Fix(debug): Use space delimiter for debug_text_only also

* Update tests-docker.yml

* Update tests-docker.yml

* run ci tests on ci yaml updates

---------

Co-authored-by: Wing Lian <wing.lian@gmail.com>

.github/ISSUE_TEMPLATE/bug-report.yaml

@@ -53,6 +53,13 @@ body:
     validations:
       required: true
 
+  - type: textarea
+    id: config
+    attributes:
+      label: Config yaml
+      description: |
+        Please attach the config yaml!
+
   - type: textarea
     id: possible-solution
     attributes:

.github/workflows/base.yml

@@ -25,6 +25,16 @@ jobs:
             python_version: "3.10"
             pytorch: 2.0.1
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 9.0+PTX"
+          - cuda: "118"
+            cuda_version: 11.8.0
+            python_version: "3.10"
+            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
         uses: actions/checkout@v3

.github/workflows/main.yml

@@ -23,33 +23,60 @@ jobs:
             python_version: "3.10"
             pytorch: 2.0.1
             axolotl_extras:
+            is_latest: true
+          - cuda: 118
+            cuda_version: 11.8.0
+            python_version: "3.10"
+            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 }}
+          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'
@@ -68,24 +95,34 @@ jobs:
             pytorch: 2.0.1
             axolotl_extras:
             is_latest: true
+          - cuda: 118
+            cuda_version: 11.8.0
+            python_version: "3.10"
+            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/tests-docker.yml

@@ -0,0 +1,46 @@
+name: e2e-docker-tests
+
+on:
+  pull_request:
+    paths:
+      - '**.py'
+      - 'requirements.txt'
+      - '.github/workflows/*.yml'
+  workflow_dispatch:
+
+jobs:
+  build-axolotl:
+    if: github.repository_owner == 'OpenAccess-AI-Collective'
+    # this job needs to be run on self-hosted GPU runners...
+    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
+    runs-on: [self-hosted, gpu, docker]
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Build Docker image
+        run: |
+          # Set up build arguments
+          BASE_TAG="main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}"
+          CUDA="${{ matrix.cuda }}"
+          PYTORCH_VERSION="${{ matrix.pytorch }}"
+          # Build the Docker image
+          docker build . \
+            --file ./docker/Dockerfile \
+            --build-arg BASE_TAG=$BASE_TAG \
+            --build-arg CUDA=$CUDA \
+            --build-arg PYTORCH_VERSION=$PYTORCH_VERSION \
+            --tag test-axolotl
+      - name: Unit Tests w docker image
+        run: |
+          docker run --rm test-axolotl pytest --ignore=tests/e2e/ /workspace/axolotl/tests/

.github/workflows/tests.yml

@@ -6,9 +6,11 @@ on:
       - "main"
     paths:
       - '**.py'
+      - 'requirements.txt'
   pull_request:
       paths:
        - '**.py'
+       - 'requirements.txt'
   workflow_dispatch:
 
 jobs:
@@ -44,7 +46,7 @@ jobs:
 
       - name: Install dependencies
         run: |
-          pip3 install -e .
+          pip3 install -U -e .
           pip3 install -r requirements-tests.txt
 
       - name: Run tests
@@ -69,8 +71,9 @@ jobs:
 
       - name: Install dependencies
         run: |
-          pip3 install -e .
-          pip3 install flash-attn
+          pip3 install --extra-index-url https://download.pytorch.org/whl/cu118 -U torch==2.0.1
+          pip3 uninstall -y transformers accelerate
+          pip3 install -U -e .[flash-attn,mamba-ssm]
           pip3 install -r requirements-tests.txt
 
       - name: Run e2e tests

.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
 

README.md

@@ -23,18 +23,22 @@ Features:
 - [Supported Features](#axolotl-supports)
 - [Quickstart](#quickstart-)
 - [Installation](#installation)
-  - [Docker Installation](#environment)
-  - [Conda/Pip venv Installation](#condapip-venv)
-  - [LambdaLabs Installation](#lambdalabs)
+  - [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)
 - [Need Help?](#need-help-)
 - [Badge](#badge-)
 - [Community Showcase](#community-showcase)
@@ -50,7 +54,7 @@ Features:
       <b>Axolotl provides a unified repository for fine-tuning <br />a variety of AI models with ease</b>
     </p>
     <p>
-      Go ahead and axolotl questions!!
+      Go ahead and Axolotl questions!!
     </p>
     <img src="https://github.com/OpenAccess-AI-Collective/axolotl/actions/workflows/pre-commit.yml/badge.svg?branch=main" alt="pre-commit">
     <img alt="PyTest Status" src="https://github.com/OpenAccess-AI-Collective/axolotl/actions/workflows/tests.yml/badge.svg?branch=main">
@@ -63,17 +67,21 @@ 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 ⚡
@@ -82,31 +90,39 @@ 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
 
 ### Environment
 
-- Docker
+#### Docker
   ```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:
 
@@ -114,18 +130,48 @@ accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
   docker compose up -d
   ```
 
-- Conda/Pip venv
+  <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=volume,src=axolotl,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**
 
   2. Install pytorch stable https://pytorch.org/get-started/locally/
 
-  3. Install axolotl along with python dependencies
+  3. Install Axolotl along with python dependencies
         ```bash
         pip3 install packaging
         pip3 install -e '.[flash-attn,deepspeed]'
         ```
+  4. (Optional) Login to Huggingface to use gated models/datasets.
+        ```bash
+        huggingface-cli login
+        ```
+        Get the token at huggingface.co/settings/tokens
 
-- LambdaLabs
+#### Runpod
+
+Use `winglian/axolotl-runpod:main-latest` or use this [direct link](https://runpod.io/gsc?template=v2ickqhz9s&ref=6i7fkpdz)
+
+#### LambdaLabs
   <details>
 
   <summary>Click to Expand</summary>
@@ -169,7 +215,30 @@ accelerate launch -m axolotl.cli.inference examples/openllama-3b/lora.yml \
   ```
   </details>
 
-- Windows: Please use WSL or Docker!
+#### 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
 
@@ -180,10 +249,17 @@ Have dataset(s) in one of the following format (JSONL recommended):
   ```json
   {"instruction": "...", "input": "...", "output": "..."}
   ```
-- `sharegpt:chat`: 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": "..."}
@@ -245,6 +321,10 @@ Have dataset(s) in one of the following format (JSONL recommended):
   ```json
   {"article": "...", "question": "...", "answer": "..."}
   ```
+- `context_qa.load_v2`: in context question answering (alternate)
+  ```json
+  {"context": "...", "question": "...", "answer": "..."}
+  ```
 - `context_qa.load_404`: in context question answering from an article, with default response for no answer from context
   ```json
   {"article": "...", "unanswerable_question": "..."}
@@ -269,11 +349,11 @@ Have dataset(s) in one of the following format (JSONL recommended):
   ```json
   {"prompt": "...", "generation": "..."}
   ```
-- `sharegpt_simple.load_role`: conversations where `role` is used instead of `from`
+- `sharegpt.load_role`: conversations where `role` is used instead of `from`
   ```json
   {"conversations": [{"role": "...", "value": "..."}]}
   ```
-- `sharegpt_simple.load_guanaco`: conversations where `from` is `prompter`/`assistant` instead of default sharegpt
+- `sharegpt.load_guanaco`: conversations where `from` is `prompter`/`assistant` instead of default sharegpt
   ```json
   {"conversations": [{"from": "...", "value": "..."}]}
   ```
@@ -286,29 +366,28 @@ Have dataset(s) in one of the following format (JSONL recommended):
 
 #### How to add custom prompts
 
-Using yaml. Example:
+For a dataset that is preprocessed for instruction purposes:
+
+```json
+{"instruction": "...", "output": "..."}
+```
+
+You can use this example in your YAML config:
+
 ```yaml
 datasets:
   - path: repo
     type:
       system_prompt: ""
-      no_input_format: |-
-        User: {instruction}<|end_of_turn|>
-        Assistant:
-      format: |-
-        User: {instruction}
-        {input}<|end_of_turn|>
-        Assistant:
+      field_system: system
+      format: "[INST] {instruction} [/INST]"
+      no_input_format: "[INST] {instruction} [/INST]"
 ```
 
-Using file:
-1. Add your method to a file in [prompt_strategies](src/axolotl/prompt_strategies). Please see other files as example.
-2. Use your custom file name as the dataset type `<prompt_strategies_file>.load_<load_fn>`.
-
 #### How to use your custom pretokenized dataset
 
 - Do not pass a `type:`
-- Dataset must contain `input_ids`, `attention_mask`, `labels` in columns
+- Columns in Dataset must be exactly `input_ids`, `attention_mask`, `labels`
 
 
 ### Config
@@ -346,11 +425,30 @@ 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
       ds_type: json # see other options below
       type: alpaca
+
+  # dataset with splits, but no train split
+  dataset:
+    - 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
@@ -378,18 +476,18 @@ 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
-# this can also be a relative path to a model on disk
+# This is the huggingface model that contains *.pt, *.safetensors, or *.bin files
+# This can also be a relative path to a model on disk
 base_model: ./llama-7b-hf
-# you can specify an ignore pattern if the model repo contains more than 1 model type (*.pt, etc)
+# You can specify an ignore pattern if the model repo contains more than 1 model type (*.pt, etc)
 base_model_ignore_patterns:
-# if the base_model repo on hf hub doesn't include configuration .json files,
-# you can set that here, or leave this empty to default to base_model
+# If the base_model repo on hf hub doesn't include configuration .json files,
+# You can set that here, or leave this empty to default to base_model
 base_model_config: ./llama-7b-hf
-# you can specify to choose a specific model revision from huggingface hub
+# You can specify to choose a specific model revision from huggingface hub
 model_revision:
 # Optional tokenizer configuration override in case you want to use a different tokenizer
 # than the one defined in the base model
@@ -404,18 +502,41 @@ trust_remote_code:
 tokenizer_use_fast:
 # Whether to use the legacy tokenizer setting, defaults to True
 tokenizer_legacy:
-# resize the model embeddings when new tokens are added to multiples of 32
-# this is reported to improve training speed on some models
+# Resize the model embeddings when new tokens are added to multiples of 32
+# This is reported to improve training speed on some models
 resize_token_embeddings_to_32x:
 
-# whether you are training a 4-bit GPTQ quantized model
+# Used to identify which the model is based on
+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
 gptq_groupsize: 128 # group size
 gptq_model_v1: false # v1 or v2
 
-# this will attempt to quantize the model down to 8 bits and use adam 8 bit optimizer
+# This will attempt to quantize the model down to 8 bits and use adam 8 bit optimizer
 load_in_8bit: true
-# use bitsandbytes 4 bit
+# Use bitsandbytes 4 bit
 load_in_4bit:
 
 # Use CUDA bf16
@@ -429,9 +550,14 @@ tf32: true # require >=ampere
 bfloat16: true # require >=ampere
 float16: true
 
-# a list of one or more datasets to finetune the model with
+# 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:
-  # hf 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>
@@ -439,18 +565,25 @@ 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
+    train_on_split: train # Optional[str] name of dataset split to load from
 
-  # custom user prompt
+    # 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
     type:
-      # the below are defaults. only set what's needed.
+      # The below are defaults. only set what's needed.
       system_prompt: ""
+      system_format: "{system}"
       field_system: system
       field_instruction: instruction
-      field_output: input
+      field_input: input
+      field_output: output
 
