perform flakey patched tests in individual runner

* fix build w pyproject to respect insalled torch version

* include in manifest

* disable duplicate code check for now

* move parser so it can be found

* add checks for correct pytorch version so this doesn't slip by again

.github/workflows/base.yml

@@ -1,6 +1,16 @@
 name: ci-cd-base
 
 on:
+  push:
+    branches:
+      - "main"
+    paths:
+      - 'Dockerfile-base'
+      - '.github/workflows/base.yml'
+  pull_request:
+    paths:
+      - 'Dockerfile-base'
+      - '.github/workflows/base.yml'
   workflow_dispatch:
 
 jobs:

.github/workflows/main.yml

@@ -49,7 +49,7 @@ jobs:
             axolotlai/axolotl
           tags: |
             type=ref,event=branch
-            type=semver,pattern={{version}}
+            type=pep440,pattern={{version}}
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
       - name: Login to Docker Hub
@@ -116,7 +116,7 @@ jobs:
             axolotlai/axolotl-cloud
           tags: |
             type=ref,event=branch
-            type=semver,pattern={{version}}
+            type=pep440,pattern={{version}}
       - name: Login to Docker Hub
         uses: docker/login-action@v3
         with:
@@ -163,7 +163,7 @@ jobs:
             axolotlai/axolotl-cloud-term
           tags: |
             type=ref,event=branch
-            type=semver,pattern={{version}}
+            type=pep440,pattern={{version}}
       - name: Login to Docker Hub
         uses: docker/login-action@v3
         with:

.github/workflows/pypi.yml

@@ -13,19 +13,13 @@ jobs:
     permissions:
       contents: write
     steps:
-      - name: Get the tag version
-        id: extract_branch
-        run: echo ::set-output name=branch::${GITHUB_REF#refs/tags/}
-        shell: bash
+      - name: Checkout code
+        uses: actions/checkout@v4
 
-      - name: Create Release
-        id: create_release
-        uses: actions/create-release@v1
+      - name: Create release
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        with:
-          tag_name: ${{ steps.extract_branch.outputs.branch }}
-          release_name: ${{ steps.extract_branch.outputs.branch }}
+        run: gh release create "$GITHUB_REF_NAME" --generate-notes
   pypi-publish:
     name: Upload release to PyPI
     runs-on: ubuntu-latest
@@ -47,7 +41,7 @@ jobs:
       - name: Install dependencies
         run: |
           pip3 install wheel packaging
-          pip3 install -e .
+          pip3 install --no-build-isolation -e .
           pip3 install -r requirements-dev.txt -r requirements-tests.txt
 
       - name: Extract tag name

.github/workflows/tests-nightly.yml

@@ -23,9 +23,15 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       fail-fast: false
+      max-parallel: 2
       matrix:
         python_version: ["3.10", "3.11"]
         pytorch_version: ["2.3.1", "2.4.1", "2.5.1"]
+        exclude:
+          - python_version: "3.10"
+            pytorch_version: "2.4.1"
+          - python_version: "3.10"
+            pytorch_version: "2.5.1"
     timeout-minutes: 20
 
     steps:
@@ -38,6 +44,11 @@ jobs:
           python-version: ${{ matrix.python_version }}
           cache: 'pip' # caching pip dependencies
 
+      - name: upgrade pip
+        run: |
+          pip3 install --upgrade pip
+          pip3 install --upgrade packaging setuptools wheel
+
       - name: Install PyTorch
         run: |
           pip3 install torch==${{ matrix.pytorch_version }} --index-url https://download.pytorch.org/whl/cpu
@@ -54,12 +65,23 @@ jobs:
         run: |
           pip3 install --upgrade pip
           pip3 install --upgrade packaging
-          pip3 install -U -e .
+          pip3 install --no-build-isolation -U -e .
+          python scripts/unsloth_install.py | sh
+          python scripts/cutcrossentropy_install.py | sh
           pip3 install -r requirements-dev.txt -r requirements-tests.txt
 
+      - name: Make sure PyTorch version wasn't clobbered
+        run: |
+          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
+
+      - name: Ensure axolotl CLI was installed
+        run: |
+          axolotl --help
+
       - name: Run tests
         run: |
-          pytest --ignore=tests/e2e/ tests/
+          pytest -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ tests/
+          pytest tests/patched/
 
       - name: cleanup pip cache
         run: |

.github/workflows/tests.yml

@@ -8,11 +8,17 @@ on:
       - '**.py'
       - 'requirements.txt'
       - '.github/workflows/*.yml'
+      - 'requirements-tests.txt'
+      - 'cicd/cicd.sh'
+      - 'cicd/Dockerfile.jinja'
   pull_request:
       paths:
        - '**.py'
        - 'requirements.txt'
        - '.github/workflows/*.yml'
+       - 'requirements-tests.txt'
+       - 'cicd/cicd.sh'
+       - 'cicd/Dockerfile.jinja'
   workflow_dispatch:
 
 # Cancel jobs on the same ref if a new one is triggered
@@ -39,9 +45,15 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       fail-fast: false
+      max-parallel: 2
       matrix:
         python_version: ["3.10", "3.11"]
         pytorch_version: ["2.3.1", "2.4.1", "2.5.1"]
+        exclude:
+          - python_version: "3.10"
+            pytorch_version: "2.4.1"
+          - python_version: "3.10"
+            pytorch_version: "2.5.1"
     timeout-minutes: 20
 
     steps:
@@ -66,12 +78,23 @@ jobs:
       - name: Install dependencies
         run: |
           pip3 show torch
-          pip3 install -U -e .
+          pip3 install --no-build-isolation -U -e .
+          python scripts/unsloth_install.py | sh
+          python scripts/cutcrossentropy_install.py | sh
           pip3 install -r requirements-dev.txt -r requirements-tests.txt
 
+      - name: Make sure PyTorch version wasn't clobbered
+        run: |
+          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
+
+      - name: Ensure axolotl CLI was installed
+        run: |
+          axolotl --help
+
       - name: Run tests
         run: |
-          pytest -n8 --ignore=tests/e2e/ tests/
+          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ tests/
+          pytest -v tests/patched/
 
       - name: cleanup pip cache
         run: |
@@ -82,6 +105,7 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       fail-fast: false
+      max-parallel: 1
       matrix:
         python_version: ["3.11"]
         pytorch_version: ["2.4.1", "2.5.1"]
@@ -100,7 +124,7 @@ jobs:
       - name: upgrade pip
         run: |
           pip3 install --upgrade pip
-          pip3 install --upgrade packaging setuptools wheel
+          pip3 install --upgrade packaging setuptools setuptools_scm build wheel
 
       - name: Install PyTorch
         run: |
@@ -109,13 +133,24 @@ jobs:
       - name: Install dependencies
         run: |
           pip3 show torch
-          python3 setup.py sdist
-          pip3 install dist/axolotl*.tar.gz
+          python -m build --no-isolation --sdist
+          pip3 install --no-build-isolation dist/axolotl*.tar.gz
+          python scripts/unsloth_install.py | sh
+          python scripts/cutcrossentropy_install.py | sh
           pip3 install -r requirements-dev.txt -r requirements-tests.txt
 
+      - name: Make sure PyTorch version wasn't clobbered
+        run: |
+          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
+
+      - name: Ensure axolotl CLI was installed
+        run: |
+          axolotl --help
+
       - name: Run tests
         run: |
-          pytest -n8 --ignore=tests/e2e/ tests/
+          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ tests/
+          pytest -v tests/patched/
 
       - name: cleanup pip cache
         run: |

.gitignore

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

MANIFEST.in

@@ -1,4 +1,5 @@
 include requirements.txt
 include README.md
 include LICENSE
+include src/setuptools_axolotl_dynamic_dependencies.py
 recursive-include axolotl *.py

README.md

@@ -10,9 +10,13 @@
     <img src="https://img.shields.io/github/license/axolotl-ai-cloud/axolotl.svg?color=blue" alt="GitHub License">
     <img src="https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/tests.yml/badge.svg" alt="tests">
     <a href="https://github.com/axolotl-ai-cloud/axolotl/releases"><img src="https://img.shields.io/github/release/axolotl-ai-cloud/axolotl.svg" alt="Releases"></a>
+    <br/>
+    <a href="https://github.com/axolotl-ai-cloud/axolotl/graphs/contributors"><img src="https://img.shields.io/github/contributors-anon/axolotl-ai-cloud/axolotl?color=yellow&style=flat-square" alt="contributors" style="height: 20px;"></a>
     <img src="https://img.shields.io/github/stars/axolotl-ai-cloud/axolotl" alt="GitHub Repo stars">
-</p>
-<p align="center">
+    <br/>
+    <a href="https://discord.com/invite/HhrNrHJPRb"><img src="https://img.shields.io/badge/discord-7289da.svg?style=flat-square&logo=discord" alt="discord" style="height: 20px;"></a>
+    <a href="https://twitter.com/axolotl_ai"><img src="https://img.shields.io/twitter/follow/axolotl_ai?style=social" alt="twitter" style="height: 20px;"></a>
+    <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">
 </p>
@@ -41,9 +45,13 @@ Features:
 ## Table of Contents
 - [Axolotl](#axolotl)
   - [Table of Contents](#table-of-contents)
-  - [Axolotl supports](#axolotl-supports)
   - [Quickstart ⚡](#quickstart-)
-    - [Usage](#usage)
+    - [Edge Builds](#edge-builds-)
+    - [Axolotl CLI Usage](#axolotl-cli-usage)
+  - [Badge ❤🏷️](#badge-️)
+  - [Contributing 🤝](#contributing-)
+  - [Sponsors 🤝❤](#sponsors-)
+  - [Axolotl supports](#axolotl-supports)
   - [Advanced Setup](#advanced-setup)
     - [Environment](#environment)
       - [Docker](#docker)
@@ -75,14 +83,6 @@ Features:
     - [Tokenization Mismatch b/w Inference \& Training](#tokenization-mismatch-bw-inference--training)
   - [Debugging Axolotl](#debugging-axolotl)
   - [Need help? 🙋](#need-help-)
-  - [Badge ❤🏷️](#badge-️)
-  - [Community Showcase](#community-showcase)
-  - [Contributing 🤝](#contributing-)
-  - [Sponsors 🤝❤](#sponsors-)
-      - [💎 Diamond Sponsors - Contact directly](#-diamond-sponsors---contact-directly)
-      - [🥇 Gold Sponsors - $5000/mo](#-gold-sponsors---5000mo)
-      - [🥈 Silver Sponsors - $1000/mo](#-silver-sponsors---1000mo)
-      - [🥉 Bronze Sponsors - $500/mo](#-bronze-sponsors---500mo)
 
 </td>
 <td>
@@ -105,6 +105,148 @@ Features:
 </tr>
 </table>
 
+## Quickstart ⚡
+
+Get started with Axolotl in just a few steps! This quickstart guide will walk you through setting up and running a basic fine-tuning task.
+
+**Requirements**: *Nvidia* GPU (Ampere architecture or newer for `bf16` and Flash Attention) or *AMD* GPU, Python >=3.10 and PyTorch >=2.3.1.
+
+```bash
+pip3 install --no-build-isolation axolotl[flash-attn,deepspeed]
+
+# download examples and optionally deepspeed configs to the local path
+axolotl fetch examples
+axolotl fetch deepspeed_configs  # OPTIONAL
+
+# finetune using lora
+axolotl train examples/llama-3/lora-1b.yml
+```
+
+### Edge Builds 🏎️
+
+If you're looking for the latest features and updates between releases, you'll need to install
+from source.
+
+```bash
+git clone https://github.com/axolotl-ai-cloud/axolotl.git
+cd axolotl
+pip3 install packaging ninja
+pip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'
+```
+
+### Axolotl CLI Usage
+We now support a new, more streamlined CLI using [click](https://click.palletsprojects.com/en/stable/).
+
+```bash
+# preprocess datasets - optional but recommended
+CUDA_VISIBLE_DEVICES="0" axolotl preprocess examples/llama-3/lora-1b.yml
+
+# finetune lora
+axolotl train examples/llama-3/lora-1b.yml
+
+# inference
+axolotl inference examples/llama-3/lora-1b.yml \
+    --lora-model-dir="./outputs/lora-out"
+
+# gradio
+axolotl inference examples/llama-3/lora-1b.yml \
+    --lora-model-dir="./outputs/lora-out" --gradio
+
+# remote yaml files - the yaml config can be hosted on a public URL
+# Note: the yaml config must directly link to the **raw** yaml
+axolotl train https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/examples/llama-3/lora-1b.yml
+```
+
+We've also added a new command for fetching `examples` and `deepspeed_configs` to your
+local machine. This will come in handy when installing `axolotl` from PyPI.
+
+```bash
+# Fetch example YAML files (stores in "examples/" folder)
+axolotl fetch examples
+
+# Fetch deepspeed config files (stores in "deepspeed_configs/" folder)
+axolotl fetch deepspeed_configs
+
+# Optionally, specify a destination folder
+axolotl fetch examples --dest path/to/folder
+```
+
+### Legacy Usage
+<details>
+
+<summary>Click to Expand</summary>
+
+While the Axolotl CLI is the preferred method for interacting with axolotl, we
+still support the legacy `-m axolotl.cli.*` usage.
+
+```bash
+# preprocess datasets - optional but recommended
+CUDA_VISIBLE_DEVICES="0" python -m axolotl.cli.preprocess examples/llama-3/lora-1b.yml
+
+# finetune lora
+accelerate launch -m axolotl.cli.train examples/llama-3/lora-1b.yml
+
+# inference
+accelerate launch -m axolotl.cli.inference examples/llama-3/lora-1b.yml \
+    --lora_model_dir="./outputs/lora-out"
+
+# gradio
+accelerate launch -m axolotl.cli.inference examples/llama-3/lora-1b.yml \
+    --lora_model_dir="./outputs/lora-out" --gradio
+
+# remote yaml files - the yaml config can be hosted on a public URL
+# Note: the yaml config must directly link to the **raw** yaml
+accelerate launch -m axolotl.cli.train https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/examples/llama-3/lora-1b.yml
+```
+
+</details>
+
+## Badge ❤🏷️
+
+Building something cool with Axolotl? Consider adding a badge to your model card.
+
+```markdown
+[<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)
+```
+
+[<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)
+
+## Sponsors 🤝❤
+
+If you love axolotl, consider sponsoring the project by reaching out directly to [wing@axolotl.ai](mailto:wing@axolotl.ai).
+
+---
+
+- [Modal](https://modal.com/) Modal lets you run data/AI jobs in the cloud, by just writing a few lines of Python. Customers use Modal to deploy Gen AI models at large scale, fine-tune LLM models, run protein folding simulations, and much more.
+
+---
+
+## Contributing 🤝
+
+Please read the [contributing guide](./.github/CONTRIBUTING.md)
+
+Bugs? Please check the [open issues](https://github.com/axolotl-ai-cloud/axolotl/issues/bug) else create a new Issue.
+
+PRs are **greatly welcome**!
+
+Please run the quickstart instructions followed by the below to setup env:
+```bash
+pip3 install -r requirements-dev.txt -r requirements-tests.txt
+pre-commit install
+
+# test
+pytest tests/
+
+# optional: run against all files
+pre-commit run --all-files
+```
+
+Thanks to all of our contributors to date. Help drive open source AI progress forward by contributing to Axolotl.
+
+<a href="https://github.com/axolotl-ai-cloud/axolotl/graphs/contributors">
+  <img src="https://contrib.rocks/image?repo=openaccess-ai-collective/axolotl" alt="contributor chart by https://contrib.rocks"/>
+</a>
+
 ## Axolotl supports
 
 |             | fp16/fp32 | lora | qlora | gptq | gptq w/flash attn | flash attn | xformers attn |
@@ -130,41 +272,6 @@ Features:
 ❌: not supported
 ❓: untested
 
-## Quickstart ⚡
-
-Get started with Axolotl in just a few steps! This quickstart guide will walk you through setting up and running a basic fine-tuning task.
-
-**Requirements**: Nvidia GPU (Ampere architecture or newer for `bf16` and Flash Attention), Python >=3.10 and PyTorch >=2.3.1.
-
-```bash
-git clone https://github.com/axolotl-ai-cloud/axolotl
-cd axolotl
-
-pip3 install packaging ninja
-pip3 install -e '.[flash-attn,deepspeed]'
-```
-
-### Usage
-```bash
-# preprocess datasets - optional but recommended
-CUDA_VISIBLE_DEVICES="" python -m axolotl.cli.preprocess examples/openllama-3b/lora.yml
-
-# finetune lora
-accelerate launch -m axolotl.cli.train examples/openllama-3b/lora.yml
-
-# inference
-accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
-    --lora_model_dir="./outputs/lora-out"
-
-# gradio
-accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
-    --lora_model_dir="./outputs/lora-out" --gradio
-
-# remote yaml files - the yaml config can be hosted on a public URL
-# Note: the yaml config must directly link to the **raw** yaml
-accelerate launch -m axolotl.cli.train https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/examples/openllama-3b/lora.yml
-```
-
 ## Advanced Setup
 
 ### Environment
@@ -213,7 +320,7 @@ docker run --privileged --gpus '"all"' --shm-size 10g --rm -it --name axolotl --
   3. Install Axolotl along with python dependencies
         ```bash
         pip3 install packaging
-        pip3 install -e '.[flash-attn,deepspeed]'
+        pip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'
         ```
   4. (Optional) Login to Huggingface to use gated models/datasets.
         ```bash
@@ -292,7 +399,7 @@ Please use WSL or Docker!
 
 Use the below instead of the install method in QuickStart.
 ```
-pip3 install -e '.'
+pip3 install --no-build-isolation -e '.'
 ```
 More info: [mac.md](/docs/mac.qmd)
 
@@ -682,86 +789,6 @@ See [this debugging guide](docs/debugging.qmd) for tips on debugging Axolotl, al
 
 ## Need help? 🙋
 
-Join our [Discord server](https://discord.gg/HhrNrHJPRb) where we our community members can help you.
+Join our [Discord server](https://discord.gg/HhrNrHJPRb) where our community members can help you.
 
-Need dedicated support? Please contact us at [✉️wing@openaccessaicollective.org](mailto:wing@openaccessaicollective.org) for dedicated support options.
-
-## Badge ❤🏷️
-
-Building something cool with Axolotl? Consider adding a badge to your model card.
-
-```markdown
-[<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)
-```
-
-[<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)
-
-## Community Showcase
-
-Check out some of the projects and models that have been built using Axolotl! Have a model you'd like to add to our Community Showcase? Open a PR with your model.
-
-Open Access AI Collective
-- [Minotaur 13b](https://huggingface.co/openaccess-ai-collective/minotaur-13b-fixed)
-- [Manticore 13b](https://huggingface.co/openaccess-ai-collective/manticore-13b)
-- [Hippogriff 30b](https://huggingface.co/openaccess-ai-collective/hippogriff-30b-chat)
-
-PocketDoc Labs
-- [Dan's PersonalityEngine 13b LoRA](https://huggingface.co/PocketDoc/Dans-PersonalityEngine-13b-LoRA)
-
-## Contributing 🤝
-
-Please read the [contributing guide](./.github/CONTRIBUTING.md)
-
-Bugs? Please check the [open issues](https://github.com/axolotl-ai-cloud/axolotl/issues/bug) else create a new Issue.
-
-PRs are **greatly welcome**!
-
-Please run the quickstart instructions followed by the below to setup env:
-```bash
-pip3 install -r requirements-dev.txt -r requirements-tests.txt
-pre-commit install
-
-# test
-pytest tests/
-
-# optional: run against all files
-pre-commit run --all-files
-```
-
-Thanks to all of our contributors to date. Help drive open source AI progress forward by contributing to Axolotl.
-
-<a href="https://github.com/axolotl-ai-cloud/axolotl/graphs/contributors">
-  <img src="https://contrib.rocks/image?repo=openaccess-ai-collective/axolotl" alt="contributor chart by https://contrib.rocks"/>
-</a>
-
-## Sponsors 🤝❤
-
-OpenAccess AI Collective is run by volunteer contributors such as [winglian](https://github.com/winglian),
-[NanoCode012](https://github.com/NanoCode012), [tmm1](https://github.com/tmm1),
-[mhenrichsen](https://github.com/mhenrichsen), [casper-hansen](https://github.com/casper-hansen),
-[hamelsmu](https://github.com/hamelsmu) and many more who help us accelerate forward by fixing bugs, answering
-community questions and implementing new features. Axolotl needs donations from sponsors for the compute needed to
-run our unit & integration tests, troubleshooting community issues, and providing bounties. If you love axolotl,
-consider sponsoring the project via [GitHub Sponsors](https://github.com/sponsors/OpenAccess-AI-Collective),
-[Ko-fi](https://ko-fi.com/axolotl_ai) or reach out directly to
-[wing@openaccessaicollective.org](mailto:wing@openaccessaicollective.org).
-
----
-
-#### 💎 Diamond Sponsors - [Contact directly](mailto:wing@openaccessaicollective.org)
-
----
-
-#### 🥇 Gold Sponsors - $5000/mo
-
----
-
-#### 🥈 Silver Sponsors - $1000/mo
-
----
-
-#### 🥉 Bronze Sponsors - $500/mo
-
- - [JarvisLabs.ai](https://jarvislabs.ai)
-
----
+Need dedicated support? Please contact us at [✉️wing@axolotl.ai](ailto:wing@axolotl.ai) for dedicated support options.

cicd/Dockerfile.jinja

@@ -4,7 +4,6 @@ ENV TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX"
 ENV AXOLOTL_EXTRAS="{{ AXOLOTL_EXTRAS }}"
 ENV AXOLOTL_ARGS="{{ AXOLOTL_ARGS }}"
 ENV CUDA="{{ CUDA }}"
-ENV BNB_CUDA_VERSION="{{ CUDA }}"
 ENV PYTORCH_VERSION="{{ PYTORCH_VERSION }}"
 ENV GITHUB_REF="{{ GITHUB_REF }}"
 ENV GITHUB_SHA="{{ GITHUB_SHA }}"
@@ -32,11 +31,14 @@ RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
     fi
 
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        pip install -e .[deepspeed,flash-attn,optimizers,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,optimizers,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \
-        pip install -e .[deepspeed,flash-attn,optimizers] $AXOLOTL_ARGS; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,optimizers] $AXOLOTL_ARGS; \
     fi
 
+RUN python scripts/unsloth_install.py | sh
+RUN python scripts/cutcrossentropy_install.py | sh
+
 # So we can test the Docker image
 RUN pip install -r requirements-dev.txt -r requirements-tests.txt
 

cicd/cicd.sh

@@ -1,6 +1,15 @@
 #!/bin/bash
 set -e
 
-pytest -n8 --ignore=tests/e2e/ /workspace/axolotl/tests/
-pytest -n1 --dist loadfile -v /workspace/axolotl/tests/e2e/patched/ /workspace/axolotl/tests/e2e/integrations/
-pytest --ignore=tests/e2e/patched/ --ignore=tests/e2e/multigpu/ --ignore=tests/e2e/integrations/ /workspace/axolotl/tests/e2e/
+python -c "import torch; assert '$PYTORCH_VERSION' in torch.__version__"
+
+pytest -v --durations=10 -n8 --ignore=tests/e2e/ --ignore=tests/patched/ /workspace/axolotl/tests/
+# pytest -v --durations=10 -n8 --dist loadfile /workspace/axolotl/tests/patched/
+pytest -v --durations=10 -n1 --dist loadfile /workspace/axolotl/tests/e2e/patched/
+pytest -v --durations=10 -n1 --dist loadfile /workspace/axolotl/tests/e2e/integrations/
+pytest -v --durations=10 --ignore=tests/e2e/patched/ --ignore=tests/e2e/multigpu/ --ignore=tests/e2e/integrations/ /workspace/axolotl/tests/e2e/
+
+tests=$(pytest --collect-only -q tests/e2e/each)
+for t in $tests; do
+    pytest $t
+done

cicd/tests.py

@@ -40,6 +40,7 @@ with open(pathlib.Path(temp_dir) / "Dockerfile", "w", encoding="utf-8") as f:
 cicd_image = (
     Image.from_dockerfile(
         pathlib.Path(temp_dir) / "Dockerfile",
+        context_mount=None,
         force_build=True,
         gpu="A10G",
     )

docker/Dockerfile

@@ -5,7 +5,6 @@ ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX"
 ARG AXOLOTL_EXTRAS=""
 ARG AXOLOTL_ARGS=""
 ARG CUDA="118"
-ENV BNB_CUDA_VERSION=$CUDA
 ARG PYTORCH_VERSION="2.1.2"
 
 ENV PYTORCH_VERSION=$PYTORCH_VERSION
@@ -21,11 +20,14 @@ WORKDIR /workspace/axolotl
 
 # If AXOLOTL_EXTRAS is set, append it in brackets
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        pip install -e .[deepspeed,flash-attn,optimizers,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,optimizers,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \
-        pip install -e .[deepspeed,flash-attn,optimizers] $AXOLOTL_ARGS; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,optimizers] $AXOLOTL_ARGS; \
     fi
 
+RUN python scripts/unsloth_install.py | sh
+RUN python scripts/cutcrossentropy_install.py | sh
+
 # So we can test the Docker image
 RUN pip install pytest
 

docker/Dockerfile-base

@@ -16,7 +16,7 @@ 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 && rm -rf /var/lib/apt/lists/* \
+    && 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 \
@@ -29,7 +29,9 @@ 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$CUDA
+    python3 -m pip install --no-cache-dir -U torch==${PYTORCH_VERSION}+cu${CUDA} --extra-index-url https://download.pytorch.org/whl/cu$CUDA && \
+    python3 -m pip install --no-cache-dir "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 && \

docker/Dockerfile-cloud

@@ -2,7 +2,7 @@ ARG BASE_TAG=main
 FROM axolotlai/axolotl:$BASE_TAG
 
 ENV HF_DATASETS_CACHE="/workspace/data/huggingface-cache/datasets"
-ENV HUGGINGFACE_HUB_CACHE="/workspace/data/huggingface-cache/hub"
+ENV HF_HUB_CACHE="/workspace/data/huggingface-cache/hub"
 ENV HF_HOME="/workspace/data/huggingface-cache/hub"
 ENV HF_HUB_ENABLE_HF_TRANSFER="1"
 

docker/Dockerfile-cloud-no-tmux

@@ -2,7 +2,7 @@ ARG BASE_TAG=main
 FROM axolotlai/axolotl:$BASE_TAG
 
 ENV HF_DATASETS_CACHE="/workspace/data/huggingface-cache/datasets"
-ENV HUGGINGFACE_HUB_CACHE="/workspace/data/huggingface-cache/hub"
+ENV HF_HUB_CACHE="/workspace/data/huggingface-cache/hub"
 ENV HF_HOME="/workspace/data/huggingface-cache/hub"
 ENV HF_HUB_ENABLE_HF_TRANSFER="1"
 

docker/Dockerfile-tests

@@ -5,7 +5,6 @@ ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX"
 ARG AXOLOTL_EXTRAS=""
 ARG AXOLOTL_ARGS=""
 ARG CUDA="118"
-ENV BNB_CUDA_VERSION=$CUDA
 ARG PYTORCH_VERSION="2.1.2"
 ARG GITHUB_REF="main"
 
@@ -25,9 +24,9 @@ RUN git fetch origin +$GITHUB_REF && \
 
 # If AXOLOTL_EXTRAS is set, append it in brackets
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        pip install -e .[deepspeed,flash-attn,mamba-ssm,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,mamba-ssm,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \
-        pip install -e .[deepspeed,flash-attn,mamba-ssm] $AXOLOTL_ARGS; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,mamba-ssm] $AXOLOTL_ARGS; \
     fi
 
 # So we can test the Docker image

docs/amd_hpc.qmd

@@ -52,7 +52,7 @@ export GPU_ARCHS="gfx90a"
 cd flash-attention
 export PYTHON_SITE_PACKAGES=$(python -c 'import site; print(site.getsitepackages()[0])')
 patch "${PYTHON_SITE_PACKAGES}/torch/utils/hipify/hipify_python.py" hipify_patch.patch
-pip install .
+pip install --no-build-isolation .
 ```
 
 ### 6. Install Axolotl
@@ -63,7 +63,7 @@ Clone and install Axolotl:
 git clone https://github.com/axolotl-ai-cloud/axolotl
 cd axolotl
 pip install packaging ninja
-pip install -e .
+pip install --no-build-isolation -e .
 ```
 
 ### 7. Apply xformers Workaround

docs/config.qmd

@@ -162,6 +162,9 @@ datasets:
 # The same applies to the `test_datasets` option and the `pretraining_dataset` option. Default is true.
 shuffle_merged_datasets: true
 
+Deduplicates datasets and test_datasets with identical entries.
+dataset_exact_deduplication: true
+
 # A list of one or more datasets to eval the model with.
 # You can use either test_datasets, or val_set_size, but not both.
 test_datasets:
@@ -406,7 +409,7 @@ lr_div_factor: # Learning rate div factor
 # - adamw_torch_fused
 # - adamw_torch_xla
 # - adamw_apex_fused
-# - adopt_adamw (only for torch version >= 2.5.1)
+# - adopt_adamw (an EXPERIMENTAL optimizer, only for torch version >= 2.5.1)
 # - adafactor
 # - adamw_anyprecision
 # - sgd

docs/debugging.qmd

@@ -71,7 +71,7 @@ Make sure you have an [editable install](https://setuptools.pypa.io/en/latest/us
 
 ```bash
 pip3 install packaging
-pip3 install -e '.[flash-attn,deepspeed]'
+pip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'
 ```
 
 #### Remote Hosts
@@ -212,7 +212,7 @@ You will now be in the container.  Next, perform an editable install of Axolotl:
 
 ```bash
 pip3 install packaging
-pip3 install -e '.[flash-attn,deepspeed]'
+pip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'
 ```
 
