Disable datasets caching when preparing dataset for packing

* update PR template so we can capture twitter or discord handles [skip ci]

* ensure that the PR template is in the correct place

.github/FUNDING.yml

@@ -3,11 +3,11 @@
 github: OpenAccess-AI-Collective # Replace with up to 4 GitHub Sponsors-enabled usernames e.g., [user1, user2]
 patreon: # Replace with a single Patreon username
 open_collective: # Replace with a single Open Collective username
-ko_fi: # Replace with a single Ko-fi username
+ko_fi: axolotl_ai # Replace with a single Ko-fi username
 tidelift: # Replace with a single Tidelift platform-name/package-name e.g., npm/babel
 community_bridge: # Replace with a single Community Bridge project-name e.g., cloud-foundry
 liberapay: # Replace with a single Liberapay username
 issuehunt: # Replace with a single IssueHunt username
 otechie: # Replace with a single Otechie username
 lfx_crowdfunding: # Replace with a single LFX Crowdfunding project-name e.g., cloud-foundry
-custom: # Replace with up to 4 custom sponsorship URLs e.g., ['link1', 'link2']
+custom: ['https://quickchart.io/qr?text=bitcoin%3Abc1qxlgwlqwfea5s2cxm42xqsfmwjct0rj8w8ea5np&size=480&centerImageUrl=https%3A%2F%2Fupload.wikimedia.org%2Fwikipedia%2Fcommons%2Fthumb%2F4%2F46%2FBitcoin.svg%2F64px-Bitcoin.svg.png'] # Replace with up to 4 custom sponsorship URLs e.g., ['link1', 'link2']

.github/PULL_REQUEST_TEMPLATE.md

@@ -20,3 +20,8 @@
 ## Types of changes
 
 <!--- What types of changes does your code introduce? Put an `x` in all the boxes that apply: -->
+
+## Social Handles (Optional)
+
+<!-- Thanks for submitting a bugfix or enhancement. -->
+<!-- We'd love to show our thanks to you on Twitter & Discord if you provide your handle -->

.github/workflows/base.yml

@@ -28,7 +28,12 @@ jobs:
           - cuda: "118"
             cuda_version: 11.8.0
             python_version: "3.10"
-            pytorch: 2.1.0
+            pytorch: 2.1.1
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 9.0+PTX"
+          - cuda: "121"
+            cuda_version: 12.1.0
+            python_version: "3.10"
+            pytorch: 2.1.1
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 9.0+PTX"
     steps:
       - name: Checkout

.github/workflows/lint.yml

@@ -0,0 +1,22 @@
+name: lint
+on:
+  # check on PRs, and manual triggers
+  pull_request:
+      paths:
+       - '**.py'
+       - 'requirements.txt'
+       - '.github/workflows/*.yml'
+       - "*.md"
+  workflow_dispatch:
+
+jobs:
+  pre-commit:
+    name: pre-commit
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v3
+      - uses: actions/setup-python@v4
+        with:
+          python-version: "3.9"
+          cache: 'pip' # caching pip dependencies
+      - uses: pre-commit/action@v3.0.0

.github/workflows/main.yml

@@ -27,38 +27,56 @@ jobs:
           - cuda: 118
             cuda_version: 11.8.0
             python_version: "3.10"
-            pytorch: 2.1.0
+            pytorch: 2.1.1
+            axolotl_extras:
+          - cuda: 121
+            cuda_version: 12.1.0
+            python_version: "3.10"
+            pytorch: 2.1.1
             axolotl_extras:
     runs-on: [self-hosted, gpu, docker]
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
       - name: Docker metadata
         id: metadata
-        uses: docker/metadata-action@v3
+        uses: docker/metadata-action@v5
         with:
           images: winglian/axolotl
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
       - name: Login to Docker Hub
-        uses: docker/login-action@v2
+        uses: docker/login-action@v3
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
           password: ${{ secrets.DOCKERHUB_TOKEN }}
-      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v2
-      - name: Build
-        uses: docker/build-push-action@v4
+      # guidance for testing before pushing: https://docs.docker.com/build/ci/github-actions/test-before-push/
+      - name: Build and export to Docker
+        uses: docker/build-push-action@v5
         with:
           context: .
+          load: true
           build-args: |
             BASE_TAG=${{ github.ref_name }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}
             CUDA=${{ matrix.cuda }}
             PYTORCH_VERSION=${{ matrix.pytorch }}
           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 }}
             ${{ (matrix.is_latest) && format('{0}-latest', steps.metadata.outputs.tags) || '' }}
           labels: ${{ steps.metadata.outputs.labels }}
+      - name: Unit Tests
+        run: |
+          docker run --rm ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }} pytest --ignore=tests/e2e/ /workspace/axolotl/tests/
+      - name: Push to Docker Hub
+        if: github.event_name != 'pull_request'
+        run: |
+          docker push ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
+          latest_tag=${{ (matrix.is_latest) && format('{0}-latest', steps.metadata.outputs.tags) || '' }}
+          if [ -n "$latest_tag" ]; then
+            docker push "$latest_tag"
+          fi
+
   build-axolotl-runpod:
     needs: build-axolotl
     if: github.repository_owner == 'OpenAccess-AI-Collective'
@@ -80,26 +98,31 @@ jobs:
           - cuda: 118
             cuda_version: 11.8.0
             python_version: "3.10"
-            pytorch: 2.1.0
+            pytorch: 2.1.1
+            axolotl_extras:
+          - cuda: 121
+            cuda_version: 12.1.0
+            python_version: "3.10"
+            pytorch: 2.1.1
             axolotl_extras:
     runs-on: [self-hosted, gpu, docker]
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
       - name: Docker metadata
         id: metadata
-        uses: docker/metadata-action@v3
+        uses: docker/metadata-action@v5
         with:
           images: winglian/axolotl-runpod
       - name: Login to Docker Hub
-        uses: docker/login-action@v2
+        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@v4
+        uses: docker/build-push-action@v5
         with:
           context: .
           build-args: |

.github/workflows/pypi.yml

@@ -34,11 +34,11 @@ jobs:
         run: echo ::set-output name=TAG_NAME::$(echo $GITHUB_REF | cut -d / -f 3)
 
       - name: Update version in setup.py
-        run: >-
+        run: |
           sed -i -E 's/version="([0-9.]+)",/version="${{ steps.tag.outputs.TAG_NAME }}",/g' setup.py
 
       - name: Build a binary wheel
-        run: >-
+        run: |
           python setup.py sdist bdist_wheel
 
       - name: Publish package distributions to PyPI

.github/workflows/tests.yml

@@ -7,10 +7,12 @@ on:
     paths:
       - '**.py'
       - 'requirements.txt'
+      - '.github/workflows/*.yml'
   pull_request:
       paths:
        - '**.py'
        - 'requirements.txt'
+       - '.github/workflows/*.yml'
   workflow_dispatch:
 
 jobs:
@@ -31,7 +33,7 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python_version: ["3.9", "3.10"]
+        python_version: ["3.9", "3.10", "3.11"]
     timeout-minutes: 10
 
     steps:
@@ -53,28 +55,54 @@ jobs:
         run: |
           pytest --ignore=tests/e2e/ tests/
 
-  e2e-test:
-    name: E2E Tests
-    runs-on: [self-hosted, gpu]
-    timeout-minutes: 20
+  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
     needs: [pre-commit, pytest]
 
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - cuda: 118
+            cuda_version: 11.8.0
+            python_version: "3.10"
+            pytorch: 2.0.1
+          - cuda: 121
+            cuda_version: 12.1.0
+            python_version: "3.10"
+            pytorch: 2.1.1
     steps:
-      - name: Check out repository code
-        uses: actions/checkout@v3
-
-      - name: Setup Python
-        uses: actions/setup-python@v4
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Docker metadata
+        id: metadata
+        uses: docker/metadata-action@v5
         with:
-          python-version: "3.10"
-#          cache: 'pip' # caching pip dependencies
-
-      - name: Install dependencies
+          images: winglian/axolotl-tests
+      - name: Build Docker image
         run: |
-          pip3 uninstall -y transformers accelerate
-          pip3 install -U -e .[flash-attn]
-          pip3 install -r requirements-tests.txt
-
-      - name: Run e2e tests
+          # 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
         run: |
-          pytest tests/e2e/
+          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
+        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/

.gitignore

@@ -1,5 +1,7 @@
 **/axolotl.egg-info
 configs
+last_run_prepared/
+.vscode
 
 # Byte-compiled / optimized / DLL files
 __pycache__/

.mypy.ini

@@ -8,6 +8,9 @@ ignore_missing_imports = True
 [mypy-axolotl.monkeypatch.*]
 ignore_errors = True
 
+[mypy-axolotl.models.mixtral.*]
+ignore_errors = True
+
 [mypy-axolotl.models.phi.*]
 ignore_errors = True
 

.vscode/README.md

@@ -0,0 +1 @@
+See [docs/debugging.md](../docs/debugging.md) for guidance on how to modify these files to debug axolotl with VSCode.

.vscode/launch.json

@@ -0,0 +1,34 @@
+{
+    // Use IntelliSense to learn about possible attributes.
+    // Hover to view descriptions of existing attributes.
+    // For more information, visit: https://go.microsoft.com/fwlink/?linkid=830387
+    "version": "0.2.0",
+    "configurations": [
+        {
+            "name": "Debug axolotl prompt - sharegpt",
+            "type": "python",
+            "module": "accelerate.commands.launch",
+            "request": "launch",
+            "args": [
+                "-m", "axolotl.cli.train", "dev_sharegpt.yml",
+                // The flags below simplify debugging by overriding the axolotl config
+                // with the debugging tips above.  Modify as needed.
+                "--dataset_processes=1",      // limits data preprocessing to one process
+                "--max_steps=1",              // limits training to just one step
+                "--batch_size=1",             // minimizes batch size
+                "--micro_batch_size=1",       // minimizes batch size
+                "--val_set_size=0",           // disables validation
+                "--sample_packing=False",     // disables sample packing which is necessary for small datasets
+                "--eval_sample_packing=False",// disables sample packing on eval set
+                "--dataset_prepared_path=temp_debug/axolotl_outputs/data", // send data outputs to a temp folder
+                "--output_dir=temp_debug/axolotl_outputs/model" // send model outputs to a temp folder
+                ],
+            "console": "integratedTerminal",      // show output in the integrated terminal
+            "cwd": "${workspaceFolder}/devtools", // set working directory to devtools from the root of the project
+            "justMyCode": true,                   // step through only axolotl code
+            "env": {"CUDA_VISIBLE_DEVICES": "0",  // Since we aren't doing distributed training, we need to limit to one GPU
+                    "HF_HOME": "${workspaceFolder}/devtools/temp_debug/.hf-cache"}, // send HF cache to a temp folder
+            "preLaunchTask": "cleanup-for-dataprep", // delete temp folders (see below)
+        }
+    ]
+}

.vscode/tasks.json

@@ -0,0 +1,27 @@
+//this file is used by launch.json
+{
+  "version": "2.0.0",
+  "tasks": [
+    // this task changes into the devtools directory and deletes the temp_debug/axolotl_outputs folder
+    {
+      "label": "delete-outputs",
+      "type": "shell",
+      "command": "rm -rf temp_debug/axolotl_outputs",
+      "options":{ "cwd": "${workspaceFolder}/devtools"},
+      "problemMatcher": []
+    },
+    // this task changes into the devtools directory and deletes the `temp_debug/.hf-cache/datasets` folder
+    {
+      "label": "delete-temp-hf-dataset-cache",
+      "type": "shell",
+      "command": "rm -rf temp_debug/.hf-cache/datasets",
+      "options":{ "cwd": "${workspaceFolder}/devtools"},
+      "problemMatcher": []
+    },
+      // this task combines the two tasks above
+    {
+     "label": "cleanup-for-dataprep",
+     "dependsOn": ["delete-outputs", "delete-temp-hf-dataset-cache"],
+    }
+  ]
+}

README.md

@@ -10,7 +10,7 @@ Features:
 - Integrated with xformer, flash attention, rope scaling, and multipacking
 - Works with single GPU or multiple GPUs via FSDP or Deepspeed
 - Easily run with Docker locally or on the cloud
-- Log results and optionally checkpoints to wandb
+- Log results and optionally checkpoints to wandb or mlflow
 - And more!
 
 
@@ -25,21 +25,26 @@ Features:
 - [Installation](#installation)
   - [Docker](#docker)
   - [Conda/Pip venv](#condapip-venv)
+  - [Runpod](#runpod)
   - [LambdaLabs](#lambdalabs)
   - [Windows](#windows)
+  - [Launching on public clouds via SkyPilot](#launching-on-public-clouds-via-skypilot)
 - [Dataset](#dataset)
   - [How to Add Custom Prompts](#how-to-add-custom-prompts)
   - [How to Use Custom Pretokenized Dataset](#how-to-use-your-custom-pretokenized-dataset)
 - [Config](#config)
   - [Train](#train)
-  - [Training w/ Deepspeed](#training-with-deepspeed)
   - [Inference](#inference)
   - [Merge LORA to Base](#merge-lora-to-base)
+  - [Special Tokens](#special-tokens)
 - [Common Errors](#common-errors-)
+  - [Tokenization Mismatch b/w Training & Inference](#tokenization-mismatch-bw-inference--training)
+- [Debugging Axolotl](#debugging-axolotl)
 - [Need Help?](#need-help-)
 - [Badge](#badge-)
 - [Community Showcase](#community-showcase)
 - [Contributing](#contributing-)
+- [Sponsors](#sponsors-)
 
 </td>
 <td>
@@ -64,17 +69,21 @@ Features:
 
 ## Axolotl supports
 
-|          | fp16/fp32 | lora | qlora | gptq | gptq w/flash attn | flash attn | xformers attn |
-|----------|:----------|:-----|-------|------|-------------------|------------|--------------|
-| llama    | ✅         | ✅    | ✅     | ✅             | ✅                 | ✅          | ✅            |
-| Pythia   | ✅         | ✅    | ✅     | ❌             | ❌                 | ❌          | ❓            |
-| cerebras | ✅         | ✅    | ✅     | ❌             | ❌                 | ❌          | ❓            |
-| btlm     | ✅         | ✅    | ✅     | ❌             | ❌                 | ❌          | ❓            |
-| mpt      | ✅         | ❌    | ❓     | ❌             | ❌                 | ❌          | ❓            |
-| falcon   | ✅         | ✅    | ✅     | ❌             | ❌                 | ❌          | ❓            |
-| gpt-j    | ✅         | ✅    | ✅     | ❌             | ❌                 | ❓          | ❓            |
-| XGen     | ✅         | ❓    | ✅     | ❓             | ❓                 | ❓          | ✅            |
-| phi      | ✅         | ✅    | ✅     | ❓             | ❓                 | ❓          | ❓            |
+|             | fp16/fp32 | lora | qlora | gptq | gptq w/flash attn | flash attn | xformers attn |
+|-------------|:----------|:-----|-------|------|-------------------|------------|--------------|
+| llama       | ✅         | ✅    | ✅     | ✅             | ✅                 | ✅          | ✅            |
+| Mistral     | ✅         | ✅    | ✅     | ✅             | ✅                 | ✅          | ✅            |
+| Mixtral-MoE | ✅         | ✅    | ✅     | ❓             | ❓                 | ❓          | ❓            |
+| Pythia      | ✅         | ✅    | ✅     | ❌             | ❌                 | ❌          | ❓            |
+| cerebras    | ✅         | ✅    | ✅     | ❌             | ❌                 | ❌          | ❓            |
+| btlm        | ✅         | ✅    | ✅     | ❌             | ❌                 | ❌          | ❓            |
+| mpt         | ✅         | ❌    | ❓     | ❌             | ❌                 | ❌          | ❓            |
+| falcon      | ✅         | ✅    | ✅     | ❌             | ❌                 | ❌          | ❓            |
+| gpt-j       | ✅         | ✅    | ✅     | ❌             | ❌                 | ❓          | ❓            |
+| XGen        | ✅         | ❓    | ✅     | ❓             | ❓                 | ❓          | ✅            |
+| phi         | ✅         | ✅    | ✅     | ❓             | ❓                 | ❓          | ❓            |
+| RWKV        | ✅         | ❓    | ❓     | ❓             | ❓                 | ❓          | ❓            |
+| Qwen        | ✅         | ✅    | ✅     | ❓             | ❓                 | ❓          | ❓            |
 
 
 ## Quickstart ⚡
@@ -83,20 +92,29 @@ Get started with Axolotl in just a few steps! This quickstart guide will walk yo
 
 **Requirements**: Python >=3.9 and Pytorch >=2.0.
 
+`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
 
 pip3 install packaging
 pip3 install -e '.[flash-attn,deepspeed]'
-pip3 install -U git+https://github.com/huggingface/peft.git
+```
 
+### Usage
+```bash
 # finetune lora
 accelerate launch -m axolotl.cli.train examples/openllama-3b/lora.yml
 
 # inference
 accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
     --lora_model_dir="./lora-out"
+
+# gradio
+accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
+    --lora_model_dir="./lora-out" --gradio
 ```
 
 ## Installation
@@ -107,7 +125,6 @@ accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
   ```bash
   docker run --gpus '"all"' --rm -it winglian/axolotl:main-py3.10-cu118-2.0.1
   ```
-  - `winglian/axolotl-runpod:main-latest`: for runpod or use this [direct link](https://runpod.io/gsc?template=v2ickqhz9s&ref=6i7fkpdz)
 
   Or run on the current files for development:
 
@@ -115,6 +132,30 @@ accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
   docker compose up -d
   ```
 
+>[!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).
+
+  <details>
+
+  <summary>Docker advanced</summary>
+
+  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
+  ```
+
+  It additionally:
+  * Prevents memory issues when running e.g. deepspeed (e.g. you could hit SIGBUS/signal 7 error) through `--ipc` and `--ulimit` args.
+  * Persists the downloaded HF data (models etc.) and your modifications to axolotl code through `--mount`/`-v` args.
+  * The `--name` argument simply makes it easier to refer to the container in vscode (`Dev Containers: Attach to Running Container...`) or in your terminal.
+  * The `--privileged` flag gives all capabilities to the container.
+  * The `--shm-size 10g` argument increases the shared memory size. Use this if you see `exitcode: -7` errors using deepspeed.
+
+  [More information on nvidia website](https://docs.nvidia.com/deeplearning/frameworks/user-guide/index.html#setincshmem)
+
+  </details>
+
 #### Conda/Pip venv
   1. Install python >=**3.9**
 
@@ -131,6 +172,10 @@ accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
         ```
         Get the token at huggingface.co/settings/tokens
 
+#### Runpod
+
+Use `winglian/axolotl-runpod:main-latest` or use this [direct link](https://runpod.io/gsc?template=v2ickqhz9s&ref=6i7fkpdz)
+
 #### LambdaLabs
   <details>
 
@@ -178,6 +223,28 @@ accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
 #### Windows
 Please use WSL or Docker!
 
+
+#### 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
+HF_TOKEN=xx sky launch axolotl.yaml --env HF_TOKEN
+
+# Managed spot (auto-recovery on preemption)
+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.
@@ -187,10 +254,17 @@ Have dataset(s) in one of the following format (JSONL recommended):
   ```json
   {"instruction": "...", "input": "...", "output": "..."}
   ```
-- `sharegpt`: conversations where `from` is `human`/`gpt`
+- `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
+    ```
 - `completion`: raw corpus
   ```json
   {"text": "..."}
@@ -356,6 +430,13 @@ 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:
+    - path: ...
+      type: sharegpt
+      conversation: chatml
+
   # local
   datasets:
     - path: data.jsonl # or json
@@ -367,6 +448,12 @@ See [examples](examples) for quick start. It is recommended to duplicate and mod
     - 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:
+    - path: s3://path_to_ds # Accepts folder with arrow/parquet or file path like above. Supports s3, gcs.
+      ...
   ```
 
 - loading
@@ -394,7 +481,7 @@ See [examples](examples) for quick start. It is recommended to duplicate and mod
 
