0.8.1 version

* llama4 support for linearized experts

* clean up fsdp2 sharding to prevent hang

* add yaml config

* cleanup example [skip ci]

.github/workflows/base.yml

@@ -40,6 +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
@@ -61,7 +79,7 @@ jobs:
         uses: docker/build-push-action@v4
         with:
           context: .
-          file: ./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

@@ -24,13 +24,13 @@ jobs:
             cuda_version: 12.4.1
             python_version: "3.11"
             pytorch: 2.5.1
-            axolotl_extras:
-            is_latest: true
+            axolotl_extras: vllm
           - 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
@@ -87,6 +87,11 @@ jobs:
             python_version: "3.11"
             pytorch: 2.5.1
             axolotl_extras:
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.6.0
+            axolotl_extras:
             is_latest: true
     runs-on: axolotl-gpu-runner
     steps:

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

@@ -4,6 +4,10 @@ on:
   pull_request:
     paths:
       - 'tests/e2e/multigpu/*.py'
+      - 'requirements.txt'
+      - 'setup.py'
+      - 'pyproject.toml'
+      - '.github/workflows/multi-gpu-e2e.yml'
   workflow_dispatch:
   schedule:
     - cron: '0 0 * * 1,4'  # Runs at 00:00 UTC every monday & thursday
@@ -20,25 +24,25 @@ 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"
             pytorch: 2.4.1
-            axolotl_extras:
+            axolotl_extras:  # no vllm support for 2.4.1
             num_gpus: 2
             nightly_build: "true"
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
             pytorch: 2.5.1
-            axolotl_extras:
-            num_gpus: 2
-            nightly_build: "true"
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.6.0
-            axolotl_extras:
+            axolotl_extras: vllm
             num_gpus: 2
             nightly_build: "true"
     runs-on: [self-hosted, modal]

.github/workflows/nightlies.yml

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

.github/workflows/precommit-autoupdate.yml

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

.github/workflows/pypi.yml

@@ -40,7 +40,7 @@ jobs:
 
       - name: Install dependencies
         run: |
-          pip3 install wheel packaging
+          pip3 install wheel packaging==23.2
           pip3 install --no-build-isolation -e .
           pip3 install -r requirements-dev.txt -r requirements-tests.txt
 

.github/workflows/tests-nightly.yml

@@ -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:
@@ -42,11 +51,11 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging setuptools wheel
+          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
 
       - name: Install PyTorch
         run: |
-          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 install --upgrade packaging
+          pip3 show torch
           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 tests/patched/
+          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: |
@@ -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
@@ -74,7 +74,7 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging setuptools wheel
+          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
 
       - name: Install PyTorch
         run: |
@@ -96,10 +96,15 @@ 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 -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: |
@@ -136,7 +141,7 @@ 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
@@ -147,7 +152,7 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging setuptools setuptools_scm build wheel
+          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 setuptools_scm build wheel
 
       - name: Install PyTorch
         run: |
@@ -170,10 +175,14 @@ 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: |
@@ -202,9 +211,9 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.5.1
+            pytorch: 2.6.0
             num_gpus: 1
-            axolotl_extras:
+            axolotl_extras: vllm
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -227,7 +236,7 @@ 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'
@@ -249,9 +258,9 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.5.1
             num_gpus: 1
-            axolotl_extras:
+            axolotl_extras: vllm
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -274,4 +283,4 @@ 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

.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/PyCQA/pylint
-    rev: v3.3.0
+-   repo: https://github.com/pylint-dev/pylint
+    rev: v3.3.6
     hooks:
     - id: pylint
 -   repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.3.0
+    rev: v1.15.0
     hooks:
     - id: mypy
       additional_dependencies:
@@ -36,7 +36,7 @@ repos:
             'pydantic>=2.5.3',
         ]
 -   repo: https://github.com/PyCQA/bandit
-    rev: 1.7.5
+    rev: 1.8.3
     hooks:
     -   id: bandit
         args: [

README.md

@@ -19,9 +19,6 @@
     <br/>
     <img src="https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/tests-nightly.yml/badge.svg" alt="tests-nightly">
     <img src="https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/multi-gpu-e2e.yml/badge.svg" alt="multigpu-semi-weekly tests">
-    <a href="https://www.phorm.ai/query?projectId=e315ba4a-4e14-421f-ab05-38a1f9076f25">
-    <img alt="phorm.ai" src="https://img.shields.io/badge/Phorm-Ask_AI-%23F2777A.svg?&logo=data:image/svg+xml;base64,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">
-  </a>
 </p>
 
 Axolotl is a tool designed to streamline post-training for various AI models.
@@ -50,13 +47,15 @@ Features:
 ## 🚀 Quick Start
 
 **Requirements**:
+
 - NVIDIA GPU (Ampere or newer for `bf16` and Flash Attention) or AMD GPU
 - Python 3.11
 - PyTorch ≥2.4.1
 
 ### Installation
 
-```shell
+```bash
+pip3 install -U packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation axolotl[flash-attn,deepspeed]
 
 # Download example axolotl configs, deepspeed configs
@@ -68,7 +67,7 @@ Other installation approaches are described [here](https://axolotl-ai-cloud.gith
 
 ### Your First Fine-tune
 
-```shell
+```bash
 # Fetch axolotl examples
 axolotl fetch examples
 
@@ -98,6 +97,7 @@ That's it! Check out our [Getting Started Guide](https://axolotl-ai-cloud.github
 - [Multi-GPU Training](https://axolotl-ai-cloud.github.io/axolotl/docs/multi-gpu.html)
 - [Multi-Node Training](https://axolotl-ai-cloud.github.io/axolotl/docs/multi-node.html)
 - [Multipacking](https://axolotl-ai-cloud.github.io/axolotl/docs/multipack.html)
+- [API Reference](https://axolotl-ai-cloud.github.io/axolotl/docs/api/) - Auto-generated code documentation
 - [FAQ](https://axolotl-ai-cloud.github.io/axolotl/docs/faq.html) - Frequently asked questions
 
 ## 🤝 Getting Help

_quarto.yml

@@ -1,12 +1,188 @@
 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: "Fine-tuning"
+  description: "We make fine-tuning accessible, scalable, and fun"
   favicon: favicon.jpg
+
   navbar:
-    title: Axolotl
+    logo: image/axolotl_logo_digital_white.svg
+    title: false
     background: dark
     pinned: false
     collapse: false
@@ -25,33 +201,79 @@ website:
       contents:
         - text: Home
           href: index.qmd
-        - section: "How-To Guides"
+
+        - section: "Getting Started"
           contents:
-          # TODO Edit folder structure after we have more docs.
             - docs/getting-started.qmd
             - docs/installation.qmd
-            - docs/debugging.qmd
             - docs/inference.qmd
-            - docs/multipack.qmd
-            - docs/fsdp_qlora.qmd
-            - docs/input_output.qmd
-            - docs/rlhf.qmd
-            - docs/nccl.qmd
-            - docs/mac.qmd
-            - docs/multi-gpu.qmd
-            - docs/multi-node.qmd
-            - docs/unsloth.qmd
-            - docs/amd_hpc.qmd
-            - docs/ray-integration.qmd
+            - docs/cli.qmd
+            - docs/config.qmd
+            - text: "API Reference"
+              href: docs/api
+
         - section: "Dataset Formats"
           contents: docs/dataset-formats/*
-        - section: "Reference"
+
+        - section: "Deployments"
           contents:
-            - docs/config.qmd
-        - docs/faq.qmd
+            - docs/docker.qmd
+            - docs/multi-gpu.qmd
+            - docs/multi-node.qmd
+            - docs/ray-integration.qmd
+            - docs/amd_hpc.qmd
+            - docs/mac.qmd
+
+        - section: "How To Guides"
+          contents:
+            - docs/multimodal.qmd
+            - docs/rlhf.qmd
+            - docs/reward_modelling.qmd
+            - docs/lr_groups.qmd
+            - docs/lora_optims.qmd
+            - docs/dataset_loading.qmd
+
+        - section: "Core Concepts"
+          contents:
+            - docs/batch_vs_grad.qmd
+            - docs/dataset_preprocessing.qmd
+            - docs/multipack.qmd
+
+        - section: "Advanced Features"
+          contents:
+            - docs/fsdp_qlora.qmd
+            - docs/unsloth.qmd
+            - docs/torchao.qmd
+            - docs/custom_integrations.qmd
+            - docs/sequence_parallelism.qmd
+
+        - section: "Troubleshooting"
+          contents:
+            - docs/faq.qmd
+            - docs/debugging.qmd
+            - docs/nccl.qmd
 
 format:
   html:
-    theme: materia
+    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