 ### Attach To Container

docs/rlhf.qmd

@@ -52,6 +52,26 @@ datasets:
     type: chat_template.argilla
 ```
 
+
+#### KTO
+
+```yaml
+rl: kto
+rl_beta: 0.5
+kto_desirable_weight: 0.2
+
+remove_unused_columns: false
+
+datasets:
+  - path: argilla/ultrafeedback-binarized-preferences-cleaned-kto
+    type: llama3.ultra
+    split: train
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: true
+```
+
 #### Using local dataset files
 ```yaml
 datasets:

docs/unsloth.qmd

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

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

@@ -2,19 +2,15 @@
  "cells": [
   {
    "cell_type": "markdown",
-   "metadata": {
-    "id": "AKjdG7tbTb-n"
-   },
+   "metadata": {},
    "source": [
-    "# Example notebook for running Axolotl on google colab"
+    "## Setting up"
    ]
   },
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "id": "RcbNpOgWRcii"
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
     "import torch\n",
@@ -22,82 +18,76 @@
     "assert (torch.cuda.is_available()==True)"
    ]
   },
-  {
-   "cell_type": "markdown",
-   "metadata": {
-    "id": "h3nLav8oTRA5"
-   },
-   "source": [
-    "## Install Axolotl and dependencies"
-   ]
-  },
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "colab": {
-     "base_uri": "https://localhost:8080/"
-    },
-    "id": "3c3yGAwnOIdi",
-    "outputId": "e3777b5a-40ef-424f-e181-62dfecd1dd01"
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
-    "!pip install -e git+https://github.com/axolotl-ai-cloud/axolotl#egg=axolotl\n",
-    "!pip install flash-attn==\"2.7.0.post2\"\n",
-    "!pip install deepspeed==\"0.13.1\"!pip install mlflow==\"2.13.0\""
+    "!pip install --no-build-isolation axolotl[deepspeed]"
    ]
   },
   {
    "cell_type": "markdown",
-   "metadata": {
-    "id": "BW2MFr7HTjub"
-   },
+   "metadata": {},
    "source": [
-    "## Create an yaml config file"
+    "## Hugging Face login (optional)"
    ]
   },
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "id": "9pkF2dSoQEUN"
-   },
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from huggingface_hub import notebook_login\n",
+    "notebook_login()"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Example configuration"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
    "outputs": [],
    "source": [
     "import yaml\n",
     "\n",
-    "# Your YAML string\n",
     "yaml_string = \"\"\"\n",
-    "base_model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T\n",
-    "model_type: LlamaForCausalLM\n",
-    "tokenizer_type: LlamaTokenizer\n",
+    "base_model: NousResearch/Meta-Llama-3.1-8B\n",
     "\n",
     "load_in_8bit: false\n",
     "load_in_4bit: true\n",
     "strict: false\n",
     "\n",
     "datasets:\n",
-    "  - path: mhenrichsen/alpaca_2k_test\n",
+    "  - path: tatsu-lab/alpaca\n",
     "    type: alpaca\n",
-    "dataset_prepared_path:\n",
+    "dataset_prepared_path: last_run_prepared\n",
     "val_set_size: 0.05\n",
-    "output_dir: ./outputs/qlora-out\n",
+    "output_dir: ./outputs/lora-out\n",
+    "\n",
+    "sequence_len: 2048\n",
+    "sample_packing: true\n",
+    "eval_sample_packing: true\n",
+    "pad_to_sequence_len: true\n",
     "\n",
     "adapter: qlora\n",
     "lora_model_dir:\n",
-    "\n",
-    "sequence_len: 4096\n",
-    "sample_packing: true\n",
-    "eval_sample_packing: false\n",
-    "pad_to_sequence_len: true\n",
-    "\n",
     "lora_r: 32\n",
     "lora_alpha: 16\n",
     "lora_dropout: 0.05\n",
-    "lora_target_modules:\n",
     "lora_target_linear: true\n",
     "lora_fan_in_fan_out:\n",
+    "lora_modules_to_save:\n",
+    "  - embed_tokens\n",
+    "  - lm_head\n",
     "\n",
     "wandb_project:\n",
     "wandb_entity:\n",
@@ -105,12 +95,12 @@
     "wandb_name:\n",
     "wandb_log_model:\n",
     "\n",
-    "gradient_accumulation_steps: 4\n",
-    "micro_batch_size: 2\n",
-    "num_epochs: 4\n",
-    "optimizer: paged_adamw_32bit\n",
+    "gradient_accumulation_steps: 2\n",
+    "micro_batch_size: 1\n",
+    "num_epochs: 1\n",
+    "optimizer: paged_adamw_8bit\n",
     "lr_scheduler: cosine\n",
-    "learning_rate: 0.0002\n",
+    "learning_rate: 2e-5\n",
     "\n",
     "train_on_inputs: false\n",
     "group_by_length: false\n",
@@ -121,13 +111,15 @@
     "gradient_checkpointing: true\n",
     "early_stopping_patience:\n",
     "resume_from_checkpoint:\n",
-    "local_rank:\n",
     "logging_steps: 1\n",
     "xformers_attention:\n",
-    "flash_attention: true\n",
+    "flash_attention: false\n",
+    "sdp_attention: true\n",
     "\n",
-    "warmup_steps: 10\n",
-    "evals_per_epoch: 4\n",
+    "warmup_steps: 1\n",
+    "max_steps: 25\n",
+    "evals_per_epoch: 1\n",
+    "eval_table_size:\n",
     "saves_per_epoch: 1\n",
     "debug:\n",
     "deepspeed:\n",
@@ -135,9 +127,10 @@
     "fsdp:\n",
     "fsdp_config:\n",
     "special_tokens:\n",
-    "\n",
+    "  pad_token: <|end_of_text|>\n",
     "\"\"\"\n",
     "\n",
+    "\n",
     "# Convert the YAML string to a Python dictionary\n",
     "yaml_dict = yaml.safe_load(yaml_string)\n",
     "\n",
@@ -146,31 +139,124 @@
     "\n",
     "# Write the YAML file\n",
     "with open(file_path, 'w') as file:\n",
-    "    yaml.dump(yaml_dict, file)\n"
+    "    yaml.dump(yaml_dict, file)"
    ]
   },
   {
    "cell_type": "markdown",
-   "metadata": {
-    "id": "bidoj8YLTusD"
-   },
+   "metadata": {},
    "source": [
-    "## Launch the training"
+    "Above we have a configuration file with base LLM model and datasets specified, among many other things. Axolotl can automatically detect whether the specified datasets are on HuggingFace repo or local machine.\n",
+    "\n",
+    "The Axolotl configuration options encompass model and dataset selection, data pre-processing, and training. Let's go through them line by line:\n",
+    "\n",
+    "*   \"base model\": String value, specifies the underlying pre-trained LLM that will be used for finetuning\n",
+    "\n",
+    "Next we have options for model weights quantization. Quantization allows for reduction in occupied memory on GPUs.\n",
+    "\n",
+    "*   \"load_in_8bit\": Boolean value, whether to quantize the model weights into 8-bit integer.\n",
+    "\n",
+    "*   \"load_in_4bit\": Boolean value, whether to quantize the model weights into 4-bit integer.\n",
+    "\n",
+    "*   \"strict\": Boolean value. If false, it allows for overriding established configuration options in the yaml file when executing in command-line interface.\n",
+    "\n",
+    "*   \"datasets\": a list of dicts that contain path and type of data sets as well as other optional configurations where datasets are concerned. Supports multiple datasets.\n",
+    "\n",
+    "*   \"val_set_size\": Either a float value less than one or an integer less than the total size of dataset. Sets the size of validation set from the whole dataset. If float, sets the proportion of the dataset assigned for validation. If integer, sets the direct size of validation set.\n",
+    "\n",
+    "*   \"output_dir\": String value. Path of trained model.\n",
+    "\n",
+    "For data preprocessing:\n",
+    "\n",
+    "*   \"sequence_len\": Integer. Specifies the maximum sequence length of the input. Typically 2048 or less.\n",
+    "\n",
+    "*   \"pad_to_sequence_len\": Boolean. Padding input to maximum sequence length.\n",
+    "\n",
+    "*   \"sample_packing\": Boolean. Specifies whether to use multi-packing with block diagonal attention.\n",
+    "\n",
+    "*   \"special_tokens\": Python dict, optional. Allows users to specify the additional special tokens to be ignored by the tokenizer.\n",
+    "\n",
+    "For LoRA configuration and its hyperparamters:\n",
+    "\n",
+    "*   \"adapter\": String. Either \"lora\" or \"qlora\", depending on user's choice.\n",
+    "\n",
+    "*   \"lora_model_dir\": String, Optional. Path to directory that contains LoRA model, if there is already a trained LoRA model the user would like to use.\n",
+    "\n",
+    "*   \"lora_r\": Integer. Refers to the rank of LoRA decomposition matrices. Higher value will reduce LoRA efficiency. Recommended to be set to 8.\n",
+    "\n",
+    "*   \"lora_alpha\": Integer. Scale the weight matrices by $\\frac{\\text{lora_alpha}}{\\text{lora_r}}$Recommended to be fixed at 16.\n",
+    "\n",
+    "*   \"lora_dropout\": Float that is 1 or less. The dropout probability of a lora layer.\n",
+    "\n",
+    "*   \"lora_target_linear\": Boolean. If true, lora will target all linear modules in the transformers architecture.\n",
+    "\n",
+    "*   \"lora_modules_to_save\": If you added new tokens to the tokenizer, you may need to save some LoRA modules because they need to know the new tokens.\n",
+    "\n",
+    "See [LoRA](https://arxiv.org/abs/2106.09685) for detailed explanation of LoRA implementation.\n",
+    "\n",
+    "For the training configurations:\n",
+    "\n",
+    "*   \"gradient_accumulation_steps\": Integer. The number of steps over which to accumulate gradient for batch training. E.g. if 2, backprop is performed every two steps.\n",
+    "\n",
+    "*   \"micro_batch_size\": Integer. Batch size per gpu / gradient_accumulation_steps\n",
+    "\n",
+    "*   \"num_epochs\": Integer. Number of epochs. One epoch is when training has looped over every batch in the whole data set once.\n",
+    "\n",
+    "*   \"optimizer\": The optimizer to use for the training.\n",
+    "\n",
+    "*   \"learning_rate\": The learning rate.\n",
+    "\n",
+    "*   \"lr_scheduler\": The learning rate scheduler to use for adjusting learning rate during training.\n",
+    "\n",
+    "*   \"train_on_inputs\": Boolean. Whether to ignore or include the user's prompt from the training labels.\n",
+    "\n",
+    "*   \"group_by_length\": Boolean. Whether to group similarly sized data to minimize padding.\n",
+    "\n",
+    "*   \"bf16\": Either \"auto\", \"true\", or \"false\". Whether to use CUDA bf16 floating point format. If set to \"auto\", will automatically apply bf16 should the gpu supports it.\n",
+    "\n",
+    "*   \"fp16\": Optional. Specifies whether to use CUDA fp16. Automatically set to true if \"bf16\" is set to true. Otherwise false.\n",
+    "\n",
+    "*   \"tf32\": Boolean. Whether to use CUDA tf32. Will override bf16.\n",
+    "\n",
+    "*   \"gradient_checkpointing\": Boolean. Whether to use gradient checkpointing https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing\n",
+    "\n",
+    "*   \"gradient_checkpointing_kwargs\": Python Dict. Fed into the trainer.\n",
+    "\n",
+    "*   \"logging_steps\": Integer. Log training information over every specified number of steps.\n",
+    "\n",
+    "*   \"flash_attention\": Boolean. Whether to use the [flash attention](https://github.com/Dao-AILab/flash-attention) mechanism.\n",
+    "\n",
+    "*   \"sdp_attention\": Boolean. Whether to use the Scaled Dot Product attention mechanism (the attention mechanism in the [original implementation](https://arxiv.org/abs/1706.03762) of transformers.)\n",
+    "\n",
+    "*   \"warmup_steps\": Integer. The number of pre-training steps where a very low learning rate is used.\n",
+    "\n",
+    "*   \"evals_per_epoch\": Integer. Number of evaluations to be performed within one training epoch.\n",
+    "\n",
+    "*   \"saves_per_epoch\": Integer. Number of times the model is saved in one training epoch.\n",
+    "\n",
+    "*   \"weight_decay\": Positive Float. Sets the \"strength\" of weight decay (i.e. setting the coefficient of L2 regularization)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "The above is but a snippet aiming to get users familiarized with the types of streamlined configuration options axolotl provides. For a full list of configuration options, see [here](https://axolotl-ai-cloud.github.io/axolotl/docs/config.html)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Train the model"
    ]
   },
   {
    "cell_type": "code",
    "execution_count": null,
-   "metadata": {
-    "colab": {
-     "base_uri": "https://localhost:8080/"
-    },
-    "id": "ydTI2Jk2RStU",
-    "outputId": "d6d0df17-4b53-439c-c802-22c0456d301b"
-   },
+   "metadata": {},
    "outputs": [],
    "source": [
-    "# By using the ! the comand will be executed as a bash command\n",
     "!accelerate launch -m axolotl.cli.train /content/test_axolotl.yaml"
    ]
   },
@@ -178,7 +264,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "## Play with inference"
+    "Predict with trained model"
    ]
   },
   {
@@ -187,36 +273,85 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# By using the ! the comand will be executed as a bash command\n",
     "!accelerate launch -m axolotl.cli.inference /content/test_axolotl.yaml \\\n",
-    "    --qlora_model_dir=\"./qlora-out\" --gradio"
+    "    --lora_model_dir=\"./outputs/lora-out\" --gradio"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Deeper Dive"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "It is also helpful to gain some familiarity over some of the core inner workings of axolotl"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Configuration Normalization"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Axolotl uses a custom Dict class, called ```DictDefault```\n",
+    "to store configurations specified in the yaml configuration file (into a Python variable named ```cfg```). The definition for this custom Dict can be found in the [utils/dict.py](https://github.com/axolotl-ai-cloud/axolotl/blob/main/src/axolotl/utils/dict.py)\n",
+    "\n",
+    "```DictDefault``` is amended such that calling a missing key from it will result in a ```None``` return type. This is important because if some configuration options aren't specified by the user, the ```None``` type allows Axolotl to perform boolean operations to determine the default settings for missing configurations. For more examples on how this is done, check out [utils/config/__init__.py](https://github.com/axolotl-ai-cloud/axolotl/blob/main/src/axolotl/utils/config/__init__.py)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Loading Models, Tokenizers, and Trainer"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "If we inspect [cli.train.py](https://github.com/axolotl-ai-cloud/axolotl/blob/main/src/axolotl/cli/train.py), we will find that most of the heavy lifting were done by the function ```train()``` which is itself imported from [src/axolotl/train.py](https://github.com/axolotl-ai-cloud/axolotl/blob/main/src/axolotl/train.py).\n",
+    "\n",
+    "```train()``` takes care of loading the appropriate tokenizer and pre-trained model through ```load_model()``` and ```load_tokenizer()``` from [src/axolotl/utils/models.py](https://github.com/axolotl-ai-cloud/axolotl/blob/main/src/axolotl/utils/models.py) respectively.\n",
+    "\n",
+    "```load_tokenizer()``` loads in the appropriate tokenizer given the desired model, as well as chat templates.\n",
+    "\n",
+    "```ModelLoader``` class follows after tokenizer has been selected. It will automatically discern the base model type, load in the desired model, as well as applying model-appropriate attention mechanism modifications (e.g. flash attention). Depending on which base model the user chooses in the configuration, ```ModelLoader``` will utilize the corresponding \"attention hijacking\" script. For example, if the user specified the base model to be ```NousResearch/Meta-Llama-3.1-8B```, which is of llama type, and set ```flash_attn``` to ```True```, ```ModelLoader``` will load in [llama_attn_hijack_flash.py](https://github.com/axolotl-ai-cloud/axolotl/blob/main/src/axolotl/monkeypatch/llama_attn_hijack_flash.py). For a list of supported attention hijacking, please refer to the directory [/src/axolotl/monkeypatch/](https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/monkeypatch)\n",
+    "\n",
+    "Another important operation encompassed in ```train()``` is setting up the training that takes into account of user-specified traning configurations (e.g. num_epochs, optimizer) through the use of ```setup_trainer()``` from [/src/axolotl/utils/trainer.py](https://github.com/axolotl-ai-cloud/axolotl/blob/main/src/axolotl/utils/trainer.py), which in turn relies on modules from [/src/axolotl/core/trainer_builder.py](https://github.com/axolotl-ai-cloud/axolotl/blob/main/src/axolotl/core/trainer_builder.py).\n",
+    "```trainer_builder.py``` provides a list of trainer object options bespoke for the task type (Causal or Reinforcement learning ('dpo', 'ipo', 'kto') )"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Monkey patch\n",
+    "\n",
+    "The [Monkey patch directory](https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/monkeypatch) is where model architecture/optimization patching scripts are stored (these are modifications that are not implemented in the official releases, hence the name monkey patch). It includes attention jacking, ReLoRA, and unsloth optimization."
    ]
   }
  ],
  "metadata": {
-  "accelerator": "GPU",
-  "colab": {
-   "gpuType": "T4",
-   "provenance": []
-  },
   "kernelspec": {
-   "display_name": "Python 3 (ipykernel)",
+   "display_name": "Python 3",
    "language": "python",
    "name": "python3"
   },
   "language_info": {
-   "codemirror_mode": {
-    "name": "ipython",
-    "version": 3
-   },
-   "file_extension": ".py",
-   "mimetype": "text/x-python",
    "name": "python",
-   "nbconvert_exporter": "python",
-   "pygments_lexer": "ipython3",
-   "version": "3.12.1"
+   "version": "3.9.6"
   }
  },
  "nbformat": 4,
- "nbformat_minor": 4
+ "nbformat_minor": 2
 }

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

@@ -0,0 +1,95 @@
+base_model: meta-llama/Llama-3.2-1B
+model_type: LlamaForCausalLM
+tokenizer_type: AutoTokenizer
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+chat_template: llama3
+rl: dpo
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_dpo_test
+    type: chat_template.default
+    field_messages: conversation
+    field_chosen: chosen
+    field_rejected: rejected
+    message_field_role: role
+    message_field_content: content
+    roles:
+      system:
+        - system
+      user:
+        - user
+      assistant:
+        - assistant
+  - path: fozziethebeat/alpaca_messages_2k_dpo_test
+    type: chat_template.default
+    field_messages: conversation
+    field_chosen: chosen
+    field_rejected: rejected
+    message_field_role: role
+    message_field_content: content
+    roles:
+      system:
+        - system
+      user:
+        - user
+      assistant:
+        - assistant
+
+dataset_exact_deduplication: true
+dataset_prepared_path:
+val_set_size: 0
+output_dir: ./outputs/lora-out
+
+sequence_len: 4096
+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_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+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

@@ -0,0 +1,76 @@
+base_model: meta-llama/Llama-3.2-1B
+model_type: LlamaForCausalLM
+tokenizer_type: AutoTokenizer
+
+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
+
+dataset_exact_deduplication: true
+test_value: true
+
+sequence_len: 4096
+sample_packing: true
+eval_sample_packing: false
+pad_to_sequence_len: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+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
+
+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.yml

@@ -0,0 +1,74 @@
+base_model: NousResearch/Llama-3.2-1B
+
+load_in_8bit: false
+load_in_4bit: false
+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
+eval_sample_packing: true
+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
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+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
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+loss_watchdog_threshold: 5.0
+loss_watchdog_patience: 3
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  pad_token: "<|end_of_text|>"

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

@@ -0,0 +1,75 @@
+base_model: meta-llama/Llama-3.2-1B
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+rl: kto
+rl_beta: 0.5
+kto_desirable_weight: 0.2
+
+datasets:
+  - path: argilla/ultrafeedback-binarized-preferences-cleaned-kto
+    type: llama3.ultra
+    split: train
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/qlora-out
+
+remove_unused_columns: false
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: false  # not supported with kto
+eval_sample_packing: false
+pad_to_sequence_len: false
+
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 20
+evals_per_epoch: 4
+eval_table_size:
+eval_max_new_tokens: 128
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  pad_token: "<|end_of_text|>"

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

@@ -1,4 +1,4 @@
-base_model: meta-llama/Llama-3.2-1B
+base_model: NousResearch/Llama-3.2-1B
 
 load_in_8bit: false
 load_in_4bit: true
@@ -22,7 +22,6 @@ pad_to_sequence_len: true
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_linear: true
 lora_fan_in_fan_out:
 lora_target_modules:
   - gate_proj

pyproject.toml

@@ -0,0 +1,26 @@
+[build-system]
+requires = ["setuptools>=64", "wheel", "setuptools_scm>=8"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "axolotl"
+dynamic = ["version", "dependencies", "optional-dependencies"]
+description = "LLM Trainer"
+readme = "README.md"
+requires-python = ">=3.10"
+
+[project.scripts]
+axolotl = "axolotl.cli.main:main"
+
+[project.urls]
+Homepage = "https://axolotl-ai-cloud.github.io/axolotl/"
+Repository = "https://github.com/axolotl-ai-cloud/axolotl.git"
+
+[tool.setuptools_scm]
+
+[tool.setuptools]
+py-modules = ["setuptools_axolotl_dynamic_dependencies"]
+include-package-data = true
+
+[tool.setuptools.cmdclass]
+build_py = "setuptools_axolotl_dynamic_dependencies.BuildPyCommand"

requirements-dev.txt

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

requirements-tests.txt

@@ -1,3 +1,5 @@
 pytest
 pytest-xdist
 pytest-retry
+pytest-sugar
+tbparse

requirements.txt

@@ -1,22 +1,30 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
+
+# START section of dependencies that don't install on Darwin/MacOS
+bitsandbytes==0.45.0
+triton>=2.3.0
+mamba-ssm==1.2.0.post1
+flash-attn==2.7.0.post2
+xformers>=0.0.23.post1
+autoawq==0.2.7.post3
+liger-kernel==0.4.2
+# END section
+
 packaging==23.2
-peft==0.13.2
-transformers==4.46.3
+peft==0.14.0
+transformers==4.47.0
 tokenizers>=0.20.1
-bitsandbytes==0.44.1
-accelerate==1.1.0
+accelerate==1.2.0
 datasets==3.1.0
-deepspeed==0.15.4
+deepspeed==0.16.1
 pydantic==2.6.3
 addict
 fire
 PyYAML>=6.0
 requests
-flash-attn==2.7.0.post2
 sentencepiece
 wandb
 einops
-xformers>=0.0.23.post1
 optimum==1.16.2
 hf_transfer
 colorama
@@ -26,23 +34,18 @@ numpy>=1.24.4,<=2.0.1
 evaluate==0.4.1
 scipy
 scikit-learn==1.4.2
-pynvml
+nvidia-ml-py==12.560.30
 art
 gradio==3.50.2
 tensorboard
 python-dotenv==1.0.1
-autoawq==0.2.7.post2
-triton>=2.3.0
-liger-kernel==0.4.1
-
-mamba-ssm==1.2.0.post1
 
 # remote filesystems
 s3fs>=2024.5.0
 gcsfs>=2024.5.0
 # adlfs
 
-trl==0.12.0
+trl==0.12.1
 zstandard==0.22.0
 fastcore
 

scripts/cutcrossentropy_install.py

@@ -0,0 +1,28 @@
+"""Script to output the correct installation command for cut-cross-entropy."""
+import importlib.util
+import sys
+
+try:
+    import torch
+except ImportError as exc:
+    raise ImportError("Install torch via `pip install torch`") from exc
+from packaging.version import Version as V
+
+v = V(torch.__version__)
+
+# no cut-cross-entropy support for torch < 2.4.0
+if v < V("2.4.0"):
+    print("")
+    sys.exit(0)
+
+cce_spec = importlib.util.find_spec("cut_cross_entropy")
+
+UNINSTALL_PREFIX = ""
+if cce_spec:
+    if not importlib.util.find_spec("cut_cross_entropy.transformers"):
+        UNINSTALL_PREFIX = "pip uninstall -y cut-cross-entropy && "
+
+print(
+    UNINSTALL_PREFIX
+    + 'pip install "cut-cross-entropy @ git+https://github.com/apple/ml-cross-entropy.git@9c297c905f55b73594b5d650722d1e78183b77bd"'
+)