-      # customizable to be single line or multi-line
-      system_format: "{system}"
+      # Customizable to be single line or multi-line
       # 'format' can include {input}
       format: |-
         User: {instruction} {input}
@@ -458,21 +591,27 @@ datasets:
       # 'no_input_format' cannot include {input}
       no_input_format: "{instruction} "
 
-      # for completions datsets, uses the provided field if not `text`
+      # For `completion` datsets only, uses the provided field instead of `text` column
       field:
 
-# axolotl attempts to save the dataset as an arrow after packing the data together so
+# 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
-# push prepared dataset to hub
+# Push prepared dataset to hub
 push_dataset_to_hub: # repo path
+# The maximum number of processes to use while preprocessing your input dataset. This defaults to `os.cpu_count()`
+# if not set.
+dataset_processes: # defaults to os.cpu_count() if not set
 # push checkpoints to hub
 hub_model_id: # repo path to push finetuned model
 # how to push checkpoints to hub
 # https://huggingface.co/docs/transformers/v4.31.0/en/main_classes/trainer#transformers.TrainingArguments.hub_strategy
 hub_strategy:
-# whether to use hf `use_auth_token` for loading datasets. Useful for fetching private datasets
-# required to be true when used in combination with `push_dataset_to_hub`
+# Whether to use hf `use_auth_token` for loading datasets. Useful for fetching private datasets
+# Required to be true when used in combination with `push_dataset_to_hub`
 hf_use_auth_token: # boolean
 # How much of the dataset to set aside as evaluation. 1 = 100%, 0.50 = 50%, etc. 0 for no eval.
 val_set_size: 0.04
@@ -481,30 +620,41 @@ dataset_shard_num:
 # Index of shard to use for whole dataset
 dataset_shard_idx:
 
-# the maximum length of an input to train with, this should typically be less than 2048
+# The maximum length of an input to train with, this should typically be less than 2048
 # as most models have a token/context limit of 2048
 sequence_len: 2048
-# pad inputs so each step uses constant sized buffers
-# this will reduce memory fragmentation and may prevent OOMs, by re-using memory more efficiently
+# Pad inputs so each step uses constant sized buffers
+# This will reduce memory fragmentation and may prevent OOMs, by re-using memory more efficiently
 pad_to_sequence_len:
-# max sequence length to concatenate training samples together up to
-# inspired by StackLLaMA. see https://huggingface.co/blog/stackllama#supervised-fine-tuning
+# Max sequence length to concatenate training samples together up to
+# Inspired by StackLLaMA. see https://huggingface.co/blog/stackllama#supervised-fine-tuning
 # FutureWarning: This will soon be DEPRECATED
 max_packed_sequence_len: 1024
-# use efficient multi-packing with block diagonal attention and per sequence position_ids. Recommend set to 'true'
+# Use efficient multi-packing with block diagonal attention and per sequence position_ids. Recommend set to 'true'
 sample_packing:
-# set to 'false' if getting errors during eval with sample_packing on.
+# Set to 'false' if getting errors during eval with sample_packing on.
 eval_sample_packing:
-# you can set these packing optimizations AFTER starting a training at least once.
+# You can set these packing optimizations AFTER starting a training at least once.
 # The trainer will provide recommended values for these values.
 sample_packing_eff_est:
 total_num_tokens:
 
-# if you want to use 'lora' or 'qlora' or leave blank to train all parameters in original model
+# 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
-# lora hyperparameters
+# 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 `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
+# For more details about the following options, see:
+# https://www.anyscale.com/blog/fine-tuning-llms-lora-or-full-parameter-an-in-depth-analysis-with-llama-2
 lora_r: 8
 lora_alpha: 16
 lora_dropout: 0.05
@@ -516,81 +666,103 @@ lora_target_modules:
 #  - gate_proj
 #  - down_proj
 #  - up_proj
-lora_target_linear: # if true, will target all linear layers
+lora_target_linear: # If true, will target all linear layers
+
+# If you added new tokens to the tokenizer, you may need to save some LoRA modules because they need to know the new tokens.
+# For LLaMA and Mistral, you need to save `embed_tokens` and `lm_head`. It may vary for other models.
+# `embed_tokens` converts tokens to embeddings, and `lm_head` converts embeddings to token probabilities.
+# https://github.com/huggingface/peft/issues/334#issuecomment-1561727994
 lora_modules_to_save:
 #  - embed_tokens
 #  - lm_head
-lora_out_dir:
+
 lora_fan_in_fan_out: false
 
 # ReLoRA configuration
-# must use either 'lora' or 'qlora' adapter, and does not support fsdp or deepspeed
-relora_steps: # number of steps per ReLoRA restart
-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
+# Must use either 'lora' or 'qlora' adapter, and does not support fsdp or deepspeed
+relora_steps: # Number of steps per ReLoRA restart
+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_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
 
-# where to save the finished model to
+# Where to save the full-finetuned model to
 output_dir: ./completed-model
 
-# whether to use torch.compile and which backend to use
+# Whether to use torch.compile and which backend to use
 torch_compile:  # bool
 torch_compile_backend:  # Optional[str]
 
-# training hyperparameters
+# Training hyperparameters
+
+# If greater than 1, backpropagation will be skipped and the gradients will be accumulated for the given number of steps.
 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: 2
-num_epochs: 3
-warmup_steps: 100
+eval_batch_size:
+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:
-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
-save_total_limit: # checkpoints saved at a time
+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
+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.
+# e.g., when 1 epoch is 1000 steps => `num_epochs: 2` and `max_steps: 100` will train for 100 steps
 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
+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
 
-# save model as safetensors (require safetensors package)
+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:
 
-# whether to mask out or include the human's prompt from the training labels
+# Whether to mask out or include the human's prompt from the training labels
 train_on_inputs: false
-# group similarly sized data to minimize padding
-# may be slower to start, as it must download and sort the entire dataset
-# note that training loss may have an oscillating pattern with this enabled
+# Group similarly sized data to minimize padding.
+# May be slower to start, as it must download and sort the entire dataset.
+# Note that training loss may have an oscillating pattern with this enabled.
 group_by_length: false
 
 # Whether to use gradient checkpointing https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing
 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
+# 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
 early_stopping_patience: 3
 
-# specify a scheduler and kwargs to use with the optimizer
+# Specify a scheduler and kwargs to use with the optimizer
 lr_scheduler: # 'one_cycle' | 'log_sweep' | empty for cosine
 lr_scheduler_kwargs:
 
-# for one_cycle optim
-lr_div_factor: # learning rate div factor
+# For one_cycle optim
+lr_div_factor: # Learning rate div factor
 
-# for log_sweep optim
+# For log_sweep optim
 log_sweep_min_lr:
 log_sweep_max_lr:
 
-# specify optimizer
+# Specify optimizer
 # Valid values are driven by the Transformers OptimizerNames class, see:
 # https://github.com/huggingface/transformers/blob/95b374952dc27d8511541d6f5a4e22c9ec11fb24/src/transformers/training_args.py#L134
 #
@@ -616,7 +788,7 @@ log_sweep_max_lr:
 # - paged_lion_32bit
 # - paged_lion_8bit
 optimizer:
-# specify weight decay
+# Specify weight decay
 weight_decay:
 # adamw hyperparams
 adam_beta1:
@@ -625,47 +797,49 @@ adam_epsilon:
 # Gradient clipping max norm
 max_grad_norm:
 
-# whether to bettertransformers
+# 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
+neftune_noise_alpha:
+
+# Whether to bettertransformers
 flash_optimum:
-# whether to use xformers attention patch https://github.com/facebookresearch/xformers:
+# Whether to use xformers attention patch https://github.com/facebookresearch/xformers:
 xformers_attention:
-# whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:
+# Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:
 flash_attention:
-# whether to use scaled-dot-product attention
+flash_attn_cross_entropy:  # Whether to use flash-attention cross entropy implementation - advanced use only
+flash_attn_rms_norm:  # Whether to use flash-attention rms norm implementation - advanced use only
+flash_attn_fuse_qkv: # Whether to fuse QKV into a single operation
+flash_attn_fuse_mlp: # Whether to fuse part of the MLP into a single operation
+# Whether to use scaled-dot-product attention
 # 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 a specific checkpoint dir
 resume_from_checkpoint:
-# if resume_from_checkpoint isn't set and you simply want it to start where it left off
-# be careful with this being turned on between different models
+# If resume_from_checkpoint isn't set and you simply want it to start where it left off.
+# Be careful with this being turned on between different models.
 auto_resume_from_checkpoints: false
 
-# don't mess with this, it's here for accelerate and torchrun
+# Don't mess with this, it's here for accelerate and torchrun
 local_rank:
 
-# add or change special tokens
+# Add or change special tokens.
+# If you add tokens here, you don't need to add them to the `tokens` list.
 special_tokens:
   # bos_token: "<s>"
   # eos_token: "</s>"
   # unk_token: "<unk>"
-# add extra tokens
+
+# Add extra tokens.
 tokens:
 
 # FSDP
 fsdp:
 fsdp_config:
 
-# Deepspeed config path
+# Deepspeed config path. e.g., deepspeed/zero3.json
 deepspeed:
 
 # Advanced DDP Arguments
@@ -691,6 +865,66 @@ strict:
 
 </details>
 
+<details>
+<summary> Understanding of batch size and gradient accumulation steps </summary>
+<br/>
+Gradient accumulation means accumulating gradients over several mini-batches and updating the model weights afterward. When the samples in each batch are diverse, this technique doesn't significantly impact learning.
+
+This method allows for effective training with larger effective batch sizes without needing proportionally larger memory. Here's why:
+
+1. **Memory Consumption with Batch Size**: The primary reason increasing the batch size impacts memory is due to the storage requirements for intermediate activations. When you forward propagate a batch through a network, you have to store the activations at each layer for each sample in the batch, because these activations are used during backpropagation to compute gradients. Therefore, larger batches mean more activations, leading to greater GPU memory consumption.
+
+2. **Gradient Accumulation**: With gradient accumulation, you're effectively simulating a larger batch size by accumulating gradients over several smaller batches (or micro-batches). However, at any given time, you're only forward and backward propagating a micro-batch. This means you only store activations for the micro-batch, not the full accumulated batch. As a result, you can simulate the effect of a larger batch size without the memory cost of storing activations for a large batch.
+
+**Example 1:**
+Micro batch size: 3
+Gradient accumulation steps: 2
+Number of GPUs: 3
+Total batch size = 3 * 2 * 3 = 18
+
+```
+| GPU 1          | GPU 2          | GPU 3          |
+|----------------|----------------|----------------|
+| S1, S2, S3     | S4, S5, S6     | S7, S8, S9     |
+| e1, e2, e3     | e4, e5, e6     | e7, e8, e9     |
+|----------------|----------------|----------------|
+| → (accumulate) | → (accumulate) | → (accumulate) |
+|----------------|----------------|----------------|
+| S10, S11, S12  | S13, S14, S15  | S16, S17, S18  |
+| e10, e11, e12  | e13, e14, e15  | e16, e17, e18  |
+|----------------|----------------|----------------|
+| → (apply)      | → (apply)      | → (apply)      |
+
+Accumulated gradient for the weight w1 after the second iteration (considering all GPUs):
+Total gradient for w1 = e1 + e2 + e3 + e4 + e5 + e6 + e7 + e8 + e9 + e10 + e11 + e12 + e13 + e14 + e15 + e16 + e17 + e18
+
+Weight update for w1:
+w1_new = w1_old - learning rate x (Total gradient for w1 / 18)
+```
+
+**Example 2:**
+Micro batch size: 2
+Gradient accumulation steps: 1
+Number of GPUs: 3
+Total batch size = 2 * 1 * 3 = 6
+
+```
+| GPU 1     | GPU 2     | GPU 3     |
+|-----------|-----------|-----------|
+| S1, S2    | S3, S4    | S5, S6    |
+| e1, e2    | e3, e4    | e5, e6    |
+|-----------|-----------|-----------|
+| → (apply) | → (apply) | → (apply) |
+
+Accumulated gradient for the weight w1 (considering all GPUs):
+Total gradient for w1 = e1 + e2 + e3 + e4 + e5 + e6
+
+Weight update for w1:
+w1_new = w1_old - learning rate × (Total gradient for w1 / 6)
+```
+
+</details>
+
 ### Train
 
