make lisa training example work on one 24gb gpu

* another attempt at github actions

* try again

.github/ISSUE_TEMPLATE/bug-report.yaml

@@ -59,6 +59,7 @@ body:
       label: Config yaml
       description: |
         Please attach the config yaml!
+      render: yaml
 
   - type: textarea
     id: possible-solution

.github/workflows/base.yml

@@ -16,17 +16,22 @@ jobs:
             cuda_version: 11.8.0
             python_version: "3.10"
             pytorch: 2.1.2
-            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 9.0+PTX"
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
           - cuda: "121"
             cuda_version: 12.1.0
             python_version: "3.10"
             pytorch: 2.1.2
-            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 9.0+PTX"
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
           - cuda: "121"
             cuda_version: 12.1.0
             python_version: "3.11"
             pytorch: 2.1.2
-            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 9.0+PTX"
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+          - cuda: "121"
+            cuda_version: 12.1.0
+            python_version: "3.11"
+            pytorch: 2.2.1
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
     steps:
       - name: Checkout
         uses: actions/checkout@v3

.github/workflows/docs.yml

@@ -0,0 +1,31 @@
+name: Publish Docs
+on:
+  push:
+    branches:
+      - main
+
+permissions:
+    contents: write
+    pages: write
+
+jobs:
+    build-deploy:
+        runs-on: ubuntu-latest
+        steps:
+        - name: Check out repository
+          uses: actions/checkout@v4
+        - name: Set up Quarto
+          uses: quarto-dev/quarto-actions/setup@v2
+        - name: Setup Python
+          uses: actions/setup-python@v3
+          with:
+            python-version: '3.10'
+        - name: install dependencies
+          run: |
+            python3 -m pip install jupyter
+        - name: Publish to GitHub Pages (and render)
+          uses: quarto-dev/quarto-actions/publish@v2
+          with:
+            target: gh-pages
+          env:
+            GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/lint.yml

@@ -17,6 +17,6 @@ jobs:
       - uses: actions/checkout@v3
       - uses: actions/setup-python@v4
         with:
-          python-version: "3.9"
+          python-version: "3.10"
           cache: 'pip' # caching pip dependencies
       - uses: pre-commit/action@v3.0.0

.github/workflows/main.yml

@@ -18,6 +18,7 @@ jobs:
             python_version: "3.10"
             pytorch: 2.1.2
             axolotl_extras:
+            axolotl_args: "--extra-index-url https://download.pytorch.org/whl/cu118"
             is_latest: true
           - cuda: 121
             cuda_version: 12.1.0
@@ -27,7 +28,7 @@ jobs:
           - cuda: 121
             cuda_version: 12.1.0
             python_version: "3.11"
-            pytorch: 2.1.2
+            pytorch: 2.2.1
             axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
@@ -54,6 +55,7 @@ jobs:
             BASE_TAG=${{ github.ref_name }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}
             CUDA=${{ matrix.cuda }}
             PYTORCH_VERSION=${{ matrix.pytorch }}
+            AXOLOTL_ARGS=${{ matrix.axolotl_args }}
           file: ./docker/Dockerfile
           push: ${{ github.event_name != 'pull_request' }}
           tags: |
@@ -61,7 +63,7 @@ jobs:
             ${{ (matrix.is_latest) && format('{0}-latest', steps.metadata.outputs.tags) || '' }}
           labels: ${{ steps.metadata.outputs.labels }}
 
-  build-axolotl-runpod:
+  build-axolotl-cloud:
     needs: build-axolotl
     if: ${{ ! contains(github.event.commits[0].message, '[skip docker]]') && github.repository_owner == 'OpenAccess-AI-Collective' }}
     # this job needs to be run on self-hosted GPU runners...
@@ -82,7 +84,7 @@ jobs:
           - cuda: 121
             cuda_version: 12.1.0
             python_version: "3.11"
-            pytorch: 2.1.2
+            pytorch: 2.2.1
             axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
@@ -111,7 +113,5 @@ jobs:
           push: ${{ github.event_name != 'pull_request' }}
           tags: |
              ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
-             winglian/axolotl-runpod:main-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
              ${{ (matrix.is_latest) && format('{0}-latest', steps.metadata.outputs.tags) || '' }}
-             ${{ (matrix.is_latest) && format('{0}-latest', 'winglian/axolotl-runpod:main') || '' }}
           labels: ${{ steps.metadata.outputs.labels }}

.github/workflows/nightlies.yml

@@ -0,0 +1,118 @@
+name: docker-nightlies
+
+on:
+  workflow_dispatch:
+  schedule:
+    - cron: '0 0 * * *'  # Runs at 00:00 UTC every day
+
+jobs:
+  build-axolotl:
+    if: ${{ ! contains(github.event.commits[0].message, '[skip docker]]') && github.repository_owner == 'OpenAccess-AI-Collective' }}
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - cuda: 118
+            cuda_version: 11.8.0
+            python_version: "3.10"
+            pytorch: 2.1.2
+            axolotl_extras:
+            axolotl_args: "--extra-index-url https://download.pytorch.org/whl/cu118"
+            is_latest: true
+          - cuda: 121
+            cuda_version: 12.1.0
+            python_version: "3.10"
+            pytorch: 2.1.2
+            axolotl_extras:
+          - cuda: 121
+            cuda_version: 12.1.0
+            python_version: "3.11"
+            pytorch: 2.2.1
+            axolotl_extras:
+    runs-on: axolotl-gpu-runner
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Docker metadata
+        id: metadata
+        uses: docker/metadata-action@v5
+        with:
+          images: winglian/axolotl
+          tags: |
+            type=raw,value={{ branch }}-{{ date 'YYYYMMDD' }}
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_TOKEN }}
+      # guidance for testing before pushing: https://docs.docker.com/build/ci/github-actions/test-before-push/
+      - name: Build and export to Docker
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          build-args: |
+            BASE_TAG=${{ github.ref_name }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}
+            CUDA=${{ matrix.cuda }}
+            PYTORCH_VERSION=${{ matrix.pytorch }}
+            AXOLOTL_ARGS=${{ matrix.axolotl_args }}
+          file: ./docker/Dockerfile
+          push: ${{ github.event_name != 'pull_request' }}
+          tags: |
+            ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
+          labels: ${{ steps.metadata.outputs.labels }}
+
+  build-axolotl-cloud:
+    needs: build-axolotl
+    if: ${{ ! contains(github.event.commits[0].message, '[skip docker]]') && github.repository_owner == 'OpenAccess-AI-Collective' }}
+    # this job needs to be run on self-hosted GPU runners...
+    strategy:
+      matrix:
+        include:
+          - cuda: 118
+            cuda_version: 11.8.0
+            python_version: "3.10"
+            pytorch: 2.1.2
+            axolotl_extras:
+            is_latest: true
+          - cuda: 121
+            cuda_version: 12.1.0
+            python_version: "3.10"
+            pytorch: 2.1.2
+            axolotl_extras:
+          - cuda: 121
+            cuda_version: 12.1.0
+            python_version: "3.11"
+            pytorch: 2.2.1
+            axolotl_extras:
+    runs-on: axolotl-gpu-runner
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Docker metadata
+        id: metadata
+        uses: docker/metadata-action@v5
+        with:
+          images: winglian/axolotl-cloud
+          tags: |
+            type=raw,value={{ branch }}-{{ date 'YYYYMMDD' }}
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_TOKEN }}
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v2
+      - name: Build
+        uses: docker/build-push-action@v5
+        with:
+          context: .
+          build-args: |
+            BASE_TAG=${{ github.ref_name }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
+            CUDA=${{ matrix.cuda }}
+          file: ./docker/Dockerfile-cloud
+          push: ${{ github.event_name != 'pull_request' }}
+          tags: |
+             ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
+          labels: ${{ steps.metadata.outputs.labels }}

.github/workflows/pypi.yml

@@ -25,7 +25,7 @@ jobs:
 
       - name: Install dependencies
         run: |
-          pip3 install wheel
+          pip3 install wheel packaging
           pip3 install -e .
           pip3 install -r requirements-tests.txt
 

.github/workflows/tests.yml

@@ -23,7 +23,7 @@ jobs:
       - uses: actions/checkout@v3
       - uses: actions/setup-python@v4
         with:
-          python-version: "3.9"
+          python-version: "3.10"
           cache: 'pip' # caching pip dependencies
       - uses: pre-commit/action@v3.0.0
 
@@ -33,8 +33,8 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python_version: ["3.9", "3.10", "3.11"]
-    timeout-minutes: 10
+        python_version: ["3.10", "3.11"]
+    timeout-minutes: 20
 
     steps:
       - name: Check out repository code
@@ -48,6 +48,8 @@ jobs:
 
       - name: Install dependencies
         run: |
+          pip3 install --upgrade pip
+          pip3 install --upgrade packaging
           pip3 install -U -e .
           pip3 install -r requirements-tests.txt
 
@@ -58,8 +60,8 @@ jobs:
   docker-e2e-tests:
     if: github.repository_owner == 'OpenAccess-AI-Collective'
     # this job needs to be run on self-hosted GPU runners...
-    runs-on: [self-hosted, gpu, docker]
-    timeout-minutes: 30
+    runs-on: [self-hosted, modal]
+    timeout-minutes: 60
     needs: [pre-commit, pytest]
 
     strategy:
@@ -70,43 +72,36 @@ jobs:
             cuda_version: 11.8.0
             python_version: "3.10"
             pytorch: 2.1.2
+            axolotl_args: "--extra-index-url https://download.pytorch.org/whl/cu118"
+            num_gpus: 1
           - cuda: 121
             cuda_version: 12.1.0
             python_version: "3.10"
             pytorch: 2.1.2
+            num_gpus: 1
+          - cuda: 121
+            cuda_version: 12.1.0
+            python_version: "3.11"
+            pytorch: 2.2.1
+            num_gpus: 1
     steps:
       - name: Checkout
         uses: actions/checkout@v4
-      - name: Docker metadata
-        id: metadata
-        uses: docker/metadata-action@v5
+      - name: Install Python
+        uses: actions/setup-python@v5
         with:
-          images: winglian/axolotl-tests
-      - name: Build Docker image
+          python-version: "3.10"
+      - name: Install Modal
         run: |
-          # Set up build arguments
-          BASE_TAG="main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}"
-          CUDA="${{ matrix.cuda }}"
-          PYTORCH_VERSION="${{ matrix.pytorch }}"
-          # Build the Docker image
-          docker build . \
-            --file ./docker/Dockerfile-tests \
-            --build-arg BASE_TAG=$BASE_TAG \
-            --build-arg CUDA=$CUDA \
-            --build-arg GITHUB_REF=$GITHUB_REF \
-            --build-arg PYTORCH_VERSION=$PYTORCH_VERSION \
-            --tag ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }} \
-            --no-cache
-      - name: Unit Tests w docker image
+          python -m pip install --upgrade pip
+          pip install modal jinja2
+      - name: Update env vars
         run: |
-          docker run --rm ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }} pytest --ignore=tests/e2e/ /workspace/axolotl/tests/
-      - name: GPU Unit Tests w docker image
+          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
+          echo "PYTORCH_VERSION=${{ matrix.pytorch}}" >> $GITHUB_ENV
+          echo "AXOLOTL_ARGS=${{ matrix.axolotl_args}}" >> $GITHUB_ENV
+          echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
+          echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
+      - name: Run tests job on Modal
         run: |
-          docker run --privileged --gpus "all" --env WANDB_DISABLED=true --rm ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }} pytest --ignore=tests/e2e/patched/ /workspace/axolotl/tests/e2e/
-      - name: GPU Unit Tests monkeypatched w docker image
-        run: |
-          docker run --privileged --gpus "all" --env WANDB_DISABLED=true --rm ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }} pytest /workspace/axolotl/tests/e2e/patched/
-      - name: Prune image from docker
-        if: github.ref != 'refs/heads/main'
-        run: |
-          docker rmi -f ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}
+          modal run cicd.tests

.gitignore

@@ -2,6 +2,7 @@
 configs
 last_run_prepared/
 .vscode
+_site/
 
 # Byte-compiled / optimized / DLL files
 __pycache__/
@@ -167,3 +168,10 @@ cython_debug/
 # WandB
 # wandb creates a folder to store logs for training runs
 wandb
+
+# Runs
+lora-out/*
+qlora-out/*
+mlruns/*
+
+/.quarto/

.mypy.ini

@@ -1,5 +1,5 @@
 [mypy]
-
+plugins = pydantic.mypy
 exclude = venv
 
 [mypy-alpaca_lora_4bit.*]

.pre-commit-config.yaml

@@ -31,6 +31,7 @@ repos:
       additional_dependencies:
         [
             'types-PyYAML',
+            'pydantic>=2.5.3',
         ]
 -   repo: https://github.com/PyCQA/bandit
     rev: 1.7.5

README.md

@@ -13,6 +13,9 @@ Features:
 - Log results and optionally checkpoints to wandb or mlflow
 - And more!
 
+<a href="https://www.phorm.ai/query?projectId=e315ba4a-4e14-421f-ab05-38a1f9076f25">
+  <img alt="phorm.ai" src="https://img.shields.io/badge/Phorm-Ask_AI-%23F2777A.svg?&logo=data:image/svg+xml;base64,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">
+</a>
 
 <table>
 <tr>
@@ -22,12 +25,14 @@ Features:
 - [Introduction](#axolotl)
 - [Supported Features](#axolotl-supports)
 - [Quickstart](#quickstart-)
-- [Installation](#installation)
+- [Environment](#environment)
   - [Docker](#docker)
   - [Conda/Pip venv](#condapip-venv)
-  - [Cloud GPU](#cloud-gpu) - Latitude.sh, RunPod
+  - [Cloud GPU](#cloud-gpu) - Latitude.sh, JarvisLabs, RunPod
   - [Bare Metal Cloud GPU](#bare-metal-cloud-gpu)
   - [Windows](#windows)
+  - [Mac](#mac)
+  - [Google Colab](#google-colab)
   - [Launching on public clouds via SkyPilot](#launching-on-public-clouds-via-skypilot)
 - [Dataset](#dataset)
   - [How to Add Custom Prompts](#how-to-add-custom-prompts)
@@ -38,8 +43,8 @@ Features:
   - [Merge LORA to Base](#merge-lora-to-base)
   - [Special Tokens](#special-tokens)
 - Advanced Topics
-  - [Multipack](./docs/multipack.md)<svg width="24" height="24" viewBox="0 0 24 24" xmlns="http://www.w3.org/2000/svg"><path d="M17 13.5v6H5v-12h6m3-3h6v6m0-6-9 9" class="icon_svg-stroke" stroke="#666" stroke-width="1.5" fill="none" fill-rule="evenodd" stroke-linecap="round" stroke-linejoin="round"></path></svg>
-  - [RLHF & DPO](./docs/rlhf.md)<svg width="24" height="24" viewBox="0 0 24 24" xmlns="http://www.w3.org/2000/svg"><path d="M17 13.5v6H5v-12h6m3-3h6v6m0-6-9 9" class="icon_svg-stroke" stroke="#666" stroke-width="1.5" fill="none" fill-rule="evenodd" stroke-linecap="round" stroke-linejoin="round"></path></svg>
+  - [Multipack](./docs/multipack.qmd)<svg width="24" height="24" viewBox="0 0 24 24" xmlns="http://www.w3.org/2000/svg"><path d="M17 13.5v6H5v-12h6m3-3h6v6m0-6-9 9" class="icon_svg-stroke" stroke="#666" stroke-width="1.5" fill="none" fill-rule="evenodd" stroke-linecap="round" stroke-linejoin="round"></path></svg>
+  - [RLHF & DPO](./docs/rlhf.qmd)<svg width="24" height="24" viewBox="0 0 24 24" xmlns="http://www.w3.org/2000/svg"><path d="M17 13.5v6H5v-12h6m3-3h6v6m0-6-9 9" class="icon_svg-stroke" stroke="#666" stroke-width="1.5" fill="none" fill-rule="evenodd" stroke-linecap="round" stroke-linejoin="round"></path></svg>
 - [Common Errors](#common-errors-)
   - [Tokenization Mismatch b/w Training & Inference](#tokenization-mismatch-bw-inference--training)
 - [Debugging Axolotl](#debugging-axolotl)
@@ -87,17 +92,18 @@ Features:
 | phi         | ✅         | ✅    | ✅     | ❓             | ❓                 | ❓          | ❓            |
 | RWKV        | ✅         | ❓    | ❓     | ❓             | ❓                 | ❓          | ❓            |
 | Qwen        | ✅         | ✅    | ✅     | ❓             | ❓                 | ❓          | ❓            |
+| Gemma       | ✅         | ✅    | ✅     | ❓             | ❓                 | ✅          | ❓            |
 
+✅: supported
+❌: 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**: Python >=3.9 and Pytorch >=2.0.
+**Requirements**: Python >=3.10 and Pytorch >=2.1.1.
 
-`pip3 install "axolotl[flash-attn,deepspeed] @ git+https://github.com/OpenAccess-AI-Collective/axolotl"`
-
-### For developers
 ```bash
 git clone https://github.com/OpenAccess-AI-Collective/axolotl
 cd axolotl
@@ -127,13 +133,14 @@ accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
 accelerate launch -m axolotl.cli.train https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/main/examples/openllama-3b/lora.yml
 ```
 
-## Installation
+## Advanced Setup
 
 ### Environment
 
 #### Docker
+
   ```bash
-  docker run --gpus '"all"' --rm -it winglian/axolotl:main-py3.10-cu118-2.0.1
+  docker run --gpus '"all"' --rm -it winglian/axolotl:main-latest
   ```
 
   Or run on the current files for development:
@@ -143,7 +150,7 @@ accelerate launch -m axolotl.cli.train https://raw.githubusercontent.com/OpenAcc
   ```
 
 >[!Tip]
-> If you want to debug axolotl or prefer to use Docker as your development environment, see the [debugging guide's section on Docker](docs/debugging.md#debugging-with-docker).
+> If you want to debug axolotl or prefer to use Docker as your development environment, see the [debugging guide's section on Docker](docs/debugging.qmd#debugging-with-docker).
 
   <details>
 
@@ -152,7 +159,7 @@ accelerate launch -m axolotl.cli.train https://raw.githubusercontent.com/OpenAcc
   A more powerful Docker command to run would be this:
 
   ```bash
-docker run --privileged --gpus '"all"' --shm-size 10g --rm -it --name axolotl --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --mount type=bind,src="${PWD}",target=/workspace/axolotl -v ${HOME}/.cache/huggingface:/root/.cache/huggingface winglian/axolotl:main-py3.10-cu118-2.0.1
+docker run --privileged --gpus '"all"' --shm-size 10g --rm -it --name axolotl --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --mount type=bind,src="${PWD}",target=/workspace/axolotl -v ${HOME}/.cache/huggingface:/root/.cache/huggingface winglian/axolotl:main-latest
   ```
 
   It additionally:
@@ -167,7 +174,7 @@ docker run --privileged --gpus '"all"' --shm-size 10g --rm -it --name axolotl --
   </details>
 
 #### Conda/Pip venv
-  1. Install python >=**3.9**
+  1. Install python >=**3.10**
 
   2. Install pytorch stable https://pytorch.org/get-started/locally/
 
@@ -187,6 +194,7 @@ docker run --privileged --gpus '"all"' --shm-size 10g --rm -it --name axolotl --
 For cloud GPU providers that support docker images, use [`winglian/axolotl-cloud:main-latest`](https://hub.docker.com/r/winglian/axolotl-cloud/tags)
 
 - on Latitude.sh use this [direct link](https://latitude.sh/blueprint/989e0e79-3bf6-41ea-a46b-1f246e309d5c)
+- on JarvisLabs.ai use this [direct link](https://jarvislabs.ai/templates/axolotl)
 - on RunPod use this [direct link](https://runpod.io/gsc?template=v2ickqhz9s&ref=6i7fkpdz)
 
 #### Bare Metal Cloud GPU
@@ -200,11 +208,11 @@ For cloud GPU providers that support docker images, use [`winglian/axolotl-cloud
   1. Install python
   ```bash
   sudo apt update
-  sudo apt install -y python3.9
+  sudo apt install -y python3.10
 
-  sudo update-alternatives --install /usr/bin/python python /usr/bin/python3.9 1
-  sudo update-alternatives --config python # pick 3.9 if given option
-  python -V # should be 3.9
+  sudo update-alternatives --install /usr/bin/python python /usr/bin/python3.10 1
+  sudo update-alternatives --config python # pick 3.10 if given option
+  python -V # should be 3.10
 
   ```
 
@@ -236,21 +244,50 @@ For cloud GPU providers that support docker images, use [`winglian/axolotl-cloud
   ```
   </details>
 
+##### GCP
+
+<details>
+
+<summary>Click to Expand</summary>
+
+Use a Deeplearning linux OS with cuda and pytorch installed. Then follow instructions on quickstart.
+
+Make sure to run the below to uninstall xla.
+```bash
+pip uninstall -y torch_xla[tpu]
+```
+
+</details>
+
 #### Windows
 Please use WSL or Docker!
 
+#### Mac
+
+Use the below instead of the install method in QuickStart.
+```
+pip3 install -e '.'
+```
+More info: [mac.md](/docs/mac.qmd)
+
+#### Google Colab
+
+Please use this example [notebook](examples/colab-notebooks/colab-axolotl-example.ipynb).
 
 #### Launching on public clouds via SkyPilot
 To launch on GPU instances (both on-demand and spot instances) on 7+ clouds (GCP, AWS, Azure, OCI, and more), you can use [SkyPilot](https://skypilot.readthedocs.io/en/latest/index.html):
+
 ```bash
 pip install "skypilot-nightly[gcp,aws,azure,oci,lambda,kubernetes,ibm,scp]"  # choose your clouds
 sky check
 ```
+
 Get the [example YAMLs](https://github.com/skypilot-org/skypilot/tree/master/llm/axolotl) of using Axolotl to finetune `mistralai/Mistral-7B-v0.1`:
 ```
 git clone https://github.com/skypilot-org/skypilot.git
 cd skypilot/llm/axolotl
 ```
+
 Use one command to launch:
 ```bash
 # On-demand
@@ -260,32 +297,33 @@ HF_TOKEN=xx sky launch axolotl.yaml --env HF_TOKEN
 HF_TOKEN=xx BUCKET=<unique-name> sky spot launch axolotl-spot.yaml --env HF_TOKEN --env BUCKET
 ```
 
-
 ### Dataset
 
 Axolotl supports a variety of dataset formats. Below are some of the formats you can use.
 Have dataset(s) in one of the following format (JSONL recommended):
 
-- `alpaca`: instruction; input(optional)
-  ```json
-  {"instruction": "...", "input": "...", "output": "..."}
-  ```
-- `sharegpt`: conversations where `from` is `human`/`gpt`. (optional: `system` to override default system prompt)
-  ```json
-  {"conversations": [{"from": "...", "value": "..."}]}
-  ```
-- `llama-2`: the json is the same format as `sharegpt` above, with the following config (see the [config section](#config) for more details)
-    ```yml
-    datasets:
-      - path: <your-path>
-        type: sharegpt
-        conversation: llama-2
-    ```
+#### Pretraining
+
 - `completion`: raw corpus
   ```json
   {"text": "..."}
   ```
 
+Note: Axolotl usually loads the entire dataset into memory. This will be challenging for large datasets. Use the following config to enable streaming:
+
+```yaml
+pretraining_dataset: # hf path only
+```
+
+#### Supervised finetuning
+
+##### Instruction
+
+- `alpaca`: instruction; input(optional)
+  ```json
+  {"instruction": "...", "input": "...", "output": "..."}
+  ```
+
 <details>
 
 <summary>See other formats</summary>
@@ -362,14 +400,37 @@ Have dataset(s) in one of the following format (JSONL recommended):
   ```json
   {"scores": "...", "critiques": "...", "instruction": "...", "answer": "...", "revision": "..."}
   ```
-- `pygmalion`: pygmalion
-  ```json
-  {"conversations": [{"role": "...", "value": "..."}]}
-  ```
 - `metharme`: instruction, adds additional eos tokens
   ```json
   {"prompt": "...", "generation": "..."}
   ```
+
+</details>
+
+##### Template-Free
+
+- `input_output`: template-free prompt construction
+  ```json
+   {"segments": [{"label": true|false, "text": "..."}]}
+  ```
+
+This is a special format that allows you to construct prompts without using templates. This is for advanced users who want more freedom with prompt construction.  See [these docs](docs/input_output.qmd) for more details.
+
+##### Conversation
+
+- `sharegpt`: conversations where `from` is `human`/`gpt`. (optional: first row with role `system` to override default system prompt)
+  ```json
+  {"conversations": [{"from": "...", "value": "..."}]}
+  ```
+
+<details>
+
+<summary>See other formats</summary>
+
+- `pygmalion`: pygmalion
+  ```json
+  {"conversations": [{"role": "...", "value": "..."}]}
+  ```
 - `sharegpt.load_role`: conversations where `role` is used instead of `from`
   ```json
   {"conversations": [{"role": "...", "value": "..."}]}
@@ -385,6 +446,8 @@ Have dataset(s) in one of the following format (JSONL recommended):
 
 </details>
 
+Note: `type: sharegpt` opens a special config `conversation:` that enables conversions to many Conversation types. See dataset section under [all yaml options](#all-yaml-options).
+
 #### How to add custom prompts
 
 For a dataset that is preprocessed for instruction purposes:
@@ -406,12 +469,16 @@ datasets:
       format: "[INST] {instruction} [/INST]"
       no_input_format: "[INST] {instruction} [/INST]"
 ```
+See full config options under [all yaml options](#all-yaml-options).
 
 #### How to use your custom pretokenized dataset
 
 - Do not pass a `type:`
 - Columns in Dataset must be exactly `input_ids`, `attention_mask`, `labels`
 
+```yaml
+- path: ...
+```
 
 ### Config
 
@@ -425,22 +492,18 @@ See [examples](examples) for quick start. It is recommended to duplicate and mod
 
 - dataset
   ```yaml
-  sequence_len: 2048 # max token length for prompt
-
-  # huggingface repo
   datasets:
+      # huggingface repo
     - path: vicgalle/alpaca-gpt4
-      type: alpaca # format from earlier
+      type: alpaca
 
-  # huggingface repo with specific configuration/subset
-  datasets:
+      # huggingface repo with specific configuration/subset
     - path: EleutherAI/pile
       name: enron_emails
       type: completion # format from earlier
       field: text # Optional[str] default: text, field to use for completion data
 
-  # huggingface repo with multiple named configurations/subsets
-  datasets:
+      # huggingface repo with multiple named configurations/subsets
     - path: bigcode/commitpackft
       name:
         - ruby
@@ -448,34 +511,29 @@ See [examples](examples) for quick start. It is recommended to duplicate and mod
         - typescript
       type: ... # unimplemented custom format
 
-  # fastchat conversation
-  # See 'conversation' options: https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
-  datasets:
+      # fastchat conversation
+      # See 'conversation' options: https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
     - path: ...
       type: sharegpt
-      conversation: chatml
+      conversation: chatml # default: vicuna_v1.1
 
-  # local
-  datasets:
+      # local
     - path: data.jsonl # or json
       ds_type: json # see other options below
       type: alpaca
 
-  # dataset with splits, but no train split
-  dataset:
+      # dataset with splits, but no train split
     - path: knowrohit07/know_sql
       type: context_qa.load_v2
       train_on_split: validation
 
-  # loading from s3 or gcs
-  # s3 creds will be loaded from the system default and gcs only supports public access
-  dataset:
+      # loading from s3 or gcs
+      # s3 creds will be loaded from the system default and gcs only supports public access
     - path: s3://path_to_ds # Accepts folder with arrow/parquet or file path like above. Supports s3, gcs.
       ...
 
-  # Loading Data From a Public URL
-  # - The file format is `json` (which includes `jsonl`) by default. For different formats, adjust the `ds_type` option accordingly.
-  dataset:
+      # Loading Data From a Public URL
+      # - The file format is `json` (which includes `jsonl`) by default. For different formats, adjust the `ds_type` option accordingly.
     - path: https://some.url.com/yourdata.jsonl # The URL should be a direct link to the file you wish to load. URLs must use HTTPS protocol, not HTTP.
       ds_type: json # this is the default, see other options below.
   ```
@@ -484,9 +542,11 @@ See [examples](examples) for quick start. It is recommended to duplicate and mod
   ```yaml
   load_in_4bit: true
   load_in_8bit: true
+
   bf16: auto # require >=ampere, auto will detect if your GPU supports this and choose automatically.
   fp16: # leave empty to use fp16 when bf16 is 'auto'. set to false if you want to fallback to fp32
   tf32: true # require >=ampere
+
   bfloat16: true # require >=ampere, use instead of bf16 when you don't want AMP (automatic mixed precision)
   float16: true # use instead of fp16 when you don't want AMP
   ```
@@ -494,7 +554,7 @@ See [examples](examples) for quick start. It is recommended to duplicate and mod
 
 - lora
   ```yaml
-  adapter: lora # qlora or leave blank for full finetune
+  adapter: lora # 'qlora' or leave blank for full finetune
   lora_r: 8
   lora_alpha: 16
   lora_dropout: 0.05
@@ -503,9 +563,9 @@ See [examples](examples) for quick start. It is recommended to duplicate and mod
     - v_proj
   ```
 