scripts/motd

@@ -13,5 +13,5 @@ cd /workspace
 rm -rf /workspace/axolotl
 git clone https://github.com/axolotl-ai-cloud/axolotl.git
 cd axolotl
-pip install --no-deps -e .
+pip install --no-build-isolation --no-deps -e .
 ```

scripts/unsloth_install.py

@@ -0,0 +1,36 @@
+# noqa
+# pylint: skip-file
+try:
+    import torch
+except ImportError:
+    raise ImportError("Install torch via `pip install torch`")
+from packaging.version import Version as V
+
+v = V(torch.__version__)
+cuda = str(torch.version.cuda)
+try:
+    is_ampere = torch.cuda.get_device_capability()[0] >= 8
+except RuntimeError:
+    is_ampere = False
+if cuda != "12.1" and cuda != "11.8" and cuda != "12.4":
+    raise RuntimeError(f"CUDA = {cuda} not supported!")
+if v <= V("2.1.0"):
+    raise RuntimeError(f"Torch = {v} too old!")
+elif v <= V("2.1.1"):
+    x = "cu{}{}-torch211"
+elif v <= V("2.1.2"):
+    x = "cu{}{}-torch212"
+elif v < V("2.3.0"):
+    x = "cu{}{}-torch220"
+elif v < V("2.4.0"):
+    x = "cu{}{}-torch230"
+elif v < V("2.5.0"):
+    x = "cu{}{}-torch240"
+elif v < V("2.6.0"):
+    x = "cu{}{}-torch250"
+else:
+    raise RuntimeError(f"Torch = {v} too new!")
+x = x.format(cuda.replace(".", ""), "-ampere" if is_ampere else "")
+print(
+    f'pip install unsloth-zoo==2024.11.7 && pip install --no-deps "unsloth[{x}]==2024.11.9"'
+)

setup.py

@@ -1,8 +1,10 @@
 """setup.py for axolotl"""
-
+import ast
+import os
 import platform
 import re
 from importlib.metadata import PackageNotFoundError, version
+from pathlib import Path
 
 from setuptools import find_packages, setup
 
@@ -91,24 +93,39 @@ def parse_requirements():
     return _install_requires, _dependency_links
 
 
+def get_package_version():
+    with open(
+        Path(os.path.dirname(os.path.abspath(__file__)))
+        / "src"
+        / "axolotl"
+        / "__init__.py",
+        "r",
+        encoding="utf-8",
+    ) as fin:
+        version_match = re.search(r"^__version__\s*=\s*(.*)$", fin.read(), re.MULTILINE)
+    version_ = ast.literal_eval(version_match.group(1))
+    return version_
+
+
 install_requires, dependency_links = parse_requirements()
 
-
 setup(
-    name="axolotl",
-    version="0.5.1",
-    description="LLM Trainer",
-    long_description="Axolotl is a tool designed to streamline the fine-tuning of various AI models, offering support for multiple configurations and architectures.",
+    version=get_package_version(),
     package_dir={"": "src"},
     packages=find_packages("src"),
     install_requires=install_requires,
     dependency_links=dependency_links,
+    entry_points={
+        "console_scripts": [
+            "axolotl=axolotl.cli.main:main",
+        ],
+    },
     extras_require={
         "flash-attn": [
             "flash-attn==2.7.0.post2",
         ],
         "deepspeed": [
-            "deepspeed==0.15.4",
+            "deepspeed==0.16.1",
             "deepspeed-kernels",
         ],
         "mamba-ssm": [

src/axolotl/__init__.py

@@ -0,0 +1,3 @@
+"""Axolotl - Train and fine-tune large language models"""
+
+__version__ = "0.6.0"

src/axolotl/cli/__init__.py

@@ -27,14 +27,17 @@ from transformers.utils import is_torch_bf16_gpu_available
 from transformers.utils.import_utils import _is_package_available
 
 from axolotl.common.cli import TrainerCliArgs, load_model_and_tokenizer
-from axolotl.integrations.base import PluginManager
 from axolotl.logging_config import configure_logging
 from axolotl.train import TrainDatasetMeta
-from axolotl.utils.chat_templates import get_chat_template
+from axolotl.utils.chat_templates import (
+    get_chat_template,
+    get_chat_template_from_config,
+)
 from axolotl.utils.comet_ import setup_comet_env_vars
 from axolotl.utils.config import (
     normalize_cfg_datasets,
     normalize_config,
+    prepare_plugins,
     validate_config,
 )
 from axolotl.utils.data import load_prepare_dpo_datasets, prepare_dataset
@@ -97,8 +100,8 @@ def print_dep_versions():
         print("*" * 40)
         print("**** Axolotl Dependency Versions *****")
         for pkg in packages:
-            version = _is_package_available(pkg, return_version=True)
-            print(f"{pkg: >{max_len}}: {version[1]: <15}")
+            pkg_version = _is_package_available(pkg, return_version=True)
+            print(f"{pkg: >{max_len}}: {pkg_version[1]: <15}")
         print("*" * 40)
 
 
@@ -136,7 +139,7 @@ def check_remote_config(config: Union[str, Path]):
         with open(output_path, "wb") as file:
             file.write(content)
         LOG.info(
-            f"Using the following config obtained from {config}:\n\n{content.decode('utf-8')}\n"
+            f"Using the following config obtained from {config}: \n\n{content.decode('utf-8')}\n"
         )
         return output_path
 
@@ -199,6 +202,10 @@ def do_inference(
         )
     elif cfg.chat_template:
         chat_template_str = get_chat_template(cfg.chat_template)
+    elif cfg.datasets[0].type == "chat_template":
+        chat_template_str = get_chat_template_from_config(
+            cfg=cfg, ds_cfg=cfg.datasets[0], tokenizer=tokenizer
+        )
 
     model = model.to(cfg.device, dtype=cfg.torch_dtype)
 
@@ -373,7 +380,7 @@ def choose_config(path: Path):
 
     if len(yaml_files) == 1:
         print(f"Using default YAML file '{yaml_files[0]}'")
-        return yaml_files[0]
+        return str(yaml_files[0])
 
     print("Choose a YAML file:")
     for idx, file in enumerate(yaml_files):
@@ -384,7 +391,7 @@ def choose_config(path: Path):
         try:
             choice = int(input("Enter the number of your choice: "))
             if 1 <= choice <= len(yaml_files):
-                chosen_file = yaml_files[choice - 1]
+                chosen_file = str(yaml_files[choice - 1])
             else:
                 print("Invalid choice. Please choose a number from the list.")
         except ValueError:
@@ -419,17 +426,14 @@ def load_cfg(config: Union[str, Path] = Path("examples/"), **kwargs):
 
     cfg.axolotl_config_path = config
 
-    if cfg.get("plugins"):
-        plugin_manager = PluginManager.get_instance()
-        for plugin_name in cfg["plugins"]:
-            plugin_manager.register(plugin_name)
-
     try:
         device_props = torch.cuda.get_device_properties("cuda")
         gpu_version = "sm_" + str(device_props.major) + str(device_props.minor)
     except:  # pylint: disable=bare-except # noqa: E722
         gpu_version = None
 
+    prepare_plugins(cfg)
+
     cfg = validate_config(
         cfg,
         capabilities={
@@ -437,6 +441,9 @@ def load_cfg(config: Union[str, Path] = Path("examples/"), **kwargs):
             "n_gpu": int(os.environ.get("WORLD_SIZE", 1)),
             "compute_capability": gpu_version,
         },
+        env_capabilities={
+            "torch_version": str(torch.__version__).split("+", maxsplit=1)[0],
+        },
     )
 
     prepare_optim_env(cfg)

src/axolotl/cli/inference.py

@@ -2,6 +2,7 @@
 CLI to run inference on a trained model
 """
 from pathlib import Path
+from typing import Union
 
 import fire
 import transformers
@@ -16,10 +17,10 @@ from axolotl.cli import (
 from axolotl.common.cli import TrainerCliArgs
 
 
-def do_cli(config: Path = Path("examples/"), gradio=False, **kwargs):
+def do_cli(config: Union[Path, str] = Path("examples/"), gradio=False, **kwargs):
     # pylint: disable=duplicate-code
     print_axolotl_text_art()
-    parsed_cfg = load_cfg(config, **kwargs)
+    parsed_cfg = load_cfg(config, inference=True, **kwargs)
     parsed_cfg.sample_packing = False
     parser = transformers.HfArgumentParser((TrainerCliArgs))
     parsed_cli_args, _ = parser.parse_args_into_dataclasses(

src/axolotl/cli/main.py

@@ -0,0 +1,233 @@
+"""CLI definition for various axolotl commands."""
+# pylint: disable=redefined-outer-name
+import subprocess  # nosec B404
+from typing import Optional
+
+import click
+
+import axolotl
+from axolotl.cli.utils import (
+    add_options_from_config,
+    add_options_from_dataclass,
+    build_command,
+    fetch_from_github,
+)
+from axolotl.common.cli import PreprocessCliArgs, TrainerCliArgs
+from axolotl.utils.config.models.input.v0_4_1 import AxolotlInputConfig
+
+
+@click.group()
+@click.version_option(version=axolotl.__version__, prog_name="axolotl")
+def cli():
+    """Axolotl CLI - Train and fine-tune large language models"""
+
+
+@cli.command()
+@click.argument("config", type=click.Path(exists=True, path_type=str))
+@add_options_from_dataclass(PreprocessCliArgs)
+@add_options_from_config(AxolotlInputConfig)
+def preprocess(config: str, **kwargs):
+    """Preprocess datasets before training."""
+    kwargs = {k: v for k, v in kwargs.items() if v is not None}
+
+    from axolotl.cli.preprocess import do_cli
+
+    do_cli(config=config, **kwargs)
+
+
+@cli.command()
+@click.argument("config", type=click.Path(exists=True, path_type=str))
+@click.option(
+    "--accelerate/--no-accelerate",
+    default=True,
+    help="Use accelerate launch for multi-GPU training",
+)
+@add_options_from_dataclass(TrainerCliArgs)
+@add_options_from_config(AxolotlInputConfig)
+def train(config: str, accelerate: bool, **kwargs):
+    """Train or fine-tune a model."""
+    kwargs = {k: v for k, v in kwargs.items() if v is not None}
+
+    if accelerate:
+        base_cmd = ["accelerate", "launch", "-m", "axolotl.cli.train"]
+        if config:
+            base_cmd.append(config)
+        cmd = build_command(base_cmd, kwargs)
+        subprocess.run(cmd, check=True)  # nosec B603
+    else:
+        from axolotl.cli.train import do_cli
+
+        do_cli(config=config, **kwargs)
+
+
+@cli.command()
+@click.argument("config", type=click.Path(exists=True, path_type=str))
+@click.option(
+    "--accelerate/--no-accelerate",
+    default=True,
+    help="Use accelerate launch for multi-GPU inference",
+)
+@click.option(
+    "--lora-model-dir",
+    type=click.Path(exists=True, path_type=str),
+    help="Directory containing LoRA model",
+)
+@click.option(
+    "--base-model",
+    type=click.Path(exists=True, path_type=str),
+    help="Path to base model for non-LoRA models",
+)
+@click.option("--gradio", is_flag=True, help="Launch Gradio interface")
+@click.option("--load-in-8bit", is_flag=True, help="Load model in 8-bit mode")
+@add_options_from_dataclass(TrainerCliArgs)
+@add_options_from_config(AxolotlInputConfig)
+def inference(
+    config: str,
+    accelerate: bool,
+    lora_model_dir: Optional[str] = None,
+    base_model: Optional[str] = None,
+    **kwargs,
+):
+    """Run inference with a trained model."""
+    kwargs = {k: v for k, v in kwargs.items() if v is not None}
+    del kwargs["inference"]  # interferes with inference.do_cli
+
+    if lora_model_dir:
+        kwargs["lora_model_dir"] = lora_model_dir
+    if base_model:
+        kwargs["output_dir"] = base_model
+
+    if accelerate:
+        base_cmd = ["accelerate", "launch", "-m", "axolotl.cli.inference"]
+        if config:
+            base_cmd.append(config)
+        cmd = build_command(base_cmd, kwargs)
+        subprocess.run(cmd, check=True)  # nosec B603
+    else:
+        from axolotl.cli.inference import do_cli
+
+        do_cli(config=config, **kwargs)
+
+
+@cli.command()
+@click.argument("config", type=click.Path(exists=True, path_type=str))
+@click.option(
+    "--accelerate/--no-accelerate",
+    default=False,
+    help="Use accelerate launch for multi-GPU operations",
+)
+@click.option(
+    "--model-dir",
+    type=click.Path(exists=True, path_type=str),
+    help="Directory containing model weights to shard",
+)
+@click.option(
+    "--save-dir",
+    type=click.Path(path_type=str),
+    help="Directory to save sharded weights",
+)
+@add_options_from_dataclass(TrainerCliArgs)
+@add_options_from_config(AxolotlInputConfig)
+def shard(config: str, accelerate: bool, **kwargs):
+    """Shard model weights."""
+    kwargs = {k: v for k, v in kwargs.items() if v is not None}
+
+    if accelerate:
+        base_cmd = ["accelerate", "launch", "-m", "axolotl.cli.shard"]
+        if config:
+            base_cmd.append(config)
+        cmd = build_command(base_cmd, kwargs)
+        subprocess.run(cmd, check=True)  # nosec B603
+    else:
+        from axolotl.cli.shard import do_cli
+
+        do_cli(config=config, **kwargs)
+
+
+@cli.command()
+@click.argument("config", type=click.Path(exists=True, path_type=str))
+@click.option(
+    "--accelerate/--no-accelerate",
+    default=True,
+    help="Use accelerate launch for weight merging",
+)
+@click.option(
+    "--model-dir",
+    type=click.Path(exists=True, path_type=str),
+    help="Directory containing sharded weights",
+)
+@click.option(
+    "--save-path", type=click.Path(path_type=str), help="Path to save merged weights"
+)
+@add_options_from_dataclass(TrainerCliArgs)
+@add_options_from_config(AxolotlInputConfig)
+def merge_sharded_fsdp_weights(config: str, accelerate: bool, **kwargs):
+    """Merge sharded FSDP model weights."""
+    kwargs = {k: v for k, v in kwargs.items() if v is not None}
+
+    if accelerate:
+        base_cmd = [
+            "accelerate",
+            "launch",
+            "-m",
+            "axolotl.cli.merge_sharded_fsdp_weights",
+        ]
+        if config:
+            base_cmd.append(config)
+        cmd = build_command(base_cmd, kwargs)
+        subprocess.run(cmd, check=True)  # nosec B603
+    else:
+        from axolotl.cli.merge_sharded_fsdp_weights import do_cli
+
+        do_cli(config=config, **kwargs)
+
+
+@cli.command()
+@click.argument("config", type=click.Path(exists=True, path_type=str))
+@click.option(
+    "--lora-model-dir",
+    type=click.Path(exists=True, path_type=str),
+    help="Directory containing the LoRA model to merge",
+)
+@click.option(
+    "--output-dir",
+    type=click.Path(path_type=str),
+    help="Directory to save the merged model",
+)
+def merge_lora(
+    config: str,
+    lora_model_dir: Optional[str] = None,
+    output_dir: Optional[str] = None,
+):
+    """Merge a trained LoRA into a base model"""
+    kwargs = {}
+    if lora_model_dir:
+        kwargs["lora_model_dir"] = lora_model_dir
+    if output_dir:
+        kwargs["output_dir"] = output_dir
+
+    from axolotl.cli.merge_lora import do_cli
+
+    do_cli(config=config, **kwargs)
+
+
+@cli.command()
+@click.argument("directory", type=click.Choice(["examples", "deepspeed_configs"]))
+@click.option("--dest", help="Destination directory")
+def fetch(directory: str, dest: Optional[str]):
+    """
+    Fetch example configs or other resources.
+
+    Available directories:
+    - examples: Example configuration files
+    - deepspeed_configs: DeepSpeed configuration files
+    """
+    fetch_from_github(f"{directory}/", dest)
+
+
+def main():
+    cli()
+
+
+if __name__ == "__main__":
+    main()

src/axolotl/cli/merge_lora.py

@@ -2,6 +2,7 @@
 CLI to run merge a trained LoRA into a base model
 """
 from pathlib import Path
+from typing import Union
 
 import fire
 import transformers
@@ -11,7 +12,7 @@ from axolotl.cli import do_merge_lora, load_cfg, print_axolotl_text_art
 from axolotl.common.cli import TrainerCliArgs
 
 
-def do_cli(config: Path = Path("examples/"), **kwargs):
+def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
     # pylint: disable=duplicate-code
     print_axolotl_text_art()
     parser = transformers.HfArgumentParser((TrainerCliArgs))

src/axolotl/cli/merge_sharded_fsdp_weights.py

@@ -177,7 +177,7 @@ def merge_fsdp_weights(
     state.wait_for_everyone()
 
 
-def do_cli(config: Path = Path("examples/"), **kwargs):
+def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
     # pylint: disable=duplicate-code
     print_axolotl_text_art()
     parser = transformers.HfArgumentParser((TrainerCliArgs))

src/axolotl/cli/utils.py

@@ -0,0 +1,218 @@
+"""Utility methods for axoltl CLI."""
+import concurrent.futures
+import dataclasses
+import hashlib
+import json
+import logging
+from pathlib import Path
+from types import NoneType
+from typing import Any, Dict, List, Optional, Tuple, Type, Union, get_args, get_origin
+
+import click
+import requests
+from pydantic import BaseModel
+
+LOG = logging.getLogger("axolotl.cli.utils")
+
+
+def add_options_from_dataclass(config_class: Type[Any]):
+    """Create Click options from the fields of a dataclass."""
+
+    def decorator(function):
+        # Process dataclass fields in reverse order for correct option ordering
+        for field in reversed(dataclasses.fields(config_class)):
+            field_type = field.type
+
+            if get_origin(field_type) is Union and type(None) in get_args(field_type):
+                field_type = next(
+                    t for t in get_args(field_type) if not isinstance(t, NoneType)
+                )
+
+            if field_type == bool:
+                field_name = field.name.replace("_", "-")
+                option_name = f"--{field_name}/--no-{field_name}"
+                function = click.option(
+                    option_name,
+                    default=field.default,
+                    help=field.metadata.get("description"),
+                )(function)
+            else:
+                option_name = f"--{field.name.replace('_', '-')}"
+                function = click.option(
+                    option_name,
+                    type=field_type,
+                    default=field.default,
+                    help=field.metadata.get("description"),
+                )(function)
+        return function
+
+    return decorator
+
+
+def add_options_from_config(config_class: Type[BaseModel]):
+    """Create Click options from the fields of a Pydantic model."""
+
+    def decorator(function):
+        # Process model fields in reverse order for correct option ordering
+        for name, field in reversed(config_class.model_fields.items()):
+            if field.annotation == bool:
+                field_name = name.replace("_", "-")
+                option_name = f"--{field_name}/--no-{field_name}"
+                function = click.option(
+                    option_name, default=None, help=field.description
+                )(function)
+            else:
+                option_name = f"--{name.replace('_', '-')}"
+                function = click.option(
+                    option_name, default=None, help=field.description
+                )(function)
+        return function
+
+    return decorator
+
+
+def build_command(base_cmd: List[str], options: Dict[str, Any]) -> List[str]:
+    """Build command list from base command and options."""
+    cmd = base_cmd.copy()
+
+    for key, value in options.items():
+        if value is None:
+            continue
+
+        key = key.replace("_", "-")
+
+        if isinstance(value, bool):
+            if value:
+                cmd.append(f"--{key}")
+        else:
+            cmd.extend([f"--{key}", str(value)])
+
+    return cmd
+
+
+def download_file(
+    file_info: tuple, raw_base_url: str, dest_path: Path, dir_prefix: str
+) -> Tuple[str, str]:
+    """
+    Download a single file and return its processing status.
+
+    Args:
+        file_info: Tuple of (file_path, remote_sha)
+        raw_base_url: Base URL for raw GitHub content
+        dest_path: Local destination directory
+        dir_prefix: Directory prefix to filter files
+
+    Returns:
+        Tuple of (file_path, status) where status is 'new', 'updated', or 'unchanged'
+    """
+    file_path, remote_sha = file_info
+    raw_url = f"{raw_base_url}/{file_path}"
+    dest_file = dest_path / file_path.split(dir_prefix)[-1]
+
+    # Check if file exists and needs updating
+    if dest_file.exists():
+        with open(dest_file, "rb") as file:
+            content = file.read()
+            # Calculate git blob SHA
+            blob = b"blob " + str(len(content)).encode() + b"\0" + content
+            local_sha = hashlib.sha1(blob, usedforsecurity=False).hexdigest()
+
+        if local_sha == remote_sha:
+            print(f"Skipping {file_path} (unchanged)")
+            return file_path, "unchanged"
+
+        print(f"Updating {file_path}")
+        status = "new"
+    else:
+        print(f"Downloading {file_path}")
+        status = "new"
+
+    # Create directories if needed
+    dest_file.parent.mkdir(parents=True, exist_ok=True)
+
+    # Download and save file
+    try:
+        response = requests.get(raw_url, timeout=30)
+        response.raise_for_status()
+
+        with open(dest_file, "wb") as file:
+            file.write(response.content)
+
+        return file_path, status
+    except (requests.RequestException, IOError) as request_error:
+        print(f"Error downloading {file_path}: {str(request_error)}")
+        return file_path, "error"
+
+
+def fetch_from_github(
+    dir_prefix: str, dest_dir: Optional[str] = None, max_workers: int = 5
+) -> None:
+    """
+    Sync files from a specific directory in the GitHub repository.
+    Only downloads files that don't exist locally or have changed.
+
+    Args:
+        dir_prefix: Directory prefix to filter files (e.g., 'examples/', 'deepspeed_configs/')
+        dest_dir: Local destination directory
+        max_workers: Maximum number of concurrent downloads
+    """
+    api_url = "https://api.github.com/repos/axolotl-ai-cloud/axolotl/git/trees/main?recursive=1"
+    raw_base_url = "https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main"
+
+    # Get repository tree with timeout
+    response = requests.get(api_url, timeout=30)
+    response.raise_for_status()
+    tree = json.loads(response.text)
+
+    # Filter for files and get their SHA
+    files = {
+        item["path"]: item["sha"]
+        for item in tree["tree"]
+        if item["type"] == "blob" and item["path"].startswith(dir_prefix)
+    }
+
+    if not files:
+        raise click.ClickException(f"No files found in {dir_prefix}")
+
+    # Default destination directory is the last part of dir_prefix
+    default_dest = Path(dir_prefix.rstrip("/"))
+    dest_path = Path(dest_dir) if dest_dir else default_dest
+
+    # Keep track of processed files for summary
+    files_processed: Dict[str, List[str]] = {
+        "new": [],
+        "updated": [],
+        "unchanged": [],
+        "error": [],
+    }
+
+    # Process files in parallel using ThreadPoolExecutor
+    with concurrent.futures.ThreadPoolExecutor(max_workers=max_workers) as executor:
+        future_to_file = {
+            executor.submit(
+                download_file,
+                (file_path, remote_sha),
+                raw_base_url,
+                dest_path,
+                dir_prefix,
+            ): file_path
+            for file_path, remote_sha in files.items()
+        }
+
+        # Process completed tasks as they finish
+        for future in concurrent.futures.as_completed(future_to_file):
+            file_path = future_to_file[future]
+            try:
+                file_path, status = future.result()
+                files_processed[status].append(file_path)
+            except (requests.RequestException, IOError) as request_error:
+                print(f"Error processing {file_path}: {str(request_error)}")
+                files_processed["error"].append(file_path)
+
+    # Log summary
+    LOG.info("\nSync Summary:")
+    LOG.info(f"New files: {len(files_processed['new'])}")
+    LOG.info(f"Updated files: {len(files_processed['updated'])}")
+    LOG.info(f"Unchanged files: {len(files_processed['unchanged'])}")
+    if files_processed["error"]:
+        LOG.info(f"Failed files: {len(files_processed['error'])}")

src/axolotl/core/tokenizer_utils.py

@@ -3,36 +3,88 @@ helper functions for fixing the embeddings/tokenizer
 """
 
 # Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
+# GNU LESSER GENERAL PUBLIC LICENSE
+# Version 3, 29 June 2007
 #
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
+# Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
+# Everyone is permitted to copy and distribute verbatim copies
+# of this license document, but changing it is not allowed.
 
 import gc
 import itertools
+import logging
+from collections import Counter
 
+import datasets
 import numpy as np
 import torch
 
+LOG = logging.getLogger("axolotl.core.tokenizer_utils")
 
-@torch.inference_mode
-def fix_untrained_tokens(model, tokenizer, train_dataset, eps=1e-16):
+
+@torch.inference_mode()
+def fix_untrained_tokens(  # pylint: disable=too-many-return-statements
+    model, tokenizer, train_dataset, ignored_tokenizer_names=None, eps=1e-16
+):
     """
-    Many of the newer models have reserved tokens that are not trained.
+    Llama-3 for eg has untrained vectors in the base model.
+    These include <|eot_id|>, <|start_header_id|>, <|end_header_id|>
+    We reset them to the mean of the rest of the tokens
     """
+    # Code licensed under LGPL
     embedding_matrix = model.get_input_embeddings().weight
     lm_head_matrix = model.get_output_embeddings().weight
+    chat_template = getattr(tokenizer, "chat_template", None)
+    tokenizer = tokenizer.tokenizer if hasattr(tokenizer, "tokenizer") else tokenizer
+
+    # Ignore some model checks for now
+    if not ignored_tokenizer_names:
+        ignored_tokenizer_names = []
+    if (
+        model.config._name_or_path  # pylint: disable=protected-access
+        in ignored_tokenizer_names
+    ):
+        return
+
+    # Sometimes the sizes can be different like in vision models
+    # Ie <image> is in input, but not in output
+    min_size = min(embedding_matrix.shape[1], lm_head_matrix.shape[1])
+    embedding_matrix = embedding_matrix[:, :min_size]
+    lm_head_matrix = lm_head_matrix[:, :min_size]
 
     # Get untrained tokens
-    indicator_untrained = torch.amax(embedding_matrix, axis=1) <= eps
+    indicator_untrained1 = torch.amax(embedding_matrix, axis=1) <= eps
+    # Check lm_head as well
+
+    # Does NOT work for Llama 3.1!!
+    indicator_untrained2 = torch.amax(lm_head_matrix, axis=1) <= eps
+
+    # We instead check for repeated vectors
+    lm_head_where = torch.where(indicator_untrained1)[0]
+    lm_head_bad = lm_head_matrix[lm_head_where]
+    lm_head_bad = lm_head_bad.cpu().float().numpy().round(3)
+    counter = Counter()
+    for row in lm_head_bad:
+        counter[hash(row.data.tobytes())] += 1
+    counter = Counter({k: c for k, c in counter.items() if c >= 2})
+
+    lm_head_where = lm_head_where.cpu().numpy()
+    final_bad_lm_head = []
+    for j, row in enumerate(lm_head_bad):
+        if hash(row.data.tobytes()) in counter:
+            final_bad_lm_head.append(lm_head_where[j])
+    indicator_untrained2 = indicator_untrained2 | torch.zeros_like(indicator_untrained2)
+    indicator_untrained2[final_bad_lm_head] = True
+
+    # Combine both checks
+    indicator_untrained = indicator_untrained1 & indicator_untrained2
+
+    # Remove pad token possibility
+    if hasattr(tokenizer, "pad_token_id"):
+        pad_token_id = tokenizer.pad_token_id
+        if pad_token_id is not None and pad_token_id < indicator_untrained.shape[0]:
+            indicator_untrained[pad_token_id] = False
+
     where_untrained = torch.where(indicator_untrained)[0]
     n_untrained = where_untrained.shape[0]
     n_trained = embedding_matrix.shape[0] - n_untrained
@@ -40,10 +92,9 @@ def fix_untrained_tokens(model, tokenizer, train_dataset, eps=1e-16):
     # Get set and actual tokens
     where_untrained = where_untrained.tolist()
     if len(where_untrained) == 0:
-        return False
+        return
 
     # Remove untrained indices where it's longer
-
     where_untrained_set = frozenset(where_untrained)
     actual_bad_tokens = tokenizer.convert_ids_to_tokens(where_untrained)
     # Remove None items in actual_bad_tokens
@@ -53,10 +104,14 @@ def fix_untrained_tokens(model, tokenizer, train_dataset, eps=1e-16):
     if_bad_first = False
     if_bad_second = False
     # Check tokenizer's chat template for any untrained tokens
-    chat_template = getattr(tokenizer, "chat_template", None)
     if chat_template is not None:
         if_bad_first = any(x in chat_template for x in actual_bad_tokens)
 
+    if isinstance(train_dataset, datasets.IterableDataset):
+        # Skip the check, since the code below assumes
+        # an indexable dataset
+        return
+
     # Check the first 250, last 250 input_ids
     size_dataset = len(train_dataset)
     size = min(size_dataset, 250)
@@ -83,7 +138,69 @@ def fix_untrained_tokens(model, tokenizer, train_dataset, eps=1e-16):
 
     # Check if bad tokens exists!
     if not if_bad_first and not if_bad_second:
-        return False
+        return
+
+    # Check if lm_head / embed_token are trainable!
+    bad_not_trainable = False
+    if not embedding_matrix.requires_grad:
+        bad_not_trainable = True
+    if not lm_head_matrix.requires_grad:
+        bad_not_trainable = True
+
+    if bad_not_trainable:  # pylint: disable=too-many-nested-blocks
+        final_bad_items = []
+
+        # Re-check the first 250, last 250 input_ids
+        size_dataset = len(train_dataset)
+        size = min(size_dataset, 250)
+        for j in range(size):
+            input_ids = train_dataset[j]
+            if "input_ids" in input_ids:
+                input_ids = input_ids["input_ids"]
+                for item in input_ids:
+                    if item in where_untrained_set:
+                        final_bad_items.append(item)
+
+        # Re-check last 250
+        left = max(size_dataset - 250, 0)
+        for j in range(left, size_dataset):
+            input_ids = train_dataset[j]
+            if "input_ids" in input_ids:
+                input_ids = input_ids["input_ids"]
+                for item in input_ids:
+                    if item in where_untrained_set:
+                        final_bad_items.append(item)
+
+        # If no bad tokens, possibly chat template itself has issues?
+        if len(final_bad_items) == 0:
+            # Recheck 2000 and last 2000 items
+            size_dataset = len(train_dataset)
+            size = min(size_dataset, 2000)
+            for j in range(size):
+                input_ids = train_dataset[j]
+                if "input_ids" in input_ids:
+                    input_ids = input_ids["input_ids"]
+                    for item in input_ids:
+                        if item in where_untrained_set:
+                            final_bad_items.append(item)
+
+            # Re-check last 2000
+            left = max(size_dataset - 2000, 0)
+            for j in range(left, size_dataset):
+                input_ids = train_dataset[j]
+                if "input_ids" in input_ids:
+                    input_ids = input_ids["input_ids"]
+                    for item in input_ids:
+                        if item in where_untrained_set:
+                            final_bad_items.append(item)
+
+            # Most likely false signal!
+            if len(final_bad_items) == 0:
+                return
+
+        raise ValueError(
+            f"Untrained tokens of [{list(set(final_bad_items))}] found, but embed_tokens & lm_head not trainable, causing NaNs. "
+        )
 
     # Count all the possible bad tokens
     final_counts = np.zeros(
@@ -97,6 +214,23 @@ def fix_untrained_tokens(model, tokenizer, train_dataset, eps=1e-16):
 
     train_dataset.map(mapping, batched=True, desc="Counting untrained tokens")
 
+    # Get counts for untrained tokens
+    counts_untrained = final_counts[where_untrained]
+    # Identify untrained tokens seen in train_dataset
+    indices_seen_in_train = np.where(counts_untrained > 0)[0]
+    tokens_to_update = [where_untrained[i] for i in indices_seen_in_train]
+
+    if len(tokens_to_update) == 0:
+        LOG.info(
+            "No untrained tokens found in train_dataset. No embeddings were modified."
+        )
+        return
+
+    # Log the token IDs that are being rescaled
+    LOG.info(
+        f"Rescaling embeddings for tokens seen in train_dataset: {tokens_to_update}"
+    )
+
     # Get sum of all items
     sum_embedding = torch.sum(embedding_matrix, dtype=torch.float32, axis=0)
     sum_lm_head = torch.sum(lm_head_matrix, dtype=torch.float32, axis=0)
@@ -113,38 +247,26 @@ def fix_untrained_tokens(model, tokenizer, train_dataset, eps=1e-16):
     mean_embedding = sum_embedding / n_trained
     mean_lm_head = sum_lm_head / n_trained
 
-    # Scale each to be equal to 1/max_frequency. Also set some to 0 if none seen
-    scaling = final_counts[where_untrained] / max(final_counts.max(), 1)
+    # Compute scaling for tokens to update
+    scaling = counts_untrained[indices_seen_in_train] / max(final_counts.max(), 1)
     scaling = torch.tensor(scaling, device=mean_embedding.device).unsqueeze(1)
-    mean_embedding = (
-        mean_embedding.repeat(
-            (
-                n_untrained,
-                1,
-            )
-        )
-        * scaling
-    )
-    mean_lm_head = (
-        mean_lm_head.repeat(
-            (
-                n_untrained,
-                1,
-            )
-        )
-        * scaling
-    )
-    where_null = scaling.ravel() == 0
-    mean_embedding[where_null] = 0
-    mean_lm_head[where_null] = 0
 
-    # Set them to the mean
-    embedding_matrix[where_untrained] = mean_embedding.to(embedding_matrix.dtype)
-    lm_head_matrix[where_untrained] = mean_lm_head.to(lm_head_matrix.dtype)
+    # Prepare mean embeddings for tokens to update
+    mean_embedding_repeated = (
+        mean_embedding.unsqueeze(0).repeat(len(tokens_to_update), 1) * scaling
+    )
+    mean_lm_head_repeated = (
+        mean_lm_head.unsqueeze(0).repeat(len(tokens_to_update), 1) * scaling
+    )
+
+    # Update embeddings only for tokens seen in train_dataset
+    embedding_matrix[tokens_to_update] = mean_embedding_repeated.to(
+        embedding_matrix.dtype
+    )
+    lm_head_matrix[tokens_to_update] = mean_lm_head_repeated.to(lm_head_matrix.dtype)
 
     # Clean up
     for _ in range(3):
         gc.collect()
         torch.cuda.empty_cache()
-
-    return True
+    return

src/axolotl/core/trainer_builder.py

@@ -22,6 +22,7 @@ from typing import Any, Dict, List, Literal, Optional, Type, Union
 import torch
 import transformers
 from datasets import Dataset
+from packaging import version
 from peft.optimizers import create_loraplus_optimizer
 from torch import nn
 from torch.optim.lr_scheduler import OneCycleLR
@@ -107,6 +108,22 @@ def _sanitize_kwargs_for_tagging(tag_names, kwargs=None):
     return kwargs
 
 
+def _sanitize_kwargs_for_ds_tagging(dataset_tags, kwargs=None):
+    if isinstance(dataset_tags, str):
+        dataset_tags = [dataset_tags]
+
+    if (dataset_tags is not None) and (kwargs is not None):
+        if "dataset_tags" not in kwargs:
+            kwargs["dataset_tags"] = dataset_tags
+        elif "dataset_tags" in kwargs and isinstance(kwargs["dataset_tags"], list):
+            kwargs["dataset_tags"].extend(dataset_tags)
+        elif "dataset_tags" in kwargs and isinstance(kwargs["dataset_tags"], str):
+            dataset_tags.append(kwargs["dataset_tags"])
+            kwargs["dataset_tags"] = dataset_tags
+
+    return kwargs
+
+
 @dataclass
 class AxolotlTrainingMixins:
     """
@@ -220,6 +237,14 @@ class AxolotlTrainingMixins:
         default=1e-6,
         metadata={"help": "loraplus learning rate for lora embedding layers."},
     )
+    embedding_lr_scale: Optional[float] = field(
+        default=None,
+        metadata={"help": "Scale the learning rate for the embedding layers."},
+    )
+    embedding_lr: Optional[float] = field(
+        default=None,
+        metadata={"help": "absolute learning rate for the embedding layers."},
+    )
     qlora: bool = field(
         default=False,
         metadata={"help": "whether this is a qlora training"},
@@ -386,7 +411,7 @@ class SchedulerMixin(Trainer):
                     min_lr_ratio=self.args.cosine_min_lr_ratio,
                 )
             else:
-                return super().create_scheduler(num_training_steps, optimizer)
+                return super().create_scheduler(num_training_steps, optimizer=optimizer)
         else:
             if use_cosine_quadratic:
                 LOG.warning("axolotl's cosine scheduler with quadratic warmup not used (e.g., because of deepspeed).")
@@ -410,10 +435,12 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
         *_args,
         bench_data_collator=None,
         eval_data_collator=None,
+        dataset_tags=None,
         **kwargs,
     ):
         self.bench_data_collator = bench_data_collator
         self.eval_data_collator = eval_data_collator
+        self.dataset_tags = dataset_tags
         super().__init__(*_args, **kwargs)
         self.train_data_collator = self.data_collator
         self._stored_metrics = defaultdict(lambda: defaultdict(list))
@@ -435,6 +462,8 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
     def create_optimizer(self):
         if (
             self.args.loraplus_lr_ratio is None
+            and self.args.embedding_lr_scale is None
+            and self.args.embedding_lr is None
             and self.args.alternate_optimizer
             not in [
                 "optimi_adamw",
@@ -449,30 +478,59 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
         opt_model = self.model_wrapped if is_sagemaker_mp_enabled() else self.model
         if self.optimizer is None:  # pylint: disable=access-member-before-definition
             decay_parameters = self.get_decay_parameter_names(opt_model)
-            optimizer_grouped_parameters = [
-                {
-                    "params": [
-                        p
-                        for n, p in opt_model.named_parameters()
-                        if (n in decay_parameters and p.requires_grad)
-                    ],
-                    "weight_decay": self.args.weight_decay,
-                },
-                {
-                    "params": [
-                        p
-                        for n, p in opt_model.named_parameters()
-                        if (n not in decay_parameters and p.requires_grad)
-                    ],
-                    "weight_decay": 0.0,
-                },
-            ]
+            params = {
+                "to_weight_decay": {},  # LayerNorm and bias
+                "embeddings": {},  # lm_head, embed_tokens,
+                "no_weight_decay": {},
+            }
 
             optimizer_cls, optimizer_kwargs = Trainer.get_optimizer_cls_and_kwargs(
                 self.args,
                 opt_model,
             )
 
+            for name, param in opt_model.named_parameters():
+                if not param.requires_grad:
+                    continue
+                if name.endswith("modules_to_save.default.weight") or any(
+                    embed_name in name for embed_name in ["embed_tokens", "lm_head"]
+                ):
+                    params["embeddings"][name] = param
+                elif name in decay_parameters:
+                    params["to_weight_decay"][name] = param
+                else:
+                    params["no_weight_decay"][name] = param
+            optimizer_grouped_parameters = []
+            if params["to_weight_decay"]:
+                optimizer_grouped_parameters.append(
+                    {
+                        "params": list(params["to_weight_decay"].values()),