 <details>
 
-<summary>All yaml options</summary>
+<summary>All yaml options (click me)</summary>
 
 ```yaml
 # This is the huggingface model that contains *.pt, *.safetensors, or *.bin files
@@ -429,6 +516,23 @@ is_falcon_derived_model:
 is_llama_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:
+  # RoPE Scaling https://github.com/huggingface/transformers/pull/24653
+  rope_scaling:
+    type: # linear | dynamic
+    factor: # float
+
+# optional overrides to the bnb 4bit quantization configuration
+# https://huggingface.co/docs/transformers/main/main_classes/quantization#transformers.BitsAndBytesConfig
+bnb_config_kwargs:
+  # These are default values
+  llm_int8_has_fp16_weight: false
+  bnb_4bit_quant_type: nf4
+  bnb_4bit_use_double_quant: true
+
 
 # Whether you are training a 4-bit GPTQ quantized model
 gptq: true
@@ -451,9 +555,14 @@ tf32: true # require >=ampere
 bfloat16: true # require >=ampere
 float16: true
 
+# Limit the memory for all available GPUs to this amount (if an integer, expressed in gigabytes); default: unset
+gpu_memory_limit: 20GiB
+# Do the LoRA/PEFT loading on CPU -- this is required if the base model is so large it takes up most or all of the available GPU VRAM, e.g. during a model and LoRA merge
+lora_on_cpu: true
+
 # A list of one or more datasets to finetune the model with
 datasets:
-  # HuggingFace dataset repo | "json" for local dataset, make sure to fill data_files
+  # HuggingFace dataset repo | s3://,gs:// path | "json" for local dataset, make sure to fill data_files
   - path: vicgalle/alpaca-gpt4
   # The type of prompt to use for training. [alpaca, sharegpt, gpteacher, oasst, reflection]
     type: alpaca # format | format:<prompt_style> (chat/instruct) | <prompt_strategies>.load_<load_fn>
@@ -461,7 +570,12 @@ datasets:
     data_files: # Optional[str] path to source data files
     shards: # Optional[int] number of shards to split data into
     name: # Optional[str] name of dataset configuration to load
-    conversation:  # Optional[str] fastchat conversation type, only used with type: sharegpt
+    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
+    field_human: # Optional[str]. Human key to use for conversation.
+    field_model: # Optional[str]. Assistant key to use for conversation.
 
   # Custom user prompt
   - path: repo
@@ -485,6 +599,12 @@ datasets:
       # For `completion` datsets only, uses the provided field instead of `text` column
       field:
 
+# use RL training: dpo, ipo, kto_pair
+rl:
+
+# Saves the desired chat template to the tokenizer_config.json for easier inferencing
+# Currently supports chatml and inst (mistral/mixtral)
+chat_template: chatml
 # Axolotl attempts to save the dataset as an arrow after packing the data together so
 # subsequent training attempts load faster, relative path
 dataset_prepared_path: data/last_run_prepared
@@ -527,10 +647,17 @@ eval_sample_packing:
 sample_packing_eff_est:
 total_num_tokens:
 
+# Passed through to transformers when loading the model when launched without accelerate
+# Use `sequential` when training w/ model parallelism to limit memory
+device_map:
+# Defines the max memory usage per gpu on the system. Passed through to transformers when loading the model.
+max_memory:
+
 # If you want to use 'lora' or 'qlora' or leave blank to train all parameters in original model
 adapter: lora
 # If you already have a lora model trained that you want to load, put that here.
-# This means after training, if you want to test the model, you should set this to the value of `lora_out_dir`.
+# This means after training, if you want to test the model, you should set this to the value of `output_dir`.
+# Note that if you merge an adapter to the base model, a new subdirectory `merged` will be created under the `output_dir`.
 lora_model_dir:
 
 # LoRA hyperparameters
@@ -557,10 +684,6 @@ lora_modules_to_save:
 #  - embed_tokens
 #  - lm_head
 
-# Once you complete training, the model will be saved to the following directory.
-# If you merge the adapter to the base model, a subdirectory `merged` will be created under this directory.
-# Make sure `lora_model_dir` points to this directory if you want to use the trained model.
-lora_out_dir:
 lora_fan_in_fan_out: false
 
 # ReLoRA configuration
@@ -570,13 +693,19 @@ relora_warmup_steps: # Number of per-restart warmup steps
 relora_cpu_offload: # True to perform lora weight merges on cpu during restarts, for modest gpu memory savings
 
 # wandb configuration if you're using it
+# Make sure your `WANDB_API_KEY` environment variable is set (recommended) or you login to wandb with `wandb login`.
 wandb_mode: # "offline" to save run metadata locally and not sync to the server, "disabled" to turn off wandb
 wandb_project: # Your wandb project name
 wandb_entity: # A wandb Team name if using a Team
 wandb_watch:
-wandb_run_id: # Set the name of your wandb run
+wandb_name: # Set the name of your wandb run
+wandb_run_id: # Set the ID of your wandb run
 wandb_log_model: # "checkpoint" to log model to wandb Artifacts every `save_steps` or "end" to log only at the end of training
 
+# mlflow configuration if you're using it
+mlflow_tracking_uri: # URI to mlflow
+mlflow_experiment_name: # Your experiment name
+
 # Where to save the full-finetuned model to
 output_dir: ./completed-model
 
@@ -591,14 +720,17 @@ gradient_accumulation_steps: 1
 # The number of samples to include in each batch. This is the number of samples sent to each GPU.
 micro_batch_size: 2
 eval_batch_size:
-num_epochs: 3
-warmup_steps: 100
+num_epochs: 4
+warmup_steps: 100  # cannot use with warmup_ratio
+warmup_ratio: 0.05  # cannot use with warmup_steps
 learning_rate: 0.00003
 lr_quadratic_warmup:
 logging_steps:
+eval_steps: # Leave empty to eval at each epoch, integers for every N steps. decimal for fraction of total steps
+evals_per_epoch: # number of times per epoch to run evals, mutually exclusive with eval_steps
 save_strategy: # Set to `no` to skip checkpoint saves
 save_steps: # Leave empty to save at each epoch
-eval_steps: # Leave empty to eval at each epoch
+saves_per_epoch: # number of times per epoch to save a checkpoint, mutually exclusive with save_steps
 save_total_limit: # Checkpoints saved at a time
 # Maximum number of iterations to train for. It precedes num_epochs which means that
 # if both are set, num_epochs will not be guaranteed.
@@ -608,6 +740,9 @@ max_steps:
 eval_table_size: # Approximate number of predictions sent to wandb depending on batch size. Enabled above 0. Default is 0
 eval_table_max_new_tokens: # Total number of tokens generated for predictions sent to wandb. Default is 128
 
+loss_watchdog_threshold: # High loss value, indicating the learning has broken down (a good estimate is ~2 times the loss at the start of training)
+loss_watchdog_patience: # Number of high-loss steps in a row before the trainer aborts (default: 3)
+
 # Save model as safetensors (require safetensors package)
 save_safetensors:
 
@@ -620,6 +755,9 @@ group_by_length: false
 
 # Whether to use gradient checkpointing https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing
 gradient_checkpointing: false
+# additional kwargs to pass to the trainer for gradient checkpointing
+# gradient_checkpointing_kwargs:
+#   use_reentrant: false
 
 # 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
@@ -628,6 +766,7 @@ early_stopping_patience: 3
 # Specify a scheduler and kwargs to use with the optimizer
 lr_scheduler: # 'one_cycle' | 'log_sweep' | empty for cosine
 lr_scheduler_kwargs:
+cosine_min_lr_ratio: # decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of peak lr
 
 # For one_cycle optim
 lr_div_factor: # Learning rate div factor
@@ -674,7 +813,7 @@ max_grad_norm:
 # Augmentation techniques
 # NEFT https://arxiv.org/abs/2310.05914, set this to a number (paper default is 5) to add noise to embeddings
 # currently only supported on Llama and Mistral
-noisy_embedding_alpha:
+neftune_noise_alpha:
 
 # Whether to bettertransformers
 flash_optimum:
@@ -689,15 +828,6 @@ flash_attn_fuse_mlp: # Whether to fuse part of the MLP into a single operation
 # Whether to use scaled-dot-product attention
 # https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html
 sdp_attention:
-# Landmark attention (only llama)
-landmark_attention:
-# xpos RoPE see https://github.com/kaiokendev/cutoff-len-is-context-len/blob/main/util/xpos_rope_llama_monkey_patch.py
-# LLaMA only
-xpos_rope:
-# RoPE Scaling https://github.com/huggingface/transformers/pull/24653
-rope_scaling:
-  type: # linear | dynamic
-  factor: # float
 
 # Resume from a specific checkpoint dir
 resume_from_checkpoint:
@@ -815,14 +945,42 @@ Run
 accelerate launch -m axolotl.cli.train your_config.yml
 ```
 
-#### Multi-GPU
+#### Preprocess dataset
+
+You can optionally pre-tokenize dataset with the following before finetuning.
+This is recommended for large datasets.
+
+- Set `dataset_prepared_path:` to a local folder for saving and loading pre-tokenized dataset.
+- (Optional): Set `push_dataset_to_hub: hf_user/repo` to push it to Huggingface.
+- (Optional): Use `--debug` to see preprocessed examples.
 
-You can optionally pre-tokenize dataset with the following before finetuning:
 ```bash
-CUDA_VISIBLE_DEVICES=0 accelerate launch -m axolotl.cli.train your_config.yml --prepare_ds_only
+python -m axolotl.cli.preprocess your_config.yml
 ```
 
-##### Config
+#### Multi-GPU
+
+Below are the options available in axolotl for training with multiple GPUs. Note that DeepSpeed
+is the recommended multi-GPU option currently because FSDP may experience
+[loss instability](https://github.com/huggingface/transformers/issues/26498).
+
+##### DeepSpeed
+
+Deepspeed is an optimization suite for multi-gpu systems allowing you to train much larger models than you
+might typically be able to fit into your GPU's VRAM. More information about the various optimization types
+for deepspeed is available at https://huggingface.co/docs/accelerate/main/en/usage_guides/deepspeed#what-is-integrated
+
+We provide several default deepspeed JSON configurations for ZeRO stage 1, 2, and 3.
+
+```yaml
+deepspeed: deepspeed/zero1.json
+```
+
+```shell
+accelerate launch -m axolotl.cli.train examples/llama-2/config.py --deepspeed deepspeed/zero1.json
+```
+
+##### FSDP
 
 - llama FSDP
 ```yaml
@@ -837,37 +995,40 @@ fsdp_config:
 
 ##### Weights & Biases Logging
 
+Make sure your `WANDB_API_KEY` environment variable is set (recommended) or you login to wandb with `wandb login`.
+
 - wandb options
 ```yaml
 wandb_mode:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 ```
 
-### Training with Deepspeed
+##### Special Tokens
 
-Deepspeed is an optimization suite for multi-gpu systems allowing you to train much larger models than you
-might typically be able to fit into your GPU's VRAM. More information about the various optimization types
-for deepspeed is available at https://huggingface.co/docs/accelerate/main/en/usage_guides/deepspeed#what-is-integrated
+It is important to have special tokens like delimiters, end-of-sequence, beginning-of-sequence in your tokenizer's vocabulary.  This will help you avoid tokenization issues and help your model train better.  You can do this in axolotl like this:
 
-We provide several default deepspeed JSON configurations for ZeRO stage 1, 2, and 3.
-
-```shell
-accelerate launch -m axolotl.cli.train examples/llama-2/config.py --deepspeed deepspeed/zero1.json
+```yml
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"
+tokens: # these are delimiters
+  - "<|im_start|>"
+  - "<|im_end|>"
 ```
 
-or
+When you include these tokens in your axolotl config, axolotl adds these tokens to the tokenizer's vocabulary.
 
-```yaml
-deepspeed: deepspeed/zero1.json
-```
+### Inference Playground
 
-### Inference
+Axolotl allows you to load your model in an interactive terminal playground for quick experimentation.
+The config file is the same config file used for training.
 
-Pass the appropriate flag to the train command:
+Pass the appropriate flag to the inference command, depending upon what kind of model was trained:
 
 - Pretrained LORA:
   ```bash
@@ -882,6 +1043,10 @@ Pass the appropriate flag to the train command:
   cat /tmp/prompt.txt | python -m axolotl.cli.inference examples/your_config.yml \
     --base_model="./completed-model" --prompter=None --load_in_8bit=True
   ```
+-- With gradio hosting
+  ```bash
+  python -m axolotl.cli.inference examples/your_config.yml --gradio
+  ```
 
 Please use `--sample_packing False` if you have it on and receive the error similar to below:
 
@@ -889,21 +1054,23 @@ Please use `--sample_packing False` if you have it on and receive the error simi
 
 ### Merge LORA to base
 
-Add below flag to train command above
+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 examples/your_config.yml --lora_model_dir="./completed-model" --load_in_8bit=False --load_in_4bit=False
+python3 -m axolotl.cli.merge_lora your_config.yml --lora_model_dir="./completed-model"
 ```
 
-If you run out of CUDA memory, you can try to merge in system RAM with
+You may need to use the `gpu_memory_limit` and/or `lora_on_cpu` config options to avoid running out of memory. If you still run out of CUDA memory, you can try to merge in system RAM with
 
 ```bash
 CUDA_VISIBLE_DEVICES="" python3 -m axolotl.cli.merge_lora ...
 ```
 
+although this will be very slow, and using the config options above are recommended instead.
+
 ## Common Errors 🧰
 
-See also the [FAQ's](./docs/faq.md).
+See also the [FAQ's](./docs/faq.md) and [debugging guide](docs/debugging.md).
 
 > 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:
 
@@ -913,6 +1080,10 @@ Please reduce any below
   - `gradient_accumulation_steps`
   - `sequence_len`
 
+If it does not help, try running without deepspeed and without accelerate (replace "accelerate launch" with "python") in the command.
+
+Using adamw_bnb_8bit might also save you some memory.
+
 > `failed (exitcode: -9)`
 
 Usually means your system has run out of system memory.
@@ -935,6 +1106,24 @@ It's safe to ignore it.
 
 See the [NCCL](docs/nccl.md) guide.
 
+
+### Tokenization Mismatch b/w Inference & Training
+
+For many formats, Axolotl constructs prompts by concatenating token ids _after_ tokenizing strings.  The reason for concatenating token ids rather than operating on strings is to maintain precise accounting for attention masks.
+
+If you decode a prompt constructed by axolotl, you might see spaces between tokens (or lack thereof) that you do not expect, especially around delimiters and special tokens.  When you are starting out with a new format, you should always do the following:
+
+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.
+4. As an additional troubleshooting step, you can look 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.
+
 ## Need help? 🙋♂️
 