@@ -31,10 +31,11 @@ RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
         sed -i 's#^datasets.*#datasets @ git+https://github.com/huggingface/datasets.git@main#' requirements.txt; \
     fi
 
+RUN pip install packaging==23.2 setuptools==75.8.0
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        pip install --no-build-isolation -e .[deepspeed,flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$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 +3,10 @@ set -e
 
 python -c "import torch; assert '$PYTORCH_VERSION' in torch.__version__"
 
-pytest -v --durations=10 -n8 --ignore=tests/e2e/ --ignore=tests/patched/ /workspace/axolotl/tests/
-# pytest -v --durations=10 -n8 --dist loadfile /workspace/axolotl/tests/patched/
-pytest -v --durations=10 /workspace/axolotl/tests/e2e/patched/
+pytest -v --durations=10 -n8 --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli /workspace/axolotl/tests/
+pytest -v --durations=10 /workspace/axolotl/tests/e2e/patched/lora_kernels  # running these with the other patches causes a failure
+pytest -v --durations=10 --ignore=tests/e2e/patched/lora_kernels /workspace/axolotl/tests/e2e/patched
 pytest -v --durations=10 -n1 /workspace/axolotl/tests/e2e/solo/
 pytest -v --durations=10 /workspace/axolotl/tests/e2e/integrations/
-pytest -v --durations=10 --ignore=tests/e2e/solo/ --ignore=tests/e2e/patched/ --ignore=tests/e2e/multigpu/ --ignore=tests/e2e/integrations/ /workspace/axolotl/tests/e2e/
+pytest -v --durations=10 /workspace/axolotl/tests/cli
+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/

cicd/e2e_tests.py

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

cicd/multigpu.py

@@ -1,6 +1,7 @@
 """
- modal application to run axolotl gpu tests in Modal
- """
+modal application to run axolotl gpu tests in Modal
+"""
+
 # pylint: disable=duplicate-code
 
 import os
@@ -37,15 +38,11 @@ temp_dir = tempfile.mkdtemp()
 with open(pathlib.Path(temp_dir) / "Dockerfile", "w", encoding="utf-8") as f:
     f.write(dockerfile_contents)
 
-cicd_image = (
-    Image.from_dockerfile(
-        pathlib.Path(temp_dir) / "Dockerfile",
-        force_build=True,
-        gpu="A10G",
-    )
-    .env(df_args)
-    .pip_install("fastapi==0.110.0", "pydantic==2.6.3")
-)
+cicd_image = Image.from_dockerfile(
+    pathlib.Path(temp_dir) / "Dockerfile",
+    force_build=True,
+    gpu="A10G",
+).env(df_args)
 
 app = App("Axolotl CI/CD", secrets=[])
 

cicd/multigpu.sh

@@ -2,4 +2,5 @@
 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/

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

@@ -28,8 +28,8 @@ 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==${PYTORCH_VERSION}+cu${CUDA} --extra-index-url https://download.pytorch.org/whl/cu$CUDA && \
+RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==23.2 setuptools==75.8.0 wheel && \
+    python3 -m pip install --no-cache-dir -U torch==${PYTORCH_VERSION}+cu${CUDA} torchvision --extra-index-url https://download.pytorch.org/whl/cu$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

docker/Dockerfile-base-nightly

@@ -0,0 +1,39 @@
+ARG CUDA_VERSION="12.8.1"
+ARG CUDNN_VERSION="8"
+ARG UBUNTU_VERSION="22.04"
+ARG MAX_JOBS=4
+
+FROM nvidia/cuda:$CUDA_VERSION-cudnn$CUDNN_VERSION-devel-ubuntu$UBUNTU_VERSION AS base-builder
+
+ENV PATH="/root/miniconda3/bin:${PATH}"
+
+ARG PYTHON_VERSION="3.11"
+ARG PYTORCH_VERSION="nightly"
+ARG CUDA="128"
+ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 9.0+PTX"
+
+ENV PYTHON_VERSION=$PYTHON_VERSION
+ENV TORCH_CUDA_ARCH_LIST=$TORCH_CUDA_ARCH_LIST
+
+RUN apt-get update \
+    && apt-get install -y wget git build-essential ninja-build git-lfs libaio-dev pkg-config && rm -rf /var/lib/apt/lists/* \
+    && wget \
+    https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh \
+    && mkdir /root/.conda \
+    && bash Miniconda3-latest-Linux-x86_64.sh -b \
+    && rm -f Miniconda3-latest-Linux-x86_64.sh \
+    && conda create -n "py${PYTHON_VERSION}" python="${PYTHON_VERSION}"
+
+ENV PATH="/root/miniconda3/envs/py${PYTHON_VERSION}/bin:${PATH}"
+
+WORKDIR /workspace
+
+RUN python3 -m pip install --upgrade pip && pip3 install packaging && \
+    python3 -m pip install --no-cache-dir -U torch --extra-index-url https://download.pytorch.org/whl/nightly/cu$CUDA && \
+    python3 -m pip install --no-cache-dir "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main" && \
+    python3 -m pip install --no-cache-dir "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main"
+
+RUN git lfs install --skip-repo && \
+    pip3 install awscli && \
+    # The base image ships with `pydantic==1.8.2` which is not working
+    pip3 install -U --no-cache-dir pydantic==1.10.10

docker/Dockerfile-cloud

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

docs/.gitignore

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

docs/amd_hpc.qmd

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

docs/cli.qmd

@@ -1,28 +1,19 @@
-# Axolotl CLI Documentation
+---
+title: "Command Line Interface (CLI)"
+format:
+  html:
+    toc: true
+    toc-expand: 2
+    toc-depth: 3
+execute:
+  enabled: false
+---
 
 The Axolotl CLI provides a streamlined interface for training and fine-tuning large language models. This guide covers
 the CLI commands, their usage, and common examples.
 
-### Table of Contents
 
-- Basic Commands
-- Command Reference
-  - fetch
-  - preprocess
-  - train
-  - inference
-  - merge-lora
-  - merge-sharded-fsdp-weights
-  - evaluate
-  - lm-eval
-- Legacy CLI Usage
-- Remote Compute with Modal Cloud
-  - Cloud Configuration
-  - Running on Modal Cloud
-  - Cloud Configuration Options
-
-
-### Basic Commands
+## Basic Commands
 
 All Axolotl commands follow this general structure:
 
@@ -32,9 +23,9 @@ axolotl <command> [config.yml] [options]
 
 The config file can be local or a URL to a raw YAML file.
 
-### Command Reference
+## Command Reference
 
-#### fetch
+### fetch
 
 Downloads example configurations and deepspeed configs to your local machine.
 
@@ -49,7 +40,7 @@ axolotl fetch deepspeed_configs
 axolotl fetch examples --dest path/to/folder
 ```
 
-#### preprocess
+### preprocess
 
 Preprocesses and tokenizes your dataset before training. This is recommended for large datasets.
 
@@ -74,7 +65,7 @@ dataset_prepared_path: Local folder for saving preprocessed data
 push_dataset_to_hub: HuggingFace repo to push preprocessed data (optional)
 ```
 
-#### train
+### train
 
 Trains or fine-tunes a model using the configuration specified in your YAML file.
 
@@ -95,7 +86,38 @@ axolotl train config.yml --no-accelerate
 axolotl train config.yml --resume-from-checkpoint path/to/checkpoint
 ```
 
-#### inference
+It is possible to run sweeps over multiple hyperparameters by passing in a sweeps config.
+
+```bash
+# Basic training with sweeps
+axolotl train config.yml --sweep path/to/sweep.yaml
+```
+
+Example sweep config:
+```yaml
+_:
+  # This section is for dependent variables we need to fix
+  - load_in_8bit: false
+    load_in_4bit: false
+    adapter: lora
+  - load_in_8bit: true
+    load_in_4bit: false
+    adapter: lora
+
+# These are independent variables
+learning_rate: [0.0003, 0.0006]
+lora_r:
+  - 16
+  - 32
+lora_alpha:
+  - 16
+  - 32
+  - 64
+```
+
+
+
+### inference
 
 Runs inference using your trained model in either CLI or Gradio interface mode.
 
@@ -115,7 +137,7 @@ cat prompt.txt | axolotl inference config.yml \
     --base-model="./completed-model"
 ```
 
-#### merge-lora
+### merge-lora
 
 Merges trained LoRA adapters into the base model.
 