+                        "weight_decay": self.args.weight_decay,
+                        "lr": optimizer_kwargs["lr"],
+                    }
+                )
+            if params["embeddings"]:
+                lr = optimizer_kwargs["lr"]  # pylint: disable=invalid-name
+                if self.args.embedding_lr_scale:
+                    lr *= self.args.embedding_lr_scale  # pylint: disable=invalid-name
+                elif self.args.embedding_lr:
+                    lr = self.args.embedding_lr  # pylint: disable=invalid-name
+                optimizer_grouped_parameters.append(
+                    {
+                        "params": list(params["embeddings"].values()),
+                        "weight_decay": 0.0,
+                        "lr": lr,
+                    }
+                )
+            if params["no_weight_decay"]:
+                optimizer_grouped_parameters.append(
+                    {
+                        "params": list(params["no_weight_decay"].values()),
+                        "weight_decay": 0.0,
+                        "lr": optimizer_kwargs["lr"],
+                    }
+                )
+
             if self.args.loraplus_lr_ratio is not None:
                 loraplus_lr_ratio = getattr(self.args, "loraplus_lr_ratio", None)
                 loraplus_lr_embedding = getattr(
@@ -485,6 +543,13 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
                     loraplus_lr_embedding=loraplus_lr_embedding,
                     **optimizer_kwargs,
                 )
+            elif (
+                self.args.embedding_lr_scale is not None
+                or self.args.embedding_lr is not None
+            ):
+                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
+                    optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs)
+                )
             elif self.args.alternate_optimizer == "optimi_adamw":
                 from optimi import AdamW
 
@@ -516,7 +581,9 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
 
                 self.optimizer = (  # pylint: disable=attribute-defined-outside-init
                     ADOPT(
-                        optimizer_grouped_parameters, decoupled=True, **optimizer_kwargs
+                        optimizer_grouped_parameters,
+                        decouple=True,
+                        **optimizer_kwargs,
                     )
                 )
 
@@ -871,6 +938,9 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
         Overwrite the `push_to_hub` method in order to force-add the tags when pushing the
         model on the Hub. Please refer to `~transformers.Trainer.push_to_hub` for more details.
         """
+        kwargs = _sanitize_kwargs_for_ds_tagging(
+            dataset_tags=self.dataset_tags, kwargs=kwargs
+        )
         kwargs = _sanitize_kwargs_for_tagging(tag_names=self.tag_names, kwargs=kwargs)
 
         return super().push_to_hub(*args, **kwargs)
@@ -888,13 +958,15 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
 
         return res
 
-    def log(self, logs: Dict[str, float]) -> None:
+    def log(self, logs: Dict[str, float], start_time: Optional[float] = None) -> None:
         """
         Log `logs` on the various objects watching training, including stored metrics.
 
         Args:
             logs (`Dict[str, float]`):
                 The values to log.
+            start_time (`Optional[float]`):
+                The start of training.
         """
         # logs either has 'loss' or 'eval_loss'
         train_eval = "train" if "loss" in logs else "eval"
@@ -902,7 +974,13 @@ class AxolotlTrainer(SchedulerMixin, Trainer):
         for key, metrics in self._stored_metrics[train_eval].items():
             logs[key] = torch.tensor(metrics).mean().item()
         del self._stored_metrics[train_eval]
-        return super().log(logs)
+
+        if version.parse(transformers.__version__) >= version.parse("4.47.0.dev0"):
+            try:
+                return super().log(logs, start_time)
+            except TypeError:
+                return super().log(logs)  # transformers<=4.46
+        return super().log(logs)  # transformers<=4.46
 
     def store_metrics(
         self, metrics: Dict[str, float], train_eval: Literal["train", "eval"] = "train"
@@ -994,8 +1072,9 @@ class AxolotlDPOTrainer(SchedulerMixin, DPOTrainer):
 
     tag_names = ["axolotl", "dpo"]
 
-    def __init__(self, *args, **kwargs):
+    def __init__(self, *args, dataset_tags=None, **kwargs):
         super().__init__(*args, **kwargs)
+        self.dataset_tags = dataset_tags
         self.optimizer = None
 
     def create_optimizer(self):
@@ -1034,6 +1113,9 @@ class AxolotlDPOTrainer(SchedulerMixin, DPOTrainer):
         Overwrite the `push_to_hub` method in order to force-add the tags when pushing the
         model on the Hub. Please refer to `~transformers.Trainer.push_to_hub` for more details.
         """
+        kwargs = _sanitize_kwargs_for_ds_tagging(
+            dataset_tags=self.dataset_tags, kwargs=kwargs
+        )
         kwargs = _sanitize_kwargs_for_tagging(tag_names=self.tag_names, kwargs=kwargs)
 
         return super().push_to_hub(*args, **kwargs)
@@ -1082,6 +1164,22 @@ class AxolotlDPOTrainer(SchedulerMixin, DPOTrainer):
         torch.cuda.empty_cache()
         return loss
 
+    def log(self, logs: Dict[str, float], start_time: Optional[float] = None) -> None:
+        # TODO remove once trl supports the updated to the Trainer.log method
+        # logs either has 'loss' or 'eval_loss'
+        train_eval = "train" if "loss" in logs else "eval"
+        # Add averaged stored metrics to logs
+        for key, metrics in self._stored_metrics[train_eval].items():
+            logs[key] = torch.tensor(metrics).mean().item()
+        del self._stored_metrics[train_eval]
+
+        if version.parse(transformers.__version__) >= version.parse("4.47.0.dev0"):
+            return super(DPOTrainer, self).log(  # pylint: disable=bad-super-call
+                logs, start_time
+            )
+        # transformers<=4.46
+        return super(DPOTrainer, self).log(logs)  # pylint: disable=bad-super-call
+
 
 class AxolotlORPOTrainer(SchedulerMixin, ORPOTrainer):
     """
@@ -1090,6 +1188,22 @@ class AxolotlORPOTrainer(SchedulerMixin, ORPOTrainer):
 
     tag_names = ["axolotl", "orpo"]
 
+    def log(self, logs: Dict[str, float], start_time: Optional[float] = None) -> None:
+        # TODO remove once trl supports the updated to the Trainer.log method
+        # logs either has 'loss' or 'eval_loss'
+        train_eval = "train" if "loss" in logs else "eval"
+        # Add averaged stored metrics to logs
+        for key, metrics in self._stored_metrics[train_eval].items():
+            logs[key] = torch.tensor(metrics).mean().item()
+        del self._stored_metrics[train_eval]
+
+        if version.parse(transformers.__version__) >= version.parse("4.47.0.dev0"):
+            return super(ORPOTrainer, self).log(  # pylint: disable=bad-super-call
+                logs, start_time
+            )
+        # transformers<=4.46
+        return super(ORPOTrainer, self).log(logs)  # pylint: disable=bad-super-call
+
 
 class AxolotlKTOTrainer(SchedulerMixin, KTOTrainer):
     """
@@ -1098,6 +1212,49 @@ class AxolotlKTOTrainer(SchedulerMixin, KTOTrainer):
 
     tag_names = ["axolotl", "kto"]
 
+    def log(self, logs: Dict[str, float], start_time: Optional[float] = None) -> None:
+        # TODO remove once trl supports the updated to the Trainer.log method
+        # logs either has 'loss' or 'eval_loss'
+        train_eval = "train" if "loss" in logs else "eval"
+        # train metrics should have no prefix, eval should have 'eval_'
+        prefix = "eval_" if train_eval == "eval" else ""
+        # accumulate average metrics from sums and lengths
+        for split in ["chosen", "rejected"]:
+            if f"count/{split}" in self._stored_metrics[train_eval]:
+                count_sum = (
+                    torch.Tensor(self._stored_metrics[train_eval][f"count/{split}"])
+                    .sum()
+                    .item()
+                )
+                for metric in ["rewards", "logps", "logits"]:
+                    logs[f"{prefix}{metric}/{split}"] = (
+                        torch.Tensor(
+                            self._stored_metrics[train_eval][f"{metric}/{split}_sum"]
+                        )
+                        .sum()
+                        .item()
+                        / count_sum
+                    )
+                    # delete obsolete metric
+                    del self._stored_metrics[train_eval][f"{metric}/{split}_sum"]
+                del self._stored_metrics[train_eval][f"count/{split}"]
+        # calculate reward margin
+        if f"{prefix}rewards/chosen" in logs and f"{prefix}rewards/rejected" in logs:
+            logs[f"{prefix}rewards/margins"] = (
+                logs[f"{prefix}rewards/chosen"] - logs[f"{prefix}rewards/rejected"]
+            )
+        # Add averaged stored metrics to logs
+        for key, metrics in self._stored_metrics[train_eval].items():
+            logs[f"{prefix}{key}"] = torch.Tensor(metrics).mean().item()
+        del self._stored_metrics[train_eval]
+
+        if version.parse(transformers.__version__) >= version.parse("4.47.0.dev0"):
+            return super(KTOTrainer, self).log(  # pylint: disable=bad-super-call
+                logs, start_time
+            )
+        # transformers<=4.46
+        return super(KTOTrainer, self).log(logs)  # pylint: disable=bad-super-call
+
 
 class AxolotlCPOTrainer(SchedulerMixin, CPOTrainer):
     """
@@ -1106,6 +1263,22 @@ class AxolotlCPOTrainer(SchedulerMixin, CPOTrainer):
 
     tag_names = ["axolotl", "cpo"]
 
+    def log(self, logs: Dict[str, float], start_time: Optional[float] = None) -> None:
+        # TODO remove once trl supports the updated to the Trainer.log method
+        # logs either has 'loss' or 'eval_loss'
+        train_eval = "train" if "loss" in logs else "eval"
+        # Add averaged stored metrics to logs
+        for key, metrics in self._stored_metrics[train_eval].items():
+            logs[key] = torch.tensor(metrics).mean().item()
+        del self._stored_metrics[train_eval]
+
+        if version.parse(transformers.__version__) >= version.parse("4.47.0.dev0"):
+            return super(CPOTrainer, self).log(  # pylint: disable=bad-super-call
+                logs, start_time
+            )
+        # transformers<=4.46
+        return super(CPOTrainer, self).log(logs)  # pylint: disable=bad-super-call
+
 
 class AxolotlRewardTrainer(SchedulerMixin, RewardTrainer):
     """
@@ -1114,6 +1287,15 @@ class AxolotlRewardTrainer(SchedulerMixin, RewardTrainer):
 
     tag_names = ["axolotl", "reward"]
 
+    def log(self, logs: Dict[str, float], start_time: Optional[float] = None) -> None:
+        # TODO remove once trl supports the updated to the Trainer.log method
+        if version.parse(transformers.__version__) >= version.parse("4.47.0.dev0"):
+            return super(RewardTrainer, self).log(  # pylint: disable=bad-super-call
+                logs, start_time
+            )
+        # transformers<=4.46
+        return super(RewardTrainer, self).log(logs)  # pylint: disable=bad-super-call
+
 
 class TrainerBuilderBase(abc.ABC):
     """
@@ -1186,8 +1368,6 @@ class TrainerBuilderBase(abc.ABC):
                 SaveAxolotlConfigtoWandBCallback(self.cfg.axolotl_config_path)
             )
         if self.cfg.use_mlflow and is_mlflow_available():
-            from transformers.integrations.integration_utils import MLflowCallback
-
             from axolotl.utils.callbacks.mlflow_ import (
                 SaveAxolotlConfigtoMlflowCallback,
             )
@@ -1195,7 +1375,6 @@ class TrainerBuilderBase(abc.ABC):
             callbacks.extend(
                 [
                     SaveAxolotlConfigtoMlflowCallback(self.cfg.axolotl_config_path),
-                    MLflowCallback,
                 ]
             )
         if self.cfg.use_comet and is_comet_available():
@@ -1212,11 +1391,17 @@ class TrainerBuilderBase(abc.ABC):
         Callbacks added after the trainer is created, usually b/c these need access to the trainer
         """
         callbacks = []
-
-        plugin_manager = PluginManager.get_instance()
-        callbacks.extend(
-            plugin_manager.add_callbacks_post_trainer(cfg=self.cfg, trainer=trainer)
-        )
+        if self.cfg.plugins:
+            plugin_manager = PluginManager.get_instance()
+            callbacks.extend(
+                [
+                    cb
+                    for cb in plugin_manager.add_callbacks_post_trainer(
+                        self.cfg, trainer
+                    )
+                    if cb
+                ]
+            )
         return callbacks
 
     def hook_pre_create_training_args(self, training_arguments_kwargs):
@@ -1263,7 +1448,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
-        callbacks = super().get_post_trainer_create_callbacks(trainer=trainer)
+        callbacks = []
         if self.cfg.use_wandb and self.cfg.eval_table_size > 0:
             LogPredictionCallback = log_prediction_callback_factory(
                 trainer, self.tokenizer, "wandb"
@@ -1301,17 +1486,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         if self.cfg.lisa_step_interval and self.cfg.lisa_n_layers:
             callbacks.append(lisa_callback_factory(trainer))
 
-        if self.cfg.plugins:
-            plugin_manager = PluginManager.get_instance()
-            callbacks.extend(
-                [
-                    cb
-                    for cb in plugin_manager.add_callbacks_post_trainer(
-                        self.cfg, trainer
-                    )
-                    if cb
-                ]
-            )
+        callbacks.extend(super().get_post_trainer_create_callbacks(trainer=trainer))
         return callbacks
 
     def _get_trainer_cls(self):
@@ -1575,6 +1750,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs[
             "loraplus_lr_embedding"
         ] = self.cfg.loraplus_lr_embedding
+        training_arguments_kwargs["embedding_lr"] = self.cfg.embedding_lr
+        training_arguments_kwargs["embedding_lr_scale"] = self.cfg.embedding_lr_scale
+
         if self.cfg.lr_scheduler in ["one_cycle", "log_sweep"]:
             training_arguments_kwargs["lr_scheduler_type"] = "cosine"
             training_arguments_kwargs[
@@ -1759,6 +1937,10 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         else:
             trainer_kwargs["tokenizer"] = self.tokenizer
 
+        if (trainer_cls is not AxolotlRewardTrainer) and self.cfg.datasets is not None:
+            trainer_kwargs["dataset_tags"] = [
+                d["path"] for d in self.cfg.datasets if not Path(d["path"]).is_dir()
+            ]
         trainer = trainer_cls(
             model=self.model,
             train_dataset=self.train_dataset,
@@ -2032,6 +2214,10 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         else:
             dpo_trainer_kwargs["tokenizer"] = self.tokenizer
 
+        if self.cfg.datasets is not None and (trainer_cls is AxolotlDPOTrainer):
+            dpo_trainer_kwargs["dataset_tags"] = [
+                d["path"] for d in self.cfg.datasets if not Path(d["path"]).is_dir()
+            ]
         dpo_trainer = trainer_cls(
             *trainer_cls_args,
             args=training_args,

src/axolotl/core/trainers/trl.py

@@ -40,7 +40,7 @@ class TRLPPOTrainer(PPOTrainer):
                 query_tensors,
                 return_prompt=False,
                 generate_ref_response=True,
-                **generation_kwargs
+                **generation_kwargs,
             )
             batch["response"] = self.tokenizer.batch_decode(response_tensors)
             batch["ref_response"] = self.tokenizer.batch_decode(ref_response_tensors)

src/axolotl/integrations/cut_cross_entropy/ACKNOWLEDGEMENTS.md

@@ -0,0 +1,325 @@
+Acknowledgements
+
+Portions of this Cut Cross Entropy Software may utilize the following copyrighted
+material, the use of which is hereby acknowledged.
+
+
+------
+
+
+PyTorch
+
+    From PyTorch:
+
+    Copyright (c) 2016-     Facebook, Inc            (Adam Paszke)
+    Copyright (c) 2014-     Facebook, Inc            (Soumith Chintala)
+    Copyright (c) 2011-2014 Idiap Research Institute (Ronan Collobert)
+    Copyright (c) 2012-2014 Deepmind Technologies    (Koray Kavukcuoglu)
+    Copyright (c) 2011-2012 NEC Laboratories America (Koray Kavukcuoglu)
+    Copyright (c) 2011-2013 NYU                      (Clement Farabet)
+    Copyright (c) 2006-2010 NEC Laboratories America (Ronan Collobert, Leon Bottou, Iain Melvin, Jason Weston)
+    Copyright (c) 2006      Idiap Research Institute (Samy Bengio)
+    Copyright (c) 2001-2004 Idiap Research Institute (Ronan Collobert, Samy Bengio, Johnny Mariethoz)
+
+    From Caffe2:
+
+    Copyright (c) 2016-present, Facebook Inc. All rights reserved.
+
+    All contributions by Facebook:
+    Copyright (c) 2016 Facebook Inc.
+
+    All contributions by Google:
+    Copyright (c) 2015 Google Inc.
+    All rights reserved.
+
+    All contributions by Yangqing Jia:
+    Copyright (c) 2015 Yangqing Jia
+    All rights reserved.
+
+    All contributions by Kakao Brain:
+    Copyright 2019-2020 Kakao Brain
+
+    All contributions by Cruise LLC:
+    Copyright (c) 2022 Cruise LLC.
+    All rights reserved.
+
+    All contributions by Arm:
+    Copyright (c) 2021, 2023-2024 Arm Limited and/or its affiliates
+
+    All contributions from Caffe:
+    Copyright(c) 2013, 2014, 2015, the respective contributors
+    All rights reserved.
+
+    All other contributions:
+    Copyright(c) 2015, 2016 the respective contributors
+    All rights reserved.
+
+    Caffe2 uses a copyright model similar to Caffe: each contributor holds
+    copyright over their contributions to Caffe2. The project versioning records
+    all such contribution and copyright details. If a contributor wants to further
+    mark their specific copyright on a particular contribution, they should
+    indicate their copyright solely in the commit message of the change when it is
+    committed.
+
+    All rights reserved.
+
+    Redistribution and use in source and binary forms, with or without
+    modification, are permitted provided that the following conditions are met:
+
+    1. Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+
+    2. Redistributions in binary form must reproduce the above copyright
+    notice, this list of conditions and the following disclaimer in the
+    documentation and/or other materials provided with the distribution.
+
+    3. Neither the names of Facebook, Deepmind Technologies, NYU, NEC Laboratories America
+    and IDIAP Research Institute nor the names of its contributors may be
+    used to endorse or promote products derived from this software without
+    specific prior written permission.
+
+    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+    AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+    IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+    ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+    LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+    CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+    SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+    INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+    CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+    ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+    POSSIBILITY OF SUCH DAMAGE.
+
+
+Triton
+
+    /*
+    * Copyright 2018-2020 Philippe Tillet
+    * Copyright 2020-2022 OpenAI
+    *
+    * Permission is hereby granted, free of charge, to any person obtaining
+    * a copy of this software and associated documentation files
+    * (the "Software"), to deal in the Software without restriction,
+    * including without limitation the rights to use, copy, modify, merge,
+    * publish, distribute, sublicense, and/or sell copies of the Software,
+    * and to permit persons to whom the Software is furnished to do so,
+    * subject to the following conditions:
+    *
+    * The above copyright notice and this permission notice shall be
+    * included in all copies or substantial portions of the Software.
+    *
+    * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+    * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+    * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
+    * IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY
+    * CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+    * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+    * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+    */
+
+
+Transformers
+
+    Copyright 2018- The Hugging Face team. All rights reserved.
+
+                                    Apache License
+                            Version 2.0, January 2004
+                            http://www.apache.org/licenses/
+
+    TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+    1. Definitions.
+
+        "License" shall mean the terms and conditions for use, reproduction,
+        and distribution as defined by Sections 1 through 9 of this document.
+
+        "Licensor" shall mean the copyright owner or entity authorized by
+        the copyright owner that is granting the License.
+
+        "Legal Entity" shall mean the union of the acting entity and all
+        other entities that control, are controlled by, or are under common
+        control with that entity. For the purposes of this definition,
+        "control" means (i) the power, direct or indirect, to cause the
+        direction or management of such entity, whether by contract or
+        otherwise, or (ii) ownership of fifty percent (50%) or more of the
+        outstanding shares, or (iii) beneficial ownership of such entity.
+
+        "You" (or "Your") shall mean an individual or Legal Entity
+        exercising permissions granted by this License.
+
+        "Source" form shall mean the preferred form for making modifications,
+        including but not limited to software source code, documentation
+        source, and configuration files.
+
+        "Object" form shall mean any form resulting from mechanical
+        transformation or translation of a Source form, including but
+        not limited to compiled object code, generated documentation,
+        and conversions to other media types.
+
+        "Work" shall mean the work of authorship, whether in Source or
+        Object form, made available under the License, as indicated by a
+        copyright notice that is included in or attached to the work
+        (an example is provided in the Appendix below).
+
+        "Derivative Works" shall mean any work, whether in Source or Object
+        form, that is based on (or derived from) the Work and for which the
+        editorial revisions, annotations, elaborations, or other modifications
+        represent, as a whole, an original work of authorship. For the purposes
+        of this License, Derivative Works shall not include works that remain
+        separable from, or merely link (or bind by name) to the interfaces of,
+        the Work and Derivative Works thereof.
+
+        "Contribution" shall mean any work of authorship, including
+        the original version of the Work and any modifications or additions
+        to that Work or Derivative Works thereof, that is intentionally
+        submitted to Licensor for inclusion in the Work by the copyright owner
+        or by an individual or Legal Entity authorized to submit on behalf of
+        the copyright owner. For the purposes of this definition, "submitted"
+        means any form of electronic, verbal, or written communication sent
+        to the Licensor or its representatives, including but not limited to
+        communication on electronic mailing lists, source code control systems,
+        and issue tracking systems that are managed by, or on behalf of, the
+        Licensor for the purpose of discussing and improving the Work, but
+        excluding communication that is conspicuously marked or otherwise
+        designated in writing by the copyright owner as "Not a Contribution."
+
+        "Contributor" shall mean Licensor and any individual or Legal Entity
+        on behalf of whom a Contribution has been received by Licensor and
+        subsequently incorporated within the Work.
+
+    2. Grant of Copyright License. Subject to the terms and conditions of
+        this License, each Contributor hereby grants to You a perpetual,
+        worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+        copyright license to reproduce, prepare Derivative Works of,
+        publicly display, publicly perform, sublicense, and distribute the
+        Work and such Derivative Works in Source or Object form.
+
+    3. Grant of Patent License. Subject to the terms and conditions of
+        this License, each Contributor hereby grants to You a perpetual,
+        worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+        (except as stated in this section) patent license to make, have made,
+        use, offer to sell, sell, import, and otherwise transfer the Work,
+        where such license applies only to those patent claims licensable
+        by such Contributor that are necessarily infringed by their
+        Contribution(s) alone or by combination of their Contribution(s)
+        with the Work to which such Contribution(s) was submitted. If You
+        institute patent litigation against any entity (including a
+        cross-claim or counterclaim in a lawsuit) alleging that the Work
+        or a Contribution incorporated within the Work constitutes direct
+        or contributory patent infringement, then any patent licenses
+        granted to You under this License for that Work shall terminate
+        as of the date such litigation is filed.
+
+    4. Redistribution. You may reproduce and distribute copies of the
+        Work or Derivative Works thereof in any medium, with or without
+        modifications, and in Source or Object form, provided that You
+        meet the following conditions:
+
+        (a) You must give any other recipients of the Work or
+            Derivative Works a copy of this License; and
+
+        (b) You must cause any modified files to carry prominent notices
+            stating that You changed the files; and
+
+        (c) You must retain, in the Source form of any Derivative Works
+            that You distribute, all copyright, patent, trademark, and
+            attribution notices from the Source form of the Work,
+            excluding those notices that do not pertain to any part of
+            the Derivative Works; and
+
+        (d) If the Work includes a "NOTICE" text file as part of its
+            distribution, then any Derivative Works that You distribute must
+            include a readable copy of the attribution notices contained
+            within such NOTICE file, excluding those notices that do not
+            pertain to any part of the Derivative Works, in at least one
+            of the following places: within a NOTICE text file distributed
+            as part of the Derivative Works; within the Source form or
+            documentation, if provided along with the Derivative Works; or,
+            within a display generated by the Derivative Works, if and
+            wherever such third-party notices normally appear. The contents
+            of the NOTICE file are for informational purposes only and
+            do not modify the License. You may add Your own attribution
+            notices within Derivative Works that You distribute, alongside
+            or as an addendum to the NOTICE text from the Work, provided
+            that such additional attribution notices cannot be construed
+            as modifying the License.
+
+        You may add Your own copyright statement to Your modifications and
+        may provide additional or different license terms and conditions
+        for use, reproduction, or distribution of Your modifications, or
+        for any such Derivative Works as a whole, provided Your use,
+        reproduction, and distribution of the Work otherwise complies with
+        the conditions stated in this License.
+
+    5. Submission of Contributions. Unless You explicitly state otherwise,
+        any Contribution intentionally submitted for inclusion in the Work
+        by You to the Licensor shall be under the terms and conditions of
+        this License, without any additional terms or conditions.
+        Notwithstanding the above, nothing herein shall supersede or modify
+        the terms of any separate license agreement you may have executed
+        with Licensor regarding such Contributions.
+
+    6. Trademarks. This License does not grant permission to use the trade
+        names, trademarks, service marks, or product names of the Licensor,
+        except as required for reasonable and customary use in describing the
+        origin of the Work and reproducing the content of the NOTICE file.
+
+    7. Disclaimer of Warranty. Unless required by applicable law or
+        agreed to in writing, Licensor provides the Work (and each
+        Contributor provides its Contributions) on an "AS IS" BASIS,
+        WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+        implied, including, without limitation, any warranties or conditions
+        of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+        PARTICULAR PURPOSE. You are solely responsible for determining the
+        appropriateness of using or redistributing the Work and assume any
+        risks associated with Your exercise of permissions under this License.
+
+    8. Limitation of Liability. In no event and under no legal theory,
+        whether in tort (including negligence), contract, or otherwise,
+        unless required by applicable law (such as deliberate and grossly
+        negligent acts) or agreed to in writing, shall any Contributor be
+        liable to You for damages, including any direct, indirect, special,
+        incidental, or consequential damages of any character arising as a
+        result of this License or out of the use or inability to use the
+        Work (including but not limited to damages for loss of goodwill,
+        work stoppage, computer failure or malfunction, or any and all
+        other commercial damages or losses), even if such Contributor
+        has been advised of the possibility of such damages.
+
+    9. Accepting Warranty or Additional Liability. While redistributing
+        the Work or Derivative Works thereof, You may choose to offer,
+        and charge a fee for, acceptance of support, warranty, indemnity,
+        or other liability obligations and/or rights consistent with this
+        License. However, in accepting such obligations, You may act only
+        on Your own behalf and on Your sole responsibility, not on behalf
+        of any other Contributor, and only if You agree to indemnify,
+        defend, and hold each Contributor harmless for any liability
+        incurred by, or claims asserted against, such Contributor by reason
+        of your accepting any such warranty or additional liability.
+
+    END OF TERMS AND CONDITIONS
+
+    APPENDIX: How to apply the Apache License to your work.
+
+        To apply the Apache License to your work, attach the following
+        boilerplate notice, with the fields enclosed by brackets "[]"
+        replaced with your own identifying information. (Don't include
+        the brackets!)  The text should be enclosed in the appropriate
+        comment syntax for the file format. We also recommend that a
+        file or class name and description of purpose be included on the
+        same "printed page" as the copyright notice for easier
+        identification within third-party archives.
+
+    Copyright [yyyy] [name of copyright owner]
+
+    Licensed under the Apache License, Version 2.0 (the "License");
+    you may not use this file except in compliance with the License.
+    You may obtain a copy of the License at
+
+        http://www.apache.org/licenses/LICENSE-2.0
+
+    Unless required by applicable law or agreed to in writing, software
+    distributed under the License is distributed on an "AS IS" BASIS,
+    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+    See the License for the specific language governing permissions and
+    limitations under the License.

src/axolotl/integrations/cut_cross_entropy/LICENSE

@@ -0,0 +1,47 @@
+Copyright (C) 2024 Apple Inc. All Rights Reserved.
+
+IMPORTANT:  This Apple software is supplied to you by Apple
+Inc. ("Apple") in consideration of your agreement to the following
+terms, and your use, installation, modification or redistribution of
+this Apple software constitutes acceptance of these terms.  If you do
+not agree with these terms, please do not use, install, modify or
+redistribute this Apple software.
+
+In consideration of your agreement to abide by the following terms, and
+subject to these terms, Apple grants you a personal, non-exclusive
+license, under Apple's copyrights in this original Apple software (the
+"Apple Software"), to use, reproduce, modify and redistribute the Apple
+Software, with or without modifications, in source and/or binary forms;
+provided that if you redistribute the Apple Software in its entirety and
+without modifications, you must retain this notice and the following
+text and disclaimers in all such redistributions of the Apple Software.
+Neither the name, trademarks, service marks or logos of Apple Inc. may
+be used to endorse or promote products derived from the Apple Software
+without specific prior written permission from Apple.  Except as
+expressly stated in this notice, no other rights or licenses, express or
+implied, are granted by Apple herein, including but not limited to any
+patent rights that may be infringed by your derivative works or by other
+works in which the Apple Software may be incorporated.
+
+The Apple Software is provided by Apple on an "AS IS" basis.  APPLE
+MAKES NO WARRANTIES, EXPRESS OR IMPLIED, INCLUDING WITHOUT LIMITATION
+THE IMPLIED WARRANTIES OF NON-INFRINGEMENT, MERCHANTABILITY AND FITNESS
+FOR A PARTICULAR PURPOSE, REGARDING THE APPLE SOFTWARE OR ITS USE AND
+OPERATION ALONE OR IN COMBINATION WITH YOUR PRODUCTS.
+
+IN NO EVENT SHALL APPLE BE LIABLE FOR ANY SPECIAL, INDIRECT, INCIDENTAL
+OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) ARISING IN ANY WAY OUT OF THE USE, REPRODUCTION,
+MODIFICATION AND/OR DISTRIBUTION OF THE APPLE SOFTWARE, HOWEVER CAUSED
+AND WHETHER UNDER THEORY OF CONTRACT, TORT (INCLUDING NEGLIGENCE),
+STRICT LIABILITY OR OTHERWISE, EVEN IF APPLE HAS BEEN ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+
+
+-------------------------------------------------------------------------------
+SOFTWARE DISTRIBUTED WITH CUT CROSS ENTROPY:
+
+The Cut Cross Entropy software includes a number of subcomponents with separate
+copyright notices and license terms - please see the file ACKNOWLEDGEMENTS.md.
+-------------------------------------------------------------------------------

src/axolotl/integrations/cut_cross_entropy/README.md

@@ -0,0 +1,10 @@
+# Cut Cross Entropy
+
+### Usage
+
+```yaml
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+cut_cross_entropy: true
+```

src/axolotl/integrations/cut_cross_entropy/__init__.py

@@ -0,0 +1,83 @@
+# Copyright 2024 Axolotl AI. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Module for the Plugin for Cut Cross Entropy integration with Axolotl.
+
+Cut Cross Entropy is an optimized implementation of cross entropy loss
+from Apple's ML team.
+"""
+import importlib
+import logging
+
+import torch
+
+from axolotl.integrations.base import BasePlugin
+from axolotl.utils import get_pytorch_version
+
+from ...utils.distributed import zero_only
+from .args import CutCrossEntropyArgs  # pylint: disable=unused-import. # noqa: F401
+
+LOG = logging.getLogger("axolotl.integrations.cut_cross_entropy")
+
+_CCE_INSTALL_MESSAGE = (
+    "Please install cut_cross_entropy with transformers support using "
+    '`pip install "cut-cross-entropy[transformers]==24.11.4"`'
+)
+
+
+class CutCrossEntropyPlugin(BasePlugin):
+    """
+    Plugin for Cut Cross Entropy integration with Axolotl.
+    """
+
+    def get_input_args(self):
+        return "axolotl.integrations.cut_cross_entropy.CutCrossEntropyArgs"
+
+    def _check_requirements(self):
+        """Check if all requirements are met."""
+        # Check PyTorch version
+
+        major, minor, _ = get_pytorch_version()
+        if (major, minor) < (2, 4):
+            raise ImportError(
+                "Cut Cross Entropy requires PyTorch >= 2.4.0. "
+                f"Current version: {torch.__version__}"
+            )
+
+        # Check if cut_cross_entropy is installed
+        cce_spec = importlib.util.find_spec("cut_cross_entropy")
+        if cce_spec is None:
+            raise ImportError(_CCE_INSTALL_MESSAGE)
+
+        cce_spec_transformers = importlib.util.find_spec(
+            "cut_cross_entropy.transformers"
+        )
+        if cce_spec_transformers is None:
+            raise ImportError(_CCE_INSTALL_MESSAGE)
+
+    def pre_model_load(self, cfg):
+        """Apply cut cross entropy before model loading if enabled."""
+        if cfg.cut_cross_entropy:
+            self._check_requirements()
+
+            from cut_cross_entropy.transformers import cce_patch
+
+            with zero_only():
+                LOG.info(
+                    f"Applying Cut Cross Entropy to model type: {cfg.model_config_type}"
+                )
+
+            # The patch checks model_type internally
+            cce_patch(cfg.model_config_type)

src/axolotl/integrations/cut_cross_entropy/args.py

@@ -0,0 +1,42 @@
+# Copyright 2024 Axolotl AI. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""
+Module for handling Cut Cross Entropy input arguments.
+"""
+import logging
+from typing import Optional
+
+from pydantic import BaseModel, model_validator
+
+LOG = logging.getLogger("axolotl.integrations.cut_cross_entropy.args")
+
+
+class CutCrossEntropyArgs(BaseModel):
+    """
+    Input args for Cut Cross Entropy.
+    """
+
+    cut_cross_entropy: Optional[bool] = None
+
+    @model_validator(mode="before")
+    @classmethod
+    def check_dtype_is_half(cls, data):
+        if data.get("cut_cross_entropy") and not (data.get("bf16") or data.get("fp16")):
+            raise ValueError(
+                "Cut Cross Entropy requires fp16/bf16 training for backward pass. "
+                "Please set `bf16` or `fp16` to `True`."
+            )
+
+        return data

src/axolotl/monkeypatch/llama_attn_hijack_flash.py

@@ -4,7 +4,6 @@
 
 import logging
 import warnings
-from functools import partial
 from typing import List, Optional, Tuple, Union
 
 import torch
@@ -94,13 +93,32 @@ def replace_llama_qkv_with_fused(model):
             set_module_name(model, name, qkv)
 
 
-def patch_llama_cross_entropy():