 Join our [Discord server](https://discord.gg/HhrNrHJPRb) where we can help you
@@ -977,3 +1166,33 @@ pre-commit install
 # test
 pytest tests/
 ```
+
+## Sponsors 🤝❤
+
+OpenAccess AI Collective is run by volunteer contributors such as [winglian](https://github.com/winglian),
+[NanoCode012](https://github.com/NanoCode012), [tmm1](https://github.com/tmm1),
+[mhenrichsen](https://github.com/mhenrichsen), [casper-hansen](https://github.com/casper-hansen),
+[hamelsmu](https://github.com/hamelsmu) and many more who help us accelerate forward by fixing bugs, answering
+community questions and implementing new features. Axolotl needs donations from sponsors for the compute needed to
+run our unit & integration tests, troubleshooting community issues, and providing bounties. If you love axolotl,
+consider sponsoring the project via [GitHub Sponsors](https://github.com/sponsors/OpenAccess-AI-Collective),
+[Ko-fi](https://ko-fi.com/axolotl_ai) or reach out directly to
+[wing@openaccessaicollective.org](mailto:wing@openaccessaicollective.org).
+
+---
+
+#### 💎 Diamond Sponsors - [Contact directly](mailto:wing@openaccessaicollective.org)
+
+---
+
+#### 🥇 Gold Sponsors - $5000/mo
+
+---
+
+#### 🥈 Silver Sponsors - $1000/mo
+
+---
+
+#### 🥉 Bronze Sponsors - $500/mo
+
+---

deepspeed/zero1.json

@@ -24,16 +24,6 @@
       "weight_decay": "auto"
     }
   },
-  "scheduler": {
-    "type": "WarmupDecayLR",
-    "params": {
-      "warmup_min_lr": "auto",
-      "warmup_max_lr": "auto",
-      "warmup_num_steps": "auto",
-      "warmup_type": "linear",
-      "total_num_steps": "auto"
-    }
-  },
   "gradient_accumulation_steps": "auto",
   "train_batch_size": "auto",
   "train_micro_batch_size_per_gpu": "auto",

deepspeed/zero2.json

@@ -28,16 +28,6 @@
       "weight_decay": "auto"
     }
   },
-  "scheduler": {
-    "type": "WarmupDecayLR",
-    "params": {
-      "warmup_min_lr": "auto",
-      "warmup_max_lr": "auto",
-      "warmup_num_steps": "auto",
-      "warmup_type": "linear",
-      "total_num_steps": "auto"
-    }
-  },
   "gradient_accumulation_steps": "auto",
   "train_batch_size": "auto",
   "train_micro_batch_size_per_gpu": "auto",

deepspeed/zero3.json

@@ -1,14 +1,6 @@
 {
   "zero_optimization": {
     "stage": 3,
-    "offload_optimizer": {
-      "device": "cpu",
-      "pin_memory": true
-    },
-    "offload_param": {
-      "device": "cpu",
-      "pin_memory": true
-    },
     "overlap_comm": true,
     "contiguous_gradients": true,
     "sub_group_size": 0,
@@ -40,15 +32,6 @@
       "weight_decay": "auto"
     }
   },
-  "scheduler": {
-    "type": "WarmupLR",
-    "params": {
-      "warmup_min_lr": "auto",
-      "warmup_max_lr": "auto",
-      "warmup_num_steps": "auto",
-      "warmup_type": "linear"
-    }
-  },
   "gradient_accumulation_steps": "auto",
   "train_batch_size": "auto",
   "train_micro_batch_size_per_gpu": "auto",

deepspeed/zero3_bf16.json

@@ -0,0 +1,39 @@
+{
+  "zero_optimization": {
+    "stage": 3,
+    "overlap_comm": true,
+    "contiguous_gradients": true,
+    "sub_group_size": 0,
+    "reduce_bucket_size": "auto",
+    "stage3_prefetch_bucket_size": "auto",
+    "stage3_param_persistence_threshold": "auto",
+    "stage3_max_live_parameters": 0,
+    "stage3_max_reuse_distance": 0,
+    "stage3_gather_16bit_weights_on_model_save": true
+  },
+  "bf16": {
+    "enabled": true
+  },
+  "fp16": {
+    "enabled": "auto",
+    "auto_cast": false,
+    "loss_scale": 0,
+    "initial_scale_power": 32,
+    "loss_scale_window": 1000,
+    "hysteresis": 2,
+    "min_loss_scale": 1
+  },
+  "optimizer": {
+    "type": "AdamW",
+    "params": {
+      "lr": "auto",
+      "betas": "auto",
+      "eps": "auto",
+      "weight_decay": "auto"
+    }
+  },
+  "gradient_accumulation_steps": "auto",
+  "train_batch_size": "auto",
+  "train_micro_batch_size_per_gpu": "auto",
+  "wall_clock_breakdown": false
+}

devtools/README.md

@@ -0,0 +1 @@
+This directory contains example config files that might be useful for debugging. Please see [docs/debugging.md](../docs/debugging.md) for more information.

devtools/dev_sharegpt.yml

@@ -0,0 +1,49 @@
+# Example config for debugging the sharegpt prompt format
+base_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: true
+load_in_4bit: false
+
+datasets:
+  - path: philschmid/guanaco-sharegpt-style
+    type: sharegpt
+    shards: 10
+val_set_size: 0
+output_dir: temp_debug/axolotl_outputs/model
+dataset_prepared_path: temp_debug/axolotl_outputs/data
+dataset_processes: 1
+
+sequence_len: 4096
+sample_packing: false
+pad_to_sequence_len: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+micro_batch_size: 1
+num_epochs: 1
+max_steps: 10
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: false
+fp16: true
+tf32: false
+
+gradient_checkpointing: true
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+weight_decay: 0.0

docker/Dockerfile

@@ -10,7 +10,7 @@ ARG PYTORCH_VERSION="2.0.1"
 ENV PYTORCH_VERSION=$PYTORCH_VERSION
 
 RUN apt-get update && \
-    apt-get install -y vim curl
+    apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev
 
 WORKDIR /workspace
 
@@ -19,13 +19,15 @@ 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 sed -i "s/torch==.*/torch==$PYTORCH_VERSION/" requirements.txt
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        pip install -e .[flash-attn,$AXOLOTL_EXTRAS]; \
+        pip install -e .[deepspeed,flash-attn,mamba-ssm,$AXOLOTL_EXTRAS]; \
     else \
-        pip install -e .[flash-attn]; \
+        pip install -e .[deepspeed,flash-attn,mamba-ssm]; \
     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

docker/Dockerfile-base

@@ -10,8 +10,10 @@ ENV PATH="/root/miniconda3/bin:${PATH}"
 ARG PYTHON_VERSION="3.9"
 ARG PYTORCH_VERSION="2.0.1"
 ARG CUDA="118"
+ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 9.0+PTX"
 
 ENV PYTHON_VERSION=$PYTHON_VERSION
+ENV TORCH_CUDA_ARCH_LIST=$TORCH_CUDA_ARCH_LIST
 
 RUN apt-get update \
     && apt-get install -y wget git build-essential ninja-build git-lfs libaio-dev && rm -rf /var/lib/apt/lists/* \
@@ -27,47 +29,9 @@ ENV PATH="/root/miniconda3/envs/py${PYTHON_VERSION}/bin:${PATH}"
 WORKDIR /workspace
 
 RUN python3 -m pip install --upgrade pip && pip3 install packaging && \
-    python3 -m pip install --no-cache-dir -U torch==${PYTORCH_VERSION}+cu${CUDA} --extra-index-url https://download.pytorch.org/whl/cu$CUDA
+    python3 -m pip install --no-cache-dir -U torch==${PYTORCH_VERSION}+cu${CUDA} deepspeed-kernels --extra-index-url https://download.pytorch.org/whl/cu$CUDA
 
-FROM base-builder AS deepspeed-builder
-
-ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 9.0+PTX"
-
-WORKDIR /workspace
-
-RUN git clone https://github.com/microsoft/DeepSpeed.git && \
-    cd DeepSpeed && \
-    MAX_CONCURRENCY=8 DS_BUILD_SPARSE_ATTN=0 DS_BUILD_OPS=1 DS_BUILD_EVOFORMER_ATTN=0 python3 setup.py bdist_wheel
-
-FROM base-builder AS bnb-builder
-
-WORKDIR /workspace
-ARG CUDA="118"
-ENV CUDA=$CUDA
-ARG MAX_JOBS="-1"
-ENV MAX_JOBS=$MAX_JOBS
-
-RUN git clone https://github.com/TimDettmers/bitsandbytes.git && \
-    cd bitsandbytes && \
-    CUDA_VERSION=$CUDA make cuda11x && \
-    python setup.py bdist_wheel
-
-FROM base-builder
-
-ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 9.0+PTX"
-ENV TORCH_CUDA_ARCH_LIST=$TORCH_CUDA_ARCH_LIST
-
-RUN mkdir -p /workspace/builds
-COPY --from=bnb-builder /workspace/bitsandbytes /workspace/builds/bitsandbytes
-
-RUN mkdir -p /workspace/wheels/bitsandbytes
-COPY --from=deepspeed-builder /workspace/DeepSpeed/dist/deepspeed-*.whl wheels
-COPY --from=bnb-builder /workspace/bitsandbytes/dist/bitsandbytes-*.whl wheels
-COPY --from=bnb-builder /workspace/bitsandbytes/bitsandbytes/libbitsandbytes*.so wheels/bitsandbytes
-
-RUN pip3 install wheels/deepspeed-*.whl
-RUN cd /workspace/builds/bitsandbytes && python3 setup.py install
-RUN git lfs install --skip-repo
-RUN pip3 install awscli && \
+RUN git lfs install --skip-repo && \
+    pip3 install awscli && \
     # The base image ships with `pydantic==1.8.2` which is not working
     pip3 install -U --no-cache-dir pydantic==1.10.10

docker/Dockerfile-runpod

@@ -4,6 +4,8 @@ FROM winglian/axolotl:$BASE_TAG
 ENV HF_DATASETS_CACHE="/workspace/data/huggingface-cache/datasets"
 ENV HUGGINGFACE_HUB_CACHE="/workspace/data/huggingface-cache/hub"
 ENV TRANSFORMERS_CACHE="/workspace/data/huggingface-cache/hub"
+ENV HF_HOME="/workspace/data/huggingface-cache/hub"
+ENV HF_HUB_ENABLE_HF_TRANSFER="1"
 
 COPY scripts/runpod-entrypoint.sh /root/runpod-entrypoint.sh
 

docker/Dockerfile-tests

@@ -0,0 +1,40 @@
+ARG BASE_TAG=main-base
+FROM winglian/axolotl-base:$BASE_TAG
+
+ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX"
+ARG AXOLOTL_EXTRAS=""
+ARG CUDA="118"
+ENV BNB_CUDA_VERSION=$CUDA
+ARG PYTORCH_VERSION="2.0.1"
+ARG GITHUB_REF="main"
+
+ENV PYTORCH_VERSION=$PYTORCH_VERSION
+
+RUN apt-get update && \
+    apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev
+
+WORKDIR /workspace
+
+RUN git clone --depth=1 https://github.com/OpenAccess-AI-Collective/axolotl.git
+
+WORKDIR /workspace/axolotl
+
+RUN git fetch origin +$GITHUB_REF && \
+    git checkout FETCH_HEAD
+
+# If AXOLOTL_EXTRAS is set, append it in brackets
+RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
+        pip install -e .[deepspeed,flash-attn,mamba-ssm,$AXOLOTL_EXTRAS]; \
+    else \
+        pip install -e .[deepspeed,flash-attn,mamba-ssm]; \
+    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

docs/debugging.md

@@ -0,0 +1,242 @@
+# Debugging 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.
+
+## Table of Contents
+
+- [General Tips](#general-tips)
+- [Debugging with VSCode](#debugging-with-vscode)
+    - [Background](#background)
+    - [Configuration](#configuration)
+    - [Customizing your debugger](#customizing-your-debugger)
+    - [Video Tutorial](#video-tutorial)
+- [Debugging With Docker](#debugging-with-docker)
+    - [Setup](#setup)
+    - [Attach To Container](#attach-to-container)
+    - [Video - Attaching To Docker On Remote Host](#video---attaching-to-docker-on-remote-host)
+
+## General Tips
+
+While debugging it's helpful to simplify your test scenario as much as possible.  Here are some tips for doing so:
+
+> [!Important]
+> All of these tips are incorporated into the [example configuration](#configuration) for debugging with VSCode below.
+
+1. **Make sure you are using the latest version of axolotl**:  This project changes often and bugs get fixed fast.  Check your git branch and make sure you have pulled the latest changes from `main`.
+1. **Eliminate concurrency**: Restrict the number of processes to 1 for both training and data preprocessing:
+    - Set `CUDA_VISIBLE_DEVICES` to a single GPU, ex: `export CUDA_VISIBLE_DEVICES=0`.
+    - Set `dataset_processes: 1` in your axolotl config or run the training command with `--dataset_processes=1`.
+2. **Use a small dataset**: Construct or use a small dataset from HF Hub. When using a small dataset, you will often have to make sure `sample_packing: False` and `eval_sample_packing: False` to avoid errors.  If you are in a pinch and don't have time to construct a small dataset but want to use from the HF Hub, you can shard the data (this will still tokenize the entire dataset, but will only use a fraction of the data for training.  For example, to shard the dataset into 20 pieces, add the following to your axolotl config):
+    ```yaml
+    dataset:
+        ...
+        shards: 20
+    ```
+3. **Use a small model**: A good example of a small model is [TinyLlama/TinyLlama-1.1B-Chat-v1.0](https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0).
+4. **Minimize iteration time**: Make sure the training loop finishes as fast as possible, with these settings.
+    - `micro_batch_size: 1`
+    - `max_steps: 1`
+    - `val_set_size: 0`
+5. **Clear Caches:** Axolotl caches certain steps and so does the underlying HuggingFace trainer.  You may want to clear some of these caches when debugging.
+    - Data preprocessing: When debugging data preprocessing, which includes prompt template formation, you may want to delete the directory set in `dataset_prepared_path:` in your axolotl config.  If you didn't set this value, the default is `last_run_prepared`.
+    - HF Hub: If you are debugging data preprocessing, you should clear the relevant HF cache [HuggingFace cache](https://huggingface.co/docs/datasets/cache), by deleting the appropriate `~/.cache/huggingface/datasets/...` folder(s).
+    - **The recommended approach is to redirect all outputs and caches to a temporary folder and delete selected subfolders before each run.  This is demonstrated in the example configuration below.**
+
+
+## Debugging with VSCode
+
+### Background
+
+The below example shows how to configure VSCode to debug data preprocessing of the `sharegpt` format.  This is the format used when you have the following in your axolotl config:
+
+```yaml
+datasets:
+  - path: <path to your sharegpt formatted dataset> # example on HF Hub: philschmid/guanaco-sharegpt-style
+    type: sharegpt
+```
+
+>[!Important]
+> If you are already familiar with advanced VSCode debugging, you can skip the below explanation and look at the files [.vscode/launch.json](../.vscode/launch.json) and [.vscode/tasks.json](../.vscode/tasks.json) for an example configuration.
+
+>[!Tip]
+> If you prefer to watch a video, rather than read, you can skip to the [video tutorial](#video-tutorial) below (but doing both is recommended).
+
+### Setup
+
+Make sure you have an [editable install](https://setuptools.pypa.io/en/latest/userguide/development_mode.html) of Axolotl, which ensures that changes you make to the code are reflected at runtime.  Run the following commands from the root of this project:
+
+```bash
+pip3 install packaging
+pip3 install -e '.[flash-attn,deepspeed]'
+```
+
+#### Remote Hosts
+
+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
+
+The easiest way to get started is to modify the [.vscode/launch.json](../.vscode/launch.json) file in this project.  This is just an example configuration, so you may need to modify or copy it to suit your needs.
+
+For example, to mimic the command `cd devtools && CUDA_VISIBLE_DEVICES=0 accelerate launch -m axolotl.cli.train dev_sharegpt.yml`, you would use the below configuration[^1].  Note that we add additional flags that override the axolotl config and incorporate the tips above (see the comments). We also set the working directory to `devtools` and set the `env` variable `HF_HOME` to a temporary folder that is later partially deleted.  This is because we want to delete the HF dataset cache before each run in order to ensure that the data preprocessing code is run from scratch.
+
+```jsonc
+// .vscode/launch.json
+{
+    "version": "0.2.0",
+    "configurations": [
+        {
+            "name": "Debug axolotl prompt - sharegpt",
+            "type": "python",
+            "module": "accelerate.commands.launch",
+            "request": "launch",
+            "args": [
+                "-m", "axolotl.cli.train", "dev_sharegpt.yml",
+                // The flags below simplify debugging by overriding the axolotl config
+                // with the debugging tips above.  Modify as needed.
+                "--dataset_processes=1",      // limits data preprocessing to one process
+                "--max_steps=1",              // limits training to just one step
+                "--batch_size=1",             // minimizes batch size
+                "--micro_batch_size=1",       // minimizes batch size
+                "--val_set_size=0",           // disables validation
+                "--sample_packing=False",     // disables sample packing which is necessary for small datasets
+                "--eval_sample_packing=False",// disables sample packing on eval set
+                "--dataset_prepared_path=temp_debug/axolotl_outputs/data", // send data outputs to a temp folder
+                "--output_dir=temp_debug/axolotl_outputs/model" // send model outputs to a temp folder
+                ],
+            "console": "integratedTerminal",      // show output in the integrated terminal
+            "cwd": "${workspaceFolder}/devtools", // set working directory to devtools from the root of the project
+            "justMyCode": true,                   // step through only axolotl code
+            "env": {"CUDA_VISIBLE_DEVICES": "0",  // Since we aren't doing distributed training, we need to limit to one GPU
+                    "HF_HOME": "${workspaceFolder}/devtools/temp_debug/.hf-cache"}, // send HF cache to a temp folder
+            "preLaunchTask": "cleanup-for-dataprep", // delete temp folders (see below)
+        }
+    ]
+}
+```
+
+**Additional notes about this configuration:**
+
+- The argument `justMyCode` is set to `true` such that you step through only the axolotl code.  If you want to step into dependencies, set this to `false`.
+- The `preLaunchTask`: `cleanup-for-dataprep` is defined in [.vscode/tasks.json](../.vscode/tasks.json) and is used to delete the following folders before debugging, which is essential to ensure that the data pre-processing code is run from scratch:
+    -  `./devtools/temp_debug/axolotl_outputs`
+    - `./devtools/temp_debug/.hf-cache/datasets`
+
+>[!Tip]
+> You may not want to delete these folders. For example, if you are debugging model training instead of data pre-processing, you may NOT want to delete the cache or output folders. You may also need to add additional tasks to the `tasks.json` file depending on your use case.
+
+Below is the [./vscode/tasks.json](../.vscode/tasks.json) file that defines the `cleanup-for-dataprep` task.  This task is run before each debugging session when you use the above configuration.  Note how there are two tasks that delete the two folders mentioned above.  The third task `cleanup-for-dataprep` is a composite task that combines the two tasks.  A composite task is necessary because VSCode does not allow you to specify multiple tasks in the `preLaunchTask` argument of the `launch.json` file.
+
+```jsonc
+// .vscode/tasks.json
+// this file is used by launch.json
+{
+    "version": "2.0.0",
+    "tasks": [
+      // this task changes into the devtools directory and deletes the temp_debug/axolotl_outputs folder
+      {
+        "label": "delete-outputs",
+        "type": "shell",
+        "command": "rm -rf temp_debug/axolotl_outputs",
+        "options":{ "cwd": "${workspaceFolder}/devtools"},
+        "problemMatcher": []
+      },
+      // this task changes into the devtools directory and deletes the `temp_debug/.hf-cache/datasets` folder
+      {
+        "label": "delete-temp-hf-dataset-cache",
+        "type": "shell",
+        "command": "rm -rf temp_debug/.hf-cache/datasets",
+        "options":{ "cwd": "${workspaceFolder}/devtools"},
+        "problemMatcher": []
+      },
+        // this task combines the two tasks above
+      {
+       "label": "cleanup-for-dataprep",
+       "dependsOn": ["delete-outputs", "delete-temp-hf-dataset-cache"],
+      }
+    ]
+}
+```
+
+### Customizing your debugger
+
+Your debugging use case may differ from the example above.  The easiest thing to do is to put your own axolotl config in the `devtools` folder and modify the `launch.json` file to use your config.  You may also want to modify the `preLaunchTask` to delete different folders or not delete anything at all.
+
+### Video Tutorial
+
+The following video tutorial walks through the above configuration and demonstrates how to debug with VSCode, (click the image below to watch):
+
+<div style="text-align: center; line-height: 0;">
+
+<a href="https://youtu.be/xUUB11yeMmc" target="_blank"
+title="How to debug Axolotl (for fine tuning LLMs)"><img
+src="https://i.ytimg.com/vi/xUUB11yeMmc/maxresdefault.jpg"
+style="border-radius: 10px; display: block; margin: auto;" width="560" height="315" /></a>
+
+<figcaption style="font-size: smaller;"><a href="https://hamel.dev">Hamel Husain's</a> tutorial: <a href="https://www.youtube.com/watch?v=xUUB11yeMmc">Debugging Axolotl w/VSCode</a></figcaption>
+
+</div>
+<br>
+
+## Debugging With Docker
+
+Using [official Axolotl Docker images](https://hub.docker.com/r/winglian/axolotl/tags) is a great way to debug your code, and is a very popular way to use Axolotl.  Attaching VSCode to Docker takes a few more steps.
+
+### Setup
+
+On the host that is running axolotl (ex: if you are using a remote host), clone the axolotl repo and change your current directory to the root:
+
+```bash
+git clone https://github.com/OpenAccess-AI-Collective/axolotl
+cd axolotl
+```
+
+>[!Tip]
+> If you already have axolotl cloned on your host, make sure you have the latest changes and change into the root of the project.
+
+Next, run the desired docker image and mount the current directory. Below is a docker command you can run to do this:[^2]
+
+```bash
+docker run --privileged --gpus '"all"' --shm-size 10g --rm -it --name axolotl --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 --mount type=bind,src="${PWD}",target=/workspace/axolotl -v ${HOME}/.cache/huggingface:/root/.cache/huggingface winglian/axolotl:main-py3.10-cu118-2.0.1
+```
+
+>[!Tip]
+> To understand which containers are available, see the [Docker section of the README](../README.md#docker) and the [DockerHub repo](https://hub.docker.com/r/winglian/axolotl/tags).  For details of how the Docker containers are built, see axolotl's [Docker CI builds](../.github/workflows/main.yml).
+
+You will now be in the container.  Next, perform an editable install of Axolotl:
+
+```bash
+pip3 install packaging
+pip3 install -e '.[flash-attn,deepspeed]'
+```
+
+### Attach To Container
+
+Next, if you are using a remote host, [Remote into this host with VSCode](https://code.visualstudio.com/docs/remote/ssh).  If you are using a local host, you can skip this step.
+
+Next, select `Dev Containers: Attach to Running Container...` using the command palette (`CMD + SHIFT + P`) in VSCode.  You will be prompted to select a container to attach to.  Select the container you just created.  You will now be in the container with a working directory that is at the root of the project.  Any changes you make to the code will be reflected both in the container and on the host.
+
+Now you are ready to debug as described above (see [Debugging with VSCode](#debugging-with-vscode)).
+
+### Video - Attaching To Docker On Remote Host
+
+Here is a short video that demonstrates how to attach to a Docker container on a remote host:
+
+<div style="text-align: center; line-height: 0;">
+
+<a href="https://youtu.be/0AuoR7QnHR0" target="_blank"
+title="Debugging Axolotl Part 2: Attaching to Docker on a Remote Host"><img
+src="https://i.ytimg.com/vi/0AuoR7QnHR0/hqdefault.jpg"
+style="border-radius: 10px; display: block; margin: auto;" width="560" height="315" /></a>
+
+<figcaption style="font-size: smaller;"><a href="https://hamel.dev">Hamel Husain's</a> tutorial: <a href="https://youtu.be/0AuoR7QnHR0">Debugging Axolotl Part 2: Attaching to Docker on a Remote Host
+</a></figcaption>
+
+</div>
+<br>
+
+[^1]: The config actually mimics the command `CUDA_VISIBLE_DEVICES=0 python -m accelerate.commands.launch -m axolotl.cli.train devtools/sharegpt.yml`, but this is the same thing.
+
+[^2]: Many of the below flags are recommended best practices by Nvidia when using nvidia-container-toolkit.  You can read more about these flags [here](https://docs.nvidia.com/deeplearning/frameworks/user-guide/index.html).

docs/faq.md

@@ -12,3 +12,7 @@ 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/rlhf.md

@@ -0,0 +1,44 @@
+# RLHF (Beta)
+
+### Overview
+
+Reinforcement Learning from Human Feedback is a method whereby a language model is optimized from data using human
+feedback. Various methods include, but not limited to:
+
+- Proximal Policy Optimization (PPO) (not yet supported in axolotl)
+- Direct Preference Optimization (DPO)
+- Identity Preference Optimization (IPO)
+
+
+### RLHF using Axolotl
+
+[!IMPORTANT]
+This is a BETA feature and many features are not fully implemented. You are encouraged to open new PRs to improve the integration and functionality.
+
+The various RL training methods are implemented in trl and wrapped via axolotl. Below are various examples with how you can use various preference datasets to train models that use ChatML
+
+#### DPO
+```yaml
+rl: true
+datasets:
+  - path: Intel/orca_dpo_pairs
+    split: train
+    type: intel_apply_chatml
+  - path: argilla/ultrafeedback-binarized-preferences
+    split: train