@@ -137,7 +159,7 @@ gpu_memory_limit: Limit GPU memory usage
 lora_on_cpu: Load LoRA weights on CPU
 ```
 
-#### merge-sharded-fsdp-weights
+### merge-sharded-fsdp-weights
 
 Merges sharded FSDP model checkpoints into a single combined checkpoint.
 
@@ -146,36 +168,38 @@ Merges sharded FSDP model checkpoints into a single combined checkpoint.
 axolotl merge-sharded-fsdp-weights config.yml
 ```
 
-#### evaluate
+### evaluate
 
-Evaluates a model's performance using metrics specified in the config.
+Evaluates a model's performance (loss etc) on the train and eval datasets.
 
 ```bash
 # Basic evaluation
 axolotl evaluate config.yml
 ```
 
-#### lm-eval
+### lm-eval
 
 Runs LM Evaluation Harness on your model.
 
 ```bash
 # Basic evaluation
 axolotl lm-eval config.yml
-
-# Evaluate specific tasks
-axolotl lm-eval config.yml --tasks arc_challenge,hellaswag
 ```
 
 Configuration options:
 
 ```yaml
-lm_eval_tasks: List of tasks to evaluate
-lm_eval_batch_size: Batch size for evaluation
-output_dir: Directory to save evaluation results
+# List of tasks to evaluate
+lm_eval_tasks:
+  - arc_challenge
+  - hellaswag
+lm_eval_batch_size: # Batch size for evaluation
+output_dir: # Directory to save evaluation results
 ```
 
-### Legacy CLI Usage
+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:
 
@@ -195,19 +219,25 @@ accelerate launch -m axolotl.cli.inference config.yml \
     --lora_model_dir="./outputs/lora-out" --gradio
 ```
 
-### Remote Compute with Modal Cloud
+::: {.callout-important}
+When overriding CLI parameters in the legacy CLI, use same notation as in yaml file (e.g., `--lora_model_dir`).
+
+**Note:** This differs from the new Click-based CLI, which uses dash notation (e.g., `--lora-model-dir`). Keep this in mind if you're referencing newer documentation or switching between CLI versions.
+:::
+
+## Remote Compute with Modal Cloud
 
 Axolotl supports running training and inference workloads on Modal cloud infrastructure. This is configured using a
 cloud YAML file alongside your regular Axolotl config.
 
-#### Cloud Configuration
+### Cloud Configuration
 
 Create a cloud config YAML with your Modal settings:
 
 ```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)
@@ -215,13 +245,17 @@ branch: main    # Git branch to use (optional)
 volumes:        # Persistent storage volumes
   - name: axolotl-cache
     mount: /workspace/cache
+  - name: axolotl-data
+    mount: /workspace/data
+  - name: axolotl-artifacts
+    mount: /workspace/artifacts
 
-env:            # Environment variables
+secrets:        # Secrets to inject
   - WANDB_API_KEY
   - HF_TOKEN
 ```
 
-#### Running on Modal Cloud
+### Running on Modal Cloud
 
 Commands that support the --cloud flag:
 
@@ -239,18 +273,30 @@ axolotl train config.yml --cloud cloud_config.yml --no-accelerate
 axolotl lm-eval config.yml --cloud cloud_config.yml
 ```
 