-<details>
+<details id="all-yaml-options">
 
-<summary>All yaml options (click me)</summary>
+<summary>All yaml options (click to expand)</summary>
 
 ```yaml
 # This is the huggingface model that contains *.pt, *.safetensors, or *.bin files
@@ -517,8 +577,8 @@ base_model_ignore_patterns:
 # You can set that here, or leave this empty to default to base_model
 base_model_config: ./llama-7b-hf
 # You can specify to choose a specific model revision from huggingface hub
-model_revision:
-# Optional tokenizer configuration override in case you want to use a different tokenizer
+revision_of_model:
+# Optional tokenizer configuration path in case you want to use a different tokenizer
 # than the one defined in the base model
 tokenizer_config:
 # If you want to specify the type of model to load, AutoModelForCausalLM is a good choice too
@@ -535,15 +595,16 @@ tokenizer_legacy:
 # This is reported to improve training speed on some models
 resize_token_embeddings_to_32x:
 
+# (Internal use only)
 # Used to identify which the model is based on
 is_falcon_derived_model:
 is_llama_derived_model:
+is_qwen_derived_model:
 # Please note that if you set this to true, `padding_side` will be set to "left" by default
 is_mistral_derived_model:
-is_qwen_derived_model:
 
 # optional overrides to the base model configuration
-model_config:
+overrides_of_model_config:
   # RoPE Scaling https://github.com/huggingface/transformers/pull/24653
   rope_scaling:
     type: # linear | dynamic
@@ -560,8 +621,6 @@ bnb_config_kwargs:
 
 # Whether you are training a 4-bit GPTQ quantized model
 gptq: true
-gptq_groupsize: 128 # group size
-gptq_model_v1: false # v1 or v2
 
 # This will attempt to quantize the model down to 8 bits and use adam 8 bit optimizer
 load_in_8bit: true
@@ -597,9 +656,13 @@ datasets:
     train_on_split: train # Optional[str] name of dataset split to load from
 
     # Optional[str] fastchat conversation type, only used with type: sharegpt
-    conversation:  # Options (see Conversation 'name'): https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
+    conversation: # Options (see Conversation 'name'): https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py
     field_human: # Optional[str]. Human key to use for conversation.
     field_model: # Optional[str]. Assistant key to use for conversation.
+    # Add additional keys from your dataset as input or output roles
+    roles:
+      input: # Optional[List[str]]. These will be masked based on train_on_input
+      output: # Optional[List[str]].
 
   # Custom user instruction prompt
   - path: repo
@@ -624,6 +687,10 @@ datasets:
       # For `completion` datsets only, uses the provided field instead of `text` column
       field:
 
+# If false, the datasets will not be shuffled and will keep their original order in `datasets`.
+# The same applies to the `test_datasets` option and the `pretraining_dataset` option. Default is true.
+shuffle_merged_datasets: true
+
 # 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:
@@ -635,7 +702,7 @@ test_datasets:
     data_files:
       - /workspace/data/eval.jsonl
 
-# use RL training: dpo, ipo, kto_pair
+# use RL training: 'dpo', 'ipo', 'kto_pair'
 rl:
 
 # Saves the desired chat template to the tokenizer_config.json for easier inferencing
@@ -655,7 +722,7 @@ dataset_processes: # defaults to os.cpu_count() if not set
 # Only needed if cached dataset is taking too much storage
 dataset_keep_in_memory:
 # push checkpoints to hub
-hub_model_id: # repo path to push finetuned model
+hub_model_id: # private repo path to push finetuned model
 # how to push checkpoints to hub
 # https://huggingface.co/docs/transformers/v4.31.0/en/main_classes/trainer#transformers.TrainingArguments.hub_strategy
 hub_strategy:
@@ -751,6 +818,7 @@ wandb_log_model: # "checkpoint" to log model to wandb Artifacts every `save_step
 # mlflow configuration if you're using it
 mlflow_tracking_uri: # URI to mlflow
 mlflow_experiment_name: # Your experiment name
+hf_mlflow_log_artifacts:  # set to true to copy each saved checkpoint on each save to mlflow artifact registry
 
 # Where to save the full-finetuned model to
 output_dir: ./completed-model
@@ -804,7 +872,7 @@ group_by_length: false
 gradient_checkpointing: false
 # additional kwargs to pass to the trainer for gradient checkpointing
 # gradient_checkpointing_kwargs:
-#   use_reentrant: false
+#   use_reentrant: true
 
 # Stop training after this many evaluation losses have increased in a row
 # https://huggingface.co/transformers/v4.2.2/_modules/transformers/trainer_callback.html#EarlyStoppingCallback
@@ -819,10 +887,6 @@ cosine_constant_lr_ratio: # freeze lr at some percentage of the step, e.g. cosin
 # For one_cycle optim
 lr_div_factor: # Learning rate div factor
 
-# For log_sweep optim
-log_sweep_min_lr:
-log_sweep_max_lr:
-
 # Specify optimizer
 # Valid values are driven by the Transformers OptimizerNames class, see:
 # https://github.com/huggingface/transformers/blob/95b374952dc27d8511541d6f5a4e22c9ec11fb24/src/transformers/training_args.py#L134
@@ -848,7 +912,26 @@ log_sweep_max_lr:
 # - paged_adamw_8bit
 # - paged_lion_32bit
 # - paged_lion_8bit
+# - galore_adamw
+# - galore_adamw_8bit
+# - galore_adafactor
+# - galore_adamw_layerwise
+# - galore_adamw_8bit_layerwise
+# - galore_adafactor_layerwise
 optimizer:
+# Dictionary of arguments to pass to the optimizer
+optim_args:
+# For Galore Optimizers the following optim_args are available
+# rank:  # type: int
+# update_proj_gap  # type: int
+# scale  # type: float
+# proj_type:  # type: str, default = std
+
+# The target modules to optimize, i.e. the module names that you would like to train, right now this is used only for GaLore algorithm
+optim_target_modules:
+# - self_attn  # for llama
+# - mlp
+
 # Specify weight decay
 weight_decay:
 # adamw hyperparams
@@ -1045,6 +1128,10 @@ fsdp_config:
   fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
 ```
 
+##### FSDP + QLoRA
+
+Axolotl supports training with FSDP and QLoRA, see [these docs](docs/fsdp_qlora.qmd) for more information.
+
 ##### Weights & Biases Logging
 
 Make sure your `WANDB_API_KEY` environment variable is set (recommended) or you login to wandb with `wandb login`.
@@ -1106,7 +1193,7 @@ Please use `--sample_packing False` if you have it on and receive the error simi
 
 ### Merge LORA to base
 
-The following command will merge your LORA adapater with your base model.  You can optionally pass the argument `--lora_model_dir` to specify the directory where your LORA adapter was saved, otherwhise, this will be inferred from `output_dir` in your axolotl config file.  The merged model is saved in the sub-directory `{lora_model_dir}/merged`.
+The following command will merge your LORA adapater with your base model. You can optionally pass the argument `--lora_model_dir` to specify the directory where your LORA adapter was saved, otherwhise, this will be inferred from `output_dir` in your axolotl config file.  The merged model is saved in the sub-directory `{lora_model_dir}/merged`.
 
 ```bash
 python3 -m axolotl.cli.merge_lora your_config.yml --lora_model_dir="./completed-model"
@@ -1122,7 +1209,7 @@ although this will be very slow, and using the config options above are recommen
 
 ## Common Errors 🧰
 
-See also the [FAQ's](./docs/faq.md) and [debugging guide](docs/debugging.md).
+See also the [FAQ's](./docs/faq.qmd) and [debugging guide](docs/debugging.qmd).
 
 > If you encounter a 'Cuda out of memory' error, it means your GPU ran out of memory during the training process. Here's how to resolve it:
 
@@ -1156,7 +1243,7 @@ It's safe to ignore it.
 
 > NCCL Timeouts during training
 
-See the [NCCL](docs/nccl.md) guide.
+See the [NCCL](docs/nccl.qmd) guide.
 
 
 ### Tokenization Mismatch b/w Inference & Training
@@ -1167,14 +1254,14 @@ If you decode a prompt constructed by axolotl, you might see spaces between toke
 
 1. Materialize some data using `python -m axolotl.cli.preprocess your_config.yml --debug`, and then decode the first few rows with your model's tokenizer.
 2. During inference, right before you pass a tensor of token ids to your model, decode these tokens back into a string.
-3. Make sure the inference string from #2 looks **exactly** like the data you fine tuned on from #1, including spaces and new lines.  If they aren't the same adjust your inference server accordingly.
+3. Make sure the inference string from #2 looks **exactly** like the data you fine tuned on from #1, including spaces and new lines.  If they aren't the same, adjust your inference server accordingly.
 4. As an additional troubleshooting step, you can look at the token ids between 1 and 2 to make sure they are identical.
 
 Having misalignment between your prompts during training and inference can cause models to perform very poorly, so it is worth checking this.  See [this blog post](https://hamel.dev/notes/llm/05_tokenizer_gotchas.html) for a concrete example.
 
 ## Debugging Axolotl
 
-See [this debugging guide](docs/debugging.md) for tips on debugging Axolotl, along with an example configuration for debugging with VSCode.
+See [this debugging guide](docs/debugging.qmd) for tips on debugging Axolotl, along with an example configuration for debugging with VSCode.
 
 ## Need help? 🙋
 
@@ -1214,11 +1301,20 @@ PRs are **greatly welcome**!
 
 Please run below to setup env
 ```bash
+git clone https://github.com/OpenAccess-AI-Collective/axolotl
+cd axolotl
+
+pip3 install packaging
+pip3 install -e '.[flash-attn,deepspeed]'
+
 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.
@@ -1255,4 +1351,6 @@ consider sponsoring the project via [GitHub Sponsors](https://github.com/sponsor
 
 #### 🥉 Bronze Sponsors - $500/mo
 
+ - [JarvisLabs.ai](https://jarvislabs.ai)
+
 ---

_quarto.yml

@@ -0,0 +1,51 @@
+project:
+  type: website
+
+website:
+  title: "Axolotl"
+  description: "Fine-tuning"
+  favicon: favicon.jpg
+  navbar:
+    title: Axolotl
+    background: dark
+    pinned: false
+    collapse: false
+    tools:
+    - icon: twitter
+      href: https://twitter.com/axolotl_ai
+    - icon: github
+      href: https://github.com/OpenAccess-AI-Collective/axolotl/
+    - icon: discord
+      href: https://discord.gg/7m9sfhzaf3
+
+  sidebar:
+      pinned: true
+      collapse-level: 2
+      style: docked
+      contents:
+        - text: Home
+          href: index.qmd
+        - section: "How-To Guides"
+          contents:
+          # TODO Edit folder structure after we have more docs.
+            - docs/debugging.qmd
+            - docs/multipack.qmd
+            - docs/fdsp_qlora.qmd
+            - docs/input_output.qmd
+            - docs/rlhf.qmd
+            - docs/nccl.qmd
+            - docs/mac.qmd
+            - docs/multi-node.qmd
+        - section: "Reference"
+          contents:
+            - docs/config.qmd
+        - docs/faq.qmd
+
+
+
+
+format:
+  html:
+    theme: materia
+    css: styles.css
+    toc: true

cicd/Dockerfile.jinja

@@ -0,0 +1,40 @@
+FROM winglian/axolotl-base:{{ BASE_TAG }}
+
+ENV TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX"
+ENV AXOLOTL_EXTRAS="{{ AXOLOTL_EXTRAS }}"
+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 }}"
+
+RUN apt-get update && \
+    apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev
+
+WORKDIR /workspace
+
+RUN git clone --depth=1 https://github.com/OpenAccess-AI-Collective/axolotl.git
+
+WORKDIR /workspace/axolotl
+
+RUN git fetch origin +$GITHUB_REF && \
+    git checkout FETCH_HEAD
+
+# If AXOLOTL_EXTRAS is set, append it in brackets
+RUN pip install causal_conv1d
+RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
+        pip install -e .[deepspeed,flash-attn,mamba-ssm,galore,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
+    else \
+        pip install -e .[deepspeed,flash-attn,mamba-ssm,galore] $AXOLOTL_ARGS; \
+    fi
+
+# So we can test the Docker image
+RUN pip install pytest
+
+# fix so that git fetch/pull from remote works
+RUN git config remote.origin.fetch "+refs/heads/*:refs/remotes/origin/*" && \
+    git config --get remote.origin.fetch
+
+# helper for huggingface-login cli
+RUN git config --global credential.helper store

cicd/cicd.sh

@@ -0,0 +1,5 @@
+#!/bin/bash
+
+pytest --ignore=tests/e2e/ /workspace/axolotl/tests/
+pytest /workspace/axolotl/tests/e2e/patched/
+pytest --ignore=tests/e2e/patched/ /workspace/axolotl/tests/e2e/

cicd/tests.py

@@ -0,0 +1,75 @@
+"""
+ modal application to run axolotl gpu tests in Modal
+ """
+import os
+import pathlib
+import tempfile
+
+import jinja2
+import modal
+from jinja2 import select_autoescape
+from modal import Image, Stub
+
+cicd_path = pathlib.Path(__file__).parent.resolve()
+
+template_loader = jinja2.FileSystemLoader(searchpath=cicd_path)
+template_env = jinja2.Environment(
+    loader=template_loader, autoescape=select_autoescape()
+)
+df_template = template_env.get_template("Dockerfile.jinja")
+
+df_args = {
+    "AXOLOTL_EXTRAS": os.environ.get("AXOLOTL_EXTRAS", ""),
+    "AXOLOTL_ARGS": os.environ.get("AXOLOTL_ARGS", ""),
+    "PYTORCH_VERSION": os.environ.get("PYTORCH_VERSION", "2.0.1"),
+    "BASE_TAG": os.environ.get("BASE_TAG", "main-base-py3.10-cu118-2.0.1"),
+    "CUDA": os.environ.get("CUDA", "118"),
+    "GITHUB_REF": os.environ.get("GITHUB_REF", "refs/heads/main"),
+    "GITHUB_SHA": os.environ.get("GITHUB_SHA", ""),
+}
+
+dockerfile_contents = df_template.render(**df_args)
+
+temp_dir = tempfile.mkdtemp()
+with open(pathlib.Path(temp_dir) / "Dockerfile", "w", encoding="utf-8") as f:
+    f.write(dockerfile_contents)
+
+cicd_image = (
+    Image.from_dockerfile(
+        pathlib.Path(temp_dir) / "Dockerfile",
+        force_build=True,
+        gpu="A10G",
+    )
+    .env(df_args)
+    .pip_install("fastapi==0.110.0", "pydantic==2.6.3")
+)
+
+stub = Stub("Axolotl CI/CD", secrets=[])
+
+
+N_GPUS = int(os.environ.get("N_GPUS", 1))
+GPU_CONFIG = modal.gpu.A10G(count=N_GPUS)
+
+
+def run_cmd(cmd: str, run_folder: str):
+    import subprocess  # nosec
+
+    # Propagate errors from subprocess.
+    if exit_code := subprocess.call(cmd.split(), cwd=run_folder):  # nosec
+        exit(exit_code)  # pylint: disable=consider-using-sys-exit
+
+
+@stub.function(
+    image=cicd_image,
+    gpu=GPU_CONFIG,
+    timeout=45 * 60,
+    cpu=8.0,
+    memory=131072,
+)
+def cicd_pytest():
+    run_cmd("./cicd/cicd.sh", "/workspace/axolotl")
+
+
+@stub.local_entrypoint()
+def main():
+    cicd_pytest.remote()

deepspeed_configs/zero1.json

@@ -16,6 +16,7 @@
     "min_loss_scale": 1
   },
   "gradient_accumulation_steps": "auto",
+  "gradient_clipping": "auto",
   "train_batch_size": "auto",
   "train_micro_batch_size_per_gpu": "auto",
   "wall_clock_breakdown": false

deepspeed_configs/zero2.json

@@ -20,6 +20,7 @@
     "min_loss_scale": 1
   },
   "gradient_accumulation_steps": "auto",
+  "gradient_clipping": "auto",
   "train_batch_size": "auto",
   "train_micro_batch_size_per_gpu": "auto",
   "wall_clock_breakdown": false

deepspeed_configs/zero3.json

@@ -24,6 +24,7 @@
     "min_loss_scale": 1
   },
   "gradient_accumulation_steps": "auto",
+  "gradient_clipping": "auto",
   "train_batch_size": "auto",
   "train_micro_batch_size_per_gpu": "auto",
   "wall_clock_breakdown": false

deepspeed_configs/zero3_bf16.json

@@ -24,6 +24,7 @@
     "min_loss_scale": 1
   },
   "gradient_accumulation_steps": "auto",
+  "gradient_clipping": "auto",
   "train_batch_size": "auto",
   "train_micro_batch_size_per_gpu": "auto",
   "wall_clock_breakdown": false

devtools/README.md

@@ -1 +1 @@
-This directory contains example config files that might be useful for debugging. Please see [docs/debugging.md](../docs/debugging.md) for more information.
+This directory contains example config files that might be useful for debugging. Please see [docs/debugging.qmd](../docs/debugging.qmd) for more information.

docker/Dockerfile

@@ -3,9 +3,10 @@ FROM winglian/axolotl-base:$BASE_TAG
 
 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="2.0.1"
+ARG PYTORCH_VERSION="2.1.2"
 
 ENV PYTORCH_VERSION=$PYTORCH_VERSION
 
@@ -19,10 +20,11 @@ RUN git clone --depth=1 https://github.com/OpenAccess-AI-Collective/axolotl.git
 WORKDIR /workspace/axolotl
 
 # If AXOLOTL_EXTRAS is set, append it in brackets
+RUN pip install causal_conv1d
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        pip install -e .[deepspeed,flash-attn,mamba-ssm,$AXOLOTL_EXTRAS]; \
+        pip install -e .[deepspeed,flash-attn,mamba-ssm,galore,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \
-        pip install -e .[deepspeed,flash-attn,mamba-ssm]; \
+        pip install -e .[deepspeed,flash-attn,mamba-ssm,galore] $AXOLOTL_ARGS; \
     fi
 
 # So we can test the Docker image

docker/Dockerfile-base

@@ -7,8 +7,8 @@ FROM nvidia/cuda:$CUDA_VERSION-cudnn$CUDNN_VERSION-devel-ubuntu$UBUNTU_VERSION a
 
 ENV PATH="/root/miniconda3/bin:${PATH}"
 
-ARG PYTHON_VERSION="3.9"
-ARG PYTORCH_VERSION="2.0.1"
+ARG PYTHON_VERSION="3.10"
+ARG PYTORCH_VERSION="2.1.2"
 ARG CUDA="118"
 ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 9.0+PTX"
 

docker/Dockerfile-tests

@@ -3,9 +3,10 @@ FROM winglian/axolotl-base:$BASE_TAG
 
 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="2.0.1"
+ARG PYTORCH_VERSION="2.1.2"
 ARG GITHUB_REF="main"
 
 ENV PYTORCH_VERSION=$PYTORCH_VERSION
@@ -24,9 +25,9 @@ RUN git fetch origin +$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]; \
+        pip install -e .[deepspeed,flash-attn,mamba-ssm,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \
-        pip install -e .[deepspeed,flash-attn,mamba-ssm]; \
+        pip install -e .[deepspeed,flash-attn,mamba-ssm] $AXOLOTL_ARGS; \
     fi
 
 # So we can test the Docker image

docs/.gitignore

@@ -0,0 +1,2 @@
+/.quarto/
+_site/

docs/config.qmd

@@ -0,0 +1,17 @@
+---
+title: Config options
+description: A complete list of all configuration options.
+---
+
+```{python}
+#|echo: false
+#|output: asis
+import re
+# Regex pattern to match the YAML block including its code fence
+pattern = r'<details[^>]*id="all-yaml-options"[^>]*>.*?<summary>All yaml options.*?```yaml(.*?)```.*?</details>'
+
+with open('../README.md', 'r') as f:
+    doc = f.read()
+match = re.search(pattern, doc, re.DOTALL)
+print("```yaml", match.group(1).strip(), "```", sep="\n")
+```

docs/debugging.qmd

@@ -1,4 +1,8 @@
-# Debugging Axolotl
+---
+title: Debugging
+description: How to debug Axolotl
+---
+
 
 This document provides some tips and tricks for debugging Axolotl.  It also provides an example configuration for debugging with VSCode.  A good debugging setup is essential to understanding how Axolotl code works behind the scenes.
 
@@ -74,7 +78,6 @@ pip3 install -e '.[flash-attn,deepspeed]'
 
 If you developing on a remote host, you can easily use VSCode to debug remotely.  To do so, you will need to follow this [remote - SSH guide](https://code.visualstudio.com/docs/remote/ssh).  You can also see the video below on [Docker and Remote SSH debugging](#video---attaching-to-docker-on-remote-host).
 
-```bash
 
 ### Configuration
 

docs/faq.md

@@ -1,18 +0,0 @@
-# Axolotl FAQ's
-
-
-> The trainer stopped and hasn't progressed in several minutes.
-
-Usually an issue with the GPU's communicating with each other. See the [NCCL doc](../docs/nccl.md)
-
-> Exitcode -9
-
-This usually happens when you run out of system RAM.
-
-> Exitcode -7 while using deepspeed
-
-Try upgrading deepspeed w: `pip install -U deepspeed`
-
-> AttributeError: 'DummyOptim' object has no attribute 'step'
-
-You may be using deepspeed with single gpu. Please don't set `deepspeed:` in yaml or cli.

docs/faq.qmd

@@ -0,0 +1,21 @@
+---
+title: FAQ
+description: Frequently asked questions
+---
+
+
+**Q: The trainer stopped and hasn't progressed in several minutes.**
+
+> A: Usually an issue with the GPUs communicating with each other. See the [NCCL doc](nccl.qmd)
+
+**Q: Exitcode -9**
+
+> A: This usually happens when you run out of system RAM.
+
+**Q: Exitcode -7 while using deepspeed**
+
+> A: Try upgrading deepspeed w: `pip install -U deepspeed`
+
+**Q: AttributeError: 'DummyOptim' object has no attribute 'step'**
+
+> A: You may be using deepspeed with single gpu. Please don't set `deepspeed:` in yaml or cli.

docs/fsdp_qlora.qmd

@@ -0,0 +1,43 @@
+---
+title: FDSP + QLoRA
+description: Use FSDP with QLoRA to fine-tune large LLMs on consumer GPUs.
+format:
+  html:
+    toc: true
+---
+
+## Background
+
+Using FSDP with QLoRA is essential for **fine-tuning larger (70b+ parameter) LLMs on consumer GPUs.**  For example, you can use FSDP + QLoRA to train a 70b model on two 24GB GPUs[^1].
+
+Below, we describe how to use this feature in Axolotl.
+
+## Usage
+
+To enable `QLoRA` with `FSDP`, you need to perform the following steps:
+
+> ![Tip]
+> See the [example config](#example-config) file in addition to reading these instructions.
+
+1. Set `adapter: qlora` in your axolotl config file.
+2. Enable FSDP in your axolotl config, as [described here](https://github.com/OpenAccess-AI-Collective/axolotl?tab=readme-ov-file#fsdp).
+3. Use one of the supported model types: `llama`, `mistral` or `mixtral`.
+
+## Example Config
+
+[examples/llama-2/qlora-fsdp.yml](../examples/llama-2/qlora-fsdp.yml) contains an example of how to enable QLoRA + FSDP in axolotl.
+
+## References
+
+- [PR #1378](https://github.com/OpenAccess-AI-Collective/axolotl/pull/1378) enabling QLoRA in FSDP in Axolotl.
+- [Blog Post](https://www.answer.ai/posts/2024-03-06-fsdp-qlora.html) from the [Answer.AI](https://www.answer.ai/) team describing the work that enabled QLoRA in FSDP.
+- Related HuggingFace PRs Enabling FDSP + QLoRA:
+    - Accelerate [PR#2544](https://github.com/huggingface/accelerate/pull/2544 )
+    - Transformers [PR#29587](https://github.com/huggingface/transformers/pull/29587)
+    - TRL [PR#1416](https://github.com/huggingface/trl/pull/1416)
+    - PEFT [PR#1550](https://github.com/huggingface/peft/pull/1550)
+
+
+
+
+[^1]: This was enabled by [this work](https://www.answer.ai/posts/2024-03-06-fsdp-qlora.html) from the Answer.AI team.

docs/input_output.qmd

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

docs/mac.qmd

@@ -0,0 +1,22 @@
+---
+title: Mac M-series
+description: Mac M-series support
+---
+
+Currently Axolotl on Mac is partially usable, many of the dependencies of Axolotl including Pytorch do not support MPS or have incomplete support.
+
+Current support:
+
+- [x] Support for all models
+- [x] Full training of models
+- [x] LoRA training
+- [x] Sample packing
+- [ ] FP16 and BF16 (awaiting AMP support for MPS in Pytorch)
+- [ ] Tri-dao's flash-attn (until it is supported use spd_attention as an alternative)
+- [ ] xformers
+- [ ] bitsandbytes (meaning no 4/8 bits loading and bnb optimizers)
+- [ ] qlora
+- [ ] DeepSpeed
+
+Untested:
+- FSDP

docs/multi-node.qmd

@@ -1,4 +1,7 @@
-# Multi Node
+---
+title: Multi Node
+description: How to use Axolotl on multiple machines
+---
 
 You will need to create a configuration for accelerate, either by using `accelerate config` and follow the instructions or you can use one of the preset below:
 

docs/multipack.qmd

@@ -1,4 +1,7 @@
-# Multipack (Sample Packing)
+---
+title: Multipack (Sample Packing)
+description: Multipack is a technique to pack multiple sequences into a single batch to increase training throughput.
+---
 
 ## Visualization of Multipack with Flash Attention
 

docs/nccl.qmd

@@ -1,4 +1,7 @@
-# NCCL
+---
+title: NCCL
+description: Troubleshooting NCCL issues
+---
 
 NVIDIA NCCL is a library to facilitate and optimize multi-GPU communication operations, such as broadcast, all-gather, reduce, all-reduce, etc. Broadly, NCCL configuration is highly environment-specific and is configured via several [environment variables](https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/env.html). A common NCCL-related problem occurs when a long-running operation times out causing the training process to abort:
 

docs/rlhf.qmd

@@ -1,4 +1,7 @@
-# RLHF (Beta)
+---
+title: "RLHF (Beta)"
+description: "Reinforcement Learning from Human Feedback is a method whereby a language model is optimized from data using human feedback."
+---
 
 ### Overview
 
@@ -34,6 +37,21 @@ datasets:
 rl: ipo
 ```
 
+#### ORPO
+
+Paper: https://arxiv.org/abs/2403.07691
+
+```yaml
+rl: orpo
+orpo_alpha: 0.1
+remove_unused_columns: false
+
+chat_template: chatml
+datasets:
+  - path: argilla/ultrafeedback-binarized-preferences-cleaned
+    type: orpo.chat_template
+```
+
 #### Using local dataset files
 ```yaml
 datasets:

examples/gemma/qlora.yml

@@ -21,8 +21,9 @@ lora_dropout: 0.05
 lora_target_linear: true
 
 sequence_len: 4096
-sample_packing: false
-pad_to_sequence_len: false
+sample_packing: true
+eval_sample_packing: false
+pad_to_sequence_len: true
 
 wandb_project:
 wandb_entity:

examples/jamba/README.md

@@ -0,0 +1,10 @@
+# Jamba
+
+- ✅ qlora w/ deepspeed Zero-2 needs at least 2x GPUs and
+  - 35GiB VRAM per GPU w minimal context length
+  - 56GiB VRAM per GPU (w multipack enabled)
+- ✅ qlora w/ deepspeed Zero-3 needs at least 2x GPUs and 67GiB VRAM (wtf?)
+- ✅ qlora single-gpu, ~51GiB VRAM
+- ✅ multipack
+- ❓ FSDP
+- ❓ 8-bit LoRA

examples/jamba/qlora.yaml

@@ -0,0 +1,62 @@
+base_model: ai21labs/Jamba-v0.1
+trust_remote_code: true
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.0
+output_dir: ./out
+
+sequence_len: 4096
+sample_packing: false
+pad_to_sequence_len: false
+eval_sample_packing: false
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+adapter: qlora
+lora_r: 8
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+low_cpu_mem_usage: true
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 2
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.00001
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch:
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+special_tokens:

examples/jamba/qlora_deepspeed.yaml

@@ -0,0 +1,62 @@
+base_model: ai21labs/Jamba-v0.1
+trust_remote_code: true
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.0
+output_dir: ./out
+
+sequence_len: 4096
+sample_packing: false
+pad_to_sequence_len: false
+eval_sample_packing: false
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+adapter: qlora
+lora_r: 8
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+low_cpu_mem_usage: true
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 2
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.00001
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch:
+saves_per_epoch: 1
+debug:
+deepspeed: deepspeed_configs/zero2.json
+weight_decay: 0.0
+special_tokens:

examples/llama-2/lisa.yml

@@ -0,0 +1,75 @@
+base_model: NousResearch/Llama-2-7b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+
+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: 0.05
+output_dir: ./lisa-out
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+adapter:
+lora_model_dir:
+lora_r:
+lora_alpha:
+lora_dropout:
+lora_target_linear:
+lora_fan_in_fan_out:
+
+lisa_n_layers: 2
+lisa_step_interval: 20
+lisa_layers_attribute: model.layers
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 5e-5 # recommendation from lisa paper for 7b
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+flash_attn_cross_entropy: false
+flash_attn_rms_norm: true
+flash_attn_fuse_qkv: false
+flash_attn_fuse_mlp: true
+
+warmup_steps: 100
+evals_per_epoch: 4
+eval_table_size:
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.1
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/llama-2/qlora-fsdp.yml