 Run
@@ -698,14 +932,42 @@ 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="" 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
@@ -720,37 +982,40 @@ 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
@@ -765,23 +1030,35 @@ 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:
+
+> RuntimeError: stack expects each tensor to be equal size, but got [1, 32, 1, 128] at entry 0 and [1, 32, 8, 128] at entry 1
 
 ### 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).
+
 > 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:
 
 Please reduce any below
@@ -790,6 +1067,10 @@ 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.
@@ -812,6 +1093,20 @@ 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.
+
 ## Need help? 🙋♂️
 
 Join our [Discord server](https://discord.gg/HhrNrHJPRb) where we can help you

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
+}

docker/Dockerfile

@@ -5,24 +5,31 @@ ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6+PTX"
 ARG AXOLOTL_EXTRAS=""
 ARG CUDA="118"
 ENV BNB_CUDA_VERSION=$CUDA
+ARG PYTORCH_VERSION="2.0.1"
+
+ENV PYTORCH_VERSION=$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
 
 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 cd axolotl && \
-    if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        pip install -e .[flash-attn,$AXOLOTL_EXTRAS]; \
+RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
+        pip install -e .[deepspeed,flash-attn,$AXOLOTL_EXTRAS]; \
     else \
-        pip install -e .[flash-attn]; \
+        pip install -e .[deepspeed,flash-attn]; \
     fi
 
+# So we can test the Docker image
+RUN pip install pytest
+
 # fix so that git fetch/pull from remote works
-RUN cd axolotl && \
-    git config remote.origin.fetch "+refs/heads/*:refs/remotes/origin/*" && \
+RUN git config remote.origin.fetch "+refs/heads/*:refs/remotes/origin/*" && \
     git config --get remote.origin.fetch
 
 # helper for huggingface-login cli

docker/Dockerfile-base

@@ -10,11 +10,13 @@ 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/*
+    && apt-get install -y wget git build-essential ninja-build git-lfs libaio-dev && rm -rf /var/lib/apt/lists/* \
     && wget \
     https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh \
     && mkdir /root/.conda \
@@ -27,52 +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
-
-# recompile apex
-RUN python3 -m pip uninstall -y apex
-RUN git clone https://github.com/NVIDIA/apex
-RUN cd apex && python3 -m pip install -v --disable-pip-version-check --no-cache-dir --no-build-isolation --config-settings "--build-option=--cpp_ext" --config-settings "--build-option=--cuda_ext" ./
-
-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 +4,7 @@ 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"
 
 COPY scripts/runpod-entrypoint.sh /root/runpod-entrypoint.sh
 

docs/faq.md

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

docs/multipack.md

@@ -0,0 +1,51 @@
+# Multipack
+
+4k context, bsz =4,
+each character represents 256 tokens
+X represents a padding token
+
+```
+   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5
+[[ A A A A A A A A A A A ]
+   B B B B B B ]
+   C C C C C C C ]
+   D D D D ]]
+
+[[ E E E E E E E E ]
+ [ F F F F ]
+ [ G G G ]
+ [ H H H H ]]
+
+[[ I I I ]
+ [ J J J ]
+ [ K K K K K]
+ [ L L L ]]
+```
+
+after padding to longest input in each step
+```
+   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5
+[[ A A A A A A A A A A A ]
+   B B B B B B X X X X X X ]
+   C C C C C C C X X X X ]
+   D D D D X X X X X X X ]]
+
+[[ E E E E E E E E ]
+ [ F F F F X X X X ]
+ [ G G G X X X X X ]
+ [ H H H H X X X X ]]
+
+[[ I I I X X ]
+ [ J J J X X ]
+ [ K K K K K ]
+ [ L L L X X ]]
+```
+
+w packing ( note it's the same effective number of tokens per step, but a true bsz of 1)
+```
+   0 1 2 3 4 5 6 7 8 9 0 1 2 3 4 5
+[[ A A A A A A A A A A A B B B B B
+   B C C C C C C C D D D D E E E E
+   E E E E F F F F F G G G H H H H
+   I I I J J J J K K K K K L L L X ]]
+```

docs/rlhf.md

@@ -0,0 +1,35 @@
+# 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
+```

examples/cerebras/btlm-ft.yml

@@ -1,5 +1,4 @@
 base_model: cerebras/btlm-3b-8k-base
-base_model_config: cerebras/btlm-3b-8k-base
 model_type: AutoModelForCausalLM
 tokenizer_type: GPT2Tokenizer
 trust_remote_code: true
@@ -15,7 +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:
@@ -36,7 +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
@@ -73,8 +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

@@ -1,5 +1,4 @@
 base_model: cerebras/Cerebras-GPT-1.3B
-base_model_config: cerebras/Cerebras-GPT-1.3B
 load_in_8bit: false
 load_in_4bit: true
 strict: false
@@ -7,8 +6,8 @@ push_dataset_to_hub:
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
     type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
+dataset_prepared_path:
+val_set_size: 0.05
 adapter: qlora
 lora_model_dir:
 sequence_len: 2048
@@ -25,7 +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
@@ -50,8 +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

@@ -1,5 +1,4 @@
 base_model: codellama/CodeLlama-13b-hf
-base_model_config: codellama/CodeLlama-13b-hf
 model_type: LlamaForCausalLM
 tokenizer_type: CodeLlamaTokenizer
 is_llama_derived_model: true
@@ -11,8 +10,8 @@ strict: false
 datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
+dataset_prepared_path:
+val_set_size: 0.05
 output_dir: ./lora-out
 
 sequence_len: 4096
@@ -30,12 +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
@@ -55,8 +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

@@ -1,5 +1,4 @@
 base_model: codellama/CodeLlama-13b-hf
-base_model_config: codellama/CodeLlama-13b-hf
 model_type: LlamaForCausalLM
 tokenizer_type: CodeLlamaTokenizer
 is_llama_derived_model: true
@@ -11,8 +10,8 @@ strict: false
 datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
+dataset_prepared_path:
+val_set_size: 0.05
 output_dir: ./qlora-out
 
 adapter: qlora
@@ -32,12 +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
@@ -57,8 +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

@@ -1,5 +1,4 @@
 base_model: codellama/CodeLlama-34b-hf
-base_model_config: codellama/CodeLlama-34b-hf
 model_type: LlamaForCausalLM
 tokenizer_type: CodeLlamaTokenizer
 is_llama_derived_model: true
@@ -11,8 +10,8 @@ strict: false
 datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
+dataset_prepared_path:
+val_set_size: 0.05
 output_dir: ./lora-out
 
 sequence_len: 4096
@@ -30,12 +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
@@ -55,8 +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

@@ -1,5 +1,4 @@
 base_model: codellama/CodeLlama-34b-hf
-base_model_config: codellama/CodeLlama-34b-hf
 model_type: LlamaForCausalLM
 tokenizer_type: CodeLlamaTokenizer
 is_llama_derived_model: true
@@ -11,8 +10,8 @@ strict: false
 datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
+dataset_prepared_path:
+val_set_size: 0.05
 output_dir: ./qlora-out
 
 adapter: qlora
@@ -32,12 +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
@@ -57,8 +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

@@ -1,5 +1,4 @@
 base_model: codellama/CodeLlama-7b-hf
-base_model_config: codellama/CodeLlama-7b-hf
 model_type: LlamaForCausalLM
 tokenizer_type: CodeLlamaTokenizer
 is_llama_derived_model: true
@@ -11,8 +10,8 @@ strict: false
 datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
+dataset_prepared_path:
+val_set_size: 0.05
 output_dir: ./lora-out
 
 sequence_len: 4096
@@ -30,12 +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
@@ -55,8 +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

@@ -1,5 +1,4 @@
 base_model: codellama/CodeLlama-7b-hf
-base_model_config: codellama/CodeLlama-7b-hf
 model_type: LlamaForCausalLM
 tokenizer_type: CodeLlamaTokenizer
 is_llama_derived_model: true
@@ -11,8 +10,8 @@ strict: false
 datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
+dataset_prepared_path:
+val_set_size: 0.05
 output_dir: ./qlora-out
 
 adapter: qlora
@@ -32,12 +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
@@ -57,8 +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

@@ -1,8 +1,8 @@
 base_model: tiiuae/falcon-7b
-base_model_config: tiiuae/falcon-7b
 trust_remote_code: true
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+is_falcon_derived_model: true
 load_in_8bit: true
 load_in_4bit: false
 gptq: false
@@ -11,8 +11,8 @@ push_dataset_to_hub:
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
     type: alpaca:chat
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
+dataset_prepared_path:
+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

@@ -1,11 +1,11 @@
 # 1b: tiiuae/falcon-rw-1b
 # 40b: tiiuae/falcon-40b
 base_model: tiiuae/falcon-7b
-base_model_config: tiiuae/falcon-7b
 # required by falcon custom model code: https://huggingface.co/tiiuae/falcon-7b/tree/main
 trust_remote_code: true
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+is_falcon_derived_model: true
 load_in_8bit: false
 # enable 4bit for QLoRA
 load_in_4bit: true
@@ -17,8 +17,8 @@ datasets:
     data_files:
       - Chain-of-Thought/formatted_cot_data/gsm8k_train.json
     type: "alpaca:chat"
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
+dataset_prepared_path:
+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

@@ -1,8 +1,8 @@
 base_model: tiiuae/falcon-7b
-base_model_config: tiiuae/falcon-7b
 trust_remote_code: true
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
+is_falcon_derived_model: true
 load_in_8bit: false
 load_in_4bit: false
 gptq: false
@@ -11,8 +11,8 @@ push_dataset_to_hub:
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
     type: alpaca:chat
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
+dataset_prepared_path:
+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

@@ -1,5 +1,4 @@
 base_model: EleutherAI/gpt-j-6b
-base_model_config: EleutherAI/gpt-j-6b
 load_in_8bit: false
 load_in_4bit: true
 strict: false
@@ -7,8 +6,8 @@ push_dataset_to_hub:
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
     type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
+dataset_prepared_path:
+val_set_size: 0.05
 adapter: qlora
 lora_model_dir:
 sequence_len: 2048
@@ -22,7 +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
@@ -47,8 +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

@@ -1,12 +1,11 @@
 base_model: huggyllama/llama-7b
-base_model_config: huggyllama/llama-7b
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 load_in_8bit: false
 datasets:
   - path: openaccess-ai-collective/jeopardy
     type: jeopardy
-dataset_prepared_path: last_run_prepared
+dataset_prepared_path:
 val_set_size: 0.02
 adapter:
 lora_model_dir:
@@ -20,12 +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
@@ -43,8 +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/README.md