-    from flash_attn.losses.cross_entropy import CrossEntropyLoss
-
-    LOG.info("patching with flash_attn.losses.cross_entropy")
-    transformers.models.llama.modeling_llama.CrossEntropyLoss = partial(
-        CrossEntropyLoss, inplace_backward=True
+def patch_fa_llama_cross_entropy():
+    LOG.info(
+        "patching transformers.loss.loss_utils.fixed_cross_entropy with flash_attn.ops.triton.cross_entropy"
     )
+    from flash_attn.ops.triton.cross_entropy import (
+        cross_entropy_loss as flash_attn_cross_entropy_loss,
+    )
+
+    def fa2_fixed_cross_entropy(
+        source,
+        target,
+        num_items_in_batch: int = None,
+        ignore_index: int = -100,
+        **kwargs,
+    ):  # pylint: disable=unused-argument
+        reduction = "sum" if num_items_in_batch is not None else "mean"
+        loss, _ = flash_attn_cross_entropy_loss(
+            source, target, ignore_index=ignore_index
+        )
+        if reduction == "sum":
+            loss = loss.sum() / num_items_in_batch
+        else:
+            loss = loss.sum() / (target != ignore_index).sum()
+        return loss
+
+    transformers.loss.loss_utils.fixed_cross_entropy = fa2_fixed_cross_entropy
 
 
 def patch_llama_rms_norm():
@@ -147,7 +165,7 @@ def replace_llama_attn_with_flash_attn(
 
     # skip only if explicitly disabled
     if cross_entropy:
-        patch_llama_cross_entropy()
+        patch_fa_llama_cross_entropy()
 
     # skip only if explicitly disabled
     if rms_norm:

src/axolotl/monkeypatch/relora.py

@@ -46,9 +46,10 @@ def reset_optimizer(
     *,
     reset_params: List[str],  # where str is the key to a torch.nn.Parameter
     optimizer_state_keys: List[str],
-    prune_ratio: float = 0.9,
+    optimizer_magnitude_pruning: float = 0.9,
 ):
-    pruning_fn = partial(magnitude_pruning_, prune_ratio=prune_ratio)
+    # pylint:disable=unused-argument
+    pruning_fn = partial(magnitude_pruning_, prune_ratio=optimizer_magnitude_pruning)
     n_zeros = 0
     n_total = 0
 
@@ -56,16 +57,22 @@ def reset_optimizer(
     if isinstance(optimizer, ZeroRedundancyOptimizer):
         optimizer_state = optimizer.optim.state
 
-    for param in reset_params:
-        param_state = optimizer_state[param]
-        if len(param_state) == 0:  # no state for this param, happens for ZeRo optimizer
-            continue
-        for key in optimizer_state_keys:
-            pruning_fn(
-                param_state[key]
-            )  # pruning fn has to be inplace to keep the same keys in the dict
-            n_total += param_state[key].numel()
-            n_zeros += torch.sum(param_state[key] == 0).item()
+    for group in optimizer.param_groups:
+        for param in group["params"]:
+            state = optimizer_state[param]
+            for key, value in state.items():
+                if key not in optimizer_state_keys:
+                    continue
+                if torch.is_tensor(value):
+                    try:
+                        pruning_fn(value)
+                        n_total += value.numel()
+                        n_zeros += torch.sum(value == 0).item()
+                    except RuntimeError as exc:
+                        if "quantile() input tensor is too large" in str(exc):
+                            pass
+                        else:
+                            raise exc
 
     _zeroed = n_zeros / (1e-7 + n_total) * 100
     LOG.info(f"Percent of optimizer states zeroed: {_zeroed:.2f}")
@@ -129,6 +136,9 @@ class ReLoRACallback(TrainerCallback):
 
             if "adam" in args.optim.lower():
                 optimizer_state_keys = ["exp_avg", "exp_avg_sq"]
+                if "8bit" in args.optim.lower():
+                    optimizer_state_keys.append("state1")
+                    optimizer_state_keys.append("state2")
             else:
                 raise ValueError(f"Optimizer {args.optim} not supported with ReLoRA")
 
@@ -160,7 +170,7 @@ class ReLoRACallback(TrainerCallback):
                     optimizer,
                     reset_params=lora_params,
                     optimizer_state_keys=optimizer_state_keys,
-                    prune_ratio=args.relora_prune_ratio,
+                    optimizer_magnitude_pruning=args.relora_prune_ratio,
                 )
 
             if self.quantized:

src/axolotl/monkeypatch/trainer_fsdp_optim.py

@@ -0,0 +1,80 @@
+"""
+fix for FSDP optimizer save in trainer w 4.47.0
+"""
+import inspect
+import logging
+
+from transformers import Trainer
+
+from axolotl.monkeypatch.unsloth_ import detab_code
+
+LOG = logging.getLogger("axolotl.monkeypatch.trainer_fsdp_save")
+
+ORIGINAL_TRAINER_CODE = """
+
+    delay_optimizer_creation = is_sagemaker_mp_enabled() or self.is_fsdp_xla_enabled
+
+"""
+
+PATCHED_TRAINER_CODE = """
+
+    delay_optimizer_creation = is_sagemaker_mp_enabled() or self.is_fsdp_xla_enabled or self.is_fsdp_enabled
+
+"""
+
+
+def get_training_loop_code() -> str:
+    training_loop = inspect.getsource(
+        Trainer._inner_training_loop  # pylint: disable=protected-access
+    )
+    return training_loop
+
+
+def check_training_loop_is_patchable() -> bool:
+    training_loop = get_training_loop_code()
+    training_loop, _ = detab_code(training_loop)
+    return ORIGINAL_TRAINER_CODE in training_loop
+
+
+def patch_training_loop_for_fsdp():
+    """
+    monkeypatch for fixing the training loop for fsdp with optimizer save
+    """
+
+    try:
+        training_loop = get_training_loop_code()
+    except OSError:
+        return
+    Trainer._original_inner_training_loop = (  # pylint: disable=protected-access
+        training_loop
+    )
+    training_loop, _ = detab_code(training_loop)
+    if ORIGINAL_TRAINER_CODE not in training_loop:
+        return
+
+    training_loop = training_loop.replace(ORIGINAL_TRAINER_CODE, PATCHED_TRAINER_CODE)
+    training_loop = training_loop.replace(
+        "def _inner_training_loop(",
+        "def _fixed_inner_training_loop(",
+        1,
+    )
+
+    # load imports necessary
+    import transformers.trainer
+
+    items_to_import = []
+    for item in dir(transformers.trainer):
+        if item in training_loop:
+            items_to_import.append(item)
+
+    exec(  # pylint: disable=exec-used  # nosec B102
+        "from transformers.trainer import ("
+        + ", ".join(x for x in items_to_import)
+        + ")",
+        globals(),
+    )
+    exec(training_loop, globals())  # pylint: disable=exec-used  # nosec B102
+    LOG.info("patching _inner_training_loop for fsdp optimizer save")
+    Trainer._inner_training_loop = (  # pylint: disable=protected-access
+        _fixed_inner_training_loop  # pylint: disable=undefined-variable  # noqa: F821
+    )

src/axolotl/monkeypatch/trainer_grad_accum.py

@@ -0,0 +1,290 @@
+"""
+fix for FSDP gradient accumulation
+see https://github.com/huggingface/transformers/pull/35128
+"""
+import inspect
+import logging
+
+from transformers import LlamaForCausalLM, Trainer
+
+from axolotl.monkeypatch.unsloth_ import detab_code
+
+LOG = logging.getLogger("axolotl.monkeypatch.trainer_grad_accum")
+
+ORIGINAL_CONTEXT_CODE = """
+    with self.compute_loss_context_manager():
+        if self.model_accepts_loss_kwargs:
+            loss = self.compute_loss(model, inputs)
+        else:
+            loss = self.compute_loss(model, inputs, num_items_in_batch=num_items_in_batch)
+"""
+
+PATCHED_CONTEXT_CODE = """
+    with self.compute_loss_context_manager():
+        if self.model_accepts_loss_kwargs:
+            loss = self.compute_loss(model, inputs, num_items_in_batch=num_items_in_batch)
+        else:
+            loss = self.compute_loss(model, inputs)
+"""
+
+ORIGINAL_LLAMA_FCLM_CODE = """
+    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+    output_hidden_states = (
+        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+    )
+    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        **kwargs,
+    )
+
+    hidden_states = outputs[0]
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
+
+    loss = None
+    if labels is not None:
+        loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs)
+"""
+
+PATCHED_LLAMA_FCLM_CODE = """
+    output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+    output_hidden_states = (
+        output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+    )
+    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+    # remove num_items_in_batch otherwise self.model attempts to pass it to flash_attention
+    num_items_in_batch = kwargs.pop("num_items_in_batch", None)
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        **kwargs,
+    )
+    hidden_states = outputs[0]
+    # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+    logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
+
+    loss = None
+    if labels is not None:
+        loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, num_items_in_batch=num_items_in_batch, **kwargs)
+"""
+
+
+def get_training_step_code() -> str:
+    training_step = inspect.getsource(
+        Trainer.training_step  # pylint: disable=protected-access
+    )
+    return training_step
+
+
+def check_training_step_is_patchable() -> bool:
+    training_step = get_training_step_code()
+    training_step, _ = detab_code(training_step)
+    return ORIGINAL_CONTEXT_CODE in training_step
+
+
+def patch_training_step_for_ga():
+    """
+    monkeypatch for fixing the training loop for gradient accumulation
+    """
+
+    try:
+        training_step = get_training_step_code()
+    except OSError:
+        return
+    Trainer._original_training_step = training_step  # pylint: disable=protected-access
+    training_step, _ = detab_code(training_step)
+    if ORIGINAL_CONTEXT_CODE not in training_step:
+        return
+    # assert (
+    #     ORIGINAL_CONTEXT_CODE in training_step
+    # ), "Original training_step code not found"
+
+    training_step = training_step.replace(ORIGINAL_CONTEXT_CODE, PATCHED_CONTEXT_CODE)
+    training_step = training_step.replace(
+        "def training_step(",
+        "def _fixed_training_step(",
+        1,
+    )
+
+    # load imports necessary
+    import transformers.trainer
+
+    items_to_import = []
+    for item in dir(transformers.trainer):
+        if item in training_step:
+            items_to_import.append(item)
+
+    exec(  # pylint: disable=exec-used  # nosec B102
+        "from transformers.trainer import ("
+        + ", ".join(x for x in items_to_import)
+        + ")",
+        globals(),
+    )
+    exec(training_step, globals())  # pylint: disable=exec-used  # nosec B102
+    LOG.info("patching training_step")
+    Trainer.training_step = (  # pylint: disable=protected-access
+        _fixed_training_step  # pylint: disable=undefined-variable  # noqa: F821
+    )
+
+
+def get_model_forward_code() -> str:
+    forward = inspect.getsource(
+        LlamaForCausalLM.forward  # pylint: disable=protected-access
+    )
+    return forward
+
+
+def check_forward_is_patchable() -> bool:
+    forward = get_model_forward_code()
+    forward, _ = detab_code(forward)
+    return ORIGINAL_LLAMA_FCLM_CODE in forward
+
+
+def patch_forward_for_ga():
+    """
+    monkeypatch for fixing the training loop for gradient accumulation
+    """
+
+    try:
+        forward = get_model_forward_code()
+    except OSError:
+        return
+    LlamaForCausalLM._original_forward = forward  # pylint: disable=protected-access
+    forward, _ = detab_code(forward)
+    if ORIGINAL_LLAMA_FCLM_CODE not in forward:
+        return
+    # assert ORIGINAL_LLAMA_FCLM_CODE in forward, "Original forward code not found"
+
+    forward = forward.replace(ORIGINAL_LLAMA_FCLM_CODE, PATCHED_LLAMA_FCLM_CODE)
+    forward = forward.replace(
+        "def forward(",
+        "def _fixed_forward(",
+        1,
+    )
+
+    # load imports necessary
+    import transformers.models.llama.modeling_llama
+
+    items_to_import = []
+    for item in dir(transformers.models.llama.modeling_llama):
+        if item in forward:
+            items_to_import.append(item)
+
+    exec(  # pylint: disable=exec-used  # nosec B102
+        "from transformers.models.llama.modeling_llama import ("
+        + ", ".join(x for x in items_to_import)
+        + ")",
+        globals(),
+    )
+    exec(forward, globals())  # pylint: disable=exec-used  # nosec B102
+    LOG.info("patching forward")
+    LlamaForCausalLM.forward = (  # pylint: disable=protected-access
+        _fixed_forward  # pylint: disable=undefined-variable  # noqa: F821
+    )
+
+
+ORIGINAL_TRAINER_CODE = """
+                context = (
+                    functools.partial(self.accelerator.no_sync, model=model)
+                    if i != len(batch_samples) - 1
+                    else contextlib.nullcontext
+                )
+                with context():
+                    tr_loss_step = self.training_step(model, inputs, num_items_in_batch)
+"""
+
+PATCHED_TRAINER_CODE = """
+                disable_deepspeed_no_sync = (
+                        self.accelerator.distributed_type == DistributedType.DEEPSPEED
+                        # and self.accelerator.deepspeed_engine_wrapped.engine.zero_optimization_partition_gradients()
+                )
+                context = (
+                    functools.partial(self.accelerator.no_sync, model=model)
+                    if i != len(batch_samples) - 1 and not disable_deepspeed_no_sync
+                    else contextlib.nullcontext
+                )
+                with context():
+                    tr_loss_step = self.training_step(model, inputs, num_items_in_batch)
+"""
+
+
+def get_training_loop_code() -> str:
+    training_loop = inspect.getsource(
+        Trainer._inner_training_loop  # pylint: disable=protected-access
+    )
+    return training_loop
+
+
+def check_training_loop_is_patchable() -> bool:
+    training_loop = get_training_loop_code()
+    training_loop, _ = detab_code(training_loop)
+    return ORIGINAL_TRAINER_CODE in training_loop
+
+
+def patch_training_loop_for_deepspeed_0_16_x():
+    """
+    monkeypatch for fixing the training loop for deepspeed GA
+
+    see https://github.com/huggingface/transformers/pull/35157
+    """
+
+    try:
+        training_loop = get_training_loop_code()
+    except OSError:
+        return
+    Trainer._original_inner_training_loop = (  # pylint: disable=protected-access
+        training_loop
+    )
+    training_loop, _ = detab_code(training_loop)
+    if ORIGINAL_TRAINER_CODE not in training_loop:
+        return
+
+    training_loop = training_loop.replace(ORIGINAL_TRAINER_CODE, PATCHED_TRAINER_CODE)
+    training_loop = training_loop.replace(
+        "def _inner_training_loop(",
+        "def _fixed_inner_training_loop(",
+        1,
+    )
+
+    # load imports necessary
+    import transformers.trainer
+
+    items_to_import = []
+    for item in dir(transformers.trainer):
+        if item in training_loop:
+            items_to_import.append(item)
+
+    exec(  # pylint: disable=exec-used  # nosec B102
+        "from transformers.trainer import ("
+        + ", ".join(x for x in items_to_import)
+        + ")",
+        globals(),
+    )
+    exec(training_loop, globals())  # pylint: disable=exec-used  # nosec B102
+    LOG.info("patching _inner_training_loop for fsdp optimizer save")
+    Trainer._inner_training_loop = (  # pylint: disable=protected-access
+        _fixed_inner_training_loop  # pylint: disable=undefined-variable  # noqa: F821
+    )

src/axolotl/monkeypatch/unsloth_.py

@@ -9,10 +9,7 @@ import torch
 from accelerate.logging import get_logger
 from peft import PeftModelForCausalLM
 from torch import nn
-from transformers.models.llama.modeling_llama import (
-    LlamaFlashAttention2,
-    LlamaForCausalLM,
-)
+from transformers.models.llama.modeling_llama import LlamaFlashAttention2
 
 LOG = get_logger("axolotl.monkeypatch.unsloth")
 
@@ -55,11 +52,6 @@ def original_apply_o(self, hidden_states):
     return attn_output
 
 
-def get_forward_code() -> str:
-    forward = inspect.getsource(LlamaForCausalLM.forward)
-    return forward
-
-
 def get_self_attn_code() -> str:
     forward = inspect.getsource(LlamaFlashAttention2.forward)
     return forward
@@ -102,12 +94,22 @@ def integrate_cross_entropy_loss_patch(model_type: str = "llama") -> None:
 
 
 def detab_code(code: str) -> Tuple[str, str]:
-    spaces = re.match(r"([\s\t]{1,})", code).group(0)
-    code = re.sub(r"^" + spaces, "", code, flags=re.MULTILINE)
+    try:
+        spaces = re.match(r"([\s\t]{1,})", code).group(0)
+        code = re.sub(r"^" + spaces, "", code, flags=re.MULTILINE)
+    except AttributeError:
+        return code, ""
     return code, spaces
 
 
+self_attn_lora_patched = False  # pylint: disable=invalid-name
+
+
 def patch_self_attn_lora():
+    global self_attn_lora_patched  # pylint: disable=global-statement
+    if self_attn_lora_patched:
+        # prevent patching multiple times
+        return
     self_attn_forward = get_self_attn_code()
     LlamaFlashAttention2._original_forward = (  # pylint: disable=protected-access
         self_attn_forward
@@ -139,6 +141,7 @@ def patch_self_attn_lora():
         globals(),
     )
     exec(self_attn_forward, globals())  # pylint: disable=exec-used  # nosec B102
+    self_attn_lora_patched = True
     LOG.info("patching unsloth attn lora", main_process_only=True)
     LlamaFlashAttention2.forward = (
         unsloth_attn_forward  # pylint: disable=undefined-variable  # noqa: F821
@@ -188,7 +191,7 @@ def integrate_lora_mlp_patch(peft_model: PeftModelForCausalLM):
             for module in layer_modules
         )
         mlp_not_dora = all(
-            getattr(module, "lora_magnitude_vector", None) is None
+            len(getattr(module, "lora_magnitude_vector", []) or []) == 0
             for module in layer_modules
         )
 
@@ -213,7 +216,7 @@ def integrate_lora_patch(peft_model: PeftModelForCausalLM, cfg):
                 for module in layer_modules
             )
             qkv_not_dora = all(
-                getattr(module, "lora_magnitude_vector", None) is None
+                len(getattr(module, "lora_magnitude_vector", []) or []) == 0
                 for module in layer_modules
             )
 
@@ -232,7 +235,7 @@ def integrate_lora_patch(peft_model: PeftModelForCausalLM, cfg):
                 for module in layer_modules
             )
             o_not_dora = all(
-                getattr(module, "lora_magnitude_vector", None) is None
+                len(getattr(module, "lora_magnitude_vector", []) or []) == 0
                 for module in layer_modules
             )
 

src/axolotl/prompt_strategies/bradley_terry/chat_template.py

@@ -28,6 +28,8 @@ class BTChatTemplateStrategy(ChatTemplateStrategy):
         :return:
         """
 
+        max_length = self.prompter.max_length
+
         self.messages = "chosen_messages"
         # pylint: disable=duplicate-code
         prompt[self.messages] = []
@@ -39,6 +41,16 @@ class BTChatTemplateStrategy(ChatTemplateStrategy):
         prompt[self.messages].append({"role": "assistant", "content": prompt["chosen"]})
         chosen_tokenized = super().tokenize_prompt(prompt)
 
+        if len(chosen_tokenized["input_ids"]) > max_length:
+            LOG.warning(
+                f"Chosen sequence exceeds max sequence length: {len(chosen_tokenized['input_ids'])}",
+            )
+
+            chosen_tokenized["input_ids"] = chosen_tokenized["input_ids"][:max_length]
+            chosen_tokenized["attention_mask"] = chosen_tokenized["attention_mask"][
+                :max_length
+            ]
+
         self.messages = "rejected_messages"
         # pylint: disable=duplicate-code
         prompt[self.messages] = []
@@ -52,6 +64,18 @@ class BTChatTemplateStrategy(ChatTemplateStrategy):
         )
         rejected_tokenized = super().tokenize_prompt(prompt)
 
+        if len(rejected_tokenized["input_ids"]) > max_length:
+            LOG.warning(
+                f"Rejected sequence exceeds max sequence length: {len(rejected_tokenized['input_ids'])}",
+            )
+
+            rejected_tokenized["input_ids"] = rejected_tokenized["input_ids"][
+                :max_length
+            ]
+            rejected_tokenized["attention_mask"] = rejected_tokenized["attention_mask"][
+                :max_length
+            ]
+
         return {
             "input_ids_chosen": chosen_tokenized["input_ids"],
             "attention_mask_chosen": chosen_tokenized["attention_mask"],
@@ -80,9 +104,9 @@ def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
         "roles": ds_cfg.get("roles"),
         "drop_system_message": ds_cfg.get("drop_system_message", False),
         # we need to add one for detecting sequences with exceeding the `sequence_len` limit.
-        "max_length": cfg.sequence_len + 1
-        if not cfg.reward_model
-        else cfg.sequence_len,
+        "max_length": (
+            cfg.sequence_len + 1 if not cfg.reward_model else cfg.sequence_len
+        ),
     }
 
     strategy_params = {

src/axolotl/prompt_strategies/chat_template.py

@@ -42,6 +42,7 @@ class ChatTemplatePrompter(Prompter):
                 "gpt": "assistant",
                 "system": "system",
             }
+
         self.message_field_role = message_field_role
         self.message_field_content = message_field_content
         self.message_field_training = message_field_training
@@ -53,21 +54,9 @@ class ChatTemplatePrompter(Prompter):
         self.drop_system_message = drop_system_message
 
     def build_prompt(self, conversation, add_generation_prompt=False, images=None):
-        turns = [
-            {
-                "role": self.roles[t[self.message_field_role]],
-                "content": t[self.message_field_content],
-                "training": t.get(self.message_field_training, None),
-            }
-            for t in conversation
-        ]
-
-        if self.drop_system_message and turns[0]["role"] == "system":
-            turns = turns[1:]
-
         if self.processor:
             text = self.processor.apply_chat_template(
-                turns,
+                conversation,
                 chat_template=self.chat_template,
                 tokenize=False,
                 add_generation_prompt=add_generation_prompt,
@@ -76,8 +65,6 @@ class ChatTemplatePrompter(Prompter):
                 text=text,
                 images=images,
                 return_tensors="pt",
-                truncation=True,
-                max_length=self.max_length,
             )
             # workaround since processor works in batches instead of single examples
             for k, val in batch.items():
@@ -88,9 +75,7 @@ class ChatTemplatePrompter(Prompter):
             return batch
 
         return self.tokenizer.apply_chat_template(
-            turns,
-            truncation=True,
-            max_length=self.max_length,
+            conversation,
             add_generation_prompt=add_generation_prompt,
             chat_template=self.chat_template,
         )
@@ -215,7 +200,14 @@ class ChatTemplateStrategy(PromptTokenizingStrategy):
         train_on_eos=None,
     ):
         super().__init__(prompter, tokenizer, train_on_inputs, sequence_len)
-        self.roles_to_train = roles_to_train if roles_to_train is not None else []
+
+        self.roles_to_train = []
+        if roles_to_train:
+            # map roles if exist in prompter.roles else use the role as is
+            self.roles_to_train = [
+                prompter.roles.get(role, role) for role in roles_to_train
+            ]
+
         self.train_on_eos = train_on_eos
         self.images = "images"
 
@@ -262,30 +254,28 @@ class ChatTemplateStrategy(PromptTokenizingStrategy):
 
             return tokenized_prompt
 
-        turns = prompt[self.messages]
+        turns = self.get_conversation_thread(prompt)
         input_ids = self.prompter.build_prompt(turns)
         labels = [IGNORE_TOKEN_ID] * len(input_ids)
 
         last_eos_idx = -1
         for index, turn in enumerate(turns):
-            role = turn.get(self.prompter.message_field_role)
-            content = turn.get(self.prompter.message_field_content)
-            train_turn = turn.get(self.prompter.message_field_training)
-            train_detail = turn.get(self.prompter.message_field_training_detail)
+            role = turn.get("role")
+            content = turn.get("content")
+            train_turn = turn.get("training")
+            train_detail = turn.get("training_detail")
 
             LOG.debug(
                 f"Processing turn {index}: role={role}, content={content}, train_turn={train_turn}, train_detail={train_detail}"
             )
 
-            should_train = (
-                train_turn
-                if train_turn is not None
-                else (
-                    bool(train_detail is not None)
-                    if train_detail is not None
-                    else self.train_on_inputs or role in self.roles_to_train
-                )
-            )
+            should_train = None
+            if train_turn is not None:
+                should_train = train_turn
+            elif train_detail is not None:
+                should_train = bool(train_detail)
+            else:
+                should_train = self.train_on_inputs or role in self.roles_to_train
 
             LOG.debug(f"Should train: {should_train}")
 
@@ -293,6 +283,9 @@ class ChatTemplateStrategy(PromptTokenizingStrategy):
                 conversation_ids=input_ids, turn=index, turn_content=turn
             )
 
+            if turn_start_idx == -1 or turn_end_idx == -1:
+                LOG.warning(f"Failed to find boundaries for turn {index}")
+
             LOG.debug(f"Turn indices: start={turn_start_idx}, end={turn_end_idx}")
 
             if should_train and turn_start_idx != -1 and turn_end_idx != -1:
@@ -313,7 +306,9 @@ class ChatTemplateStrategy(PromptTokenizingStrategy):
                     labels[turn_start_idx:turn_end_idx] = input_ids[
                         turn_start_idx:turn_end_idx
                     ]
-                    LOG.debug(f"Labels set for range {turn_start_idx}:{turn_end_idx}")
+                    LOG.debug(
+                        f"Set labels for training from {turn_start_idx} to {turn_end_idx}"
+                    )
 
                 LOG.debug(f"Labels after processing turn {index}: {labels}")
 
@@ -351,52 +346,73 @@ class ChatTemplateStrategy(PromptTokenizingStrategy):
                 return i
         return -1
 
-    def find_turn(self, conversation_ids, turn, turn_content):
+    def find_turn(self, conversation_ids: list[int], turn: int, turn_content: dict):
         """
         Locate the starting and ending indices of the specified turn in a conversation.
-
-        Args:
-            conversation_ids (list[int]): Token IDs representing the conversation.
-            turn (int): The turn number to locate (based on EOS tokens).
-            turn_content (str): String containing the content of the turn.
-
-        Returns:
-            tuple: (start_idx, end_idx) indices of the start and end of the turn content.
-                   Returns (-1, -1) if the turn content is not found.
         """
-        content = turn_content.get(self.prompter.message_field_content, "")
+        content = turn_content.get("content")
         content_ids = self.tokenizer.encode(content, add_special_tokens=False)
 
-        eos_token_id = self.tokenizer.eos_token_id
-        eos_count = 0
-        start_search_idx = 0
+        LOG.debug(f"content_ids (length {len(content_ids)}): {content_ids}")
 
-        # Locate the starting index after the specified number of EOS tokens
-        for i, token_id in enumerate(conversation_ids):
-            if token_id == eos_token_id:
-                eos_count += 1
-                if eos_count == turn:
-                    start_search_idx = (
-                        i + 1
-                    )  # Start searching after the specified turn's EOS token
-                    break
+        if not content_ids:
+            LOG.warning(f"Empty content for turn {turn}")
+            return -1, -1
 
-        # Find the start index of the content within the conversation
-        start_idx = -1
-        for i in range(start_search_idx, len(conversation_ids) - len(content_ids) + 1):
-            if conversation_ids[i : i + len(content_ids)] == content_ids:
-                start_idx = i
-                break
-
-        if start_idx != -1:
-            end_idx = start_idx + len(content_ids)
+        # For first turn, start from beginning
+        if turn == 0:
+            start_search_idx = 0
         else:
-            end_idx = -1
+            # For subsequent turns, find the previous EOS token
+            eos_token_id = self.tokenizer.eos_token_id
+            eos_count = 0
+            start_search_idx = 0
 
-        return start_idx, end_idx
+            for i, token_id in enumerate(conversation_ids):
+                if token_id == eos_token_id:
+                    eos_count += 1
+                    if eos_count == turn:  # Find the nth EOS token where n = turn
+                        start_search_idx = i + 1
+                        break
+
+        # we can optimize this to only search for a few tokens from start_search_idx
+        # but it would risk missing the content if it's not found within the first few tokens or
+        # if start_search_idx cannot be found above.
+        last_index = len(conversation_ids) - len(content_ids) + 1
+
+        if last_index < start_search_idx:
+            LOG.warning(
+                f"last_index to search is less than start_search_idx for turn {turn}"
+            )
+            return -1, -1
+
+        # Search for content starting from start_search_idx
+        first_elem = content_ids[0]
+        for i in range(start_search_idx, last_index):
+            # Quick check of first element before doing full comparison
+            if conversation_ids[i] == first_elem:
+                # Check if the rest of the content matches
+                if conversation_ids[i : i + len(content_ids)] == content_ids:
+                    LOG.debug(f"Found turn {turn} content at position {i}")
+                    return i, i + len(content_ids)
+
+        return -1, -1
 
     def get_conversation_thread(self, prompt):
-        return prompt[self.messages]
+        turns = [
+            {
+                "role": self.prompter.roles[t[self.prompter.message_field_role]],
+                "content": t[self.prompter.message_field_content],
+                "training": t.get(self.prompter.message_field_training),
+                "training_detail": t.get(self.prompter.message_field_training_detail),
+            }
+            for t in prompt[self.messages]
+        ]
+
+        if self.prompter.drop_system_message and turns[0]["role"] == "system":
+            turns = turns[1:]
+
+        return turns
 
     def get_images(self, prompt):
         return prompt.get(self.images, None)

src/axolotl/train.py

@@ -260,9 +260,28 @@ def train(
 
     if not cfg.hub_model_id:
         try:
-            trainer.create_model_card(
-                model_name=cfg.output_dir.lstrip("./").encode("utf-8").decode("utf-8")
-            )
+            model_card_kwarg = {
+                "model_name": cfg.output_dir.lstrip("./")
+                .encode("utf-8")
+                .decode("utf-8")
+            }
+            if cfg.datasets is not None:
+                if cfg.rl is not None or cfg.reward_model:
+                    dataset_tags = [
+                        d["path"] for d in cfg.datasets if not Path(d["path"]).is_dir()
+                    ]
+                    if dataset_tags:
+                        # guard as create_model_card may fail if dataset_tags is empty list
+                        model_card_kwarg["dataset_name"] = dataset_tags
+                else:
+                    dataset_tags = [
+                        d["path"] for d in cfg.datasets if not Path(d["path"]).is_dir()
+                    ]
+                    if dataset_tags:
+                        # guard as create_model_card may fail if dataset_tags is empty list
+                        model_card_kwarg["dataset_tags"] = dataset_tags
+
+            trainer.create_model_card(**model_card_kwarg)
         except (AttributeError, UnicodeDecodeError):
             pass
     elif cfg.hub_model_id:

src/axolotl/utils/__init__.py

@@ -1,7 +1,11 @@
 """
 Basic utils for Axolotl
 """
+
 import importlib.util
+import re
+
+import torch
 
 
 def is_mlflow_available():
@@ -10,3 +14,23 @@ def is_mlflow_available():
 
 def is_comet_available():
     return importlib.util.find_spec("comet_ml") is not None
+
+
+# pylint: disable=duplicate-code
+def get_pytorch_version() -> tuple[int, int, int]:
+    """
+    Get Pytorch version as a tuple of (major, minor, patch).
+    """
+    torch_version = torch.__version__
+    version_match = re.match(r"^(\d+)\.(\d+)(?:\.(\d+))?", torch_version)
+
+    if not version_match:
+        raise ValueError("Invalid version format")
+
+    major, minor, patch = version_match.groups()
+    major, minor = int(major), int(minor)
+    patch = int(patch) if patch is not None else 0  # Default patch to 0 if not present
+    return major, minor, patch
+
+
+# pylint: enable=duplicate-code

src/axolotl/utils/bench.py

@@ -1,9 +1,23 @@
 """Benchmarking and measurement utilities"""
 import functools
 
-import pynvml
 import torch
-from pynvml.nvml import NVMLError
+from transformers.utils.import_utils import is_torch_npu_available
+
+from axolotl.utils.distributed import get_device_type
+
+try:
+    from pynvml import (
+        NVMLError,
+        nvmlDeviceGetHandleByIndex,
+        nvmlDeviceGetMemoryInfo,
+        nvmlInit,
+    )
+except ImportError:
+    NVMLError = None
+    nvmlDeviceGetHandleByIndex = None
+    nvmlDeviceGetMemoryInfo = None
+    nvmlInit = None
 
 
 def check_cuda_device(default_value):
@@ -53,24 +67,35 @@ def mps_memory_usage_all():
     return usage, reserved - usage, 0
 
 
+def npu_memory_usage_all(device=0):
+    usage = torch.npu.memory_allocated(device) / 1024.0**3
+    reserved = torch.npu.memory_reserved(device) / 1024.0**3
+    return usage, reserved - usage, 0
+
+
 @check_cuda_device(0.0)
 def gpu_memory_usage_smi(device=0):
     if isinstance(device, torch.device):
         device = device.index
     if isinstance(device, str) and device.startswith("cuda:"):
         device = int(device[5:])
+    if not nvmlInit:
+        return 0.0
     try:
-        pynvml.nvmlInit()
-        handle = pynvml.nvmlDeviceGetHandleByIndex(device)
-        info = pynvml.nvmlDeviceGetMemoryInfo(handle)
+        nvmlInit()
+        handle = nvmlDeviceGetHandleByIndex(device)
+        info = nvmlDeviceGetMemoryInfo(handle)
         return info.used / 1024.0**3
     except NVMLError:
         return 0.0
 
 
 def log_gpu_memory_usage(log, msg, device):
+    cur_device = get_device_type()
     if torch.backends.mps.is_available():
         usage, cache, misc = mps_memory_usage_all()
+    elif "npu" in str(cur_device) and is_torch_npu_available():
+        usage, cache, misc = npu_memory_usage_all(device)
     else:
         usage, cache, misc = gpu_memory_usage_all(device)
     extras = []
@@ -79,6 +104,7 @@ def log_gpu_memory_usage(log, msg, device):
     if misc > 0:
         extras.append(f"+{misc:.03f}GB misc")
     log.info(
-        f"GPU memory usage {msg}: {usage:.03f}GB ({', '.join(extras)})", stacklevel=2
+        f"{str(cur_device)} memory usage {msg}: {usage:.03f}GB ({', '.join(extras)})",
+        stacklevel=2,
     )
     return usage, cache, misc

src/axolotl/utils/callbacks/__init__.py

@@ -28,6 +28,7 @@ from transformers import (
     TrainingArguments,
 )
 from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR, IntervalStrategy
+from trl.models import unwrap_model_for_generation
 
 from axolotl.utils import is_comet_available, is_mlflow_available
 from axolotl.utils.bench import log_gpu_memory_usage
@@ -46,6 +47,7 @@ from axolotl.utils.distributed import (
 if TYPE_CHECKING:
     from axolotl.core.trainer_builder import AxolotlTrainingArguments
 
+
 IGNORE_INDEX = -100
 LOG = logging.getLogger("axolotl.callbacks")
 
@@ -375,7 +377,10 @@ def causal_lm_bench_eval_callback_factory(trainer: Trainer, tokenizer):
             for metric in self.cfg.eval_causal_lm_metrics:
                 if metric == "perplexity":
                     max_seq_len = self.cfg.eval_max_new_tokens
-                    metrics[metric] = Perplexity(trainer.model, tokenizer, max_seq_len)
+                    metrics[metric] = Perplexity(
+                        tokenizer=tokenizer,
+                        max_seq_len=max_seq_len,
+                    )
                 else:
                     try:
                         metrics[metric] = evaluate.load(metric)
@@ -392,8 +397,11 @@ def causal_lm_bench_eval_callback_factory(trainer: Trainer, tokenizer):
             eval_dataloader,
             **kwargs,  # pylint: disable=unused-argument
         ):
-            trainer.model.eval()
-            device = torch.device(self.cfg.device)
+            trainer.model_wrapped.eval()
+
+            device = torch.device(
+                self.cfg.device
+            )  # Use this instead of trainer.model_wrapped.device as it may return cpu if fsdp offloaded
 
             # pylint: disable=duplicate-code
             generation_config = GenerationConfig(
@@ -430,6 +438,10 @@ def causal_lm_bench_eval_callback_factory(trainer: Trainer, tokenizer):
                         for k in metric._feature_names()  # pylint: disable=protected-access
                         if k in kwargs
                     }
+
+                    if isinstance(metric, Perplexity):
+                        metric_kwargs["model"] = trainer.model_wrapped
+
                     metric_score = metric.compute(**metric_kwargs)
                     return (
                         metric_score["score"]
@@ -465,89 +477,97 @@ def causal_lm_bench_eval_callback_factory(trainer: Trainer, tokenizer):
             def predict_with_generate():
                 eval_src, eval_pred, eval_ref = [], [], []
 
-                for batch in tqdm(eval_dataloader):
-                    batch_labels = batch["labels"].to(device)
-                    batch_input_ids = batch["input_ids"].to(device)
+                with unwrap_model_for_generation(
+                    trainer.model_wrapped, trainer.accelerator
+                ) as unwrapped_model:
+                    for batch in tqdm(eval_dataloader, disable=not is_main_process()):
+                        batch_labels = batch["labels"].to(device)
+                        batch_input_ids = batch["input_ids"].to(device)
 
-                    if "position_ids" in batch:
-                        batch_pos_ids = batch["position_ids"].tolist()
-                    else:
-                        batch_pos_ids = [None] * len(batch["input_ids"])
-
-                    prompt_token_ids_list = []
-                    completion_token_ids_list = []
-
-                    for input_ids_all, labels_all, pos_ids in zip(
-                        batch_input_ids,
-                        batch_labels,