@@ -0,0 +1,76 @@
+base_model: NousResearch/Llama-2-7b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+datasets:
+  - path: yahma/alpaca-cleaned
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.05
+output_dir: ./qlora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 512
+sample_packing: false
+pad_to_sequence_len: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules:
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 4
+num_epochs: 4
+optimizer: adamw_torch
+lr_scheduler: cosine
+learning_rate: 0.00001
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 4
+eval_table_size:
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+  - full_shard
+fsdp_config:
+  fsdp_limit_all_gathers: true
+  fsdp_sync_module_states: true
+  fsdp_offload_params: true
+  fsdp_use_orig_params: false
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+special_tokens:

examples/mistral/Mistral-7b-example/README.md

@@ -1,12 +0,0 @@
-# Description
-This repository presents an in-depth guide for fine-tuning Mistral-7b or any other compatible model using Axolotl, tailored specifically for chatbot development. It streamlines the process of fine-tuning and uploading the enhanced model to HuggingFace 🤗, thereby serving as an invaluable tool for developers in the AI and chatbot domain.
-
-**What’s Inside:**
-
-Beginner-Friendly Instructions: Comprehensive steps to guide you through fine-tuning your chosen model, including details on the data structure (jsonl), configuration, and the code itself.
-
-Hardware Utilized: For reference, the fine-tuning in this guide was performed using 4x NVIDIA GeForce RTX 3090 (rented 2.1.2-cuda12.1-cudnn8-devel).
-
-**Uploading to HuggingFace 🤗:**
-To upload your fine-tuned model to Hugging Face, include the following files:
-![Screenshot 2024-01-19 213932](https://github.com/OpenAccess-AI-Collective/axolotl/assets/138583191/d660eb84-2d76-46a1-9846-cf0aeb3006d9)