@@ -9,12 +9,16 @@ gradient_accumulation_steps: 2
 micro_batch_size: 1
 
 ```shell
-accelerate launch scripts/finetune.py examples/llama-2/qlora.yml
-
+accelerate launch -m axolotl.cli.train examples/llama-2/qlora.yml
 ```
 or
 
 ```shell
-accelerate launch scripts/finetune.py examples/llama-2/lora.yml
-
+accelerate launch -m axolotl.cli.train examples/llama-2/lora.yml
+```
+
+To launch a full finetuning with 16-bit precision:
+
+```shell
+accelerate launch -m axolotl.cli.train examples/llama-2/fft_optimized.yml
 ```

examples/llama-2/fft_optimized.yml

@@ -0,0 +1,72 @@
+base_model: NousResearch/Llama-2-7b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.05
+output_dir: ./out
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+adapter:
+lora_model_dir:
+lora_r:
+lora_alpha:
+lora_dropout:
+lora_target_linear:
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+flash_attn_cross_entropy: false
+flash_attn_rms_norm: true
+flash_attn_fuse_qkv: false
+flash_attn_fuse_mlp: true
+
+warmup_steps: 100
+evals_per_epoch: 4
+eval_table_size:
+saves_per_epoch: 1
+debug:
+deepspeed: #deepspeed/zero2.json # multi-gpu only
+weight_decay: 0.1
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/llama-2/gptq-lora.yml

@@ -1,5 +1,4 @@
 base_model: TheBloke/Llama-2-7B-GPTQ
-base_model_config: TheBloke/Llama-2-7B-GPTQ
 is_llama_derived_model: false
 gptq: true
 gptq_disable_exllama: true
@@ -15,8 +14,8 @@ hf_use_auth_token: true
 datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
+dataset_prepared_path:
+val_set_size: 0.05
 adapter: lora
 lora_model_dir:
 sequence_len: 4096
@@ -33,12 +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
@@ -63,8 +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

@@ -1,5 +1,4 @@
 base_model: NousResearch/Llama-2-7b-hf
-base_model_config: NousResearch/Llama-2-7b-hf
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 is_llama_derived_model: true
@@ -11,8 +10,8 @@ strict: false
 datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
+dataset_prepared_path:
+val_set_size: 0.05
 output_dir: ./lora-out
 
 sequence_len: 4096
@@ -30,12 +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
@@ -55,10 +54,10 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
-eval_table_size: 5
+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

@@ -1,5 +1,4 @@
 base_model: NousResearch/Llama-2-7b-hf
-base_model_config: NousResearch/Llama-2-7b-hf
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 is_llama_derived_model: true
@@ -11,8 +10,8 @@ strict: false
 datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
+dataset_prepared_path:
+val_set_size: 0.05
 output_dir: ./qlora-out
 
 adapter: qlora
@@ -32,12 +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
@@ -57,9 +56,9 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
-eval_table_size: 5
-save_steps:
+evals_per_epoch: 4
+eval_table_size:
+saves_per_epoch: 1
 debug:
 deepspeed:
 weight_decay: 0.0

examples/llama-2/relora.yml

@@ -1,5 +1,4 @@
 base_model: NousResearch/Llama-2-7b-hf
-base_model_config: NousResearch/Llama-2-7b-hf
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 is_llama_derived_model: true
@@ -11,8 +10,8 @@ strict: false
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
     type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
+dataset_prepared_path:
+val_set_size: 0.05
 output_dir: ./relora-out
 
 adapter: qlora
@@ -36,12 +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
@@ -61,8 +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/README.md

@@ -0,0 +1,12 @@
+**Mistral 7B** is a language model with a total of 7.3 billion parameters, showcasing a notable performance across a variety of benchmarks.
+
+Fine Tune:
+```shell
+accelerate launch -m axolotl.cli.train examples/mistral/config.yml
+
+```
+
+If you run into CUDA OOM, use deepspeed with config zero2.json:
+```shell
+accelerate launch -m axolotl.cli.train examples/mistral/config.yml --deepspeed deepspeed/zero2.json
+```

examples/mistral/config.yml

@@ -0,0 +1,62 @@
+base_model: mistralai/Mistral-7B-v0.1
+model_type: MistralForCausalLM
+tokenizer_type: LlamaTokenizer
+is_mistral_derived_model: true
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./out
+
+sequence_len: 8192
+sample_packing: true
+pad_to_sequence_len: true
+eval_sample_packing: false
+
+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: 0.000005
+
+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
+eval_table_size:
+eval_table_max_new_tokens: 128
+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/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

@@ -0,0 +1,81 @@
+base_model: mistralai/Mistral-7B-v0.1
+model_type: MistralForCausalLM
+tokenizer_type: LlamaTokenizer
+is_mistral_derived_model: true
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./qlora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 8192
+sample_packing: true
+pad_to_sequence_len: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16: false
+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:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/mpt-7b/config.yml

@@ -1,12 +1,11 @@
 base_model: mosaicml/mpt-7b
-base_model_config: mosaicml/mpt-7b
 tokenizer_type: AutoTokenizer
 trust_remote_code: true  # required for mpt as their model class is not merged into transformers yet
 load_in_8bit: false
 datasets:
   - path: vicgalle/alpaca-gpt4
     type: alpaca
-dataset_prepared_path: last_run_prepared
+dataset_prepared_path:
 val_set_size: 0.02
 adapter:
 lora_model_dir:
@@ -22,12 +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
@@ -45,8 +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

@@ -1,5 +1,4 @@
 base_model: openlm-research/open_llama_3b_v2
-base_model_config: openlm-research/open_llama_3b_v2
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 load_in_8bit: false
@@ -9,7 +8,7 @@ push_dataset_to_hub:
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
     type: alpaca
-dataset_prepared_path: last_run_prepared
+dataset_prepared_path:
 val_set_size: 0.02
 adapter:
 lora_model_dir:
@@ -24,7 +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
@@ -50,8 +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

@@ -1,5 +1,4 @@
 base_model: openlm-research/open_llama_3b_v2
-base_model_config: openlm-research/open_llama_3b_v2
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 load_in_8bit: true
@@ -9,7 +8,7 @@ push_dataset_to_hub:
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
     type: alpaca
-dataset_prepared_path: last_run_prepared
+dataset_prepared_path:
 val_set_size: 0.02
 adapter: lora
 lora_model_dir:
@@ -30,7 +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
@@ -55,8 +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

@@ -1,5 +1,4 @@
 base_model: openlm-research/open_llama_3b_v2
-base_model_config: openlm-research/open_llama_3b_v2
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 load_in_8bit: false
@@ -9,8 +8,8 @@ push_dataset_to_hub:
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
     type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
+dataset_prepared_path:
+val_set_size: 0.05
 adapter: qlora
 lora_model_dir:
 sequence_len: 1024
@@ -24,7 +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
@@ -49,8 +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

@@ -1,6 +1,5 @@
 base_model: microsoft/phi-1_5
-base_model_config: microsoft/phi-1_5
-model_type: MixFormerSequentialForCausalLM
+model_type: PhiForCausalLM
 tokenizer_type: AutoTokenizer
 is_llama_derived_model: false
 trust_remote_code: true
@@ -13,7 +12,7 @@ datasets:
   - path: garage-bAInd/Open-Platypus
     type: alpaca
 
-dataset_prepared_path: last_run_prepared
+dataset_prepared_path:
 val_set_size: 0.05
 output_dir: ./phi-sft-out
 
@@ -32,7 +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
@@ -60,8 +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

@@ -1,5 +1,4 @@
 base_model: microsoft/phi-1_5
-base_model_config: microsoft/phi-1_5
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
 is_llama_derived_model: false
@@ -13,7 +12,7 @@ datasets:
   - path: garage-bAInd/Open-Platypus
     type: alpaca
 
-dataset_prepared_path: last_run_prepared
+dataset_prepared_path:
 val_set_size: 0.05
 output_dir: ./phi-sft-out
 
@@ -32,7 +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
@@ -60,8 +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/pythia-12b/config.yml

@@ -1,5 +1,4 @@
 base_model: EleutherAI/pythia-12b-deduped
-base_model_config: EleutherAI/pythia-12b-deduped
 base_model_ignore_patterns: pytorch*  # prefer safetensors
 model_type: GPTNeoXForCausalLM
 tokenizer_type: AutoTokenizer
@@ -10,7 +9,7 @@ device_map: auto
 datasets:
   - path: vicgalle/alpaca-gpt4
     type: alpaca
-dataset_prepared_path: last_run_prepared
+dataset_prepared_path:
 val_set_size: 0.05
 adapter:
 lora_model_dir:
@@ -25,7 +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

@@ -1,10 +1,9 @@
 base_model: EleutherAI/pythia-1.4b-deduped
-base_model_config: EleutherAI/pythia-1.4b-deduped
 load_in_8bit: true
 datasets:
   - path: teknium/GPT4-LLM-Cleaned
     type: alpaca
-dataset_prepared_path: last_run_prepared
+dataset_prepared_path:
 val_set_size: 0.05
 adapter: lora
 lora_model_dir:
@@ -19,20 +18,20 @@ 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
-bf16: True
-tf32: True
+bf16: true
+tf32: true
 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

@@ -1,5 +1,4 @@
 base_model: togethercomputer/RedPajama-INCITE-Chat-3B-v1
-base_model_config: togethercomputer/RedPajama-INCITE-Chat-3B-v1
 model_type: GPTNeoXForCausalLM
 tokenizer_type: AutoTokenizer
 trust_remote_code:
@@ -7,7 +6,7 @@ load_in_8bit: false
 datasets:
   - path: vicgalle/alpaca-gpt4
     type: alpaca
-dataset_prepared_path: last_run_prepared
+dataset_prepared_path:
 val_set_size: 0.02
 adapter:
 lora_model_dir:
@@ -23,12 +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
@@ -46,8 +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

@@ -1,11 +1,10 @@
 base_model: replit/replit-code-v1-3b
-base_model_config: replit/replit-code-v1-3b
 trust_remote_code: true
 load_in_8bit: false
 datasets:
   - path: vicgalle/alpaca-gpt4
     type: alpaca
-dataset_prepared_path: last_run_prepared
+dataset_prepared_path:
 val_set_size: 0.05
 adapter: lora
 lora_model_dir:
@@ -22,12 +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:
@@ -46,8 +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

@@ -1,6 +1,4 @@
-base_model: PY007/TinyLlama-1.1B-step-50K-105b
-base_model_config: 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
@@ -12,12 +10,13 @@ strict: false
 datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
+dataset_prepared_path:
+val_set_size: 0.05
 output_dir: ./lora-out
 
 sequence_len: 4096
 sample_packing: true
+pad_to_sequence_len: true
 
 adapter: lora
 lora_model_dir:
@@ -30,12 +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
@@ -55,15 +54,11 @@ xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
-eval_steps: 20
-eval_table_size: 5
-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

@@ -0,0 +1,58 @@
+base_model: TinyLlama/TinyLlama-1.1B-Chat-v1.0
+
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+max_steps: 200
+pretraining_dataset:
+  path: c4
+  name: en
+dataset_prepared_path:
+val_set_size: 0.0
+output_dir: ./model-out
+
+sequence_len: 2048
+sample_packing: true
+
+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: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch:
+eval_table_size:
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:

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

@@ -1,7 +1,6 @@
 # An example finetuning Saleforce's XGen-7b model with 8k context using qlora
 # on Tim Dettmer's Guanaco dataset.
 base_model: Salesforce/xgen-7b-8k-base
-base_model_config: Salesforce/xgen-7b-8k-base
 trust_remote_code: true
 model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
@@ -16,8 +15,8 @@ datasets:
     data_files:
       - openassistant_best_replies_train.jsonl
     type: "completion"
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.01
+dataset_prepared_path:
+val_set_size: 0.05
 # enable QLoRA
 adapter: qlora
 lora_model_dir:
@@ -39,7 +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
 