-                        batch_pos_ids,
-                    ):
-                        if pos_ids is None:
-                            pos_ranges = [(0, len(input_ids_all) - 1)]
+                        if "position_ids" in batch:
+                            batch_pos_ids = batch["position_ids"].tolist()
                         else:
-                            pos_ranges = find_ranges(pos_ids)
+                            batch_pos_ids = [None] * len(batch["input_ids"])
 
-                        for pos_range in pos_ranges:
-                            start, end = pos_range
-                            if start == end:
-                                continue
+                        prompt_token_ids_list = []
+                        completion_token_ids_list = []
 
-                            input_ids = input_ids_all[start : end + 1]
-                            labels = labels_all[start : end + 1]
+                        for input_ids_all, labels_all, pos_ids in zip(
+                            batch_input_ids,
+                            batch_labels,
+                            batch_pos_ids,
+                        ):
+                            if pos_ids is None:
+                                pos_ranges = [(0, len(input_ids_all) - 1)]
+                            else:
+                                pos_ranges = find_ranges(pos_ids)
 
-                            tokens_without_loss = labels == IGNORE_INDEX
-                            tokens_with_loss = labels != IGNORE_INDEX
-                            tokens_exclude_padding = input_ids != tokenizer.pad_token_id
-                            prompt_token_includes = (
-                                tokens_without_loss & tokens_exclude_padding
+                            for pos_range in pos_ranges:
+                                start, end = pos_range
+                                if start == end:
+                                    continue
+
+                                input_ids = input_ids_all[start : end + 1]
+                                labels = labels_all[start : end + 1]
+
+                                tokens_without_loss = labels == IGNORE_INDEX
+                                tokens_with_loss = labels != IGNORE_INDEX
+                                tokens_exclude_padding = (
+                                    input_ids != tokenizer.pad_token_id
+                                )
+                                prompt_token_includes = (
+                                    tokens_without_loss & tokens_exclude_padding
+                                )
+
+                                prompt_token_ids = input_ids[prompt_token_includes]
+                                prompt_token_ids_list.append(prompt_token_ids)
+
+                                completion_token_ids = input_ids[tokens_with_loss]
+                                completion_token_ids_list.append(completion_token_ids)
+
+                        prompt_texts = tokenizer.batch_decode(
+                            prompt_token_ids_list, skip_special_tokens=True
+                        )
+                        completion_texts = tokenizer.batch_decode(
+                            completion_token_ids_list, skip_special_tokens=True
+                        )
+
+                        with torch.no_grad():
+                            prompt_encoding = tokenizer(
+                                prompt_texts, padding=True, return_tensors="pt"
+                            ).to(device)
+
+                            predictions = unwrapped_model.generate(
+                                **prompt_encoding, generation_config=generation_config
                             )
 
-                            prompt_token_ids = input_ids[prompt_token_includes]
-                            prompt_token_ids_list.append(prompt_token_ids)
+                            del prompt_encoding
 
-                            completion_token_ids = input_ids[tokens_with_loss]
-                            completion_token_ids_list.append(completion_token_ids)
+                        prediction_all_tokens = predictions["sequences"].cpu().tolist()
+                        prediction_without_prompt_tokens_list = []
+                        for prompt_token_ids, prediction_tokens in zip(
+                            prompt_token_ids_list, prediction_all_tokens
+                        ):
+                            prediction_without_prompt_tokens = prediction_tokens[
+                                len(prompt_token_ids) :
+                            ]
+                            prediction_without_prompt_tokens_list.append(
+                                prediction_without_prompt_tokens
+                            )
 
-                    prompt_texts = tokenizer.batch_decode(
-                        prompt_token_ids_list, skip_special_tokens=True
-                    )
-                    completion_texts = tokenizer.batch_decode(
-                        completion_token_ids_list, skip_special_tokens=True
-                    )
-
-                    with torch.no_grad():
-                        prompt_encoding = tokenizer(
-                            prompt_texts, padding=True, return_tensors="pt"
-                        ).to(self.cfg.device)
-                        predictions = trainer.model.generate(
-                            **prompt_encoding, generation_config=generation_config
+                        predicted_texts = tokenizer.batch_decode(
+                            prediction_without_prompt_tokens_list,
+                            skip_special_tokens=True,
                         )
 
-                    prediction_all_tokens = predictions["sequences"].cpu().tolist()
-                    prediction_without_prompt_tokens_list = []
-                    for prompt_token_ids, prediction_tokens in zip(
-                        prompt_token_ids_list, prediction_all_tokens
-                    ):
-                        prediction_without_prompt_tokens = prediction_tokens[
-                            len(prompt_token_ids) :
-                        ]
-                        prediction_without_prompt_tokens_list.append(
-                            prediction_without_prompt_tokens
-                        )
-
-                    predicted_texts = tokenizer.batch_decode(
-                        prediction_without_prompt_tokens_list, skip_special_tokens=True
-                    )
-
-                    eval_src.extend(prompt_texts)
-                    eval_pred.extend(predicted_texts)
-                    eval_ref.extend(completion_texts)
+                        eval_src.extend(prompt_texts)
+                        eval_pred.extend(predicted_texts)
+                        eval_ref.extend(completion_texts)
 
                 return eval_src, eval_pred, eval_ref
 
-            if is_main_process():
-                eval_preds = predict_with_generate()
-                trainer.log(evaluate_preds(*eval_preds))
+            eval_preds = predict_with_generate()
+            trainer.log(evaluate_preds(*eval_preds))
 
             return control
 

src/axolotl/utils/callbacks/perplexity.py

@@ -8,6 +8,8 @@ from transformers.modeling_outputs import CausalLMOutput
 from transformers.modeling_utils import PreTrainedModel
 from transformers.tokenization_utils import PreTrainedTokenizer
 
+from axolotl.utils.distributed import is_main_process
+
 
 class Perplexity:
     """
@@ -17,16 +19,13 @@ class Perplexity:
 
     def __init__(
         self,
-        model: PreTrainedModel,
         tokenizer: PreTrainedTokenizer,
         max_seq_len: int,
         stride: int = 512,
     ) -> None:
         self.max_seq_len = max_seq_len
         self.stride = stride
-        self.model = model
         self.tokenizer = tokenizer
-        self.device = model.device
         self.name = "perplexity"
 
     def _feature_names(self) -> List[str]:
@@ -34,6 +33,7 @@ class Perplexity:
 
     def compute(
         self,
+        model: PreTrainedModel,
         references: Optional[List[str]] = None,
     ) -> Dict[str, float]:
         """
@@ -41,17 +41,21 @@ class Perplexity:
         """
         assert references is not None, "Missing parameter: references"
 
+        model.eval()
+
         references_tokenized = self.tokenizer(
             references, return_tensors="pt", padding=True, truncation=True
         )
         input_ids: Tensor = references_tokenized["input_ids"]  # type: ignore
-        input_ids = input_ids.to(self.device)
+        input_ids = input_ids.to(model.device)
 
         sequence_length = input_ids.size(1)
 
         losses = []
         prev_end_loc = 0
-        for begin_loc in tqdm(range(0, sequence_length, self.stride)):
+        for begin_loc in tqdm(
+            range(0, sequence_length, self.stride), disable=not is_main_process()
+        ):
             end_loc = min(begin_loc + self.max_seq_len, sequence_length)
             trg_len = end_loc - prev_end_loc
             input_ids_slice = input_ids[:, begin_loc:end_loc]
@@ -59,7 +63,7 @@ class Perplexity:
             labels_slice[:, :-trg_len] = -100
 
             with torch.no_grad():
-                outputs: CausalLMOutput = self.model(
+                outputs: CausalLMOutput = model(
                     input_ids=input_ids_slice, labels=labels_slice
                 )
 

src/axolotl/utils/collators/mm_chat.py

@@ -1,8 +1,10 @@
 """
 Collators for multi-modal chat messages and packing
 """
+
+from copy import deepcopy
 from dataclasses import dataclass
-from typing import Any, Dict, List, Optional, Union
+from typing import Any, Optional, Union
 
 from PIL import Image
 from transformers import PreTrainedTokenizerBase, ProcessorMixin
@@ -30,8 +32,8 @@ class MultiModalChatDataCollator(DataCollatorMixin):
             raise ValueError("Packing is currently not supported.")
 
     def torch_call(
-        self, examples: List[Union[List[int], Any, Dict[str, Any]]]
-    ) -> Dict[str, Any]:
+        self, examples: list[Union[list[int], Any, dict[str, Any]]]
+    ) -> dict[str, Any]:
         # Handle dict or lists with proper padding and conversion to tensor.
 
         return self.__class__.process_rows(
@@ -46,6 +48,120 @@ class MultiModalChatDataCollator(DataCollatorMixin):
         # *** This is COPIED from the trl example sft_vlm.py code ***
         # use this as a starting point
 
+        def _preprocess(examples: list[dict]) -> list[dict]:
+            """
+            Preprocess conversation examples to ensure consistent format.
+
+            Converts different conversation formats to OpenAI format with 'messages'.
+            Supports two formats:
+            1. OpenAI format with 'messages'
+            2. Legacy format with 'conversations'
+
+            Args:
+                examples: list of conversation dictionaries
+
+            Returns:
+                dict in OpenAI format with 'messages' key
+
+            Raises:
+                ValueError: If the conversation format is not supported
+            """
+            role_mapping = {
+                "human": "user",
+                "gpt": "assistant",
+            }
+
+            def normalize_role(role: str) -> str:
+                """Normalize role names to OpenAI format. Default to original role if not found."""
+                return role_mapping.get(role, role)
+
+            def convert_legacy_format(example: dict) -> dict:
+                """Convert legacy 'conversations' format to OpenAI 'messages' format."""
+                messages = [
+                    {
+                        "role": normalize_role(convo["from"]),
+                        "content": convo["value"],
+                    }
+                    for convo in example["conversations"]
+                ]
+
+                # Create new dict without 'conversations' key
+                result = deepcopy(example)
+                result.pop("conversations")
+                return {"messages": messages, **result}
+
+            processed_examples = []
+            for example in examples:
+                # OpenAI format
+                if "messages" in example:
+                    processed_examples.append(example)
+
+                # Legacy format
+                elif "conversations" in example:
+                    processed_examples.append(convert_legacy_format(example))
+
+                else:
+                    raise ValueError(
+                        "Only `messages` and `conversations` message keys are currently supported."
+                    )
+
+            return processed_examples
+
+        def _process_images(examples, max_images):
+            """
+            Process images from examples, ensuring consistency in image presence and applying max_images limit.
+
+            Args:
+                examples: List of dictionaries that may contain 'images' key
+                max_images: Maximum number of images to keep per example (0 means no limit)
+
+            Returns:
+                Either None (if no images) or List[Image objects] (if all examples have images)
+
+            Raises:
+                ValueError: If there's a mix of None and non-None images
+            """
+
+            def get_image(example):
+                if "images" not in example:
+                    return None
+                images = example["images"]
+                if isinstance(images, str):
+                    return Image.open(images)
+                return images
+
+            images = [get_image(example) for example in examples]
+
+            # Count None and non-None images
+            none_count = sum(1 for img in images if img is None)
+
+            # All images are None
+            if none_count == len(images):
+                return None
+
+            # Mix of None and non-None images
+            if none_count > 0:
+                raise ValueError(
+                    "All images should be either None or not None. "
+                    "Please provide images for all examples or None."
+                )
+
+            # Apply max_images limit if specified
+            if max_images > 0:
+                images = [
+                    (
+                        img_batch[:max_images]
+                        if isinstance(img_batch, (list, tuple))
+                        else img_batch
+                    )
+                    for img_batch in images
+                ]
+
+            return images
+
+        # Preprocess the examples
+        examples = _preprocess(examples)
+
         # Get the texts and images, and apply the chat template
         texts = [
             processor.apply_chat_template(
@@ -53,15 +169,8 @@ class MultiModalChatDataCollator(DataCollatorMixin):
             )
             for example in examples
         ]
-        images = [
-            Image.open(example["images"])
-            if isinstance(example["images"], str)
-            else example["images"]
-            for example in examples
-        ]
 
-        if max_images > 0:
-            images = [img_batch[:max_images] for img_batch in images]
+        images = _process_images(examples, max_images=max_images)
 
         # Tokenize the texts and process the images
         batch = processor(text=texts, images=images, return_tensors="pt", padding=True)

src/axolotl/utils/config/__init__.py

@@ -5,7 +5,9 @@ from typing import Optional
 
 import torch
 from transformers.utils import is_torch_bf16_gpu_available
+from transformers.utils.import_utils import is_torch_npu_available
 
+from axolotl.integrations.base import PluginManager
 from axolotl.integrations.config import merge_input_args
 from axolotl.utils.bench import log_gpu_memory_usage
 from axolotl.utils.config.models.input.v0_4_1 import (
@@ -29,7 +31,10 @@ def choose_device(cfg):
             if torch.backends.mps.is_available():
                 return "mps"
 
-            raise SystemError("No CUDA/mps device found")
+            if is_torch_npu_available():
+                return f"npu:{cfg.local_rank}"
+
+            raise SystemError("No CUDA/mps/npu device found")
         except Exception:  # pylint: disable=broad-exception-caught
             return "cpu"
 
@@ -39,6 +44,8 @@ def choose_device(cfg):
     else:
         if cfg.device.startswith("cuda"):
             cfg.device_map = {"": torch.cuda.current_device()}
+        elif cfg.device.startswith("npu"):
+            cfg.device_map = {"npu": torch.npu.current_device()}
         else:
             cfg.device_map = {"": cfg.device}
 
@@ -146,7 +153,7 @@ def normalize_config(cfg):
     cfg.is_llama_derived_model = (
         (
             hasattr(model_config, "model_type")
-            and model_config.model_type == ["llama", "mllama_text_model"]
+            and model_config.model_type in ["llama", "mllama_text_model"]
         )
         or cfg.is_llama_derived_model
         or "llama" in cfg.base_model.lower()
@@ -223,7 +230,11 @@ def normalize_cfg_datasets(cfg):
                     cfg.datasets[idx].chat_template_jinja = cfg.chat_template_jinja
 
 
-def validate_config(cfg: DictDefault, capabilities: Optional[dict] = None):
+def validate_config(
+    cfg: DictDefault,
+    capabilities: Optional[dict] = None,
+    env_capabilities: Optional[dict] = None,
+):
     AxolotlConfigWCapabilities = AxolotlConfigWCapabilitiesBase
     AxolotlInputConfig = AxolotlInputConfigBase
 
@@ -233,14 +244,35 @@ def validate_config(cfg: DictDefault, capabilities: Optional[dict] = None):
             AxolotlInputConfig,  # pylint: disable=invalid-name
         ) = merge_input_args()
 
-    if capabilities:
+    if capabilities or env_capabilities:
+        if (capabilities and not env_capabilities) or (
+            env_capabilities and not capabilities
+        ):
+            raise ValueError(
+                "Both capabilities and env_capabilities must be provided or not provided."
+            )
+
         return DictDefault(
             dict(
                 AxolotlConfigWCapabilities(
-                    **cfg.to_dict(), capabilities=capabilities
+                    **cfg.to_dict(),
+                    capabilities=capabilities,
+                    env_capabilities=env_capabilities,
                 ).model_dump(exclude_none=True)
             )
         )
+
     return DictDefault(
         dict(AxolotlInputConfig(**cfg.to_dict()).model_dump(exclude_none=True))
     )
+
+
+def prepare_plugins(cfg):
+    """
+    Prepare the plugins for the configuration
+    """
+
+    if cfg.get("plugins"):
+        plugin_manager = PluginManager.get_instance()
+        for plugin_name in cfg["plugins"]:
+            plugin_manager.register(plugin_name)

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

@@ -7,9 +7,9 @@ Module for pydantic models for configuration
 import logging
 import os
 from enum import Enum
-from importlib.metadata import version
 from typing import Annotated, Any, Dict, List, Literal, Optional, Tuple, Union
 
+from packaging import version
 from pydantic import (
     BaseModel,
     Field,
@@ -20,8 +20,9 @@ from pydantic import (
 )
 from transformers import SchedulerType
 from transformers.training_args import OptimizerNames
+from transformers.utils.import_utils import is_torch_npu_available
 
-from axolotl.utils.config.models.internals import GPUCapabilities
+from axolotl.utils.config.models.internals import EnvCapabilities, GPUCapabilities
 
 LOG = logging.getLogger("axolotl.utils.config.models.input")
 
@@ -322,11 +323,13 @@ class LoraConfig(BaseModel):
     @model_validator(mode="before")
     @classmethod
     def validate_adapter(cls, data):
-        if not data.get("adapter") and (
-            data.get("load_in_8bit") or data.get("load_in_4bit")
+        if (
+            not data.get("adapter")
+            and not data.get("inference")
+            and (data.get("load_in_8bit") or data.get("load_in_4bit"))
         ):
             raise ValueError(
-                "load_in_8bit and load_in_4bit are not supported without setting an adapter."
+                "load_in_8bit and load_in_4bit are not supported without setting an adapter for training."
                 "If you want to full finetune, please turn off load_in_8bit and load_in_4bit."
             )
         return data
@@ -430,6 +433,8 @@ class HyperparametersConfig(BaseModel):
     group_by_length: Optional[bool] = None
 
     learning_rate: Union[str, float]
+    embedding_lr: Optional[float] = None
+    embedding_lr_scale: Optional[float] = None
     weight_decay: Optional[float] = 0.0
     optimizer: Optional[
         Union[
@@ -622,6 +627,7 @@ class AxolotlInputConfig(
         json_schema_extra={"description": "streaming dataset to use for pretraining"},
     )
     dataset_processes: Optional[int] = Field(default=os.cpu_count())
+    dataset_exact_deduplication: Optional[bool] = None
     dataset_keep_in_memory: Optional[bool] = None
     dataloader_pin_memory: Optional[bool] = None
     dataloader_num_workers: Optional[int] = None
@@ -1314,6 +1320,7 @@ class AxolotlInputConfig(
             and data.get("gradient_checkpointing_kwargs", {})
             and data.get("gradient_checkpointing_kwargs", {}).get("use_reentrant")
             is False
+            and data.get("deepspeed", "") is not None
             and "zero3" in data.get("deepspeed", "")
         ):
             # may result in:
@@ -1425,21 +1432,6 @@ class AxolotlInputConfig(
                 )
         return data
 
-    @model_validator(mode="before")
-    @classmethod
-    def check_unsloth_xformers_version(cls, data):
-        if (
-            data.get("unsloth_lora_mlp")
-            or data.get("unsloth_lora_qkv")
-            or data.get("unsloth_lora_o")
-        ):
-            xformers_version = version("xformers")
-            if xformers_version == "0.0.27":
-                raise ValueError(
-                    "xformers version 0.0.27 is not supported with unsloth. Please downgrade to 0.0.26.post1"
-                )
-        return data
-
     @model_validator(mode="before")
     @classmethod
     def check_torch_compile_deepspeed(cls, data):
@@ -1449,11 +1441,67 @@ class AxolotlInputConfig(
             )
         return data
 
+    @model_validator(mode="before")
+    @classmethod
+    def check_npu_config(cls, data):
+        if is_torch_npu_available():
+            # check attention config
+            attn_list = ["flash_attention", "sdp_attention", "s2_attention"]
+            for attn in attn_list:
+                if data.get(attn):
+                    raise NotImplementedError(
+                        f"{attn} is currently not supported in Ascend npu, please disable this configuration."
+                    )
+
+            # check quant config
+            if data.get("optimizer") is not None and "bit" in data.get("optimizer"):
+                optimizer = data.get("optimizer")
+                raise NotImplementedError(
+                    f"{optimizer} is currently not supported in Ascend npu, choose another one please."
+                )
+
+            quant_list = ["load_in_8bit", "load_in_4bit"]
+            for quant in quant_list:
+                if data.get(quant):
+                    raise NotImplementedError(
+                        f"Quantification is currently not supported in Ascend npu, please disable {quant}."
+                    )
+
+            # check dtype config
+            if data.get("tf32"):
+                raise NotImplementedError(
+                    "tf32 dtype is currently not supported in Ascend npu, please disable this configuration"
+                )
+
+        return data
+
+    @model_validator(mode="before")
+    @classmethod
+    def check_kto_config(cls, data):
+        if data.get("rl") == "kto":
+            if data.get("sample_packing") or data.get("eval_sample_packing"):
+                raise ValueError("sample_packing is not supported with kto")
+
+            if data.get("remove_unused_columns") is not False:
+                raise ValueError("Set `remove_unused_columns: False` when using kto")
+
+            if data.get("gradient_checkpointing") and not (
+                data.get("gradient_checkpointing_kwargs")
+                and isinstance(data.get("gradient_checkpointing_kwargs"), dict)
+                and data["gradient_checkpointing_kwargs"].get("use_reentrant")
+            ):
+                raise ValueError(
+                    "Set `gradient_checkpointing_kwargs: {use_reentrant: true}` for when kto is enabled"
+                )
+
+        return data
+
 
 class AxolotlConfigWCapabilities(AxolotlInputConfig):
     """wrapper to valdiate gpu capabilities with the configured options"""
 
     capabilities: GPUCapabilities
+    env_capabilities: EnvCapabilities
 
     @model_validator(mode="after")
     def check_bf16(self):
@@ -1494,19 +1542,6 @@ class AxolotlConfigWCapabilities(AxolotlInputConfig):
 
         return data
 
-    @model_validator(mode="before")
-    @classmethod
-    def check_hopper_8bit_lora(cls, data):
-        is_sm_90: bool = (
-            data["capabilities"]
-            and data["capabilities"].get("compute_capability") == "sm_90"
-        )
-        if data.get("adapter") and data.get("load_in_8bit") and is_sm_90:
-            # see https://github.com/bitsandbytes-foundation/bitsandbytes/issues/538#issuecomment-2262945464
-            raise ValueError("8-bit LoRA is not supported on Hopper GPUs")
-
-        return data
-
     @model_validator(mode="before")
     @classmethod
     def check_fsdp_deepspeed(cls, data):
@@ -1528,3 +1563,21 @@ class AxolotlConfigWCapabilities(AxolotlInputConfig):
                     "unsloth_lora_mlp, unsloth_lora_qkv, and unsloth_lora_o are not compatible with multi-GPU training."
                 )
         return data
+
+    @model_validator(mode="before")
+    @classmethod
+    def check_adopt_torch_version(cls, data):
+        if (data.get("optimizer") is not None) and ("adopt" in data.get("optimizer")):
+            env_capabilities = data.get("env_capabilities", {})
+            torch_version = env_capabilities.get("torch_version")
+
+            if torch_version is None:
+                import torch
+
+                torch_version = str(torch.__version__).split("+", maxsplit=1)[0]
+
+            if version.parse(torch_version) < version.parse("2.5.1"):
+                raise ValueError(
+                    "ADOPT optimizer is incompatible with torch version < 2.5.1"
+                )
+        return data

src/axolotl/utils/config/models/internals/__init__.py

@@ -12,3 +12,9 @@ class GPUCapabilities(BaseModel):
     n_gpu: int = Field(default=1)
     n_node: int = Field(default=1)
     compute_capability: Optional[str] = Field(default=None)
+
+
+class EnvCapabilities(BaseModel):
+    """model to manage the environment capabilities statically"""
+
+    torch_version: Optional[str] = Field(default=None)

src/axolotl/utils/data/rl.py

@@ -13,7 +13,7 @@ from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
 from axolotl.prompt_strategies.dpo import load as load_dpo
 from axolotl.prompt_strategies.kto import load as load_kto
 from axolotl.prompt_strategies.orpo import load as load_orpo
-from axolotl.utils.data.utils import md5
+from axolotl.utils.data.utils import deduplicate_and_log_datasets, md5
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import is_main_process, zero_first
 from axolotl.utils.models import load_tokenizer
@@ -208,4 +208,9 @@ def load_prepare_dpo_datasets(cfg):
         if eval_dataset and not eval_is_preprocessed:
             _save_preprocessed_ds(cfg, cfg.test_datasets, eval_dataset)
 
+    if cfg.dataset_exact_deduplication:
+        train_dataset, eval_dataset, _ = deduplicate_and_log_datasets(
+            train_dataset=train_dataset, eval_dataset=eval_dataset
+        )
+
     return train_dataset, eval_dataset

src/axolotl/utils/data/sft.py

@@ -2,11 +2,9 @@
 
 import functools
 import logging
-import time
 from pathlib import Path
 from typing import List, Optional, Tuple, Union
 
-import requests
 from datasets import (
     Dataset,
     DatasetDict,
@@ -44,7 +42,11 @@ from axolotl.prompters import (
     UnsupportedPrompter,
 )
 from axolotl.utils.data.pretraining import wrap_pretraining_dataset
-from axolotl.utils.data.utils import md5
+from axolotl.utils.data.utils import (
+    deduplicate_and_log_datasets,
+    md5,
+    retry_on_request_exceptions,
+)
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import is_local_main_process, zero_first
 from axolotl.utils.trainer import (
@@ -55,27 +57,6 @@ from axolotl.utils.trainer import (
 LOG = logging.getLogger("axolotl")
 
 
-def retry_on_request_exceptions(max_retries=3, delay=1):
-    def decorator(func):
-        @functools.wraps(func)
-        def wrapper(*args, **kwargs):  # pylint: disable=inconsistent-return-statements
-            for attempt in range(max_retries):
-                try:
-                    return func(*args, **kwargs)
-                except (
-                    requests.exceptions.ReadTimeout,
-                    requests.exceptions.ConnectionError,
-                ) as exc:
-                    if attempt < max_retries - 1:
-                        time.sleep(delay)
-                    else:
-                        raise exc
-
-        return wrapper
-
-    return decorator
-
-
 @retry_on_request_exceptions(max_retries=3, delay=5)
 def prepare_dataset(cfg, tokenizer, processor=None):
     prompters = []
@@ -136,8 +117,9 @@ def prepare_dataset(cfg, tokenizer, processor=None):
         # https://discuss.huggingface.co/t/how-to-use-huggingface-trainer-streaming-datasets-without-wrapping-it-with-torchdatas-iterablewrapper/25230
         train_dataset = train_dataset.with_format("torch")
         eval_dataset = None
+        if cfg.dataset_exact_deduplication:
+            LOG.info("Deduplication not available for pretrained datasets")
         return train_dataset, eval_dataset, cfg.max_steps, prompters
-
     if eval_dataset and cfg.sample_packing and cfg.eval_sample_packing is not False:
         total_eval_steps = calculate_total_num_steps(cfg, eval_dataset, update=False)
         if total_eval_steps == 0:
@@ -584,7 +566,8 @@ def load_prepare_datasets(
         )
         train_fingerprint = md5(to_hash_train)
         test_fingerprint = md5(to_hash_test)
-
+        if cfg.dataset_exact_deduplication:
+            _, _, dataset = deduplicate_and_log_datasets(dataset=dataset)
         dataset = dataset.train_test_split(
             test_size=val_set_size,
             shuffle=False,
@@ -596,12 +579,17 @@ def load_prepare_datasets(
         train_dataset = dataset["train"]
         eval_dataset = dataset["test"]
     elif split == "test":
+        if cfg.dataset_exact_deduplication:
+            _, eval_dataset, _ = deduplicate_and_log_datasets(eval_dataset=dataset)
+        else:
+            eval_dataset = dataset
         train_dataset = None
-        eval_dataset = dataset
     else:
-        train_dataset = dataset
+        if cfg.dataset_exact_deduplication:
+            train_dataset, _, _ = deduplicate_and_log_datasets(train_dataset=dataset)
+        else:
+            train_dataset = dataset
         eval_dataset = None
-
     return train_dataset, eval_dataset, prompters
 
 

src/axolotl/utils/data/utils.py

@@ -1,6 +1,55 @@
 """data handling helpers"""
-
+import functools
 import hashlib
+import logging
+import time
+from enum import Enum
+
+import huggingface_hub
+import requests
+from datasets import Dataset
+
+LOG = logging.getLogger("axolotl")
+
+
+class RetryStrategy(Enum):
+    """
+    Enum for retry strategies.
+    """
+
+    CONSTANT = 1
+    LINEAR = 2
+    EXPONENTIAL = 3
+
+
+def retry_on_request_exceptions(
+    max_retries=3, delay=1, retry_strategy: RetryStrategy = RetryStrategy.LINEAR
+):
+    def decorator(func):
+        @functools.wraps(func)
+        def wrapper(*args, **kwargs):  # pylint: disable=inconsistent-return-statements
+            for attempt in range(max_retries):
+                try:
+                    return func(*args, **kwargs)
+                except (
+                    requests.exceptions.ReadTimeout,
+                    requests.exceptions.ConnectionError,
+                    huggingface_hub.errors.HfHubHTTPError,
+                ) as exc:
+                    if attempt < max_retries - 1:
+                        if retry_strategy == RetryStrategy.EXPONENTIAL:
+                            step_delay = delay * 2**attempt
+                        elif retry_strategy == RetryStrategy.LINEAR:
+                            step_delay = delay * (attempt + 1)
+                        else:
+                            step_delay = delay  # Use constant delay.
+                        time.sleep(step_delay)
+                    else:
+                        raise exc
+
+        return wrapper
+
+    return decorator
 
 
 def md5(to_hash: str, encoding: str = "utf-8") -> str:
@@ -8,3 +57,96 @@ def md5(to_hash: str, encoding: str = "utf-8") -> str:
         return hashlib.md5(to_hash.encode(encoding), usedforsecurity=False).hexdigest()
     except TypeError:
         return hashlib.md5(to_hash.encode(encoding)).hexdigest()  # nosec
+
+
+def sha256(to_hash: str, encoding: str = "utf-8") -> str:
+    return hashlib.sha256(to_hash.encode(encoding)).hexdigest()
+
+
+def deduplicate_dataset(
+    dataset: Dataset, seen_hashes: dict[str, list[int]], other_dataset: Dataset = None
+) -> Dataset:
+    unique_indices = []
+
+    for idx, row in enumerate(dataset):
+        row_hash = sha256(str(row))  # Using SHA256 for collision resistance.
+        if row_hash not in seen_hashes:
+            seen_hashes[row_hash] = [idx]
+            unique_indices.append(idx)
+        else:
+            # Check for collision by looking up the original dataset indices
+            original_indices = seen_hashes[row_hash]
+            is_duplicate = False
+            for original_idx in original_indices:
+                if (
+                    not idx == original_idx
+                    and original_idx < len(dataset)
+                    and str(dataset[original_idx]) == str(row)
+                ):
+                    is_duplicate = True
+                    break
+                # Check in the other dataset if provided
+                if other_dataset is not None:
+                    if original_idx < len(other_dataset) and str(
+                        other_dataset[original_idx]
+                    ) == str(row):
+                        is_duplicate = True
+                        break
+            if not is_duplicate:
+                seen_hashes[row_hash].append(idx)
+                unique_indices.append(idx)
+                continue
+    return dataset.select(unique_indices)
+
+
+def deduplicate_and_log_datasets(
+    *,
+    train_dataset: Dataset = None,
+    eval_dataset: Dataset = None,
+    dataset: Dataset = None,
+) -> tuple[Dataset, Dataset, Dataset]:
+    """
+    Deduplicates train, eval, and an optional dataset if provided, logging original and new sizes.
+
+    Returns:
+        tuple: Deduplicated train, eval, and additional datasets.
+    """
+    seen_hashes: dict[str, list[int]] = {}
+
+    # Handle cases where datasets are None
+    if train_dataset is not None:
+        LOG.info(
+            f"Starting deduplication for train dataset. Original size: {len(train_dataset)}"
+        )
+        train_dataset = deduplicate_dataset(
+            dataset=train_dataset, seen_hashes=seen_hashes
+        )
+        LOG.info(
+            f"Deduplication complete for train dataset. New size: {len(train_dataset)}"
+        )
+    else:
+        LOG.info("Train dataset is None. Skipping deduplication.")
+
+    if eval_dataset is not None:
+        LOG.info(
+            f"Starting deduplication for eval dataset. Original size: {len(eval_dataset)}"
+        )
+        eval_dataset = deduplicate_dataset(
+            dataset=eval_dataset, seen_hashes=seen_hashes, other_dataset=train_dataset
+        )
+        LOG.info(
+            f"Deduplication complete for eval dataset. New size: {len(eval_dataset)}"
+        )
+    else:
+        LOG.info("Eval dataset is None. Skipping deduplication.")
+
+    if dataset is not None and (eval_dataset is None and train_dataset is None):
+        LOG.info(
+            f"Starting deduplication for combined dataset. Original size: {len(dataset)}"
+        )
+        dataset = deduplicate_dataset(dataset=dataset, seen_hashes=seen_hashes)
+        LOG.info(
+            f"Deduplication complete for combined dataset. New size: {len(dataset)}"
+        )
+
+    return train_dataset, eval_dataset, dataset

src/axolotl/utils/distributed.py

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

src/axolotl/utils/models.py

@@ -2,10 +2,12 @@
 
 # pylint: disable=too-many-lines
 import gc
+import importlib
 import logging
 import math
 import os
 import types
+from functools import cached_property
 from typing import Any, Dict, Optional, Tuple, Union  # noqa: F401
 
 import addict
@@ -55,7 +57,7 @@ from axolotl.prompt_tokenizers import LLAMA_DEFAULT_EOS_TOKEN
 from axolotl.utils.bench import log_gpu_memory_usage
 from axolotl.utils.chat_templates import get_chat_template_from_config
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.distributed import zero_only
+from axolotl.utils.distributed import get_device_count, get_device_type, zero_only
 from axolotl.utils.gradient_checkpointing import hf_grad_checkpoint_unsloth_wrapper
 from axolotl.utils.lora_embeddings import get_linear_embedding_layers
 from axolotl.utils.model_shard_quant import load_sharded_model, load_sharded_model_quant
@@ -378,12 +380,34 @@ class ModelLoader:
         plugin_manager = PluginManager.get_instance()
         plugin_manager.pre_model_load(self.cfg)
 
+        if self.cfg.fsdp:
+            from axolotl.monkeypatch.trainer_fsdp_optim import (
+                patch_training_loop_for_fsdp,
+            )
+
+            patch_training_loop_for_fsdp()
+        elif self.cfg.deepspeed and self.cfg.gradient_accumulation_steps > 1:
+            from axolotl.monkeypatch.trainer_grad_accum import (
+                patch_training_loop_for_deepspeed_0_16_x,
+            )
+
+            patch_training_loop_for_deepspeed_0_16_x()
+
         if self.cfg.gradient_checkpointing == "unsloth":
             transformers.modeling_utils.checkpoint = hf_grad_checkpoint_unsloth_wrapper
 
         if self.cfg.flash_attention:
             self.patch_attention()
 
+        if self.cfg.model_config_type == "llama":
+            from axolotl.monkeypatch.trainer_grad_accum import (
+                patch_forward_for_ga,
+                patch_training_step_for_ga,
+            )
+
+            patch_forward_for_ga()
+            patch_training_step_for_ga()
+
         if self.cfg.sample_packing and self.cfg.s2_attention:
             raise ValueError(
                 "Received `sample_packing=true` and `s2_attention=true`; however, \
@@ -395,10 +419,14 @@ class ModelLoader:
             and self.cfg.flash_attention
             and self.cfg.sample_packing
         ):
-            has_remote_code = (
-                "auto_map" in self.model_config