examples/mistral/Mistral-7b-example/code.ipynb

@@ -1,970 +0,0 @@
-{
- "cells": [
-  {
-   "cell_type": "code",
-   "execution_count": 1,
-   "id": "3fe31229-8f6b-48bc-a86d-af8e5466d11c",
-   "metadata": {
-    "scrolled": true
-   },
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "GPU available? True\n",
-      "BF16 is supported? True\n"
-     ]
-    }
-   ],
-   "source": [
-    "# Check if GPU is available I used 4x NVIDIA GeForce RTX 3090 (rented 2.1.2-cuda12.1-cudnn8-devel)\n",
-    "import torch\n",
-    "print('GPU available?', torch.cuda.is_available())\n",
-    "print('BF16 is supported?', torch.cuda.is_bf16_supported())"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 2,
-   "id": "1dee845b-f3cb-4b1e-bdd9-1a918eac140b",
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Collecting huggingface_hub\n",
-      "  Downloading huggingface_hub-0.20.1-py3-none-any.whl.metadata (12 kB)\n",
-      "Requirement already satisfied: filelock in /opt/conda/lib/python3.10/site-packages (from huggingface_hub) (3.9.0)\n",
-      "Requirement already satisfied: fsspec>=2023.5.0 in /opt/conda/lib/python3.10/site-packages (from huggingface_hub) (2023.10.0)\n",
-      "Requirement already satisfied: requests in /opt/conda/lib/python3.10/site-packages (from huggingface_hub) (2.31.0)\n",
-      "Requirement already satisfied: tqdm>=4.42.1 in /opt/conda/lib/python3.10/site-packages (from huggingface_hub) (4.65.0)\n",
-      "Requirement already satisfied: pyyaml>=5.1 in /opt/conda/lib/python3.10/site-packages (from huggingface_hub) (6.0.1)\n",
-      "Requirement already satisfied: typing-extensions>=3.7.4.3 in /opt/conda/lib/python3.10/site-packages (from huggingface_hub) (4.7.1)\n",
-      "Requirement already satisfied: packaging>=20.9 in /opt/conda/lib/python3.10/site-packages (from huggingface_hub) (23.1)\n",
-      "Requirement already satisfied: charset-normalizer<4,>=2 in /opt/conda/lib/python3.10/site-packages (from requests->huggingface_hub) (2.0.4)\n",
-      "Requirement already satisfied: idna<4,>=2.5 in /opt/conda/lib/python3.10/site-packages (from requests->huggingface_hub) (3.4)\n",
-      "Requirement already satisfied: urllib3<3,>=1.21.1 in /opt/conda/lib/python3.10/site-packages (from requests->huggingface_hub) (1.26.18)\n",
-      "Requirement already satisfied: certifi>=2017.4.17 in /opt/conda/lib/python3.10/site-packages (from requests->huggingface_hub) (2023.7.22)\n",
-      "Downloading huggingface_hub-0.20.1-py3-none-any.whl (330 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m330.1/330.1 kB\u001b[0m \u001b[31m8.8 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m:00:01\u001b[0m\n",
-      "\u001b[?25hInstalling collected packages: huggingface_hub\n",
-      "Successfully installed huggingface_hub-0.20.1\n",
-      "\u001b[33mWARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv\u001b[0m\u001b[33m\n",
-      "\u001b[0m"
-     ]
-    }
-   ],
-   "source": [
-    "!pip install huggingface_hub"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 4,
-   "id": "88731672-9050-4034-8266-11aaace2a44e",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "from huggingface_hub import notebook_login"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 5,
-   "id": "6b5aa7d7-3b18-4c14-afd4-043c2c545259",
-   "metadata": {},
-   "outputs": [
-    {
-     "data": {
-      "application/vnd.jupyter.widget-view+json": {
-       "model_id": "60df98d7b0294289aad8b6c8cd023c3b",
-       "version_major": 2,
-       "version_minor": 0
-      },
-      "text/plain": [
-       "VBox(children=(HTML(value='<center> <img\\nsrc=https://huggingface.co/front/assets/huggingface_logo-noborder.sv…"
-      ]
-     },
-     "metadata": {},
-     "output_type": "display_data"
-    }
-   ],
-   "source": [
-    "#Login to huggingface so you can push the model to hub later\n",
-    "import sys\n",
-    "stdout = sys.stdout\n",
-    "notebook_login()"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 6,
-   "id": "b74d0635-5033-4494-b7bd-ff6822103d93",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "#I noticed that when you use notebook_login() nothing gets printed after so we use sys \n",
-    "sys.stdout = stdout"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 7,
-   "id": "e3c3b088-45e7-484b-ae39-66beabc48da8",
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Cloning into 'axolotl'...\n",
-      "remote: Enumerating objects: 235, done.\u001b[K\n",
-      "remote: Counting objects: 100% (235/235), done.\u001b[K\n",
-      "remote: Compressing objects: 100% (207/207), done.\u001b[K\n",
-      "remote: Total 235 (delta 48), reused 123 (delta 13), pack-reused 0\u001b[K\n",
-      "Receiving objects: 100% (235/235), 1.46 MiB | 11.65 MiB/s, done.\n",
-      "Resolving deltas: 100% (48/48), done.\n"
-     ]
-    }
-   ],
-   "source": [
-    "#axolotl\n",
-    "!git clone -b main --depth 1 https://github.com/OpenAccess-AI-Collective/axolotl"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 8,
-   "id": "66927751-4fd6-4477-97fc-6ab08c9d9a74",
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "/axolotl\n"
-     ]
-    }
-   ],
-   "source": [
-    "cd axolotl"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 9,
-   "id": "fcccf8da-353b-4d70-8f55-5cfe08c7e6b9",
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Requirement already satisfied: packaging in /opt/conda/lib/python3.10/site-packages (23.1)\n",
-      "\u001b[33mWARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv\u001b[0m\u001b[33m\n",
-      "\u001b[0mObtaining file:///axolotl\n",
-      "  Preparing metadata (setup.py) ... \u001b[?25ldone\n",
-      "\u001b[?25hCollecting auto-gptq==0.5.1\n",
-      "  Downloading auto_gptq-0.5.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (20 kB)\n",
-      "Requirement already satisfied: packaging in /opt/conda/lib/python3.10/site-packages (23.1)\n",
-      "Collecting peft==0.6.0\n",
-      "  Downloading peft-0.6.0-py3-none-any.whl.metadata (23 kB)\n",
-      "Collecting transformers==4.36.2\n",
-      "  Downloading transformers-4.36.2-py3-none-any.whl.metadata (126 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m126.8/126.8 kB\u001b[0m \u001b[31m9.1 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hCollecting tokenizers==0.15.0\n",
-      "  Downloading tokenizers-0.15.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (6.7 kB)\n",
-      "Collecting bitsandbytes>=0.41.1\n",
-      "  Downloading bitsandbytes-0.41.3.post2-py3-none-any.whl.metadata (9.8 kB)\n",
-      "Collecting accelerate==0.24.1\n",
-      "  Downloading accelerate-0.24.1-py3-none-any.whl.metadata (18 kB)\n",
-      "Collecting addict\n",
-      "  Downloading addict-2.4.0-py3-none-any.whl (3.8 kB)\n",
-      "Collecting fire\n",
-      "  Downloading fire-0.5.0.tar.gz (88 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m88.3/88.3 kB\u001b[0m \u001b[31m28.4 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25h  Preparing metadata (setup.py) ... \u001b[?25ldone\n",
-      "\u001b[?25hRequirement already satisfied: PyYAML>=6.0 in /opt/conda/lib/python3.10/site-packages (6.0.1)\n",
-      "Collecting datasets>=2.15.0\n",
-      "  Downloading datasets-2.16.0-py3-none-any.whl.metadata (20 kB)\n",
-      "Collecting sentencepiece\n",
-      "  Downloading sentencepiece-0.1.99-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (1.3 MB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m1.3/1.3 MB\u001b[0m \u001b[31m47.9 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hCollecting wandb\n",
-      "  Downloading wandb-0.16.1-py3-none-any.whl.metadata (9.8 kB)\n",
-      "Collecting einops\n",
-      "  Downloading einops-0.7.0-py3-none-any.whl.metadata (13 kB)\n",
-      "Collecting optimum==1.13.2\n",
-      "  Downloading optimum-1.13.2.tar.gz (300 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m301.0/301.0 kB\u001b[0m \u001b[31m72.3 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25h  Installing build dependencies ... \u001b[?25ldone\n",
-      "\u001b[?25h  Getting requirements to build wheel ... \u001b[?25ldone\n",
-      "\u001b[?25h  Preparing metadata (pyproject.toml) ... \u001b[?25ldone\n",
-      "\u001b[?25hCollecting hf_transfer\n",
-      "  Downloading hf_transfer-0.1.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (1.5 kB)\n",
-      "Collecting colorama\n",
-      "  Downloading colorama-0.4.6-py2.py3-none-any.whl (25 kB)\n",
-      "Collecting numba\n",
-      "  Downloading numba-0.58.1-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl.metadata (2.7 kB)\n",
-      "Requirement already satisfied: numpy>=1.24.4 in /opt/conda/lib/python3.10/site-packages (1.26.0)\n",
-      "Collecting bert-score==0.3.13\n",
-      "  Downloading bert_score-0.3.13-py3-none-any.whl (61 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m61.1/61.1 kB\u001b[0m \u001b[31m20.1 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hCollecting evaluate==0.4.0\n",
-      "  Downloading evaluate-0.4.0-py3-none-any.whl (81 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m81.4/81.4 kB\u001b[0m \u001b[31m26.0 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hCollecting rouge-score==0.1.2\n",
-      "  Downloading rouge_score-0.1.2.tar.gz (17 kB)\n",
-      "  Preparing metadata (setup.py) ... \u001b[?25ldone\n",
-      "\u001b[?25hCollecting scipy\n",
-      "  Downloading scipy-1.11.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (60 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m60.4/60.4 kB\u001b[0m \u001b[31m17.3 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hCollecting scikit-learn==1.2.2\n",
-      "  Downloading scikit_learn-1.2.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (9.6 MB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m9.6/9.6 MB\u001b[0m \u001b[31m83.2 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m00:01\u001b[0m00:01\u001b[0mm\n",
-      "\u001b[?25hCollecting pynvml\n",
-      "  Downloading pynvml-11.5.0-py3-none-any.whl (53 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m53.1/53.1 kB\u001b[0m \u001b[31m13.6 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hCollecting art\n",
-      "  Downloading art-6.1-py3-none-any.whl.metadata (69 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m69.9/69.9 kB\u001b[0m \u001b[31m21.5 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hCollecting fschat==0.2.34\n",
-      "  Downloading fschat-0.2.34-py3-none-any.whl.metadata (20 kB)\n",
-      "Collecting gradio==3.50.2\n",
-      "  Downloading gradio-3.50.2-py3-none-any.whl.metadata (17 kB)\n",
-      "Collecting tensorboard\n",
-      "  Downloading tensorboard-2.15.1-py3-none-any.whl.metadata (1.7 kB)\n",
-      "Collecting s3fs\n",
-      "  Downloading s3fs-2023.12.2-py3-none-any.whl.metadata (1.6 kB)\n",
-      "Collecting gcsfs\n",
-      "  Downloading gcsfs-2023.12.2.post1-py2.py3-none-any.whl.metadata (1.6 kB)\n",
-      "Collecting xformers==0.0.23\n",
-      "  Downloading xformers-0.0.23-cp310-cp310-manylinux2014_x86_64.whl.metadata (1.0 kB)\n",
-      "Collecting deepspeed\n",
-      "  Downloading deepspeed-0.12.6.tar.gz (1.2 MB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m1.2/1.2 MB\u001b[0m \u001b[31m109.7 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25h  Preparing metadata (setup.py) ... \u001b[?25ldone\n",
-      "\u001b[?25hCollecting flash-attn==2.3.3\n",
-      "  Downloading flash_attn-2.3.3.tar.gz (2.3 MB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m2.3/2.3 MB\u001b[0m \u001b[31m111.5 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25h  Preparing metadata (setup.py) ... \u001b[?25ldone\n",
-      "\u001b[?25hRequirement already satisfied: psutil in /opt/conda/lib/python3.10/site-packages (from accelerate==0.24.1) (5.9.0)\n",
-      "Requirement already satisfied: torch>=1.10.0 in /opt/conda/lib/python3.10/site-packages (from accelerate==0.24.1) (2.1.1)\n",
-      "Requirement already satisfied: huggingface-hub in /opt/conda/lib/python3.10/site-packages (from accelerate==0.24.1) (0.20.1)\n",
-      "Collecting rouge (from auto-gptq==0.5.1)\n",
-      "  Downloading rouge-1.0.1-py3-none-any.whl (13 kB)\n",
-      "Collecting gekko (from auto-gptq==0.5.1)\n",
-      "  Downloading gekko-1.0.6-py3-none-any.whl (12.2 MB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m12.2/12.2 MB\u001b[0m \u001b[31m77.9 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m00:01\u001b[0m0:01\u001b[0m\n",
-      "\u001b[?25hCollecting safetensors (from auto-gptq==0.5.1)\n",
-      "  Downloading safetensors-0.4.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (3.8 kB)\n",
-      "Requirement already satisfied: tqdm in /opt/conda/lib/python3.10/site-packages (from auto-gptq==0.5.1) (4.65.0)\n",
-      "Collecting pandas>=1.0.1 (from bert-score==0.3.13)\n",
-      "  Downloading pandas-2.1.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (18 kB)\n",
-      "Requirement already satisfied: requests in /opt/conda/lib/python3.10/site-packages (from bert-score==0.3.13) (2.31.0)\n",
-      "Collecting matplotlib (from bert-score==0.3.13)\n",
-      "  Downloading matplotlib-3.8.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (5.8 kB)\n",
-      "Collecting dill (from evaluate==0.4.0)\n",
-      "  Downloading dill-0.3.7-py3-none-any.whl.metadata (9.9 kB)\n",
-      "Collecting xxhash (from evaluate==0.4.0)\n",
-      "  Downloading xxhash-3.4.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (12 kB)\n",
-      "Collecting multiprocess (from evaluate==0.4.0)\n",
-      "  Downloading multiprocess-0.70.15-py310-none-any.whl.metadata (7.2 kB)\n",
-      "Requirement already satisfied: fsspec>=2021.05.0 in /opt/conda/lib/python3.10/site-packages (from fsspec[http]>=2021.05.0->evaluate==0.4.0) (2023.10.0)\n",
-      "Collecting responses<0.19 (from evaluate==0.4.0)\n",
-      "  Downloading responses-0.18.0-py3-none-any.whl (38 kB)\n",
-      "Collecting ninja (from flash-attn==2.3.3)\n",
-      "  Downloading ninja-1.11.1.1-py2.py3-none-manylinux1_x86_64.manylinux_2_5_x86_64.whl.metadata (5.3 kB)\n",
-      "Collecting aiohttp (from fschat==0.2.34)\n",
-      "  Downloading aiohttp-3.9.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (7.4 kB)\n",
-      "Collecting fastapi (from fschat==0.2.34)\n",
-      "  Downloading fastapi-0.108.0-py3-none-any.whl.metadata (24 kB)\n",
-      "Collecting httpx (from fschat==0.2.34)\n",
-      "  Downloading httpx-0.26.0-py3-none-any.whl.metadata (7.6 kB)\n",
-      "Collecting markdown2[all] (from fschat==0.2.34)\n",
-      "  Downloading markdown2-2.4.12-py2.py3-none-any.whl.metadata (2.0 kB)\n",
-      "Collecting nh3 (from fschat==0.2.34)\n",
-      "  Downloading nh3-0.2.15-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (1.7 kB)\n",
-      "Requirement already satisfied: prompt-toolkit>=3.0.0 in /opt/conda/lib/python3.10/site-packages (from fschat==0.2.34) (3.0.36)\n",
-      "Collecting pydantic<2,>=1 (from fschat==0.2.34)\n",
-      "  Downloading pydantic-1.10.13-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (149 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m149.6/149.6 kB\u001b[0m \u001b[31m42.3 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hCollecting rich>=10.0.0 (from fschat==0.2.34)\n",
-      "  Downloading rich-13.7.0-py3-none-any.whl.metadata (18 kB)\n",
-      "Collecting shortuuid (from fschat==0.2.34)\n",
-      "  Downloading shortuuid-1.0.11-py3-none-any.whl (10 kB)\n",
-      "Collecting tiktoken (from fschat==0.2.34)\n",
-      "  Downloading tiktoken-0.5.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (6.6 kB)\n",
-      "Collecting uvicorn (from fschat==0.2.34)\n",
-      "  Downloading uvicorn-0.25.0-py3-none-any.whl.metadata (6.4 kB)\n",
-      "Collecting aiofiles<24.0,>=22.0 (from gradio==3.50.2)\n",
-      "  Downloading aiofiles-23.2.1-py3-none-any.whl.metadata (9.7 kB)\n",
-      "Collecting altair<6.0,>=4.2.0 (from gradio==3.50.2)\n",
-      "  Downloading altair-5.2.0-py3-none-any.whl.metadata (8.7 kB)\n",
-      "Collecting ffmpy (from gradio==3.50.2)\n",
-      "  Downloading ffmpy-0.3.1.tar.gz (5.5 kB)\n",
-      "  Preparing metadata (setup.py) ... \u001b[?25ldone\n",
-      "\u001b[?25hCollecting gradio-client==0.6.1 (from gradio==3.50.2)\n",
-      "  Downloading gradio_client-0.6.1-py3-none-any.whl.metadata (7.1 kB)\n",
-      "Collecting importlib-resources<7.0,>=1.3 (from gradio==3.50.2)\n",
-      "  Downloading importlib_resources-6.1.1-py3-none-any.whl.metadata (4.1 kB)\n",
-      "Requirement already satisfied: jinja2<4.0 in /opt/conda/lib/python3.10/site-packages (from gradio==3.50.2) (3.1.2)\n",
-      "Requirement already satisfied: markupsafe~=2.0 in /opt/conda/lib/python3.10/site-packages (from gradio==3.50.2) (2.1.1)\n",
-      "Collecting orjson~=3.0 (from gradio==3.50.2)\n",
-      "  Downloading orjson-3.9.10-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (49 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m49.3/49.3 kB\u001b[0m \u001b[31m14.0 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hRequirement already satisfied: pillow<11.0,>=8.0 in /opt/conda/lib/python3.10/site-packages (from gradio==3.50.2) (10.0.1)\n",
-      "Collecting pydub (from gradio==3.50.2)\n",
-      "  Downloading pydub-0.25.1-py2.py3-none-any.whl (32 kB)\n",
-      "Collecting python-multipart (from gradio==3.50.2)\n",
-      "  Downloading python_multipart-0.0.6-py3-none-any.whl (45 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m45.7/45.7 kB\u001b[0m \u001b[31m13.5 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hCollecting semantic-version~=2.0 (from gradio==3.50.2)\n",
-      "  Downloading semantic_version-2.10.0-py2.py3-none-any.whl (15 kB)\n",
-      "Requirement already satisfied: typing-extensions~=4.0 in /opt/conda/lib/python3.10/site-packages (from gradio==3.50.2) (4.7.1)\n",
-      "Collecting websockets<12.0,>=10.0 (from gradio==3.50.2)\n",
-      "  Downloading websockets-11.0.3-cp310-cp310-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl (129 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m129.9/129.9 kB\u001b[0m \u001b[31m30.8 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hCollecting coloredlogs (from optimum==1.13.2)\n",
-      "  Downloading coloredlogs-15.0.1-py2.py3-none-any.whl (46 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m46.0/46.0 kB\u001b[0m \u001b[31m11.6 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hRequirement already satisfied: sympy in /opt/conda/lib/python3.10/site-packages (from optimum==1.13.2) (1.11.1)\n",
-      "Collecting absl-py (from rouge-score==0.1.2)\n",
-      "  Downloading absl_py-2.0.0-py3-none-any.whl.metadata (2.3 kB)\n",
-      "Collecting nltk (from rouge-score==0.1.2)\n",
-      "  Downloading nltk-3.8.1-py3-none-any.whl (1.5 MB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m1.5/1.5 MB\u001b[0m \u001b[31m90.4 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hRequirement already satisfied: six>=1.14.0 in /opt/conda/lib/python3.10/site-packages (from rouge-score==0.1.2) (1.16.0)\n",
-      "Collecting joblib>=1.1.1 (from scikit-learn==1.2.2)\n",
-      "  Downloading joblib-1.3.2-py3-none-any.whl.metadata (5.4 kB)\n",
-      "Collecting threadpoolctl>=2.0.0 (from scikit-learn==1.2.2)\n",
-      "  Downloading threadpoolctl-3.2.0-py3-none-any.whl.metadata (10.0 kB)\n",
-      "Requirement already satisfied: filelock in /opt/conda/lib/python3.10/site-packages (from transformers==4.36.2) (3.9.0)\n",
-      "Collecting regex!=2019.12.17 (from transformers==4.36.2)\n",
-      "  Downloading regex-2023.12.25-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (40 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m40.9/40.9 kB\u001b[0m \u001b[31m12.0 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hRequirement already satisfied: networkx in /opt/conda/lib/python3.10/site-packages (from torch>=1.10.0->accelerate==0.24.1) (3.1)\n",
-      "Collecting pyarrow>=8.0.0 (from datasets>=2.15.0)\n",
-      "  Downloading pyarrow-14.0.2-cp310-cp310-manylinux_2_28_x86_64.whl.metadata (3.0 kB)\n",
-      "Collecting pyarrow-hotfix (from datasets>=2.15.0)\n",
-      "  Downloading pyarrow_hotfix-0.6-py3-none-any.whl.metadata (3.6 kB)\n",
-      "Collecting hjson (from deepspeed)\n",
-      "  Downloading hjson-3.1.0-py3-none-any.whl (54 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m54.0/54.0 kB\u001b[0m \u001b[31m19.4 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hCollecting py-cpuinfo (from deepspeed)\n",
-      "  Downloading py_cpuinfo-9.0.0-py3-none-any.whl (22 kB)\n",
-      "Collecting termcolor (from fire)\n",
-      "  Downloading termcolor-2.4.0-py3-none-any.whl.metadata (6.1 kB)\n",
-      "Requirement already satisfied: decorator>4.1.2 in /opt/conda/lib/python3.10/site-packages (from gcsfs) (5.1.1)\n",
-      "INFO: pip is looking at multiple versions of gcsfs to determine which version is compatible with other requirements. This could take a while.\n",
-      "Collecting gcsfs\n",
-      "  Downloading gcsfs-2023.12.1-py2.py3-none-any.whl.metadata (1.6 kB)\n",
-      "  Downloading gcsfs-2023.12.0-py2.py3-none-any.whl.metadata (1.6 kB)\n",
-      "  Downloading gcsfs-2023.10.0-py2.py3-none-any.whl.metadata (1.6 kB)\n",
-      "Collecting google-auth>=1.2 (from gcsfs)\n",
-      "  Downloading google_auth-2.25.2-py2.py3-none-any.whl.metadata (4.7 kB)\n",
-      "Collecting google-auth-oauthlib (from gcsfs)\n",
-      "  Downloading google_auth_oauthlib-1.2.0-py2.py3-none-any.whl.metadata (2.7 kB)\n",
-      "Collecting google-cloud-storage (from gcsfs)\n",
-      "  Downloading google_cloud_storage-2.14.0-py2.py3-none-any.whl.metadata (6.1 kB)\n",
-      "Collecting llvmlite<0.42,>=0.41.0dev0 (from numba)\n",
-      "  Downloading llvmlite-0.41.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (4.8 kB)\n",
-      "Collecting aiobotocore<3.0.0,>=2.5.4 (from s3fs)\n",
-      "  Downloading aiobotocore-2.9.0-py3-none-any.whl.metadata (20 kB)\n",
-      "INFO: pip is looking at multiple versions of s3fs to determine which version is compatible with other requirements. This could take a while.\n",
-      "Collecting s3fs\n",
-      "  Downloading s3fs-2023.12.1-py3-none-any.whl.metadata (1.6 kB)\n",
-      "  Downloading s3fs-2023.10.0-py3-none-any.whl.metadata (1.6 kB)\n",
-      "Collecting aiobotocore~=2.7.0 (from s3fs)\n",
-      "  Downloading aiobotocore-2.7.0-py3-none-any.whl.metadata (20 kB)\n",
-      "Collecting grpcio>=1.48.2 (from tensorboard)\n",
-      "  Downloading grpcio-1.60.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (4.0 kB)\n",
-      "Collecting markdown>=2.6.8 (from tensorboard)\n",
-      "  Downloading Markdown-3.5.1-py3-none-any.whl.metadata (7.1 kB)\n",
-      "Collecting protobuf<4.24,>=3.19.6 (from tensorboard)\n",
-      "  Downloading protobuf-4.23.4-cp37-abi3-manylinux2014_x86_64.whl.metadata (540 bytes)\n",
-      "Requirement already satisfied: setuptools>=41.0.0 in /opt/conda/lib/python3.10/site-packages (from tensorboard) (68.0.0)\n",
-      "Collecting tensorboard-data-server<0.8.0,>=0.7.0 (from tensorboard)\n",
-      "  Downloading tensorboard_data_server-0.7.2-py3-none-manylinux_2_31_x86_64.whl.metadata (1.1 kB)\n",
-      "Collecting werkzeug>=1.0.1 (from tensorboard)\n",
-      "  Downloading werkzeug-3.0.1-py3-none-any.whl.metadata (4.1 kB)\n",
-      "Requirement already satisfied: Click!=8.0.0,>=7.1 in /opt/conda/lib/python3.10/site-packages (from wandb) (8.1.7)\n",
-      "Collecting GitPython!=3.1.29,>=1.0.0 (from wandb)\n",
-      "  Downloading GitPython-3.1.40-py3-none-any.whl.metadata (12 kB)\n",
-      "Collecting sentry-sdk>=1.0.0 (from wandb)\n",
-      "  Downloading sentry_sdk-1.39.1-py2.py3-none-any.whl.metadata (9.7 kB)\n",
-      "Collecting docker-pycreds>=0.4.0 (from wandb)\n",
-      "  Downloading docker_pycreds-0.4.0-py2.py3-none-any.whl (9.0 kB)\n",
-      "Collecting setproctitle (from wandb)\n",
-      "  Downloading setproctitle-1.3.3-cp310-cp310-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (9.9 kB)\n",
-      "Collecting appdirs>=1.4.3 (from wandb)\n",
-      "  Downloading appdirs-1.4.4-py2.py3-none-any.whl (9.6 kB)\n",
-      "Collecting botocore<1.31.65,>=1.31.16 (from aiobotocore~=2.7.0->s3fs)\n",
-      "  Downloading botocore-1.31.64-py3-none-any.whl.metadata (6.1 kB)\n",
-      "Collecting wrapt<2.0.0,>=1.10.10 (from aiobotocore~=2.7.0->s3fs)\n",
-      "  Downloading wrapt-1.16.0-cp310-cp310-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (6.6 kB)\n",
-      "Collecting aioitertools<1.0.0,>=0.5.1 (from aiobotocore~=2.7.0->s3fs)\n",
-      "  Downloading aioitertools-0.11.0-py3-none-any.whl (23 kB)\n",
-      "Requirement already satisfied: attrs>=17.3.0 in /opt/conda/lib/python3.10/site-packages (from aiohttp->fschat==0.2.34) (23.1.0)\n",
-      "Collecting multidict<7.0,>=4.5 (from aiohttp->fschat==0.2.34)\n",
-      "  Downloading multidict-6.0.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (114 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m114.5/114.5 kB\u001b[0m \u001b[31m37.1 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hCollecting yarl<2.0,>=1.0 (from aiohttp->fschat==0.2.34)\n",
-      "  Downloading yarl-1.9.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (31 kB)\n",
-      "Collecting frozenlist>=1.1.1 (from aiohttp->fschat==0.2.34)\n",
-      "  Downloading frozenlist-1.4.1-cp310-cp310-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (12 kB)\n",
-      "Collecting aiosignal>=1.1.2 (from aiohttp->fschat==0.2.34)\n",
-      "  Downloading aiosignal-1.3.1-py3-none-any.whl (7.6 kB)\n",
-      "Collecting async-timeout<5.0,>=4.0 (from aiohttp->fschat==0.2.34)\n",
-      "  Downloading async_timeout-4.0.3-py3-none-any.whl.metadata (4.2 kB)\n",
-      "Requirement already satisfied: jsonschema>=3.0 in /opt/conda/lib/python3.10/site-packages (from altair<6.0,>=4.2.0->gradio==3.50.2) (4.20.0)\n",
-      "Requirement already satisfied: toolz in /opt/conda/lib/python3.10/site-packages (from altair<6.0,>=4.2.0->gradio==3.50.2) (0.12.0)\n",
-      "Collecting gitdb<5,>=4.0.1 (from GitPython!=3.1.29,>=1.0.0->wandb)\n",
-      "  Downloading gitdb-4.0.11-py3-none-any.whl.metadata (1.2 kB)\n",
-      "Collecting cachetools<6.0,>=2.0.0 (from google-auth>=1.2->gcsfs)\n",
-      "  Downloading cachetools-5.3.2-py3-none-any.whl.metadata (5.2 kB)\n",
-      "Collecting pyasn1-modules>=0.2.1 (from google-auth>=1.2->gcsfs)\n",
-      "  Downloading pyasn1_modules-0.3.0-py2.py3-none-any.whl (181 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m181.3/181.3 kB\u001b[0m \u001b[31m59.8 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hCollecting rsa<5,>=3.1.4 (from google-auth>=1.2->gcsfs)\n",
-      "  Downloading rsa-4.9-py3-none-any.whl (34 kB)\n",
-      "Collecting requests-oauthlib>=0.7.0 (from google-auth-oauthlib->gcsfs)\n",
-      "  Downloading requests_oauthlib-1.3.1-py2.py3-none-any.whl (23 kB)\n",
-      "Collecting contourpy>=1.0.1 (from matplotlib->bert-score==0.3.13)\n",
-      "  Downloading contourpy-1.2.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (5.8 kB)\n",
-      "Collecting cycler>=0.10 (from matplotlib->bert-score==0.3.13)\n",
-      "  Downloading cycler-0.12.1-py3-none-any.whl.metadata (3.8 kB)\n",
-      "Collecting fonttools>=4.22.0 (from matplotlib->bert-score==0.3.13)\n",
-      "  Downloading fonttools-4.47.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (157 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m157.2/157.2 kB\u001b[0m \u001b[31m41.3 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hCollecting kiwisolver>=1.3.1 (from matplotlib->bert-score==0.3.13)\n",
-      "  Downloading kiwisolver-1.4.5-cp310-cp310-manylinux_2_12_x86_64.manylinux2010_x86_64.whl.metadata (6.4 kB)\n",
-      "Collecting pyparsing>=2.3.1 (from matplotlib->bert-score==0.3.13)\n",
-      "  Downloading pyparsing-3.1.1-py3-none-any.whl.metadata (5.1 kB)\n",
-      "Requirement already satisfied: python-dateutil>=2.7 in /opt/conda/lib/python3.10/site-packages (from matplotlib->bert-score==0.3.13) (2.8.2)\n",
-      "Requirement already satisfied: pytz>=2020.1 in /opt/conda/lib/python3.10/site-packages (from pandas>=1.0.1->bert-score==0.3.13) (2023.3.post1)\n",
-      "Collecting tzdata>=2022.1 (from pandas>=1.0.1->bert-score==0.3.13)\n",
-      "  Downloading tzdata-2023.3-py2.py3-none-any.whl (341 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m341.8/341.8 kB\u001b[0m \u001b[31m72.5 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hRequirement already satisfied: wcwidth in /opt/conda/lib/python3.10/site-packages (from prompt-toolkit>=3.0.0->fschat==0.2.34) (0.2.5)\n",
-      "Requirement already satisfied: charset-normalizer<4,>=2 in /opt/conda/lib/python3.10/site-packages (from requests->bert-score==0.3.13) (2.0.4)\n",
-      "Requirement already satisfied: idna<4,>=2.5 in /opt/conda/lib/python3.10/site-packages (from requests->bert-score==0.3.13) (3.4)\n",
-      "Requirement already satisfied: urllib3<3,>=1.21.1 in /opt/conda/lib/python3.10/site-packages (from requests->bert-score==0.3.13) (1.26.18)\n",
-      "Requirement already satisfied: certifi>=2017.4.17 in /opt/conda/lib/python3.10/site-packages (from requests->bert-score==0.3.13) (2023.7.22)\n",
-      "Collecting markdown-it-py>=2.2.0 (from rich>=10.0.0->fschat==0.2.34)\n",
-      "  Downloading markdown_it_py-3.0.0-py3-none-any.whl.metadata (6.9 kB)\n",
-      "Requirement already satisfied: pygments<3.0.0,>=2.13.0 in /opt/conda/lib/python3.10/site-packages (from rich>=10.0.0->fschat==0.2.34) (2.15.1)\n",
-      "Collecting h11>=0.8 (from uvicorn->fschat==0.2.34)\n",
-      "  Downloading h11-0.14.0-py3-none-any.whl (58 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m58.3/58.3 kB\u001b[0m \u001b[31m21.6 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hCollecting humanfriendly>=9.1 (from coloredlogs->optimum==1.13.2)\n",
-      "  Downloading humanfriendly-10.0-py2.py3-none-any.whl (86 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m86.8/86.8 kB\u001b[0m \u001b[31m27.2 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hCollecting starlette<0.33.0,>=0.29.0 (from fastapi->fschat==0.2.34)\n",
-      "  Downloading starlette-0.32.0.post1-py3-none-any.whl.metadata (5.8 kB)\n",
-      "Collecting typing-extensions~=4.0 (from gradio==3.50.2)\n",
-      "  Downloading typing_extensions-4.9.0-py3-none-any.whl.metadata (3.0 kB)\n",
-      "Collecting google-api-core!=2.0.*,!=2.1.*,!=2.2.*,!=2.3.0,<3.0.0dev,>=1.31.5 (from google-cloud-storage->gcsfs)\n",
-      "  Downloading google_api_core-2.15.0-py3-none-any.whl.metadata (2.7 kB)\n",
-      "Collecting google-cloud-core<3.0dev,>=2.3.0 (from google-cloud-storage->gcsfs)\n",
-      "  Downloading google_cloud_core-2.4.1-py2.py3-none-any.whl.metadata (2.7 kB)\n",
-      "Collecting google-resumable-media>=2.6.0 (from google-cloud-storage->gcsfs)\n",
-      "  Downloading google_resumable_media-2.7.0-py2.py3-none-any.whl.metadata (2.2 kB)\n",
-      "Collecting google-crc32c<2.0dev,>=1.0 (from google-cloud-storage->gcsfs)\n",
-      "  Downloading google_crc32c-1.5.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (32 kB)\n",
-      "Requirement already satisfied: anyio in /opt/conda/lib/python3.10/site-packages (from httpx->fschat==0.2.34) (4.2.0)\n",
-      "Collecting httpcore==1.* (from httpx->fschat==0.2.34)\n",
-      "  Downloading httpcore-1.0.2-py3-none-any.whl.metadata (20 kB)\n",
-      "Requirement already satisfied: sniffio in /opt/conda/lib/python3.10/site-packages (from httpx->fschat==0.2.34) (1.3.0)\n",
-      "Collecting wavedrom (from markdown2[all]->fschat==0.2.34)\n",
-      "  Downloading wavedrom-2.0.3.post3.tar.gz (137 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m137.7/137.7 kB\u001b[0m \u001b[31m47.1 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25h  Preparing metadata (setup.py) ... \u001b[?25ldone\n",
-      "\u001b[?25hRequirement already satisfied: mpmath>=0.19 in /opt/conda/lib/python3.10/site-packages (from sympy->optimum==1.13.2) (1.3.0)\n",
-      "Collecting jmespath<2.0.0,>=0.7.1 (from botocore<1.31.65,>=1.31.16->aiobotocore~=2.7.0->s3fs)\n",
-      "  Downloading jmespath-1.0.1-py3-none-any.whl (20 kB)\n",
-      "Collecting smmap<6,>=3.0.1 (from gitdb<5,>=4.0.1->GitPython!=3.1.29,>=1.0.0->wandb)\n",
-      "  Downloading smmap-5.0.1-py3-none-any.whl.metadata (4.3 kB)\n",
-      "Collecting googleapis-common-protos<2.0.dev0,>=1.56.2 (from google-api-core!=2.0.*,!=2.1.*,!=2.2.*,!=2.3.0,<3.0.0dev,>=1.31.5->google-cloud-storage->gcsfs)\n",
-      "  Downloading googleapis_common_protos-1.62.0-py2.py3-none-any.whl.metadata (1.5 kB)\n",
-      "Requirement already satisfied: jsonschema-specifications>=2023.03.6 in /opt/conda/lib/python3.10/site-packages (from jsonschema>=3.0->altair<6.0,>=4.2.0->gradio==3.50.2) (2023.12.1)\n",
-      "Requirement already satisfied: referencing>=0.28.4 in /opt/conda/lib/python3.10/site-packages (from jsonschema>=3.0->altair<6.0,>=4.2.0->gradio==3.50.2) (0.32.0)\n",
-      "Requirement already satisfied: rpds-py>=0.7.1 in /opt/conda/lib/python3.10/site-packages (from jsonschema>=3.0->altair<6.0,>=4.2.0->gradio==3.50.2) (0.15.2)\n",
-      "Collecting mdurl~=0.1 (from markdown-it-py>=2.2.0->rich>=10.0.0->fschat==0.2.34)\n",
-      "  Downloading mdurl-0.1.2-py3-none-any.whl (10.0 kB)\n",
-      "Collecting pyasn1<0.6.0,>=0.4.6 (from pyasn1-modules>=0.2.1->google-auth>=1.2->gcsfs)\n",
-      "  Downloading pyasn1-0.5.1-py2.py3-none-any.whl.metadata (8.6 kB)\n",
-      "Collecting oauthlib>=3.0.0 (from requests-oauthlib>=0.7.0->google-auth-oauthlib->gcsfs)\n",
-      "  Downloading oauthlib-3.2.2-py3-none-any.whl (151 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m151.7/151.7 kB\u001b[0m \u001b[31m50.3 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hRequirement already satisfied: exceptiongroup>=1.0.2 in /opt/conda/lib/python3.10/site-packages (from anyio->httpx->fschat==0.2.34) (1.0.4)\n",
-      "Collecting svgwrite (from wavedrom->markdown2[all]->fschat==0.2.34)\n",
-      "  Downloading svgwrite-1.4.3-py3-none-any.whl (67 kB)\n",
-      "\u001b[2K     \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m67.1/67.1 kB\u001b[0m \u001b[31m21.8 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading accelerate-0.24.1-py3-none-any.whl (261 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m261.4/261.4 kB\u001b[0m \u001b[31m53.4 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading auto_gptq-0.5.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (4.8 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m4.8/4.8 MB\u001b[0m \u001b[31m89.9 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0mta \u001b[36m0:00:01\u001b[0m\n",
-      "\u001b[?25hDownloading fschat-0.2.34-py3-none-any.whl (220 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m220.1/220.1 kB\u001b[0m \u001b[31m63.9 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading gradio-3.50.2-py3-none-any.whl (20.3 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m20.3/20.3 MB\u001b[0m \u001b[31m82.8 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m:00:01\u001b[0m00:01\u001b[0m\n",
-      "\u001b[?25hDownloading peft-0.6.0-py3-none-any.whl (134 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m134.9/134.9 kB\u001b[0m \u001b[31m40.0 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading tokenizers-0.15.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (3.8 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m3.8/3.8 MB\u001b[0m \u001b[31m87.1 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0mta \u001b[36m0:00:01\u001b[0m\n",
-      "\u001b[?25hDownloading transformers-4.36.2-py3-none-any.whl (8.2 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m8.2/8.2 MB\u001b[0m \u001b[31m90.1 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0mta \u001b[36m0:00:01\u001b[0m\n",
-      "\u001b[?25hDownloading xformers-0.0.23-cp310-cp310-manylinux2014_x86_64.whl (213.0 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m213.0/213.0 MB\u001b[0m \u001b[31m36.6 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m00:01\u001b[0m00:01\u001b[0m\n",
-      "\u001b[?25hDownloading gradio_client-0.6.1-py3-none-any.whl (299 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m299.2/299.2 kB\u001b[0m \u001b[31m64.9 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading bitsandbytes-0.41.3.post2-py3-none-any.whl (92.6 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m92.6/92.6 MB\u001b[0m \u001b[31m56.6 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m:00:01\u001b[0m00:01\u001b[0m\n",
-      "\u001b[?25hDownloading datasets-2.16.0-py3-none-any.whl (507 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m507.1/507.1 kB\u001b[0m \u001b[31m87.3 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading scipy-1.11.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (36.4 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m36.4/36.4 MB\u001b[0m \u001b[31m77.4 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m:00:01\u001b[0m00:01\u001b[0m\n",
-      "\u001b[?25hDownloading art-6.1-py3-none-any.whl (599 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m599.8/599.8 kB\u001b[0m \u001b[31m96.0 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading einops-0.7.0-py3-none-any.whl (44 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m44.6/44.6 kB\u001b[0m \u001b[31m13.3 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading gcsfs-2023.10.0-py2.py3-none-any.whl (33 kB)\n",
-      "Downloading hf_transfer-0.1.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (3.9 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m3.9/3.9 MB\u001b[0m \u001b[31m99.7 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m:00:01\u001b[0m\n",
-      "\u001b[?25hDownloading numba-0.58.1-cp310-cp310-manylinux2014_x86_64.manylinux_2_17_x86_64.whl (3.6 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m3.6/3.6 MB\u001b[0m \u001b[31m100.1 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m00:01\u001b[0m\n",
-      "\u001b[?25hDownloading s3fs-2023.10.0-py3-none-any.whl (28 kB)\n",
-      "Downloading tensorboard-2.15.1-py3-none-any.whl (5.5 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m5.5/5.5 MB\u001b[0m \u001b[31m96.9 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0mta \u001b[36m0:00:01\u001b[0m\n",
-      "\u001b[?25hDownloading wandb-0.16.1-py3-none-any.whl (2.1 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m2.1/2.1 MB\u001b[0m \u001b[31m99.4 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading absl_py-2.0.0-py3-none-any.whl (130 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m130.2/130.2 kB\u001b[0m \u001b[31m36.6 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading aiobotocore-2.7.0-py3-none-any.whl (73 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m73.5/73.5 kB\u001b[0m \u001b[31m25.1 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading aiofiles-23.2.1-py3-none-any.whl (15 kB)\n",
-      "Downloading aiohttp-3.9.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (1.2 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m1.2/1.2 MB\u001b[0m \u001b[31m99.6 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading altair-5.2.0-py3-none-any.whl (996 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m996.9/996.9 kB\u001b[0m \u001b[31m110.4 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading dill-0.3.7-py3-none-any.whl (115 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m115.3/115.3 kB\u001b[0m \u001b[31m34.9 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading GitPython-3.1.40-py3-none-any.whl (190 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m190.6/190.6 kB\u001b[0m \u001b[31m47.0 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading google_auth-2.25.2-py2.py3-none-any.whl (184 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m184.2/184.2 kB\u001b[0m \u001b[31m44.9 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading google_auth_oauthlib-1.2.0-py2.py3-none-any.whl (24 kB)\n",
-      "Downloading grpcio-1.60.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (5.4 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m5.4/5.4 MB\u001b[0m \u001b[31m102.4 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0ma \u001b[36m0:00:01\u001b[0m\n",
-      "\u001b[?25hDownloading importlib_resources-6.1.1-py3-none-any.whl (33 kB)\n",
-      "Downloading joblib-1.3.2-py3-none-any.whl (302 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m302.2/302.2 kB\u001b[0m \u001b[31m64.0 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading llvmlite-0.41.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (43.6 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m43.6/43.6 MB\u001b[0m \u001b[31m74.3 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m:00:01\u001b[0m00:01\u001b[0m\n",
-      "\u001b[?25hDownloading Markdown-3.5.1-py3-none-any.whl (102 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m102.2/102.2 kB\u001b[0m \u001b[31m34.4 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading matplotlib-3.8.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (11.6 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m11.6/11.6 MB\u001b[0m \u001b[31m99.9 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m:00:01\u001b[0m0:01\u001b[0m\n",
-      "\u001b[?25hDownloading orjson-3.9.10-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (138 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m138.7/138.7 kB\u001b[0m \u001b[31m38.8 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading pandas-2.1.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (12.3 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m12.3/12.3 MB\u001b[0m \u001b[31m96.4 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m:00:01\u001b[0m0:01\u001b[0m\n",
-      "\u001b[?25hDownloading protobuf-4.23.4-cp37-abi3-manylinux2014_x86_64.whl (304 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m304.5/304.5 kB\u001b[0m \u001b[31m68.5 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading pyarrow-14.0.2-cp310-cp310-manylinux_2_28_x86_64.whl (38.0 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m38.0/38.0 MB\u001b[0m \u001b[31m78.5 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m:00:01\u001b[0m00:01\u001b[0m\n",
-      "\u001b[?25hDownloading pydantic-1.10.13-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (3.1 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m3.1/3.1 MB\u001b[0m \u001b[31m95.6 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading regex-2023.12.25-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (773 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m774.0/774.0 kB\u001b[0m \u001b[31m116.8 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading rich-13.7.0-py3-none-any.whl (240 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m240.6/240.6 kB\u001b[0m \u001b[31m59.9 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading safetensors-0.4.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (1.3 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m1.3/1.3 MB\u001b[0m \u001b[31m102.1 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading sentry_sdk-1.39.1-py2.py3-none-any.whl (254 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m254.1/254.1 kB\u001b[0m \u001b[31m71.9 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading tensorboard_data_server-0.7.2-py3-none-manylinux_2_31_x86_64.whl (6.6 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m6.6/6.6 MB\u001b[0m \u001b[31m104.2 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0ma \u001b[36m0:00:01\u001b[0m\n",
-      "\u001b[?25hDownloading threadpoolctl-3.2.0-py3-none-any.whl (15 kB)\n",
-      "Downloading uvicorn-0.25.0-py3-none-any.whl (60 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m60.3/60.3 kB\u001b[0m \u001b[31m19.1 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading werkzeug-3.0.1-py3-none-any.whl (226 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m226.7/226.7 kB\u001b[0m \u001b[31m67.5 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading fastapi-0.108.0-py3-none-any.whl (92 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m92.0/92.0 kB\u001b[0m \u001b[31m33.8 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading typing_extensions-4.9.0-py3-none-any.whl (32 kB)\n",
-      "Downloading google_cloud_storage-2.14.0-py2.py3-none-any.whl (121 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m121.6/121.6 kB\u001b[0m \u001b[31m36.9 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading httpx-0.26.0-py3-none-any.whl (75 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m75.9/75.9 kB\u001b[0m \u001b[31m24.7 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading httpcore-1.0.2-py3-none-any.whl (76 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m76.9/76.9 kB\u001b[0m \u001b[31m28.4 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading multiprocess-0.70.15-py310-none-any.whl (134 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m134.8/134.8 kB\u001b[0m \u001b[31m48.7 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading nh3-0.2.15-cp37-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (1.7 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m1.7/1.7 MB\u001b[0m \u001b[31m108.1 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading ninja-1.11.1.1-py2.py3-none-manylinux1_x86_64.manylinux_2_5_x86_64.whl (307 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m307.2/307.2 kB\u001b[0m \u001b[31m66.5 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading pyarrow_hotfix-0.6-py3-none-any.whl (7.9 kB)\n",
-      "Downloading setproctitle-1.3.3-cp310-cp310-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl (30 kB)\n",
-      "Downloading termcolor-2.4.0-py3-none-any.whl (7.7 kB)\n",
-      "Downloading tiktoken-0.5.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (2.0 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m2.0/2.0 MB\u001b[0m \u001b[31m101.9 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading xxhash-3.4.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (194 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m194.1/194.1 kB\u001b[0m \u001b[31m44.4 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading async_timeout-4.0.3-py3-none-any.whl (5.7 kB)\n",
-      "Downloading botocore-1.31.64-py3-none-any.whl (11.3 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m11.3/11.3 MB\u001b[0m \u001b[31m98.8 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m:00:01\u001b[0m0:01\u001b[0m\n",
-      "\u001b[?25hDownloading cachetools-5.3.2-py3-none-any.whl (9.3 kB)\n",
-      "Downloading contourpy-1.2.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (310 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m310.7/310.7 kB\u001b[0m \u001b[31m69.7 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading cycler-0.12.1-py3-none-any.whl (8.3 kB)\n",
-      "Downloading fonttools-4.47.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (4.6 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m4.6/4.6 MB\u001b[0m \u001b[31m102.7 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0ma \u001b[36m0:00:01\u001b[0m\n",
-      "\u001b[?25hDownloading frozenlist-1.4.1-cp310-cp310-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl (239 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m239.5/239.5 kB\u001b[0m \u001b[31m71.4 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading gitdb-4.0.11-py3-none-any.whl (62 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m62.7/62.7 kB\u001b[0m \u001b[31m23.3 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading google_api_core-2.15.0-py3-none-any.whl (121 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m122.0/122.0 kB\u001b[0m \u001b[31m32.4 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading google_cloud_core-2.4.1-py2.py3-none-any.whl (29 kB)\n",
-      "Downloading google_resumable_media-2.7.0-py2.py3-none-any.whl (80 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m80.6/80.6 kB\u001b[0m \u001b[31m22.5 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading kiwisolver-1.4.5-cp310-cp310-manylinux_2_12_x86_64.manylinux2010_x86_64.whl (1.6 MB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m1.6/1.6 MB\u001b[0m \u001b[31m102.0 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading markdown_it_py-3.0.0-py3-none-any.whl (87 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m87.5/87.5 kB\u001b[0m \u001b[31m25.0 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading pyparsing-3.1.1-py3-none-any.whl (103 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m103.1/103.1 kB\u001b[0m \u001b[31m32.7 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading starlette-0.32.0.post1-py3-none-any.whl (70 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m70.0/70.0 kB\u001b[0m \u001b[31m19.1 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading wrapt-1.16.0-cp310-cp310-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_17_x86_64.manylinux2014_x86_64.whl (80 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m80.3/80.3 kB\u001b[0m \u001b[31m30.3 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading yarl-1.9.4-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (301 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m301.6/301.6 kB\u001b[0m \u001b[31m80.7 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading markdown2-2.4.12-py2.py3-none-any.whl (41 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m41.2/41.2 kB\u001b[0m \u001b[31m12.5 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading googleapis_common_protos-1.62.0-py2.py3-none-any.whl (228 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m228.7/228.7 kB\u001b[0m \u001b[31m57.7 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading pyasn1-0.5.1-py2.py3-none-any.whl (84 kB)\n",
-      "\u001b[2K   \u001b[90m━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━\u001b[0m \u001b[32m84.9/84.9 kB\u001b[0m \u001b[31m30.5 MB/s\u001b[0m eta \u001b[36m0:00:00\u001b[0m\n",
-      "\u001b[?25hDownloading smmap-5.0.1-py3-none-any.whl (24 kB)\n",
-      "Building wheels for collected packages: flash-attn, optimum, rouge-score, deepspeed, fire, ffmpy, wavedrom\n",
-      "  Building wheel for flash-attn (setup.py) ... \u001b[?25ldone\n",
-      "\u001b[?25h  Created wheel for flash-attn: filename=flash_attn-2.3.3-cp310-cp310-linux_x86_64.whl size=57042553 sha256=b1df92cb5bd7657d38b789dd48e907aa3e0bd2715c817eb85f3c4320bb11fb3f\n",
-      "  Stored in directory: /root/.cache/pip/wheels/e5/e6/fa/941802ec61d1afd320d27160ab1db98e6dba65381f84b76d4a\n",
-      "  Building wheel for optimum (pyproject.toml) ... \u001b[?25ldone\n",
-      "\u001b[?25h  Created wheel for optimum: filename=optimum-1.13.2-py3-none-any.whl size=395599 sha256=ff3a73120e1b6eeeda28f76e3fc8cd4cd826e5d66c869b7848ba150e7af79c62\n",
-      "  Stored in directory: /root/.cache/pip/wheels/6e/b7/2c/79405d98f0943373d8546daeae25a3d377f7659ca0cbe48699\n",
-      "  Building wheel for rouge-score (setup.py) ... \u001b[?25ldone\n",
-      "\u001b[?25h  Created wheel for rouge-score: filename=rouge_score-0.1.2-py3-none-any.whl size=24932 sha256=8118ecbbcd3529085e794c803f0ddb182fc6c6d3e8a494103b49a94abf1bec37\n",
-      "  Stored in directory: /root/.cache/pip/wheels/5f/dd/89/461065a73be61a532ff8599a28e9beef17985c9e9c31e541b4\n",
-      "  Building wheel for deepspeed (setup.py) ... \u001b[?25ldone\n",
-      "\u001b[?25h  Created wheel for deepspeed: filename=deepspeed-0.12.6-py3-none-any.whl size=1306729 sha256=35c46b6f0275b0d3063522e0af4f3cbd9ec1c310114d8917d87cbe2bf43346e2\n",
-      "  Stored in directory: /root/.cache/pip/wheels/a3/dc/a2/f585faaed4dec84108916dcc8e8a7c129a216df8202ca32984\n",
-      "  Building wheel for fire (setup.py) ... \u001b[?25ldone\n",
-      "\u001b[?25h  Created wheel for fire: filename=fire-0.5.0-py2.py3-none-any.whl size=116934 sha256=e76d5185f237f34ec69bb8aa657497bef07408978e4f7efdaef48663bb8cd4ef\n",
-      "  Stored in directory: /root/.cache/pip/wheels/90/d4/f7/9404e5db0116bd4d43e5666eaa3e70ab53723e1e3ea40c9a95\n",
-      "  Building wheel for ffmpy (setup.py) ... \u001b[?25ldone\n",
-      "\u001b[?25h  Created wheel for ffmpy: filename=ffmpy-0.3.1-py3-none-any.whl size=5579 sha256=da3b54dc0ac1a825a1a233315970ac80b8b4c53ebd9cb2a2cfdeab118f453a64\n",
-      "  Stored in directory: /root/.cache/pip/wheels/01/a6/d1/1c0828c304a4283b2c1639a09ad86f83d7c487ef34c6b4a1bf\n",
-      "  Building wheel for wavedrom (setup.py) ... \u001b[?25ldone\n",
-      "\u001b[?25h  Created wheel for wavedrom: filename=wavedrom-2.0.3.post3-py2.py3-none-any.whl size=30052 sha256=7f0cbd15d63ee9c120190bac122ab51bbbfc91ee374bc3c046fadb320816c17e\n",
-      "  Stored in directory: /root/.cache/pip/wheels/9c/52/8c/38b454b42f712f325e26f633287484c7dc1ad469e1580c5954\n",
-      "Successfully built flash-attn optimum rouge-score deepspeed fire ffmpy wavedrom\n",
-      "Installing collected packages: sentencepiece, pydub, py-cpuinfo, ninja, nh3, hjson, ffmpy, bitsandbytes, appdirs, addict, xxhash, wrapt, werkzeug, websockets, tzdata, typing-extensions, threadpoolctl, termcolor, tensorboard-data-server, svgwrite, smmap, shortuuid, setproctitle, sentry-sdk, semantic-version, scipy, safetensors, rouge, regex, python-multipart, pyparsing, pynvml, pyasn1, pyarrow-hotfix, pyarrow, protobuf, orjson, oauthlib, multidict, mdurl, markdown2, markdown, llvmlite, kiwisolver, joblib, jmespath, importlib-resources, humanfriendly, hf_transfer, h11, grpcio, google-crc32c, gekko, frozenlist, fonttools, einops, docker-pycreds, dill, cycler, contourpy, colorama, cachetools, async-timeout, art, aioitertools, aiofiles, absl-py, yarl, wavedrom, uvicorn, tiktoken, scikit-learn, rsa, responses, requests-oauthlib, pydantic, pyasn1-modules, pandas, numba, nltk, multiprocess, matplotlib, markdown-it-py, httpcore, googleapis-common-protos, google-resumable-media, gitdb, fire, coloredlogs, botocore, aiosignal, xformers, tokenizers, starlette, rouge-score, rich, httpx, google-auth, GitPython, flash-attn, deepspeed, aiohttp, accelerate, wandb, transformers, gradio-client, google-auth-oauthlib, google-api-core, fastapi, altair, aiobotocore, tensorboard, s3fs, peft, gradio, google-cloud-core, fschat, datasets, bert-score, optimum, google-cloud-storage, evaluate, auto-gptq, gcsfs, axolotl\n",
-      "  Attempting uninstall: typing-extensions\n",
-      "    Found existing installation: typing_extensions 4.7.1\n",
-      "    Uninstalling typing_extensions-4.7.1:\n",
-      "      Successfully uninstalled typing_extensions-4.7.1\n",
-      "  Running setup.py develop for axolotl\n",
-      "Successfully installed GitPython-3.1.40 absl-py-2.0.0 accelerate-0.24.1 addict-2.4.0 aiobotocore-2.7.0 aiofiles-23.2.1 aiohttp-3.9.1 aioitertools-0.11.0 aiosignal-1.3.1 altair-5.2.0 appdirs-1.4.4 art-6.1 async-timeout-4.0.3 auto-gptq-0.5.1 axolotl-0.3.0 bert-score-0.3.13 bitsandbytes-0.41.3.post2 botocore-1.31.64 cachetools-5.3.2 colorama-0.4.6 coloredlogs-15.0.1 contourpy-1.2.0 cycler-0.12.1 datasets-2.16.0 deepspeed-0.12.6 dill-0.3.7 docker-pycreds-0.4.0 einops-0.7.0 evaluate-0.4.0 fastapi-0.108.0 ffmpy-0.3.1 fire-0.5.0 flash-attn-2.3.3 fonttools-4.47.0 frozenlist-1.4.1 fschat-0.2.34 gcsfs-2023.10.0 gekko-1.0.6 gitdb-4.0.11 google-api-core-2.15.0 google-auth-2.25.2 google-auth-oauthlib-1.2.0 google-cloud-core-2.4.1 google-cloud-storage-2.14.0 google-crc32c-1.5.0 google-resumable-media-2.7.0 googleapis-common-protos-1.62.0 gradio-3.50.2 gradio-client-0.6.1 grpcio-1.60.0 h11-0.14.0 hf_transfer-0.1.4 hjson-3.1.0 httpcore-1.0.2 httpx-0.26.0 humanfriendly-10.0 importlib-resources-6.1.1 jmespath-1.0.1 joblib-1.3.2 kiwisolver-1.4.5 llvmlite-0.41.1 markdown-3.5.1 markdown-it-py-3.0.0 markdown2-2.4.12 matplotlib-3.8.2 mdurl-0.1.2 multidict-6.0.4 multiprocess-0.70.15 nh3-0.2.15 ninja-1.11.1.1 nltk-3.8.1 numba-0.58.1 oauthlib-3.2.2 optimum-1.13.2 orjson-3.9.10 pandas-2.1.4 peft-0.6.0 protobuf-4.23.4 py-cpuinfo-9.0.0 pyarrow-14.0.2 pyarrow-hotfix-0.6 pyasn1-0.5.1 pyasn1-modules-0.3.0 pydantic-1.10.13 pydub-0.25.1 pynvml-11.5.0 pyparsing-3.1.1 python-multipart-0.0.6 regex-2023.12.25 requests-oauthlib-1.3.1 responses-0.18.0 rich-13.7.0 rouge-1.0.1 rouge-score-0.1.2 rsa-4.9 s3fs-2023.10.0 safetensors-0.4.1 scikit-learn-1.2.2 scipy-1.11.4 semantic-version-2.10.0 sentencepiece-0.1.99 sentry-sdk-1.39.1 setproctitle-1.3.3 shortuuid-1.0.11 smmap-5.0.1 starlette-0.32.0.post1 svgwrite-1.4.3 tensorboard-2.15.1 tensorboard-data-server-0.7.2 termcolor-2.4.0 threadpoolctl-3.2.0 tiktoken-0.5.2 tokenizers-0.15.0 transformers-4.36.2 typing-extensions-4.8.0 tzdata-2023.3 uvicorn-0.25.0 wandb-0.16.1 wavedrom-2.0.3.post3 websockets-11.0.3 werkzeug-3.0.1 wrapt-1.16.0 xformers-0.0.23 xxhash-3.4.1 yarl-1.9.4\n",
-      "\u001b[33mWARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv\u001b[0m\u001b[33m\n",
-      "\u001b[0mCollecting git+https://github.com/huggingface/peft.git\n",
-      "  Cloning https://github.com/huggingface/peft.git to /tmp/pip-req-build-hka8xgk2\n",
-      "  Running command git clone --filter=blob:none --quiet https://github.com/huggingface/peft.git /tmp/pip-req-build-hka8xgk2\n",
-      "  Resolved https://github.com/huggingface/peft.git to commit cf04d0353f0343cbf66627228c4495f51669af34\n",
-      "  Installing build dependencies ... \u001b[?25ldone\n",
-      "\u001b[?25h  Getting requirements to build wheel ... \u001b[?25ldone\n",
-      "\u001b[?25h  Preparing metadata (pyproject.toml) ... \u001b[?25ldone\n",
-      "\u001b[?25hRequirement already satisfied: numpy>=1.17 in /opt/conda/lib/python3.10/site-packages (from peft==0.7.2.dev0) (1.26.0)\n",
-      "Requirement already satisfied: packaging>=20.0 in /opt/conda/lib/python3.10/site-packages (from peft==0.7.2.dev0) (23.1)\n",
-      "Requirement already satisfied: psutil in /opt/conda/lib/python3.10/site-packages (from peft==0.7.2.dev0) (5.9.0)\n",
-      "Requirement already satisfied: pyyaml in /opt/conda/lib/python3.10/site-packages (from peft==0.7.2.dev0) (6.0.1)\n",
-      "Requirement already satisfied: torch>=1.13.0 in /opt/conda/lib/python3.10/site-packages (from peft==0.7.2.dev0) (2.1.1)\n",
-      "Requirement already satisfied: transformers in /opt/conda/lib/python3.10/site-packages (from peft==0.7.2.dev0) (4.36.2)\n",
-      "Requirement already satisfied: tqdm in /opt/conda/lib/python3.10/site-packages (from peft==0.7.2.dev0) (4.65.0)\n",
-      "Requirement already satisfied: accelerate>=0.21.0 in /opt/conda/lib/python3.10/site-packages (from peft==0.7.2.dev0) (0.24.1)\n",
-      "Requirement already satisfied: safetensors in /opt/conda/lib/python3.10/site-packages (from peft==0.7.2.dev0) (0.4.1)\n",
-      "Requirement already satisfied: huggingface-hub>=0.17.0 in /opt/conda/lib/python3.10/site-packages (from peft==0.7.2.dev0) (0.20.1)\n",
-      "Requirement already satisfied: filelock in /opt/conda/lib/python3.10/site-packages (from huggingface-hub>=0.17.0->peft==0.7.2.dev0) (3.9.0)\n",
-      "Requirement already satisfied: fsspec>=2023.5.0 in /opt/conda/lib/python3.10/site-packages (from huggingface-hub>=0.17.0->peft==0.7.2.dev0) (2023.10.0)\n",
-      "Requirement already satisfied: requests in /opt/conda/lib/python3.10/site-packages (from huggingface-hub>=0.17.0->peft==0.7.2.dev0) (2.31.0)\n",
-      "Requirement already satisfied: typing-extensions>=3.7.4.3 in /opt/conda/lib/python3.10/site-packages (from huggingface-hub>=0.17.0->peft==0.7.2.dev0) (4.8.0)\n",
-      "Requirement already satisfied: sympy in /opt/conda/lib/python3.10/site-packages (from torch>=1.13.0->peft==0.7.2.dev0) (1.11.1)\n",
-      "Requirement already satisfied: networkx in /opt/conda/lib/python3.10/site-packages (from torch>=1.13.0->peft==0.7.2.dev0) (3.1)\n",
-      "Requirement already satisfied: jinja2 in /opt/conda/lib/python3.10/site-packages (from torch>=1.13.0->peft==0.7.2.dev0) (3.1.2)\n",
-      "Requirement already satisfied: regex!=2019.12.17 in /opt/conda/lib/python3.10/site-packages (from transformers->peft==0.7.2.dev0) (2023.12.25)\n",
-      "Requirement already satisfied: tokenizers<0.19,>=0.14 in /opt/conda/lib/python3.10/site-packages (from transformers->peft==0.7.2.dev0) (0.15.0)\n",
-      "Requirement already satisfied: MarkupSafe>=2.0 in /opt/conda/lib/python3.10/site-packages (from jinja2->torch>=1.13.0->peft==0.7.2.dev0) (2.1.1)\n",
-      "Requirement already satisfied: charset-normalizer<4,>=2 in /opt/conda/lib/python3.10/site-packages (from requests->huggingface-hub>=0.17.0->peft==0.7.2.dev0) (2.0.4)\n",
-      "Requirement already satisfied: idna<4,>=2.5 in /opt/conda/lib/python3.10/site-packages (from requests->huggingface-hub>=0.17.0->peft==0.7.2.dev0) (3.4)\n",
-      "Requirement already satisfied: urllib3<3,>=1.21.1 in /opt/conda/lib/python3.10/site-packages (from requests->huggingface-hub>=0.17.0->peft==0.7.2.dev0) (1.26.18)\n",
-      "Requirement already satisfied: certifi>=2017.4.17 in /opt/conda/lib/python3.10/site-packages (from requests->huggingface-hub>=0.17.0->peft==0.7.2.dev0) (2023.7.22)\n",
-      "Requirement already satisfied: mpmath>=0.19 in /opt/conda/lib/python3.10/site-packages (from sympy->torch>=1.13.0->peft==0.7.2.dev0) (1.3.0)\n",
-      "Building wheels for collected packages: peft\n",
-      "  Building wheel for peft (pyproject.toml) ... \u001b[?25ldone\n",
-      "\u001b[?25h  Created wheel for peft: filename=peft-0.7.2.dev0-py3-none-any.whl size=169456 sha256=4c70d23e759fa6abb3827fb2f3a8683be3b24d78777d0f403bbc2c0548e5dd4b\n",
-      "  Stored in directory: /tmp/pip-ephem-wheel-cache-my5ncou6/wheels/d7/c7/de/1368fac8590e1b103ddc2ec2a28ad51d83aded1a3830e8a087\n",
-      "Successfully built peft\n",
-      "Installing collected packages: peft\n",
-      "  Attempting uninstall: peft\n",
-      "    Found existing installation: peft 0.6.0\n",
-      "    Uninstalling peft-0.6.0:\n",
-      "      Successfully uninstalled peft-0.6.0\n",
-      "\u001b[31mERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.\n",
-      "axolotl 0.3.0 requires peft==0.6.0, but you have peft 0.7.2.dev0 which is incompatible.\u001b[0m\u001b[31m\n",
-      "\u001b[0mSuccessfully installed peft-0.7.2.dev0\n",
-      "\u001b[33mWARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv\u001b[0m\u001b[33m\n",
-      "\u001b[0m"
-     ]
-    }
-   ],
-   "source": [
-    "#instaling what is needed inside axolotl file\n",
-    "!pip install packaging\n",
-    "!pip install -e '.[flash-attn,deepspeed]'\n",
-    "!pip install -U git+https://github.com/huggingface/peft.git"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 16,
-   "id": "82d1a380-1e87-48fe-89fe-25331326014d",
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "The following values were not passed to `accelerate launch` and had defaults used instead:\n",
-      "\t`--num_processes` was set to a value of `3`\n",
-      "\t\tMore than one GPU was found, enabling multi-GPU training.\n",
-      "\t\tIf this was unintended please pass in `--num_processes=1`.\n",
-      "\t`--num_machines` was set to a value of `1`\n",
-      "\t`--mixed_precision` was set to a value of `'no'`\n",
-      "\t`--dynamo_backend` was set to a value of `'no'`\n",
-      "To avoid this warning pass in values for each of the problematic parameters or run `accelerate config`.\n",
-      "/opt/conda/lib/python3.10/site-packages/transformers/deepspeed.py:23: FutureWarning: transformers.deepspeed module is deprecated and will be removed in a future version. Please import deepspeed modules directly from transformers.integrations\n",
-      "  warnings.warn(\n",
-      "[2023-12-28 15:44:09,979] [INFO] [datasets.<module>:58] [PID:2814] PyTorch version 2.1.1 available.\n",
-      "/opt/conda/lib/python3.10/site-packages/transformers/deepspeed.py:23: FutureWarning: transformers.deepspeed module is deprecated and will be removed in a future version. Please import deepspeed modules directly from transformers.integrations\n",
-      "  warnings.warn(\n",
-      "/opt/conda/lib/python3.10/site-packages/transformers/deepspeed.py:23: FutureWarning: transformers.deepspeed module is deprecated and will be removed in a future version. Please import deepspeed modules directly from transformers.integrations\n",
-      "  warnings.warn(\n",
-      "[2023-12-28 15:44:10,011] [INFO] [datasets.<module>:58] [PID:2812] PyTorch version 2.1.1 available.\n",
-      "[2023-12-28 15:44:10,013] [INFO] [datasets.<module>:58] [PID:2813] PyTorch version 2.1.1 available.\n",
-      "[2023-12-28 15:44:10,805] [INFO] [axolotl.normalize_config:150] [PID:2814] [RANK:2] GPU memory usage baseline: 0.000GB (+0.317GB misc)\u001b[39m\n",
-      "[2023-12-28 15:44:10,830] [INFO] [real_accelerator.py:161:get_accelerator] Setting ds_accelerator to cuda (auto detect)\n",
-      "[2023-12-28 15:44:10,842] [INFO] [axolotl.normalize_config:150] [PID:2813] [RANK:1] GPU memory usage baseline: 0.000GB (+0.317GB misc)\u001b[39m\n",
-      "[2023-12-28 15:44:10,865] [INFO] [real_accelerator.py:161:get_accelerator] Setting ds_accelerator to cuda (auto detect)\n",
-      "[2023-12-28 15:44:10,869] [INFO] [axolotl.normalize_config:150] [PID:2812] [RANK:0] GPU memory usage baseline: 0.000GB (+0.351GB misc)\u001b[39m\n",
-      "[2023-12-28 15:44:10,887] [INFO] [real_accelerator.py:161:get_accelerator] Setting ds_accelerator to cuda (auto detect)\n",
-      "[2023-12-28 15:44:10,961] [INFO] [comm.py:637:init_distributed] cdb=None\n",
-      "[2023-12-28 15:44:10,994] [INFO] [comm.py:637:init_distributed] cdb=None\n",
-      "[2023-12-28 15:44:11,015] [INFO] [comm.py:637:init_distributed] cdb=None\n",
-      "[2023-12-28 15:44:11,015] [INFO] [comm.py:668:init_distributed] Initializing TorchBackend in DeepSpeed with backend nccl\n",
-      "                                 dP            dP   dP \n",
-      "                                 88            88   88 \n",
-      "      .d8888b. dP.  .dP .d8888b. 88 .d8888b. d8888P 88 \n",
-      "      88'  `88  `8bd8'  88'  `88 88 88'  `88   88   88 \n",
-      "      88.  .88  .d88b.  88.  .88 88 88.  .88   88   88 \n",
-      "      `88888P8 dP'  `dP `88888P' dP `88888P'   dP   dP \n",
-      "                                                       \n",
-      "                                                       \n",
-      "\n",
-      "[2023-12-28 15:44:11,412] [DEBUG] [axolotl.load_tokenizer:184] [PID:2812] [RANK:0] EOS: 2 / </s>\u001b[39m\n",
-      "[2023-12-28 15:44:11,412] [DEBUG] [axolotl.load_tokenizer:185] [PID:2812] [RANK:0] BOS: 1 / <s>\u001b[39m\n",
-      "[2023-12-28 15:44:11,412] [DEBUG] [axolotl.load_tokenizer:186] [PID:2812] [RANK:0] PAD: 2 / </s>\u001b[39m\n",
-      "[2023-12-28 15:44:11,412] [DEBUG] [axolotl.load_tokenizer:187] [PID:2812] [RANK:0] UNK: 0 / <unk>\u001b[39m\n",
-      "[2023-12-28 15:44:11,413] [INFO] [axolotl.load_tokenized_prepared_datasets:143] [PID:2812] [RANK:0] Loading prepared dataset from disk at tilemachos/GF_new.json/1adc45d2edc1e98ce657814412c6593c...\u001b[39m\n",
-      "[2023-12-28 15:44:11,415] [INFO] [axolotl.load_tokenized_prepared_datasets:145] [PID:2812] [RANK:0] Prepared dataset loaded from disk...\u001b[39m\n",
-      "[2023-12-28 15:44:11,432] [DEBUG] [axolotl.load_tokenizer:184] [PID:2814] [RANK:2] EOS: 2 / </s>\u001b[39m\n",
-      "[2023-12-28 15:44:11,432] [DEBUG] [axolotl.load_tokenizer:185] [PID:2814] [RANK:2] BOS: 1 / <s>\u001b[39m\n",
-      "[2023-12-28 15:44:11,432] [DEBUG] [axolotl.load_tokenizer:186] [PID:2814] [RANK:2] PAD: 2 / </s>\u001b[39m\n",
-      "[2023-12-28 15:44:11,432] [DEBUG] [axolotl.load_tokenizer:187] [PID:2814] [RANK:2] UNK: 0 / <unk>\u001b[39m\n",
-      "[2023-12-28 15:44:11,530] [DEBUG] [axolotl.load_tokenizer:184] [PID:2813] [RANK:1] EOS: 2 / </s>\u001b[39m\n",
-      "[2023-12-28 15:44:11,531] [DEBUG] [axolotl.load_tokenizer:185] [PID:2813] [RANK:1] BOS: 1 / <s>\u001b[39m\n",
-      "[2023-12-28 15:44:11,531] [DEBUG] [axolotl.load_tokenizer:186] [PID:2813] [RANK:1] PAD: 2 / </s>\u001b[39m\n",
-      "[2023-12-28 15:44:11,531] [DEBUG] [axolotl.load_tokenizer:187] [PID:2813] [RANK:1] UNK: 0 / <unk>\u001b[39m\n",
-      "[2023-12-28 15:44:12,158] [INFO] [axolotl.load_tokenized_prepared_datasets:143] [PID:2813] [RANK:1] Loading prepared dataset from disk at tilemachos/GF_new.json/1adc45d2edc1e98ce657814412c6593c...\u001b[39m\n",
-      "[2023-12-28 15:44:12,158] [INFO] [axolotl.load_tokenized_prepared_datasets:143] [PID:2814] [RANK:2] Loading prepared dataset from disk at tilemachos/GF_new.json/1adc45d2edc1e98ce657814412c6593c...\u001b[39m\n",
-      "[2023-12-28 15:44:12,160] [INFO] [axolotl.load_tokenized_prepared_datasets:145] [PID:2813] [RANK:1] Prepared dataset loaded from disk...\u001b[39m\n",
-      "[2023-12-28 15:44:12,161] [INFO] [axolotl.load_tokenized_prepared_datasets:145] [PID:2814] [RANK:2] Prepared dataset loaded from disk...\u001b[39m\n",
-      "[2023-12-28 15:44:12,236] [DEBUG] [axolotl.log:60] [PID:2812] [RANK:0] total_num_tokens: 28120\u001b[39m\n",
-      "[2023-12-28 15:44:12,238] [DEBUG] [axolotl.log:60] [PID:2812] [RANK:0] `total_supervised_tokens: 7990`\u001b[39m\n",
-      "[2023-12-28 15:44:12,238] [DEBUG] [axolotl.log:60] [PID:2812] [RANK:0] total_num_steps: 6\u001b[39m\n",
-      "[2023-12-28 15:44:12,242] [DEBUG] [axolotl.train.log:60] [PID:2812] [RANK:0] loading tokenizer... mistralai/Mistral-7B-v0.1\u001b[39m\n",
-      "[2023-12-28 15:44:12,518] [DEBUG] [axolotl.load_tokenizer:184] [PID:2812] [RANK:0] EOS: 2 / </s>\u001b[39m\n",
-      "[2023-12-28 15:44:12,518] [DEBUG] [axolotl.load_tokenizer:185] [PID:2812] [RANK:0] BOS: 1 / <s>\u001b[39m\n",
-      "[2023-12-28 15:44:12,518] [DEBUG] [axolotl.load_tokenizer:186] [PID:2812] [RANK:0] PAD: 2 / </s>\u001b[39m\n",
-      "[2023-12-28 15:44:12,518] [DEBUG] [axolotl.load_tokenizer:187] [PID:2812] [RANK:0] UNK: 0 / <unk>\u001b[39m\n",
-      "[2023-12-28 15:44:12,518] [DEBUG] [axolotl.train.log:60] [PID:2812] [RANK:0] loading model and peft_config...\u001b[39m\n",
-      "[2023-12-28 15:44:12,589] [DEBUG] [axolotl.load_tokenizer:184] [PID:2814] [RANK:2] EOS: 2 / </s>\u001b[39m\n",
-      "[2023-12-28 15:44:12,589] [DEBUG] [axolotl.load_tokenizer:185] [PID:2814] [RANK:2] BOS: 1 / <s>\u001b[39m\n",
-      "[2023-12-28 15:44:12,589] [DEBUG] [axolotl.load_tokenizer:186] [PID:2814] [RANK:2] PAD: 2 / </s>\u001b[39m\n",
-      "[2023-12-28 15:44:12,589] [DEBUG] [axolotl.load_tokenizer:187] [PID:2814] [RANK:2] UNK: 0 / <unk>\u001b[39m\n",
-      "[2023-12-28 15:44:12,599] [DEBUG] [axolotl.load_tokenizer:184] [PID:2813] [RANK:1] EOS: 2 / </s>\u001b[39m\n",
-      "[2023-12-28 15:44:12,599] [DEBUG] [axolotl.load_tokenizer:185] [PID:2813] [RANK:1] BOS: 1 / <s>\u001b[39m\n",
-      "[2023-12-28 15:44:12,599] [DEBUG] [axolotl.load_tokenizer:186] [PID:2813] [RANK:1] PAD: 2 / </s>\u001b[39m\n",
-      "[2023-12-28 15:44:12,599] [DEBUG] [axolotl.load_tokenizer:187] [PID:2813] [RANK:1] UNK: 0 / <unk>\u001b[39m\n",
-      "[2023-12-28 15:44:13,049] [INFO] [partition_parameters.py:348:__exit__] finished initializing model - num_params = 291, num_elems = 7.24B\n",
-      "Loading checkpoint shards: 100%|██████████████████| 2/2 [00:11<00:00,  5.81s/it]\n",
-      "Loading checkpoint shards: 100%|██████████████████| 2/2 [00:11<00:00,  5.98s/it]\n",
-      "[2023-12-28 15:44:25,395] [INFO] [axolotl.load_model:503] [PID:2813] [RANK:1] GPU memory usage after model load: 7.576GB (+0.524GB cache, +0.708GB misc)\u001b[39m\n",
-      "[2023-12-28 15:44:25,399] [INFO] [axolotl.load_model:526] [PID:2813] [RANK:1] converting PEFT model w/ prepare_model_for_kbit_training\u001b[39m\n",
-      "[2023-12-28 15:44:25,403] [INFO] [axolotl.load_model:538] [PID:2813] [RANK:1] converting modules to torch.bfloat16 for flash attention\u001b[39m\n",
-      "trainable params: 3,407,872 || all params: 7,245,139,968 || trainable%: 0.04703666202518836\n",
-      "[2023-12-28 15:44:25,480] [INFO] [axolotl.load_model:568] [PID:2813] [RANK:1] GPU memory usage after adapters: 7.589GB (+1.501GB cache, +0.708GB misc)\u001b[39m\n",
-      "[2023-12-28 15:44:25,572] [INFO] [axolotl.load_model:503] [PID:2814] [RANK:2] GPU memory usage after model load: 7.576GB (+0.410GB cache, +0.708GB misc)\u001b[39m\n",
-      "[2023-12-28 15:44:25,576] [INFO] [axolotl.load_model:526] [PID:2814] [RANK:2] converting PEFT model w/ prepare_model_for_kbit_training\u001b[39m\n",
-      "[2023-12-28 15:44:25,580] [INFO] [axolotl.load_model:538] [PID:2814] [RANK:2] converting modules to torch.bfloat16 for flash attention\u001b[39m\n",
-      "trainable params: 3,407,872 || all params: 7,245,139,968 || trainable%: 0.04703666202518836\n",
-      "[2023-12-28 15:44:25,660] [INFO] [axolotl.load_model:568] [PID:2814] [RANK:2] GPU memory usage after adapters: 7.589GB (+1.388GB cache, +0.708GB misc)\u001b[39m\n",
-      "Loading checkpoint shards: 100%|██████████████████| 2/2 [00:12<00:00,  6.30s/it]\n",
-      "[2023-12-28 15:44:26,170] [INFO] [axolotl.load_model:503] [PID:2812] [RANK:0] GPU memory usage after model load: 7.576GB (+0.776GB cache, +0.741GB misc)\u001b[39m\n",
-      "[2023-12-28 15:44:26,177] [INFO] [axolotl.load_model:526] [PID:2812] [RANK:0] converting PEFT model w/ prepare_model_for_kbit_training\u001b[39m\n",
-      "[2023-12-28 15:44:26,181] [INFO] [axolotl.load_model:538] [PID:2812] [RANK:0] converting modules to torch.bfloat16 for flash attention\u001b[39m\n",
-      "trainable params: 3,407,872 || all params: 7,245,139,968 || trainable%: 0.04703666202518836\n",
-      "[2023-12-28 15:44:26,259] [INFO] [axolotl.load_model:568] [PID:2812] [RANK:0] GPU memory usage after adapters: 7.589GB (+1.753GB cache, +0.741GB misc)\u001b[39m\n",
-      "[2023-12-28 15:44:26,293] [INFO] [axolotl.train.log:60] [PID:2812] [RANK:0] Pre-saving adapter config to ./out\u001b[39m\n",
-      "[2023-12-28 15:44:26,296] [INFO] [axolotl.train.log:60] [PID:2812] [RANK:0] Starting trainer...\u001b[39m\n",
-      "Using /root/.cache/torch_extensions/py310_cu121 as PyTorch extensions root...\n",
-      "Using /root/.cache/torch_extensions/py310_cu121 as PyTorch extensions root...\n",
-      "Using /root/.cache/torch_extensions/py310_cu121 as PyTorch extensions root...\n",
-      "Detected CUDA files, patching ldflags\n",
-      "Emitting ninja build file /root/.cache/torch_extensions/py310_cu121/fused_adam/build.ninja...\n",
-      "Building extension module fused_adam...\n",
-      "Allowing ninja to set a default number of workers... (overridable by setting the environment variable MAX_JOBS=N)\n",
-      "ninja: no work to do.\n",
-      "Loading extension module fused_adam...\n",
-      "Time to load fused_adam op: 0.05891108512878418 seconds\n",
-      "Loading extension module fused_adam...\n",
-      "Time to load fused_adam op: 0.10173463821411133 seconds\n",
-      "Loading extension module fused_adam...\n",
-      "Time to load fused_adam op: 0.10152459144592285 seconds\n",
-      "/opt/conda/lib/python3.10/site-packages/deepspeed/ops/adam/fused_adam.py:96: UserWarning: The torch.cuda.*DtypeTensor constructors are no longer recommended. It's best to use methods such as torch.tensor(data, dtype=*, device='cuda') to create tensors. (Triggered internally at /opt/conda/conda-bld/pytorch_1699449201336/work/torch/csrc/tensor/python_tensor.cpp:83.)\n",
-      "  self._dummy_overflow_buf = get_accelerator().IntTensor([0])\n",
-      "/opt/conda/lib/python3.10/site-packages/deepspeed/ops/adam/fused_adam.py:96: UserWarning: The torch.cuda.*DtypeTensor constructors are no longer recommended. It's best to use methods such as torch.tensor(data, dtype=*, device='cuda') to create tensors. (Triggered internally at /opt/conda/conda-bld/pytorch_1699449201336/work/torch/csrc/tensor/python_tensor.cpp:83.)\n",
-      "  self._dummy_overflow_buf = get_accelerator().IntTensor([0])\n",
-      "/opt/conda/lib/python3.10/site-packages/deepspeed/ops/adam/fused_adam.py:96: UserWarning: The torch.cuda.*DtypeTensor constructors are no longer recommended. It's best to use methods such as torch.tensor(data, dtype=*, device='cuda') to create tensors. (Triggered internally at /opt/conda/conda-bld/pytorch_1699449201336/work/torch/csrc/tensor/python_tensor.cpp:83.)\n",
-      "  self._dummy_overflow_buf = get_accelerator().IntTensor([0])\n",
-      "Parameter Offload: Total persistent parameters: 3674112 in 193 params\n",
-      "  0%|                                                    | 0/17 [00:00<?, ?it/s]/opt/conda/lib/python3.10/site-packages/torch/utils/checkpoint.py:429: UserWarning: torch.utils.checkpoint: please pass in use_reentrant=True or use_reentrant=False explicitly. The default value of use_reentrant will be updated to be False in the future. To maintain current behavior, pass use_reentrant=True. It is recommended that you use use_reentrant=False. Refer to docs for more details on the differences between the two variants.\n",
-      "  warnings.warn(\n",
-      "/opt/conda/lib/python3.10/site-packages/torch/utils/checkpoint.py:429: UserWarning: torch.utils.checkpoint: please pass in use_reentrant=True or use_reentrant=False explicitly. The default value of use_reentrant will be updated to be False in the future. To maintain current behavior, pass use_reentrant=True. It is recommended that you use use_reentrant=False. Refer to docs for more details on the differences between the two variants.\n",
-      "  warnings.warn(\n",
-      "/opt/conda/lib/python3.10/site-packages/torch/utils/checkpoint.py:429: UserWarning: torch.utils.checkpoint: please pass in use_reentrant=True or use_reentrant=False explicitly. The default value of use_reentrant will be updated to be False in the future. To maintain current behavior, pass use_reentrant=True. It is recommended that you use use_reentrant=False. Refer to docs for more details on the differences between the two variants.\n",
-      "  warnings.warn(\n",
-      "/opt/conda/lib/python3.10/site-packages/bitsandbytes/autograd/_functions.py:322: UserWarning: MatMul8bitLt: inputs will be cast from torch.bfloat16 to float16 during quantization\n",
-      "  warnings.warn(f\"MatMul8bitLt: inputs will be cast from {A.dtype} to float16 during quantization\")\n",
-      "/opt/conda/lib/python3.10/site-packages/bitsandbytes/autograd/_functions.py:322: UserWarning: MatMul8bitLt: inputs will be cast from torch.bfloat16 to float16 during quantization\n",
-      "  warnings.warn(f\"MatMul8bitLt: inputs will be cast from {A.dtype} to float16 during quantization\")\n",
-      "/opt/conda/lib/python3.10/site-packages/bitsandbytes/autograd/_functions.py:322: UserWarning: MatMul8bitLt: inputs will be cast from torch.bfloat16 to float16 during quantization\n",
-      "  warnings.warn(f\"MatMul8bitLt: inputs will be cast from {A.dtype} to float16 during quantization\")\n",
-      "{'loss': 2.0448, 'learning_rate': 2e-05, 'epoch': 0.06}                         \n",
-      "  6%|██▌                                         | 1/17 [00:28<07:32, 28.30s/it]\n",
-      "  0%|                                                     | 0/3 [00:00<?, ?it/s]\u001b[A\n",
-      " 67%|██████████████████████████████               | 2/3 [00:03<00:01,  1.85s/it]\u001b[A\n",
-      "                                                                                \u001b[A\n",
-      "\u001b[A{'eval_loss': 1.9694719314575195, 'eval_runtime': 11.391, 'eval_samples_per_second': 1.492, 'eval_steps_per_second': 0.263, 'epoch': 0.06}\n",
-      "  6%|██▌                                         | 1/17 [00:39<07:32, 28.30s/it]\n",
-      "100%|█████████████████████████████████████████████| 3/3 [00:07<00:00,  2.65s/it]\u001b[A\n",
-      "                                                                                \u001b[A[2023-12-28 15:45:35,358] [INFO] [axolotl.callbacks.on_step_end:122] [PID:2812] [RANK:0] GPU memory usage while training: 12.210GB (+4.259GB cache, +0.776GB misc)\u001b[39m\n",
-      " 12%|█████▏                                      | 2/17 [01:04<08:18, 33.20s/it][2023-12-28 15:45:35,358] [INFO] [axolotl.callbacks.on_step_end:122] [PID:2814] [RANK:2] GPU memory usage while training: 12.269GB (+4.522GB cache, +0.743GB misc)\u001b[39m\n",
-      "[2023-12-28 15:45:35,358] [INFO] [axolotl.callbacks.on_step_end:122] [PID:2813] [RANK:1] GPU memory usage while training: 12.283GB (+4.493GB cache, +0.743GB misc)\u001b[39m\n",
-      "{'loss': 2.0022, 'learning_rate': 4e-05, 'epoch': 0.12}                         \n",
-      "{'loss': 2.1054, 'learning_rate': 6e-05, 'epoch': 0.17}                         \n",
-      "{'loss': 1.9004, 'learning_rate': 8e-05, 'epoch': 0.23}                         \n",
-      "{'loss': 1.8794, 'learning_rate': 0.0001, 'epoch': 0.29}                        \n",
-      " 29%|████████████▉                               | 5/17 [02:20<05:23, 26.92s/it]\n",
-      "  0%|                                                     | 0/3 [00:00<?, ?it/s]\u001b[A\n",
-      " 67%|██████████████████████████████               | 2/3 [00:03<00:01,  1.88s/it]\u001b[A\n",
-      "                                                                                \u001b[A\n",
-      "\u001b[A{'eval_loss': 1.7912336587905884, 'eval_runtime': 11.3106, 'eval_samples_per_second': 1.503, 'eval_steps_per_second': 0.265, 'epoch': 0.29}\n",
-      " 29%|████████████▉                               | 5/17 [02:32<05:23, 26.92s/it]\n",
-      "100%|█████████████████████████████████████████████| 3/3 [00:07<00:00,  2.67s/it]\u001b[A\n",
-      "{'loss': 1.7871, 'learning_rate': 0.00012, 'epoch': 0.35}                       \u001b[A\n",
-      "{'loss': 1.7758, 'learning_rate': 0.00014, 'epoch': 0.4}                        \n",
-      "{'loss': 1.4645, 'learning_rate': 0.00016, 'epoch': 0.46}                       \n",
-      "{'loss': 1.4009, 'learning_rate': 0.00018, 'epoch': 0.52}                       \n",
-      "{'loss': 1.3927, 'learning_rate': 0.0002, 'epoch': 0.58}                        \n",
-      " 59%|█████████████████████████▎                 | 10/17 [04:38<03:04, 26.33s/it]\n",
-      "  0%|                                                     | 0/3 [00:00<?, ?it/s]\u001b[A\n",
-      " 67%|██████████████████████████████               | 2/3 [00:03<00:01,  1.89s/it]\u001b[A\n",
-      "                                                                                \u001b[A\n",
-      "\u001b[A{'eval_loss': 1.1426481008529663, 'eval_runtime': 11.3344, 'eval_samples_per_second': 1.5, 'eval_steps_per_second': 0.265, 'epoch': 0.58}\n",
-      " 59%|█████████████████████████▎                 | 10/17 [04:49<03:04, 26.33s/it]\n",
-      "100%|█████████████████████████████████████████████| 3/3 [00:07<00:00,  2.68s/it]\u001b[A\n",
-      "{'loss': 1.0122, 'learning_rate': 0.0001900968867902419, 'epoch': 0.63}         \u001b[A\n",
-      "{'loss': 1.0019, 'learning_rate': 0.00016234898018587337, 'epoch': 0.69}        \n",
-      "{'loss': 0.8976, 'learning_rate': 0.00012225209339563145, 'epoch': 0.75}        \n",
-      "{'loss': 0.9301, 'learning_rate': 7.774790660436858e-05, 'epoch': 0.81}         \n",
-      "{'loss': 0.8595, 'learning_rate': 3.7651019814126654e-05, 'epoch': 0.87}        \n",
-      " 88%|█████████████████████████████████████▉     | 15/17 [06:55<00:52, 26.17s/it]\n",
-      "  0%|                                                     | 0/3 [00:00<?, ?it/s]\u001b[A\n",
-      " 67%|██████████████████████████████               | 2/3 [00:03<00:01,  1.88s/it]\u001b[A\n",
-      "                                                                                \u001b[A\n",
-      "\u001b[A{'eval_loss': 0.8175248503684998, 'eval_runtime': 11.2932, 'eval_samples_per_second': 1.505, 'eval_steps_per_second': 0.266, 'epoch': 0.87}\n",
-      " 88%|█████████████████████████████████████▉     | 15/17 [07:06<00:52, 26.17s/it]\n",
-      "100%|█████████████████████████████████████████████| 3/3 [00:07<00:00,  2.67s/it]\u001b[A\n",
-      "{'loss': 0.7931, 'learning_rate': 9.903113209758096e-06, 'epoch': 0.92}         \u001b[A\n",
-      "{'loss': 0.6909, 'learning_rate': 0.0, 'epoch': 0.98}                           \n",
-      "100%|███████████████████████████████████████████| 17/17 [07:56<00:00, 28.03s/it]/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py:1879: UserWarning: Positional args are being deprecated, use kwargs instead. Refer to https://pytorch.org/docs/master/generated/torch.nn.Module.html#torch.nn.Module.state_dict for details.\n",
-      "  warnings.warn(\n",
-      "/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py:1879: UserWarning: Positional args are being deprecated, use kwargs instead. Refer to https://pytorch.org/docs/master/generated/torch.nn.Module.html#torch.nn.Module.state_dict for details.\n",
-      "  warnings.warn(\n",
-      "/opt/conda/lib/python3.10/site-packages/torch/nn/modules/module.py:1879: UserWarning: Positional args are being deprecated, use kwargs instead. Refer to https://pytorch.org/docs/master/generated/torch.nn.Module.html#torch.nn.Module.state_dict for details.\n",
-      "  warnings.warn(\n",
-      "{'train_runtime': 489.0649, 'train_samples_per_second': 0.63, 'train_steps_per_second': 0.035, 'train_loss': 1.408153467318591, 'epoch': 0.98}\n",
-      "100%|███████████████████████████████████████████| 17/17 [08:09<00:00, 28.77s/it]\n",
-      "[2023-12-28 15:52:39,488] [INFO] [axolotl.train.log:60] [PID:2812] [RANK:0] Training Completed!!! Saving pre-trained model to ./out\u001b[39m\n",
-      "\u001b[0m\u001b[0m\u001b[0m"
-     ]
-    }
-   ],
-   "source": [
-    "\"\"\"\n",
-    "Training using the config.yml file and using deepspeed:zero3_bf16 the most aggressive optimization out of zero1,zero2,zero3 stages which partitions \n",
-    "not only optimizer states but also gradients and parameters across GPUs. The bf16 indicate mixed precision training using bfloat16.\n",
-    "For more information read axolotl's readme\n",
-    "\"\"\"\n",
-    "!accelerate launch -m axolotl.cli.train /folder/config.yml --deepspeed deepspeed_configs/zero3_bf16.json"
-   ]
-  }
- ],
- "metadata": {
-  "kernelspec": {
-   "display_name": "Python 3 (ipykernel)",
-   "language": "python",
-   "name": "python3"
-  },
-  "language_info": {
-   "codemirror_mode": {
-    "name": "ipython",
-    "version": 3
-   },
-   "file_extension": ".py",
-   "mimetype": "text/x-python",
-   "name": "python",
-   "nbconvert_exporter": "python",
-   "pygments_lexer": "ipython3",
-   "version": "3.10.13"
-  }
- },
- "nbformat": 4,
- "nbformat_minor": 5
-}

