7 +4,6 @@ base_model: NousResearch/Llama-3.2-1B
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
@@ -24,7 +23,6 @@ pad_to_sequence_len: true
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
-lora_fan_in_fan_out:
 lora_target_modules:
   - gate_proj
   - down_proj
@@ -47,18 +45,12 @@ optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 loss_watchdog_threshold: 5.0
@@ -66,13 +58,7 @@ loss_watchdog_patience: 3
 
 warmup_steps: 10
 evals_per_epoch: 4
-eval_table_size:
-eval_max_new_tokens: 128
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: "<|end_of_text|>"

examples/llama-3/qlora-fsdp-405b.yaml

@@ -5,7 +5,6 @@ tokenizer_type: AutoTokenizer
 # hub_model_id: username/custom_model_name
 
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: tatsu-lab/alpaca
@@ -24,7 +23,6 @@ pad_to_sequence_len: true
 lora_r: 16
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_modules:
 lora_target_linear: true
 
 gradient_accumulation_steps: 4
@@ -34,8 +32,6 @@ optimizer: adamw_torch_fused
 lr_scheduler: cosine
 learning_rate: 0.00001
 
-train_on_inputs: false
-group_by_length: false
 bf16: true
 tf32: true
 

examples/llama-3/qlora-fsdp-70b.yaml

@@ -7,7 +7,6 @@ tokenizer_type: AutoTokenizer  # PreTrainedTokenizerFast
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: tatsu-lab/alpaca
@@ -26,9 +25,7 @@ pad_to_sequence_len: true
 lora_r: 8
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -43,28 +40,19 @@ optimizer: adamw_torch_fused
 lr_scheduler: cosine
 learning_rate: 0.00001
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: 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:
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
 fsdp:
   - full_shard

examples/llama-3/qlora.yml

@@ -7,7 +7,6 @@ tokenizer_type: AutoTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: aaditya/alpaca_subset_1
@@ -26,9 +25,7 @@ pad_to_sequence_len: true
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -43,28 +40,17 @@ optimizer: paged_adamw_32bit
 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:
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: "<|end_of_text|>"

examples/llama-4/README.md

@@ -0,0 +1,28 @@
+# Llama 4 by Meta AI
+
+## Flash Attention vs Flex Attention
+
+While Flash Attention to support is "enabled" for Llama-4, the upstream implementation is not correct and usage of Flex Attention is recommended.
+
+## Available Examples
+
+### Llama 4 Scout 17Bx16Experts (109B)
+
+Flex Attention
+- [Text Single GPU (H100) QLoRA](./scout-qlora-single-h100-flex.yaml)
+- [Text Multi GPU QLoRA w/ FSDP2](./scout-qlora-flexattn-fsdp2.yaml)
+
+[//]: # (Flash Attention (Do not use))
+
+[//]: # (- [Multi-Modal/Vision QLoRA w/ FSDP1](./scout-vision-qlora-fsdp.yaml))
+
+[//]: # (- [Text Single GPU (H100) QLoRA](./scout-qlora-single-h100.yaml))
+
+[//]: # (- [Text Multi GPU QLoRA w/ FSDP1](./scout-qlora-fsdp1.yaml))
+
+Our Single H100 implementation for Llama 4 Scout uses only 64.5GB VRAM for post-training with 4k context length @ 519 tokens/second. [WandB logs here](https://wandb.ai/axolotl-ai/llama4-flexattn-qlora/runs/wpie7dkj)
+Multi-GPU (4xH100) for Llama 4 Scout uses 62.8GB VRAM/GPU @ 4k contenxt length @ 280tps/gpu, [WandB logs here](https://wandb.ai/axolotl-ai/llama4-flexattn-qlora/runs/2lkezdj8)
+
+### Llama 4 Maverick 17Bx128Experts (400B)
+
+Coming Soon