@@ -52,7 +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:
@@ -79,8 +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

@@ -1,24 +1,22 @@
---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
+auto-gptq==0.5.1
 packaging
-peft @ git+https://github.com/huggingface/peft.git
-transformers @ git+https://github.com/huggingface/transformers.git
+peft==0.6.0
+transformers @ git+https://github.com/huggingface/transformers.git@3cefac1d974db5e2825a0cb2b842883a628be7a0
+tokenizers==0.15.0
 bitsandbytes>=0.41.1
-accelerate @ git+https://github.com/huggingface/accelerate
+accelerate==0.24.1
 deepspeed
 addict
-evaluate
 fire
 PyYAML>=6.0
-datasets
-flash-attn>=2.2.1
+datasets>=2.15.0
+flash-attn==2.3.3
 sentencepiece
 wandb
 einops
-xformers
-optimum
+xformers==0.0.22
+optimum==1.13.2
 hf_transfer
 colorama
 numba
@@ -31,3 +29,13 @@ scipy
 scikit-learn==1.2.2
 pynvml
 art
+fschat==0.2.34
+gradio==3.50.2
+tensorboard
+
+# remote filesystems
+s3fs
+gcsfs
+# adlfs
+
+trl @ git+https://github.com/huggingface/trl.git@main

scripts/finetune.py

@@ -7,6 +7,7 @@ import transformers
 
 from axolotl.cli import (
     check_accelerate_default_config,
+    check_user_token,
     do_inference,
     do_merge_lora,
     load_cfg,
@@ -31,6 +32,7 @@ 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))
     parsed_cli_args, _ = parser.parse_args_into_dataclasses(
         return_remaining_strings=True
@@ -43,8 +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
 
 
@@ -21,6 +23,15 @@ def parse_requirements():
             ):
                 # Handle standard packages
                 _install_requires.append(line)
+
+    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
 
 
@@ -38,10 +49,13 @@ setup(
     dependency_links=dependency_links,
     extras_require={
         "flash-attn": [
-            "flash-attn>=2.2.1",
+            "flash-attn==2.3.3",
         ],
         "deepspeed": [
             "deepspeed",
         ],
+        "mamba-ssm": [
+            "mamba-ssm==1.0.1",
+        ],
     },
 )

src/axolotl/cli/__init__.py

@@ -2,19 +2,25 @@
 
 import importlib
 import logging
+import math
 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 transformers import GenerationConfig, TextStreamer
+from datasets import concatenate_datasets, load_dataset
+from huggingface_hub import HfApi
+from huggingface_hub.utils import LocalTokenNotFoundError
+from transformers import GenerationConfig, TextIteratorStreamer, TextStreamer
 
 from axolotl.common.cli import TrainerCliArgs, load_model_and_tokenizer
 from axolotl.logging_config import configure_logging
@@ -25,6 +31,7 @@ from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import is_main_process
 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__), ".."))
@@ -42,14 +49,14 @@ 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)
 
 
 def get_multi_line_input() -> Optional[str]:
-    print("Give me an instruction (Ctrl + D to finish): ")
+    print("Give me an instruction (Ctrl + D to submit): ")
     instruction = ""
     for line in sys.stdin:
         instruction += line  # pylint: disable=consider-using-join
@@ -66,14 +73,15 @@ 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"))
 
@@ -98,15 +106,7 @@ 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)
@@ -151,6 +151,83 @@ 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"))
 
@@ -192,6 +269,7 @@ def load_cfg(config: Path = Path("examples/"), **kwargs):
     # load the config from the yaml file
     with open(config, encoding="utf-8") as file:
         cfg: DictDefault = DictDefault(yaml.safe_load(file))
+    cfg.axolotl_config_path = config
     # if there are any options passed in the cli, if it is something that seems valid from the yaml,
     # then overwrite the value
     cfg_keys = cfg.keys()
@@ -206,6 +284,8 @@ def load_cfg(config: Path = Path("examples/"), **kwargs):
 
     validate_config(cfg)
 
+    prepare_optim_env(cfg)
+
     normalize_config(cfg)
 
     setup_wandb_env_vars(cfg)
@@ -219,7 +299,9 @@ 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...")
@@ -235,6 +317,98 @@ 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,
@@ -247,3 +421,16 @@ def check_accelerate_default_config():
         LOG.warning(
             f"accelerate config file found at {config_args.default_yaml_config_file}. This can lead to unexpected errors"
         )
+
+
+def check_user_token():
+    # Verify if token is valid
+    api = HfApi()
+    try:
+        user_info = api.whoami()
+        return bool(user_info)
+    except LocalTokenNotFoundError:
+        LOG.warning(
+            "Error verifying HuggingFace token. Remember to log in using `huggingface-cli login` and get your access token from https://huggingface.co/settings/tokens if you want to use gated models or datasets."
+        )
+        return False

src/axolotl/cli/inference.py

@@ -6,21 +6,30 @@ 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)
+    parsed_cfg.sample_packing = False
     parser = transformers.HfArgumentParser((TrainerCliArgs))
     parsed_cli_args, _ = parser.parse_args_into_dataclasses(
         return_remaining_strings=True
     )
     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 +18,22 @@ 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."
+        )
 
     do_merge_lora(cfg=parsed_cfg, cli_args=parsed_cli_args)
 

src/axolotl/cli/preprocess.py

@@ -0,0 +1,53 @@
+"""
+CLI to run training on a model
+"""
+import logging
+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,
+    print_axolotl_text_art,
+)
+from axolotl.common.cli import PreprocessCliArgs
+from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
+
+LOG = logging.getLogger("axolotl.cli.preprocess")
+
+
+def do_cli(config: Path = Path("examples/"), **kwargs):
+    # pylint: disable=duplicate-code
+    print_axolotl_text_art()
+    parsed_cfg = load_cfg(config, **kwargs)
+    check_accelerate_default_config()
+    check_user_token()
+    parser = transformers.HfArgumentParser((PreprocessCliArgs))
+    parsed_cli_args, _ = parser.parse_args_into_dataclasses(
+        return_remaining_strings=True
+    )
+    if not parsed_cfg.dataset_prepared_path:
+        msg = (
+            Fore.RED
+            + "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
+
+    _ = 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 __name__ == "__main__":
+    fire.Fire(do_cli)

src/axolotl/cli/train.py

@@ -1,6 +1,7 @@
 """
 CLI to run training on a model
 """
+import logging
 from pathlib import Path
 
 import fire
@@ -8,27 +9,32 @@ import transformers
 
 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.train import train
 
+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
     )
-
-    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/common/const.py

@@ -0,0 +1,5 @@
+"""
+Various shared constants
+"""
+
+DEFAULT_DATASET_PREPARED_PATH = "last_run_prepared"