examples/mistral/Mistral-7b-example/data.jsonl

@@ -1,10 +0,0 @@
-{"conversations": [{"from": "Customer", "value": "\"<Customer>: Who is the Founder of Apple\""}, {"from": "gpt", "value": "\"<Chatbot>: The founder of Apple is Steve Jobs\""}]}
-{"conversations": [{"from": "Customer", "value": "\"<Customer>: What is the capital of France?\""}, {"from": "gpt", "value": "\"<Chatbot>: The capital of France is Paris.\""}]}
-{"conversations": [{"from": "Customer", "value": "\"<Customer>: How far is the Moon from Earth?\""}, {"from": "gpt", "value": "\"<Chatbot>: The Moon is approximately 384,400 kilometers from Earth.\""}]}
-{"conversations": [{"from": "Customer", "value": "\"<Customer>: What is the tallest mountain in the world?\""}, {"from": "gpt", "value": "\"<Chatbot>: The tallest mountain in the world is Mount Everest.\""}]}
-{"conversations": [{"from": "Customer", "value": "\"<Customer>: Who wrote Romeo and Juliet?\""}, {"from": "gpt", "value": "\"<Chatbot>: Romeo and Juliet was written by William Shakespeare.\""}]}
-{"conversations": [{"from": "Customer", "value": "\"<Customer>: What is the boiling point of water?\""}, {"from": "gpt", "value": "\"<Chatbot>: The boiling point of water is 100 degrees Celsius.\""}]}
-{"conversations": [{"from": "Customer", "value": "\"<Customer>: When was the first man on the moon?\""}, {"from": "gpt", "value": "\"<Chatbot>: The first man landed on the moon in 1969.\""}]}
-{"conversations": [{"from": "Customer", "value": "\"<Customer>: What is the largest ocean?\""}, {"from": "gpt", "value": "\"<Chatbot>: The largest ocean is the Pacific Ocean.\""}]}
-{"conversations": [{"from": "Customer", "value": "\"<Customer>: Who invented the telephone?\""}, {"from": "gpt", "value": "\"<Chatbot>: The telephone was invented by Alexander Graham Bell.\""}]}
-{"conversations": [{"from": "Customer", "value": "\"<Customer>: What is the formula for water?\""}, {"from": "gpt", "value": "\"<Chatbot>: The chemical formula for water is H2O.\""}]}

examples/mistral/config.yml

@@ -56,6 +56,3 @@ weight_decay: 0.0
 fsdp:
 fsdp_config:
 special_tokens:
-  bos_token: "<s>"
-  eos_token: "</s>"
-  unk_token: "<unk>"

examples/mistral/lora-mps.yml

@@ -0,0 +1,79 @@
+base_model: mistralai/Mistral-7B-v0.1
+model_type: MistralForCausalLM
+tokenizer_type: LlamaTokenizer
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0
+output_dir: ./lora-out
+eval_sample_packing: false
+
+adapter: lora
+lora_model_dir:
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 8
+micro_batch_size: 1
+num_epochs: 2
+optimizer: adamw_torch
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16: false
+tf32: true
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: false
+sdp_attention: true
+
+loss_watchdog_threshold: 5.0
+loss_watchdog_patience: 3
+
+warmup_steps: 10
+evals_per_epoch: 4
+eval_table_size:
+eval_table_max_new_tokens: 128
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:

examples/mistral/lora.yml

@@ -1,4 +1,3 @@
-#Mistral-7b
 base_model: mistralai/Mistral-7B-v0.1
 model_type: MistralForCausalLM
 tokenizer_type: LlamaTokenizer
@@ -8,26 +7,32 @@ load_in_4bit: false
 strict: false
 
 datasets:
-  - path: tilemachos/Demo-Dataset #Path to json dataset file in huggingface
-    #for type,conversation arguments read axolotl readme and pick what is suited for your project, I wanted a chatbot and put sharegpt and chatml
-    type: sharegpt
-    conversation: chatml
-dataset_prepared_path: tilemachos/Demo-Dataset #Path to json dataset file in huggingface
-val_set_size: 0.05
-output_dir: ./out
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./lora-out
 