-#### Cloud Configuration Options
+### Cloud Configuration Options
 
 ```yaml
-provider: compute provider, currently only `modal` is supported
-gpu: GPU type to use
-gpu_count: Number of GPUs (default: 1)
-memory: RAM in GB (default: 128)
-timeout: Maximum runtime in seconds
-timeout_preprocess: Preprocessing timeout
-branch: Git branch to use
-docker_tag: Custom Docker image tag
-volumes: List of persistent storage volumes
-env: Environment variables to pass
-secrets: Secrets to inject
+provider:    # compute provider, currently only `modal` is supported
+gpu:         # GPU type to use
+gpu_count:   # Number of GPUs (default: 1)
+memory:      # RAM in GB (default: 128)
+timeout:     # Maximum runtime in seconds
+timeout_preprocess: # Preprocessing timeout
+branch:      # Git branch to use
+docker_tag:  # Custom Docker image tag
+volumes:     # List of persistent storage volumes
+
+# 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

@@ -1,5 +1,5 @@
 ---
-title: Config options
+title: Config Reference
 description: A complete list of all configuration options.
 ---
 
@@ -30,6 +30,11 @@ tokenizer_legacy:
 # Resize the model embeddings when new tokens are added to multiples of 32
 # This is reported to improve training speed on some models
 resize_token_embeddings_to_32x:
+# Optional[bool] Whether to shrink the embeddings to len(tokenizer). By default, we won't shrink.
+shrink_embeddings:
+# 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
@@ -83,6 +88,12 @@ gpu_memory_limit: 20GiB
 # Do the LoRA/PEFT loading on CPU -- this is required if the base model is so large it takes up most or all of the available GPU VRAM, e.g. during a model and LoRA merge
 lora_on_cpu: true
 
+# List[str]. Add plugins to extend the pipeline.
+# See `src/axolotl/integrations` for the available plugins or doc below for more details.
+# https://axolotl-ai-cloud.github.io/axolotl/docs/custom_integrations.html
+plugins:
+  # - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
 # A list of one or more datasets to finetune the model with
 datasets:
   # HuggingFace dataset repo | s3://,gs:// path | "json" for local dataset, make sure to fill data_files
@@ -91,9 +102,14 @@ datasets:
     type: alpaca # format | format:<prompt_style> (chat/instruct) | <prompt_strategies>.load_<load_fn>
     ds_type: # Optional[str] (json|arrow|parquet|text|csv) defines the datatype when path is a file
     data_files: # Optional[str] path to source data files
-    shards: # Optional[int] number of shards to split data into
+
+    shards: # Optional[int] split dataset into N pieces (use with shards_idx)
+    shards_idx: # Optional[int] = 0 the index of sharded dataset to use
+
+    preprocess_shards: # Optional[int] process dataset in N sequential chunks for memory efficiency (exclusive with `shards`)
+
     name: # Optional[str] name of dataset configuration to load
-    train_on_split: train # Optional[str] name of dataset split to load from
+    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
 
@@ -137,22 +153,37 @@ datasets:
 
     # Key containing the messages (default: "messages")
     field_messages: messages
-    # Key for role in each message (default: "role")
-    message_field_role: role
-    # Key for content in each message (default:  "content")
-    message_field_content: content
 
-    # Optional[Dict[str, List]]. Roles mapping in the messages. The default is:
+    # Mapping of properties from the input dataset to the chat template.
+    # (default: message_property_mappings={'role':'role', 'content':'content'})
+    # If a property exists in the template but not in this mapping, the system will attempt
+    # to load it directly from the message using the property name as the key.
+    # Example: In the mapping below, 'from' is loaded from input dataset and used as 'role',
+    # while 'value' is loaded and used as 'content' in the chat template.
+    message_property_mappings:
+      role: from
+      content: value
+      # ...
+
+    # Optional[Dict[str, List]]. Roles mapping in the messages.
+    # The 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"]
       system: ["system"]
       tool: ["tool"]
 
+    # Optional[bool]. Whether to drop the system turn from the dataset. Only works with chat_template.
+    # This does not drop the default system message from chat_template if it exists. If you wish to,
+    # we recommend using a custom jinja template with the default system message removed or
+    # adding a system turn with empty content.
+    drop_system_message:
+
     # IMPORTANT: The following fields determine which parts of the conversation to train on.
     # Priority order: message_field_training > message_field_training_detail > train_on_inputs or role in roles_to_train
     # See examples at `docs/dataset-formats/conversation.qmd`
-    # Note: If the below 4 fields are empty, defaults to training only on the last message.
+    # Note: If the below 4 fields are set to empty, defaults to training only on the last message.
 
     # Optional[List[str]]. Roles to train on. The tokens from these roles will be considered for the loss.
     roles_to_train: ["assistant"]  # default
@@ -160,6 +191,7 @@ datasets:
     # - all: train on all EOS tokens
     # - turn (default): train on the EOS token at the end of each trainable turn
     # - last: train on the last EOS token in the conversation
+    # TIP: Please make sure that your `tokenizer.eos_token` is same as EOS/EOT token in template. Otherwise, set `eos_token` under `special_tokens`.
     train_on_eos: last
     # The key in the message turn that indicates via boolean whether tokens of a turn should be considered for training. Useful to selectively train on certain turns besides the `roles_to_train`.
     message_field_training: training
@@ -186,10 +218,46 @@ test_datasets:
     data_files:
       - /workspace/data/eval.jsonl
 
-# use RL training: 'dpo', 'ipo', 'kto'
+# use RL training: 'dpo', 'ipo', 'kto', 'simpo', 'orpo', 'grpo'
 rl:
-# whether to perform weighting if doing DPO training. Boolean.
-dpo_use_weighting:
+rl_beta:  # Optional[float]. The beta parameter for the RL training.
+
+# dpo
+dpo_use_weighting:  # Optional[bool]. Whether to perform weighting.
+rpo_alpha: # Optional[float]. Weighting of NLL term in loss from RPO paper.
+
+# orpo
+orpo_alpha: 0.1  # Parameter controlling the relative ratio loss weight in the ORPO loss. Passed to `beta` in `ORPOConfig` due to trl mapping.
+
+# kto
+kto_desirable_weight: # Optional[float]. Factor for desirable loss term in KTO loss.
+kto_undesirable_weight: # Optional[float]. Factor for undesirable loss term in KTO loss.
+
+# simpo
+cpo_alpha: 1.0  # Weight of the BC regularizer
+simpo_gamma: 0.5  # Target reward margin for the SimPO loss
+
+# grpo
+trl:
+  use_vllm: # Optional[bool]. Whether to use VLLM for RL training.
+  vllm_server_host: # Optional[str]. Host of the vLLM server to connect to.
+  vllm_server_port: # Optional[int]. Port of the vLLM server to connect to.
+  vllm_server_timeout: # Optional[int]. Total timeout (in seconds) to wait for the vLLM server to respond.
+  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.
+
+  reward_funcs: # Optional[list[str]]. List of reward functions to load. Paths must be importable from current dir.
+  reward_weights: # Optional[list[float]]. List of reward weights for the reward functions.
+
+  num_generations: # Optional[int]. Number of generations to sample.
+  log_completions: # Optional[bool]. Whether to log completions.
+
+  sync_ref_model: # Optional[bool]. Whether to sync the reference model.
+  ref_model_mixup_alpha: # Optional[float]. Mixup alpha for the reference model.
+  ref_model_sync_steps: # Optional[int]. Sync steps for the reference model.
+
 
 # reward modelling: `True` or `False`
 reward_model:
@@ -207,13 +275,13 @@ process_reward_model:
 chat_template: tokenizer_default
 # custom jinja template for chat template. This will be only used if chat_template is set to `jinja` or `null` (in which case chat_template is automatically set to `jinja`). Default is null.
 chat_template_jinja: null
-# Changes the default system message
-default_system_message: You are a helpful assistant. Please give a long and detailed answer. # Currently only supports chatml.
+# Changes the default system message. Currently only supports chatml.
+default_system_message: You are a helpful assistant. Please give a long and detailed answer.
 # Axolotl attempts to save the dataset as an arrow after packing the data together so
 # subsequent training attempts load faster, relative path
 dataset_prepared_path: data/last_run_prepared
 # Push prepared dataset to hub
-push_dataset_to_hub: # repo path
+push_dataset_to_hub: # Optional[str] repo_org/repo_name
 # The maximum number of processes to use while preprocessing your input dataset. This defaults to `os.cpu_count()`
 # if not set.
 dataset_processes: # defaults to os.cpu_count() if not set
@@ -254,9 +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:
 
@@ -288,7 +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.
@@ -300,6 +392,13 @@ lora_modules_to_save:
 
 lora_fan_in_fan_out: false
 
+# Apply custom LoRA autograd functions and activation function Triton kernels for
+# speed and memory savings
+# See: https://axolotl-ai-cloud.github.io/axolotl/docs/lora_optims.html
+lora_mlp_kernel: true
+lora_qkv_kernel: true
+lora_o_kernel: true
+
 # LoRA+ hyperparameters
 # For more details about the following options, see:
 # https://arxiv.org/abs/2402.12354  and `src/axolotl/core/train_builder.py`
@@ -348,6 +447,9 @@ comet_mode: # Create a new experiment ("create") or log to an existing one ("get
 comet_online: # Set to True to log data to Comet server, or False for offline storage. Default is True.
 comet_experiment_config: # Dictionary for additional configuration settings, see the doc for more details.
 
+# Tensorboard
+use_tensorboard: # Optional[bool]
+
 # Where to save the full-finetuned model to
 output_dir: ./completed-model
 
@@ -382,8 +484,15 @@ save_total_limit: # Checkpoints saved at a time
 # e.g., when 1 epoch is 1000 steps => `num_epochs: 2` and `max_steps: 100` will train for 100 steps
 max_steps:
 
+# bool of whether to include tokens trainer per second in the training metrics. This iterates over the entire dataset once, so it takes some time.
+include_tokens_per_second: # Optional[bool]
+
+# whether to find batch size that fits in memory. Passed to underlying transformers Trainer
+auto_find_batch_size: # Optional[bool]
+
 eval_table_size: # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0
 eval_max_new_tokens: # Total number of tokens generated for predictions sent to wandb. Default is 128
+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.
@@ -403,7 +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:
@@ -414,7 +524,7 @@ gradient_checkpointing: false
 early_stopping_patience: 3
 
 # Specify a scheduler and kwargs to use with the optimizer
-lr_scheduler: # 'one_cycle' | 'log_sweep' | empty for cosine
+lr_scheduler: # 'one_cycle' | 'rex' | 'log_sweep' | empty for cosine
 lr_scheduler_kwargs:
 cosine_min_lr_ratio: # decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of peak lr
 cosine_constant_lr_ratio: # freeze lr at some percentage of the step, e.g. cosine_constant_lr_ratio=0.8 means start cosine_min_lr at 80% of training step (https://arxiv.org/pdf/2308.04014.pdf)
@@ -424,36 +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:
@@ -482,27 +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_rms_norm:  # 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_mlp: # Whether to fuse part of the MLP into a single operation
-# Whether to use scaled-dot-product attention
+flash_attn_cross_entropy:  # Optional[bool]. Whether to use flash-attention cross entropy 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: # Optional[bool]. Whether to fuse QKV into a single operation
+flash_attn_fuse_mlp: # Optional[bool]. Whether to fuse part of the MLP into a single operation
+# 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:
-# Resume from a specific checkpoint dir
+
+# Optional[bool]. Whether to use low_cpu_mem_usage
+low_cpu_mem_usage:
+# 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:
 
@@ -517,6 +664,13 @@ special_tokens:
 # Add extra tokens.
 tokens:
 
+# Mapping token_id to new_token_string to override reserved added_tokens in the tokenizer.
+# Only works for tokens that are not part of the base vocab (aka are added_tokens).
+# Can be checked if they exist in tokenizer.json added_tokens.
+added_tokens_overrides:  # Dict[int, str]
+#  128041: "<|im_start|>"
+#  128042: "<|im_end|>"
+
 # FSDP
 fsdp:
 fsdp_config:
@@ -529,6 +683,17 @@ 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://axolotl-ai-cloud.github.io/axolotl/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