-                and "AutoModelForCausalLM" in self.model_config["auto_map"]
-            )
+            if "auto_map" in self.model_config:
+                try:
+                    auto_map_config = self.model_config["auto_map"]
+                except TypeError:
+                    auto_map_config = self.model_config.auto_map
+                has_remote_code = "AutoModelForCausalLM" in auto_map_config
+            else:
+                has_remote_code = False
             if has_remote_code and self.cfg.trust_remote_code is False:
                 # if explicitly set in the YAML, we should prefer that, for example if explicitly disabled
                 has_remote_code = self.cfg.trust_remote_code
@@ -409,7 +437,7 @@ class ModelLoader:
             )
 
             if self.cfg.is_llama_derived_model:
-                self.patch_loss()
+                self.patch_loss_llama()
                 if self.cfg.unsloth_lora_qkv or self.cfg.unsloth_lora_o:
                     from axolotl.monkeypatch.unsloth_ import patch_self_attn_lora
 
@@ -451,27 +479,34 @@ class ModelLoader:
 
                 replace_stablelm_attn_with_flash_attn(self.cfg.base_model)
 
-    def patch_loss(self) -> None:
+    @cached_property
+    def has_flash_attn(self) -> bool:
+        """Check if flash attention is installed"""
+        return importlib.util.find_spec("flash_attn") is not None
+
+    def patch_loss_llama(self) -> None:
         """
         Patch loss functions
         """
-        from axolotl.monkeypatch.llama_attn_hijack_flash import (
-            patch_llama_cross_entropy,
-            patch_llama_rms_norm,
-        )
+        if self.has_flash_attn:
+            from axolotl.monkeypatch.llama_attn_hijack_flash import (
+                patch_fa_llama_cross_entropy,
+                patch_llama_rms_norm,
+            )
 
-        if self.cfg.flash_attn_cross_entropy:
-            patch_llama_cross_entropy()
-        if self.cfg.flash_attn_rms_norm:
+        if self.cfg.flash_attn_cross_entropy and self.has_flash_attn:
+            patch_fa_llama_cross_entropy()
+        elif self.cfg.unsloth_cross_entropy_loss:
+            from axolotl.monkeypatch.unsloth_ import integrate_cross_entropy_loss_patch
+
+            integrate_cross_entropy_loss_patch(model_type="llama")
+
+        if self.cfg.flash_attn_rms_norm and self.has_flash_attn:
             patch_llama_rms_norm()
         elif self.cfg.unsloth_rms_norm:
             from axolotl.monkeypatch.unsloth_ import patch_unsloth_layernorm
 
             patch_unsloth_layernorm()
-        if self.cfg.unsloth_cross_entropy_loss:
-            from axolotl.monkeypatch.unsloth_ import integrate_cross_entropy_loss_patch
-
-            integrate_cross_entropy_loss_patch(model_type="llama")
         if self.cfg.unsloth_lora_qkv or self.cfg.unsloth_lora_o:
             from axolotl.monkeypatch.unsloth_ import patch_self_attn_lora
 
@@ -481,6 +516,7 @@ class ModelLoader:
         """
         Modify all llama derived models in one block
         """
+        self.patch_loss_llama()
 
         if self.cfg.flash_attention:
             from axolotl.monkeypatch.llama_attn_hijack_flash import (
@@ -528,16 +564,6 @@ class ModelLoader:
                 "Shifted-sparse attention not currently implemented without flash attention."
             )
 
-        if self.cfg.unsloth_cross_entropy_loss:
-            from axolotl.monkeypatch.unsloth_ import integrate_cross_entropy_loss_patch
-
-            integrate_cross_entropy_loss_patch(model_type="llama")
-
-        if self.cfg.unsloth_lora_qkv or self.cfg.unsloth_lora_o:
-            from axolotl.monkeypatch.unsloth_ import patch_self_attn_lora
-
-            patch_self_attn_lora()
-
     def set_auto_model_loader(self) -> None:
         """set self.AutoModelLoader
         - default value: AutoModelForCausalLM (set at __init__)
@@ -570,7 +596,8 @@ class ModelLoader:
             )
 
             max_memory = {}
-            for i in range(torch.cuda.device_count()):
+            num_device = get_device_count()
+            for i in range(num_device):
                 max_memory[i] = gpu_memory_limit
             max_memory["cpu"] = "256GiB"  # something sufficiently large to fit anything
 
@@ -595,8 +622,11 @@ class ModelLoader:
         self.model_kwargs["device_map"] = device_map
         self.model_kwargs["torch_dtype"] = self.cfg.torch_dtype
 
-        if torch.backends.mps.is_available():
+        cur_device = get_device_type()
+        if "mps" in str(cur_device):
             self.model_kwargs["device_map"] = "mps:0"
+        elif "npu" in str(cur_device):
+            self.model_kwargs["device_map"] = "npu:0"
 
         # TODO can we put the reference model on it's own gpu? I think we have to move logits around to calculate loss
         # if cfg.rl:
@@ -1050,7 +1080,11 @@ class ModelLoader:
         self.ajust_model_config()
 
         # log device memory usage
-        if hasattr(self.model, "device") and self.model.device.type in ("cuda", "mps"):
+        if hasattr(self.model, "device") and self.model.device.type in (
+            "cuda",
+            "mps",
+            "npu",
+        ):
             log_gpu_memory_usage(LOG, "after model load", self.model.device)
 
         # make sure these are fp32 per Ramesh et al. (2021)
@@ -1076,14 +1110,17 @@ class ModelLoader:
 
         self.prepare_model(qlora_fsdp)
 
-        # LlamaRMSNorm layers are in fp32 after kbit_training or full finetune, so we need to
-        # convert them back to fp16/bf16 for flash-attn compatibility.
-        if (needs_fa2_dtype or self.cfg.flash_attention) and not qlora_fsdp:
-            LOG.info(
-                "converting modules to %s for flash attention", self.cfg.torch_dtype
-            )
+        should_convert = (
+            # LlamaRMSNorm layers are in fp32 after kbit_training or full finetune, so we need to
+            # convert them back to fp16/bf16 for flash-attn compatibility.
+            ((needs_fa2_dtype or self.cfg.flash_attention) and not qlora_fsdp)
+            or self.cfg.cut_cross_entropy  # Cut cross entropy requires embedding layers to be in fp16/bf16 for backward pass
+        )
+
+        if should_convert:
+            LOG.info("Converting modules to %s", self.cfg.torch_dtype)
             self.convert_embedding_modules_dtype(
-                embedding_modules,
+                embedding_modules=embedding_modules,
                 dist_dtype=self.cfg.torch_dtype,
                 before_kbit_train_or_finetune=False,
             )
@@ -1118,9 +1155,9 @@ class ModelLoader:
             and not skip_move_to_device
         ):
             # TODO revaldate this conditional
-            self.model.to(f"cuda:{self.cfg.local_rank}")
+            self.model.to(f"{str(get_device_type())}:{self.cfg.local_rank}")
 
-        if torch.cuda.device_count() > 1 and int(os.getenv("WORLD_SIZE", "1")) == 1:
+        if get_device_count() > 1 and int(os.getenv("WORLD_SIZE", "1")) == 1:
             setattr(self.model, "is_parallelizable", True)
             setattr(self.model, "model_parallel", True)
 

src/axolotl/utils/optimizers/adopt.py

@@ -6,21 +6,29 @@ Taniguchi, Shohei and Harada, Keno and Minegishi, Gouki and Oshima, Yuta and Jeo
 """
 # mypy: ignore-errors
 # pylint: skip-file
+# flake8: noqa
 # mypy: allow-untyped-decorators
 # mypy: allow-untyped-defs
-from typing import List, Optional, Tuple, Union, cast
+from typing import Callable, List, Optional, Tuple, Union, cast
 
 import torch
 from torch import Tensor
-from torch.optim.optimizer import (
+from torch.optim.optimizer import (  # DeviceDict,; _capturable_doc,; _differentiable_doc,; _foreach_doc,; _fused_doc,; _maximize_doc,; _stack_if_compiling,
+    DeviceDict,
     Optimizer,
     ParamsT,
+    _capturable_doc,
     _default_to_fused_or_foreach,
     _device_dtype_check_for_fused,
+    _differentiable_doc,
     _disable_dynamo_if_unsupported,
+    _foreach_doc,
+    _fused_doc,
     _get_capturable_supported_devices,
     _get_scalar_dtype,
     _get_value,
+    _maximize_doc,
+    _stack_if_compiling,
     _use_grad_for_differentiable,
     _view_as_real,
 )
@@ -35,8 +43,9 @@ class ADOPT(Optimizer):
         lr: Union[float, Tensor] = 1e-3,
         betas: Tuple[float, float] = (0.9, 0.9999),
         eps: float = 1e-6,
+        clip_lambda: Optional[Callable[[int], float]] = lambda step: step**0.25,
         weight_decay: float = 0.0,
-        decoupled: bool = False,
+        decouple: bool = False,
         *,
         foreach: Optional[bool] = None,
         maximize: bool = False,
@@ -62,12 +71,14 @@ class ADOPT(Optimizer):
         if not 0.0 <= weight_decay:
             raise ValueError(f"Invalid weight_decay value: {weight_decay}")
 
+        self.clip_lambda = clip_lambda
+
         defaults = dict(
             lr=lr,
             betas=betas,
             eps=eps,
             weight_decay=weight_decay,
-            decoupled=decoupled,
+            decouple=decouple,
             maximize=maximize,
             foreach=foreach,
             capturable=capturable,
@@ -219,8 +230,9 @@ class ADOPT(Optimizer):
                 beta1=beta1,
                 beta2=beta2,
                 lr=group["lr"],
+                clip_lambda=self.clip_lambda,
                 weight_decay=group["weight_decay"],
-                decoupled=group["decoupled"],
+                decouple=group["decouple"],
                 eps=group["eps"],
                 maximize=group["maximize"],
                 foreach=group["foreach"],
@@ -247,8 +259,9 @@ def _single_tensor_adopt(
     beta1: float,
     beta2: float,
     lr: Union[float, Tensor],
+    clip_lambda: Optional[Callable[[int], float]],
     weight_decay: float,
-    decoupled: bool,
+    decouple: bool,
     eps: float,
     maximize: bool,
     capturable: bool,
@@ -276,14 +289,10 @@ def _single_tensor_adopt(
                 and param.device.type in capturable_supported_devices
             ), f"If capturable=True, params and state_steps must be on supported devices: {capturable_supported_devices}."
 
-        # update step
-        step_t += 1
+        step = step_t if capturable or differentiable else _get_value(step_t)
 
-        if weight_decay != 0:
-            if decoupled:
-                param.add_(param, alpha=-lr * weight_decay)
-            else:
-                grad = grad.add(param, alpha=weight_decay)
+        if weight_decay != 0 and not decouple:
+            grad = grad.add(param, alpha=weight_decay)
 
         if torch.is_complex(param):
             grad = torch.view_as_real(grad)
@@ -293,20 +302,29 @@ def _single_tensor_adopt(
                 exp_avg_sq = torch.view_as_real(exp_avg_sq)
             param = torch.view_as_real(param)
 
-        step = step_t if capturable or differentiable else _get_value(step_t)
-        if step == 1:
+        if step == 0:
             exp_avg_sq.addcmul_(grad, grad.conj())
+            # update step
+            step_t += 1
             continue
 
+        if weight_decay != 0 and decouple:
+            param.add_(param, alpha=-lr * weight_decay)
+
         denom = torch.clamp(exp_avg_sq.sqrt(), eps)
-        if step == 2:
-            exp_avg.addcdiv_(grad, denom)
-        else:
-            exp_avg.mul_(beta1).addcdiv_(grad, denom, value=1 - beta1)
+        normed_grad = grad.div(denom)
+        if clip_lambda is not None:
+            clip = clip_lambda(step)
+            normed_grad.clamp_(-clip, clip)
+
+        exp_avg.lerp_(normed_grad, 1 - beta1)
 
         param.add_(exp_avg, alpha=-lr)
         exp_avg_sq.mul_(beta2).addcmul_(grad, grad.conj(), value=1 - beta2)
 
+        # update step
+        step_t += 1
+
 
 def _multi_tensor_adopt(
     params: List[Tensor],
@@ -321,8 +339,9 @@ def _multi_tensor_adopt(
     beta1: float,
     beta2: float,
     lr: Union[float, Tensor],
+    clip_lambda: Optional[Callable[[int], float]],
     weight_decay: float,
-    decoupled: bool,
+    decouple: bool,
     eps: float,
     maximize: bool,
     capturable: bool,
@@ -376,6 +395,51 @@ def _multi_tensor_adopt(
         if maximize:
             device_grads = torch._foreach_neg(device_grads)  # type: ignore[assignment]
 
+        if weight_decay != 0 and not decouple:
+            # Re-use the intermediate memory (device_grads) already allocated for maximize
+            if maximize:
+                torch._foreach_add_(device_grads, device_params, alpha=weight_decay)
+            else:
+                device_grads = torch._foreach_add(  # type: ignore[assignment]
+                    device_grads, device_params, alpha=weight_decay
+                )
+
+        if device_state_steps[0] == 0:
+            torch._foreach_addcmul_(device_exp_avg_sqs, device_grads, device_grads)
+
+            # Update steps
+            # If steps are on CPU, foreach will fall back to the slow path, which is a for-loop calling t.add(1) over
+            # and over. 1 will then be wrapped into a Tensor over and over again, which is slower than if we just
+            # wrapped it once now. The alpha is required to assure we go to the right overload.
+            if not torch._utils.is_compiling() and device_state_steps[0].is_cpu:
+                torch._foreach_add_(
+                    device_state_steps, torch.tensor(1.0, device="cpu"), alpha=1.0
+                )
+            else:
+                torch._foreach_add_(device_state_steps, 1)
+
+            continue
+
+        if weight_decay != 0 and decouple:
+            torch._foreach_add_(device_params, device_params, alpha=-lr * weight_decay)
+
+        exp_avg_sq_sqrt = torch._foreach_sqrt(device_exp_avg_sqs)
+        torch._foreach_maximum_(exp_avg_sq_sqrt, eps)
+
+        normed_grad = torch._foreach_div(device_grads, exp_avg_sq_sqrt)
+        if clip_lambda is not None:
+            clip = clip_lambda(device_state_steps[0])
+            torch._foreach_maximum_(normed_grad, -clip)
+            torch._foreach_minimum_(normed_grad, clip)
+
+        torch._foreach_lerp_(device_exp_avgs, normed_grad, 1 - beta1)
+
+        torch._foreach_add_(device_params, device_exp_avgs, alpha=-lr)
+        torch._foreach_mul_(device_exp_avg_sqs, beta2)
+        torch._foreach_addcmul_(
+            device_exp_avg_sqs, device_grads, device_grads, value=1 - beta2
+        )
+
         # Update steps
         # If steps are on CPU, foreach will fall back to the slow path, which is a for-loop calling t.add(1) over
         # and over. 1 will then be wrapped into a Tensor over and over again, which is slower than if we just
@@ -387,41 +451,6 @@ def _multi_tensor_adopt(
         else:
             torch._foreach_add_(device_state_steps, 1)
 
-        if weight_decay != 0:
-            if decoupled:
-                torch._foreach_add_(
-                    device_params, device_params, alpha=-lr * weight_decay
-                )
-            else:
-                # Re-use the intermediate memory (device_grads) already allocated for maximize
-                if maximize:
-                    torch._foreach_add_(device_grads, device_params, alpha=weight_decay)
-                else:
-                    device_grads = torch._foreach_add(  # type: ignore[assignment]
-                        device_grads, device_params, alpha=weight_decay
-                    )
-
-        if device_state_steps[0] == 1:
-            torch._foreach_addcmul_(device_exp_avg_sqs, device_grads, device_grads)
-            continue
-
-        exp_avg_sq_sqrt = torch._foreach_sqrt(device_exp_avg_sqs)
-        exp_avg_sq_sqrt = torch._foreach_maximum(exp_avg_sq_sqrt, eps)
-
-        if device_state_steps[0] == 2:
-            torch._foreach_addcdiv_(device_exp_avgs, device_grads, exp_avg_sq_sqrt)
-        else:
-            torch._foreach_mul_(device_exp_avgs, beta1)
-            torch._foreach_addcdiv_(
-                device_exp_avgs, device_grads, exp_avg_sq_sqrt, value=1 - beta1
-            )
-
-        torch._foreach_add_(device_params, device_exp_avgs, alpha=-lr)
-        torch._foreach_mul_(device_exp_avg_sqs, beta2)
-        torch._foreach_addcmul_(
-            device_exp_avg_sqs, device_grads, device_grads, value=1 - beta2
-        )
-
 
 @_disable_dynamo_if_unsupported(single_tensor_fn=_single_tensor_adopt)
 def adopt(
@@ -443,8 +472,9 @@ def adopt(
     beta1: float,
     beta2: float,
     lr: Union[float, Tensor],
+    clip_lambda: Optional[Callable[[int], float]],
     weight_decay: float,
-    decoupled: bool,
+    decouple: bool,
     eps: float,
     maximize: bool,
 ):
@@ -497,8 +527,9 @@ def adopt(
         beta1=beta1,
         beta2=beta2,
         lr=lr,
+        clip_lambda=clip_lambda,
         weight_decay=weight_decay,
-        decoupled=decoupled,
+        decouple=decouple,
         eps=eps,
         maximize=maximize,
         capturable=capturable,

src/setuptools_axolotl_dynamic_dependencies.py

@@ -0,0 +1,104 @@
+"""
+dynamic requirements for axolotl
+"""
+import platform
+import re
+from importlib.metadata import PackageNotFoundError, version
+
+from setuptools.command.build_py import build_py as _build_py
+
+
+# pylint: disable=duplicate-code
+def parse_requirements():
+    _install_requires = []
+    _dependency_links = []
+    with open("./requirements.txt", encoding="utf-8") as requirements_file:
+        lines = [r.strip() for r in requirements_file.readlines()]
+        for line in lines:
+            is_extras = (
+                "flash-attn" in line
+                or "flash-attention" in line
+                or "deepspeed" in line
+                or "mamba-ssm" in line
+                or "lion-pytorch" in line
+            )
+            if line.startswith("--extra-index-url"):
+                # Handle custom index URLs
+                _, url = line.split()
+                _dependency_links.append(url)
+            elif not is_extras and line and line[0] != "#":
+                # Handle standard packages
+                _install_requires.append(line)
+
+    try:
+        xformers_version = [req for req in _install_requires if "xformers" in req][0]
+        torchao_version = [req for req in _install_requires if "torchao" in req][0]
+        autoawq_version = [req for req in _install_requires if "autoawq" in req][0]
+
+        if "Darwin" in platform.system():
+            # don't install xformers on MacOS
+            _install_requires.pop(_install_requires.index(xformers_version))
+        else:
+            # detect the version of torch already installed
+            # and set it so dependencies don't clobber the torch version
+            try:
+                torch_version = version("torch")
+            except PackageNotFoundError:
+                torch_version = "2.5.1"
+            _install_requires.append(f"torch=={torch_version}")
+
+            version_match = re.match(r"^(\d+)\.(\d+)(?:\.(\d+))?", torch_version)
+            if version_match:
+                major, minor, patch = version_match.groups()
+                major, minor = int(major), int(minor)
+                patch = (
+                    int(patch) if patch is not None else 0
+                )  # Default patch to 0 if not present
+            else:
+                raise ValueError("Invalid version format")
+
+            if (major, minor) >= (2, 5):
+                _install_requires.pop(_install_requires.index(xformers_version))
+                if patch == 0:
+                    _install_requires.append("xformers==0.0.28.post2")
+                else:
+                    _install_requires.append("xformers==0.0.28.post3")
+                _install_requires.pop(_install_requires.index(autoawq_version))
+            elif (major, minor) >= (2, 4):
+                if patch == 0:
+                    _install_requires.pop(_install_requires.index(xformers_version))
+                    _install_requires.append("xformers>=0.0.27")
+                else:
+                    _install_requires.pop(_install_requires.index(xformers_version))
+                    _install_requires.append("xformers==0.0.28.post1")
+            elif (major, minor) >= (2, 3):
+                _install_requires.pop(_install_requires.index(torchao_version))
+                if patch == 0:
+                    _install_requires.pop(_install_requires.index(xformers_version))
+                    _install_requires.append("xformers>=0.0.26.post1")
+                else:
+                    _install_requires.pop(_install_requires.index(xformers_version))
+                    _install_requires.append("xformers>=0.0.27")
+            elif (major, minor) >= (2, 2):
+                _install_requires.pop(_install_requires.index(torchao_version))
+                _install_requires.pop(_install_requires.index(xformers_version))
+                _install_requires.append("xformers>=0.0.25.post1")
+            else:
+                _install_requires.pop(_install_requires.index(torchao_version))
+                _install_requires.pop(_install_requires.index(xformers_version))
+                _install_requires.append("xformers>=0.0.23.post1")
+
+    except PackageNotFoundError:
+        pass
+    return _install_requires, _dependency_links
+
+
+class BuildPyCommand(_build_py):
+    """
+    custom build_py command to parse dynamic requirements
+    """
+
+    def finalize_options(self):
+        super().finalize_options()
+        install_requires, _ = parse_requirements()
+        self.distribution.install_requires = install_requires

tests/cli/conftest.py

@@ -0,0 +1,36 @@
+"""Shared pytest fixtures for cli module."""
+import pytest
+from click.testing import CliRunner
+
+VALID_TEST_CONFIG = """
+base_model: HuggingFaceTB/SmolLM2-135M
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+sequence_len: 2048
+max_steps: 1
+micro_batch_size: 1
+gradient_accumulation_steps: 1
+learning_rate: 1e-3
+special_tokens:
+  pad_token: <|endoftext|>
+"""
+
+
+@pytest.fixture
+def cli_runner():
+    return CliRunner()
+
+
+@pytest.fixture
+def valid_test_config():
+    return VALID_TEST_CONFIG
+
+
+@pytest.fixture
+def config_path(tmp_path):
+    """Creates a temporary config file"""
+    path = tmp_path / "config.yml"
+    path.write_text(VALID_TEST_CONFIG)
+
+    return path

tests/cli/test_cli_fetch.py

@@ -0,0 +1,38 @@
+"""pytest tests for axolotl CLI fetch command."""
+from unittest.mock import patch
+
+from axolotl.cli.main import fetch
+
+
+def test_fetch_cli_examples(cli_runner):
+    """Test fetch command with examples directory"""
+    with patch("axolotl.cli.main.fetch_from_github") as mock_fetch:
+        result = cli_runner.invoke(fetch, ["examples"])
+
+        assert result.exit_code == 0
+        mock_fetch.assert_called_once_with("examples/", None)
+
+
+def test_fetch_cli_deepspeed(cli_runner):
+    """Test fetch command with deepspeed_configs directory"""
+    with patch("axolotl.cli.main.fetch_from_github") as mock_fetch:
+        result = cli_runner.invoke(fetch, ["deepspeed_configs"])
+
+        assert result.exit_code == 0
+        mock_fetch.assert_called_once_with("deepspeed_configs/", None)
+
+
+def test_fetch_cli_with_dest(cli_runner, tmp_path):
+    """Test fetch command with custom destination"""
+    with patch("axolotl.cli.main.fetch_from_github") as mock_fetch:
+        custom_dir = tmp_path / "tmp_examples"
+        result = cli_runner.invoke(fetch, ["examples", "--dest", str(custom_dir)])
+
+        assert result.exit_code == 0
+        mock_fetch.assert_called_once_with("examples/", str(custom_dir))
+
+
+def test_fetch_cli_invalid_directory(cli_runner):
+    """Test fetch command with invalid directory choice"""
+    result = cli_runner.invoke(fetch, ["invalid"])
+    assert result.exit_code != 0

tests/cli/test_cli_inference.py

@@ -0,0 +1,30 @@
+"""pytest tests for axolotl CLI inference command."""
+from unittest.mock import patch
+
+from axolotl.cli.main import cli
+
+
+def test_inference_basic(cli_runner, config_path):
+    """Test basic inference"""
+    with patch("axolotl.cli.inference.do_inference") as mock:
+        result = cli_runner.invoke(
+            cli,
+            ["inference", str(config_path), "--no-accelerate"],
+            catch_exceptions=False,
+        )
+
+        assert mock.called
+        assert result.exit_code == 0
+
+
+def test_inference_gradio(cli_runner, config_path):
+    """Test basic inference (gradio path)"""
+    with patch("axolotl.cli.inference.do_inference_gradio") as mock:
+        result = cli_runner.invoke(
+            cli,
+            ["inference", str(config_path), "--no-accelerate", "--gradio"],
+            catch_exceptions=False,
+        )
+
+        assert mock.called
+        assert result.exit_code == 0

tests/cli/test_cli_interface.py

@@ -0,0 +1,47 @@
+"""General pytest tests for axolotl.cli.main interface."""
+from axolotl.cli.main import build_command, cli
+
+
+def test_build_command():
+    """Test converting dict of options to CLI arguments"""
+    base_cmd = ["accelerate", "launch"]
+    options = {
+        "learning_rate": 1e-4,
+        "batch_size": 8,
+        "debug": True,
+        "use_fp16": False,
+        "null_value": None,
+    }
+
+    result = build_command(base_cmd, options)
+    assert result == [
+        "accelerate",
+        "launch",
+        "--learning-rate",
+        "0.0001",
+        "--batch-size",
+        "8",
+        "--debug",
+    ]
+
+
+def test_invalid_command_options(cli_runner):
+    """Test handling of invalid command options"""
+    result = cli_runner.invoke(
+        cli,
+        [
+            "train",
+            "config.yml",
+            "--invalid-option",
+            "value",
+        ],
+    )
+    assert result.exit_code != 0
+    assert "No such option" in result.output
+
+
+def test_required_config_argument(cli_runner):
+    """Test commands fail properly when config argument is missing"""
+    result = cli_runner.invoke(cli, ["train"])
+    assert result.exit_code != 0
+    assert "Missing argument 'CONFIG'" in result.output

tests/cli/test_cli_merge_lora.py

@@ -0,0 +1,56 @@
+"""pytest tests for axolotl CLI merge_lora command."""
+from unittest.mock import patch
+
+from axolotl.cli.main import cli
+
+
+def test_merge_lora_basic(cli_runner, config_path):
+    """Test basic merge_lora command"""
+    with patch("axolotl.cli.merge_lora.do_cli") as mock_do_cli:
+        result = cli_runner.invoke(cli, ["merge-lora", str(config_path)])
+        assert result.exit_code == 0
+
+        mock_do_cli.assert_called_once()
+        assert mock_do_cli.call_args.kwargs["config"] == str(config_path)
+
+
+def test_merge_lora_with_dirs(cli_runner, config_path, tmp_path):
+    """Test merge_lora with custom lora and output directories"""
+    lora_dir = tmp_path / "lora"
+    output_dir = tmp_path / "output"
+    lora_dir.mkdir()
+
+    with patch("axolotl.cli.merge_lora.do_cli") as mock_do_cli:
+        result = cli_runner.invoke(
+            cli,
+            [
+                "merge-lora",
+                str(config_path),
+                "--lora-model-dir",
+                str(lora_dir),
+                "--output-dir",
+                str(output_dir),
+            ],
+        )
+        assert result.exit_code == 0
+
+        mock_do_cli.assert_called_once()
+        assert mock_do_cli.call_args.kwargs["config"] == str(config_path)
+        assert mock_do_cli.call_args.kwargs["lora_model_dir"] == str(lora_dir)
+        assert mock_do_cli.call_args.kwargs["output_dir"] == str(output_dir)
+
+
+def test_merge_lora_nonexistent_config(cli_runner, tmp_path):
+    """Test merge_lora with nonexistent config"""
+    config_path = tmp_path / "nonexistent.yml"
+    result = cli_runner.invoke(cli, ["merge-lora", str(config_path)])
+    assert result.exit_code != 0
+
+
+def test_merge_lora_nonexistent_lora_dir(cli_runner, config_path, tmp_path):
+    """Test merge_lora with nonexistent lora directory"""
+    lora_dir = tmp_path / "nonexistent"
+    result = cli_runner.invoke(
+        cli, ["merge-lora", str(config_path), "--lora-model-dir", str(lora_dir)]
+    )
+    assert result.exit_code != 0

tests/cli/test_cli_merge_sharded_fsdp_weights.py

@@ -0,0 +1,60 @@
+"""pytest tests for axolotl CLI merge_sharded_fsdp_weights command."""
+# pylint: disable=duplicate-code
+from unittest.mock import patch
+
+from axolotl.cli.main import cli
+
+
+def test_merge_sharded_fsdp_weights_no_accelerate(cli_runner, config_path):
+    """Test merge_sharded_fsdp_weights command without accelerate"""
+    with patch("axolotl.cli.merge_sharded_fsdp_weights.do_cli") as mock:
+        result = cli_runner.invoke(
+            cli, ["merge-sharded-fsdp-weights", str(config_path), "--no-accelerate"]
+        )
+
+        assert mock.called
+        assert mock.call_args.kwargs["config"] == str(config_path)
+        assert result.exit_code == 0
+
+
+def test_merge_sharded_fsdp_weights_with_model_dir(cli_runner, config_path, tmp_path):
+    """Test merge_sharded_fsdp_weights command with model_dir option"""
+    model_dir = tmp_path / "model"
+    model_dir.mkdir()
+
+    with patch("axolotl.cli.merge_sharded_fsdp_weights.do_cli") as mock:
+        result = cli_runner.invoke(
+            cli,
+            [
+                "merge-sharded-fsdp-weights",
+                str(config_path),
+                "--no-accelerate",
+                "--model-dir",
+                str(model_dir),
+            ],
+        )
+
+        assert mock.called
+        assert mock.call_args.kwargs["config"] == str(config_path)
+        assert mock.call_args.kwargs["model_dir"] == str(model_dir)
+        assert result.exit_code == 0
+
+
+def test_merge_sharded_fsdp_weights_with_save_path(cli_runner, config_path):
+    """Test merge_sharded_fsdp_weights command with save_path option"""
+    with patch("axolotl.cli.merge_sharded_fsdp_weights.do_cli") as mock:
+        result = cli_runner.invoke(
+            cli,
+            [
+                "merge-sharded-fsdp-weights",
+                str(config_path),
+                "--no-accelerate",
+                "--save-path",
+                "/path/to/save",
+            ],
+        )
+
+        assert mock.called
+        assert mock.call_args.kwargs["config"] == str(config_path)
+        assert mock.call_args.kwargs["save_path"] == "/path/to/save"
+        assert result.exit_code == 0

tests/cli/test_cli_preprocess.py

@@ -0,0 +1,71 @@
+"""pytest tests for axolotl CLI preprocess command."""
+import shutil
+from pathlib import Path
+from unittest.mock import patch
+
+import pytest
+
+from axolotl.cli.main import cli
+
+
+@pytest.fixture(autouse=True)
+def cleanup_last_run_prepared():
+    yield
+
+    if Path("last_run_prepared").exists():
+        shutil.rmtree("last_run_prepared")
+
+
+def test_preprocess_config_not_found(cli_runner):
+    """Test preprocess fails when config not found"""
+    result = cli_runner.invoke(cli, ["preprocess", "nonexistent.yml"])
+    assert result.exit_code != 0
+
+
+def test_preprocess_basic(cli_runner, config_path):
+    """Test basic preprocessing with minimal config"""
+    with patch("axolotl.cli.preprocess.do_cli") as mock_do_cli:
+        result = cli_runner.invoke(cli, ["preprocess", str(config_path)])
+        assert result.exit_code == 0
+
+        mock_do_cli.assert_called_once()
+        assert mock_do_cli.call_args.kwargs["config"] == str(config_path)
+        assert mock_do_cli.call_args.kwargs["download"] is True
+
+
+def test_preprocess_without_download(cli_runner, config_path):
+    """Test preprocessing without model download"""
+    with patch("axolotl.cli.preprocess.do_cli") as mock_do_cli:
+        result = cli_runner.invoke(
+            cli, ["preprocess", str(config_path), "--no-download"]
+        )
+        assert result.exit_code == 0
+
+        mock_do_cli.assert_called_once()
+        assert mock_do_cli.call_args.kwargs["config"] == str(config_path)
+        assert mock_do_cli.call_args.kwargs["download"] is False
+
+
+def test_preprocess_custom_path(cli_runner, tmp_path, valid_test_config):
+    """Test preprocessing with custom dataset path"""
+    config_path = tmp_path / "config.yml"
+    custom_path = tmp_path / "custom_prepared"
+    config_path.write_text(valid_test_config)
+
+    with patch("axolotl.cli.preprocess.do_cli") as mock_do_cli:
+        result = cli_runner.invoke(
+            cli,
+            [
+                "preprocess",
+                str(config_path),
+                "--dataset-prepared-path",
+                str(custom_path.absolute()),
+            ],
+        )
+        assert result.exit_code == 0
+
+        mock_do_cli.assert_called_once()
+        assert mock_do_cli.call_args.kwargs["config"] == str(config_path)
+        assert mock_do_cli.call_args.kwargs["dataset_prepared_path"] == str(
+            custom_path.absolute()
+        )

tests/cli/test_cli_shard.py

@@ -0,0 +1,76 @@
+"""pytest tests for axolotl CLI shard command."""
+# pylint: disable=duplicate-code
+from unittest.mock import patch
+
+from axolotl.cli.main import cli
+
+
+def test_shard_with_accelerate(cli_runner, config_path):
+    """Test shard command with accelerate"""
+    with patch("subprocess.run") as mock:
+        result = cli_runner.invoke(cli, ["shard", str(config_path), "--accelerate"])
+
+        assert mock.called
+        assert mock.call_args.args[0] == [
+            "accelerate",
+            "launch",
+            "-m",
+            "axolotl.cli.shard",
+            str(config_path),
+            "--debug-num-examples",
+            "0",
+        ]
+        assert mock.call_args.kwargs == {"check": True}
+        assert result.exit_code == 0
+
+
+def test_shard_no_accelerate(cli_runner, config_path):
+    """Test shard command without accelerate"""
+    with patch("axolotl.cli.shard.do_cli") as mock:
+        result = cli_runner.invoke(cli, ["shard", str(config_path), "--no-accelerate"])
+
+        assert mock.called
+        assert result.exit_code == 0
+
+
+def test_shard_with_model_dir(cli_runner, config_path, tmp_path):
+    """Test shard command with model_dir option"""
+    model_dir = tmp_path / "model"
+    model_dir.mkdir()
+
+    with patch("axolotl.cli.shard.do_cli") as mock:
+        result = cli_runner.invoke(
+            cli,
+            [
+                "shard",
+                str(config_path),
+                "--no-accelerate",
+                "--model-dir",
+                str(model_dir),
+            ],
+            catch_exceptions=False,
+        )
+
+        assert mock.called
+        assert mock.call_args.kwargs["config"] == str(config_path)
+        assert mock.call_args.kwargs["model_dir"] == str(model_dir)
+        assert result.exit_code == 0
+
+
+def test_shard_with_save_dir(cli_runner, config_path):
+    with patch("axolotl.cli.shard.do_cli") as mock:
+        result = cli_runner.invoke(
+            cli,
+            [
+                "shard",
+                str(config_path),
+                "--no-accelerate",
+                "--save-dir",
+                "/path/to/save",
+            ],
+        )
+
+        assert mock.called
+        assert mock.call_args.kwargs["config"] == str(config_path)
+        assert mock.call_args.kwargs["save_dir"] == "/path/to/save"
+        assert result.exit_code == 0

tests/cli/test_cli_train.py

@@ -0,0 +1,98 @@
+"""pytest tests for axolotl CLI train command."""
+from unittest.mock import MagicMock, patch
+
+from axolotl.cli.main import cli
+
+
+def test_train_cli_validation(cli_runner):
+    """Test CLI validation"""
+    # Test missing config file
+    result = cli_runner.invoke(cli, ["train", "--no-accelerate"])
+    assert result.exit_code != 0
+
+    # Test non-existent config file
+    result = cli_runner.invoke(cli, ["train", "nonexistent.yml", "--no-accelerate"])
+    assert result.exit_code != 0
+    assert "Error: Invalid value for 'CONFIG'" in result.output
+
+
+def test_train_basic_execution(cli_runner, tmp_path, valid_test_config):
+    """Test basic successful execution"""
+    config_path = tmp_path / "config.yml"
+    config_path.write_text(valid_test_config)
+
+    with patch("subprocess.run") as mock:
+        result = cli_runner.invoke(cli, ["train", str(config_path)])
+
+        assert mock.called
+        assert mock.call_args.args[0] == [
+            "accelerate",
+            "launch",
+            "-m",
+            "axolotl.cli.train",
+            str(config_path),
+            "--debug-num-examples",
+            "0",
+        ]
+        assert mock.call_args.kwargs == {"check": True}
+        assert result.exit_code == 0
+
+
+def test_train_basic_execution_no_accelerate(cli_runner, tmp_path, valid_test_config):
+    """Test basic successful execution"""
+    config_path = tmp_path / "config.yml"
+    config_path.write_text(valid_test_config)
+
+    with patch("axolotl.cli.train.train") as mock_train:
+        mock_train.return_value = (MagicMock(), MagicMock())
+
+        result = cli_runner.invoke(
+            cli,
+            [
+                "train",
+                str(config_path),
+                "--learning-rate",
+                "1e-4",
+                "--micro-batch-size",
+                "2",
+                "--no-accelerate",
+            ],
+            catch_exceptions=False,
+        )
+
+        assert result.exit_code == 0
+        mock_train.assert_called_once()
+
+
+def test_train_cli_overrides(cli_runner, tmp_path, valid_test_config):
+    """Test CLI arguments properly override config values"""
+    config_path = tmp_path / "config.yml"
+    output_dir = tmp_path / "model-out"
+
+    test_config = valid_test_config.replace(
+        "output_dir: model-out", f"output_dir: {output_dir}"
+    )
+    config_path.write_text(test_config)
+
+    with patch("axolotl.cli.train.train") as mock_train:
+        mock_train.return_value = (MagicMock(), MagicMock())
+
+        result = cli_runner.invoke(
+            cli,
+            [
+                "train",
+                str(config_path),
+                "--learning-rate",
+                "1e-4",
+                "--micro-batch-size",
+                "2",
+                "--no-accelerate",
+            ],
+            catch_exceptions=False,
+        )
+
+        assert result.exit_code == 0
+        mock_train.assert_called_once()
+        cfg = mock_train.call_args[1]["cfg"]
+        assert cfg["learning_rate"] == 1e-4
+        assert cfg["micro_batch_size"] == 2

tests/cli/test_cli_version.py

@@ -0,0 +1,10 @@
+"""pytest tests for axolotl CLI --version"""
+from axolotl.cli.main import cli
+
+
+def test_print_version(cli_runner):
+    """Test that version is printed when --version is used."""
+
+    result = cli_runner.invoke(cli, ["--version"])
+    assert result.exit_code == 0
+    assert "axolotl, version " in result.output

tests/cli/test_utils.py

@@ -0,0 +1,72 @@
+"""pytest tests for axolotl CLI utils."""
+# pylint: disable=redefined-outer-name
+import json
+from unittest.mock import Mock, patch
+
+import click
+import pytest
+import requests
+
+from axolotl.cli.utils import fetch_from_github
+
+# Sample GitHub API response
+MOCK_TREE_RESPONSE = {
+    "tree": [
+        {"path": "examples/config1.yml", "type": "blob", "sha": "abc123"},
+        {"path": "examples/config2.yml", "type": "blob", "sha": "def456"},
+        {"path": "other/file.txt", "type": "blob", "sha": "xyz789"},