-#using lora for lower cost
 adapter: lora
-lora_r: 8
+lora_model_dir:
+
+sequence_len: 8192
+sample_packing: true
+pad_to_sequence_len: true
+
+lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
 lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
   - q_proj
   - v_proj
-
-sequence_len: 512
-sample_packing: false
-pad_to_sequence_len: true
+  - k_proj
+  - o_proj
 
 wandb_project:
 wandb_entity:
@@ -35,18 +40,17 @@ wandb_watch:
 wandb_name:
 wandb_log_model:
 
-#only 2 epochs because of small dataset
-gradient_accumulation_steps: 3
+gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 2
+num_epochs: 1
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
 train_on_inputs: false
 group_by_length: false
-bf16: true
-fp16: false
+bf16: auto
+fp16:
 tf32: false
 
 gradient_checkpointing: true
@@ -57,18 +61,17 @@ logging_steps: 1
 xformers_attention:
 flash_attention: true
 
+loss_watchdog_threshold: 5.0
+loss_watchdog_patience: 3
+
 warmup_steps: 10
 evals_per_epoch: 4
 eval_table_size:
 eval_max_new_tokens: 128
 saves_per_epoch: 1
 debug:
-#default deepspeed, can use more aggresive if needed like zero2, zero3
-deepspeed: deepspeed_configs/zero1.json
+deepspeed:
 weight_decay: 0.0
 fsdp:
 fsdp_config:
 special_tokens:
-  bos_token: "<s>"
-  eos_token: "</s>"
-  unk_token: "<unk>"

examples/mistral/mixtral-qlora-fsdp.yml

@@ -0,0 +1,74 @@
+base_model: mistralai/Mixtral-8x7B-v0.1
+model_type: AutoModelForCausalLM
+tokenizer_type: LlamaTokenizer
+trust_remote_code: true
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.02
+output_dir: ./qlora-out
+
+model_config:
+  output_router_logits: true
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 1024
+sample_packing: false
+pad_to_sequence_len: false
+
+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: 1
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+loss_watchdog_threshold: 5.0
+loss_watchdog_patience: 3
+
+warmup_steps: 10
+evals_per_epoch: 4
+eval_table_size:
+eval_max_new_tokens: 128
+saves_per_epoch: 1
+debug:
+weight_decay: 0.0
+fsdp:
+  - full_shard
+fsdp_config:
+  fsdp_transformer_layer_cls_to_wrap: MixtralSparseMoeBlock
+special_tokens:

examples/mistral/mixtral.yml

@@ -16,12 +16,12 @@ output_dir: ./qlora-out
 
 ## You can optionally freeze the entire model and unfreeze a subset of parameters
 unfrozen_parameters:
-#  - lm_head.*
-#  - model.embed_tokens.*
-#  - model.layers.2[0-9]+.block_sparse_moe.gate.*
-#  - model.layers.2[0-9]+.block_sparse_moe.experts.*
-#  - model.layers.3[0-9]+.block_sparse_moe.gate.*
-#  - model.layers.3[0-9]+.block_sparse_moe.experts.*
+#  - ^lm_head.weight$
+#  - ^model.embed_tokens.weight$[:32000]
+#  - model.layers.2[0-9]+.block_sparse_moe.gate
+#  - model.layers.2[0-9]+.block_sparse_moe.experts
+#  - model.layers.3[0-9]+.block_sparse_moe.gate
+#  - model.layers.3[0-9]+.block_sparse_moe.experts
 
 model_config:
   output_router_logits: true

examples/mistral/qlora.yml

@@ -75,6 +75,3 @@ weight_decay: 0.0
 fsdp:
 fsdp_config:
 special_tokens:
-  bos_token: "<s>"
-  eos_token: "</s>"
-  unk_token: "<unk>"

examples/qwen/README.md

@@ -0,0 +1,10 @@
+# Qwen
+
+TODO
+
+# Qwen2 MoE
+
+✅ multipack
+✅ qwen2_moe 4-bit QLoRA
+✅ qwen2_moe 16-bit LoRA
+❓ qwen2_moe 8-bit LoRA

examples/qwen/qwen2-moe-lora.yaml

@@ -0,0 +1,64 @@
+base_model: Qwen/Qwen1.5-MoE-A2.7B
+trust_remote_code: true
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./out
+
+sequence_len: 1024  # supports up to 32k
+sample_packing: false
+pad_to_sequence_len: false
+
+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: 1
+num_epochs: 4
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:

examples/qwen/qwen2-moe-qlora.yaml

@@ -0,0 +1,64 @@
+base_model: Qwen/Qwen1.5-MoE-A2.7B
+trust_remote_code: true
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./out
+
+sequence_len: 1024  # supports up to 32k
+sample_packing: false
+pad_to_sequence_len: false
+
+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: 1
+num_epochs: 4
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:

examples/stablelm-2/1.6b/fft.yml

@@ -0,0 +1,69 @@
+base_model: stabilityai/stablelm-2-1_6b
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+trust_remote_code: true
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.05
+output_dir: ./out
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+adapter:
+lora_model_dir:
+lora_r:
+lora_alpha:
+lora_dropout:
+lora_target_linear:
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+flash_attn_cross_entropy: false
+flash_attn_rms_norm: true
+flash_attn_fuse_qkv: false
+flash_attn_fuse_mlp: true
+
+warmup_steps: 100
+evals_per_epoch: 4
+eval_table_size:
+saves_per_epoch: 1
+debug:
+deepspeed: #deepspeed_configs/zero2.json # multi-gpu only
+weight_decay: 0.1
+fsdp:
+fsdp_config:
+special_tokens:

examples/stablelm-2/1.6b/lora.yml

@@ -0,0 +1,66 @@
+base_model: stabilityai/stablelm-2-1_6b
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+trust_remote_code: true
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./lora-out
+
+sequence_len: 4096
+sample_packing: true
+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: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+flash_attn_cross_entropy: false
+flash_attn_rms_norm: true
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:

examples/stablelm-2/README.md

@@ -0,0 +1,36 @@
+# StableLM 2
+
+This repository contains examples for training and processing using StableLM-2. It also includes a section to help you estimate the GPU requirements for your specific use case.
+
+## Estimating GPU Requirements
+
+| type          | deepspeed | batch size | context length | vRAM GPU (GBs) |
+|---------------|-----------|------------|----------------|----------------|
+| full finetune | N/A       | 1          | 4096           | ~21.5GBs       |
+| full finetune | zero2     | 1          | 4096           | ~20GBs         |
+| lora          | N/A       | 1          | 4096           | ~16.6GBs       |
+
+The above are estimates and might differ slight depending on the setup for example whether you pack your sequence lengths or not (the above assumes you do to length 4096).
+
+This blog post from Hamel Husain was a great resource for estimating these numbers: https://hamel.dev/notes/llm/03_estimating_vram.html
+
+## Training
+We have example scripts here for both full finetuning and lora using the popular alpaca dataset:
+
+```shell
+# preprocess the dataset
+CUDA_VISIBLE_DEVICES="" python -m axolotl.cli.preprocess examples/stablelm-2/1.6b/lora.yml
+```
+
+Single GPU Training:
+```shell
+python -m axolotl.cli.train examples/stablelm-2/fft.yml --deepspeed deepspeed_configs/zero2.json
+# OR
+python -m axolotl.cli.train examples/stablelm-2/1.6b/lora.yml
+```
+
+Multinode GPU Training with `accelerate`:
+```shell
+# make sure you've configured accelerate properly
+accelerate launch -m axolotl.cli.train examples/stablelm-2/1.6b/fft.yml --deepspeed deepspeed_configs/zero2.json
+```

examples/starcoder2/qlora.yml

@@ -0,0 +1,69 @@
+base_model: bigcode/starcoder2-3b
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+
+
+dataset_prepared_path:
+val_set_size: 0.2
+output_dir: ./qlora
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 8192
+sample_packing: true
+pad_to_sequence_len: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules:
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 8
+micro_batch_size: 2
+num_epochs: 3
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16: false
+tf32: false
+
+gradient_checkpointing: 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_steps:
+eval_table_size:
+saves_per_epoch: 4
+save_steps:
+save_total_limit: 2
+debug:
+deepspeed:
+weight_decay:
+fsdp:
+fsdp_config:
+special_tokens:

examples/tiny-llama/lora.yml

@@ -15,6 +15,7 @@ output_dir: ./lora-out
 
 sequence_len: 4096
 sample_packing: true
+eval_sample_packing: false
 pad_to_sequence_len: true
 
 adapter: lora

index.qmd

@@ -0,0 +1,19 @@
+
+
+```{python}
+#|output: asis
+#|echo: false
+
+# This cell steals the README as the home page for now, but excludes the table of contents (quarto adds its own)
+import re
+pattern = re.compile(
+    r"<table>\s*<tr>\s*<td>\s*## Table of Contents.*?</td>\s*</tr>\s*</table>",
+    re.DOTALL | re.IGNORECASE
+)
+
+with open('README.md', 'r') as f:
+    txt = f.read()
+
+cleaned = pattern.sub("", txt)
+print(cleaned)
+```

requirements.txt

@@ -1,17 +1,18 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
 packaging==23.2
-peft @ git+https://github.com/huggingface/peft.git
-transformers @ git+https://github.com/huggingface/transformers.git@ae49b218c3d718df90d8e4a109016450fb8f0632
+peft==0.10.0
+transformers @ git+https://github.com/huggingface/transformers.git@43d17c18360ac9c3d3491389328e2fe55fe8f9ce
 tokenizers==0.15.0
-bitsandbytes>=0.41.1
-accelerate==0.26.1
-deepspeed>=0.13.1
+bitsandbytes==0.43.0
+accelerate==0.28.0
+deepspeed==0.13.1
+pydantic==2.6.3
 addict
 fire
 PyYAML>=6.0
 requests
 datasets>=2.15.0
-flash-attn==2.3.3
+flash-attn==2.5.5
 sentencepiece
 wandb
 einops
@@ -21,22 +22,21 @@ hf_transfer
 colorama
 numba
 numpy>=1.24.4
-mlflow
 # qlora things
 evaluate==0.4.1
 scipy
 scikit-learn==1.2.2
 pynvml
 art
-fschat==0.2.34
+fschat==0.2.36
 gradio==3.50.2
 tensorboard
 
-mamba-ssm==1.1.1
+mamba-ssm==1.2.0.post1
 
 # remote filesystems
 s3fs
 gcsfs
 # adlfs
 
-trl>=0.7.9
+trl @ git+https://github.com/huggingface/trl.git@0ee349dcd43b0f4b3169449f16751c38ac4a609f

setup.py

@@ -18,6 +18,7 @@ def parse_requirements():
                 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
@@ -67,20 +68,29 @@ setup(
     dependency_links=dependency_links,
     extras_require={
         "flash-attn": [
-            "flash-attn==2.5.0",
+            "flash-attn==2.5.5",
         ],
         "fused-dense-lib": [
             "fused-dense-lib  @ git+https://github.com/Dao-AILab/flash-attention@v2.3.3#subdirectory=csrc/fused_dense_lib",
         ],
         "deepspeed": [
-            "deepspeed>=0.13.1",
+            "deepspeed==0.13.1",
             "deepspeed-kernels",
         ],
         "mamba-ssm": [
-            "mamba-ssm==1.0.1",
+            "mamba-ssm==1.2.0.post1",
         ],
         "auto-gptq": [
             "auto-gptq==0.5.1",
         ],
+        "mlflow": [
+            "mlflow",
+        ],
+        "lion-pytorch": [
+            "lion-pytorch==0.1.2",
+        ],
+        "galore": [
+            "galore_torch",
+        ],
     },
 )

src/axolotl/cli/__init__.py

@@ -13,7 +13,6 @@ from threading import Thread
 from typing import Any, Dict, List, Optional, Union
 from urllib.parse import urlparse
 
-import gradio as gr
 import requests
 import torch
 import yaml
@@ -24,6 +23,7 @@ from art import text2art
 from huggingface_hub import HfApi
 from huggingface_hub.utils import LocalTokenNotFoundError
 from transformers import GenerationConfig, TextIteratorStreamer, TextStreamer
+from transformers.utils import is_torch_bf16_gpu_available
 
 from axolotl.common.cli import TrainerCliArgs, load_model_and_tokenizer
 from axolotl.logging_config import configure_logging
@@ -214,6 +214,8 @@ def do_inference_gradio(
     cfg: DictDefault,
     cli_args: TrainerCliArgs,
 ):
+    import gradio as gr
+
     model, tokenizer = load_model_and_tokenizer(cfg=cfg, cli_args=cli_args)
     prompter = cli_args.prompter
     default_tokens = {"unk_token": "<unk>", "bos_token": "<s>", "eos_token": "</s>"}
@@ -328,7 +330,6 @@ def load_cfg(config: Union[str, Path] = Path("examples/"), **kwargs):
     # load the config from the yaml file
     with open(config, encoding="utf-8") as file:
         cfg: DictDefault = DictDefault(yaml.safe_load(file))
-    cfg.axolotl_config_path = config
     # if there are any options passed in the cli, if it is something that seems valid from the yaml,
     # then overwrite the value
     cfg_keys = cfg.keys()
@@ -341,7 +342,22 @@ def load_cfg(config: Union[str, Path] = Path("examples/"), **kwargs):
             else:
                 cfg[k] = kwargs[k]
 
-    validate_config(cfg)
+    cfg.axolotl_config_path = config
+
+    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
+
+    cfg = validate_config(
+        cfg,
+        capabilities={
+            "bf16": is_torch_bf16_gpu_available(),
+            "n_gpu": os.environ.get("WORLD_SIZE", 1),
+            "compute_capability": gpu_version,
+        },
+    )
 
     prepare_optim_env(cfg)
 

src/axolotl/cli/preprocess.py

@@ -54,7 +54,7 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
         LOG.warning(msg)
         parsed_cfg.dataset_prepared_path = DEFAULT_DATASET_PREPARED_PATH
 
-    if parsed_cfg.rl:
+    if parsed_cfg.rl and parsed_cfg.rl != "orpo":
         load_rl_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
     else:
         load_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)

src/axolotl/cli/train.py

@@ -47,7 +47,7 @@ def do_train(cfg, cli_args) -> Tuple[PreTrainedModel, PreTrainedTokenizer]:
     else:
         register_chatml_template()
 
-    if cfg.rl:
+    if cfg.rl and cfg.rl != "orpo":
         dataset_meta = load_rl_datasets(cfg=cfg, cli_args=cli_args)
     else:
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)

src/axolotl/core/trainer_builder.py

@@ -5,14 +5,16 @@ Builder for the training args and trainer
 
 import abc
 import importlib
+import importlib.util
 import logging
 import math
 import sys
 from abc import abstractmethod
+from collections import defaultdict
 from dataclasses import dataclass, field
 from functools import wraps
 from pathlib import Path
-from typing import List, Optional, Type, Union
+from typing import Dict, List, Literal, Optional, Type, Union
 
 import torch
 import transformers
@@ -26,21 +28,24 @@ from transformers import (
     TrainingArguments,
 )
 from transformers.trainer_utils import seed_worker
+from transformers.utils import is_sagemaker_mp_enabled
 from trl import DPOTrainer
+from trl.trainer.utils import pad_to_length
 
+from axolotl.loraplus import create_loraplus_optimizer
 from axolotl.monkeypatch.multipack import SUPPORTED_MULTIPACK_MODEL_TYPES
 from axolotl.monkeypatch.relora import ReLoRACallback, ReLoRAScheduler
 from axolotl.utils.callbacks import (
     EvalFirstStepCallback,
     GPUStatsCallback,
     LossWatchDogCallback,
-    SaveAxolotlConfigtoMlflowCallback,
     SaveAxolotlConfigtoWandBCallback,
     SaveBetterTransformerModelCallback,
     bench_eval_callback_factory,
     causal_lm_bench_eval_callback_factory,
     log_prediction_callback_factory,
 )
+from axolotl.utils.callbacks.lisa import lisa_callback_factory
 from axolotl.utils.collators import (
     BatchSamplerDataCollatorForSeq2Seq,
     DataCollatorForSeq2Seq,
@@ -54,6 +59,9 @@ from axolotl.utils.schedulers import (
     get_cosine_schedule_with_warmup_decay_constant,
 )
 
+if is_sagemaker_mp_enabled():
+    import smdistributed.modelparallel.torch as smp
+
 try:
     import torch._dynamo  # pylint: disable=ungrouped-imports
 except ImportError:
@@ -62,6 +70,10 @@ except ImportError:
 LOG = logging.getLogger("axolotl.core.trainer_builder")
 
 
+def is_mlflow_available():
+    return importlib.util.find_spec("mlflow") is not None
+
+
 def _sanitize_kwargs_for_tagging(tag_names, kwargs=None):
     if isinstance(tag_names, str):
         tag_names = [tag_names]
@@ -175,6 +187,32 @@ class AxolotlTrainingArguments(TrainingArguments):
             "help": "Starting constant learning rate step is cosine_constant_lr_ratio * max_steps"
         },
     )
+    loraplus_lr_ratio: Optional[float] = field(
+        default=None, metadata={"help": "loraplus learning rate ratio lr_B / lr_A."}
+    )
+    loraplus_lr_embedding: Optional[float] = field(
+        default=1e-6,
+        metadata={"help": "loraplus learning rate for lora embedding layers."},
+    )
+    qlora: bool = field(
+        default=False,
+        metadata={"help": "whether this is a qlora training"},
+    )
+    orpo_alpha: Optional[float] = field(
+        default=None,
+    )
+    lisa_n_layers: Optional[int] = field(
+        default=None,
+        metadata={"help": "the number of activate layers in LISA"},
+    )
+    lisa_step_interval: Optional[int] = field(
+        default=None,
+        metadata={"help": "how often to switch layers in LISA"},
+    )
+    lisa_layers_attribute: Optional[str] = field(
+        default=None,
+        metadata={"help": "path under the model to access the layers"},
+    )
 
 
 class AxolotlTrainer(Trainer):
@@ -191,13 +229,44 @@ class AxolotlTrainer(Trainer):
         num_epochs=1,
         bench_data_collator=None,
         eval_data_collator=None,
-        **kwargs
+        **kwargs,
     ):
         self.num_epochs = num_epochs
         self.bench_data_collator = bench_data_collator
         self.eval_data_collator = eval_data_collator
         super().__init__(*_args, **kwargs)
         self.train_data_collator = self.data_collator
+        self._stored_metrics = defaultdict(lambda: defaultdict(list))
+        if self.args.orpo_alpha:
+            self.loss_fct = torch.nn.CrossEntropyLoss(reduction="none")
+
+    def create_optimizer(self):
+        if self.args.loraplus_lr_ratio is None:
+            return super().create_optimizer()
+
+        opt_model = self.model_wrapped if is_sagemaker_mp_enabled() else self.model
+        if self.optimizer is None:  # pylint: disable=access-member-before-definition
+            optimizer_cls, optimizer_kwargs = Trainer.get_optimizer_cls_and_kwargs(
+                self.args,
+                opt_model,
+            )
+
+            loraplus_lr_ratio = getattr(self.args, "loraplus_lr_ratio", None)
+            loraplus_lr_embedding = getattr(self.args, "loraplus_lr_embedding", None)
+            self.optimizer = create_loraplus_optimizer(  # pylint: disable=attribute-defined-outside-init
+                opt_model,
+                optimizer_cls,
+                optimizer_kwargs,
+                loraplus_lr_ratio,
+                loraplus_lr_embedding,
+            )
+
+        if is_sagemaker_mp_enabled():
+            self.optimizer = smp.DistributedOptimizer(  # pylint: disable=attribute-defined-outside-init
+                self.optimizer
+            )
+
+        return self.optimizer
 
     def create_scheduler(
         self, num_training_steps: int, optimizer: torch.optim.Optimizer = None
@@ -413,8 +482,165 @@ class AxolotlTrainer(Trainer):
         #     outputs = model(**inputs)
         #     loss = trainer_weighted_loss(outputs, labels, shift_labels=True)
         #     return (loss, outputs) if return_outputs else loss
+        if self.args.orpo_alpha:
+            return self.orpo_compute_loss(model, inputs, return_outputs=return_outputs)
         return super().compute_loss(model, inputs, return_outputs=return_outputs)
 
+    @staticmethod
+    def orpo_concatenate_inputs(inputs, label_pad_token=-100, pad_token=0, device=None):
+        concatenated_batch = {}
+
+        max_length = max(
+            inputs["input_ids"].shape[1], inputs["rejected_input_ids"].shape[1]
+        )
+        # Concatenate positive and negative inputs
+        concatenated_batch["input_ids"] = pad_to_length(
+            inputs["input_ids"], max_length, pad_token
+        )
+        concatenated_batch["rejected_input_ids"] = pad_to_length(
+            inputs["rejected_input_ids"], max_length, pad_token
+        )
+        concatenated_batch["labels"] = pad_to_length(
+            inputs["labels"], max_length, label_pad_token
+        )
+        concatenated_batch["rejected_labels"] = pad_to_length(
+            inputs["rejected_labels"], max_length, label_pad_token
+        )
+        concatenated_batch["attention_mask"] = pad_to_length(
+            inputs["attention_mask"], max_length, 0
+        )
+        concatenated_batch["rejected_attention_mask"] = pad_to_length(
+            inputs["rejected_attention_mask"], max_length, 0
+        )
+        concatenated_batch["prompt_attention_mask"] = pad_to_length(
+            inputs["prompt_attention_mask"], max_length, 0
+        ).to(device=device)
+
+        input_ids = torch.cat(
+            [concatenated_batch["input_ids"], concatenated_batch["rejected_input_ids"]],
+            dim=0,
+        ).to(device=device)
+        attention_mask = torch.cat(
+            [
+                concatenated_batch["attention_mask"],
+                concatenated_batch["rejected_attention_mask"],
+            ],
+            dim=0,
+        ).to(device=device)
+        labels = torch.cat(
+            [concatenated_batch["labels"], concatenated_batch["rejected_labels"]], dim=0
+        ).to(device=device)
+
+        return {
+            "input_ids": input_ids,
+            "labels": labels,
+            "attention_mask": attention_mask,
+            "prompt_attention_mask": concatenated_batch["prompt_attention_mask"],
+        }
+
+    def orpo_compute_custom_loss(self, logits, labels):
+        logits = logits.contiguous()
+        loss = 0.0
+
+        if labels is not None:
+            # move labels to correct device to enable model parallelism
+            labels = labels.to(logits.device)
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+
+            # Flatten the tokens
+            loss = self.loss_fct(shift_logits.transpose(2, 1), shift_labels).mean(
+                dim=-1
+            )
+
+        return loss
+
+    def orpo_compute_logps(
+        self, prompt_attention_mask, chosen_inputs, chosen_attention_mask, logits
+    ):
+        # Get the shape of chosen_attention_mask[:, :-1]
+        chosen_shape = chosen_attention_mask[:, :-1].shape
+
+        # Calculate the padding size
+        pad_length = chosen_shape[1] - (prompt_attention_mask.shape[1] - 1)
+
+        # Pad prompt_attention_mask with zeros to match the desired shape
+        prompt_attention_mask_padded = torch.nn.functional.pad(
+            prompt_attention_mask[:, 1:], (0, pad_length), mode="constant", value=0
+        )
+
+        # Perform the subtraction operation
+        mask = chosen_attention_mask[:, :-1] > prompt_attention_mask_padded
+
+        per_token_logps = torch.gather(
+            logits[:, :-1, :].log_softmax(-1),
+            dim=2,
+            index=(mask * chosen_inputs[:, 1:]).unsqueeze(2),
+        ).squeeze(2)
+        return torch.mul(per_token_logps, mask).sum(dim=1) / mask.sum(dim=1)
+
+    def orpo_compute_loss(self, model, inputs, return_outputs=False):
+        concat_inputs = AxolotlTrainer.orpo_concatenate_inputs(
+            inputs,
+            label_pad_token=-100,
+            pad_token=self.tokenizer.pad_token_id,
+            device=self.accelerator.device,
+        )
+
+        # Perform a single forward pass
+        outputs = model(
+            **{
+                "input_ids": concat_inputs["input_ids"],
+                "attention_mask": concat_inputs["attention_mask"],
+                "labels": concat_inputs["labels"],
+            },
+            output_hidden_states=True,
+        )
+
+        # Split the outputs for positive and negative examples
+        outputs_pos, outputs_neg = outputs.logits.chunk(2)
+
+        # Calculate NLL loss
+        pos_loss = self.orpo_compute_custom_loss(
+            logits=outputs_pos, labels=concat_inputs["input_ids"].chunk(2)[0]
+        )
+
+        # Calculate Log Probability
+        pos_prob = self.orpo_compute_logps(
+            prompt_attention_mask=concat_inputs["prompt_attention_mask"],
+            chosen_inputs=concat_inputs["input_ids"].chunk(2)[0],
+            chosen_attention_mask=concat_inputs["attention_mask"].chunk(2)[0],
+            logits=outputs_pos,
+        )
+        neg_prob = self.orpo_compute_logps(
+            prompt_attention_mask=concat_inputs["prompt_attention_mask"],
+            chosen_inputs=concat_inputs["input_ids"].chunk(2)[1],
+            chosen_attention_mask=concat_inputs["attention_mask"].chunk(2)[1],
+            logits=outputs_neg,
+        )
+
+        # Calculate log odds
+        log_odds = (pos_prob - neg_prob) - (
+            torch.log(1 - torch.exp(pos_prob)) - torch.log(1 - torch.exp(neg_prob))
+        )
+        sig_ratio = torch.nn.functional.sigmoid(log_odds)
+        ratio = torch.log(sig_ratio)
+
+        # Calculate the Final Loss
+        loss = torch.mean(pos_loss - self.args.orpo_alpha * ratio).to(
+            dtype=torch.bfloat16
+        )
+
+        metrics = {}
+        metrics["chosen_geometric_mean"] = torch.mean(pos_prob).cpu().item()
+        metrics["rejected_geometric_mean"] = torch.mean(neg_prob).cpu().item()
+        metrics["log_odds_ratio"] = torch.mean(ratio).cpu().item()
+        metrics["log_odds"] = torch.mean(log_odds).cpu().item()
+        self.store_metrics(metrics, train_eval="train")
+
+        return (loss, outputs_pos) if return_outputs else loss
+
     @wraps(Trainer.push_to_hub)
     def push_to_hub(self, *args, **kwargs) -> str:
         """
@@ -425,6 +651,41 @@ class AxolotlTrainer(Trainer):
 
         return super().push_to_hub(*args, **kwargs)
 
+    @wraps(Trainer.create_accelerator_and_postprocess)
+    def create_accelerator_and_postprocess(self):
+        res = super().create_accelerator_and_postprocess()
+
+        if self.is_fsdp_enabled:
+            if (
+                "limit_all_gathers" in self.args.fsdp_config
+                and self.args.fsdp_config["limit_all_gathers"]
+            ):
+                self.accelerator.state.fsdp_plugin.limit_all_gathers = True
+
+        return res
+
+    def log(self, logs: Dict[str, float]) -> None:
+        """
+        Log `logs` on the various objects watching training, including stored metrics.
+
+        Args:
+            logs (`Dict[str, float]`):
+                The values to log.
+        """
+        # 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]
+        return super().log(logs)
+
+    def store_metrics(
+        self, metrics: Dict[str, float], train_eval: Literal["train", "eval"] = "train"
+    ) -> None:
+        for key, value in metrics.items():
+            self._stored_metrics[train_eval][key].append(value)
+
 
 class AxolotlMambaTrainer(AxolotlTrainer):
     """
@@ -557,6 +818,12 @@ class TrainerBuilderBase(abc.ABC):
         self.model = model
         self.tokenizer = tokenizer
 
+        # in case the model supports tagging, add the axolotl tag.
+        # This makes sure the tag is correctly pushed even if a user calls
+        # model.push_to_hub instad of  trainer.push_to_hub.
+        if hasattr(model, "add_model_tags"):
+            model.add_model_tags(["axolotl"])
+
     @property
     def model_ref(self):
         return self._model_ref
@@ -648,7 +915,11 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             callbacks.append(
                 SaveAxolotlConfigtoWandBCallback(self.cfg.axolotl_config_path)
             )
-        if self.cfg.use_mlflow:
+        if self.cfg.use_mlflow and is_mlflow_available():
+            from axolotl.utils.callbacks.mlflow_ import (
+                SaveAxolotlConfigtoMlflowCallback,
+            )
+
             callbacks.append(
                 SaveAxolotlConfigtoMlflowCallback(self.cfg.axolotl_config_path)
             )
@@ -680,6 +951,8 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             )
             callbacks.append(early_stop_cb)
 
+        if self.cfg.lisa_step_interval and self.cfg.lisa_n_layers:
+            callbacks.append(lisa_callback_factory(trainer))
         return callbacks
 
     def _get_trainer_cls(self):
@@ -731,15 +1004,14 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 training_arguments_kwargs[
                     "gradient_checkpointing_kwargs"
                 ] = self.cfg.gradient_checkpointing_kwargs
-            else:
-                training_arguments_kwargs["gradient_checkpointing_kwargs"] = {
-                    "use_reentrant": False
-                }
         if self.cfg.fsdp:
             training_arguments_kwargs["fsdp"] = self.cfg.fsdp
             if self.cfg.fsdp_config:
                 training_arguments_kwargs["fsdp_config"] = dict(self.cfg.fsdp_config)
 