+
 # Path to torch distx for optim 'adamw_anyprecision'
 torchdistx_path:
 

docs/custom_integrations.qmd

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

docs/dataset-formats/conversation.qmd

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

docs/dataset-formats/index.qmd

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

docs/dataset-formats/pretraining.qmd

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

docs/dataset-formats/template_free.qmd

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

docs/dataset_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

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

docs/debugging.qmd

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

docs/docker.qmd

@@ -0,0 +1,139 @@
+---
+title: "Docker"
+format:
+  html:
+    toc: true
+    toc-depth: 4
+---
+
+This section describes the different Docker images that are released by AxolotlAI at [Docker Hub](https://hub.docker.com/u/axolotlai).
+
+## Base
+
+The base image is the most minimal image that can install Axolotl. It is based on the `nvidia/cuda` image. It includes python, torch, git, git-lfs, awscli, pydantic, and more.
+
+#### Image
+
+```
+axolotlai/axolotl-base
+```
+
+Link: [Docker Hub](https://hub.docker.com/r/axolotlai/axolotl-base)
+
+#### Tags format
+
+```bash
+main-base-py{python_version}-cu{cuda_version}-{pytorch_version}
+```
+
+Tags examples:
+
+- `main-base-py3.11-cu124-2.6.0`
+- `main-base-py3.11-cu124-2.5.1`
+- `main-base-py3.11-cu124-2.4.1`
+
+## Main
+
+The main image is the image that is used to run Axolotl. It is based on the `axolotlai/axolotl-base` image and includes the Axolotl codebase, dependencies, and more.
+
+#### Image
+
+```
+axolotlai/axolotl
+```
+
+Link: [Docker Hub](https://hub.docker.com/r/axolotlai/axolotl)
+
+#### Tags format {#sec-main-tags}
+
+```bash
+# on push to main
+main-py{python_version}-cu{cuda_version}-{pytorch_version}
+
+# latest main (currently torch 2.5.1, python 3.11, cuda 12.4)
+main-latest
+
+# nightly build
+{branch}-{date_in_YYYYMMDD}-py{python_version}-cu{cuda_version}-{pytorch_version}
+
+# tagged release
+{version}
+```
+
+:::{.callout-tip}
+
+There may be some extra tags appended to the image, like `-vllm` which installs those packages.
+
+:::
+
+Tags examples:
+
+- `main-py3.11-cu124-2.6.0`
+- `main-py3.11-cu124-2.5.1`
+- `main-py3.11-cu124-2.4.1`
+- `main-latest`
+- `main-20250303-py3.11-cu124-2.6.0`
+- `main-20250303-py3.11-cu124-2.5.1`
+- `main-20250303-py3.11-cu124-2.4.1`
+- `0.7.1`
+
+## Cloud
+
+The cloud image is the image that is used to run Axolotl in the cloud. It is based on the `axolotlai/axolotl` image and sets ENV variables like HuggingFace cache directories for volume mounts, tmux, and more for different cloud providers.
+
+:::{.callout-tip}
+
+Jupyter lab is run by default. Set `JUPYTER_DISABLE=1` in the environment variables to disable it.
+
+:::
+
+#### Image
+
+```
+axolotlai/axolotl-cloud
+```
+
+Link: [Docker Hub](https://hub.docker.com/r/axolotlai/axolotl-cloud)
+
+#### Tags format
+
+This uses the same tags as the [`main` image](#sec-main-tags).
+
+#### Environment variables
+
+- `JUPYTER_DISABLE`: Disable Jupyter lab.
+- `JUPYTER_PASSWORD`: Set a password for the Jupyter lab.
+- `PUBLIC_KEY` / `SSH_KEY`: Add a public key for the SSH service.
+
+#### Volume mounts
+
+:::{.callout-tip}
+
+We recommend mounting volumes to `/workspace/data` for data persistence. `/workspace/axolotl` contains the source code and is ephemeral.
+
+:::
+
+- `/workspace/data/axolotl-artifacts`: Directory to store Axolotl artifacts.
+- `/workspace/data/huggingface-cache`: Directory to store HuggingFace cache.
+
+## Cloud-no-tmux
+
+This is the same as the [`cloud` image](#sec-cloud) but without tmux.
+
+#### Image
+
+```
+axolotlai/axolotl-cloud-term
+```
+
+Link: [Docker Hub](https://hub.docker.com/r/axolotlai/axolotl-cloud-term)
+
+:::{.callout-note}
+
+The naming may be a bit confusing as it has `-term` appended to the end.
+
+:::
+
+#### Tags format
+
+This uses the same tags as the [`cloud` image](#sec-cloud-tags).

docs/faq.qmd

@@ -3,6 +3,7 @@ title: FAQ
 description: Frequently asked questions
 ---
 
+### General
 
 **Q: The trainer stopped and hasn't progressed in several minutes.**
 
@@ -18,4 +19,64 @@ description: Frequently asked questions
 
 **Q: AttributeError: 'DummyOptim' object has no attribute 'step'**
 
-> A: You may be using deepspeed with single gpu. Please don't set `deepspeed:` in yaml or cli.
+**Q: ModuleNotFoundError: No module named 'mpi4py' using single GPU with deepspeed**
+
+> A: You may be using deepspeed with single gpu. Please remove the `deepspeed:` section in the yaml file or `--deepspeed` CLI flag.
+
+**Q: The codes is stuck on saving preprocessed datasets.**
+
+> A: This is usually an issue with the GPU. This can be resolved through setting the os environment variable `CUDA_VISIBLE_DEVICES=0`. If you are on runpod, this is usually a pod issue. Starting a new pod should take care of it.
+
+**Q: Received mismatch error on merge adapters / loading adapters between torch.Size of checkpoint and model.**
+
+> A: This is likely due to vocab size mismatch. By default, Axolotl expands the model's embeddings if the tokenizer has more tokens than the model. Please use the `axolotl merge-lora` command to merge the adapters instead of using your own scripts.
+
+> On the other hand, if the model has more tokens than the tokenizer, Axolotl does not shrink the model's embeddings unless `shrink_embeddings: true` is set in the config.
+
+**Q: How to call Axolotl via custom python scripts?**
+
+> A: 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
+
+**Q: `jinja2.exceptions.UndefinedError: 'dict object' has no attribute 'content' / 'role' / ____`**
+
+> A: This means that the property mapping for the stated attribute does not exist when building `chat_template` prompt. For example, if `no attribute 'content'`, please check you have added the correct mapping for `content` under `message_property_mappings`.
+
+**Q: `Empty template generated for turn ___`**
+
+> A: The `content` is empty for that turn.
+
+**Q: `Could not find content start/end boundary for turn __`**
+
+> A: The specific turn's start/end could not be detected. Please ensure you have set the `eos_token` following your `chat_template`. Otherwise, this could be a `chat_template` which doesn't use proper boundaries for each turn (like system). On the rare occurrence, make sure your content is not `[[dummy_message]]`. Please let us know about this.
+
+**Q: `Content end boundary is before start boundary for turn ___`**
+
+> A: This is an edge case which should not occur. Please create an Issue if this happens.
+
+**Q: `Content end boundary is the same as start boundary for turn ___. This is likely an empty turn.`**
+
+> A: This is likely an empty turn.
+
+**Q: The EOS/EOT token is incorrectly being masked or not being masked.**
+
+> A: This is because of the mismatch between `tokenizer.eos_token` and EOS/EOT token in template. Please make sure to set `eos_token` under `special_tokens` to the same EOS/EOT token as in template.
+
+**Q: "`chat_template` choice is `tokenizer_default` but tokenizer's `chat_template` is null. Please add a `chat_template` in tokenizer config"**
+
+> A: This is because the tokenizer does not have a chat template. Please add a chat template in the tokenizer config. See [chat_template](dataset-formats/conversation.qmd#chat-template) for more details.

docs/getting-started.qmd

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

docs/inference.qmd

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

docs/input_output.qmd

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

docs/installation.qmd

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

docs/lora_optims.qmd

@@ -0,0 +1,132 @@
+---
+title: "LoRA Optimizations"
+description: "Custom autograd functions and Triton kernels in Axolotl for optimized LoRA fine-tuning"
+---
+
+Inspired by [Unsloth](https://github.com/unslothai/unsloth), we've implemented two
+optimizations for LoRA and QLoRA fine-tuning, supporting both single GPU and multi-GPU
+(in the DDP and DeepSpeed settings) training. These include (1) SwiGLU and GEGLU activation function
+Triton kernels, and (2) LoRA MLP and attention custom autograd functions. Our goal was
+to leverage operator fusion and tensor re-use in order to improve speed and reduce
+memory usage during the forward and backward passes of these calculations.
+
+We currently support several common model architectures, including (but not limited to):
+
+- `llama`
+- `mistral`
+- `qwen2`
+- `gemma`
+- `gemma2`
+- `gemma3`
+
+<details>
+
+The set of models we support is currently limited by our attention patching strategy,
+which assumes (and replaces) specific code blocks for query / key / value and output
+projections:
+
+```python
+ORIGINAL_QKV_CODE = """
+    query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+    key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+    value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+""".lstrip(
+    "\n"
+)
+
+ORIGINAL_O_CODE = """
+    attn_output = self.o_proj(attn_output)
+""".lstrip(
+    "\n"
+)
+```
+
+Is replaced with:
+
+```python
+PATCHED_QKV_CODE = """
+    query_states, key_states, value_states = self.apply_qkv(hidden_states)
+    query_states = query_states.view(hidden_shape).transpose(1, 2)
+    key_states = key_states.view(hidden_shape).transpose(1, 2)
+    value_states = value_states.view(hidden_shape).transpose(1, 2)
+""".lstrip(
+    "\n"
+)
+
+PATCHED_O_CODE = """
+    attn_output = self.apply_o(attn_output)
+""".lstrip(
+    "\n"
+)
+```
+
+Where `apply_qkv` and `apply_o` are defined in the `axolotl.kernels.lora` module.
+
+We welcome testing of other model architectures and / or PRs to expand our patching
+logic to be compatible with more of them.
+
+</details>
+
+::: {.callout-tip}
+Check out our [LoRA optimizations blog](https://axolotlai.substack.com/p/accelerating-lora-fine-tuning-with).
+:::
+
+## Usage
+
+These optimizations can be enabled in your Axolotl config YAML file. The
+`lora_mlp_kernel` option enables the optimized MLP path, while `lora_qkv_kernel` and
+`lora_o_kernel` enable the fused query-key-value projection and optimized output
+projection, respectively.
+
+```yaml
+lora_mlp_kernel: true
+lora_qkv_kernel: true
+lora_o_kernel: true
+```
+
+## Requirements
+
+- One or more NVIDIA or AMD GPUs (in order to use the Triton kernels)
+    - Note: Set `TORCH_ROCM_AOTRITON_ENABLE_EXPERIMENTAL=1` to enable [memory-efficient attention on AMD GPUs](https://github.com/ROCm/aotriton/issues/16#issuecomment-2346675491)
+- Targeted LoRA adapters cannot use Dropout
+    - This may limit model expressivity / cause overfitting
+- Targeted LoRA adapters cannot have bias terms
+    - This may limit model expressivity
+
+Models with pre-existing LoRA adapters that use Dropout or have bias terms may need to
+be re-finetuned without these features in order to be useful.
+
+## Implementation details
+
+### Custom autograd functions
+
+The LoRA MLP autograd function optimizes the entire MLP computation path. It fuses the
+LoRA and base weight computations together and provides a single, efficient backward
+pass for the entire MLP block.
+
+For attention components, similar optimizations are provided through a function that
+handles the query, key, and value projections, and a function that handles the output
+projection. They are designed to work with the existing `transformers` attention
+implementation via some monkey-patching logic.
+
+### Triton kernels
+
+Two activation functions (SwiGLU and GeGLU) are implemented with Triton kernels for
+improved speed and memory performance. These kernels handle both the forward and
+backward passes.
+
+### Integration
+
+The custom autograd functions and Triton kernels are designed to work together. The
+autograd function manages the high-level computation flow and gradient tracking, while
+calling the Triton kernels for the activation function computation. During the backward
+pass, the kernel computes both the activation output and the required gradients, which
+the autograd function then uses to compute the final gradients for the entire
+computation path.
+
+## Future Work
+
+- Support for additional model architectures
+- Support for the FSDP setting
+- Support for dropout and bias
+- Additional operator fusions

docs/mac.qmd

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

docs/multi-gpu.qmd

@@ -1,5 +1,5 @@
 ---
-title: "Multi-GPU Training Guide"
+title: "Multi-GPU"
 format:
   html:
     toc: true
@@ -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}
@@ -35,7 +36,11 @@ deepspeed: deepspeed_configs/zero1.json
 ### Usage {#sec-deepspeed-usage}
 