src/axolotl/core/trainer_builder.py

@@ -0,0 +1,943 @@
+"""
+Builder for the training args and trainer
+"""
+
+import abc
+import importlib
+import logging
+import math
+import sys
+from abc import abstractmethod
+from dataclasses import dataclass, field
+from functools import wraps
+from pathlib import Path
+from typing import Optional
+
+import torch
+import transformers
+from datasets import Dataset
+from torch.optim.lr_scheduler import OneCycleLR
+from torch.utils.data import BatchSampler, DataLoader, RandomSampler, SequentialSampler
+from transformers import EarlyStoppingCallback, Trainer, TrainingArguments
+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 (
+    BatchSamplerDataCollatorForSeq2Seq,
+    MambaDataCollator,
+)
+from axolotl.utils.samplers import MultipackBatchSampler
+from axolotl.utils.schedulers import get_cosine_schedule_with_quadratic_warmup
+
+try:
+    import torch._dynamo  # pylint: disable=ungrouped-imports
+except ImportError:
+    pass
+
+LOG = logging.getLogger("axolotl.core.trainer_builder")
+
+
+@dataclass
+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."},
+    )
+    eval_sample_packing: Optional[bool] = field(
+        default=None,
+        metadata={"help": "Use sample packing for efficient evals."},
+    )
+    sample_packing_efficiency: float = field(
+        default=1.0,
+        metadata={"help": "Sample packing efficiency for calculating batch length."},
+    )
+    max_seq_length: int = field(
+        default=2048,
+        metadata={"help": "The maximum sequence length the model can handle"},
+    )
+    sample_packing_seq_len_multiplier: int = field(
+        default=1,
+        metadata={"help": "the multiplier for the max len for packed sequences"},
+    )
+    relora_steps: Optional[int] = field(
+        default=None,
+        metadata={"help": "how often to reset for ReLoRA"},
+    )
+    relora_warmup_steps: Optional[int] = field(
+        default=None,
+        metadata={"help": "how many warmup steps to take after reset for ReLoRA"},
+    )
+    bench_split: Optional[str] = field(
+        default="eval", metadata={"help": "The benchmark split to run on"}
+    )
+    bench_dataset: Optional[str] = field(
+        default="pharaouk/dharma-1/dharma_1_mini.json",
+        metadata={
+            "help": "Benchmark dataset to use: options are `mmlu-zs`, `mmlu-fs`, or the full path to the dataset file"
+        },
+    )
+    do_bench_eval: Optional[bool] = field(
+        default=False, metadata={"help": "Whether to run the Benchmark evaluation."}
+    )
+    max_bench_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "If set, only evaluates on `max_bench_samples` of the benchmark dataset."
+        },
+    )
+    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"},
+    )
+
+
+class AxolotlTrainer(Trainer):
+    """
+    Extend the base Trainer for axolotl helpers
+    """
+
+    args = None  # type: AxolotlTrainingArguments
+    tag_names = ["axolotl"]
+
+    def __init__(self, *args, num_epochs=1, bench_data_collator=None, **kwargs):
+        self.num_epochs = num_epochs
+        self.bench_data_collator = bench_data_collator
+        super().__init__(*args, **kwargs)
+
+    def create_scheduler(
+        self, num_training_steps: int, optimizer: torch.optim.Optimizer = None
+    ):
+        """
+        Setup the scheduler. The optimizer of the trainer must have been set up either before this method is called or
+        passed as an argument.
+
+        Args:
+            num_training_steps (int): The number of training steps to do.
+            optimizer (torch.optim.Optimizer): The training optimizer
+        """
+
+        # fmt: off
+        if self.lr_scheduler is None:  # type: ignore  # pylint: disable=access-member-before-definition
+            # fmt: on
+            if (
+                self.args.lr_scheduler_type == "cosine"
+                and self.args.lr_quadratic_warmup is True
+            ):
+                self.lr_scheduler = get_cosine_schedule_with_quadratic_warmup(  # pylint: disable=attribute-defined-outside-init
+                    optimizer,
+                    num_warmup_steps=self.args.get_warmup_steps(num_training_steps),
+                    num_training_steps=num_training_steps,
+                )
+            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.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=(
+                    self.train_dataset.data.column("position_ids")
+                    .to_pandas()
+                    .apply(lambda x: x[-1] + 1)
+                    .values
+                ),
+                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.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=(
+                    eval_dataset.data.column("position_ids")
+                    .to_pandas()
+                    .apply(lambda x: x[-1] + 1)
+                    .values
+                ),
+                packing_efficiency_estimate=self.args.sample_packing_efficiency,
+            )
+        return super()._get_eval_sampler(eval_dataset)
+
+    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) -> 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)
+            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(
+        self, bench_dataset: Dataset
+    ) -> Optional[torch.utils.data.Sampler]:
+        if self.args.world_size <= 1:
+            return SequentialSampler(bench_dataset)
+        return None
+
+    def get_bench_dataloader(
+        self,
+        bench_dataset: Dataset,
+    ) -> 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)
+            dataloader_params["drop_last"] = self.args.dataloader_drop_last
+
+        return DataLoader(bench_dataset, **dataloader_params)
+        # return self.accelerator.prepare(DataLoader(bench_dataset, **dataloader_params))
+
+    def compute_loss(self, model, inputs, return_outputs=False):
+        # use one's weighted cross entropy loss calc
+        # if self.args.sample_packing:
+        #     labels = inputs.pop("labels")
+        #     outputs = model(**inputs)
+        #     loss = trainer_weighted_loss(outputs, labels, shift_labels=True)
+        #     return (loss, outputs) if return_outputs else loss
+        return super().compute_loss(model, inputs, return_outputs=return_outputs)
+
+    def _sanitize_kwargs_for_tagging(self, tag_names, kwargs=None):
+        if isinstance(tag_names, str):
+            tag_names = [tag_names]
+
+        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
+
+        return kwargs
+
+    @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
+        )
+
+        return super().push_to_hub(*args, **kwargs)
+
+
+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):
+    """
+    Trainer subclass that uses the OneCycleLR scheduler
+    """
+
+    tag_names = ["axolotl", "onecycle"]
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.lr_scheduler = None
+
+    def create_scheduler(
+        self,
+        num_training_steps: int,
+        optimizer: Optional[torch.optim.Optimizer] = None,
+    ):
+        optimizer = self.optimizer if optimizer is None else optimizer
+        num_warmup_steps = self.args.get_warmup_steps(num_training_steps)
+        pct_start = num_warmup_steps / num_training_steps
+
+        self.lr_scheduler = OneCycleLR(
+            optimizer,
+            max_lr=self.args.learning_rate,
+            total_steps=num_training_steps,
+            pct_start=pct_start,
+            div_factor=6,
+        )
+
+        return self.lr_scheduler
+
+
+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
+
+    def create_scheduler(
+        self,
+        num_training_steps: int,
+        optimizer: Optional[torch.optim.Optimizer] = None,
+    ):
+        optimizer = self.optimizer if optimizer is None else optimizer
+        lr_scheduler = super().create_scheduler(num_training_steps, optimizer)
+
+        if self.args.relora_steps:
+            warmup_steps = (
+                self.args.relora_warmup_steps if self.args.relora_warmup_steps else 10
+            )
+            self.lr_scheduler = ReLoRAScheduler(
+                optimizer,
+                lr_scheduler,
+                self.args.relora_steps,
+                warmup_steps,
+            )
+        else:
+            self.lr_scheduler = lr_scheduler
+
+        return self.lr_scheduler
+
+
+class TrainerBuilderBase(abc.ABC):
+    """
+    Base class for trainer builder
+    """
+
+    _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
+
+    @train_dataset.setter
+    def train_dataset(self, dataset):
+        self._train_dataset = dataset
+
+    @property
+    def eval_dataset(self):
+        return self._eval_dataset
+
+    @eval_dataset.setter
+    def eval_dataset(self, dataset):
+        self._eval_dataset = dataset
+
+    @abstractmethod
+    def build(self, total_num_steps):
+        pass
+
+    @abstractmethod
+    def get_callbacks(self):
+        pass
+
+    @abstractmethod
+    def get_post_trainer_create_callbacks(self, trainer):
+        """
+        Callbacks added after the trainer is created, usually b/c these need access to the trainer
+        """
+
+
+class HFCausalTrainerBuilder(TrainerBuilderBase):
+    """
+    Build the HuggingFace training args/trainer for Causal models
+    """
+
+    def hook_pre_create_training_args(self, training_arguments_kwargs):
+        # TODO
+        return training_arguments_kwargs
+
+    def hook_post_create_training_args(self, training_arguments):
+        # TODO
+        return training_arguments
+
+    def hook_pre_create_trainer(self, trainer_kwargs, trainer_cls):
+        # TODO
+        return trainer_kwargs, trainer_cls
+
+    def hook_post_create_trainer(self, trainer):
+        # TODO
+        return trainer
+
+    def get_callbacks(self):
+        callbacks = []
+        callbacks.append(GPUStatsCallback(self.cfg))
+        callbacks.append(EvalFirstStepCallback)
+
+        if self.cfg.relora_steps:
+            callbacks.append(ReLoRACallback(self.cfg))
+
+        if (
+            hasattr(self.model, "use_bettertransformer")
+            and self.model.use_bettertransformer is True
+        ):
+            callbacks.append(SaveBetterTransformerModelCallback)
+
+        if self.cfg.use_wandb:
+            callbacks.append(
+                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):
+        callbacks = []
+        if self.cfg.use_wandb and self.cfg.eval_table_size > 0:
+            LogPredictionCallback = log_prediction_callback_factory(
+                trainer, self.tokenizer
+            )
+            callbacks.append(LogPredictionCallback(self.cfg))
+
+        if self.cfg.do_bench_eval:
+            callbacks.append(bench_eval_callback_factory(trainer, self.tokenizer))
+
+        if self.cfg.early_stopping_patience:
+            early_stop_cb = EarlyStoppingCallback(
+                self.cfg.early_stopping_patience,
+            )
+            callbacks.append(early_stop_cb)
+
+        return callbacks
+
+    def _get_trainer_cls(self):
+        if self.cfg.lr_scheduler == "one_cycle" and (
+            self.cfg.fsdp or self.cfg.adapter == "qlora"
+        ):
+            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 = 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
+            else max(min(int(0.005 * total_num_steps), 10), 1)
+        )
+
+        training_arguments_kwargs = {}
+        if self.cfg.bf16 == "full":
+            training_arguments_kwargs["bf16_full_eval"] = True
+        else:
+            training_arguments_kwargs["bf16"] = self.cfg.bf16
+        training_arguments_kwargs["fp16"] = (
+            self.cfg.fp16 and not self.cfg.bf16
+        ) or False
+        training_arguments_kwargs["tf32"] = self.cfg.tf32
+        training_arguments_kwargs["warmup_steps"] = warmup_steps
+        training_arguments_kwargs["logging_steps"] = logging_steps
+
+        if self.cfg.seed:
+            training_arguments_kwargs["seed"] = self.cfg.seed
+
+        if self.cfg.gradient_checkpointing:
+            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:
+                training_arguments_kwargs["fsdp_config"] = dict(self.cfg.fsdp_config)
+
+        # deepspeed
+        if self.cfg.deepspeed:
+            training_arguments_kwargs["deepspeed"] = self.cfg.deepspeed
+
+        if self.cfg.lr_quadratic_warmup is not None:
+            training_arguments_kwargs[
+                "lr_quadratic_warmup"
+            ] = self.cfg.lr_quadratic_warmup
+
+        if self.cfg.adam_beta1:
+            training_arguments_kwargs["adam_beta1"] = self.cfg.adam_beta1
+        if self.cfg.adam_beta2:
+            training_arguments_kwargs["adam_beta2"] = self.cfg.adam_beta2
+        if self.cfg.adam_epsilon:
+            training_arguments_kwargs["adam_epsilon"] = self.cfg.adam_epsilon
+        if self.cfg.max_grad_norm:
+            training_arguments_kwargs["max_grad_norm"] = self.cfg.max_grad_norm
+
+        if self.cfg.hub_model_id:
+            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 is not None:
+            training_arguments_kwargs["save_safetensors"] = self.cfg.save_safetensors
+
+        if self.cfg.sample_packing_eff_est:
+            training_arguments_kwargs[
+                "sample_packing_efficiency"
+            ] = self.cfg.sample_packing_eff_est
+
+        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.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
+        else:
+            # we have an eval set, but no steps defined, default to use epoch
+            training_arguments_kwargs["evaluation_strategy"] = "epoch"
+
+        if self.cfg.save_steps:
+            training_arguments_kwargs["save_strategy"] = "steps"
+            training_arguments_kwargs["save_steps"] = self.cfg.save_steps
+        elif self.cfg.save_strategy:
+            training_arguments_kwargs["save_strategy"] = self.cfg.save_strategy
+        else:
+            # default to saving each epoch if not defined
+            training_arguments_kwargs["save_strategy"] = "epoch"
+
+        if self.cfg.do_bench_eval:
+            training_arguments_kwargs["do_bench_eval"] = self.cfg.do_bench_eval
+            if self.cfg.bench_dataset:
+                training_arguments_kwargs["bench_dataset"] = self.cfg.bench_dataset
+        if self.cfg.metric_for_best_model:
+            training_arguments_kwargs[
+                "metric_for_best_model"
+            ] = self.cfg.metric_for_best_model
+        if self.cfg.greater_is_better:
+            training_arguments_kwargs["greater_is_better"] = self.cfg.greater_is_better
+
+        if self.cfg.torch_compile:
+            if torch.__version__ < "2.1.0":  # pylint: disable=protected-access
+                LOG.warning("torch>=2.1.0 required for torch_compile to work properly")
+            elif torch._dynamo:  # pylint: disable=protected-access
+                torch._dynamo.config.suppress_errors = (  # pylint: disable=protected-access
+                    True
+                )
+                training_arguments_kwargs["torch_compile"] = self.cfg.torch_compile
+                if self.cfg.torch_compile_backend:
+                    training_arguments_kwargs[
+                        "torch_compile_backend"
+                    ] = self.cfg.torch_compile_backend
+
+        # DDP Config
+        if self.cfg.ddp_timeout:
+            training_arguments_kwargs["ddp_timeout"] = self.cfg.ddp_timeout
+        # see https://pytorch.org/docs/stable/generated/torch.nn.parallel.DistributedDataParallel.html
+        if self.cfg.ddp_bucket_cap_mb:
+            training_arguments_kwargs["ddp_bucket_cap_mb"] = self.cfg.ddp_bucket_cap_mb
+        if self.cfg.ddp_broadcast_buffers is not None:
+            training_arguments_kwargs[
+                "ddp_broadcast_buffers"
+            ] = self.cfg.ddp_broadcast_buffers
+
+        # these are all the "standard" kwargs that are def used
+        training_arguments_kwargs["max_steps"] = (
+            total_num_steps if self.cfg.max_steps else -1
+        )
+        training_arguments_kwargs["max_seq_length"] = self.cfg.sequence_len
+        training_arguments_kwargs[
+            "per_device_train_batch_size"
+        ] = self.cfg.micro_batch_size
+        training_arguments_kwargs[
+            "per_device_eval_batch_size"
+        ] = self.cfg.eval_batch_size
+        training_arguments_kwargs[
+            "gradient_accumulation_steps"
+        ] = self.cfg.gradient_accumulation_steps
+        training_arguments_kwargs[
+            "eval_accumulation_steps"
+        ] = self.cfg.gradient_accumulation_steps
+        training_arguments_kwargs["num_train_epochs"] = self.cfg.num_epochs
+        training_arguments_kwargs["learning_rate"] = self.cfg.learning_rate
+        training_arguments_kwargs["output_dir"] = self.cfg.output_dir
+        training_arguments_kwargs["save_total_limit"] = (
+            self.cfg.save_total_limit if self.cfg.save_total_limit else 4
+        )
+        training_arguments_kwargs["load_best_model_at_end"] = (
+            (
+                self.cfg.load_best_model_at_end is not False
+                or self.cfg.early_stopping_patience
+            )
+            and self.cfg.val_set_size > 0
+            and self.cfg.save_steps
+            and self.cfg.eval_steps
+            and self.cfg.save_steps % self.cfg.eval_steps == 0
+        ) or False
+        training_arguments_kwargs["ddp_find_unused_parameters"] = (
+            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
+        training_arguments_kwargs["run_name"] = (
+            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"
+        )
+        training_arguments_kwargs["lr_scheduler_type"] = (
+            self.cfg.lr_scheduler
+            if self.cfg.lr_scheduler
+            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["weight_decay"] = (
+            self.cfg.weight_decay if self.cfg.weight_decay is not None else 0.0
+        )
+        training_arguments_kwargs["sample_packing"] = (
+            self.cfg.sample_packing if self.cfg.sample_packing else False
+        )
+        training_arguments_kwargs["eval_sample_packing"] = (
+            self.cfg.sample_packing
+            if self.cfg.eval_sample_packing is not False
+            else False
+        )
+        training_arguments_kwargs[
+            "sample_packing_seq_len_multiplier"
+        ] = self.cfg.micro_batch_size
+        training_arguments_kwargs["relora_steps"] = self.cfg.relora_steps
+        training_arguments_kwargs["relora_warmup_steps"] = self.cfg.relora_warmup_steps
+        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,
+            )
+        )
+        training_args = self.hook_post_create_training_args(training_args)
+        trainer_kwargs = {}
+
+        if self.cfg.optimizer == "adamw_anyprecision":
+            if Path(self.cfg.torchdistx_path).exists():
+                sys.path.append(self.cfg.torchdistx_path)
+                importlib.import_module("torchdistx")
+
+        data_collator_kwargs = {
+            "padding": True,  # True/"longest" is the default
+        }
+        if self.cfg.pad_to_sequence_len:
+            data_collator_kwargs["pad_to_multiple_of"] = 64 * math.ceil(
+                self.cfg.sequence_len / 64
+            )
+        else:
+            # A100 is best at 64, while others at 8. Let's use the larger so we don't have to check
+            # https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html
+            data_collator_kwargs["pad_to_multiple_of"] = 64
+
+        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),
+            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.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, **kwargs):
+        if training_args.pretraining:
+            return None
+
+        if self.cfg.model_config_type == "mamba":
+            return MambaDataCollator(tokenizer=self.tokenizer)
+
+        return BatchSamplerDataCollatorForSeq2Seq(
+            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
+
+        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
@@ -22,7 +22,7 @@ class TokenizedPromptDataset(Dataset):
     """
     Dataset that returns tokenized prompts from a stream of text files.
         Args:
-            prompt_tokenizer (PromptTokenizingStrategy): The prompt tokenizing method for proccessing the data.
+            prompt_tokenizer (PromptTokenizingStrategy): The prompt tokenizing method for processing the data.
             dataset (dataset.Dataset): Dataset with text files.
     """
 