+        if self.cfg.adapter == "qlora":
+            training_arguments_kwargs["qlora"] = True
+
         # deepspeed
         if self.cfg.deepspeed:
             training_arguments_kwargs["deepspeed"] = self.cfg.deepspeed
@@ -794,6 +1066,11 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         elif self.cfg.sample_packing and self.cfg.eval_sample_packing is False:
             training_arguments_kwargs["dataloader_drop_last"] = True
 
+        if self.cfg.remove_unused_columns is not None:
+            training_arguments_kwargs[
+                "remove_unused_columns"
+            ] = self.cfg.remove_unused_columns
+
         if not self.cfg.test_datasets and self.cfg.val_set_size == 0:
             # no eval set, so don't eval
             training_arguments_kwargs["evaluation_strategy"] = "no"
@@ -907,6 +1184,22 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs["optim"] = (
             self.cfg.optimizer if self.cfg.optimizer else "adamw_hf"
         )
+        if self.cfg.optim_args:
+            if isinstance(self.cfg.optim_args, dict):
+                optim_args = ",".join(
+                    [f"{key}={value}" for key, value in self.cfg.optim_args.items()]
+                )
+            else:
+                optim_args = self.cfg.optim_args
+            training_arguments_kwargs["optim_args"] = optim_args
+        if self.cfg.optim_target_modules:
+            training_arguments_kwargs[
+                "optim_target_modules"
+            ] = self.cfg.optim_target_modules
+        training_arguments_kwargs["loraplus_lr_ratio"] = self.cfg.loraplus_lr_ratio
+        training_arguments_kwargs[
+            "loraplus_lr_embedding"
+        ] = self.cfg.loraplus_lr_embedding
         training_arguments_kwargs["lr_scheduler_type"] = (
             self.cfg.lr_scheduler
             if self.cfg.lr_scheduler
@@ -951,29 +1244,65 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                     "relora_prune_ratio"
                 ] = self.cfg.relora_prune_ratio
 
+        if self.cfg.lisa_step_interval and self.cfg.lisa_n_layers:
+            training_arguments_kwargs["lisa_n_layers"] = self.cfg.lisa_n_layers
+            training_arguments_kwargs[
+                "lisa_step_interval"
+            ] = self.cfg.lisa_step_interval
+            training_arguments_kwargs[
+                "lisa_layers_attribute"
+            ] = self.cfg.lisa_layers_attribute
+
         training_arguments_kwargs = self.hook_pre_create_training_args(
             training_arguments_kwargs
         )
         training_arguments_kwargs["model_type"] = self.cfg.model_config_type
         training_arguments_kwargs["pretraining"] = bool(self.cfg.pretraining_dataset)
 
+        if self.cfg.rl == "orpo":
+            training_arguments_kwargs["orpo_alpha"] = self.cfg.orpo_alpha
+
         if self.cfg.neftune_noise_alpha is not None:
             training_arguments_kwargs[
                 "neftune_noise_alpha"
             ] = self.cfg.neftune_noise_alpha
 
+        trainer_kwargs = {}
+
+        if self.cfg.optimizer == "lion_pytorch":
+            from lion_pytorch import Lion
+
+            lion_kwargs = {"lr": training_arguments_kwargs["learning_rate"]}
+            if "weight_decay" in training_arguments_kwargs:
+                lion_kwargs["weight_decay"] = training_arguments_kwargs["weight_decay"]
+
+            if (
+                "adam_beta1" in training_arguments_kwargs
+                and "adam_beta2" in training_arguments_kwargs
+            ):
+                lion_kwargs["betas"] = (
+                    training_arguments_kwargs["adam_beta1"],
+                    training_arguments_kwargs["adam_beta2"],
+                )
+
+            trainer_kwargs["optimizers"] = (
+                Lion(params=self.model.parameters(), **lion_kwargs),
+                None,
+            )
+            # Set default so transformers doesn't throw
+            training_arguments_kwargs["optim"] = "adamw_hf"
+
+        if self.cfg.optimizer == "adamw_anyprecision":
+            if Path(self.cfg.torchdistx_path).exists():
+                sys.path.append(self.cfg.torchdistx_path)
+                importlib.import_module("torchdistx")
+
         training_args = (
             AxolotlTrainingArguments(  # pylint: disable=unexpected-keyword-arg
                 **training_arguments_kwargs,
             )
         )
         training_args = self.hook_post_create_training_args(training_args)
-        trainer_kwargs = {}
-
-        if self.cfg.optimizer == "adamw_anyprecision":
-            if Path(self.cfg.torchdistx_path).exists():
-                sys.path.append(self.cfg.torchdistx_path)
-                importlib.import_module("torchdistx")
 
         data_collator_kwargs = {
             "padding": True,  # True/"longest" is the default
@@ -996,6 +1325,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             train_dataset=self.train_dataset,
             eval_dataset=self.eval_dataset,
             args=training_args,
+            tokenizer=self.tokenizer,
             data_collator=self.build_collator(training_args, **data_collator_kwargs),
             eval_data_collator=self.build_collator(
                 training_args, is_eval=True, **data_collator_kwargs

src/axolotl/logging_config.py

@@ -30,6 +30,7 @@ class ColorfulFormatter(Formatter):
 
 DEFAULT_LOGGING_CONFIG: Dict[str, Any] = {
     "version": 1,
+    "disable_existing_loggers": False,
     "formatters": {
         "simple": {
             "format": "[%(asctime)s] [%(levelname)s] [%(name)s.%(funcName)s:%(lineno)d] [PID:%(process)d] %(message)s",

src/axolotl/loraplus.py

@@ -0,0 +1,133 @@
+"""Module for LoRA+"""
+
+# MIT License
+#
+# Copyright (c) 2024 nikhil-ghosh-berkeley
+# https://github.com/nikhil-ghosh-berkeley/loraplus
+
+import logging
+from functools import reduce
+
+from peft.tuners import lora
+from torch import nn
+from transformers.pytorch_utils import ALL_LAYERNORM_LAYERS
+from transformers.trainer_pt_utils import get_parameter_names
+
+LOG = logging.getLogger("axolotl.loraplus")
+
+
+def get_module(name, opt_model):
+    """
+    Retrieve a module from a model using its parameter name.
+    Args:
+        name (str): Full name of the parameter, typically including module path.
+        opt_model (torch.nn.Module): The model from which to retrieve the module.
+
+    Returns:
+        Module corresponding to the given name.
+    """
+    parent_idx = 2 if "lora" in name else 1
+    module_names = name.split(sep=".")[:-parent_idx]
+    module = reduce(getattr, module_names, opt_model)
+    return module
+
+
+def create_loraplus_optimizer(
+    opt_model,
+    optimizer_cls,
+    optimizer_kwargs,
+    loraplus_lr_ratio,
+    loraplus_lr_embedding=None,
+):
+    """
+    Creates an optimizer for the given model, applying LoRA-specific learning rate adjustments to different parameter groups.
+
+    Args:
+        opt_model (torch.nn.Module): The model for which the optimizer is being created.
+        optimizer_cls (class): The class of the optimizer to be used (e.g., torch.optim.Adam).
+        optimizer_kwargs (dict): A dictionary of keyword arguments for the optimizer's initialization.
+        loraplus_lr_ratio (float): The learning rate ratio to be applied to LoRA parameters.
+        loraplus_lr_embedding (float, optional): A specific learning rate for embedding parameters, with a default value if not provided.
+
+    Returns:
+        An instance of the specified optimizer class configured with the model's parameters organized into groups with custom learning rates.
+    """
+
+    assert loraplus_lr_ratio is not None, "loraplus_lr_ratio must be provided."
+
+    if loraplus_lr_embedding is None:
+        loraplus_lr_embedding = 1e-6
+
+    decay_parameters = get_parameter_names(opt_model, ALL_LAYERNORM_LAYERS)
+    decay_parameters = [name for name in decay_parameters if "bias" not in name]
+    param_groups = {
+        "groupA": {},
+        "groupB": {},
+        "groupB_no_decay": {},
+        "embedding": {},
+    }
+
+    for name, param in opt_model.named_parameters():
+        if not param.requires_grad:
+            continue
+
+        module = get_module(name, opt_model)
+        if isinstance(module, lora.Embedding):
+            param_groups["embedding"][name] = param
+        elif "lora_B" in name or param.ndim == 1:
+            if name in decay_parameters:
+                param_groups["groupB"][name] = param
+            else:
+                param_groups["groupB_no_decay"][name] = param
+        else:
+            param_groups["groupA"][name] = param
+
+    assigned_param_groups = ""
+    for group, group_params in param_groups.items():
+        assigned_param_groups += f"{group}\n {list(group_params.keys())}\n\n"
+    LOG.info(assigned_param_groups)
+
+    lr = optimizer_kwargs["lr"]  # pylint: disable=invalid-name
+    weight_decay = optimizer_kwargs.get("weight_decay", 0.0)
+
+    optimizer_grouped_parameters = [
+        {
+            "params": list(param_groups["groupA"].values()),
+            "weight_decay": weight_decay,
+            "lr": lr,
+        },
+        {
+            "params": list(param_groups["embedding"].values()),
+            "weight_decay": weight_decay,
+            "lr": loraplus_lr_embedding,
+        },
+        {
+            "params": list(param_groups["groupB"].values()),
+            "weight_decay": weight_decay,
+            "lr": lr * loraplus_lr_ratio,
+        },
+        {
+            "params": list(param_groups["groupB_no_decay"].values()),
+            "weight_decay": 0.0,
+            "lr": lr * loraplus_lr_ratio,
+        },
+    ]
+
+    optimizer = optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs)
+    if optimizer_cls.__name__ == "Adam8bit":
+        import bitsandbytes
+
+        manager = bitsandbytes.optim.GlobalOptimManager.get_instance()
+
+        skipped = 0
+        for module in opt_model.modules():
+            if isinstance(module, nn.Embedding):
+                skipped += sum(
+                    {p.data_ptr(): p.numel() for p in module.parameters()}.values()
+                )
+                LOG.info(f"skipped {module}: {skipped/2**20}M params")
+                manager.register_module_override(module, "weight", {"optim_bits": 32})
+                LOG.debug(f"bitsandbytes: will optimize {module} in fp32")
+        LOG.info(f"skipped: {skipped/2**20}M params")
+
+    return optimizer

src/axolotl/monkeypatch/fastchat_conversation_turns.py

@@ -106,7 +106,7 @@ def get_turns(  # pylint: disable=too-many-return-statements
         if self.system_message:
             contains_sys_msg = True
             if self.messages:
-                # There is no clear guidance on how to handle system messages in Mistral so we just prepend it to the first human instruction seperated by a newline
+                # There is no clear guidance on how to handle system messages in Mistral so we just prepend it to the first human instruction separated by a newline
                 first_role, first_msg = self.messages[0]
                 if first_role == self.roles[0]:
                     system_prompt = self.system_template.format(

src/axolotl/monkeypatch/llama_attn_hijack_flash.py

@@ -44,6 +44,18 @@ except ImportError:
 LOG = logging.getLogger("axolotl")
 
 
+def is_xformers_swiglu_available() -> bool:
+    from xformers.ops.common import get_xformers_operator
+
+    try:
+        get_xformers_operator("swiglu_packedw")()
+        return True
+    except RuntimeError as exc:
+        if "No such operator xformers::swiglu_packedw " in str(exc):
+            return False
+        return True
+
+
 def replace_llama_mlp_with_swiglu(model):
     for name, module in model.named_modules():
         if isinstance(module, LlamaMLP):
@@ -272,12 +284,7 @@ def flashattn_forward_with_s2attn(
     # [bsz, nh, q_len, hd]
     # pylint: disable=duplicate-code
 
-    kv_seq_len = key_states.shape[-2]
-    if past_key_value is not None:
-        kv_seq_len += past_key_value[0].shape[-2]
-    cos, sin = self.rotary_emb(
-        value_states, seq_len=kv_seq_len, position_ids=position_ids
-    )
+    cos, sin = self.rotary_emb(value_states, position_ids=position_ids)
     query_states, key_states = apply_rotary_pos_emb(
         query_states, key_states, cos, sin, position_ids
     )
@@ -423,13 +430,7 @@ def flashattn_forward(
     # [bsz, q_len, nh, hd]
     # [bsz, nh, q_len, hd]
 
-    kv_seq_len = key_states.shape[-2]
-    if past_key_value is not None:
-        kv_seq_len += past_key_value[0].shape[-2]
-
-    cos, sin = self.rotary_emb(
-        value_states, seq_len=kv_seq_len, position_ids=position_ids
-    )
+    cos, sin = self.rotary_emb(value_states, position_ids=position_ids)
     query_states, key_states = apply_rotary_pos_emb(
         query_states, key_states, cos, sin, position_ids
     )

src/axolotl/monkeypatch/llama_attn_hijack_xformers.py

@@ -80,11 +80,7 @@ def xformers_forward(
     # [bsz, q_len, nh, hd]
     # [bsz, nh, q_len, hd]
 
-    kv_seq_len = key_states.shape[-2]
-    if past_key_value is not None:
-        kv_seq_len += past_key_value[0].shape[-2]
-
-    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+    cos, sin = self.rotary_emb(value_states)
     query_states, key_states = apply_rotary_pos_emb(
         query_states, key_states, cos, sin, position_ids
     )

src/axolotl/monkeypatch/multipack.py

@@ -1,15 +1,27 @@
 """multipack patching for v2 of sample packing"""
+import importlib
 
 import transformers
+from accelerate import init_empty_weights
+from transformers import AutoConfig, AutoModelForCausalLM
 from transformers.integrations import is_deepspeed_zero3_enabled
 
 from axolotl.monkeypatch.mixtral import patch_mixtral_moe_forward_zero3
 from axolotl.monkeypatch.utils import get_unpad_data
 
-SUPPORTED_MULTIPACK_MODEL_TYPES = ["mixtral", "qwen2", "falcon", "phi", "gemma"]
+SUPPORTED_MULTIPACK_MODEL_TYPES = [
+    "mixtral",
+    "qwen2",
+    "qwen2_moe",
+    "falcon",
+    "phi",
+    "gemma",
+    "gemmoe",
+    "starcoder2",
+]
 
 
-def patch_for_multipack(model_type):
+def patch_for_multipack(model_type, model_name=None):
     if model_type == "mixtral":
         transformers.models.mixtral.modeling_mixtral._get_unpad_data = (  # pylint: disable=protected-access
             get_unpad_data
@@ -20,6 +32,10 @@ def patch_for_multipack(model_type):
         transformers.models.qwen2.modeling_qwen2._get_unpad_data = (  # pylint: disable=protected-access
             get_unpad_data
         )
+    elif model_type == "qwen2_moe":
+        transformers.models.qwen2_moe.modeling_qwen2_moe._get_unpad_data = (  # pylint: disable=protected-access
+            get_unpad_data
+        )
     elif model_type == "falcon":
         transformers.models.falcon.modeling_falcon._get_unpad_data = (  # pylint: disable=protected-access
             get_unpad_data
@@ -32,3 +48,21 @@ def patch_for_multipack(model_type):
         transformers.models.gemma.modeling_gemma._get_unpad_data = (  # pylint: disable=protected-access
             get_unpad_data
         )
+    elif model_type == "starcoder2":
+        transformers.models.starcoder2.modeling_starcoder2._get_unpad_data = (  # pylint: disable=protected-access
+            get_unpad_data
+        )
+    elif model_type == "gemmoe":
+        patch_remote(model_name, ".configuration_gemmoe", ".modeling_gemmoe")
+    elif model_type == "jamba":
+        patch_remote(model_name, ".configuration_jamba", ".modeling_jamba")
+
+
+def patch_remote(model_name, config_name, modeling_name):
+    model_config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
+    # we need to load the model here in order for modeling_* to be available
+    with init_empty_weights():
+        AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)
+    module_name = model_config.__class__.__module__.replace(config_name, modeling_name)
+    modeling_arch = importlib.import_module(module_name)
+    modeling_arch._get_unpad_data = get_unpad_data  # pylint: disable=protected-access

src/axolotl/monkeypatch/relora.py

@@ -267,7 +267,7 @@ class ReLoRAScheduler(LRScheduler):
         original = self.inner_schedule.get_lr()
         step = self.last_epoch
 
-        if step < self.relora_steps:
+        if step < self.relora_steps - self.warmup_steps:
             scale = 1
         else:
             per_relora_progress = step % self.relora_steps

src/axolotl/prompt_strategies/base.py

@@ -0,0 +1,20 @@
+"""
+module for base dataset transform strategies
+"""
+
+import importlib
+import logging
+
+LOG = logging.getLogger("axolotl")
+
+
+def load(strategy, cfg, module_base=None, **kwargs):
+    try:
+        load_fn = strategy.split(".")[-1]
+        strategy = ".".join(strategy.split(".")[:-1])
+        mod = importlib.import_module(f".{strategy}", module_base)
+        func = getattr(mod, load_fn)
+        return func(cfg, **kwargs)
+    except Exception:  # pylint: disable=broad-exception-caught
+        LOG.warning(f"unable to load strategy {strategy}")
+        return None

src/axolotl/prompt_strategies/chat_template.py

@@ -0,0 +1,78 @@
+"""
+HF Chat Templates prompt strategy
+"""
+from typing import Any, Dict, Optional
+
+from axolotl.prompt_tokenizers import PromptTokenizingStrategy
+from axolotl.prompters import Prompter
+from axolotl.utils.chat_templates import chat_templates
+
+
+class ChatTemplatePrompter(Prompter):
+    """prompter for HF chat templates"""
+
+    def __init__(self, tokenizer, chat_template=None, max_length=2048):
+        self.tokenizer = tokenizer
+        self.chat_template = chat_template
+        self.max_length = max_length
+
+    def build_prompt(self, conversation, add_generation_prompt=False):
+        return self.tokenizer.apply_chat_template(
+            conversation,
+            truncation=True,
+            max_length=self.max_length,
+            add_generation_prompt=add_generation_prompt,
+            chat_template=self.chat_template,
+        )
+
+
+class ChatTemplateStrategy(PromptTokenizingStrategy):
+    """
+    Tokenizing strategy for instruction-based prompts.
+    """
+
+    def tokenize_prompt(self, prompt):
+        turns = self.get_conversation_thread(prompt)
+        prompt_ids = self.prompter.build_prompt([turns[0]], add_generation_prompt=True)
+        input_ids = self.prompter.build_prompt(turns)
+
+        if not self.train_on_inputs:
+            user_prompt_len = len(prompt_ids)
+            labels = [-100] * user_prompt_len + input_ids[user_prompt_len:]
+        else:
+            labels = input_ids
+
+        tokenized_prompt = {
+            "input_ids": input_ids,
+            "labels": labels,
+            "attention_mask": [1] * len(input_ids),
+        }
+
+        return tokenized_prompt
+
+    def get_conversation_thread(self, prompt):
+        conversations = prompt["conversations"]
+        # remap roles - allow for assistant turn
+        role_map = {
+            "human": "user",
+            "user": "user",
+            "assistant": "assistant",
+            "gpt": "assistant",
+        }
+        turns = [
+            {"role": role_map[t["from"]], "content": t["value"]} for t in conversations
+        ]
+        return turns
+
+
+def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
+    chat_template = (
+        ds_cfg["chat_template"] if ds_cfg and "chat_template" in ds_cfg else "chatml"
+    )
+    strategy = ChatTemplateStrategy(
+        ChatTemplatePrompter(tokenizer, chat_templates(chat_template)),
+        tokenizer,
+        cfg.train_on_inputs,
+        cfg.sequence_len,
+    )
+    return strategy

src/axolotl/prompt_strategies/dpo/__init__.py

@@ -1,21 +1,8 @@
 """
 module for DPO style dataset transform strategies
 """
+from functools import partial
 
-import importlib
-import logging
+from ..base import load as load_base
 
-LOG = logging.getLogger("axolotl")
-
-
-def load(strategy, cfg):
-    try:
-        load_fn = strategy.split(".")[-1]
-        strategy = ".".join(strategy.split(".")[:-1])
-        mod = importlib.import_module(f".{strategy}", "axolotl.prompt_strategies.dpo")
-        func = getattr(mod, load_fn)
-        load_kwargs = {}
-        return func(cfg, **load_kwargs)
-    except Exception:  # pylint: disable=broad-exception-caught
-        LOG.warning(f"unable to load strategy {strategy}")
-        return None
+load = partial(load_base, module_base="axolotl.prompt_strategies.dpo")

src/axolotl/prompt_strategies/dpo/chatml.py

@@ -5,6 +5,7 @@ DPO strategies for chatml
 
 def argilla(
     cfg,
+    **kwargs,
 ):  # pylint: disable=possibly-unused-variable,unused-argument
     def transform_fn(sample):
         if "system" in sample and sample["system"]:
@@ -23,8 +24,28 @@ def argilla(
     return transform_fn
 
 
+def argilla_chat(
+    cfg,
+    **kwargs,
+):  # pylint: disable=possibly-unused-variable,unused-argument
+    """
+    for argilla/dpo-mix-7k conversations
+    """
+
+    def transform_fn(sample):
+        sample[
+            "prompt"
+        ] = f"<|im_start|>user\n{sample['chosen'][0]['content']}<|im_end|>\n<|im_start|>assistant\n"
+        sample["chosen"] = f"{sample['chosen'][1]['content']}<|im_end|>"
+        sample["rejected"] = f"{sample['rejected'][1]['content']}<|im_end|>"
+        return sample
+
+    return transform_fn
+
+
 def icr(
     cfg,
+    **kwargs,
 ):  # pylint: disable=possibly-unused-variable,unused-argument
     """
     chatml transforms for datasets with system, input, chosen, rejected
@@ -48,7 +69,7 @@ def icr(
     return transform_fn
 
 
-def intel(cfg):  # pylint: disable=possibly-unused-variable,unused-argument
+def intel(cfg, **kwargs):  # pylint: disable=possibly-unused-variable,unused-argument
     """
     For Intel Orca DPO Pairs
     """
@@ -70,7 +91,9 @@ def intel(cfg):  # pylint: disable=possibly-unused-variable,unused-argument
     return transform_fn
 
 
-def prompt_pairs(cfg):  # pylint: disable=possibly-unused-variable,unused-argument
+def prompt_pairs(
+    cfg, **kwargs
+):  # pylint: disable=possibly-unused-variable,unused-argument
     def transform_fn(sample):
         if "system" in sample and sample["system"]:
             sample["prompt"] = (
@@ -88,7 +111,7 @@ def prompt_pairs(cfg):  # pylint: disable=possibly-unused-variable,unused-argume
     return transform_fn
 
 
-def ultra(cfg):  # pylint: disable=possibly-unused-variable,unused-argument
+def ultra(cfg, **kwargs):  # pylint: disable=possibly-unused-variable,unused-argument
     """
     for ultrafeedback binarized conversations
     """

src/axolotl/prompt_strategies/dpo/user_defined.py

@@ -0,0 +1,41 @@
+"""
+User-defined DPO strategies
+"""
+
+
+def default(cfg, dataset_idx=0, **kwargs):  # pylint: disable=unused-argument
+    ds_cfg = cfg["datasets"][dataset_idx]["type"]
+    if not isinstance(ds_cfg, dict):
+        raise ValueError(
+            f"User-defined dataset type must be a dictionary. Got: {ds_cfg}"
+        )
+    field_prompt = ds_cfg.get("field_prompt", "prompt")
+    field_system = ds_cfg.get("field_system", "system")
+    field_chosen = ds_cfg.get("field_chosen", "chosen")
+    field_rejected = ds_cfg.get("field_rejected", "rejected")
+    prompt_format = ds_cfg.get("prompt_format")
+    if not prompt_format:
+        prompt_format = "{" + field_prompt + "}"
+    chosen_format = ds_cfg.get("chosen_format")
+    if not chosen_format:
+        chosen_format = "{" + field_chosen + "}"
+    rejected_format = ds_cfg.get("rejected_format")
+    if not rejected_format:
+        rejected_format = "{" + field_rejected + "}"
+
+    def transform_fn(sample):
+        if (
+            "{" + field_system + "}" in prompt_format
+            and field_system in sample
+            and sample[field_system]
+        ):
+            sample["prompt"] = prompt_format.format(
+                system=sample[field_system], prompt=sample[field_prompt]
+            )
+        else:
+            sample["prompt"] = prompt_format.format(prompt=sample["prompt"])
+        sample["chosen"] = chosen_format.format(chosen=sample[field_chosen])
+        sample["rejected"] = rejected_format.format(rejected=sample[field_rejected])
+        return sample
+
+    return transform_fn

src/axolotl/prompt_strategies/dpo/zephyr.py

@@ -3,7 +3,7 @@ DPO strategies for zephyr
 """
 
 
-def nectar(cfg):  # pylint: disable=possibly-unused-variable,unused-argument
+def nectar(cfg, **kwargs):  # pylint: disable=possibly-unused-variable,unused-argument
     def transform_fn(sample):
         data = {}
         data["prompt"] = (

src/axolotl/prompt_strategies/input_output.py

@@ -0,0 +1,54 @@
+"""Module for plain input/output prompt pairs"""
+from typing import Generator, Tuple
+
+from axolotl.prompt_tokenizers import PromptTokenizingStrategy
+from axolotl.prompters import IGNORE_TOKEN_ID, Prompter
+
+
+class RawInputOutputStrategy(PromptTokenizingStrategy):
+    """Prompt Strategy class for input/output pairs"""
+
+    def __init__(self, *args, eos_token=None, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.eos_token = eos_token
+        if not eos_token:
+            self.eos_token = self.tokenizer.eos_token
+
+    def tokenize_prompt(self, prompt):
+        # pylint: disable=duplicate-code
+        input_ids = []
+        labels = []
+        for label, text in self.prompter.build_prompt(prompt["segments"]):
+            tokenized_output = self.tokenizer(
+                text, add_special_tokens=False, return_tensors=None
+            )["input_ids"]
+            input_ids += tokenized_output
+            if label or self.train_on_inputs:
+                labels += tokenized_output
+            else:
+                labels += [IGNORE_TOKEN_ID] * len(tokenized_output)
+
+        tokenized_prompt = {
+            "input_ids": input_ids,
+            "labels": labels,
+            "attention_mask": [1] * len(input_ids),
+        }
+
+        return tokenized_prompt
+
+
+class RawInputOutputPrompter(Prompter):
+    """prompter for raw i/o data"""
+
+    def build_prompt(self, source) -> Generator[Tuple[bool, str], None, None]:
+        for segment in source:
+            yield segment["label"], segment["text"]
+
+
+def load(tokenizer, cfg):
+    return RawInputOutputStrategy(
+        RawInputOutputPrompter(),
+        tokenizer,
+        cfg.train_on_inputs,
+        cfg.sequence_len,
+    )

src/axolotl/prompt_strategies/orpo/__init__.py

@@ -0,0 +1,9 @@
+"""
+module for ORPO style dataset transform strategies
+"""
+
+from functools import partial
+
+from ..base import load as load_base
+
+load = partial(load_base, module="axolotl.prompt_strategies.orpo")

src/axolotl/prompt_strategies/orpo/chat_template.py

@@ -0,0 +1,188 @@
+"""chatml prompt tokenization strategy for ORPO"""
+from typing import Any, Dict, Generator, List, Optional, Tuple
+
+from pydantic import BaseModel
+
+from axolotl.prompt_tokenizers import IGNORE_INDEX, PromptTokenizingStrategy
+from axolotl.prompters import Prompter
+from axolotl.utils.chat_templates import chat_templates
+
+
+class Message(BaseModel):
+    """message/turn"""
+
+    role: str
+    content: str
+    label: Optional[bool] = None
+
+
+class MessageList(BaseModel):
+    """conversation"""
+
+    messages: List[Message]
+
+
+def load(
+    tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None, **kwargs
+):  # pylint: disable=possibly-unused-variable,unused-argument
+    """
+    chatml transforms for datasets with system, input, chosen, rejected
+    """
+
+    chat_template = chat_templates("chatml")
+    if ds_cfg and "chat_template" in ds_cfg:
+        chat_template = ds_cfg["chat_template"]
+        try:
+            chat_template = chat_templates(chat_template)
+        except ValueError:
+            pass
+    tokenizer.chat_template = chat_template
+
+    return ORPOTokenizingStrategy(
+        ORPOPrompter(chat_template, tokenizer),
+        tokenizer,
+        cfg.train_on_inputs,
+        cfg.sequence_len,
+        dataset_parser=ORPODatasetParsingStrategy(),
+    )
+
+
+class ORPODatasetParsingStrategy:
+    """Strategy to parse chosen rejected dataset into messagelist"""
+
+    def get_chosen_conversation_thread(self, prompt) -> MessageList:
+        """Dataset structure mappings"""
+
+        messages: List[Message] = []
+        if system := prompt.get("system", None):
+            messages.append(Message(role="system", content=system, label=False))
+        messages.append(Message(role="user", content=prompt["prompt"], label=False))
+        messages.append(
+            Message(
+                role="assistant", content=prompt["chosen"][1]["content"], label=True
+            )
+        )
+        return MessageList(messages=messages)
+
+    def get_rejected_conversation_thread(self, prompt) -> MessageList:
+        """Dataset structure mappings"""
+
+        messages: List[Message] = []
+        if system := prompt.get("system", None):
+            messages.append(Message(role="system", content=system, label=False))
+        messages.append(Message(role="user", content=prompt["prompt"], label=False))
+        messages.append(
+            Message(
+                role="assistant", content=prompt["rejected"][1]["content"], label=True
+            )
+        )
+        return MessageList(messages=messages)
+
+
+class ORPOTokenizingStrategy(PromptTokenizingStrategy):
+    """
+    rejected_input_ids
+    input_ids
+    rejected_attention_mask
+    attention_mask
+    rejected_labels
+    labels
+    """
+
+    def __init__(
+        self,
+        *args,
+        dataset_parser=None,
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+        self.dataset_parser = dataset_parser
+
+    def tokenize_prompt(self, prompt):
+        # pass the rejected prompt/row to the Prompter to get the formatted prompt
+        prompt_len = 0
+        rejected_message_list = self.dataset_parser.get_rejected_conversation_thread(
+            prompt
+        )
+        input_ids = []
+        labels = []
+        for _, (part, label) in enumerate(
+            self.prompter.build_prompt(rejected_message_list)
+        ):
+            if not part:
+                continue
+            _input_ids = self.tokenizer.encode(part, add_special_tokens=False)
+            prev_idx = len(input_ids)
+            input_ids += _input_ids[prev_idx:]
+            if label:
+                labels += input_ids[prev_idx:]
+            else:
+                labels += [IGNORE_INDEX] * (len(input_ids) - prev_idx)
+                prompt_len = len(input_ids)
+        # remap the input_ids, attention_mask and labels
+        rejected_input_ids = input_ids
+        rejected_labels = labels
+        # pass the chosen prompt/row to the Prompter to get the formatted prompt
+        chosen_message_list = self.dataset_parser.get_chosen_conversation_thread(prompt)
+        input_ids = []
+        labels = []
+        for _, (part, label) in enumerate(
+            self.prompter.build_prompt(chosen_message_list)
+        ):
+            if not part:
+                continue
+            _input_ids = self.tokenizer.encode(part, add_special_tokens=False)
+            prev_idx = len(input_ids)
+            input_ids += _input_ids[prev_idx:]
+            if label:
+                labels += input_ids[prev_idx:]
+            else:
+                labels += [IGNORE_INDEX] * (len(input_ids) - prev_idx)
+
+        return {
+            "rejected_input_ids": rejected_input_ids,
+            "rejected_labels": rejected_labels,
+            "rejected_attention_mask": [1] * len(rejected_labels),
+            "input_ids": input_ids,
+            "labels": labels,
+            "attention_mask": [1] * len(labels),
+            "prompt_attention_mask": [1] * prompt_len
+            + [0] * (len(labels) - prompt_len),
+        }
+
+
+class ORPOPrompter(Prompter):
+    """Single Turn prompter for ORPO"""
+
+    def __init__(self, chat_template, tokenizer):
+        self.chat_template = chat_template
+        self.tokenizer = tokenizer
+
+    def build_prompt(
+        self,
+        message_list: MessageList,
+    ) -> Generator[Tuple[str, bool], None, None]:
+        conversation = []
+        for message in message_list.messages:
+            conversation.append(message.model_dump())
+            if message.role == "system":
+                yield self.tokenizer.apply_chat_template(
+                    conversation,
+                    add_generation_prompt=False,
+                    chat_template=self.chat_template,
+                    tokenize=False,
+                ), False
+            if message.role == "user":
+                yield self.tokenizer.apply_chat_template(
+                    conversation,
+                    add_generation_prompt=True,
+                    chat_template=self.chat_template,
+                    tokenize=False,
+                ), False
+            if message.role == "assistant":
+                yield self.tokenizer.apply_chat_template(
+                    conversation,
+                    add_generation_prompt=False,
+                    chat_template=self.chat_template,
+                    tokenize=False,
+                ), True

src/axolotl/prompt_strategies/sharegpt.py

@@ -1,10 +1,18 @@
 """Module containing the SimpleShareGPTPromptTokenizingStrategy class"""
+
+import logging
 from typing import Any, Dict, Optional
 
 from fastchat.conversation import Conversation, SeparatorStyle, register_conv_template
 
 from axolotl.prompt_tokenizers import ShareGPTPromptTokenizingStrategy
 from axolotl.prompters import ShareGPTPrompterV2
+from axolotl.utils.tokenization import (
+    chatml_to_conversation,
+    merge_consecutive_messages,
+)
+
+LOG = logging.getLogger("axolotl")
 
 
 def register_chatml_template(system_message=None):
@@ -19,6 +27,16 @@ def register_chatml_template(system_message=None):
             sep="<|im_end|>",
         )
     )
+    register_conv_template(
+        Conversation(
+            name="chatml_glaive",
+            system_template="<|im_start|>system\n{system_message}",
+            system_message=system_message,
+            roles=["<|im_start|>user", "<|im_start|>assistant", "<|im_start|>tool"],
+            sep_style=SeparatorStyle.CHATML,
+            sep="<|im_end|>",
+        )
+    )
 
 
 def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
@@ -27,11 +45,13 @@ def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
     )
     field_human = ds_cfg["field_human"] if ds_cfg and "field_human" in ds_cfg else None
     field_model = ds_cfg["field_model"] if ds_cfg and "field_model" in ds_cfg else None
+    roles = ds_cfg["roles"].to_dict() if ds_cfg and "roles" in ds_cfg else None
     strategy = SimpleShareGPTPromptTokenizingStrategy(
         ShareGPTPrompterV2(
             conversation=conversation,
             role_key_model=field_model,
             role_key_human=field_human,
+            roles=roles,
         ),
         tokenizer,
         cfg.train_on_inputs,
@@ -77,12 +97,26 @@ def load_guanaco(tokenizer, cfg):
     )
 