 ```{.bash}
-accelerate launch -m axolotl.cli.train examples/llama-2/config.yml --deepspeed deepspeed_configs/zero1.json
+# Passing arg via config
+axolotl train config.yml
+
+# Passing arg via cli
+axolotl train config.yml --deepspeed deepspeed_configs/zero1.json
 ```
 
 ### ZeRO Stages {#sec-zero-stages}
@@ -62,6 +67,28 @@ 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).
@@ -70,25 +97,7 @@ For combining FSDP with QLoRA, see our [dedicated guide](fsdp_qlora.qmd).
 
 ### Liger Kernel Integration {#sec-liger}
 
-::: {.callout-note}
-Liger Kernel provides efficient Triton kernels for LLM training, offering:
-
-- 20% increase in multi-GPU training throughput
-- 60% reduction in memory usage
-- Compatibility with both FSDP and DeepSpeed
-:::
-
-Configuration:
-
-```{.yaml}
-plugins:
-  - axolotl.integrations.liger.LigerPlugin
-liger_rope: true
-liger_rms_norm: true
-liger_glu_activation: true
-liger_layer_norm: true
-liger_fused_linear_cross_entropy: true
-```
+Please see [docs](custom_integrations.qmd#liger) for more info.
 
 ## Troubleshooting {#sec-troubleshooting}
 

docs/multi-node.qmd

@@ -13,7 +13,7 @@ You will also need to have the same configuration file for your model on each ma
 Make sure the main machine is reachable by other machines.
 :::
 
-# Accelerate
+## Accelerate
 
 You will need to create a configuration for accelerate, either by using `accelerate config` and follow the instructions or you can use one of the preset below:
 
@@ -51,17 +51,17 @@ fsdp_config:
 
 All you have to do now is launch using accelerate as you would usually do on each machine and voila, the processes will start once you have launched accelerate on every machine.
 
-# Raytrain
+## Raytrain
 
 Please see ray train doc [here](ray-integration.qmd).
 
-# Torchrun
+## Torchrun
 
 If you are using Infiniband, we recommend torchrun to utilize the full bandwidth.
 
 Set the following env (change buffersize/socketname depending on your system):
 
-```yaml
+```bash
 export NCCL_IB_DISABLE=0
 export NCCL_SOCKET_IFNAME="eth0,en,eth,em,bond"
 export NCCL_BUFFSIZE=2097152

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,
-you'll need to use the following in YAML in combination with the rest of the required hyperparams.
+Multimodal support is limited and doesn't have full feature parity.
+
+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/nccl.qmd

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

docs/ray-integration.qmd

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

docs/reward_modelling.qmd

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

docs/rlhf.qmd

@@ -1,26 +1,40 @@
 ---
 title: "RLHF (Beta)"
 description: "Reinforcement Learning from Human Feedback is a method whereby a language model is optimized from data using human feedback."
+back-to-top-navigation: true
+toc: true
+toc-expand: 2
+toc-depth: 4
 ---
 
-### Overview
+## Overview
 
 Reinforcement Learning from Human Feedback is a method whereby a language model is optimized from data using human
 feedback. Various methods include, but not limited to:
 
+- [Direct Preference Optimization (DPO)](#dpo)
+- [Identity Preference Optimization (IPO)](#ipo)
+- [Kahneman-Tversky Optimization (KTO)](#kto)
+- [Odds Ratio Preference Optimization (ORPO)](#orpo)
 - Proximal Policy Optimization (PPO) (not yet supported in axolotl)
-- Direct Preference Optimization (DPO)
-- Identity Preference Optimization (IPO)
 
 
-### RLHF using Axolotl
+## RLHF using Axolotl
 
->[!IMPORTANT]
->This is a BETA feature and many features are not fully implemented. You are encouraged to open new PRs to improve the integration and functionality.
+::: {.callout-important}
+This is a BETA feature and many features are not fully implemented. You are encouraged to open new PRs to improve the integration and functionality.
+:::
 
-The various RL training methods are implemented in trl and wrapped via axolotl. Below are various examples with how you can use various preference datasets to train models that use ChatML
+We rely on the [TRL](https://github.com/huggingface/trl) library for implementations of various RL training methods, which we wrap around to expose in axolotl. Each method has their own supported ways of loading datasets and prompt formats.
+
+::: {.callout-tip}
+You can find what each method supports by going into `src/axolotl/prompt_strategies/{method}` where `{method}` is one of our supported methods. The `type: ` can be retrieved from `{method}.{function_name}`.
+:::
+
+### DPO
+
+Example config:
 
-#### DPO
 ```yaml
 rl: dpo
 datasets:
@@ -32,12 +46,265 @@ datasets:
     type: chatml
 ```
 
-#### IPO
+DPO supports the following types with the following dataset format:
+
+#### chatml.argilla
+
+```json
+{
+    "system": "...", // optional
+    "instruction": "...",
+    "chosen_response": "...",
+    "rejected_response": "..."
+}
+```
+
+#### chatml.argilla_chat
+
+```json
+{
+    "chosen": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ],
+    "rejected": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ]
+}
+```
+
+#### chatml.icr
+
+```json
+{
+    "system": "...", // optional
+    "input": "...",
+    "chosen": "...",
+    "rejected": "..."
+}
+```
+
+#### chatml.intel
+
+```json
+{
+    "system": "...", // optional
+    "question": "...",
+    "chosen": "...",
+    "rejected": "..."
+}
+```
+
+#### chatml.prompt_pairs
+
+```json
+{
+    "system": "...", // optional
+    "prompt": "...",
+    "chosen": "...",
+    "rejected": "..."
+}
+```
+
+#### chatml.ultra
+
+```json
+{
+    "system": "...", // optional
+    "prompt": "...",
+    "chosen": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ],
+    "rejected": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ]
+}
+```
+
+#### llama3.argilla
+
+```json
+{
+    "system": "...", // optional
+    "instruction": "...",
+    "chosen_response": "...",
+    "rejected_response": "..."
+}
+```
+
+#### llama3.argilla_chat
+
+```json
+{
+    "chosen": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ],
+    "rejected": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ]
+}
+```
+
+#### llama3.icr
+
+```json
+{
+    "system": "...", // optional
+    "input": "...",
+    "chosen": "...",
+    "rejected": "..."
+}
+```
+
+#### llama3.intel
+
+```json
+{
+    "system": "...", // optional
+    "question": "...",
+    "chosen": "...",
+    "rejected": "..."
+}
+```
+
+#### llama3.prompt_pairs
+
+```json
+{
+    "system": "...", // optional
+    "prompt": "...",
+    "chosen": "...",
+    "rejected": "..."
+}
+```
+
+#### llama3.ultra
+
+```json
+{
+    "system": "...", // optional
+    "prompt": "...",
+    "chosen": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ],
+    "rejected": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ]
+}
+```
+
+#### zephyr.nectar
+
+```json
+{
+    "prompt": "...",
+    "answers": [
+        {
+            "answer": "...",
+            "rank": 1
+        },
+        {
+            "answer": "...",
+            "rank": 2
+        }
+        // ... more answers with ranks
+    ]
+}
+```
+
+#### chat_template.default
+
+```yaml
+rl: dpo
+datasets:
+  - path: ...
+    split: train
+    type: chat_template.default
+    field_messages: "messages"
+    field_chosen: "chosen"
+    field_rejected: "rejected"
+    message_property_mappings:
+      role: role
+      content: content
+    roles:
+      user: ["user"]
+      assistant: ["assistant"]
+      system: ["system"]
+```
+
+Sample input format:
+
+```json
+{
+    "messages": [
+        {
+            "role": "system",
+            "content": "..."
+        },
+        {
+            "role": "user",
+            "content": "..."
+        },
+        // ... more messages
+    ],
+    "chosen": {
+        "role": "assistant",
+        "content": "..."
+    },
+    "rejected": {
+        "role": "assistant",
+        "content": "..."
+    }
+}
+```
+
+#### user_defined.default
+
+For custom behaviors,
+
+```yaml
+rl: dpo
+datasets:
+  - path: ...
+    split: train
+    type: user_defined.default
+
+    field_prompt: "prompt"
+    field_system: "system"
+    field_chosen: "chosen"
+    field_rejected: "rejected"
+    prompt_format: "{prompt}"
+    chosen_format: "{chosen}"
+    rejected_format: "{rejected}"
+```
+
+The input format is a simple JSON input with customizable fields based on the above config.
+
+```json
+{
+    "system": "...",  // optional
+    "prompt": "...",
+    "chosen": "...",
+    "rejected": "..."
+}
+```
+
+### IPO
+
+As IPO is just DPO with a different loss function, all supported dataset formats for [DPO](#dpo) are also supported for IPO.
+
 ```yaml
 rl: ipo
 ```
 
-#### ORPO
+### ORPO
 
 Paper: https://arxiv.org/abs/2403.07691
 
@@ -52,13 +319,34 @@ datasets:
     type: chat_template.argilla
 ```
 