@@ -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
@@ -55,7 +61,7 @@ class ConstantLengthDataset(IterableDataset):
     """
     Iterable dataset that returns constant length chunks of tokens from stream of text files.
         Args:
-            tokenizer (Tokenizer): The processor used for proccessing the data.
+            tokenizer (Tokenizer): The processor used for processing the data.
             dataset (dataset.Dataset): Dataset with text files.
             seq_length (int): Length of token sequences to return.
     """

src/axolotl/models/mamba/__init__.py

@@ -0,0 +1,12 @@
+"""
+Modeling module for Mamba models
+"""
+
+
+def fix_mamba_attn_for_loss():
+    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/models/phi/modeling_mixformer_sequential.py

@@ -711,12 +711,8 @@ class ParallelBlock(nn.Module):
         self.resid_dropout = nn.Dropout(config.resid_pdrop)
         self.block_idx = block_idx
 
-        self.mixer = MHA(config=config, **mixer, layer_idx=block_idx)
-        mlp_cls = mlp.pop("mlp_cls")
-        if mlp_cls == "fused_mlp":
-            self.mlp = FusedMLP(config=config, **mlp)
-        else:
-            self.mlp = MLP(config=config, **mlp)
+        self.mixer = MHA(config, layer_idx=block_idx)
+        self.mlp = MLP(config)
 
     def forward(
         self,

src/axolotl/monkeypatch/btlm_attn_hijack_flash.py

@@ -7,6 +7,7 @@ import logging
 from typing import Optional, Tuple
 
 import torch
+from accelerate import init_empty_weights
 from flash_attn.flash_attn_interface import flash_attn_func
 from transformers import AutoConfig, AutoModelForCausalLM
 
@@ -17,7 +18,8 @@ def replace_btlm_attn_with_flash_attn(model_name="cerebras/btlm-3b-8k-base"):
     # this is a wonky hack to get the remotely loaded module
     model_config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
     # we need to load the model here in order for modeling_btlm to be available
-    AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)
+    with init_empty_weights():
+        AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)
     module_name = model_config.__class__.__module__.replace(
         ".configuration_btlm", ".modeling_btlm"
     )

src/axolotl/monkeypatch/falcon_attn_hijack_flash.py

@@ -1,101 +0,0 @@
-"""
-Flash Attention monkey patch for Falcon
-
-copied from https://github.com/pacman100/DHS-LLM-Workshop/blob/main/chat_assistant/training/falcon_flash_attn_monkey_patch.py
-"""
-
-from typing import Optional, Tuple
-
-import torch
-import transformers
-from flash_attn import flash_attn_func
-
-
-def forward(
-    self,
-    hidden_states: torch.Tensor,
-    alibi: Optional[torch.Tensor],
-    attention_mask: torch.Tensor,  # pylint: disable=unused-argument
-    layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
-    head_mask: Optional[torch.Tensor] = None,  # pylint: disable=unused-argument
-    use_cache: bool = False,
-    output_attentions: bool = False,  # pylint: disable=unused-argument
-):
-    fused_qkv = self.query_key_value(
-        hidden_states
-    )  # [batch_size, seq_length, 3 x hidden_size]
-    num_kv_heads = (
-        self.num_heads if self.new_decoder_architecture else self.num_kv_heads
-    )
-    # 3 x [batch_size, seq_length, num_heads, head_dim]
-    (
-        query_layer,
-        key_layer,
-        value_layer,
-    ) = self._split_heads(  # pylint: disable=protected-access
-        fused_qkv
-    )
-
-    batch_size, query_length, _, _ = query_layer.shape
-
-    query_layer = query_layer.transpose(1, 2).reshape(
-        batch_size * self.num_heads, query_length, self.head_dim
-    )
-    key_layer = key_layer.transpose(1, 2).reshape(
-        batch_size * num_kv_heads,
-        query_length,
-        self.head_dim,
-    )
-    value_layer = value_layer.transpose(1, 2).reshape(
-        batch_size * num_kv_heads, query_length, self.head_dim
-    )
-
-    past_kv_length = 0 if layer_past is None else layer_past[0].shape[1]
-    query_layer, key_layer = self.maybe_rotary(query_layer, key_layer, past_kv_length)
-
-    if layer_past is not None:
-        past_key, past_value = layer_past
-        # concatenate along seq_length dimension:
-        #  - key: [batch_size * self.num_heads, kv_length, head_dim]
-        #  - value: [batch_size * self.num_heads, kv_length, head_dim]
-        key_layer = torch.cat((past_key, key_layer), dim=1)
-        value_layer = torch.cat((past_value, value_layer), dim=1)
-
-    # unused
-    # _, kv_length, _ = key_layer.shape
-    if use_cache:
-        present = (key_layer, value_layer)
-    else:
-        present = None
-    # unused
-    # attention_mask_float = (attention_mask * 1.0).masked_fill(attention_mask, float("-1e9")).to(query_layer.dtype)
-    query_layer_ = (
-        query_layer.reshape(batch_size, self.num_heads, -1, self.head_dim)
-        .transpose(1, 2)
-        .to(torch.bfloat16)
-    )
-    key_layer_ = (
-        key_layer.reshape(batch_size, num_kv_heads, -1, self.head_dim)
-        .transpose(1, 2)
-        .to(torch.bfloat16)
-    )
-    value_layer_ = (
-        value_layer.reshape(batch_size, num_kv_heads, -1, self.head_dim)
-        .transpose(1, 2)
-        .to(torch.bfloat16)
-    )
-
-    if alibi is not None:
-        raise ValueError("`alibi` is not supported when `use_flash_attn` is True")
-
-    # below output will have shape (batch_size, seqlen, nheads, headdim)
-    attn_output = flash_attn_func(query_layer_, key_layer_, value_layer_, causal=True)
-    attn_output = attn_output.reshape(
-        batch_size, query_length, self.num_heads * self.head_dim
-    )
-    output_tensor = self.dense(attn_output)
-    return output_tensor, present
-
-
-def replace_falcon_attn_with_flash_attn():
-    transformers.models.falcon.modeling_falcon.FalconAttention.forward = forward