 
+def load_glaive(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
+    conversation = (
+        ds_cfg["conversation"]
+        if ds_cfg and "conversation" in ds_cfg
+        else "chatml_glaive"
+    )
+    return GlaiveShareGPTPromptTokenizingStrategy(
+        ShareGPTPrompterV2(conversation=conversation),
+        tokenizer,
+        cfg.train_on_inputs,
+        cfg.sequence_len,
+    )
+
+
 class SimpleShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
     """
     basic sharegpt strategy to grab conversations from the sample row
     """
 
-    _strict = True
+    _strict = False
 
     @property
     def strict(self):
@@ -96,10 +130,30 @@ class SimpleShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
         conversations = prompt["conversations"]
         if self.strict:
             return conversations
-        # remap roles - allow for assistant turn
-        role_map = {"human": "human", "assistant": "gpt", "gpt": "gpt"}
+        role_key = "from"
+        if "role" in conversations[0].keys():
+            role_key = "role"
+        value_key = "value"
+        if "text" in conversations[0].keys():
+            value_key = "text"
+        elif "content" in conversations[0].keys():
+            value_key = "content"
+        # remap roles - allow for assistant turn"
+        role_map = {
+            "user": "human",
+            "human": "human",
+            "assistant": "gpt",
+            "gpt": "gpt",
+            "system": "system",
+        }
         turns = [
-            {"from": role_map[t["from"]], "value": t["value"]} for t in conversations
+            {
+                "from": (
+                    role_map[t[role_key]] if t[role_key] in role_map else t[role_key]
+                ),
+                "value": t[value_key],
+            }
+            for t in conversations
         ]
         return turns
 
@@ -143,3 +197,15 @@ class UltrachatShareGPTPromptTokenizingStrategy(SimpleShareGPTPromptTokenizingSt
             {"from": role_map[t["role"]], "value": t["content"]} for t in conversations
         ]
         return turns
+
+
+class GlaiveShareGPTPromptTokenizingStrategy(SimpleShareGPTPromptTokenizingStrategy):
+    """
+    sharegpt strategy that remaps glaive data to sharegpt format
+    """
+
+    def get_conversation_thread(self, prompt):
+        conversation = chatml_to_conversation(prompt)
+        conversation = merge_consecutive_messages(conversation)
+
+        return conversation

src/axolotl/prompt_tokenizers.py

@@ -11,7 +11,7 @@ from transformers import BatchEncoding, PreTrainedTokenizer
 from axolotl.monkeypatch.fastchat_conversation_turns import (
     add_get_turns_to_conversation,
 )
-from axolotl.prompters import IGNORE_TOKEN_ID
+from axolotl.prompters import IGNORE_TOKEN_ID, Prompter
 
 LOG = logging.getLogger("axolotl")
 
@@ -37,7 +37,7 @@ class PromptTokenizingStrategy(abc.ABC):
 
     def __init__(
         self,
-        prompter,
+        prompter: Prompter,
         tokenizer,
         train_on_inputs: bool = False,
         sequence_len: int = 2048,
@@ -340,6 +340,23 @@ class ShareGPTPromptTokenizingStrategy(PromptTokenizingStrategy):
             self.prompter._conversation.copy()  # pylint: disable=protected-access
         )
 
+        input_roles = {conversation.roles[0]}
+        output_roles = {conversation.roles[1]}
+
+        if len(conversation.roles) == 3:
+            tool_role_label = conversation.roles[2]
+            input_roles.add(tool_role_label)
+
+        # Add roles from the config
+        if self.prompter.roles:
+            if "input" in self.prompter.roles and self.prompter.roles["input"]:
+                for role in self.prompter.roles["input"]:
+                    input_roles.add(role)
+
+            if "output" in self.prompter.roles and self.prompter.roles["output"]:
+                for role in self.prompter.roles["output"]:
+                    output_roles.add(role)
+
         # support for custom roles from the dataset, only useful for vicuna style prompts/roles
         role_remap = []
         if (
@@ -360,11 +377,18 @@ class ShareGPTPromptTokenizingStrategy(PromptTokenizingStrategy):
                     LOG.warning(f"expected tuple, got {part}")
                     continue
 
-                user, assistant = conversation.roles
                 role, content = part
 
                 # Uses "in" because role contains extra characters
-                if user in role:
+                input_turn = any(r.lower() in role.lower() for r in input_roles)
+                output_turn = any(r.lower() in role.lower() for r in output_roles)
+                empty_role = role.strip() == ""
+
+                if not any([input_turn, output_turn, empty_role]):
+                    LOG.warning(f"unhandled role: {role}")
+                    continue
+
+                if input_turn:
                     role = (
                         role.replace(role_remap[0]["from"], role_remap[0]["to"])
                         if role_remap
@@ -384,7 +408,7 @@ class ShareGPTPromptTokenizingStrategy(PromptTokenizingStrategy):
                     else:
                         # everything from this is masked out from the labels
                         labels = [IGNORE_TOKEN_ID] * len(res["input_ids"])
-                elif assistant in role:
+                elif output_turn:
                     role = (
                         role.replace(role_remap[1]["from"], role_remap[1]["to"])
                         if role_remap
@@ -415,7 +439,7 @@ class ShareGPTPromptTokenizingStrategy(PromptTokenizingStrategy):
                         labels[:len_role] = [IGNORE_TOKEN_ID] * min(
                             len_role, len(labels)
                         )
-                elif role == "":
+                elif empty_role:
                     turn = content
                     # this is only ever the first part, should include the bos token and the user query
                     res = self._tokenize(
@@ -426,9 +450,6 @@ class ShareGPTPromptTokenizingStrategy(PromptTokenizingStrategy):
                     else:
                         # everything from this is masked out from the labels
                         labels = [IGNORE_TOKEN_ID] * len(res["input_ids"])
-                else:
-                    LOG.warning(f"unhandled role: {role}")
-                    continue
 
                 # pylint: disable=duplicate-code
                 result, current_len = parse_tokenized_to_result(

src/axolotl/prompters.py

@@ -259,6 +259,12 @@ SHAREGPT_ASSERTION_FAILED_ROLE = (
     "Role did not alternate between turns (gpt and human). Please check your data."
 )
 
+CONVERSATION_ROLE_FORMAT = {
+    "chatml": "<|im_start|>{ROLE}",
+    "zephyr": "<|{ROLE}|>",
+    "vicuna_v1.1": "{ROLE}",
+}
+
 
 class ShareGPTPrompter(Prompter):  # pylint: disable=too-few-public-methods
     """
@@ -267,6 +273,10 @@ class ShareGPTPrompter(Prompter):  # pylint: disable=too-few-public-methods
 
     role_key_human = "human"
     role_key_model = "gpt"
+    # Optional, only used for tool usage datasets.
+    role_key_tool: Optional[str] = None
+    # Optional, role input/output mapping
+    roles: Optional[dict] = None
 
     def __init__(
         self,
@@ -274,6 +284,8 @@ class ShareGPTPrompter(Prompter):  # pylint: disable=too-few-public-methods
         conversation: Optional[Union[str, Conversation]] = None,
         role_key_human: Optional[str] = None,
         role_key_model: Optional[str] = None,
+        role_key_tool: Optional[str] = None,
+        roles: Optional[dict] = None,
     ):
         if conversation:
             if isinstance(conversation, Conversation):
@@ -286,6 +298,10 @@ class ShareGPTPrompter(Prompter):  # pylint: disable=too-few-public-methods
             self.role_key_human = role_key_human
         if role_key_model:
             self.role_key_model = role_key_model
+        if role_key_tool:
+            self.role_key_tool = role_key_tool
+        if roles:
+            self.roles = roles
 
     def _build_result(self, source):
         if len(source) < 2:
@@ -303,6 +319,8 @@ class ShareGPTPrompter(Prompter):  # pylint: disable=too-few-public-methods
             source.pop(0)
 
         roles = {self.role_key_human: conv.roles[0], self.role_key_model: conv.roles[1]}
+        if self.role_key_tool:
+            roles[self.role_key_tool] = conv.roles[2]
 
         try:
             # Apply prompt templates
@@ -315,11 +333,23 @@ class ShareGPTPrompter(Prompter):  # pylint: disable=too-few-public-methods
 
         conv.messages = []
         for _, sentence in enumerate(source):
-            role = roles[sentence["from"]]
-            if len(conv.messages) > 0 and (
-                (role == conv.messages[-1][0]) or (role not in conv.roles)
-            ):
+            from_role = sentence["from"]
+            if from_role in roles:
+                role = roles[from_role]
+            else:
+                if self._conversation.name not in CONVERSATION_ROLE_FORMAT:
+                    raise NotImplementedError(
+                        f"Role ({role}) not in default roles, and {self._conversation.name} does not support role remapping yet."
+                        "Please help us by creating an Issue to add support for this conversation type."
+                    )
+
+                role = CONVERSATION_ROLE_FORMAT[self._conversation.name].format(
+                    ROLE=from_role
+                )
+
+            if len(conv.messages) > 0 and ((role == conv.messages[-1][0])):
                 LOG.warning(f"{SHAREGPT_ASSERTION_FAILED_ROLE}: {sentence}")
+
             conv.append_message(role, sentence["value"])
 
         return conv.get_turns()
@@ -347,11 +377,13 @@ class ShareGPTPrompterV2(ShareGPTPrompter):
         conversation: Optional[Union[str, Conversation]] = None,
         role_key_human: Optional[str] = None,
         role_key_model: Optional[str] = None,
+        roles: Optional[dict] = None,
     ):
         super().__init__(
             conversation=conversation,
             role_key_human=role_key_human,
             role_key_model=role_key_model,
+            roles=roles,
         )
 
 

src/axolotl/train.py

@@ -19,7 +19,7 @@ from transformers.integrations.deepspeed import is_deepspeed_zero3_enabled
 from axolotl.common.cli import TrainerCliArgs
 from axolotl.logging_config import configure_logging
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.freeze import freeze_parameters_except
+from axolotl.utils.freeze import freeze_layers_except
 from axolotl.utils.models import load_model, load_tokenizer
 from axolotl.utils.trainer import setup_trainer
 
@@ -85,7 +85,7 @@ def train(
     model.generation_config.do_sample = True
 
     model_ref = None
-    if cfg.rl:
+    if cfg.rl and cfg.rl != "orpo":
         if cfg.adapter and not cfg.rl_adapter_ref_model:
             # use built-in trl autounwrap
             LOG.debug("Passing model_ref: None to RL trainer")
@@ -99,7 +99,7 @@ def train(
     safe_serialization = cfg.save_safetensors is True
 
     if cfg.unfrozen_parameters:
-        freeze_parameters_except(model, cfg.unfrozen_parameters)
+        freeze_layers_except(model, cfg.unfrozen_parameters)
 
     trainer = setup_trainer(
         cfg,
@@ -110,9 +110,6 @@ def train(
         total_num_steps,
     )
 
-    if hasattr(model, "config"):
-        model.config.use_cache = False
-
     # go ahead and presave, so we have the adapter config available to inspect
     if peft_config:
         LOG.info(f"Pre-saving adapter config to {cfg.output_dir}")

src/axolotl/utils/bench.py

@@ -24,9 +24,9 @@ def check_cuda_device(default_value):
                 or not torch.cuda.is_available()
                 or device == "auto"
                 or torch.device(device).type == "cpu"
+                or torch.device(device).type == "meta"
             ):
                 return default_value
-
             return func(*args, **kwargs)
 
         return wrapper

src/axolotl/utils/callbacks/__init__.py

@@ -9,7 +9,6 @@ from tempfile import NamedTemporaryFile
 from typing import TYPE_CHECKING, Dict, List
 
 import evaluate
-import mlflow
 import numpy as np
 import pandas as pd
 import torch
@@ -42,8 +41,8 @@ from axolotl.utils.distributed import (
 if TYPE_CHECKING:
     from axolotl.core.trainer_builder import AxolotlTrainingArguments
 
-LOG = logging.getLogger("axolotl.callbacks")
 IGNORE_INDEX = -100
+LOG = logging.getLogger("axolotl.callbacks")
 
 
 class EvalFirstStepCallback(
@@ -756,31 +755,3 @@ class SaveAxolotlConfigtoWandBCallback(TrainerCallback):
             except (FileNotFoundError, ConnectionError) as err:
                 LOG.warning(f"Error while saving Axolotl config to WandB: {err}")
         return control
-
-
-class SaveAxolotlConfigtoMlflowCallback(TrainerCallback):
-    """Callback to save axolotl config to mlflow"""
-
-    def __init__(self, axolotl_config_path):
-        self.axolotl_config_path = axolotl_config_path
-
-    def on_train_begin(
-        self,
-        args: AxolotlTrainingArguments,  # pylint: disable=unused-argument
-        state: TrainerState,  # pylint: disable=unused-argument
-        control: TrainerControl,
-        **kwargs,  # pylint: disable=unused-argument
-    ):
-        if is_main_process():
-            try:
-                with NamedTemporaryFile(
-                    mode="w", delete=False, suffix=".yml", prefix="axolotl_config_"
-                ) as temp_file:
-                    copyfile(self.axolotl_config_path, temp_file.name)
-                    mlflow.log_artifact(temp_file.name, artifact_path="")
-                    LOG.info(
-                        "The Axolotl config has been saved to the MLflow artifacts."
-                    )
-            except (FileNotFoundError, ConnectionError) as err:
-                LOG.warning(f"Error while saving Axolotl config to MLflow: {err}")
-        return control

src/axolotl/utils/callbacks/lisa.py

@@ -0,0 +1,91 @@
+"""
+module for LISA
+
+Adapted from https://github.com/OptimalScale/LMFlow/pull/701 for HF transformers & Axolotl
+Arxiv: https://arxiv.org/abs/2403.17919
+License: Apache 2.0
+"""
+
+import logging
+from functools import reduce
+from typing import TYPE_CHECKING
+
+import numpy as np
+from transformers import TrainerCallback
+
+if TYPE_CHECKING:
+    from axolotl.core.trainer_builder import AxolotlTrainer
+
+LOG = logging.getLogger("axolotl.callbacks.lisa")
+
+
+def lisa_callback_factory(trainer: "AxolotlTrainer"):
+    class LISACallback(TrainerCallback):
+        """trainer callback for lisa layer switching"""
+
+        def __init__(
+            self, n_layers, step_interval, trainer, layers_attribute="model.layers"
+        ):
+            super().__init__()
+            self.n_layers = n_layers
+            self.step_interval = step_interval
+            self.layers_attribute = layers_attribute
+            self.trainer = trainer
+
+            reduce(getattr, self.layers_attribute.split("."), self.trainer.model)
+
+            self.total_layers = len(
+                reduce(getattr, self.layers_attribute.split("."), self.trainer.model)
+            )
+            self.active_layers_indices = []
+
+            layers = reduce(
+                getattr, self.layers_attribute.split("."), self.trainer.model
+            )
+            LOG.info(
+                f"LISA will activate {self.n_layers}/{len(layers)} layers ({self.n_layers*100/len(layers)}%) every {self.step_interval} steps"
+            )
+
+        def freeze_all_layers(self):
+            layers = reduce(
+                getattr, self.layers_attribute.split("."), self.trainer.model
+            )
+            for layer in layers:
+                for param in layer.parameters():
+                    param.requires_grad = False
+
+        def on_step_begin(
+            self, args, state, control, **kwargs
+        ):  # pylint: disable=unused-argument
+            # Check if it's time to switch active layers, including at step 0
+            if state.global_step % self.step_interval == 0 or state.global_step == 1:
+                self.switch_active_layers()
+
+        def switch_active_layers(self):
+            # First, disable gradients for all layers
+            self.freeze_all_layers()
+
+            # Randomly select n_layers to activate
+            layers = reduce(
+                getattr, self.layers_attribute.split("."), self.trainer.model
+            )
+            self.active_layers_indices = np.random.choice(
+                range(self.total_layers), self.n_layers, replace=False
+            )
+            LOG.info(
+                f"Activating layers at indices: {self.active_layers_indices} for the next steps."
+            )
+
+            # Enable gradients only for the selected layers
+            for idx in self.active_layers_indices:
+                for param in layers[idx].parameters():
+                    param.requires_grad = True
+
+    lisa_callback = LISACallback(
+        n_layers=trainer.args.lisa_n_layers,
+        step_interval=trainer.args.lisa_step_interval,
+        trainer=trainer,
+        layers_attribute=trainer.args.lisa_layers_attribute,
+    )
+
+    return lisa_callback

src/axolotl/utils/callbacks/mlflow_.py

@@ -0,0 +1,44 @@
+"""MLFlow module for trainer callbacks"""
+import logging
+from shutil import copyfile
+from tempfile import NamedTemporaryFile
+from typing import TYPE_CHECKING
+
+import mlflow
+from transformers import TrainerCallback, TrainerControl, TrainerState
+
+from axolotl.utils.distributed import is_main_process
+
+if TYPE_CHECKING:
+    from axolotl.core.trainer_builder import AxolotlTrainingArguments
+
+LOG = logging.getLogger("axolotl.callbacks")
+
+
+class SaveAxolotlConfigtoMlflowCallback(TrainerCallback):
+    # pylint: disable=duplicate-code
+    """Callback to save axolotl config to mlflow"""
+
+    def __init__(self, axolotl_config_path):
+        self.axolotl_config_path = axolotl_config_path
+
+    def on_train_begin(
+        self,
+        args: "AxolotlTrainingArguments",  # pylint: disable=unused-argument
+        state: TrainerState,  # pylint: disable=unused-argument
+        control: TrainerControl,
+        **kwargs,  # pylint: disable=unused-argument
+    ):
+        if is_main_process():
+            try:
+                with NamedTemporaryFile(
+                    mode="w", delete=False, suffix=".yml", prefix="axolotl_config_"
+                ) as temp_file:
+                    copyfile(self.axolotl_config_path, temp_file.name)
+                    mlflow.log_artifact(temp_file.name, artifact_path="")
+                    LOG.info(
+                        "The Axolotl config has been saved to the MLflow artifacts."
+                    )
+            except (FileNotFoundError, ConnectionError) as err:
+                LOG.warning(f"Error while saving Axolotl config to MLflow: {err}")
+        return control

src/axolotl/utils/chat_templates.py

@@ -21,7 +21,8 @@ def chat_templates(user_choice: str):
     templates = {
         "alpaca": "{% for message in messages %}{% if message['role'] == 'user' %}{{ '### Instruction: ' + message['content'] + '\n\n' }}{% elif message['role'] == 'assistant' %}{{ '### Response: ' + message['content'] + eos_token}}{% endif %}{% endfor %}",
         "inst": "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if message['role'] == 'user' %}{{ '[INST] ' + message['content'] + ' [/INST]' }}{% elif message['role'] == 'assistant' %}{{ message['content'] + eos_token}}{% else %}{{ raise_exception('Only user and assistant roles are supported!') }}{% endif %}{% endfor %}",  # I don't know what this one is called. Used by Mistral/Mixtral.
-        "chatml": "{% if messages[0]['role'] == 'system' %}{% set loop_messages = messages[1:] %}{% set system_message = messages[0]['content'] %}{% else %}{% set loop_messages = messages %}{% set system_message = 'You are a helpful assistant.' %}{% endif %}{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in loop_messages %}{% if loop.index0 == 0 %}{{'<|im_start|>system\n' + system_message + '<|im_end|>\n'}}{% endif %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}",
+        "chatml": "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}",
+        "gemma": "{{ bos_token }}{% if messages[0]['role'] == 'system' %}{{ raise_exception('System role not supported') }}{% endif %}{% for message in messages %}{% if (message['role'] == 'user') != (loop.index0 % 2 == 0) %}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{% endif %}{% if (message['role'] == 'assistant') %}{% set role = 'model' %}{% else %}{% set role = message['role'] %}{% endif %}{{ '<start_of_turn>' + role + '\n' + message['content'] | trim + '<end_of_turn>\n' }}{% endfor %}{% if add_generation_prompt %}{{'<start_of_turn>model\n'}}{% endif %}",
     }
 
     if user_choice in templates:

src/axolotl/utils/config/__init__.py

@@ -3,11 +3,16 @@ import json
 import logging
 import os
 from pathlib import Path
+from typing import Optional
 
 import torch
 from transformers.utils import is_torch_bf16_gpu_available
 
 from axolotl.utils.bench import log_gpu_memory_usage
+from axolotl.utils.config.models.input.v0_4_1 import (
+    AxolotlConfigWCapabilities,
+    AxolotlInputConfig,
+)
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.models import load_model_config
 
@@ -114,12 +119,16 @@ def normalize_config(cfg):
     model_config = load_model_config(cfg)
     cfg.model_config_type = model_config.model_type
 
+    cfg.tokenizer_config = (
+        cfg.tokenizer_config or cfg.base_model_config or cfg.base_model
+    )
+
     # figure out if the model is llama
     cfg.is_llama_derived_model = (
         (hasattr(model_config, "model_type") and model_config.model_type == "llama")
         or cfg.is_llama_derived_model
         or "llama" in cfg.base_model.lower()
-        or (cfg.model_type and "llama" in cfg.model_type.lower())
+        or (cfg.type_of_model and "llama" in cfg.type_of_model.lower())
     )
 
     # figure out if the model is falcon
@@ -135,7 +144,7 @@ def normalize_config(cfg):
         )
         or cfg.is_falcon_derived_model
         or "falcon" in cfg.base_model.lower()
-        or (cfg.model_type and "rwforcausallm" in cfg.model_type.lower())
+        or (cfg.type_of_model and "rwforcausallm" in cfg.type_of_model.lower())
     )
 
     cfg.is_mistral_derived_model = (
@@ -148,7 +157,7 @@ def normalize_config(cfg):
         )
         or cfg.is_mistral_derived_model
         or "mistral" in cfg.base_model.lower().split("/")[-1]
-        or (cfg.model_type and "mistral" in cfg.model_type.lower())
+        or (cfg.type_of_model and "mistral" in cfg.type_of_model.lower())
     )
 
     cfg.is_qwen_derived_model = (
@@ -159,9 +168,6 @@ def normalize_config(cfg):
         ]
     ) or cfg.is_qwen_derived_model
 
-    if isinstance(cfg.learning_rate, str):
-        cfg.learning_rate = float(cfg.learning_rate)
-
     if isinstance(cfg.pretraining_dataset, dict):
         cfg.pretraining_dataset = [cfg.pretraining_dataset]
 
@@ -189,9 +195,28 @@ def normalize_cfg_datasets(cfg):
                         f"updating dataset {ds_cfg.path} with `conversation: chatml` to match your chat_template"
                     )
                     cfg.datasets[idx].conversation = "chatml"
+                if ds_cfg.type == "orpo.chat_template" and not ds_cfg.chat_template:
+                    LOG.info(
+                        f"updating dataset {ds_cfg.path} with `chat_template: chatml` to match your chat_template"
+                    )
+                    cfg.datasets[idx].chat_template = "chatml"
 
 
-def validate_config(cfg):
+def validate_config(cfg: DictDefault, capabilities: Optional[dict] = None):
+    if capabilities:
+        return DictDefault(
+            dict(
+                AxolotlConfigWCapabilities(
+                    **cfg.to_dict(), capabilities=capabilities
+                ).model_dump(exclude_none=True)
+            )
+        )
+    return DictDefault(
+        dict(AxolotlInputConfig(**cfg.to_dict()).model_dump(exclude_none=True))
+    )
+
+
+def legacy_validate_config(cfg):
     """
     This is a "pre-validation" step that handles the yaml configuration before we have any
     information about the model architecture
@@ -363,11 +388,11 @@ def validate_config(cfg):
             "hub_model_id is set without any models being saved. To save a model, set either save_steps or saves_per_epoch."
         )
 
-    if cfg.gptq and cfg.model_revision:
+    if cfg.gptq and cfg.revision_of_model:
         raise ValueError(
-            "model_revision is not supported for GPTQ models. "
+            "revision_of_model is not supported for GPTQ models. "
             + "Please download the model from HuggingFace Hub manually for correct branch, "
-            + "point to its path, and remove model_revision from the config."
+            + "point to its path, and remove revision_of_model from the config."
         )
 
     # if cfg.sample_packing and cfg.sdp_attention:
@@ -480,9 +505,6 @@ def validate_config(cfg):
     if cfg.rope_scaling:
         LOG.warning("`rope_scaling` should now be be a key under `model_config`")
 
-    if cfg.warmup_steps and cfg.warmup_ratio:
-        raise ValueError("warmup_steps and warmup_ratio are mutually exclusive")
-
     if cfg.wandb_run_id and not cfg.wandb_name:
         cfg.wandb_name = cfg.wandb_run_id
 

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

@@ -0,0 +1,14 @@
+"""module for gpu capabilities"""
+from typing import Optional
+
+from pydantic import BaseModel, Field
+
+
+class GPUCapabilities(BaseModel):
+    """model to manage the gpu capabilities statically"""
+
+    bf16: bool = Field(default=False)
+    fp8: bool = Field(default=False)
+    n_gpu: int = Field(default=1)
+    n_node: int = Field(default=1)
+    compute_capability: Optional[str] = Field(default=None)

src/axolotl/utils/data.py

@@ -1,4 +1,5 @@
 """Module containing data utilities"""
+
 import functools
 import hashlib
 import logging
@@ -134,7 +135,7 @@ def load_tokenized_prepared_datasets(
     split="train",
 ) -> Tuple[DatasetDict, List[Prompter]]:
     cfg_datasets = cfg.test_datasets if split == "test" else cfg.datasets
-    tokenizer_name = tokenizer.__class__.__name__
+    tokenizer_name = cfg.tokenizer_config
     ds_hash = str(
         md5(
             (
@@ -223,7 +224,7 @@ def load_tokenized_prepared_datasets(
                     token=use_auth_token,
                 )
                 ds_from_hub = True
-            except (FileNotFoundError, ConnectionError, HFValidationError):
+            except (FileNotFoundError, ConnectionError, HFValidationError, ValueError):
                 pass
 
             ds_from_cloud = False
@@ -290,14 +291,17 @@ def load_tokenized_prepared_datasets(
             local_path = Path(config_dataset.path)
             if local_path.exists():
                 if local_path.is_dir():
-                    # TODO dirs with arrow or parquet files could be loaded with `load_from_disk`
-                    ds = load_dataset(
-                        config_dataset.path,
-                        name=config_dataset.name,
-                        data_files=config_dataset.data_files,
-                        streaming=False,
-                        split=None,
-                    )
+                    if config_dataset.data_files:
+                        ds_type = get_ds_type(config_dataset)
+                        ds = load_dataset(
+                            ds_type,
+                            name=config_dataset.name,
+                            data_files=config_dataset.data_files,
+                            streaming=False,
+                            split=None,
+                        )
+                    else:
+                        ds = load_from_disk(config_dataset.path)
                 elif local_path.is_file():
                     ds_type = get_ds_type(config_dataset)
 
@@ -415,8 +419,11 @@ def load_tokenized_prepared_datasets(
         dataset = concatenate_datasets(datasets)
 
         if len(datasets) > 1:
-            LOG.info("shuffle merged datasets")
-            dataset = dataset.shuffle(seed=seed)
+            if cfg.shuffle_merged_datasets:
+                LOG.debug("shuffle merged datasets")
+                dataset = dataset.shuffle(seed=seed)
+            else:
+                LOG.debug("NOT shuffling merged datasets")
 
         dataset, _ = process_datasets_for_packing(cfg, dataset, None)
 
@@ -819,7 +826,11 @@ def wrap_pretraining_dataset(
     else:
         encode = functools.partial(encode_pretraining, tokenizer, max_tokens)
 
-    dataset = dataset.shuffle(seed=seed, buffer_size=buffer_size)
+    if cfg.shuffle_merged_datasets:
+        dataset = dataset.shuffle(seed=seed, buffer_size=buffer_size)
+    else:
+        LOG.debug("NOT shuffling merged pretraining datasets")
+
     dataset = dataset.map(
         encode,
         batched=True,
@@ -937,7 +948,9 @@ def load_prepare_dpo_datasets(cfg):
         for i, data_set in enumerate(split_datasets):
             _type = dataset_cfgs[i]["type"]
             if _type:
-                ds_transform_fn = load_dpo(_type, _cfg)
+                if isinstance(_type, DictDefault):
+                    _type = "user_defined.default"
+                ds_transform_fn = load_dpo(_type, _cfg, dataset_idx=i)
                 split_datasets[i] = data_set.map(
                     ds_transform_fn,
                     desc="Mapping RL Dataset",

src/axolotl/utils/dict.py

@@ -12,4 +12,4 @@ class DictDefault(Dict):
         return None
 
     def __or__(self, other):
-        return DictDefault(super().__or__(other))
+        return DictDefault(super().__ror__(other))