+    type: argilla_apply_chatml
+```
+
+#### IPO
+```yaml
+rl: ipo
+```
+
+#### Trl autounwrap for peft
+
+Trl supports autounwrapping peft models, so that a ref model does not need to be additionally loaded, leading to less VRAM needed. This is on by default. To turn it off, pass the following config.
+
+```yaml
+# load ref model when adapter training.
+rl_adapter_ref_model: true
+```

examples/cerebras/btlm-ft.yml

@@ -14,7 +14,7 @@ datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
 dataset_prepared_path: last_prepared_run
-val_set_size: 0.01
+val_set_size: 0.05
 
 adapter:
 lora_model_dir:
@@ -35,7 +35,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 output_dir: btlm-out
@@ -72,8 +72,8 @@ gptq_groupsize:
 gptq_model_v1:
 
 warmup_steps: 32
-eval_steps:
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 save_total_limit:
 
 debug:

examples/cerebras/qlora.yml

@@ -7,7 +7,7 @@ datasets:
   - path: teknium/GPT4-LLM-Cleaned
     type: alpaca
 dataset_prepared_path:
-val_set_size: 0.01
+val_set_size: 0.05
 adapter: qlora
 lora_model_dir:
 sequence_len: 2048
@@ -24,7 +24,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./qlora-out
 batch_size: 4
@@ -49,8 +49,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 10
-eval_steps: 20
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

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

@@ -11,7 +11,7 @@ datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
 dataset_prepared_path:
-val_set_size: 0.01
+val_set_size: 0.05
 output_dir: ./lora-out
 
 sequence_len: 4096
@@ -29,12 +29,12 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 3
+num_epochs: 4
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -54,8 +54,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

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

@@ -11,7 +11,7 @@ datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
 dataset_prepared_path:
-val_set_size: 0.01
+val_set_size: 0.05
 output_dir: ./qlora-out
 
 adapter: qlora
@@ -31,12 +31,12 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 3
+num_epochs: 4
 optimizer: paged_adamw_32bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -56,8 +56,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

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

@@ -11,7 +11,7 @@ datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
 dataset_prepared_path:
-val_set_size: 0.01
+val_set_size: 0.05
 output_dir: ./lora-out
 
 sequence_len: 4096
@@ -29,12 +29,12 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 3
+num_epochs: 4
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -54,8 +54,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

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

@@ -11,7 +11,7 @@ datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
 dataset_prepared_path:
-val_set_size: 0.01
+val_set_size: 0.05
 output_dir: ./qlora-out
 
 adapter: qlora
@@ -31,12 +31,12 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 3
+num_epochs: 4
 optimizer: paged_adamw_32bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -56,8 +56,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

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

@@ -11,7 +11,7 @@ datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
 dataset_prepared_path:
-val_set_size: 0.01
+val_set_size: 0.05
 output_dir: ./lora-out
 
 sequence_len: 4096
@@ -29,12 +29,12 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 3
+num_epochs: 4
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -54,8 +54,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

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

@@ -11,7 +11,7 @@ datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
 dataset_prepared_path:
-val_set_size: 0.01
+val_set_size: 0.05
 output_dir: ./qlora-out
 
 adapter: qlora
@@ -31,12 +31,12 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 3
+num_epochs: 4
 optimizer: paged_adamw_32bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -56,8 +56,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

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

@@ -12,7 +12,7 @@ datasets:
   - path: teknium/GPT4-LLM-Cleaned
     type: alpaca:chat
 dataset_prepared_path:
-val_set_size: 0.01
+val_set_size: 0.05
 adapter: lora
 lora_model_dir:
 sequence_len: 2048
@@ -26,7 +26,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./falcon-7b
 batch_size: 2
@@ -51,8 +51,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 40
-eval_steps: 5
-save_steps: 43
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

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

@@ -18,7 +18,7 @@ datasets:
       - Chain-of-Thought/formatted_cot_data/gsm8k_train.json
     type: "alpaca:chat"
 dataset_prepared_path:
-val_set_size: 0.01
+val_set_size: 0.05
 # enable QLoRA
 adapter: qlora
 lora_model_dir:
@@ -40,7 +40,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./qlora-out
 
@@ -53,7 +53,7 @@ output_dir: ./qlora-out
 # decrease if OOM, increase for max VRAM utilization
 micro_batch_size: 1
 gradient_accumulation_steps: 2
-num_epochs: 3
+num_epochs: 4
 # Optimizer for QLoRA
 optimizer: paged_adamw_32bit
 torchdistx_path:
@@ -80,8 +80,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 10
-eval_steps: 5
-save_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.000001

examples/falcon/config-7b.yml

@@ -12,7 +12,7 @@ datasets:
   - path: teknium/GPT4-LLM-Cleaned
     type: alpaca:chat
 dataset_prepared_path:
-val_set_size: 0.01
+val_set_size: 0.05
 adapter:
 lora_model_dir:
 sequence_len: 2048
@@ -26,7 +26,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./falcon-7b
 batch_size: 2
@@ -51,8 +51,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 40
-eval_steps: 5
-save_steps: 43
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/gptj/qlora.yml

@@ -7,7 +7,7 @@ datasets:
   - path: teknium/GPT4-LLM-Cleaned
     type: alpaca
 dataset_prepared_path:
-val_set_size: 0.01
+val_set_size: 0.05
 adapter: qlora
 lora_model_dir:
 sequence_len: 2048
@@ -21,7 +21,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./qlora-out
 gradient_accumulation_steps: 2
@@ -46,8 +46,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 10
-eval_steps: 20
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/jeopardy-bot/config.yml

@@ -19,12 +19,12 @@ lora_fan_in_fan_out: false
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./jeopardy-bot-7b
 gradient_accumulation_steps: 1
 micro_batch_size: 1
-num_epochs: 3
+num_epochs: 4
 optimizer: adamw_bnb_8bit
 torchdistx_path:
 lr_scheduler: cosine
@@ -42,8 +42,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
-eval_steps: 110
-save_steps: 660
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/llama-2/fft_optimized.yml

@@ -11,7 +11,7 @@ datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
 dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
+val_set_size: 0.05
 output_dir: ./out
 
 sequence_len: 4096
@@ -29,7 +29,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 1
@@ -58,9 +58,9 @@ flash_attn_fuse_qkv: false
 flash_attn_fuse_mlp: true
 
 warmup_steps: 100
-eval_steps: 0.05
+evals_per_epoch: 4
 eval_table_size:
-save_steps:
+saves_per_epoch: 1
 debug:
 deepspeed: #deepspeed/zero2.json # multi-gpu only
 weight_decay: 0.1

examples/llama-2/gptq-lora.yml

@@ -15,7 +15,7 @@ datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
 dataset_prepared_path:
-val_set_size: 0.01
+val_set_size: 0.05
 adapter: lora
 lora_model_dir:
 sequence_len: 4096
@@ -32,12 +32,12 @@ lora_target_linear:
 lora_fan_in_fan_out:
 wandb_project:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./model-out
 gradient_accumulation_steps: 1
 micro_batch_size: 1
-num_epochs: 3
+num_epochs: 4
 optimizer: adamw_torch
 adam_beta2: 0.95
 adam_eps: 0.00001
@@ -62,8 +62,8 @@ flash_attention:
 sdp_attention:
 flash_optimum:
 warmup_steps: 100
-eval_steps:
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/llama-2/lora.yml

@@ -11,7 +11,7 @@ datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
 dataset_prepared_path:
-val_set_size: 0.01
+val_set_size: 0.05
 output_dir: ./lora-out
 
 sequence_len: 4096
@@ -29,12 +29,12 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 3
+num_epochs: 4
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -54,10 +54,10 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
+evals_per_epoch: 4
 eval_table_size:
 eval_table_max_new_tokens: 128
-save_steps:
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/llama-2/qlora.yml

@@ -11,7 +11,7 @@ datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
 dataset_prepared_path:
-val_set_size: 0.01
+val_set_size: 0.05
 output_dir: ./qlora-out
 
 adapter: qlora
@@ -31,12 +31,12 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 3
+num_epochs: 4
 optimizer: paged_adamw_32bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -56,9 +56,9 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
+evals_per_epoch: 4
 eval_table_size:
-save_steps:
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/llama-2/relora.yml

@@ -11,7 +11,7 @@ datasets:
   - path: teknium/GPT4-LLM-Cleaned
     type: alpaca
 dataset_prepared_path:
-val_set_size: 0.01
+val_set_size: 0.05
 output_dir: ./relora-out
 
 adapter: qlora
@@ -35,12 +35,12 @@ relora_cpu_offload: false
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 4
-num_epochs: 3
+num_epochs: 4
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -60,8 +60,8 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
-save_steps: 50
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/mamba/config.yml

@@ -0,0 +1,61 @@
+base_model: state-spaces/mamba-2.8b
+model_type: MambaLMHeadModel
+tokenizer_type: AutoTokenizer
+tokenizer_config: EleutherAI/gpt-neox-20b
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.0
+output_dir: ./out
+
+sequence_len: 2048
+sample_packing: false
+pad_to_sequence_len: false
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 2
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+learning_rate: 5e-5
+
+train_on_inputs: false
+group_by_length: true
+
+bf16: true
+fp16: false
+tf32: true
+
+gradient_checkpointing: false
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention:
+
+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:
+tokens:
+save_safetensors: False

examples/mistral/config.yml

@@ -11,22 +11,23 @@ datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
 dataset_prepared_path:
-val_set_size: 0.01
+val_set_size: 0.05
 output_dir: ./out
 
 sequence_len: 8192
 sample_packing: true
 pad_to_sequence_len: true
+eval_sample_packing: false
 
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 3
+num_epochs: 4
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.000005
@@ -46,10 +47,10 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
+evals_per_epoch: 4
 eval_table_size:
 eval_table_max_new_tokens: 128
-save_steps:
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/mistral/mixtral.yml

@@ -0,0 +1,91 @@
+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: 0.0
+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.*
+
+model_config:
+  output_router_logits: true
+
+adapter: qlora
+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
+#  - q_proj
+#  - k_proj
+#  - v_proj
+#  - o_proj
+#  - w1
+#  - w2
+#  - w3
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+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: true
+fp16: false
+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_table_max_new_tokens: 128
+saves_per_epoch: 1
+debug:
+deepspeed: deepspeed/zero2.json
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:

examples/mistral/qlora.yml

@@ -11,7 +11,7 @@ datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
 dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
+val_set_size: 0.1
 output_dir: ./qlora-out
 
 adapter: qlora
@@ -38,7 +38,7 @@ lora_target_modules:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
@@ -62,11 +62,14 @@ logging_steps: 1
 xformers_attention:
 flash_attention: true
 
+loss_watchdog_threshold: 5.0
+loss_watchdog_patience: 3
+
 warmup_steps: 10
-eval_steps: 20
+evals_per_epoch: 4
 eval_table_size:
 eval_table_max_new_tokens: 128
-save_steps:
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/mpt-7b/config.yml

@@ -21,12 +21,12 @@ lora_fan_in_fan_out: false
 wandb_project: mpt-alpaca-7b
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./mpt-alpaca-7b
 gradient_accumulation_steps: 1
 micro_batch_size: 1
-num_epochs: 3
+num_epochs: 4
 optimizer: adamw_bnb_8bit
 torchdistx_path:
 lr_scheduler: cosine
@@ -44,8 +44,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
-eval_steps: 110
-save_steps: 660
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0001

examples/openllama-3b/config.yml

@@ -23,7 +23,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./openllama-out
 gradient_accumulation_steps: 1
@@ -49,8 +49,8 @@ flash_attention: true
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/openllama-3b/lora.yml

@@ -29,7 +29,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./lora-out
 gradient_accumulation_steps: 1
@@ -54,8 +54,8 @@ flash_attention: true
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/openllama-3b/qlora.yml

@@ -9,7 +9,7 @@ datasets:
   - path: teknium/GPT4-LLM-Cleaned
     type: alpaca
 dataset_prepared_path:
-val_set_size: 0.01
+val_set_size: 0.05
 adapter: qlora
 lora_model_dir:
 sequence_len: 1024
@@ -23,7 +23,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./qlora-out
 gradient_accumulation_steps: 1
@@ -48,8 +48,8 @@ flash_attention: true
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/phi/phi-ft.yml

@@ -1,5 +1,5 @@
 base_model: microsoft/phi-1_5
-model_type: MixFormerSequentialForCausalLM
+model_type: PhiForCausalLM
 tokenizer_type: AutoTokenizer
 is_llama_derived_model: false
 trust_remote_code: true
@@ -31,7 +31,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 1
@@ -59,8 +59,8 @@ xformers_attention:
 flash_attention:
 
 warmup_steps: 100
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/phi/phi-qlora.yml

@@ -31,7 +31,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 1
@@ -59,8 +59,8 @@ xformers_attention:
 flash_attention:
 
 warmup_steps: 100
-eval_steps: 0.05
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.1

examples/phi/phi2-ft.yml

@@ -0,0 +1,74 @@
+base_model: microsoft/phi-2
+model_revision:  834565c  # pin model repo to the previous architecture
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+trust_remote_code: true
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: garage-bAInd/Open-Platypus
+    type: alpaca
+
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./phi-sft-out
+
+sequence_len: 2048
+sample_packing: false  # currently unsupported
+pad_to_sequence_len:
+
+adapter:
+lora_model_dir:
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0.1
+lora_target_linear: true
+lora_fan_in_fan_out:
+lora_modules_to_save:
+  - embd
+  - lm_head
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 4
+optimizer: paged_adamw_8bit
+adam_beta2: 0.95
+adam_epsilon: 0.00001
+max_grad_norm: 1.0
+lr_scheduler: cosine
+learning_rate: 1e-5
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: true
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 100
+evals_per_epoch: 4
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.1
+fsdp:
+fsdp_config:
+resize_token_embeddings_to_32x: true
+special_tokens:
+  pad_token: "<|endoftext|>"

examples/pythia-12b/config.yml

@@ -24,7 +24,7 @@ lora_fan_in_fan_out: true  # pythia/GPTNeoX lora specific
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./pythia-12b
 gradient_accumulation_steps: 1

examples/pythia/lora.yml

@@ -18,12 +18,12 @@ lora_fan_in_fan_out: true  # pythia/GPTNeoX lora specific
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./lora-alpaca-pythia
 gradient_accumulation_steps: 1
 micro_batch_size: 4
-num_epochs: 3
+num_epochs: 4
 learning_rate: 0.00001
 train_on_inputs: false
 group_by_length: false
@@ -33,5 +33,5 @@ early_stopping_patience:
 resume_from_checkpoint:
 local_rank:
 weight_decay: 0.1
-eval_steps: 20
+evals_per_epoch: 4
 logging_steps: 1

examples/qwen/lora.yml

@@ -0,0 +1,68 @@
+base_model: Qwen/Qwen-7B
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+
+is_qwen_derived_model: true
+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: 2048  # supports up to 8192
+sample_packing: false
+pad_to_sequence_len:
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: false
+
+gradient_checkpointing: false
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention:
+
+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/qwen/qlora.yml

@@ -0,0 +1,68 @@
+base_model: Qwen/Qwen-7B
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+
+is_qwen_derived_model: true
+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: ./lora-out
+
+sequence_len: 2048  # supports up to 8192
+sample_packing: false
+pad_to_sequence_len:
+
+adapter: qlora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: false
+
+gradient_checkpointing: false
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention:
+
+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/redpajama/config-3b.yml

@@ -22,12 +22,12 @@ lora_fan_in_fan_out: false
 wandb_project: redpajama-alpaca-3b
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./redpajama-alpaca-3b
 batch_size: 4
 micro_batch_size: 1
-num_epochs: 3
+num_epochs: 4
 optimizer: adamw_bnb_8bit
 torchdistx_path:
 lr_scheduler: cosine
@@ -45,8 +45,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
-eval_steps: 110
-save_steps: 660
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0001

examples/replit-3b/config-lora.yml

@@ -21,12 +21,12 @@ lora_fan_in_fan_out:
 wandb_project: lora-replit
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./lora-replit
 batch_size: 8
 micro_batch_size: 1
-num_epochs: 3
+num_epochs: 4
 optimizer:
 torchdistx_path:
 lr_scheduler:
@@ -45,8 +45,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
-eval_steps: 50
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0

examples/tiny-llama/README.md

@@ -0,0 +1,17 @@
+# Overview
+
+This is a simple example of how to finetune TinyLlama1.1B using either lora or qlora:
+
+LoRa:
+
+```
+accelerate launch -m axolotl.cli.train examples/tiny-llama/lora.yml
+```
+
+qLoRa:
+
+```
+accelerate launch -m axolotl.cli.train examples/tiny-llama/qlora.yml
+```
+
+Both take about 10 minutes to complete on a 4090.

examples/tiny-llama/lora.yml

@@ -1,5 +1,4 @@
-base_model: PY007/TinyLlama-1.1B-step-50K-105b
-
+base_model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 is_llama_derived_model: true
@@ -12,11 +11,12 @@ datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
 dataset_prepared_path:
-val_set_size: 0.01
+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:
@@ -29,12 +29,12 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
 micro_batch_size: 2
-num_epochs: 3
+num_epochs: 4
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
@@ -54,15 +54,11 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
-eval_table_size:
-save_steps:
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0
 fsdp:
 fsdp_config:
 special_tokens:
-  bos_token: "<s>"
-  eos_token: "</s>"
-  unk_token: "<unk>"

examples/tiny-llama/pretrain.yml

@@ -1,33 +1,28 @@
-base_model: mistralai/Mistral-7B-v0.1
-model_type: MistralForCausalLM
-tokenizer_type: LlamaTokenizer
-is_mistral_derived_model: true
-multimodal: true
+base_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
 
-vision_tower: openai/clip-vit-large-patch14
-tune_mm_mlp_adapter: true
-mm_vision_select_layer: -2
-mm_projector_type: mlp2x_gelu
-mm_image_folder: ./llava/
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+is_llama_derived_model: true
 
 load_in_8bit: false
 load_in_4bit: false
 strict: false
 
-datasets:
-  - path: liuhaotian/LLaVA-CC3M-Pretrain-595K
+max_steps: 200
+pretraining_dataset:
+  path: c4
+  name: en
 dataset_prepared_path:
-val_set_size: 0.01
-output_dir: ./out
+val_set_size: 0.0
+output_dir: ./model-out
 
 sequence_len: 2048
 sample_packing: true
-pad_to_sequence_len: true
 
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 4
@@ -35,7 +30,7 @@ micro_batch_size: 2
 num_epochs: 4
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
-learning_rate: 0.002
+learning_rate: 0.0002
 
 train_on_inputs: false
 group_by_length: false
@@ -52,12 +47,12 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 0.05
-save_steps:
+evals_per_epoch:
+eval_table_size:
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0
 fsdp:
 fsdp_config:
 special_tokens:
-  pad_token: "<unk>"

examples/tiny-llama/qlora.yml

@@ -0,0 +1,66 @@
+base_model: TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+is_llama_derived_model: 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: ./qlora-out
+
+adapter: qlora
+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_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: 2
+num_epochs: 4
+optimizer: paged_adamw_32bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+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: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:

examples/xgen-7b/xgen-7b-8k-qlora.yml

@@ -16,7 +16,7 @@ datasets:
       - openassistant_best_replies_train.jsonl
     type: "completion"
 dataset_prepared_path:
-val_set_size: 0.01
+val_set_size: 0.05
 # enable QLoRA
 adapter: qlora
 lora_model_dir:
@@ -38,7 +38,7 @@ lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
-wandb_run_id:
+wandb_name:
 wandb_log_model:
 output_dir: ./qlora-out
 
@@ -51,7 +51,7 @@ output_dir: ./qlora-out
 # decrease if OOM, increase for max VRAM utilization
 micro_batch_size: 1
 gradient_accumulation_steps: 1
-num_epochs: 3
+num_epochs: 4
 # Optimizer for QLoRA
 optimizer: paged_adamw_32bit
 torchdistx_path:
@@ -78,8 +78,8 @@ flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 10
-eval_steps: 50
-save_steps: 50
+evals_per_epoch: 4
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/yi-34B-chat/README.md

@@ -0,0 +1,5 @@
+# Overview
+
+This is an example of a Yi-34B-Chat configuration. It demonstrates that it is possible to finetune a 34B model on a GPU with 24GB of VRAM.
+
+Tested on an RTX 4090 with `python -m axolotl.cli.train examples/mistral/qlora.yml`, a single epoch of finetuning on the alpaca dataset using qlora runs in 47 mins, using 97% of available memory.

examples/yi-34B-chat/qlora.yml

@@ -0,0 +1,76 @@
+base_model: 01-ai/Yi-34B-Chat
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+is_mistral_derived_model: false
+is_llama_derived_model: true
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+sequence_len: 1024
+bf16: true
+fp16: false
+tf32: false
+flash_attention: true
+special_tokens:
+  bos_token: "<|startoftext|>"
+  eos_token: "<|endoftext|>"
+  unk_token: "<unk>"
+
+# Data
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+warmup_steps: 10
+
+# Iterations
+num_epochs: 1
+
+# Evaluation
+val_set_size: 0.1
+evals_per_epoch: 5
+eval_table_size:
+eval_table_max_new_tokens: 128
+eval_sample_packing: false
+eval_batch_size: 1
+
+# LoRA
+output_dir: ./qlora-out
+adapter: qlora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+lora_target_modules:
+
+# Sampling
+sample_packing: false
+pad_to_sequence_len: false
+
+# Batching
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+gradient_checkpointing: true
+
+# wandb
+wandb_project:
+
+# Optimizer
+optimizer: paged_adamw_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+# Misc
+train_on_inputs: false
+group_by_length: false
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+debug:
+deepspeed:
+weight_decay: 0
+fsdp:
+fsdp_config:

requirements.txt

@@ -1,27 +1,26 @@
---extra-index-url https://download.pytorch.org/whl/cu118
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
-torch==2.0.1
-auto-gptq
-packaging
-peft @ git+https://github.com/huggingface/peft.git
-transformers @ git+https://github.com/huggingface/transformers.git@acc394c4f5e1283c19783581790b3dc3105a3697
+packaging==23.2
+peft==0.7.0
+transformers @ git+https://github.com/huggingface/transformers.git@3cefac1d974db5e2825a0cb2b842883a628be7a0
+tokenizers==0.15.0
 bitsandbytes>=0.41.1
-accelerate @ git+https://github.com/huggingface/accelerate@80da9cfb09bb3cc9f1b385cb55d6b90d025a5fd9
+accelerate @ git+https://github.com/huggingface/accelerate.git@0d2280dadc6a93413a5496613b7fdda3a4d2551b
 deepspeed
 addict
 fire
 PyYAML>=6.0
-datasets
-flash-attn>=2.3.0
+datasets>=2.15.0
+flash-attn==2.3.3
 sentencepiece
 wandb
 einops
-xformers>=0.0.22
-optimum
+xformers==0.0.22
+optimum==1.13.2
 hf_transfer
 colorama
 numba
 numpy>=1.24.4