+ORPO supports the following types with the following dataset format:
 
-#### KTO
+#### chat_template.argilla
+
+```json
+{
+    "system": "...",  // optional
+    "prompt": "...",  // if available, will be taken as user message for single-turn instead of from list below
+
+    // chosen/rejected should be same till last content and only even-number of alternating user/assistant turns
+    "chosen": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ],
+    "rejected": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ]
+}
+```
+
+### KTO
 
 ```yaml
 rl: kto
-rl_beta: 0.5
-kto_desirable_weight: 0.2
+rl_beta: 0.1  # default
+kto_desirable_weight: 1.0  # default
+kto_undesirable_weight: 1.0  # default
 
 remove_unused_columns: false
 
@@ -72,7 +360,243 @@ gradient_checkpointing_kwargs:
   use_reentrant: true
 ```
 
-#### Using local dataset files
+KTO supports the following types with the following dataset format:
+
+#### chatml.argilla
+
+```json
+{
+    "system": "...", // optional
+    "instruction": "...",
+    "completion": "..."
+}
+```
+
+#### chatml.argilla_chat
+
+```json
+{
+    "chosen": [
+        {"role": "user", "content": "..."}
+    ],
+    "completion": [
+        {"role": "assistant", "content": "..."}
+    ]
+}
+```
+
+#### chatml.intel
+
+```json
+{
+    "system": "...", // optional
+    "question": "...",
+    "completion": "..."
+}
+```
+
+#### chatml.prompt_pairs
+
+```json
+{
+    "system": "...", // optional
+    "prompt": "...",
+    "completion": "..."
+}
+```
+
+#### chatml.ultra
+
+```json
+{
+    "system": "...", // optional
+    "prompt": "...",
+    "completion": "..."
+}
+```
+
+#### llama3.argilla
+
+```json
+{
+    "system": "...", // optional
+    "instruction": "...",
+    "completion": "..."
+}
+```
+
+#### llama3.argilla_chat
+
+```json
+{
+    "completion": [
+        {"role": "user", "content": "..."},
+        {"role": "assistant", "content": "..."}
+    ]
+}
+```
+
+#### llama3.intel
+
+```json
+{
+    "system": "...", // optional
+    "question": "...",
+    "completion": "..."
+}
+```
+
+#### llama3.prompt_pairs
+
+```json
+{
+    "system": "...", // optional
+    "prompt": "...",
+    "completion": "..."
+}
+```
+
+#### llama3.ultra
+
+```json
+{
+    "system": "...", // optional
+    "prompt": "...",
+    "completion": "..."
+}
+```
+
+#### user_defined.default
+
+For custom behaviors,
+
+```yaml
+rl: kto
+datasets:
+  - path: ...
+    split: train
+    type: user_defined.default
+
+    field_prompt: "prompt"
+    field_system: "system"
+    field_completion: "completion"
+    field_label: "label"
+    prompt_format: "{prompt}"
+    completion_format: "{completion}"
+```
+
+The input format is a simple JSON input with customizable fields based on the above config.
+
+```json
+{
+    "system": "...",  // optional
+    "prompt": "...",
+    "completion": "...",
+    "label": "..."
+}
+```
+
+### GRPO
+
+::: {.callout-tip}
+Check out our [GRPO cookbook](https://github.com/axolotl-ai-cloud/axolotl-cookbook/tree/main/grpo#training-an-r1-style-large-language-model-using-grpo).
+:::
+
+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 example, to load OpenAI's GSM8K and use a random reward for completions:
+
+```python
+# rewards.py
+import random
+
+def rand_reward_func(completions, **kwargs) -> list[float]:
+    return [random.uniform(0, 1) for _ in completions]
+
+def oai_gsm8k_transform(cfg, *args, **kwargs):
+    def transform_fn(example, tokenizer=None):
+        label = example["answer"].split("####")[-1].strip().replace(",", "")
+        return {
+            "prompt": [{"role": "user", "content": example["question"]},],
+            "answer": label,
+        }
+    return transform_fn, {"remove_columns": ["question"]}
+```
+
+```yaml
+rl: grpo
+
+trl:
+    beta: 0.001
+    max_completion_length: 256
+    use_vllm: True
+    num_generations: 4
+    reward_funcs: ["rewards.rand_reward_func"]    # format: '{file_name}.{fn_name}'
+    reward_weights: [1.0]
+datasets:
+  - path: openai/gsm8k
+    name: main
+    type: rewards.oai_gsm8k_transform  # format: '{file_name}.{fn_name}'
+```
+
+To see other examples of custom reward functions, please see [TRL GRPO Docs](https://github.com/huggingface/trl/blob/main/docs/source/grpo_trainer.md#using-a-custom-reward-function).
+
+To see description of the configs, please see [TRLConfig](https://github.com/axolotl-ai-cloud/axolotl/blob/main/src/axolotl/utils/config/models/input/v0_4_1/trl.py).
+
+### SimPO
+
+SimPO uses [CPOTrainer](https://huggingface.co/docs/trl/main/en/cpo_trainer) but with alternative loss function.
+
+```yaml
+rl: simpo
+rl_beta: 0.1  # default in CPOTrainer
+cpo_alpha: 1.0  # default in CPOTrainer
+simpo_gamma: 0.5  # default in CPOTrainer
+```
+
+This method uses the same dataset format as [DPO](#dpo).
+
+### Using local dataset files
+
 ```yaml
 datasets:
   - ds_type: json
@@ -82,9 +606,9 @@ datasets:
     type: chatml.intel
 ```
 