examples/llama-4/do-no-use-fa2/maverick-qlora-fsdp1.yaml

@@ -0,0 +1,88 @@
+base_model: axolotl-quants/Llama-4-Maverick-17B-128E-Linearized-bnb-nf4-bf16
+model_type: Llama4ForConditionalGeneration
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+
+liger_glu_activation: true
+liger_rms_norm: true
+liger_layer_norm: true
+
+llama4_linearized_experts: true
+load_in_4bit: true
+adapter: qlora
+lora_r: 32
+lora_alpha: 64
+lora_target_modules:
+  - self_attn.q_proj
+  - self_attn.k_proj
+  - self_attn.v_proj
+  - self_attn.o_proj
+  - shared_expert.gate_proj
+  - shared_expert.up_proj
+  - shared_expert.down_proj
+  # - experts.gate_projs.[0-9]+$
+  # - experts.up_projs.[0-9]+$
+  # - experts.down_projs.[0-9]+$
+lora_modules_to_save:
+# - lm_head
+# - embed_tokens
+
+chat_template: llama4
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 1e-4
+
+bf16: true
+tf32: true
+
+logging_steps: 1
+flash_attention: true
+
+gradient_checkpointing: offload
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+
+warmup_steps: 20
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+fsdp:
+  - auto_wrap
+  - full_shard
+fsdp_config:
+  fsdp_transformer_layer_cls_to_wrap: Llama4TextDecoderLayer
+  fsdp_limit_all_gathers: true
+  fsdp_sync_module_states: true
+  fsdp_offload_params: true
+  fsdp_use_orig_params: false
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_sharding_strategy: FULL_SHARD
+special_tokens:
+  pad_token: <|finetune_right_pad_id|>
+  eos_token: <|eot|>

examples/llama-4/do-no-use-fa2/scout-qlora-fsdp1.yaml

@@ -0,0 +1,92 @@
+base_model: axolotl-quants/Llama-4-Scout-17B-16E-Linearized-bnb-nf4-bf16
+model_type: Llama4ForConditionalGeneration
+  # Automatically upload checkpoint and final model to HF
+  # hub_model_id: username/custom_model_name
+
+
+# torch_compile: true
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+
+liger_glu_activation: true
+liger_rms_norm: true
+liger_layer_norm: true
+
+llama4_linearized_experts: true
+load_in_4bit: true
+adapter: qlora
+lora_r: 32
+lora_alpha: 64
+lora_target_modules:
+  - self_attn.q_proj
+  - self_attn.k_proj
+  - self_attn.v_proj
+  - self_attn.o_proj
+  - shared_expert.gate_proj
+  - shared_expert.up_proj
+  - shared_expert.down_proj
+    # - experts.gate_projs.[0-9]+$
+    # - experts.up_projs.[0-9]+$
+    # - experts.down_projs.[0-9]+$
+lora_modules_to_save:
+  - lm_head
+  - embed_tokens
+
+chat_template: llama4
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+bf16: true
+tf32: true
+
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 100
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+fsdp:
+  - auto_wrap
+  - full_shard
+fsdp_config:
+  fsdp_transformer_layer_cls_to_wrap: Llama4TextDecoderLayer
+  fsdp_limit_all_gathers: true
+  fsdp_sync_module_states: true
+  fsdp_offload_params: true
+  fsdp_use_orig_params: false
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_activation_checkpointing: true
+special_tokens:
+  pad_token: <|finetune_right_pad_id|>
+  eos_token: <|eot|>

examples/llama-4/do-no-use-fa2/scout-qlora-single-h100.yaml

@@ -0,0 +1,85 @@
+base_model: axolotl-quants/Llama-4-Scout-17B-16E-Linearized-bnb-nf4-bf16
+model_type: Llama4ForConditionalGeneration
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+
+liger_glu_activation: true
+liger_rms_norm: true
+liger_layer_norm: true
+
+llama4_linearized_experts: true
+load_in_4bit: true
+adapter: qlora
+lora_r: 32
+lora_alpha: 64
+lora_target_modules:
+  - self_attn.q_proj
+  - self_attn.k_proj
+  - self_attn.v_proj
+  - self_attn.o_proj
+  - shared_expert.gate_proj
+  - shared_expert.up_proj
+  - shared_expert.down_proj
+  # - experts.gate_projs.[0-9]+$
+  # - experts.up_projs.[0-9]+$
+  # - experts.down_projs.[0-9]+$
+lora_modules_to_save:
+  # - lm_head
+  # - embed_tokens
+
+lora_mlp_kernel: true
+lora_qkv_kernel: true
+lora_o_kernel: true
+
+chat_template: llama4
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+sequence_len: 4096  # up to 8k will work on a single H100
+sample_packing: true
+pad_to_sequence_len: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 1e-4
+
+bf16: true
+tf32: true
+
+logging_steps: 1
+flash_attention: true
+
+gradient_checkpointing: offload
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+
+warmup_steps: 20
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+  pad_token: <|finetune_right_pad_id|>
+  eos_token: <|eot|>