+mlflow
 # qlora things
 bert-score==0.3.13
 evaluate==0.4.0
@@ -30,4 +29,15 @@ scipy
 scikit-learn==1.2.2
 pynvml
 art
-fschat==0.2.29
+fschat==0.2.34
+gradio==3.50.2
+tensorboard
+
+mamba-ssm==1.1.1
+
+# remote filesystems
+s3fs
+gcsfs
+# adlfs
+
+trl>=0.7.9

scripts/finetune.py

@@ -45,8 +45,6 @@ def do_cli(config: Path = Path("examples/"), **kwargs):
         shard(cfg=parsed_cfg, cli_args=parsed_cli_args)
     else:
         dataset_meta = load_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
-        if parsed_cli_args.prepare_ds_only:
-            return
         train(cfg=parsed_cfg, cli_args=parsed_cli_args, dataset_meta=dataset_meta)
 
 

setup.py

@@ -1,5 +1,7 @@
 """setup.py for axolotl"""
 
+from importlib.metadata import PackageNotFoundError, version
+
 from setuptools import find_packages, setup
 
 
@@ -9,25 +11,27 @@ def parse_requirements():
     with open("./requirements.txt", encoding="utf-8") as requirements_file:
         lines = [r.strip() for r in requirements_file.readlines()]
         for line in lines:
+            is_extras = (
+                "flash-attn" in line
+                or "flash-attention" in line
+                or "deepspeed" in line
+                or "mamba-ssm" in line
+            )
             if line.startswith("--extra-index-url"):
                 # Handle custom index URLs
                 _, url = line.split()
                 _dependency_links.append(url)
-            elif (
-                "flash-attn" not in line
-                and "deepspeed" not in line
-                and line
-                and line[0] != "#"
-            ):
+            elif not is_extras and line and line[0] != "#":
                 # Handle standard packages
                 _install_requires.append(line)
 
-    # TODO(wing) remove once xformers release supports torch 2.1.0
-    if "torch==2.1.0" in _install_requires:
-        _install_requires.pop(_install_requires.index("xformers>=0.0.22"))
-        _install_requires.append(
-            "xformers @ git+https://github.com/facebookresearch/xformers.git@main"
-        )
+    try:
+        torch_version = version("torch")
+        if torch_version.startswith("2.1.1"):
+            _install_requires.pop(_install_requires.index("xformers==0.0.22"))
+            _install_requires.append("xformers==0.0.23")
+    except PackageNotFoundError:
+        pass
 
     return _install_requires, _dependency_links
 
@@ -46,10 +50,19 @@ setup(
     dependency_links=dependency_links,
     extras_require={
         "flash-attn": [
-            "flash-attn>=2.3.0",
+            "flash-attn==2.3.3",
+        ],
+        "fused-dense-lib": [
+            "fused-dense-lib  @ git+https://github.com/Dao-AILab/flash-attention@v2.3.3#subdirectory=csrc/fused_dense_lib",
         ],
         "deepspeed": [
             "deepspeed",
         ],
+        "mamba-ssm": [
+            "mamba-ssm==1.0.1",
+        ],
+        "auto-gptq": [
+            "auto-gptq==0.5.1",
+        ],
     },
 )

src/axolotl/cli/__init__.py

@@ -7,27 +7,36 @@ import os
 import random
 import sys
 from pathlib import Path
+from threading import Thread
 from typing import Any, Dict, List, Optional, Union
 
+import gradio as gr
 import torch
 import yaml
 
 # add src to the pythonpath so we don't need to pip install this
 from accelerate.commands.config import config_args
 from art import text2art
+from datasets import concatenate_datasets, load_dataset
 from huggingface_hub import HfApi
 from huggingface_hub.utils import LocalTokenNotFoundError
-from transformers import GenerationConfig, TextStreamer
+from transformers import GenerationConfig, TextIteratorStreamer, TextStreamer
 
 from axolotl.common.cli import TrainerCliArgs, load_model_and_tokenizer
 from axolotl.logging_config import configure_logging
 from axolotl.train import TrainDatasetMeta
-from axolotl.utils.config import normalize_config, validate_config
+from axolotl.utils.config import (
+    normalize_cfg_datasets,
+    normalize_config,
+    validate_config,
+)
 from axolotl.utils.data import prepare_dataset
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import is_main_process
+from axolotl.utils.mlflow_ import setup_mlflow_env_vars
 from axolotl.utils.models import load_tokenizer
 from axolotl.utils.tokenization import check_dataset_labels
+from axolotl.utils.trainer import prepare_optim_env
 from axolotl.utils.wandb_ import setup_wandb_env_vars
 
 project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
@@ -45,7 +54,7 @@ def print_axolotl_text_art(suffix=None):
     ascii_text = "  axolotl"
     if suffix:
         ascii_text += f"  x  {suffix}"
-    ascii_art = text2art(" axolotl", font=font)
+    ascii_art = text2art(ascii_text, font=font)
 
     if is_main_process():
         print(ascii_art)
@@ -69,14 +78,15 @@ def do_merge_lora(
     safe_serialization = cfg.save_safetensors is True
 
     LOG.info("running merge of LoRA with base model")
-    model = model.merge_and_unload()
-    model.to(dtype=torch.float16)
+    model = model.merge_and_unload(progressbar=True)
+    model.to(dtype=cfg.torch_dtype)
 
     if cfg.local_rank == 0:
         LOG.info(f"saving merged model to: {str(Path(cfg.output_dir) / 'merged')}")
         model.save_pretrained(
             str(Path(cfg.output_dir) / "merged"),
             safe_serialization=safe_serialization,
+            progressbar=True,
         )
         tokenizer.save_pretrained(str(Path(cfg.output_dir) / "merged"))
 
@@ -101,15 +111,7 @@ def do_inference(
             importlib.import_module("axolotl.prompters"), prompter
         )
 
-    if cfg.landmark_attention:
-        from axolotl.monkeypatch.llama_landmark_attn import set_model_mem_id
-
-        set_model_mem_id(model, tokenizer)
-        model.set_mem_cache_args(
-            max_seq_len=255, mem_freq=50, top_k=5, max_cache_size=None
-        )
-
-    model = model.to(cfg.device)
+    model = model.to(cfg.device, dtype=cfg.torch_dtype)
 
     while True:
         print("=" * 80)
@@ -154,6 +156,83 @@ def do_inference(
         print(tokenizer.decode(generated["sequences"].cpu().tolist()[0]))
 
 
+def do_inference_gradio(
+    *,
+    cfg: DictDefault,
+    cli_args: TrainerCliArgs,
+):
+    model, tokenizer = load_model_and_tokenizer(cfg=cfg, cli_args=cli_args)
+    prompter = cli_args.prompter
+    default_tokens = {"unk_token": "<unk>", "bos_token": "<s>", "eos_token": "</s>"}
+
+    for token, symbol in default_tokens.items():
+        # If the token isn't already specified in the config, add it
+        if not (cfg.special_tokens and token in cfg.special_tokens):
+            tokenizer.add_special_tokens({token: symbol})
+
+    prompter_module = None
+    if prompter:
+        prompter_module = getattr(
+            importlib.import_module("axolotl.prompters"), prompter
+        )
+
+    model = model.to(cfg.device, dtype=cfg.torch_dtype)
+
+    def generate(instruction):
+        if not instruction:
+            return
+        if prompter_module:
+            # pylint: disable=stop-iteration-return
+            prompt: str = next(
+                prompter_module().build_prompt(instruction=instruction.strip("\n"))
+            )
+        else:
+            prompt = instruction.strip()
+        batch = tokenizer(prompt, return_tensors="pt", add_special_tokens=True)
+
+        model.eval()
+        with torch.no_grad():
+            generation_config = GenerationConfig(
+                repetition_penalty=1.1,
+                max_new_tokens=1024,
+                temperature=0.9,
+                top_p=0.95,
+                top_k=40,
+                bos_token_id=tokenizer.bos_token_id,
+                eos_token_id=tokenizer.eos_token_id,
+                pad_token_id=tokenizer.pad_token_id,
+                do_sample=True,
+                use_cache=True,
+                return_dict_in_generate=True,
+                output_attentions=False,
+                output_hidden_states=False,
+                output_scores=False,
+            )
+            streamer = TextIteratorStreamer(tokenizer)
+            generation_kwargs = {
+                "inputs": batch["input_ids"].to(cfg.device),
+                "generation_config": generation_config,
+                "streamer": streamer,
+            }
+
+            thread = Thread(target=model.generate, kwargs=generation_kwargs)
+            thread.start()
+
+            all_text = ""
+
+            for new_text in streamer:
+                all_text += new_text
+                yield all_text
+
+    demo = gr.Interface(
+        fn=generate,
+        inputs="textbox",
+        outputs="text",
+        title=cfg.get("gradio_title", "Axolotl Gradio Interface"),
+    )
+    demo.queue().launch(show_api=False, share=True)
+
+
 def choose_config(path: Path):
     yaml_files = list(path.glob("*.yml"))
 
@@ -210,51 +289,19 @@ def load_cfg(config: Path = Path("examples/"), **kwargs):
 
     validate_config(cfg)
 
+    prepare_optim_env(cfg)
+
     normalize_config(cfg)
 
+    normalize_cfg_datasets(cfg)
+
     setup_wandb_env_vars(cfg)
+
+    setup_mlflow_env_vars(cfg)
+
     return cfg
 
 
-def load_mm_dataset(
-    *,
-    cfg: DictDefault,
-    cli_args: TrainerCliArgs,  # pylint: disable=unused-argument
-    model,
-):
-    # pylint: disable=duplicate-code
-    from llava.train.train import DataArguments, LazySupervisedDataset
-
-    vision_tower = model.get_vision_tower()
-    data_args = DataArguments(
-        data_path=cfg.datasets[0]["path"],
-        lazy_preprocess=cfg.mm_lazy_preprocess
-        if cfg.mm_lazy_preprocess is not None
-        else True,
-        is_multimodal=True,
-        image_folder=cfg.mm_image_folder or None,
-        image_aspect_ratio=cfg.mm_image_aspect_ratio or "square",
-        image_grid_pinpoints=cfg.mm_image_grid_pinpoints or None,
-    )
-    data_args.image_processor = vision_tower.image_processor
-    tokenizer = load_tokenizer(cfg)
-    train_dataset = LazySupervisedDataset(
-        tokenizer=tokenizer,
-        data_path=data_args["data_path"],
-        data_args=data_args,
-    )
-
-    total_num_steps = int(
-        math.ceil(len(train_dataset) * cfg.num_epochs / cfg.batch_size)
-    )
-
-    return TrainDatasetMeta(
-        train_dataset=train_dataset,
-        eval_dataset=None,
-        total_num_steps=total_num_steps,
-    )
-
-
 def load_datasets(
     *,
     cfg: DictDefault,
@@ -262,7 +309,9 @@ def load_datasets(
 ) -> TrainDatasetMeta:
     tokenizer = load_tokenizer(cfg)
 
-    train_dataset, eval_dataset, total_num_steps = prepare_dataset(cfg, tokenizer)
+    train_dataset, eval_dataset, total_num_steps, prompters = prepare_dataset(
+        cfg, tokenizer
+    )
 
     if cli_args.debug or cfg.debug:
         LOG.info("check_dataset_labels...")
@@ -278,6 +327,98 @@ def load_datasets(
             text_only=cli_args.debug_text_only,
         )
 
+        LOG.info("printing prompters...")
+        for prompter in prompters:
+            LOG.info(prompter)
+
+    return TrainDatasetMeta(
+        train_dataset=train_dataset,
+        eval_dataset=eval_dataset,
+        total_num_steps=total_num_steps,
+    )
+
+
+def load_rl_datasets(
+    *,
+    cfg: DictDefault,
+    cli_args: TrainerCliArgs,  # pylint: disable=unused-argument
+) -> TrainDatasetMeta:
+    train_datasets: List[Any] = []
+    for i, ds_cfg in enumerate(cfg.datasets):
+        train_datasets.insert(i, load_dataset(ds_cfg["path"], split=ds_cfg["split"]))
+    # eval_dataset = load_dataset(
+    #     cfg.test_datasets[0]["path"], split=cfg.test_datasets[0]["split"]
+    # )
+    eval_dataset = None
+
+    def argilla_apply_chatml(sample):  # pylint: disable=possibly-unused-variable
+        if "system" in sample and sample["system"]:
+            sample["prompt"] = (
+                f"<|im_start|>system\n{sample['system']}<|im_end|>\n"
+                f"<|im_start|>user\n{sample['instruction']}<|im_end|>\n<|im_start|>assistant\n"
+            )
+        else:
+            sample[
+                "prompt"
+            ] = f"<|im_start|>user\n{sample['instruction']}<|im_end|>\n<|im_start|>assistant\n"
+        sample["chosen"] = f"{sample['chosen_response']}<|im_end|>"
+        sample["rejected"] = f"{sample['rejected_response']}<|im_end|>"
+        return sample
+
+    def intel_apply_chatml(sample):  # pylint: disable=possibly-unused-variable
+        if "system" in sample and sample["system"]:
+            sample["prompt"] = (
+                f"<|im_start|>system\n{sample['system']}<|im_end|>\n"
+                f"<|im_start|>user\n{sample['question']}<|im_end|>\n<|im_start|>assistant\n"
+            )
+        else:
+            sample[
+                "prompt"
+            ] = f"<|im_start|>user\n{sample['question']}<|im_end|>\n<|im_start|>assistant\n"
+        sample["chosen"] = f"{sample['chosen']}<|im_end|>"
+        sample["rejected"] = f"{sample['rejected']}<|im_end|>"
+        return sample
+
+    def apply_chatml(sample):  # pylint: disable=possibly-unused-variable
+        if "system" in sample and sample["system"]:
+            sample["prompt"] = (
+                f"<|im_start|>system\n{sample['system']}<|im_end|>\n"
+                f"<|im_start|>user\n{sample['prompt']}<|im_end|>\n<|im_start|>assistant\n"
+            )
+        else:
+            sample[
+                "prompt"
+            ] = f"<|im_start|>user\n{sample['prompt']}<|im_end|>\n<|im_start|>assistant\n"
+        sample["chosen"] = f"{sample['chosen']}<|im_end|>"
+        sample["rejected"] = f"{sample['rejected']}<|im_end|>"
+        return sample
+
+    def ultra_apply_chatml(sample):  # pylint: disable=possibly-unused-variable
+        if "system" in sample and sample["system"]:
+            sample["prompt"] = (
+                f"<|im_start|>system\n{sample['system']}<|im_end|>\n"
+                f"<|im_start|>user\n{sample['prompt']}<|im_end|>\n<|im_start|>assistant\n"
+            )
+        else:
+            sample[
+                "prompt"
+            ] = f"<|im_start|>user\n{sample['prompt']}<|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
+
+    for i, data_set in enumerate(train_datasets):
+        _type = cfg.datasets[i]["type"]
+        ds_type_fn = locals()[_type]
+        train_datasets[i] = data_set.map(ds_type_fn)
+    train_dataset = concatenate_datasets(train_datasets)
+
+    # eval_dataset = eval_dataset.map(intel_apply_chatml)
+
+    total_num_steps = int(
+        math.ceil(len(train_dataset) * cfg.num_epochs / cfg.batch_size)
+    )
+
     return TrainDatasetMeta(
         train_dataset=train_dataset,
         eval_dataset=eval_dataset,

src/axolotl/cli/inference.py

@@ -6,11 +6,16 @@ from pathlib import Path
 import fire
 import transformers
 
-from axolotl.cli import do_inference, load_cfg, print_axolotl_text_art
+from axolotl.cli import (
+    do_inference,
+    do_inference_gradio,
+    load_cfg,
+    print_axolotl_text_art,
+)
 from axolotl.common.cli import TrainerCliArgs
 
 
-def do_cli(config: Path = Path("examples/"), **kwargs):
+def do_cli(config: Path = Path("examples/"), gradio=False, **kwargs):
     # pylint: disable=duplicate-code
     print_axolotl_text_art()
     parsed_cfg = load_cfg(config, **kwargs)
@@ -21,7 +26,10 @@ def do_cli(config: Path = Path("examples/"), **kwargs):
     )
     parsed_cli_args.inference = True
 
-    do_inference(cfg=parsed_cfg, cli_args=parsed_cli_args)
+    if gradio:
+        do_inference_gradio(cfg=parsed_cfg, cli_args=parsed_cli_args)
+    else:
+        do_inference(cfg=parsed_cfg, cli_args=parsed_cli_args)
 
 
 if __name__ == "__main__":

src/axolotl/cli/merge_lora.py

@@ -18,7 +18,26 @@ def do_cli(config: Path = Path("examples/"), **kwargs):
         return_remaining_strings=True
     )
     parsed_cli_args.merge_lora = True
-    parsed_cfg = load_cfg(config, merge_lora=True, **kwargs)
+
+    parsed_cfg = load_cfg(
+        config,
+        merge_lora=True,
+        load_in_8bit=False,
+        load_in_4bit=False,
+        flash_attention=False,
+        **kwargs,
+    )
+
+    if not parsed_cfg.lora_model_dir and parsed_cfg.output_dir:
+        parsed_cfg.lora_model_dir = parsed_cfg.output_dir
+    if not Path(parsed_cfg.lora_model_dir).exists():
+        raise ValueError(
+            f"Target directory for merge: `{parsed_cfg.lora_model_dir}` does not exist."
+        )
+
+    parsed_cfg.load_in_4bit = False
+    parsed_cfg.load_in_8bit = False
+    parsed_cfg.flash_attention = False
 
     do_merge_lora(cfg=parsed_cfg, cli_args=parsed_cli_args)
 

src/axolotl/cli/preprocess.py

@@ -12,15 +12,13 @@ from axolotl.cli import (
     check_accelerate_default_config,
     check_user_token,
     load_cfg,
-    load_mm_dataset,
+    load_datasets,
     print_axolotl_text_art,
 )
-from axolotl.common.cli import TrainerCliArgs
+from axolotl.common.cli import PreprocessCliArgs
 from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
-from axolotl.train import train
-from axolotl.utils.models import load_model, load_tokenizer
 
-LOG = logging.getLogger("axolotl.cli.train")
+LOG = logging.getLogger("axolotl.cli.preprocess")
 
 
 def do_cli(config: Path = Path("examples/"), **kwargs):
@@ -29,27 +27,27 @@ def do_cli(config: Path = Path("examples/"), **kwargs):
     parsed_cfg = load_cfg(config, **kwargs)
     check_accelerate_default_config()
     check_user_token()
-    parser = transformers.HfArgumentParser((TrainerCliArgs))
+    parser = transformers.HfArgumentParser((PreprocessCliArgs))
     parsed_cli_args, _ = parser.parse_args_into_dataclasses(
         return_remaining_strings=True
     )
-    if parsed_cli_args.prepare_ds_only and not parsed_cfg.dataset_prepared_path:
+
+    if not parsed_cfg.dataset_prepared_path:
         msg = (
             Fore.RED
-            + "--prepare_ds_only called without dataset_prepared_path set."
+            + "preprocess CLI called without dataset_prepared_path set, "
+            + f"using default path: {DEFAULT_DATASET_PREPARED_PATH}"
             + Fore.RESET
         )
         LOG.warning(msg)
         parsed_cfg.dataset_prepared_path = DEFAULT_DATASET_PREPARED_PATH
 
-    tokenizer = load_tokenizer(parsed_cfg)
-    model, _ = load_model(parsed_cfg, tokenizer)
-    dataset_meta = load_mm_dataset(
-        cfg=parsed_cfg, cli_args=parsed_cli_args, model=model
+    _ = load_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
+    LOG.info(
+        Fore.GREEN
+        + f"Success! Preprocessed data path: `dataset_prepared_path: {parsed_cfg.dataset_prepared_path}`"
+        + Fore.RESET
     )
-    if parsed_cli_args.prepare_ds_only:
-        return
-    train(cfg=parsed_cfg, cli_args=parsed_cli_args, dataset_meta=dataset_meta)
 
 
 if __name__ == "__main__":

src/axolotl/cli/train.py

@@ -6,17 +6,16 @@ from pathlib import Path
 
 import fire
 import transformers
-from colorama import Fore
 
 from axolotl.cli import (
     check_accelerate_default_config,
     check_user_token,
     load_cfg,
     load_datasets,
+    load_rl_datasets,
     print_axolotl_text_art,
 )
 from axolotl.common.cli import TrainerCliArgs
-from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
 from axolotl.train import train
 
 LOG = logging.getLogger("axolotl.cli.train")
@@ -24,26 +23,19 @@ LOG = logging.getLogger("axolotl.cli.train")
 
 def do_cli(config: Path = Path("examples/"), **kwargs):
     # pylint: disable=duplicate-code
-    print_axolotl_text_art()
     parsed_cfg = load_cfg(config, **kwargs)
+    print_axolotl_text_art()
     check_accelerate_default_config()
     check_user_token()
     parser = transformers.HfArgumentParser((TrainerCliArgs))
     parsed_cli_args, _ = parser.parse_args_into_dataclasses(
         return_remaining_strings=True
     )
-    if parsed_cli_args.prepare_ds_only and not parsed_cfg.dataset_prepared_path:
-        msg = (
-            Fore.RED
-            + "--prepare_ds_only called without dataset_prepared_path set."
-            + Fore.RESET
-        )
-        LOG.warning(msg)
-        parsed_cfg.dataset_prepared_path = DEFAULT_DATASET_PREPARED_PATH
 
-    dataset_meta = load_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
-    if parsed_cli_args.prepare_ds_only:
-        return
+    if parsed_cfg.rl:
+        dataset_meta = load_rl_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
+    else:
+        dataset_meta = load_datasets(cfg=parsed_cfg, cli_args=parsed_cli_args)
     train(cfg=parsed_cfg, cli_args=parsed_cli_args, dataset_meta=dataset_meta)
 
 

src/axolotl/common/cli.py

@@ -25,11 +25,22 @@ class TrainerCliArgs:
     debug_num_examples: int = field(default=5)
     inference: bool = field(default=False)
     merge_lora: bool = field(default=False)
-    prepare_ds_only: bool = field(default=False)
     prompter: Optional[str] = field(default=None)
     shard: bool = field(default=False)
 
 
+@dataclass
+class PreprocessCliArgs:
+    """
+    dataclass representing arguments for preprocessing only
+    """
+
+    debug: bool = field(default=False)
+    debug_text_only: bool = field(default=False)
+    debug_num_examples: int = field(default=1)
+    prompter: Optional[str] = field(default=None)
+
+
 def load_model_and_tokenizer(
     *,
     cfg: DictDefault,

src/axolotl/core/trainer_builder.py

@@ -1,3 +1,4 @@
+# pylint: disable=too-many-lines
 """
 Builder for the training args and trainer
 """
@@ -6,48 +7,48 @@ import abc
 import importlib
 import logging
 import math
-import os
 import sys
 from abc import abstractmethod
 from dataclasses import dataclass, field
-from functools import partial
+from functools import wraps
 from pathlib import Path
-from typing import Optional, Union
+from typing import Optional
 
 import torch
 import transformers
 from datasets import Dataset
 from torch.optim.lr_scheduler import OneCycleLR
-from torch.utils.data import DataLoader, DistributedSampler, SequentialSampler
+from torch.utils.data import BatchSampler, DataLoader, RandomSampler, SequentialSampler
 from transformers import EarlyStoppingCallback, Trainer, TrainingArguments
-from transformers.trainer_pt_utils import SequentialDistributedSampler
+from transformers.trainer_utils import seed_worker
+from trl import DPOTrainer
 
 from axolotl.monkeypatch.relora import ReLoRACallback, ReLoRAScheduler
 from axolotl.utils.callbacks import (
     EvalFirstStepCallback,
     GPUStatsCallback,
+    LossWatchDogCallback,
     SaveAxolotlConfigtoWandBCallback,
     SaveBetterTransformerModelCallback,
     bench_eval_callback_factory,
     log_prediction_callback_factory,
 )
-from axolotl.utils.collators import DataCollatorForSeq2Seq
-from axolotl.utils.dataloader import MultipackDistributedDataloader
-from axolotl.utils.schedulers import get_cosine_schedule_with_quadratic_warmup
+from axolotl.utils.collators import (
+    BatchSamplerDataCollatorForSeq2Seq,
+    DataCollatorForSeq2Seq,
+    MambaDataCollator,
+)
+from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
+from axolotl.utils.schedulers import (
+    get_cosine_schedule_with_min_lr,
+    get_cosine_schedule_with_quadratic_warmup,
+)
 
 try:
     import torch._dynamo  # pylint: disable=ungrouped-imports
 except ImportError:
     pass
 
-try:
-    from llava.train.llava_trainer import get_mm_adapter_state_maybe_zero_3
-except ImportError:
-
-    def get_mm_adapter_state_maybe_zero_3(named_params, keys_to_match):
-        raise ImportError("missing LLaVA package")
-
-
 LOG = logging.getLogger("axolotl.core.trainer_builder")
 
 
@@ -57,10 +58,19 @@ class AxolotlTrainingArguments(TrainingArguments):
     Extend the base TrainingArguments for axolotl helpers
     """
 
+    model_type: Optional[str] = field(
+        default=None, metadata={"help": "HF model configuration model_type."}
+    )
     lr_quadratic_warmup: bool = field(
         default=False,
         metadata={"help": "Use quadratic warmup for cosine scheduling."},
     )
+    pretraining: bool = field(
+        default=False,
+        metadata={
+            "help": "Indicates to trainer whether we are doing continued pretraining."
+        },
+    )
     sample_packing: bool = field(
         default=False,
         metadata={"help": "Use sample packing for efficient training."},
@@ -110,6 +120,14 @@ class AxolotlTrainingArguments(TrainingArguments):
     bench_source_max_len: int = field(
         default=2048, metadata={"help": "Maximum source sequence length for bench."}
     )
+    dataloader_prefetch_factor: Optional[int] = field(
+        default=None,
+        metadata={"help": "prefetch_factor argument to the dataloader"},
+    )