+    ]
+}
+
+
+@pytest.fixture
+def mock_responses():
+    """Mock responses for API and file downloads"""
+
+    def mock_get(url, timeout=None):  # pylint: disable=unused-argument
+        response = Mock()
+        if "api.github.com" in url:
+            response.text = json.dumps(MOCK_TREE_RESPONSE)
+        else:
+            response.content = b"file content"
+        return response
+
+    return mock_get
+
+
+def test_fetch_from_github_new_files(tmp_path, mock_responses):
+    """Test fetching new files"""
+    with patch("requests.get", mock_responses):
+        fetch_from_github("examples/", tmp_path)
+
+        # Verify files were created
+        assert (tmp_path / "config1.yml").exists()
+        assert (tmp_path / "config2.yml").exists()
+        assert not (tmp_path / "file.txt").exists()
+
+
+def test_fetch_from_github_unchanged_files(tmp_path, mock_responses):
+    """Test handling of unchanged files"""
+    # Create existing file with matching SHA
+    existing_file = tmp_path / "config1.yml"
+    existing_file.write_bytes(b"file content")
+
+    with patch("requests.get", mock_responses):
+        fetch_from_github("examples/", tmp_path)
+
+        # File should not be downloaded again
+        assert existing_file.read_bytes() == b"file content"
+
+
+def test_fetch_from_github_invalid_prefix(mock_responses):
+    """Test error handling for invalid directory prefix"""
+    with patch("requests.get", mock_responses):
+        with pytest.raises(click.ClickException):
+            fetch_from_github("nonexistent/", None)
+
+
+def test_fetch_from_github_network_error():
+    """Test handling of network errors"""
+    with patch("requests.get", side_effect=requests.RequestException):
+        with pytest.raises(requests.RequestException):
+            fetch_from_github("examples/", None)

tests/conftest.py

@@ -0,0 +1,171 @@
+"""
+shared pytest fixtures
+"""
+import functools
+import importlib
+import shutil
+import sys
+import tempfile
+import time
+
+import pytest
+import requests
+from huggingface_hub import snapshot_download
+
+
+def retry_on_request_exceptions(max_retries=3, delay=1):
+    # pylint: disable=duplicate-code
+    def decorator(func):
+        @functools.wraps(func)
+        def wrapper(*args, **kwargs):  # pylint: disable=inconsistent-return-statements
+            for attempt in range(max_retries):
+                try:
+                    return func(*args, **kwargs)
+                except (
+                    requests.exceptions.ReadTimeout,
+                    requests.exceptions.ConnectionError,
+                ) as exc:
+                    if attempt < max_retries - 1:
+                        time.sleep(delay)
+                    else:
+                        raise exc
+
+        return wrapper
+
+    return decorator
+
+
+@retry_on_request_exceptions(max_retries=3, delay=5)
+def snapshot_download_w_retry(*args, **kwargs):
+    return snapshot_download(*args, **kwargs)
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_smollm2_135m_model():
+    # download the model
+    snapshot_download_w_retry("HuggingFaceTB/SmolLM2-135M")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_llama_68m_random_model():
+    # download the model
+    snapshot_download_w_retry("JackFram/llama-68m")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_qwen_2_5_half_billion_model():
+    # download the model
+    snapshot_download_w_retry("Qwen/Qwen2.5-0.5B")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_tatsu_lab_alpaca_dataset():
+    # download the dataset
+    snapshot_download_w_retry("tatsu-lab/alpaca", repo_type="dataset")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_mhenrichsen_alpaca_2k_dataset():
+    # download the dataset
+    snapshot_download_w_retry("mhenrichsen/alpaca_2k_test", repo_type="dataset")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_mhenrichsen_alpaca_2k_w_revision_dataset():
+    # download the dataset
+    snapshot_download_w_retry(
+        "mhenrichsen/alpaca_2k_test", repo_type="dataset", revision="d05c1cb"
+    )
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_mlabonne_finetome_100k_dataset():
+    # download the dataset
+    snapshot_download_w_retry("mlabonne/FineTome-100k", repo_type="dataset")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_argilla_distilabel_capybara_dpo_7k_binarized_dataset():
+    # download the dataset
+    snapshot_download_w_retry(
+        "argilla/distilabel-capybara-dpo-7k-binarized", repo_type="dataset"
+    )
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_argilla_ultrafeedback_binarized_preferences_cleaned_dataset():
+    # download the dataset
+    snapshot_download_w_retry(
+        "argilla/ultrafeedback-binarized-preferences-cleaned", repo_type="dataset"
+    )
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_arcee_ai_distilabel_intel_orca_dpo_pairs_dataset():
+    # download the dataset
+    snapshot_download_w_retry(
+        "arcee-ai/distilabel-intel-orca-dpo-pairs-binarized", repo_type="dataset"
+    )
+
+
+@pytest.fixture
+def temp_dir():
+    # Create a temporary directory
+    _temp_dir = tempfile.mkdtemp()
+    yield _temp_dir
+    # Clean up the directory after the test
+    shutil.rmtree(_temp_dir)
+
+
+@pytest.fixture(scope="function", autouse=True)
+def cleanup_monkeypatches():
+    from transformers import Trainer
+    from transformers.models.llama.modeling_llama import (
+        LlamaAttention,
+        LlamaFlashAttention2,
+        LlamaForCausalLM,
+    )
+
+    original_fa2_forward = LlamaFlashAttention2.forward
+    original_llama_attn_forward = LlamaAttention.forward
+    original_llama_forward = LlamaForCausalLM.forward
+    original_trainer_inner_training_loop = (
+        Trainer._inner_training_loop  # pylint: disable=protected-access
+    )
+    original_trainer_training_step = Trainer.training_step
+    # monkey patches can happen inside the tests
+    yield
+    # Reset LlamaFlashAttention2 forward
+    LlamaFlashAttention2.forward = original_fa2_forward
+    LlamaAttention.forward = original_llama_attn_forward
+    LlamaForCausalLM.forward = original_llama_forward
+    Trainer._inner_training_loop = (  # pylint: disable=protected-access
+        original_trainer_inner_training_loop
+    )
+    Trainer.training_step = original_trainer_training_step
+
+    # Reset other known monkeypatches
+    modules_to_reset: list[tuple[str, list[str]]] = [
+        ("transformers.models.llama",),
+        (
+            "transformers.models.llama.modeling_llama",
+            ["LlamaFlashAttention2", "LlamaAttention"],
+        ),
+        ("transformers.trainer",),
+        ("transformers", ["Trainer"]),
+        ("transformers.loss.loss_utils",),
+    ]
+    for module_name_tuple in modules_to_reset:
+        module_name = module_name_tuple[0]
+
+        spec = importlib.util.spec_from_file_location(
+            module_name, sys.modules[module_name].__file__
+        )
+        sys.modules[module_name] = importlib.util.module_from_spec(spec)
+        spec.loader.exec_module(sys.modules[module_name])
+
+        sys.modules[module_name] = importlib.reload(sys.modules[module_name])
+        if len(module_name_tuple) > 1:
+            module_globals = module_name_tuple[1]
+            for module_global in module_globals:
+                globals().pop(module_global, None)

tests/constants.py

@@ -0,0 +1,32 @@
+# constants.py
+"""
+This module contains constants and configuration dictionaries used for
+datasets and other utilities in the Axolotl project, specifically for testing.
+"""
+# Configuration for Alpaca Messages Dataset
+ALPACA_MESSAGES_CONFIG_OG = {
+    "path": "fozziethebeat/alpaca_messages_2k_dpo_test",
+    "type": "chat_template.default",
+    "chat_template": "llama3",
+    "field_messages": "conversation",
+    "field_chosen": "chosen",
+    "field_rejected": "rejected",
+    "message_field_role": "role",
+    "message_field_content": "content",
+    "roles": {
+        "system": ["system"],
+        "user": ["user"],
+        "assistant": ["assistant"],
+    },
+}
+
+# Revision configuration extending the original
+ALPACA_MESSAGES_CONFIG_REVISION = ALPACA_MESSAGES_CONFIG_OG.copy()
+ALPACA_MESSAGES_CONFIG_REVISION["revision"] = "ea82cff"
+
+
+SPECIAL_TOKENS = {
+    "bos_token": "<s>",
+    "eos_token": "</s>",
+    "unk_token": "<unk>",
+}

tests/core/test_trainer_builder.py

@@ -14,9 +14,7 @@ from axolotl.utils.models import load_model, load_tokenizer
 def fixture_cfg():
     cfg = DictDefault(
         {
-            "base_model": "TinyLlama/TinyLlama-1.1B-Chat-v0.6",
-            "model_type": "AutoModelForCausalLM",
-            "tokenizer_type": "LlamaTokenizer",
+            "base_model": "HuggingFaceTB/SmolLM2-135M",
             "micro_batch_size": 1,
             "gradient_accumulation_steps": 1,
             "learning_rate": 0.00005,
@@ -33,6 +31,9 @@ def fixture_cfg():
             "dataloader_num_workers": 1,
             "dataloader_pin_memory": True,
             "model_config_type": "llama",
+            "special_tokens": {
+                "pad_token": "<|endoftext|>",
+            },
         }
     )
 

tests/e2e/conftest.py

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

tests/e2e/each/test_fa_xentropy.py

@@ -4,8 +4,6 @@ E2E tests for lora llama
 
 import logging
 import os
-import unittest
-from importlib import reload
 from pathlib import Path
 
 import pytest
@@ -17,63 +15,61 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import with_temp_dir
+from ..utils import check_tensorboard
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
 
 
-@pytest.fixture(autouse=True)
-def reload_transformers():
-    import transformers.models.llama.modeling_llama
-
-    yield
-    reload(transformers.models.llama.modeling_llama)
-
-
-class TestFAXentropyLlama(unittest.TestCase):
+class TestFAXentropyLlama:
     """
     Test case for Llama models using LoRA w multipack
     """
 
-    @with_temp_dir
-    def test_lora_packing_fa_cross_entropy(self, temp_dir):
+    @pytest.mark.parametrize(
+        "gradient_accumulation_steps",
+        [1, 4],
+    )
+    def test_lora_packing_fa_cross_entropy(self, temp_dir, gradient_accumulation_steps):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
-                "tokenizer_type": "LlamaTokenizer",
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
                 "sequence_len": 1024,
                 "sample_packing": True,
                 "flash_attention": True,
                 "flash_attn_cross_entropy": True,
                 "load_in_8bit": True,
                 "adapter": "lora",
-                "lora_r": 32,
-                "lora_alpha": 64,
+                "lora_r": 8,
+                "lora_alpha": 16,
                 "lora_dropout": 0.05,
                 "lora_target_linear": True,
-                "val_set_size": 0.2,
+                "val_set_size": 0.05,
                 "special_tokens": {
-                    "unk_token": "<unk>",
-                    "bos_token": "<s>",
-                    "eos_token": "</s>",
+                    "pad_token": "<|endoftext|>",
                 },
+                "chat_template": "chatml",
                 "datasets": [
                     {
-                        "path": "mhenrichsen/alpaca_2k_test",
-                        "type": "alpaca",
+                        "path": "mlabonne/FineTome-100k",
+                        "field_messages": "conversations",
+                        "message_field_content": "value",
+                        "message_field_role": "from",
+                        "type": "chat_template",
+                        "split": "train[:2%]",
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 10,
-                "save_steps": 10,
-                "micro_batch_size": 8,
-                "gradient_accumulation_steps": 1,
+                "max_steps": 5,
+                "save_steps": 5,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": gradient_accumulation_steps,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
-                "optimizer": "adamw_torch",
+                "optimizer": "adamw_8bit",
                 "lr_scheduler": "cosine",
+                "use_tensorboard": True,
             }
         )
         if is_torch_bf16_gpu_available():
@@ -87,3 +83,7 @@ class TestFAXentropyLlama(unittest.TestCase):
 
         train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
         assert (Path(temp_dir) / "adapter_model.bin").exists()
+
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 1.5, "Train Loss is too high"
+        )

tests/e2e/each/test_lora_llama_multipack.py

@@ -98,7 +98,7 @@ class TestLoraLlama(unittest.TestCase):
                 "lora_alpha": 64,
                 "lora_dropout": 0.05,
                 "lora_target_linear": True,
-                "val_set_size": 0.1,
+                "val_set_size": 0.02,
                 "special_tokens": {
                     "unk_token": "<unk>",
                     "bos_token": "<s>",

tests/e2e/each/test_resume.py

@@ -6,7 +6,6 @@ import logging
 import os
 import re
 import subprocess
-import unittest
 from pathlib import Path
 
 from transformers.utils import is_torch_bf16_gpu_available
@@ -17,35 +16,35 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import most_recent_subdir, with_temp_dir
+from ..utils import most_recent_subdir
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
 
 
-class TestResumeLlama(unittest.TestCase):
+class TestResumeLlama:
     """
     Test case for resuming training of llama models
     """
 
-    @with_temp_dir
-    def test_resume_qlora_packed(self, temp_dir):
+    def test_resume_lora_packed(self, temp_dir):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
-                "tokenizer_type": "LlamaTokenizer",
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
                 "sequence_len": 1024,
                 "sample_packing": True,
                 "flash_attention": True,
-                "load_in_4bit": True,
-                "adapter": "qlora",
-                "lora_r": 32,
-                "lora_alpha": 64,
+                "load_in_8bit": True,
+                "adapter": "lora",
+                "lora_r": 8,
+                "lora_alpha": 16,
                 "lora_dropout": 0.05,
                 "lora_target_linear": True,
-                "val_set_size": 0.1,
-                "special_tokens": {},
+                "val_set_size": 0.001,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                },
                 "datasets": [
                     {
                         "path": "vicgalle/alpaca-gpt4",
@@ -57,11 +56,11 @@ class TestResumeLlama(unittest.TestCase):
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
-                "optimizer": "adamw_torch",
+                "optimizer": "adamw_8bit",
                 "lr_scheduler": "cosine",
-                "save_steps": 10,
+                "save_steps": 3,
                 "save_total_limit": 5,
-                "max_steps": 40,
+                "max_steps": 15,
                 "use_tensorboard": True,
             }
         )
@@ -77,7 +76,7 @@ class TestResumeLlama(unittest.TestCase):
 
         resume_cfg = cfg | DictDefault(
             {
-                "resume_from_checkpoint": f"{temp_dir}/checkpoint-30/",
+                "resume_from_checkpoint": f"{temp_dir}/checkpoint-9/",
             }
         )
         normalize_config(resume_cfg)
@@ -93,4 +92,4 @@ class TestResumeLlama(unittest.TestCase):
         )
         pattern = r"first_step\s+(\d+)"
         first_steps = int(re.findall(pattern, res.stdout)[0])
-        assert first_steps == 31
+        assert first_steps == 10

tests/e2e/each/test_unsloth_qlora.py

@@ -0,0 +1,186 @@
+"""
+e2e tests for unsloth qlora
+"""
+import logging
+import os
+from pathlib import Path
+
+import pytest
+
+from axolotl.cli import load_datasets
+from axolotl.common.cli import TrainerCliArgs
+from axolotl.train import train
+from axolotl.utils.config import normalize_config
+from axolotl.utils.dict import DictDefault
+
+from ..utils import check_tensorboard
+
+LOG = logging.getLogger("axolotl.tests.e2e")
+os.environ["WANDB_DISABLED"] = "true"
+
+
+# pylint: disable=duplicate-code
+class TestUnslothQLoRA:
+    """
+    Test class for Unsloth QLoRA Llama models
+    """
+
+    @pytest.mark.parametrize(
+        "sample_packing",
+        [True, False],
+    )
+    def test_unsloth_llama_qlora_fa2(self, temp_dir, sample_packing):
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "sequence_len": 1024,
+                "sample_packing": sample_packing,
+                "flash_attention": True,
+                "unsloth_lora_mlp": True,
+                "unsloth_lora_qkv": True,
+                "unsloth_lora_o": True,
+                "load_in_4bit": True,
+                "adapter": "qlora",
+                "lora_r": 16,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "val_set_size": 0.05,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                },
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 5,
+                "save_steps": 10,
+                "micro_batch_size": 4,
+                "gradient_accumulation_steps": 2,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_8bit",
+                "lr_scheduler": "cosine",
+                "use_tensorboard": True,
+                "bf16": "auto",
+            }
+        )
+
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+        assert (Path(temp_dir) / "adapter_model.bin").exists()
+
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 2.0, "Train Loss is too high"
+        )
+
+    def test_unsloth_llama_qlora_unpacked(self, temp_dir):
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "sequence_len": 1024,
+                "unsloth_lora_mlp": True,
+                "unsloth_lora_qkv": True,
+                "unsloth_lora_o": True,
+                "sample_packing": False,
+                "load_in_4bit": True,
+                "adapter": "qlora",
+                "lora_r": 16,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "val_set_size": 0.05,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                },
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 5,
+                "save_steps": 10,
+                "micro_batch_size": 4,
+                "gradient_accumulation_steps": 2,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_8bit",
+                "lr_scheduler": "cosine",
+                "use_tensorboard": True,
+                "bf16": "auto",
+            }
+        )
+
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+        assert (Path(temp_dir) / "adapter_model.bin").exists()
+
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 2.0, "Train Loss is too high"
+        )
+
+    @pytest.mark.parametrize(
+        "sdp_attention",
+        [True, False],
+    )
+    def test_unsloth_llama_qlora_unpacked_no_fa2_fp16(self, temp_dir, sdp_attention):
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "sequence_len": 1024,
+                "unsloth_lora_mlp": True,
+                "unsloth_lora_qkv": True,
+                "unsloth_lora_o": True,
+                "sample_packing": False,
+                "load_in_4bit": True,
+                "adapter": "qlora",
+                "lora_r": 16,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "val_set_size": 0.05,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                },
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 5,
+                "save_steps": 10,
+                "micro_batch_size": 4,
+                "gradient_accumulation_steps": 2,
+                "sdp_attention": sdp_attention,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_8bit",
+                "lr_scheduler": "cosine",
+                "use_tensorboard": True,
+                "fp16": True,
+            }
+        )
+
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+        assert (Path(temp_dir) / "adapter_model.bin").exists()
+
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 2.0, "Train Loss is too high"
+        )

tests/e2e/integrations/liger.py

@@ -7,7 +7,7 @@ from pathlib import Path
 from axolotl.cli import load_datasets
 from axolotl.common.cli import TrainerCliArgs
 from axolotl.train import train
-from axolotl.utils.config import normalize_config
+from axolotl.utils.config import normalize_config, prepare_plugins
 from axolotl.utils.dict import DictDefault
 
 from ..utils import with_temp_dir
@@ -54,8 +54,10 @@ class LigerIntegrationTestCase(unittest.TestCase):
                 "lr_scheduler": "cosine",
                 "save_safetensors": True,
                 "bf16": "auto",
+                "max_steps": 10,
             }
         )
+        prepare_plugins(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
@@ -99,8 +101,10 @@ class LigerIntegrationTestCase(unittest.TestCase):
                 "lr_scheduler": "cosine",
                 "save_safetensors": True,
                 "bf16": "auto",
+                "max_steps": 10,
             }
         )
+        prepare_plugins(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)

tests/e2e/integrations/test_cut_cross_entropy.py

@@ -0,0 +1,98 @@
+"""
+Simple end-to-end test for Cut Cross Entropy integration
+"""
+
+from pathlib import Path
+
+import pytest
+
+from axolotl.cli import load_datasets
+from axolotl.common.cli import TrainerCliArgs
+from axolotl.train import train
+from axolotl.utils import get_pytorch_version
+from axolotl.utils.config import normalize_config, prepare_plugins
+from axolotl.utils.dict import DictDefault
+
+# pylint: disable=duplicate-code
+
+
+@pytest.fixture()
+def min_cfg(temp_dir):
+    return {
+        "base_model": "HuggingFaceTB/SmolLM2-135M",
+        "plugins": [
+            "axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin",
+        ],
+        "cut_cross_entropy": True,
+        "sequence_len": 1024,
+        "val_set_size": 0.1,
+        "special_tokens": {
+            "pad_token": "<|endoftext|>",
+        },
+        "datasets": [
+            {
+                "path": "mhenrichsen/alpaca_2k_test",
+                "type": "alpaca",
+            },
+        ],
+        "num_epochs": 1,
+        "micro_batch_size": 8,
+        "gradient_accumulation_steps": 1,
+        "learning_rate": 0.00001,
+        "optimizer": "adamw_torch",
+        "output_dir": temp_dir,
+        "lr_scheduler": "cosine",
+        "save_safetensors": True,
+        "max_steps": 10,
+        "bf16": "auto",
+    }
+
+
+class TestCutCrossEntropyIntegration:
+    """
+    e2e tests for cut_cross_entropy integration with Axolotl
+    """
+
+    # pylint: disable=redefined-outer-name
+    def test_llama_w_cce(self, min_cfg, temp_dir):
+        cfg = DictDefault(min_cfg)
+        prepare_plugins(cfg)
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        major, minor, _ = get_pytorch_version()
+        if (major, minor) < (2, 4):
+            with pytest.raises(ImportError):
+                train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+        else:
+            train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+            assert (Path(temp_dir) / "model.safetensors").exists()
+
+    @pytest.mark.parametrize(
+        "attention_type",
+        [
+            "flash_attention",
+            "sdp_attention",
+            # "xformers_attention",
+        ],
+    )
+    def test_llama_w_cce_and_attention(self, min_cfg, temp_dir, attention_type):
+        cfg = DictDefault(
+            min_cfg
+            | {
+                attention_type: True,
+            }
+        )
+        prepare_plugins(cfg)
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        major, minor, _ = get_pytorch_version()
+        if (major, minor) < (2, 4):
+            with pytest.raises(ImportError):
+                train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+        else:
+            train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+            assert (Path(temp_dir) / "model.safetensors").exists()

tests/e2e/multigpu/test_eval.py

@@ -11,6 +11,8 @@ from transformers.testing_utils import get_torch_dist_unique_port
 
 from axolotl.utils.dict import DictDefault
 
+from ..utils import check_tensorboard
+
 LOG = logging.getLogger("axolotl.tests.e2e.multigpu")
 os.environ["WANDB_DISABLED"] = "true"
 
@@ -26,7 +28,7 @@ class TestMultiGPUEval:
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
                 "load_in_8bit": False,
                 "load_in_4bit": True,
                 "strict": False,
@@ -40,8 +42,8 @@ class TestMultiGPUEval:
                 "lora_dropout": 0.05,
                 "lora_target_linear": True,
                 "lora_modules_to_save": ["embed_tokens", "lm_head"],
-                "val_set_size": 0.1,
-                "special_tokens": {"pad_token": "<|end_of_text|>"},
+                "val_set_size": 0.004,
+                "special_tokens": {"pad_token": "<|endoftext|>"},
                 "datasets": [
                     {
                         "path": "teknium/GPT4-LLM-Cleaned",
@@ -66,6 +68,7 @@ class TestMultiGPUEval:
                 "saves_per_epoch": 1,
                 "logging_steps": 1,
                 "weight_decay": 0.0,
+                "use_tensorboard": True,
             }
         )
 
@@ -88,11 +91,13 @@ class TestMultiGPUEval:
             ]
         )
 
+        check_tensorboard(temp_dir + "/runs", "eval/loss", 2.5, "Eval Loss is too high")
+
     def test_eval(self, temp_dir):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
                 "load_in_8bit": False,
                 "load_in_4bit": True,
                 "strict": False,
@@ -106,8 +111,8 @@ class TestMultiGPUEval:
                 "lora_dropout": 0.05,
                 "lora_target_linear": True,
                 "lora_modules_to_save": ["embed_tokens", "lm_head"],
-                "val_set_size": 0.1,
-                "special_tokens": {"pad_token": "<|end_of_text|>"},
+                "val_set_size": 0.0004,
+                "special_tokens": {"pad_token": "<|endoftext|>"},
                 "datasets": [
                     {
                         "path": "teknium/GPT4-LLM-Cleaned",
@@ -132,6 +137,7 @@ class TestMultiGPUEval:
                 "saves_per_epoch": 1,
                 "logging_steps": 1,
                 "weight_decay": 0.0,
+                "use_tensorboard": True,
             }
         )
 
@@ -153,3 +159,5 @@ class TestMultiGPUEval:
                 str(Path(temp_dir) / "config.yaml"),
             ]
         )
+
+        check_tensorboard(temp_dir + "/runs", "eval/loss", 2.9, "Eval Loss is too high")

tests/e2e/multigpu/test_llama.py

@@ -9,13 +9,12 @@ from pathlib import Path
 import pytest
 import yaml
 from accelerate.test_utils import execute_subprocess_async
+from e2e.utils import check_tensorboard
 from huggingface_hub import snapshot_download
 from transformers.testing_utils import get_torch_dist_unique_port
 
 from axolotl.utils.dict import DictDefault
 
-from ..utils import is_hopper
-
 LOG = logging.getLogger("axolotl.tests.e2e.multigpu")
 os.environ["WANDB_DISABLED"] = "true"
 
@@ -55,7 +54,7 @@ class TestMultiGPULlama:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 15,
+                "max_steps": 2,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -63,6 +62,7 @@ class TestMultiGPULlama:
                 "optimizer": "adamw_8bit",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
+                "use_tensorboard": True,
             }
         )
 
@@ -85,9 +85,13 @@ class TestMultiGPULlama:
             ]
         )
 
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
+        )
+
     @pytest.mark.parametrize(
         "gradient_accumulation_steps",
-        [1, 4],
+        [1, 2],
     )
     def test_lora_ddp_packed(self, temp_dir, gradient_accumulation_steps):
         # pylint: disable=duplicate-code
@@ -114,14 +118,15 @@ class TestMultiGPULlama:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 15,
-                "micro_batch_size": 4,
+                "max_steps": 2,
+                "micro_batch_size": 1,
                 "gradient_accumulation_steps": gradient_accumulation_steps,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_8bit",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
+                "use_tensorboard": True,
             }
         )
 
@@ -144,7 +149,10 @@ class TestMultiGPULlama:
             ]
         )
 
-    @pytest.mark.skipif(is_hopper(), reason="h100 doesn't support 8-bit lora")
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
+        )
+
     def test_dpo_lora_ddp(self, temp_dir):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
@@ -183,7 +191,7 @@ class TestMultiGPULlama:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 15,
+                "max_steps": 2,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -192,6 +200,7 @@ class TestMultiGPULlama:
                 "optimizer": "adamw_8bit",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
+                "use_tensorboard": True,
             }
         )
 
@@ -214,6 +223,10 @@ class TestMultiGPULlama:
             ]
         )
 
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
+        )
+
     def test_dpo_qlora_ddp(self, temp_dir):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
@@ -252,8 +265,8 @@ class TestMultiGPULlama:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 15,
-                "micro_batch_size": 4,
+                "max_steps": 2,
+                "micro_batch_size": 2,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
                 "warmup_steps": 0,
@@ -261,6 +274,7 @@ class TestMultiGPULlama:
                 "optimizer": "adamw_8bit",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
+                "use_tensorboard": True,
             }
         )
 
@@ -283,9 +297,13 @@ class TestMultiGPULlama:
             ]
         )
 
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
+        )
+
     @pytest.mark.parametrize(
         "gradient_accumulation_steps",
-        [1, 4],
+        [1, 2],
     )
     def test_fsdp(self, temp_dir, gradient_accumulation_steps):
         # pylint: disable=duplicate-code
@@ -304,8 +322,8 @@ class TestMultiGPULlama:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 10,
-                "micro_batch_size": 4,
+                "max_steps": 2,
+                "micro_batch_size": 2,
                 "gradient_accumulation_steps": gradient_accumulation_steps,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
@@ -326,6 +344,7 @@ class TestMultiGPULlama:
                     "fsdp_state_dict_type": "FULL_STATE_DICT",
                     "fsdp_auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
                 },
+                "use_tensorboard": True,
             }
         )
 
@@ -348,6 +367,10 @@ class TestMultiGPULlama:
             ]
         )
 
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
+        )
+
     @pytest.mark.parametrize(
         "fsdp_state_dict_type",
         ["FULL_STATE_DICT", "SHARDED_STATE_DICT"],
@@ -371,7 +394,7 @@ class TestMultiGPULlama:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 15,
+                "max_steps": 2,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -393,6 +416,7 @@ class TestMultiGPULlama:
                     "fsdp_state_dict_type": fsdp_state_dict_type,
                     "fsdp_auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
                 },
+                "use_tensorboard": True,
             }
         )
 
@@ -415,6 +439,10 @@ class TestMultiGPULlama:
             ]
         )
 
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
+        )
+
     def test_fsdp_qlora_prequant_packed(self, temp_dir):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
@@ -447,7 +475,7 @@ class TestMultiGPULlama:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 15,
+                "max_steps": 2,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
                 "output_dir": temp_dir,
@@ -469,6 +497,7 @@ class TestMultiGPULlama:
                     "fsdp_state_dict_type": "SHARDED_STATE_DICT",
                     "fsdp_auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
                 },
+                "use_tensorboard": True,
             }
         )
 
@@ -491,12 +520,41 @@ class TestMultiGPULlama:
             ]
         )
 
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
+        )
+
     @pytest.mark.parametrize(
         "gradient_accumulation_steps",
-        [1, 4],
+        [1, 2],
     )
-    def test_ds_zero3_packed(self, temp_dir, gradient_accumulation_steps):
+    @pytest.mark.parametrize(
+        "deepspeed",
+        [
+            "deepspeed_configs/zero3_bf16.json",
+            "deepspeed_configs/zero3_bf16_cpuoffload_all.json",
+            # "deepspeed_configs/zero3_bf16_cpuoffload_params.json",
+        ],
+    )
+    @pytest.mark.parametrize(
+        "qlora",
+        [True, False],
+    )
+    def test_ds_zero3_packed(
+        self, temp_dir, gradient_accumulation_steps, deepspeed, qlora
+    ):
         # pylint: disable=duplicate-code
+        if qlora:
+            adapter = {
+                "adapter": "qlora",
+                "lora_r": 8,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "load_in_4bit": True,
+            }
+        else:
+            adapter = {}
         cfg = DictDefault(
             {
                 "base_model": "HuggingFaceTB/SmolLM2-135M",
@@ -514,15 +572,17 @@ class TestMultiGPULlama:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 15,
-                "micro_batch_size": 4,
+                "max_steps": 2,
+                "micro_batch_size": 1,
                 "gradient_accumulation_steps": gradient_accumulation_steps,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
-                "deepspeed": str(AXOLOTL_ROOT / "deepspeed_configs/zero3_bf16.json"),
+                "deepspeed": str(AXOLOTL_ROOT / deepspeed),
+                "use_tensorboard": True,
+                **adapter,
             }
         )
 
@@ -545,19 +605,35 @@ class TestMultiGPULlama:
             ]
         )
 
-    def test_ds_zero3_qlora_packed(self, temp_dir):
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
+        )
+
+    @pytest.mark.parametrize(
+        "gradient_accumulation_steps",
+        [1, 2],
+    )
+    @pytest.mark.parametrize(
+        "qlora",
+        [True, False],
+    )
+    def test_ds_zero2_packed(self, temp_dir, gradient_accumulation_steps, qlora):
         # pylint: disable=duplicate-code
-        cfg = DictDefault(
-            {
-                "base_model": "HuggingFaceTB/SmolLM2-135M",
-                "load_in_4bit": True,
+        if qlora:
+            adapter = {
                 "adapter": "qlora",
                 "lora_r": 8,
                 "lora_alpha": 16,
                 "lora_dropout": 0.05,
                 "lora_target_linear": True,
+                "load_in_4bit": True,
+            }
+        else:
+            adapter = {}
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
                 "sample_packing": True,
-                "eval_sample_packing": False,
                 "pad_to_sequence_len": True,
                 "sequence_len": 2048,
                 "val_set_size": 0.05,
@@ -571,15 +647,17 @@ class TestMultiGPULlama:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 15,
-                "micro_batch_size": 4,
-                "gradient_accumulation_steps": 4,
+                "max_steps": 2,
+                "micro_batch_size": 1,
+                "gradient_accumulation_steps": gradient_accumulation_steps,
                 "output_dir": temp_dir,
-                "learning_rate": 0.0001,
+                "learning_rate": 0.00001,
                 "optimizer": "adamw_torch",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
-                "deepspeed": str(AXOLOTL_ROOT / "deepspeed_configs/zero3_bf16.json"),
+                "deepspeed": str(AXOLOTL_ROOT / "deepspeed_configs/zero2.json"),
+                "use_tensorboard": True,
+                **adapter,
             }
         )
 
@@ -601,3 +679,82 @@ class TestMultiGPULlama:
                 str(Path(temp_dir) / "config.yaml"),
             ]
         )
+
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
+        )
+
+    @pytest.mark.parametrize(
+        "gradient_accumulation_steps",
+        [1, 2],
+    )
+    @pytest.mark.parametrize(
+        "qlora",
+        [True, False],
+    )
+    def test_ds_zero1_packed(self, temp_dir, gradient_accumulation_steps, qlora):
+        # pylint: disable=duplicate-code
+        if qlora:
+            adapter = {
+                "adapter": "qlora",
+                "lora_r": 8,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "load_in_4bit": True,
+            }
+        else:
+            adapter = {}
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
+                "sequence_len": 2048,
+                "val_set_size": 0.05,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                },
+                "datasets": [
+                    {
+                        "path": "tatsu-lab/alpaca",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 2,
+                "micro_batch_size": 1,
+                "gradient_accumulation_steps": gradient_accumulation_steps,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_torch",
+                "lr_scheduler": "cosine",
+                "flash_attention": True,
+                "deepspeed": str(AXOLOTL_ROOT / "deepspeed_configs/zero1.json"),
+                "use_tensorboard": True,
+                **adapter,
+            }
+        )
+
+        # write cfg to yaml file
+        Path(temp_dir).mkdir(parents=True, exist_ok=True)
+        with open(Path(temp_dir) / "config.yaml", "w", encoding="utf-8") as fout:
+            fout.write(yaml.dump(cfg.to_dict(), Dumper=yaml.Dumper))
+
+        execute_subprocess_async(
+            [
+                "accelerate",
+                "launch",
+                "--num-processes",
+                "2",
+                "--main_process_port",
+                f"{get_torch_dist_unique_port()}",
+                "-m",
+                "axolotl.cli.train",
+                str(Path(temp_dir) / "config.yaml"),
+            ]
+        )
+
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
+        )

tests/e2e/patched/test_4d_multipack_llama.py

@@ -42,7 +42,7 @@ class Test4dMultipackLlama(unittest.TestCase):
                 "lora_dropout": 0.05,
                 "lora_target_linear": True,
                 "sequence_len": 1024,
-                "val_set_size": 0.1,
+                "val_set_size": 0.02,
                 "datasets": [
                     {
                         "path": "mhenrichsen/alpaca_2k_test",
@@ -86,7 +86,7 @@ class Test4dMultipackLlama(unittest.TestCase):
                 "lora_alpha": 16,
                 "lora_dropout": 0.05,
                 "lora_target_linear": True,
-                "val_set_size": 0.1,
+                "val_set_size": 0.02,
                 "datasets": [
                     {
                         "path": "mhenrichsen/alpaca_2k_test",