examples/llama-4/do-no-use-fa2/scout-vision-qlora-fsdp.yaml

@@ -0,0 +1,88 @@
+base_model: axolotl-quants/Llama-4-Scout-17B-16E-Linearized-bnb-nf4-bf16
+model_type: Llama4ForConditionalGeneration
+processor_type: Llama4Processor
+  # Automatically upload checkpoint and final model to HF
+  # hub_model_id: username/custom_model_name
+
+
+# these 3 lines are needed for now to handle vision chat templates w images
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+sequence_len: 4096
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+
+liger_glu_activation: true
+liger_rms_norm: true
+liger_layer_norm: true
+
+llama4_linearized_experts: true  # use Axolotl's customized model
+load_in_4bit: true
+adapter: qlora
+lora_r: 32
+lora_alpha: 64
+lora_target_modules:
+  - self_attn.q_proj
+  - self_attn.k_proj
+  - self_attn.v_proj
+  - self_attn.o_proj
+  - shared_expert.gate_proj
+  - shared_expert.up_proj
+  - shared_expert.down_proj
+  - vision_adapter.mlp.fc1
+  - vision_adapter.mlp.fc2
+  # - experts.gate_projs.[0-9]+$
+  # - experts.up_projs.[0-9]+$
+  # - experts.down_projs.[0-9]+$
+lora_modules_to_save:
+  - lm_head
+  - embed_tokens
+
+chat_template: llama4
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+    field_messages: messages
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+bf16: true
+tf32: true
+
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 100
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+fsdp:
+  - auto_wrap
+  - full_shard
+fsdp_config:
+  fsdp_transformer_layer_cls_to_wrap: Llama4TextDecoderLayer
+  fsdp_limit_all_gathers: true
+  fsdp_sync_module_states: true
+  fsdp_offload_params: true
+  fsdp_use_orig_params: false
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_activation_checkpointing: true
+special_tokens:
+  pad_token: <|finetune_right_pad_id|>
+  eos_token: <|eot|>

examples/llama-4/scout-qlora-flexattn-fsdp2.yaml

@@ -0,0 +1,86 @@
+base_model: axolotl-quants/Llama-4-Scout-17B-16E-Linearized-bnb-nf4-bf16
+model_type: Llama4ForConditionalGeneration
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+
+liger_glu_activation: true
+liger_rms_norm: true
+liger_layer_norm: true
+
+llama4_linearized_experts: true
+load_in_4bit: true
+adapter: qlora
+lora_r: 32
+lora_alpha: 64
+lora_target_modules:
+  - self_attn.q_proj
+  - self_attn.k_proj
+  - self_attn.v_proj
+  - self_attn.o_proj
+  - shared_expert.gate_proj
+  - shared_expert.up_proj
+  - shared_expert.down_proj
+  # - experts.gate_projs.[0-9]+$
+  # - experts.up_projs.[0-9]+$
+  # - experts.down_projs.[0-9]+$
+lora_modules_to_save:
+  # - lm_head
+  # - embed_tokens
+
+chat_template: llama4
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+gradient_accumulation_steps: 1
+micro_batch_size: 2
+num_epochs: 3
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 1e-4
+
+bf16: true
+tf32: true
+
+logging_steps: 1
+flex_attention: true
+flex_attn_compile_kwargs:
+  dynamic: false
+  mode: max-autotune-no-cudagraphs
+
+warmup_steps: 10
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+fsdp:
+  - auto_wrap
+  - full_shard
+fsdp_config:
+  fsdp_version: 2
+  fsdp_offload_params: false
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: Llama4TextDecoderLayer
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_reshard_after_forward: true
+  fsdp_activation_checkpointing: true
+special_tokens:
+  pad_token: <|finetune_right_pad_id|>
+  eos_token: <|eot|>