+    cosine_min_lr_ratio: Optional[float] = field(
+        default=None,
+        metadata={"help": "Minimum learning rate is min_lr_ratio * learning_rate"},
+    )
 
 
 class AxolotlTrainer(Trainer):
@@ -118,10 +136,21 @@ class AxolotlTrainer(Trainer):
     """
 
     args = None  # type: AxolotlTrainingArguments
+    tag_names = ["axolotl"]
 
-    def __init__(self, *args, bench_data_collator=None, **kwargs):
+    def __init__(
+        self,
+        *_args,
+        num_epochs=1,
+        bench_data_collator=None,
+        eval_data_collator=None,
+        **kwargs
+    ):
+        self.num_epochs = num_epochs
         self.bench_data_collator = bench_data_collator
-        super().__init__(*args, **kwargs)
+        self.eval_data_collator = eval_data_collator
+        super().__init__(*_args, **kwargs)
+        self.train_data_collator = self.data_collator
 
     def create_scheduler(
         self, num_training_steps: int, optimizer: torch.optim.Optimizer = None
@@ -147,74 +176,120 @@ class AxolotlTrainer(Trainer):
                     num_warmup_steps=self.args.get_warmup_steps(num_training_steps),
                     num_training_steps=num_training_steps,
                 )
+            elif self.args.lr_scheduler_type == "cosine" and self.args.cosine_min_lr_ratio is not None:
+                assert 0 <= self.args.cosine_min_lr_ratio <= 1.0, "cosine_min_lr_ratio must be between 0.0 and 1.0"
+                if self.args.deepspeed:
+                    LOG.warning("Using cosine scheduler with deepspeed. This may be ignored if a scheduler is set \
+                                in the deepspeed JSON")
+                self.lr_scheduler = get_cosine_schedule_with_min_lr(  # pylint: disable=attribute-defined-outside-init
+                    optimizer,
+                    num_warmup_steps=self.args.get_warmup_steps(num_training_steps),
+                    num_training_steps=num_training_steps,
+                    min_lr_ratio=self.args.cosine_min_lr_ratio,
+                )
             else:
                 return super().create_scheduler(num_training_steps, optimizer)
         return self.lr_scheduler
 
     def _get_train_sampler(self) -> Optional[torch.utils.data.Sampler]:
-        if self.args.world_size > 1 and self.args.sample_packing:
-            return DistributedSampler(
-                self.train_dataset,
-                num_replicas=self.args.world_size,
-                rank=self.args.process_index,
-                seed=self.args.seed,
+        if self.args.sample_packing and not self.args.pretraining:
+            return MultipackBatchSampler(
+                RandomSampler(self.train_dataset),
+                self.args.train_batch_size,
+                drop_last=True,
+                batch_max_len=self._train_batch_size * self.args.max_seq_length,
+                lengths=get_dataset_lengths(self.train_dataset),
+                packing_efficiency_estimate=self.args.sample_packing_efficiency,
             )
         return super()._get_train_sampler()
 
     def _get_eval_sampler(
         self, eval_dataset: Dataset
     ) -> Optional[torch.utils.data.Sampler]:
-        if (
-            self.args.world_size > 1
-            and self.args.sample_packing
-            and self.args.eval_sample_packing is not False
-        ):
-            return SequentialDistributedSampler(
-                eval_dataset,
-                num_replicas=self.args.world_size,
-                rank=self.args.process_index,
-                batch_size=self.args.per_device_eval_batch_size,
+        if self.args.sample_packing and self.args.eval_sample_packing is not False:
+            return MultipackBatchSampler(
+                SequentialSampler(eval_dataset),
+                self.args.per_device_eval_batch_size,
+                drop_last=True,
+                batch_max_len=self.args.eval_batch_size * self.args.max_seq_length,
+                lengths=get_dataset_lengths(eval_dataset),
+                packing_efficiency_estimate=self.args.sample_packing_efficiency,
             )
         return super()._get_eval_sampler(eval_dataset)
 
-    def get_train_dataloader(self) -> Union[DataLoader, MultipackDistributedDataloader]:
-        if self.args.sample_packing:
-            train_sampler = self._get_train_sampler()
-            return self.accelerator.prepare(
-                MultipackDistributedDataloader(
-                    self.train_dataset,
-                    batch_size=self._train_batch_size,
-                    seq_max_length=self.args.max_seq_length,
-                    collate_fn=self.data_collator,
-                    sampler=train_sampler,
-                    packing_efficiency_estimate=self.args.sample_packing_efficiency,
-                    sample_packing_seq_len_multiplier=self.args.sample_packing_seq_len_multiplier,
-                    device_count=int(os.environ.get("WORLD_SIZE", 1)),
-                )
+    def get_train_dataloader(self) -> DataLoader:
+        if self.args.sample_packing and not self.args.pretraining:
+            train_dataset = self.train_dataset
+            train_dataset = train_dataset.remove_columns(["length"])
+            data_collator = self.data_collator
+            dataloader_params = {
+                "batch_size": self._train_batch_size,
+                "collate_fn": data_collator,
+                "num_workers": self.args.dataloader_num_workers,
+                "pin_memory": self.args.dataloader_pin_memory,
+            }
+            if self.args.dataloader_prefetch_factor:
+                dataloader_params[
+                    "prefetch_factor"
+                ] = self.args.dataloader_prefetch_factor
+
+            sampler = self._get_train_sampler()
+            if isinstance(sampler, BatchSampler):
+                dataloader_params["batch_sampler"] = sampler
+                del dataloader_params["batch_size"]
+            else:
+                dataloader_params["sampler"] = sampler
+                dataloader_params["drop_last"] = self.args.dataloader_drop_last
+            dataloader_params["worker_init_fn"] = seed_worker
+
+            self.accelerator.even_batches = False
+            return self.accelerator.prepare_data_loader(
+                DataLoader(train_dataset, **dataloader_params)
             )
         return super().get_train_dataloader()
 
-    def get_eval_dataloader(
-        self, eval_dataset: Optional[Dataset] = None
-    ) -> Union[DataLoader, MultipackDistributedDataloader]:
+    def get_eval_dataloader(self, eval_dataset: Optional[Dataset] = None) -> DataLoader:
+        if self.args.sample_packing and self.args.eval_sample_packing is False:
+            self.data_collator = (  # pylint: disable=attribute-defined-outside-init
+                self.eval_data_collator
+            )
+            dataloader = super().get_eval_dataloader(eval_dataset)
+            self.data_collator = (  # pylint: disable=attribute-defined-outside-init
+                self.train_data_collator
+            )
+            return dataloader
+
         if self.args.sample_packing and self.args.eval_sample_packing is not False:
             eval_dataset = (
                 eval_dataset if eval_dataset is not None else self.eval_dataset
             )
 
             eval_sampler = self._get_eval_sampler(eval_dataset)
-            return self.accelerator.prepare(
-                MultipackDistributedDataloader(
-                    eval_dataset,
-                    batch_size=self.args.eval_batch_size,
-                    seq_max_length=self.args.max_seq_length,
-                    collate_fn=self.data_collator,
-                    sampler=eval_sampler,
-                    packing_efficiency_estimate=self.args.sample_packing_efficiency,
-                    sample_packing_seq_len_multiplier=self.args.eval_batch_size,
-                    device_count=int(os.environ.get("WORLD_SIZE", 1)),
-                )
+            eval_dataset = eval_dataset.remove_columns(["length"])
+            data_collator = self.data_collator
+            dataloader_params = {
+                "batch_size": self.args.eval_batch_size,
+                "collate_fn": data_collator,
+                "num_workers": self.args.dataloader_num_workers,
+                "pin_memory": self.args.dataloader_pin_memory,
+            }
+            if self.args.dataloader_prefetch_factor:
+                dataloader_params[
+                    "prefetch_factor"
+                ] = self.args.dataloader_prefetch_factor
+
+            if isinstance(eval_sampler, BatchSampler):
+                dataloader_params["batch_sampler"] = eval_sampler
+                del dataloader_params["batch_size"]
+            else:
+                dataloader_params["sampler"] = eval_sampler
+                dataloader_params["drop_last"] = self.args.dataloader_drop_last
+
+            self.accelerator.even_batches = False
+            return self.accelerator.prepare_data_loader(
+                DataLoader(eval_dataset, **dataloader_params)
             )
+
         return super().get_eval_dataloader(eval_dataset)
 
     def _get_bench_sampler(
@@ -227,13 +302,15 @@ class AxolotlTrainer(Trainer):
     def get_bench_dataloader(
         self,
         bench_dataset: Dataset,
-    ) -> Union[DataLoader, MultipackDistributedDataloader]:
+    ) -> DataLoader:
         dataloader_params = {
             "batch_size": self.args.eval_batch_size,
             "collate_fn": self.bench_data_collator,
             "num_workers": self.args.dataloader_num_workers,
             "pin_memory": self.args.dataloader_pin_memory,
         }
+        if self.args.dataloader_prefetch_factor:
+            dataloader_params["prefetch_factor"] = self.args.dataloader_prefetch_factor
 
         if not isinstance(bench_dataset, torch.utils.data.IterableDataset):
             dataloader_params["sampler"] = self._get_bench_sampler(bench_dataset)
@@ -251,35 +328,60 @@ class AxolotlTrainer(Trainer):
         #     return (loss, outputs) if return_outputs else loss
         return super().compute_loss(model, inputs, return_outputs=return_outputs)
 
-    def _save_checkpoint(self, model, trial, metrics=None):
-        if getattr(self.args, "tune_mm_mlp_adapter", False):
-            from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR
+    def _sanitize_kwargs_for_tagging(self, tag_names, kwargs=None):
+        if isinstance(tag_names, str):
+            tag_names = [tag_names]
 
-            checkpoint_folder = f"{PREFIX_CHECKPOINT_DIR}-{self.state.global_step}"
+        if kwargs is not None:
+            if "tags" not in kwargs:
+                kwargs["tags"] = tag_names
+            elif "tags" in kwargs and isinstance(kwargs["tags"], list):
+                kwargs["tags"].extend(tag_names)
+            elif "tags" in kwargs and isinstance(kwargs["tags"], str):
+                tag_names.append(kwargs["tags"])
+                kwargs["tags"] = tag_names
 
-            run_dir = self._get_output_dir(trial=trial)
-            output_dir = os.path.join(run_dir, checkpoint_folder)
+        return kwargs
 
-            # Only save Adapter
-            keys_to_match = ["mm_projector", "vision_resampler"]
-            if getattr(self.args, "use_im_start_end", False):
-                keys_to_match.extend(["embed_tokens", "embed_in"])
+    @wraps(Trainer.push_to_hub)
+    def push_to_hub(self, *args, **kwargs) -> str:
+        """
+        Overwrite the `push_to_hub` method in order to force-add the tags when pushing the
+        model on the Hub. Please refer to `~transformers.Trainer.push_to_hub` for more details.
+        """
+        kwargs = self._sanitize_kwargs_for_tagging(
+            tag_names=self.tag_names, kwargs=kwargs
+        )
 
-            weight_to_save = get_mm_adapter_state_maybe_zero_3(
-                self.model.named_parameters(), keys_to_match
-            )
+        return super().push_to_hub(*args, **kwargs)
 
-            if self.args.local_rank in (0, -1):
-                self.model.config.save_pretrained(output_dir)
-                torch.save(weight_to_save, os.path.join(output_dir, "mm_projector.bin"))
-        else:
-            super()._save_checkpoint(model, trial, metrics)
 
-    def _save(self, output_dir: Optional[str] = None, state_dict=None):
-        if getattr(self.args, "tune_mm_mlp_adapter", False):
-            pass
-        else:
-            super()._save(output_dir, state_dict)
+class AxolotlMambaTrainer(AxolotlTrainer):
+    """
+    Mamba specific trainer to handle loss calculation
+    """
+
+    tag_names = ["axolotl", "mamba"]
+
+    def compute_loss(
+        self,
+        model,
+        inputs,
+        return_outputs=False,  # pylint: disable=unused-argument
+    ):
+        input_ids = inputs.pop("input_ids")
+        lm_logits = model(input_ids).logits
+
+        labels = input_ids.to(lm_logits.device)
+        shift_logits = lm_logits[:, :-1, :].contiguous()
+        labels = labels[:, 1:].contiguous()
+
+        loss_fct = torch.nn.CrossEntropyLoss()
+        lm_loss = loss_fct(
+            shift_logits.view(-1, shift_logits.size(-1)), labels.view(-1)
+        )
+
+        return lm_loss
 
 
 class OneCycleLRSchedulerTrainer(AxolotlTrainer):
@@ -287,6 +389,8 @@ class OneCycleLRSchedulerTrainer(AxolotlTrainer):
     Trainer subclass that uses the OneCycleLR scheduler
     """
 
+    tag_names = ["axolotl", "onecycle"]
+
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         self.lr_scheduler = None
@@ -316,6 +420,8 @@ class ReLoRATrainer(AxolotlTrainer):
     Trainer subclass that uses the OneCycleLR scheduler
     """
 
+    tag_names = ["axolotl", "relora"]
+
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         self.lr_scheduler = None
@@ -351,12 +457,21 @@ class TrainerBuilderBase(abc.ABC):
 
     _train_dataset = None
     _eval_dataset = None
+    _model_ref = None
 
     def __init__(self, cfg, model, tokenizer):
         self.cfg = cfg
         self.model = model
         self.tokenizer = tokenizer
 
+    @property
+    def model_ref(self):
+        return self._model_ref
+
+    @model_ref.setter
+    def model_ref(self, model):
+        self._model_ref = model
+
     @property
     def train_dataset(self):
         return self._train_dataset
@@ -428,6 +543,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 SaveAxolotlConfigtoWandBCallback(self.cfg.axolotl_config_path)
             )
 
+        if self.cfg.loss_watchdog_threshold is not None:
+            callbacks.append(LossWatchDogCallback(self.cfg))
+
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
@@ -456,14 +574,19 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             return OneCycleLRSchedulerTrainer
         if self.cfg.relora_steps:
             return ReLoRATrainer
+        if self.cfg.model_config_type == "mamba":
+            return AxolotlMambaTrainer
         return AxolotlTrainer
 
     def build(self, total_num_steps):
-        warmup_steps = (
-            self.cfg.warmup_steps
-            if self.cfg.warmup_steps is not None
-            else min(int(0.03 * total_num_steps), 100)
-        )
+        warmup_steps = None
+        if self.cfg.warmup_steps is not None:
+            warmup_steps = self.cfg.warmup_steps
+        elif self.cfg.warmup_ratio is not None:
+            warmup_steps = max(int(self.cfg.warmup_ratio * total_num_steps), 0)
+        else:
+            warmup_steps = min(int(0.03 * total_num_steps), 100)
+
         logging_steps = (
             self.cfg.logging_steps
             if self.cfg.logging_steps is not None
@@ -489,6 +612,14 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             training_arguments_kwargs[
                 "gradient_checkpointing"
             ] = self.cfg.gradient_checkpointing
+            if self.cfg.gradient_checkpointing_kwargs:
+                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:
@@ -516,11 +647,12 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             training_arguments_kwargs["hub_model_id"] = self.cfg.hub_model_id
             training_arguments_kwargs["push_to_hub"] = True
             training_arguments_kwargs["hub_private_repo"] = True
+            training_arguments_kwargs["hub_always_push"] = True
 
             if self.cfg.hub_strategy:
                 training_arguments_kwargs["hub_strategy"] = self.cfg.hub_strategy
 
-        if self.cfg.save_safetensors:
+        if self.cfg.save_safetensors is not None:
             training_arguments_kwargs["save_safetensors"] = self.cfg.save_safetensors
 
         if self.cfg.sample_packing_eff_est:
@@ -528,16 +660,35 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 "sample_packing_efficiency"
             ] = self.cfg.sample_packing_eff_est
 
-        if self.cfg.eval_steps:
+        if self.cfg.dataloader_pin_memory is not None:
+            training_arguments_kwargs[
+                "dataloader_pin_memory"
+            ] = self.cfg.dataloader_pin_memory
+        if self.cfg.dataloader_num_workers is not None:
+            training_arguments_kwargs[
+                "dataloader_num_workers"
+            ] = self.cfg.dataloader_num_workers
+        if self.cfg.dataloader_prefetch_factor is not None:
+            training_arguments_kwargs[
+                "dataloader_prefetch_factor"
+            ] = self.cfg.dataloader_prefetch_factor
+        if self.cfg.dataloader_drop_last is not None:
+            training_arguments_kwargs[
+                "dataloader_drop_last"
+            ] = self.cfg.dataloader_drop_last
+        elif self.cfg.sample_packing and self.cfg.eval_sample_packing is False:
+            training_arguments_kwargs["dataloader_drop_last"] = True
+
+        if self.cfg.val_set_size == 0:
+            # no eval set, so don't eval
+            training_arguments_kwargs["evaluation_strategy"] = "no"
+        elif self.cfg.eval_steps:
             training_arguments_kwargs["evaluation_strategy"] = "steps"
             training_arguments_kwargs["eval_steps"] = self.cfg.eval_steps
         elif self.cfg.evaluation_strategy:
             training_arguments_kwargs[
                 "evaluation_strategy"
             ] = self.cfg.evaluation_strategy
-        elif self.cfg.val_set_size == 0:
-            # no eval set, so don't eval
-            training_arguments_kwargs["evaluation_strategy"] = "no"
         else:
             # we have an eval set, but no steps defined, default to use epoch
             training_arguments_kwargs["evaluation_strategy"] = "epoch"
@@ -623,9 +774,14 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             False if self.cfg.ddp else None
         )
         training_arguments_kwargs["group_by_length"] = self.cfg.group_by_length
-        training_arguments_kwargs["report_to"] = "wandb" if self.cfg.use_wandb else None
+        report_to = None
+        if self.cfg.use_wandb:
+            report_to = "wandb"
+        if self.cfg.use_mlflow:
+            report_to = "mlflow"
+        training_arguments_kwargs["report_to"] = report_to
         training_arguments_kwargs["run_name"] = (
-            self.cfg.wandb_run_id if self.cfg.use_wandb else None
+            self.cfg.wandb_name if self.cfg.use_wandb else None
         )
         training_arguments_kwargs["optim"] = (
             self.cfg.optimizer if self.cfg.optimizer else "adamw_hf"
@@ -636,6 +792,10 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             and self.cfg.lr_scheduler not in ("one_cycle", "log_sweep")
             else "cosine"
         )
+        training_arguments_kwargs["lr_scheduler_kwargs"] = (
+            self.cfg.lr_scheduler_kwargs if self.cfg.lr_scheduler_kwargs else {}
+        )
+        training_arguments_kwargs["cosine_min_lr_ratio"] = self.cfg.cosine_min_lr_ratio
         training_arguments_kwargs["weight_decay"] = (
             self.cfg.weight_decay if self.cfg.weight_decay is not None else 0.0
         )
@@ -643,7 +803,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             self.cfg.sample_packing if self.cfg.sample_packing else False
         )
         training_arguments_kwargs["eval_sample_packing"] = (
-            self.cfg.sample_packing if self.cfg.sample_packing else False
+            self.cfg.sample_packing
+            if self.cfg.eval_sample_packing is not False
+            else False
         )
         training_arguments_kwargs[
             "sample_packing_seq_len_multiplier"
@@ -653,6 +815,14 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         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.neftune_noise_alpha is not None:
+            training_arguments_kwargs[
+                "neftune_noise_alpha"
+            ] = self.cfg.neftune_noise_alpha
+
         training_args = (
             AxolotlTrainingArguments(  # pylint: disable=unexpected-keyword-arg
                 **training_arguments_kwargs,
@@ -666,48 +836,6 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 sys.path.append(self.cfg.torchdistx_path)
                 importlib.import_module("torchdistx")
 
-        if self.cfg.is_llama_derived_model and self.cfg.landmark_attention:
-            from axolotl.monkeypatch.llama_landmark_attn import (
-                add_mem_tokens,
-                get_mem_id,
-                set_model_mem_id,
-            )
-
-            set_model_mem_id(self.model, self.tokenizer)
-
-            LOG.info("Adding landmark attention tokens to dataset")
-
-            for dataset in [self.train_dataset, self.eval_dataset]:
-                dataset = dataset.map(
-                    partial(
-                        add_mem_tokens, mem_freq=50, mem_id=get_mem_id(self.tokenizer)
-                    ),
-                    batched=False,
-                    num_proc=32,
-                )
-
-        trainer_cls = self._get_trainer_cls()
-        trainer_kwargs, trainer_cls = self.hook_pre_create_trainer(
-            trainer_kwargs, trainer_cls
-        )
-        trainer_collator_kwargs = self.build_data_collator()
-
-        trainer = trainer_cls(
-            model=self.model,
-            train_dataset=self.train_dataset,
-            eval_dataset=self.eval_dataset,
-            args=training_args,
-            callbacks=self.get_callbacks(),
-            **trainer_collator_kwargs,
-            **trainer_kwargs,
-        )
-        trainer = self.hook_post_create_trainer(trainer)
-        for callback in self.get_post_trainer_create_callbacks(trainer):
-            trainer.add_callback(callback)
-
-        return trainer
-
-    def build_data_collator(self):
         data_collator_kwargs = {
             "padding": True,  # True/"longest" is the default
         }
@@ -720,27 +848,158 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             # https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html
             data_collator_kwargs["pad_to_multiple_of"] = 64
 
-        collator_kwargs = {}
-        if self.cfg.multimodal:
-            from llava.train.train import DataCollatorForSupervisedDataset
-
-            collator_kwargs["data_collator"] = DataCollatorForSupervisedDataset(
-                tokenizer=self.tokenizer,
-            )
-        else:
-            collator_kwargs["data_collator"] = DataCollatorForSeq2Seq(
+        trainer_cls = self._get_trainer_cls()
+        trainer_kwargs, trainer_cls = self.hook_pre_create_trainer(
+            trainer_kwargs, trainer_cls
+        )
+        trainer = trainer_cls(
+            model=self.model,
+            train_dataset=self.train_dataset,
+            eval_dataset=self.eval_dataset,
+            args=training_args,
+            data_collator=self.build_collator(training_args, **data_collator_kwargs),
+            eval_data_collator=self.build_collator(
+                training_args, is_eval=True, **data_collator_kwargs
+            ),
+            bench_data_collator=transformers.DataCollatorForSeq2Seq(
                 self.tokenizer,
                 return_tensors="pt",
                 **data_collator_kwargs,
-            )
+            ),
+            callbacks=self.get_callbacks(),
+            num_epochs=self.cfg.num_epochs,
+            **trainer_kwargs,
+        )
+        trainer = self.hook_post_create_trainer(trainer)
+        for callback in self.get_post_trainer_create_callbacks(trainer):
+            trainer.add_callback(callback)
 
-        if self.cfg.do_bench_eval:
-            collator_kwargs[
-                "bench_data_collator"
-            ] = transformers.DataCollatorForSeq2Seq(
+        if self.cfg.deepspeed and self.cfg.sample_packing:
+            trainer.accelerator.state.deepspeed_plugin.deepspeed_config[
+                "train_micro_batch_size_per_gpu"
+            ] = self.cfg.micro_batch_size
+
+        return trainer
+
+    def build_collator(
+        self, training_args: AxolotlTrainingArguments, is_eval=False, **kwargs
+    ):
+        if training_args.pretraining:
+            return None
+
+        if self.cfg.model_config_type == "mamba":
+            return MambaDataCollator(tokenizer=self.tokenizer)
+
+        use_batch_sampler_collator = False
+        if is_eval is False and training_args.sample_packing:
+            use_batch_sampler_collator = True
+        if is_eval and training_args.eval_sample_packing:
+            use_batch_sampler_collator = True
+
+        if use_batch_sampler_collator:
+            return BatchSamplerDataCollatorForSeq2Seq(
                 self.tokenizer,
                 return_tensors="pt",
-                **data_collator_kwargs,
+                **kwargs,
             )
 
-        return collator_kwargs
+        return DataCollatorForSeq2Seq(
+            self.tokenizer,
+            return_tensors="pt",
+            **kwargs,
+        )
+
+
+class HFDPOTrainerBuilder(TrainerBuilderBase):
+    """
+    Trainer factory class for DPO Trainer
+    """
+
+    def get_callbacks(self):
+        callbacks = []
+        return callbacks
+
+    def get_post_trainer_create_callbacks(self, trainer):
+        callbacks = []
+        return callbacks
+
+    def build_training_arguments(self, total_num_steps):
+        training_args_kwargs = {}
+        for arg in [
+            "adam_beta1",
+            "adam_beta2",
+            "adam_epsilon",
+            "dataloader_num_workers",
+            "dataloader_pin_memory",
+        ]:
+            if hasattr(self.cfg, arg) and getattr(self.cfg, arg) is not None:
+                training_args_kwargs[arg] = getattr(self.cfg, arg)
+        training_args = TrainingArguments(
+            per_device_train_batch_size=self.cfg.micro_batch_size,
+            max_steps=total_num_steps,
+            remove_unused_columns=False,
+            gradient_accumulation_steps=self.cfg.gradient_accumulation_steps,
+            learning_rate=self.cfg.learning_rate,
+            evaluation_strategy="no",
+            # eval_steps=self.cfg.eval_steps,
+            save_strategy="steps",
+            save_steps=self.cfg.save_steps,
+            output_dir=self.cfg.output_dir,
+            warmup_steps=self.cfg.warmup_steps,
+            bf16=True,
+            gradient_checkpointing=self.cfg.gradient_checkpointing,
+            gradient_checkpointing_kwargs={"use_reentrant": False},
+            logging_first_step=True,
+            logging_steps=1,
+            optim=self.cfg.optimizer,
+            save_total_limit=self.cfg.save_total_limit or 5,
+            **training_args_kwargs,
+        )
+
+        return training_args
+
+    def build(self, total_num_steps):
+        training_args = self.build_training_arguments(total_num_steps)
+        dpo_trainer_kwargs = {}
+        if self.cfg.rl == "ipo":
+            dpo_trainer_kwargs["loss_type"] = "ipo"
+            if self.cfg.dpo_label_smoothing:
+                dpo_trainer_kwargs["label_smoothing"] = self.cfg.dpo_label_smoothing
+        elif self.cfg.rl == "kto_pair":
+            dpo_trainer_kwargs["loss_type"] = "kto_pair"
+
+        dpo_trainer = DPOTrainer(
+            self.model,
+            self.model_ref,
+            args=training_args,
+            beta=self.cfg.dpo_beta or 0.1,
+            train_dataset=self.train_dataset,
+            # eval_dataset=self.eval_dataset,
+            eval_dataset=None,
+            tokenizer=self.tokenizer,
+            max_length=self.cfg.sequence_len,
+            max_target_length=None,
+            max_prompt_length=self.cfg.sequence_len,
+            generate_during_eval=True,
+            **dpo_trainer_kwargs,
+        )
+
+        return dpo_trainer
+
+
+class HFPPOTrainerBuilder(TrainerBuilderBase):
+    """
+    HF Factory class for PPO Trainer
+    """
+
+    def get_callbacks(self):
+        callbacks = []
+        return callbacks
+
+    def get_post_trainer_create_callbacks(self, trainer):
+        callbacks = []
+        return callbacks
+
+    def build(self, total_num_steps):
+        # build PPOConfig
+        pass

src/axolotl/core/trainers/trl.py

@@ -0,0 +1,66 @@
+"""
+module for TRL PPO training
+"""
+import torch
+from tqdm import tqdm
+from trl import PPOTrainer
+
+
+class TRLPPOTrainer(PPOTrainer):
+    """
+    wrapper for ppo trainer to handle customizations
+    """
+
+    def train(
+        self,
+        reward_pipe,
+        resume_from_checkpoint=None,  # pylint: disable=unused-argument
+    ):
+        generation_kwargs = {
+            "min_length": -1,
+            "top_k": 0.0,
+            "top_p": 1.0,
+            "do_sample": True,
+            "pad_token_id": self.tokenizer.eos_token_id,
+            "max_new_tokens": 32,
+        }
+        sent_kwargs = {
+            "return_all_scores": True,
+            "function_to_apply": "none",
+            "batch_size": 16,
+        }
+
+        for epoch, batch in tqdm(  # pylint: disable=unused-variable
+            enumerate(self.dataloader)
+        ):
+            query_tensors = batch["input_ids"]
+
+            # generate model response
+            response_tensors, ref_response_tensors = self.generate(
+                query_tensors,
+                return_prompt=False,
+                generate_ref_response=True,
+                **generation_kwargs
+            )
+            batch["response"] = self.tokenizer.batch_decode(response_tensors)
+            batch["ref_response"] = self.tokenizer.batch_decode(ref_response_tensors)
+
+            # Compute sentiment score
+            texts = [q + r for q, r in zip(batch["query"], batch["response"])]
+            pipe_outputs = reward_pipe(texts, **sent_kwargs)
+            rewards = [torch.tensor(output[1]["score"]) for output in pipe_outputs]
+            ref_texts = [q + r for q, r in zip(batch["query"], batch["ref_response"])]
+            ref_pipe_outputs = reward_pipe(ref_texts, **sent_kwargs)
+            ref_rewards = [
+                torch.tensor(output[1]["score"]) for output in ref_pipe_outputs
+            ]
+            batch["ref_rewards"] = ref_rewards
+
+            # Run PPO step
+            stats = self.step(query_tensors, response_tensors, rewards)
+            self.log_stats(
+                stats,
+                batch,
+                rewards,
+                columns_to_log=["query", "response", "ref_response", "ref_rewards"],
+            )

src/axolotl/datasets.py

@@ -2,7 +2,7 @@
 
 import logging
 import os
-from typing import List
+from typing import List, Optional
 
 import torch
 from datasets import Dataset, IterableDataset
@@ -30,14 +30,20 @@ class TokenizedPromptDataset(Dataset):
         self,
         prompt_tokenizer: PromptTokenizingStrategy,
         dataset: IterableDataset,
+        process_count: Optional[int] = None,
         **kwargs,
     ):
         self.prompt_tokenizer = prompt_tokenizer
+        self.process_count = process_count
         super().__init__(self.process(dataset).data, **kwargs)
 
     def process(self, dataset):
         features = dataset.features.keys()
-        num_proc = min(64, os.cpu_count())
+        num_proc = (
+            min(64, self.process_count)
+            if self.process_count
+            else min(64, os.cpu_count())
+        )
         map_kwargs = {}
         if self.prompt_tokenizer.supports_batched:
             map_kwargs["batched"] = True

src/axolotl/models/llava/llava_mistral.py

@@ -1,167 +0,0 @@
-"""
-LLaVA Mistral classes
-"""
-
-from typing import List, Optional, Tuple, Union
-
-import torch
-from llava.model.llava_arch import LlavaMetaForCausalLM, LlavaMetaModel
-from torch import nn
-from torch.nn import CrossEntropyLoss
-from transformers import (
-    AutoConfig,
-    AutoModelForCausalLM,
-    MistralConfig,
-    MistralForCausalLM,
-    MistralModel,
-)
-from transformers.modeling_outputs import CausalLMOutputWithPast
-
-
-class LlavaMistralConfig(MistralConfig):
-    """
-    HF Transformers Config for Mistral w LLaVA
-    """
-
-    model_type = "llava_mistral"
-
-
-class LlavaMistralModel(LlavaMetaModel, MistralModel):
-    """
-    HF Transformers Model for Mistral w LLaVA
-    """
-
-    config_class = LlavaMistralConfig
-
-    def __init__(
-        self, config: LlavaMistralConfig
-    ):  # pylint: disable=useless-parent-delegation
-        super().__init__(config)
-
-
-class LlavaMistralForCausalLM(MistralForCausalLM, LlavaMetaForCausalLM):
-    """
-    HF Transformers Causal Model for Mistral w LLaVA
-    """
-
-    config_class = LlavaMistralConfig
-
-    def __init__(self, config: LlavaMistralConfig):
-        super().__init__(config)
-        self.model = LlavaMistralModel(config)
-
-        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
-
-        # Initialize weights and apply final processing
-        self.post_init()
-
-    def get_model(self):
-        return self.model
-
-    def forward(
-        self,
-        input_ids: torch.LongTensor = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        past_key_values: Optional[List[torch.FloatTensor]] = None,
-        inputs_embeds: Optional[torch.FloatTensor] = None,
-        labels: Optional[torch.LongTensor] = None,
-        use_cache: Optional[bool] = None,
-        output_attentions: Optional[bool] = None,
-        output_hidden_states: Optional[bool] = None,
-        images: Optional[torch.FloatTensor] = None,
-        return_dict: Optional[bool] = None,
-    ) -> Union[Tuple, CausalLMOutputWithPast]:
-        output_attentions = (
-            output_attentions
-            if output_attentions is not None
-            else self.config.output_attentions
-        )
-        output_hidden_states = (
-            output_hidden_states
-            if output_hidden_states is not None
-            else self.config.output_hidden_states
-        )
-        return_dict = (
-            return_dict if return_dict is not None else self.config.use_return_dict
-        )
-
-        (
-            input_ids,
-            attention_mask,
-            past_key_values,
-            inputs_embeds,
-            labels,
-        ) = self.prepare_inputs_labels_for_multimodal(
-            input_ids, attention_mask, past_key_values, labels, images
-        )
-
-        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
-        outputs = self.model(
-            input_ids=input_ids,
-            attention_mask=attention_mask,
-            past_key_values=past_key_values,
-            inputs_embeds=inputs_embeds,
-            use_cache=use_cache,
-            output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
-            return_dict=return_dict,
-        )
-
-        hidden_states = outputs[0]
-        logits = self.lm_head(hidden_states)
-
-        loss = None
-        if labels is not None:
-            # Shift so that tokens < n predict n
-            shift_logits = logits[..., :-1, :].contiguous()
-            shift_labels = labels[..., 1:].contiguous()
-            # Flatten the tokens
-            loss_fct = CrossEntropyLoss()
-            shift_logits = shift_logits.view(-1, self.config.vocab_size)
-            shift_labels = shift_labels.view(-1)
-            # Enable model/pipeline parallelism
-            shift_labels = shift_labels.to(shift_logits.device)
-            loss = loss_fct(shift_logits, shift_labels)
-
-        if not return_dict:
-            output = (logits,) + outputs[1:]
-            return (loss,) + output if loss is not None else output
-
-        return CausalLMOutputWithPast(
-            loss=loss,
-            logits=logits,
-            past_key_values=outputs.past_key_values,
-            hidden_states=outputs.hidden_states,
-            attentions=outputs.attentions,
-        )
-
-    def prepare_inputs_for_generation(
-        self,
-        input_ids,
-        past_key_values=None,
-        attention_mask=None,
-        inputs_embeds=None,
-        **kwargs
-    ):
-        if past_key_values:
-            input_ids = input_ids[:, -1:]
-
-        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
-        if inputs_embeds is not None and past_key_values is None:
-            model_inputs = {"inputs_embeds": inputs_embeds}
-        else:
-            model_inputs = {"input_ids": input_ids}
-
-        model_inputs.update(
-            {
-                "past_key_values": past_key_values,
-                "use_cache": kwargs.get("use_cache"),
-                "attention_mask": attention_mask,
-                "images": kwargs.get("images", None),
-            }
-        )
-        return model_inputs
-
-
-AutoConfig.register("llava_mistral", LlavaMistralConfig)
-AutoModelForCausalLM.register(LlavaMistralConfig, LlavaMistralForCausalLM)

src/axolotl/models/mamba/__init__.py

@@ -0,0 +1,24 @@
+"""
+Modeling module for Mamba models
+"""
+
+import importlib
+
+
+def check_mamba_ssm_installed():
+    mamba_ssm_spec = importlib.util.find_spec("mamba_ssm")
+    if mamba_ssm_spec is None:
+        raise ImportError(
+            "MambaLMHeadModel requires mamba_ssm. Please install it with `pip install -e .[mamba-ssm]`"
+        )
+
+
+def fix_mamba_attn_for_loss():
+    check_mamba_ssm_installed()
+
+    from mamba_ssm.models import mixer_seq_simple
+
+    from .modeling_mamba import MambaLMHeadModel as MambaLMHeadModelFixed
+
+    mixer_seq_simple.MambaLMHeadModel = MambaLMHeadModelFixed
+    return mixer_seq_simple.MambaLMHeadModel  # pylint: disable=invalid-name

src/axolotl/models/mamba/configuration_mamba.py

@@ -0,0 +1,42 @@
+"""
+HF Transformers MambaConfig
+"""
+from transformers import PretrainedConfig
+
+
+class MambaConfig(PretrainedConfig):
+    """
+    modeling configuration for state space model/mamba
+    """
+
+    model_type = "mamba"
+
+    def __init__(
+        self,
+        vocab_size=50280,
+        d_model=2560,
+        n_layer=64,
+        rms_norm=True,
+        residual_in_fp32=True,
+        fused_add_norm=True,
+        pad_vocab_size_multiple=8,
+        pad_token_id=50277,
+        bos_token_id=0,
+        eos_token_id=0,
+        tie_word_embeddings=False,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.d_model = d_model
+        self.n_layer = n_layer
+        self.rms_norm = rms_norm
+        self.residual_in_fp32 = residual_in_fp32
+        self.fused_add_norm = fused_add_norm
+        self.pad_vocab_size_multiple = pad_vocab_size_multiple
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )

src/axolotl/models/mamba/modeling_mamba.py

@@ -0,0 +1,128 @@
+# pylint: skip-file
+import os
+from collections import namedtuple
+from functools import partial
+from typing import Optional, Union
+
+import torch
+from mamba_ssm.models.mixer_seq_simple import MixerModel, _init_weights
+from mamba_ssm.utils.generation import GenerationMixin
+from mamba_ssm.utils.hf import load_config_hf, load_state_dict_hf
+from torch import nn
+from torch.nn import CrossEntropyLoss
+
+from axolotl.models.mamba.configuration_mamba import MambaConfig
+
+
+class MambaLMHeadModel(nn.Module, GenerationMixin):
+    def __init__(
+        self,
+        d_model: int,
+        n_layer: int,
+        vocab_size: int,
+        initializer_cfg=None,
+        pad_vocab_size_multiple: int = 1,
+        device=None,
+        dtype=None,
+        **backbone_kwargs,
+    ) -> None:
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super().__init__()
+        if vocab_size % pad_vocab_size_multiple != 0:
+            vocab_size += pad_vocab_size_multiple - (
+                vocab_size % pad_vocab_size_multiple
+            )
+        self.config = MambaConfig(
+            vocab_size=vocab_size,
+            d_model=d_model,
+            n_layer=n_layer,
+            pad_vocab_size_multiple=pad_vocab_size_multiple,
+        )
+        self.backbone = MixerModel(
+            d_model=d_model,
+            n_layer=n_layer,
+            vocab_size=vocab_size,
+            initializer_cfg=initializer_cfg,
+            **backbone_kwargs,
+            **factory_kwargs,
+        )
+        self.lm_head = nn.Linear(d_model, vocab_size, bias=False, **factory_kwargs)
+
+        # Initialize weights and apply final processing
+        self.apply(
+            partial(
+                _init_weights,
+                n_layer=n_layer,
+                **(initializer_cfg if initializer_cfg is not None else {}),
+            )
+        )
+        self.tie_weights()
+
+    def tie_weights(self):
+        self.lm_head.weight = self.backbone.embedding.weight
+
+    def allocate_inference_cache(self, batch_size, max_seqlen, dtype=None, **kwargs):
+        return self.backbone.allocate_inference_cache(
+            batch_size, max_seqlen, dtype=dtype, **kwargs
+        )
+
+    def forward(
+        self,
+        input_ids,
+        position_ids=None,
+        inference_params=None,
+        num_last_tokens=0,
+        labels=None,
+        **kwargs,
+    ):
+        """
+        "position_ids" is just to be compatible with Transformer generation. We don't use it.
+        num_last_tokens: if > 0, only return the logits for the last n tokens
+        """
+        hidden_states = self.backbone(input_ids, inference_params=inference_params)
+        if num_last_tokens > 0:
+            hidden_states = hidden_states[:, -num_last_tokens:]
+        lm_logits = self.lm_head(hidden_states)
+
+        CausalLMOutput = namedtuple("CausalLMOutput", ["logits"])
+        return CausalLMOutput(logits=lm_logits)
+
+        loss = None
+        if labels is not None:
+            logits = lm_logits
+            # Shift so that tokens < n predict n
+            shift_logits = logits[..., :-1, :].contiguous()
+            shift_labels = labels[..., 1:].contiguous()
+            # Flatten the tokens
+            loss_fct = CrossEntropyLoss()
+            shift_logits = shift_logits.view(-1, self.config.vocab_size)
+            shift_labels = shift_labels.view(-1)
+            # Enable model parallelism
+            shift_labels = shift_labels.to(shift_logits.device)
+            loss = loss_fct(shift_logits, shift_labels)
+            CausalLMOutput = namedtuple("CausalLMOutput", ["logits", "loss"])
+            print(loss)
+            return CausalLMOutput(logits=lm_logits, loss=loss)
+
+        else:
+            CausalLMOutput = namedtuple("CausalLMOutput", ["logits"])
+            return CausalLMOutput(logits=lm_logits)
+
+    def save_pretrained(
+        self,
+        save_directory: Union[str, os.PathLike],
+        state_dict: Optional[dict] = None,
+        safe_serialization: Optional[bool] = None,  # pylint: disable=unused-argument
+    ):
+        if state_dict is None:
+            state_dict = self.state_dict()
+        torch.save(state_dict, os.path.join(save_directory, "pytorch_model.bin"))
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name, device=None, dtype=None, **kwargs):
+        config = load_config_hf(pretrained_model_name)
+        model = cls(**config, device=device, dtype=dtype, **kwargs)
+        model.load_state_dict(
+            load_state_dict_hf(pretrained_model_name, device={"": device}, dtype=dtype)
+        )
+        return model

src/axolotl/models/phi/__init__.py

@@ -3,4 +3,6 @@ MixFormers model architecture used for phi models
 """
 
 from .configuration_mixformer_sequential import MixFormerSequentialConfig  # noqa
+from .configuration_phi import PhiConfig  # noqa
 from .modeling_mixformer_sequential import MixFormerSequentialForCausalLM  # noqa
+from .modeling_phi import PhiForCausalLM  # noqa

src/axolotl/models/phi/configuration_phi.py

@@ -0,0 +1,65 @@
+# pylint: skip-file
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+import math
+from typing import Optional
+
+from transformers import PretrainedConfig
+
+
+class PhiConfig(PretrainedConfig):
+    """Phi configuration."""
+
+    model_type = "phi"
+    attribute_map = {
+        "max_position_embeddings": "n_positions",
+        "hidden_size": "n_embd",
+        "num_attention_heads": "n_head",
+        "num_hidden_layers": "n_layer",
+    }
+
+    def __init__(
+        self,
+        vocab_size: int = 50304,
+        n_positions: int = 2048,
+        n_embd: int = 1024,
+        n_layer: int = 20,
+        n_inner: Optional[int] = None,
+        n_head: int = 16,
+        n_head_kv: Optional[int] = None,
+        rotary_dim: Optional[int] = 32,
+        activation_function: Optional[str] = "gelu_new",
+        flash_attn: bool = False,
+        flash_rotary: bool = False,
+        fused_dense: bool = False,
+        attn_pdrop: float = 0.0,
+        embd_pdrop: float = 0.0,
+        resid_pdrop: float = 0.0,
+        layer_norm_epsilon: float = 1e-5,
+        initializer_range: float = 0.02,
+        tie_word_embeddings: bool = False,
+        pad_vocab_size_multiple: int = 64,
+        **kwargs
+    ) -> None:
+        self.vocab_size = int(
+            math.ceil(vocab_size / pad_vocab_size_multiple) * pad_vocab_size_multiple
+        )
+        self.n_positions = n_positions
+        self.n_embd = n_embd
+        self.n_layer = n_layer
+        self.n_inner = n_inner
+        self.n_head = n_head
+        self.n_head_kv = n_head_kv
+        self.rotary_dim = min(rotary_dim, n_embd // n_head)
+        self.activation_function = activation_function
+        self.flash_attn = flash_attn
+        self.flash_rotary = flash_rotary
+        self.fused_dense = fused_dense
+        self.attn_pdrop = attn_pdrop
+        self.embd_pdrop = embd_pdrop
+        self.resid_pdrop = resid_pdrop
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.initializer_range = initializer_range
+
+        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)

src/axolotl/monkeypatch/fastchat_conversation_turns.py

@@ -82,15 +82,44 @@ def get_turns(  # pylint: disable=too-many-return-statements
             else:
                 yield role + ":", ""
         return
-    if self.sep_style == SeparatorStyle.LLAMA2:
-        seps = [self.sep, self.sep2]
+    if self.sep_style == SeparatorStyle.LLAMA2 and self.name != "mistral":
         if self.system_message:
+            if self.messages:
+                # For llama, the system message is incorporated into the first human instruction
+                first_role, first_msg = self.messages[0]
+                if first_role == self.roles[0]:
+                    system_prompt += first_msg
+                    self.messages.pop(0)
             yield "", system_prompt
-        else:
-            yield "", "[INST] "
-        for i, (role, message) in enumerate(self.messages[1:]):
+        for i, (role, message) in enumerate(self.messages):
             if message:
-                yield role + " ", message + seps[i % 2]
+                if (i % 2 == 0 and not self.system_message) or (
+                    i % 2 != 0 and self.system_message
+                ):
+                    role = "<s> " + role
+                yield role + " ", message
+            else:
+                yield role, ""
+        return
+    if self.sep_style == SeparatorStyle.LLAMA2 and self.name == "mistral":
+        contains_sys_msg = False
+        if self.system_message:
+            contains_sys_msg = True
+            if self.messages:
+                # There is no clear guidance on how to handle system messages in Mistral so we just prepend it to the first human instruction seperated by a newline
+                first_role, first_msg = self.messages[0]
+                if first_role == self.roles[0]:
+                    system_prompt = self.system_template.format(
+                        system_message=" " + self.system_message
+                    )
+                    system_prompt += first_msg
+                    self.messages.pop(0)
+            yield "", system_prompt
+        for i, (role, message) in enumerate(self.messages):
+            if message and i == 0 and not contains_sys_msg:
+                yield "", system_prompt.strip() + " " + message  # if there is no system message, we need to make sure there is the a `<s> [INST]` at the beginning of the first instruction.
+            elif message:
+                yield role + " ", message
             else:
                 yield role, ""
         return
@@ -118,6 +147,15 @@ def get_turns(  # pylint: disable=too-many-return-statements
             else:
                 yield role + "\n", ""
         return
+    if self.sep_style == SeparatorStyle.CHATGLM3:
+        if self.system_message:
+            yield "", system_prompt
+        for role, message in self.messages:
+            if message:
+                yield role + "\n", " " + message
+            else:
+                yield role
+        return
     if self.sep_style == SeparatorStyle.CHATINTERN:
         # source: https://huggingface.co/internlm/internlm-chat-7b-8k/blob/bd546fa984b4b0b86958f56bf37f94aa75ab8831/modeling_internlm.py#L771
         seps = [self.sep, self.sep2]

src/axolotl/monkeypatch/llama_attn_hijack_flash.py

@@ -321,6 +321,8 @@ def flashattn_forward(
         # only on first autoregressive step q,k,v have same seqlen
         is_causal = key_states.shape == query_states.shape
 
+    dropout_rate = 0.0 if not self.training else getattr(self, "attention_dropout", 0.0)
+
     if cu_seqlens is not None and max_seqlen is not None and cu_seqlens.dim() == 1:
         # special handling using sample packing
         qkv = torch.stack(
@@ -330,7 +332,12 @@ def flashattn_forward(
         qkv = rearrange(qkv, "b s ... -> (b s) ...")
 
         output = flash_attn_varlen_qkvpacked_func(
-            qkv, cu_seqlens, max_seqlen, 0.0, softmax_scale=None, causal=True
+            qkv,
+            cu_seqlens,
+            max_seqlen,
+            dropout_p=dropout_rate,
+            softmax_scale=None,
+            causal=True,
         )
         output = rearrange(output, "(b s) ... -> b s ...", b=bsz)
     elif query_states.shape == key_states.shape:
@@ -353,7 +360,7 @@ def flashattn_forward(
             qkv_unpad,
             cu_seqlens_q,
             max_seqlen_q,
-            0.0,
+            dropout_p=dropout_rate,
             softmax_scale=None,
             causal=is_causal,
         )
@@ -366,6 +373,7 @@ def flashattn_forward(
             output = flash_attn_kvpacked_func(
                 query_states,
                 torch.stack([key_states, value_states], 2),
+                dropout_p=dropout_rate,
                 causal=is_causal,
             )
         else:
@@ -398,7 +406,7 @@ def flashattn_forward(
                 cu_seqlens_k,
                 max_seqlen_q,
                 max_seqlen_k,
-                0.0,
+                dropout_p=dropout_rate,
                 softmax_scale=None,
                 causal=is_causal,
             )

src/axolotl/monkeypatch/llama_attn_hijack_sdp.py

@@ -25,6 +25,8 @@ def sdp_attention_forward(
     past_key_value: Optional[Tuple[torch.Tensor]] = None,
     output_attentions: bool = False,
     use_cache: bool = False,
+    padding_mask: Optional[torch.LongTensor] = None,  # pylint: disable=unused-argument
+    **kwargs,  # pylint: disable=unused-argument
 ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
     # pylint: disable=duplicate-code
     bsz, q_len, _ = hidden_states.size()

src/axolotl/monkeypatch/llama_attn_hijack_xformers.py

@@ -29,6 +29,8 @@ def xformers_forward(
     past_key_value: Optional[Tuple[torch.Tensor]] = None,
     output_attentions: bool = False,
     use_cache: bool = False,
+    padding_mask: Optional[torch.LongTensor] = None,  # pylint: disable=unused-argument
+    **kwargs,  # pylint: disable=unused-argument
 ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
     # pylint: disable=duplicate-code
     bsz, q_len, _ = hidden_states.size()

src/axolotl/monkeypatch/llama_embeddings_hijack.py

@@ -1,40 +0,0 @@
-"""
-patch to add noisy embeddings per https://arxiv.org/abs/2310.05914
-"""
-
-import torch
-import transformers.models.llama.modeling_llama
-from transformers.utils import logging
-
-logger = logging.get_logger(__name__)
-
-
-def replace_llama_embeddings_with_uniform_distribution(noise_alpha=5):
-    # pylint: disable=duplicate-code
-    def noised_embed(orig_embed, noise_alpha, model):
-        def new_func(input_ids):
-            # during training, we add noise to the embedding
-            # during generation, we don't add noise to the embedding
-            if model.training:
-                embed_init = orig_embed(input_ids)
-                dims = torch.tensor(embed_init.size(1) * embed_init.size(2))
-                mag_norm = noise_alpha / torch.sqrt(dims)
-                return embed_init + torch.zeros_like(embed_init).uniform_(
-                    -mag_norm, mag_norm
-                )
-            return orig_embed(input_ids)
-
-        return new_func
-
-    def post_init(orig_post_init):
-        def new_func(self):
-            orig_post_init(self)
-            self.embed_tokens.forward = noised_embed(
-                self.embed_tokens.forward, noise_alpha, self
-            )
-
-        return new_func
-
-    transformers.models.llama.modeling_llama.LlamaModel.post_init = post_init(
-        transformers.models.llama.modeling_llama.LlamaModel.post_init
-    )

src/axolotl/monkeypatch/mistral_attn_hijack_flash.py

@@ -201,6 +201,8 @@ def flashattn_forward(
         # only on first autoregressive step q,k,v have same seqlen
         is_causal = key_states.shape == query_states.shape
 
+    dropout_rate = 0.0 if not self.training else getattr(self, "attention_dropout", 0.0)
+
     if cu_seqlens is not None and max_seqlen is not None and cu_seqlens.dim() == 1:
         # special handling using sample packing
         qkv = torch.stack(
@@ -213,7 +215,7 @@ def flashattn_forward(
             qkv,
             cu_seqlens,
             max_seqlen,
-            0.0,
+            dropout_p=dropout_rate,
             softmax_scale=None,
             causal=True,
             window_size=window_size,
@@ -239,7 +241,7 @@ def flashattn_forward(
             qkv_unpad,
             cu_seqlens_q,
             max_seqlen_q,
-            0.0,
+            dropout_p=dropout_rate,
             softmax_scale=None,
             causal=is_causal,
             window_size=window_size,
@@ -253,6 +255,7 @@ def flashattn_forward(
             output = flash_attn_kvpacked_func(
                 query_states,
                 torch.stack([key_states, value_states], 2),
+                dropout_p=dropout_rate,
                 causal=is_causal,
                 window_size=window_size,
             )
@@ -286,7 +289,7 @@ def flashattn_forward(
                 cu_seqlens_k,
                 max_seqlen_q,
                 max_seqlen_k,
-                0.0,
+                dropout_p=dropout_rate,
                 softmax_scale=None,
                 causal=is_causal,
                 window_size=window_size,

src/axolotl/monkeypatch/mistral_embeddings_hijack.py

@@ -1,40 +0,0 @@
-"""
-patch to add noisy embeddings per https://arxiv.org/abs/2310.05914
-"""
-
-import torch
-import transformers.models.mistral.modeling_mistral
-from transformers.utils import logging
-
-logger = logging.get_logger(__name__)
-
-
-def replace_mistral_embeddings_with_uniform_distribution(noise_alpha=5):
-    # pylint: disable=duplicate-code
-    def noised_embed(orig_embed, noise_alpha, model):
-        def new_func(input_ids):
-            # during training, we add noise to the embedding
-            # during generation, we don't add noise to the embedding
-            if model.training:
-                embed_init = orig_embed(input_ids)
-                dims = torch.tensor(embed_init.size(1) * embed_init.size(2))
-                mag_norm = noise_alpha / torch.sqrt(dims)
-                return embed_init + torch.zeros_like(embed_init).uniform_(
-                    -mag_norm, mag_norm
-                )
-            return orig_embed(input_ids)
-
-        return new_func
-
-    def post_init(orig_post_init):
-        def new_func(self):
-            orig_post_init(self)
-            self.embed_tokens.forward = noised_embed(
-                self.embed_tokens.forward, noise_alpha, self
-            )
-
-        return new_func
-
-    transformers.models.mistral.modeling_mistral.MistralModel.post_init = post_init(
-        transformers.models.mistral.modeling_mistral.MistralModel.post_init
-    )

src/axolotl/monkeypatch/mixtral/__init__.py

@@ -0,0 +1,22 @@
+"""
+Patches to support multipack for mixtral
+"""
+import transformers
+
+
+def replace_mixtral_attn_with_multipack_flash_attn():
+    from .modeling_mixtral import (
+        MixtralMultipackFlashAttention2,
+        mixtral_decoder_layer_forward,
+        mixtral_model_forward,
+    )
+
+    transformers.models.mixtral.modeling_mixtral.MixtralDecoderLayer.forward = (
+        mixtral_decoder_layer_forward
+    )
+    transformers.models.mixtral.modeling_mixtral.MixtralModel.forward = (
+        mixtral_model_forward
+    )
+    transformers.models.mixtral.modeling_mixtral.MIXTRAL_ATTENTION_CLASSES[
+        "flash_attention_2"
+    ] = MixtralMultipackFlashAttention2

src/axolotl/monkeypatch/mixtral/modeling_mixtral.py

@@ -0,0 +1,383 @@
+"""
+Mixtral modeling for multipack
+"""
+# pylint: disable=missing-module-docstring,unused-argument,protected-access,pointless-string-statement,duplicate-code
+import logging
+import warnings
+from typing import List, Optional, Tuple, Union
+
+import torch
+from einops import rearrange
+from flash_attn import flash_attn_varlen_qkvpacked_func
+from transformers import Cache, DynamicCache
+from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
+from transformers.modeling_outputs import MoeModelOutputWithPast
+from transformers.models.mixtral.modeling_mixtral import (
+    MixtralFlashAttention2,
+    apply_rotary_pos_emb,
+    repeat_kv,
+)
+
+from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
+
+LOG = logging.getLogger("axolotl.monkeypatch.mixtral")
+
+
+class MixtralMultipackFlashAttention2(MixtralFlashAttention2):
+    """
+    Custom multipack implementation w flash attention 2
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._flash_attn_uses_top_left_mask = True
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        cu_seqlens: Optional[torch.Tensor] = None,
+        max_seqlen: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        if "padding_mask" in kwargs:
+            warnings.warn(
+                "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+            )
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(
+            bsz, q_len, self.num_heads, self.head_dim
+        ).transpose(1, 2)
+        key_states = key_states.view(
+            bsz, q_len, self.num_key_value_heads, self.head_dim
+        ).transpose(1, 2)
+        value_states = value_states.view(
+            bsz, q_len, self.num_key_value_heads, self.head_dim
+        ).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            if self.layer_idx is None:
+                raise ValueError(
+                    f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
+                    "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
+                    "with a layer index."
+                )
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        query_states, key_states = apply_rotary_pos_emb(
+            query_states, key_states, cos, sin, position_ids
+        )
+
+        if past_key_value is not None:
+            cache_kwargs = {"sin": sin, "cos": cos}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(
+                key_states, value_states, self.layer_idx, cache_kwargs
+            )
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        if cu_seqlens is not None and max_seqlen is not None and cu_seqlens.dim() == 1:
+            # special handling using sample packing
+            qkv = torch.stack(
+                [query_states, key_states, value_states], dim=2
+            )  # [bsz, nh, 3, q_len, hd]
+            qkv = qkv.transpose(1, 3)  # [bsz, q_len, 3, nh, hd]
+            qkv = rearrange(qkv, "b s ... -> (b s) ...")
+
+            attn_output = flash_attn_varlen_qkvpacked_func(
+                qkv,
+                cu_seqlens,
+                max_seqlen,
+                dropout_p=self.attention_dropout,
+                softmax_scale=None,
+                causal=True,
+            )
+            attn_output = rearrange(attn_output, "(b s) ... -> b s ...", b=bsz)
+
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+
+def mixtral_decoder_layer_forward(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    output_attentions: Optional[bool] = False,
+    output_router_logits: Optional[bool] = False,
+    use_cache: Optional[bool] = False,
+    cu_seqlens: Optional[torch.Tensor] = None,
+    max_seqlen: Optional[torch.Tensor] = None,
+    **kwargs,
+) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+    if "padding_mask" in kwargs:
+        warnings.warn(
+            "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
+        )
+    """
+    Args:
+        hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+        attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+            `(batch, sequence_length)` where padding elements are indicated by 0.
+        past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+            returned tensors for more detail.
+        output_router_logits (`bool`, *optional*):
+            Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
+            should not be returned during inference.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+            (see `past_key_values`).
+    """
+
+    residual = hidden_states
+
+    hidden_states = self.input_layernorm(hidden_states)
+
+    # Self Attention
+    hidden_states, self_attn_weights, present_key_value = self.self_attn(
+        hidden_states=hidden_states,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_value=past_key_value,
+        output_attentions=output_attentions,
+        use_cache=use_cache,
+        cu_seqlens=cu_seqlens,
+        max_seqlen=max_seqlen,
+    )
+    hidden_states = residual + hidden_states
+
+    # Fully Connected
+    residual = hidden_states
+    hidden_states = self.post_attention_layernorm(hidden_states)
+    hidden_states, router_logits = self.block_sparse_moe(hidden_states)
+    hidden_states = residual + hidden_states
+
+    outputs = (hidden_states,)
+
+    if output_attentions:
+        outputs += (self_attn_weights,)
+
+    if use_cache:
+        outputs += (present_key_value,)
+
+    if output_router_logits:
+        outputs += (router_logits,)
+
+    return outputs
+
+
+def mixtral_model_forward(
+    self,
+    input_ids: torch.LongTensor = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[List[torch.FloatTensor]] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    output_router_logits: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+) -> Union[Tuple, MoeModelOutputWithPast]:
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_router_logits = (
+        output_router_logits
+        if output_router_logits is not None
+        else self.config.output_router_logits
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+
+    # retrieve input_ids and inputs_embeds
+    if input_ids is not None and inputs_embeds is not None:
+        raise ValueError(
+            "You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time"
+        )
+    if input_ids is not None:
+        batch_size, seq_length = input_ids.shape
+    elif inputs_embeds is not None:
+        batch_size, seq_length, _ = inputs_embeds.shape
+    else:
+        raise ValueError(
+            "You have to specify either decoder_input_ids or decoder_inputs_embeds"
+        )
+
+    past_key_values_length = 0
+
+    if use_cache:
+        use_legacy_cache = not isinstance(past_key_values, Cache)
+        if use_legacy_cache:
+            past_key_values = DynamicCache.from_legacy_cache(past_key_values)
+        past_key_values_length = past_key_values.get_usable_length(seq_length)
+
+    cu_seqlens = None
+    max_seqlen = None
+    if position_ids is None:
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+        position_ids = torch.arange(
+            past_key_values_length,
+            seq_length + past_key_values_length,
+            dtype=torch.long,
+            device=device,
+        )
+        position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
+    else:
+        position_ids = position_ids.view(-1, seq_length).long()
+        cu_seqlens, max_seqlen = get_cu_seqlens_from_pos_ids(position_ids)
+        cu_seqlens = cu_seqlens.squeeze()
+
+    if inputs_embeds is None:
+        inputs_embeds = self.embed_tokens(input_ids)
+
+    if (
+        attention_mask is not None
+        and self._attn_implementation == "flash_attention_2"
+        and use_cache
+    ):
+        is_padding_right = attention_mask[:, -1].sum().item() != batch_size
+        if is_padding_right:
+            raise ValueError(
+                "You are attempting to perform batched generation with padding_side='right'"
+                " this may lead to unexpected behaviour for Flash Attention version of Mixtral. Make sure to "
+                " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
+            )
+
+    if self._attn_implementation == "flash_attention_2":
+        # 2d mask is passed through the layers
+        attention_mask = (
+            attention_mask
+            if (attention_mask is not None and 0 in attention_mask)
+            else None
+        )
+    else:
+        # 4d mask is passed through the layers
+        attention_mask = _prepare_4d_causal_attention_mask(
+            attention_mask,
+            (batch_size, seq_length),
+            inputs_embeds,
+            past_key_values_length,
+            sliding_window=self.config.sliding_window,
+        )
+
+    hidden_states = inputs_embeds
+
+    if self.gradient_checkpointing and self.training:
+        if use_cache:
+            LOG.warning_once(
+                "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+            )
+            use_cache = False
+
+    # decoder layers
+    all_hidden_states = () if output_hidden_states else None
+    all_self_attns = () if output_attentions else None
+    all_router_logits = () if output_router_logits else None
+    next_decoder_cache = None
+
+    for decoder_layer in self.layers:
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        if self.gradient_checkpointing and self.training:
+            layer_outputs = self._gradient_checkpointing_func(
+                decoder_layer.__call__,
+                hidden_states,
+                attention_mask,
+                position_ids,
+                past_key_values,
+                output_attentions,
+                output_router_logits,
+                use_cache,
+                cu_seqlens,
+                max_seqlen,
+            )
+        else:
+            layer_outputs = decoder_layer(
+                hidden_states,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_value=past_key_values,
+                output_attentions=output_attentions,
+                output_router_logits=output_router_logits,
+                use_cache=use_cache,
+                cu_seqlens=cu_seqlens,
+                max_seqlen=max_seqlen,
+            )
+
+        hidden_states = layer_outputs[0]
+
+        if use_cache:
+            next_decoder_cache = layer_outputs[2 if output_attentions else 1]
+
+        if output_attentions:
+            all_self_attns += (layer_outputs[1],)
+
+        if output_router_logits:
+            all_router_logits += (layer_outputs[-1],)
+
+    hidden_states = self.norm(hidden_states)
+
+    # add hidden states from the last decoder layer
+    if output_hidden_states:
+        all_hidden_states += (hidden_states,)
+
+    next_cache = None
+    if use_cache:
+        next_cache = (
+            next_decoder_cache.to_legacy_cache()
+            if use_legacy_cache
+            else next_decoder_cache
+        )
+
+    if not return_dict:
+        return tuple(
+            v
+            for v in [
+                hidden_states,
+                next_cache,
+                all_hidden_states,
+                all_self_attns,
+                all_router_logits,
+            ]
+            if v is not None
+        )
+
+    return MoeModelOutputWithPast(
+        last_hidden_state=hidden_states,
+        past_key_values=next_cache,
+        hidden_states=all_hidden_states,
+        attentions=all_self_attns,
+        router_logits=all_router_logits,
+    )

src/axolotl/monkeypatch/utils.py

@@ -55,6 +55,7 @@ def get_cu_seqlens(attn_mask):
     return torch.stack(results).to(dtype=torch.int32), torch.stack(max_seq_lens)
 
 
+@torch.jit.script
 def get_cu_seqlens_from_pos_ids(position_ids):
     """generate a cumulative sequence length mask for flash attention using pos ids"""
     if len(position_ids.shape) == 1:
@@ -81,7 +82,7 @@ def get_cu_seqlens_from_pos_ids(position_ids):
         # Get the indices where the sequence starts
         start_indices = torch.cat(
             [
-                (seq_starts).nonzero(as_tuple=True)[0],
+                torch.nonzero(seq_starts).unbind(dim=1)[0],
                 torch.tensor([len(adjusted_row)], dtype=torch.int32, device=device),
             ]
         )

src/axolotl/monkeypatch/xpos_rope_llama_monkey_patch.py

@@ -1,94 +0,0 @@
-# pylint: skip-file
-"""
-Copied from https://github.com/kaiokendev/cutoff-len-is-context-len/blob/main/util/xpos_rope_llama_monkey_patch.py
-"""
-import torch
-import transformers
-import transformers.models.llama.modeling_llama
-from einops import rearrange
-
-
-class XposRotaryEmbedding(torch.nn.Module):
-    def __init__(
-        self,
-        dim,
-        max_position_embeddings=2048,
-        base=10000,
-        device=None,
-        scale_base=2048,
-        use_xpos=True,
-    ):
-        super().__init__()
-        self.max_seq_len_cached = max_position_embeddings
-        self.scale_base = scale_base
-
-        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
-        t = torch.arange(self.max_seq_len_cached, device=device).type_as(inv_freq)
-        freqs = torch.einsum("i , j -> i j", t, inv_freq)
-        freqs = torch.cat((freqs, freqs), dim=-1)
-
-        self.register_buffer("inv_freq", inv_freq, persistent=False)
-        self.register_buffer("freqs_cached", freqs, persistent=False)
-
-        if not use_xpos:
-            self.register_buffer("scale", None)
-            self.register_buffer("scale_cached", torch.ones(1))
-            return
-
-        scale = (torch.arange(0, dim, 2) + 0.4 * dim) / (1.4 * dim)
-        power = (t - (self.max_seq_len_cached // 2)) / self.scale_base
-        scale_cached = scale ** rearrange(power, "n -> n 1")
-        scale_cached = torch.cat((scale_cached, scale_cached), dim=-1)
-
-        self.register_buffer("scale", scale, persistent=False)
-        self.register_buffer("scale_cached", scale_cached, persistent=False)
-
-    def forward(
-        self,
-        x,
-        seq_len,
-    ):
-        if seq_len > self.max_seq_len_cached:
-            self.max_seq_len_cached = seq_len
-            t = torch.arange(self.max_seq_len_cached, device=x.device).type_as(
-                self.inv_freq
-            )
-            freqs = torch.einsum("i , j -> i j", t, self.inv_freq)
-            freqs = torch.cat((freqs, freqs), dim=-1).to(dtype=x.dtype)
-
-            self.register_buffer("freqs_cached", freqs)
-
-            if self.scale is None:
-                self.register_buffer(
-                    "scale_cached", torch.ones(1, device=x.device).to(dtype=x.dtype)
-                )
-
-                return self.freqs_cached.to(dtype=x.dtype), self.scale_cached
-
-            power = (t - (seq_len // 2)) / self.scale_base
-            scale = self.scale ** rearrange(power, "n -> n 1")
-            scale = torch.cat((scale, scale), dim=-1).to(dtype=x.dtype)
-            self.register_buffer("scale_cached", scale)
-
-        return self.freqs_cached.to(dtype=x.dtype), self.scale_cached.to(dtype=x.dtype)
-
-
-def rotate_half(x):
-    x1, x2 = x.chunk(2, dim=-1)
-    return torch.cat((-x2, x1), dim=-1)
-
-
-def apply_rotary_pos_emb(q, k, freqs, scale=1, position_ids=None):
-    freqs = freqs[position_ids, :]
-    if scale.shape[-1] != 1:
-        scale = scale[position_ids, :]
-
-    q_embed = (q * freqs.cos() * scale) + (rotate_half(q) * freqs.sin() * scale)
-    k_embed = (k * freqs.cos() * 1 / scale) + (rotate_half(k) * freqs.sin() * 1 / scale)
-
-    return q_embed, k_embed
-
-
-def replace_llama_rope_with_xpos_rope():
-    transformers.models.llama.modeling_llama.LlamaRotaryEmbedding = XposRotaryEmbedding
-    transformers.models.llama.modeling_llama.apply_rotary_pos_emb = apply_rotary_pos_emb