-#### Trl autounwrap for peft
+### TRL auto-unwrapping for PEFT
 
-Trl supports autounwrapping peft models, so that a ref model does not need to be additionally loaded, leading to less VRAM needed. This is on by default. To turn it off, pass the following config.
+TRL supports auto-unwrapping PEFT models for RL training paradigms which rely on a reference model. This significantly reduces memory pressure as an additional refreference model does not need to be loaded, and reference model log-probabilities can be obtained by disabling PEFT adapters. This is enabled by default. To turn it off, pass the following config:
 
 ```yaml
 # load ref model when adapter training.

docs/sequence_parallelism.qmd

@@ -0,0 +1,97 @@
+---
+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
+```
+
+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

docs/torchao.qmd

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

docs/unsloth.qmd

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

examples/cerebras/btlm-ft.yml

@@ -8,9 +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
@@ -34,7 +31,6 @@ lora_alpha:
 lora_dropout:
 lora_target_modules:
 lora_target_linear:
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -58,16 +54,12 @@ 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 +72,6 @@ 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

@@ -22,7 +22,6 @@ lora_target_modules:
   - c_attn
   - c_proj
 lora_target_linear:
-lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
@@ -36,15 +35,10 @@ 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 +47,6 @@ 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

@@ -26,7 +26,6 @@ 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 +40,18 @@ 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

@@ -26,9 +26,7 @@ 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 +41,18 @@ 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

@@ -26,7 +26,6 @@ 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 +40,18 @@ 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

@@ -26,9 +26,7 @@ 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 +41,18 @@ 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

@@ -26,7 +26,6 @@ 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 +40,18 @@ 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

@@ -26,9 +26,7 @@ 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 +41,18 @@ 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 +1,59 @@
+base_model: CohereForAI/c4ai-command-r7b-12-2024
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+# 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

@@ -3,9 +3,6 @@ base_model: LnL-AI/dbrx-base-converted-v2
 # hub_model_id: username/custom_model_name
 
 trust_remote_code: true
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 datasets:
@@ -48,26 +45,20 @@ 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

@@ -48,26 +48,20 @@ 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

@@ -3,9 +3,6 @@ base_model: LnL-AI/dbrx-base-converted-v2
 # hub_model_id: username/custom_model_name
 
 trust_remote_code: true
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 datasets:
@@ -35,25 +32,19 @@ 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/deepseek-v2/fft-fsdp-16b.yaml

@@ -2,9 +2,6 @@ base_model: deepseek-ai/DeepSeek-V2-Lite
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 trust_remote_code: true
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 datasets:
@@ -31,27 +28,19 @@ 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

@@ -21,8 +21,9 @@ datasets:
     type: chat_template
     split: train[:20%]
     field_messages: conversations
-    message_field_role: from
-    message_field_content: value
+    message_property_mappings:
+      role: from
+      content: value
 
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0
@@ -51,27 +52,19 @@ 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

@@ -25,9 +25,7 @@ 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 +39,10 @@ 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 +51,7 @@ 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

@@ -38,9 +38,7 @@ 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 +65,7 @@ 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 +73,6 @@ 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 +81,7 @@ 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

@@ -7,9 +7,6 @@ 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:
@@ -25,9 +22,7 @@ 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 +36,10 @@ 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 +48,7 @@ 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

@@ -42,28 +42,16 @@ 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

@@ -16,8 +16,9 @@ datasets:
     type: chat_template
     drop_system_message: true
     field_messages: conversations
-    message_field_role: from
-    message_field_content: value
+    message_property_mappings:
+      role: from
+      content: value
 
 val_set_size: 0.0
 output_dir: ./outputs/out
@@ -47,28 +48,16 @@ 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

@@ -5,9 +5,6 @@ num_labels: 1
 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
@@ -38,8 +35,6 @@ 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 +42,12 @@ 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 +1,67 @@
+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
+strict: false
+
+# 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 +1,61 @@
+base_model: google/gemma-3-4b-it
+strict: false
+
+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 +1,63 @@
+base_model: google/gemma-3-4b-it
+processor_type: AutoProcessor
+strict: false
+
+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

@@ -18,9 +18,7 @@ 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 +32,10 @@ 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 +44,6 @@ 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

@@ -40,26 +40,18 @@ 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

@@ -39,26 +39,20 @@ 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

@@ -13,8 +13,9 @@ datasets:
     type: chat_template
     drop_system_message: true
     field_messages: conversations
-    message_field_role: from
-    message_field_content: value
+    message_property_mappings:
+      role: from
+      content: value
 
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.0
@@ -38,8 +39,6 @@ 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

@@ -4,9 +4,6 @@ model_type: LlamaForCausalLM
 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:
@@ -26,7 +23,6 @@ lora_r:
 lora_alpha:
 lora_dropout:
 lora_target_linear:
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -41,18 +37,12 @@ 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 +51,8 @@ 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

@@ -10,8 +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
@@ -33,7 +31,6 @@ lora_target_modules:
   - q_proj
   - v_proj
 lora_target_linear:
-lora_fan_in_fan_out:
 wandb_project:
 wandb_watch:
 wandb_name:
@@ -50,26 +47,19 @@ 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

@@ -4,9 +4,6 @@ model_type: LlamaForCausalLM
 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:
@@ -26,7 +23,6 @@ lora_r:
 lora_alpha:
 lora_dropout:
 lora_target_linear:
-lora_fan_in_fan_out:
 
 lisa_n_layers: 4
 lisa_step_interval: 20
@@ -45,18 +41,12 @@ 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 +55,8 @@ 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

@@ -4,9 +4,6 @@ model_type: LlamaForCausalLM
 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:
@@ -26,7 +23,6 @@ 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 +40,16 @@ 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

@@ -26,7 +26,6 @@ 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 +40,16 @@ 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

@@ -26,9 +26,7 @@ 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 +41,19 @@ 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

@@ -26,9 +26,7 @@ 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 +41,16 @@ 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

@@ -24,9 +24,7 @@ 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 +43,18 @@ 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

@@ -45,14 +45,11 @@ 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 +57,4 @@ 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

@@ -17,8 +17,9 @@ datasets:
     type: chat_template
     split: train[:20%]
     field_messages: conversations
-    message_field_role: from
-    message_field_content: value
+    message_property_mappings:
+      role: from
+      content: value
 
 dataset_prepared_path: last_run_prepared
 val_set_size: 0.02
@@ -41,27 +42,19 @@ 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

@@ -1,9 +1,6 @@
 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:
@@ -30,29 +27,19 @@ 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

@@ -17,8 +17,9 @@ datasets:
     field_messages: conversation
     field_chosen: chosen
     field_rejected: rejected
-    message_field_role: role
-    message_field_content: content
+    message_property_mappings:
+      role: role
+      content: content
     roles:
       system:
         - system
@@ -41,7 +42,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -56,28 +56,15 @@ 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

@@ -14,8 +14,9 @@ datasets:
   - path: fozziethebeat/alpaca_messages_2k_test
     type: chat_template
     field_messages: messages
-    message_field_role: role
-    message_field_content: content
+    message_property_mappings:
+      role: role
+      content: content
     roles:
       user:
         - user
@@ -36,7 +37,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -51,30 +51,17 @@ 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

@@ -17,8 +17,9 @@ datasets:
     field_messages: conversation
     field_chosen: chosen
     field_rejected: rejected
-    message_field_role: role
-    message_field_content: content
+    message_property_mappings:
+      role: role
+      content: content
     roles:
       system:
         - system
@@ -31,8 +32,9 @@ datasets:
     field_messages: conversation
     field_chosen: chosen
     field_rejected: rejected
-    message_field_role: role
-    message_field_content: content
+    message_property_mappings:
+      role: role
+      content: content
     roles:
       system:
         - system
@@ -56,7 +58,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -71,28 +72,15 @@ 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

@@ -19,7 +19,6 @@ 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 +31,6 @@ 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 +48,17 @@ 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

@@ -0,0 +1,68 @@
+base_model: NousResearch/Llama-3.2-1B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+strict: false
+
+datasets:
+  - path: teknium/GPT4-LLM-Cleaned
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/lora-out
+
+adapter: lora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+pad_to_sequence_len: true
+
+lora_r: 16
+lora_alpha: 32
+# Currently, we don't support dropout with our custom Triton kernels
+# lora_dropout: 0.05
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+# These options enable our custom Triton kernels / autograd
+# functions for MLP and attention calculations
+lora_mlp_kernel: true
+lora_qkv_kernel: true
+lora_o_kernel: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+loss_watchdog_threshold: 5.0
+loss_watchdog_patience: 3
+
+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-ray.yml

@@ -1,9 +1,6 @@
 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:
@@ -24,7 +21,6 @@ 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 +43,12 @@ 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 +57,9 @@ 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 +1,66 @@
+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
+strict: 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|>