examples/llama-4/scout-qlora-single-h100-flex.yaml

@@ -0,0 +1,85 @@
+base_model: axolotl-quants/Llama-4-Scout-17B-16E-Linearized-bnb-nf4-bf16
+model_type: Llama4ForConditionalGeneration
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+liger_glu_activation: true
+liger_rms_norm: true
+liger_layer_norm: true
+cut_cross_entropy: true
+
+llama4_linearized_experts: true  # needed with custom linearized experts model
+load_in_4bit: true
+adapter: qlora
+lora_r: 32
+lora_alpha: 64
+lora_target_modules:
+  - self_attn.q_proj
+  - self_attn.k_proj
+  - self_attn.v_proj
+  - self_attn.o_proj
+  - shared_expert.gate_proj
+  - shared_expert.up_proj
+  - shared_expert.down_proj
+  # - experts.gate_projs.[0-9]+$  # optionally train the moe experts
+  # - experts.up_projs.[0-9]+$
+  # - experts.down_projs.[0-9]+$
+lora_modules_to_save:
+  # - lm_head  # needed if modifying vocabulary
+  # - embed_tokens
+
+lora_mlp_kernel: true
+lora_qkv_kernel: true
+lora_o_kernel: true
+
+chat_template: llama4
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+sequence_len: 4096  # up to 8k will work on a single H100
+sample_packing: true
+pad_to_sequence_len: true
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 1e-4
+
+bf16: true
+tf32: true
+
+torch_compile: true
+flex_attention: true
+flex_attn_compile_kwargs:
+  dynamic: false
+  mode: max-autotune-no-cudagraphs
+
+gradient_checkpointing: offload
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+
+logging_steps: 1
+warmup_steps: 20
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+weight_decay: 0.0
+special_tokens:
+  pad_token: <|finetune_right_pad_id|>
+  eos_token: <|eot|>

examples/llama-4/scout-vision-qlora-fsdp2-flex.yaml

@@ -0,0 +1,89 @@
+base_model: axolotl-quants/Llama-4-Scout-17B-16E-Linearized-bnb-nf4-bf16
+model_type: Llama4ForConditionalGeneration
+processor_type: Llama4Processor
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+# these 3 lines are needed for now to handle vision chat templates w images
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+sequence_len: 4096
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+
+liger_glu_activation: true
+liger_rms_norm: true
+liger_layer_norm: true
+
+llama4_linearized_experts: true  # use Axolotl's customized model
+load_in_4bit: true
+adapter: qlora
+lora_r: 32
+lora_alpha: 64
+lora_target_modules:
+  - self_attn.q_proj
+  - self_attn.k_proj
+  - self_attn.v_proj
+  - self_attn.o_proj
+  - shared_expert.gate_proj
+  - shared_expert.up_proj
+  - shared_expert.down_proj
+  - vision_adapter.mlp.fc1
+  - vision_adapter.mlp.fc2
+  # - experts.gate_projs.[0-9]+$
+  # - experts.up_projs.[0-9]+$
+  # - experts.down_projs.[0-9]+$
+lora_modules_to_save:
+  - lm_head
+  - embed_tokens
+
+chat_template: llama4
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+    field_messages: messages
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 1e-4
+
+bf16: true
+tf32: true
+
+logging_steps: 1
+flex_attention: true
+flex_attn_compile_kwargs:
+  dynamic: false
+  mode: max-autotune-no-cudagraphs
+
+warmup_steps: 10
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+fsdp:
+  - auto_wrap
+  - full_shard
+fsdp_config:
+  fsdp_version: 2
+  fsdp_offload_params: false
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: Llama4TextDecoderLayer
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_reshard_after_forward: true
+  fsdp_activation_checkpointing: true
+special_tokens:
+  pad_token: <|finetune_right_pad_id|>
+  eos_token: <|eot|>