src/axolotl/monkeypatch/fastchat_conversation_turns.py

@@ -0,0 +1,212 @@
+"""
+monkeypatch to add a get_turns method
+"""
+
+import logging
+from typing import Generator, Tuple
+
+from fastchat.conversation import SeparatorStyle
+
+LOG = logging.getLogger("axolotl.monkeypatch.fastchat_conversation_turns")
+
+
+def get_prompt(self) -> str:
+    ret = ""
+    for role, msg in self.get_turns():
+        ret += role + msg
+    return ret
+
+
+def get_turns(  # pylint: disable=too-many-return-statements
+    self,
+) -> Generator[Tuple[str, str], None, None]:
+    """Get the prompt for generation."""
+    system_prompt = self.system_template.format(system_message=self.system_message)
+    if self.sep_style == SeparatorStyle.ADD_COLON_SINGLE:
+        yield "", system_prompt + self.sep
+        for role, message in self.messages:
+            if message:
+                yield role + ": ", message + self.sep
+            else:
+                yield role + ":", ""
+        return
+    if self.sep_style == SeparatorStyle.ADD_COLON_TWO:
+        seps = [self.sep, self.sep2]
+        yield "", system_prompt + seps[0]
+        for i, (role, message) in enumerate(self.messages):
+            if message:
+                yield role + ": ", message + seps[i % 2]
+            else:
+                yield role + ":", ""
+        return
+    if self.sep_style == SeparatorStyle.ADD_COLON_SPACE_SINGLE:
+        yield "", system_prompt + self.sep
+        for role, message in self.messages:
+            if message:
+                yield role + ": ", message + self.sep
+            else:
+                yield role + ": ", ""  # must be end with a space
+        return
+    if self.sep_style == SeparatorStyle.ADD_NEW_LINE_SINGLE:
+        yield "", "" if system_prompt == "" else system_prompt + self.sep
+        for role, message in self.messages:
+            if message:
+                yield role + "\n", message + self.sep
+            else:
+                yield role + "\n", ""
+        return
+    if self.sep_style == SeparatorStyle.NO_COLON_SINGLE:
+        yield "", system_prompt
+        for role, message in self.messages:
+            if message:
+                yield role, message + self.sep
+            else:
+                yield role, ""
+        return
+    if self.sep_style == SeparatorStyle.NO_COLON_TWO:
+        seps = [self.sep, self.sep2]
+        yield "", system_prompt
+        for i, (role, message) in enumerate(self.messages):
+            if message:
+                yield role, message + seps[i % 2]
+            else:
+                yield role, ""
+        return
+    if self.sep_style == SeparatorStyle.RWKV:
+        yield "", system_prompt
+        for i, (role, message) in enumerate(self.messages):
+            if message:
+                yield role + ": ", message.replace("\r\n", "\n").replace(
+                    "\n\n", "\n"
+                ) + "\n\n"
+            else:
+                yield role + ":", ""
+        return
+    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
+        for i, (role, message) in enumerate(self.messages):
+            if message:
+                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
+    if self.sep_style == SeparatorStyle.CHATGLM:
+        # source: https://huggingface.co/THUDM/chatglm-6b/blob/1d240ba371910e9282298d4592532d7f0f3e9f3e/modeling_chatglm.py#L1302-L1308
+        # source2: https://huggingface.co/THUDM/chatglm2-6b/blob/e186c891cf64310ac66ef10a87e6635fa6c2a579/modeling_chatglm.py#L926
+        round_add_n = 1 if self.name == "chatglm2" else 0
+        if system_prompt:
+            yield "", system_prompt + self.sep
+
+        for i, (role, message) in enumerate(self.messages):
+            if i % 2 == 0:
+                yield "", f"[Round {i//2 + round_add_n}]{self.sep}"
+
+            if message:
+                yield f"{role}：", f"{message}{self.sep}"
+            else:
+                yield f"{role}：", ""
+        return
+    if self.sep_style == SeparatorStyle.CHATML:
+        yield "", "" if system_prompt == "" else system_prompt + self.sep + "\n"
+        for role, message in self.messages:
+            if message:
+                yield role + "\n", message + self.sep + "\n"
+            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]
+        yield "", system_prompt
+        for i, (role, message) in enumerate(self.messages):
+            prefix = "<s>" if i % 2 == 0 else ""
+            if message:
+                yield prefix + role + ":", message + seps[i % 2] + "\n"
+            else:
+                yield role + ":", ""
+        return
+    if self.sep_style == SeparatorStyle.DOLLY:
+        seps = [self.sep, self.sep2]
+        yield "", system_prompt
+        for i, (role, message) in enumerate(self.messages):
+            if message:
+                suffix = "\n\n" if i % 2 == 1 else ""
+                yield role + ":\n", message + seps[i % 2] + suffix
+            else:
+                yield role + ":\n", ""
+        return
+    if self.sep_style == SeparatorStyle.PHOENIX:
+        yield "", system_prompt
+        for role, message in self.messages:
+            if message:
+                yield role + ": ", "<s>" + message + "</s>"
+            else:
+                yield role + ": " + "<s>", ""
+        return
+    if self.sep_style == SeparatorStyle.ROBIN:
+        yield "", system_prompt + self.sep
+        for role, message in self.messages:
+            if message:
+                yield role + ":\n", message + self.sep
+            else:
+                yield role + ":\n", ""
+        return
+    if self.sep_style == SeparatorStyle.FALCON_CHAT:
+        if self.system_message:
+            yield "", system_prompt + self.sep
+        for role, message in self.messages:
+            if message:
+                yield role + ": ", message + self.sep
+            else:
+                yield role + ":", ""
+    else:
+        raise ValueError(f"Invalid style: {self.sep_style}")
+
+
+def add_get_turns_to_conversation():
+    import fastchat.conversation
+
+    fastchat.conversation.Conversation.get_turns = get_turns
+    fastchat.conversation.Conversation.get_prompt = get_prompt

src/axolotl/monkeypatch/llama_attn_hijack_flash.py

@@ -12,15 +12,19 @@ import torch.nn.functional as F
 import transformers
 from einops import rearrange
 from flash_attn.bert_padding import pad_input, unpad_input
-from torch import nn
-from transformers import LlamaConfig
 from transformers.modeling_outputs import BaseModelOutputWithPast
+from transformers.models.llama.modeling_llama import LlamaAttention
 from transformers.models.llama.modeling_llama import (
     LlamaDecoderLayer as OriginalLlamaDecoderLayer,
 )
-from transformers.models.llama.modeling_llama import apply_rotary_pos_emb, repeat_kv
+from transformers.models.llama.modeling_llama import (
+    LlamaMLP,
+    apply_rotary_pos_emb,
+    repeat_kv,
+)
+from xformers.ops import SwiGLU
 
-from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
+from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids, set_module_name
 
 try:
     from flash_attn.flash_attn_interface import (  # pylint: disable=ungrouped-imports
@@ -40,7 +44,33 @@ except ImportError:
 LOG = logging.getLogger("axolotl")
 
 
-def replace_llama_attn_with_flash_attn(packed: Optional[bool] = False):
+def replace_llama_mlp_with_swiglu(model):
+    for name, module in model.named_modules():
+        if isinstance(module, LlamaMLP):
+            mlp = FusedMLP(
+                module.config, module.gate_proj, module.up_proj, module.down_proj
+            )
+            set_module_name(model, name, mlp)
+
+
+def replace_llama_qkv_with_fused(model):
+    for name, module in model.named_modules():
+        if isinstance(module, LlamaAttention):
+            qkv = FusedAttention(
+                module.config,
+                module.q_proj,
+                module.k_proj,
+                module.v_proj,
+                module.o_proj,
+            )
+            set_module_name(model, name, qkv)
+
+
+def replace_llama_attn_with_flash_attn(
+    packed: Optional[bool] = False,
+    cross_entropy: Optional[bool] = False,
+    rms_norm: Optional[bool] = False,
+):
     transformers.models.llama.modeling_llama.LlamaModel._prepare_decoder_attention_mask = (  # pylint: disable=protected-access
         _prepare_decoder_attention_mask
     )
@@ -51,46 +81,123 @@ def replace_llama_attn_with_flash_attn(packed: Optional[bool] = False):
             llama_model_forward
         )
 
-    try:
-        from flash_attn.losses.cross_entropy import CrossEntropyLoss
+    # skip only if explicitly disabled
+    if cross_entropy:
+        try:
+            from flash_attn.losses.cross_entropy import CrossEntropyLoss
 
-        LOG.info("patching with flash_attn.losses.cross_entropy")
-        transformers.models.llama.modeling_llama.CrossEntropyLoss = partial(
-            CrossEntropyLoss, inplace_backward=True
+            LOG.info("patching with flash_attn.losses.cross_entropy")
+            transformers.models.llama.modeling_llama.CrossEntropyLoss = partial(
+                CrossEntropyLoss, inplace_backward=True
+            )
+        except ImportError:
+            LOG.info(
+                "optimized flash-attention CrossEntropyLoss not found (run `pip install 'git+https://github.com/Dao-AILab/flash-attention.git#egg=xentropy_cuda_lib&subdirectory=csrc/xentropy'`)"
+            )
+
+    # skip only if explicitly disabled
+    if rms_norm:
+        try:
+            from flash_attn.ops.rms_norm import RMSNorm
+
+            class LlamaRMSNorm(RMSNorm):
+                """Patched LLamaRMSNorm"""
+
+                def __init__(self, hidden_size, eps=1e-6):
+                    super().__init__(hidden_size, eps=eps)
+
+            LOG.info("patching with flash_attn.ops.rms_norm")
+            transformers.models.llama.modeling_llama.LlamaRMSNorm = LlamaRMSNorm
+        except ImportError:
+            LOG.info(
+                "optimized flash-attention RMSNorm not found (run `pip install 'git+https://github.com/Dao-AILab/flash-attention.git#egg=dropout_layer_norm&subdirectory=csrc/layer_norm'`)"
+            )
+
+
+class FusedAttention(LlamaAttention):
+    """
+    Fused QKV Attention layer for incrementally improved training efficiency
+    """
+
+    def __init__(
+        self,
+        config,
+        q: torch.nn.Linear,  # pylint: disable=invalid-name
+        k: torch.nn.Linear,  # pylint: disable=invalid-name
+        v: torch.nn.Linear,  # pylint: disable=invalid-name
+        o: torch.nn.Linear,  # pylint: disable=invalid-name
+    ):
+        super().__init__(config)
+        self.config = config
+        self.init_device = next(iter(q.state_dict().values())).device
+
+        # define equivalent fused qkv projection
+        self.out_features: List[int] = [q.out_features, k.out_features, v.out_features]
+        self.qkv_proj = torch.nn.Linear(
+            q.in_features, sum(self.out_features), device=self.init_device, bias=False
         )
-    except ImportError:
-        LOG.info(
-            "optimized flash-attention CrossEntropyLoss not found (run `pip install 'git+https://github.com/Dao-AILab/flash-attention.git#egg=xentropy_cuda_lib&subdirectory=csrc/xentropy'`)"
+        self.o_proj = o
+
+        # overwrite initialized weights with pretrained weights
+        self.qkv_proj.weight.data = torch.cat(
+            (q.weight.data, k.weight.data, v.weight.data), dim=0
         )
 
-    try:
-        from flash_attn.ops.rms_norm import RMSNorm
-
-        class LlamaRMSNorm(RMSNorm):
-            """Patched LLamaRMSNorm"""
-
-            def __init__(self, hidden_size, eps=1e-6):
-                super().__init__(hidden_size, eps=eps)
-
-        LOG.info("patching with flash_attn.ops.rms_norm")
-        transformers.models.llama.modeling_llama.LlamaRMSNorm = LlamaRMSNorm
-    except ImportError:
-        LOG.info(
-            "optimized flash-attention RMSNorm not found (run `pip install 'git+https://github.com/Dao-AILab/flash-attention.git#egg=dropout_layer_norm&subdirectory=csrc/layer_norm'`)"
+    def _post_training(self, model, name):
+        q_proj, k_proj, v_proj = torch.split(
+            self.qkv_proj.weight.data, self.out_features, dim=0
         )
 
+        new_attn = LlamaAttention(self.config)
+        new_attn.q_proj.weight.data = q_proj
+        new_attn.k_proj.weight.data = k_proj
+        new_attn.v_proj.weight.data = v_proj
+        new_attn.o_proj.weight.data = self.o_proj.weight.data
 
-class GaussianDropout(nn.Module):
-    def __init__(self, p=0.5):
-        super(GaussianDropout, self).__init__()
-        if p <= 0 or p >= 1:
-            raise Exception("p value should accomplish 0 < p < 1")
-        self.p = p
+        set_module_name(model, name, new_attn)
 
-    def forward(self, x):
-        stddev = (self.p / (1.0 - self.p)) ** 0.5
-        epsilon = torch.randn_like(x) * stddev
-        return x * epsilon
+
+class FusedMLP(torch.nn.Module):
+    """
+    Fused MLP layer for incrementally improved training efficiency
+    """
+
+    def __init__(
+        self,
+        config,
+        gate_proj: torch.nn.Linear,
+        up_proj: torch.nn.Linear,
+        down_proj: torch.nn.Linear,
+    ):
+        super().__init__()
+        self.config = config
+        self.swiglu = SwiGLU(
+            in_features=config.hidden_size,
+            hidden_features=config.intermediate_size,
+            bias=False,
+            _pack_weights=True,
+        )
+        # overwrite initialized weights with pretrained weights
+        self.swiglu.w12.weight.data = torch.cat(
+            (gate_proj.weight.data, up_proj.weight.data), dim=0
+        )
+        self.swiglu.w3.weight.data = down_proj.weight.data
+
+    def _post_training(self, model, name):
+        w1, w2 = torch.split(  # pylint: disable=invalid-name
+            self.swiglu.w12.weight.data, self.config.intermediate_size, dim=0
+        )
+
+        # Assign the split weights back to the original layers
+        new_mlp = LlamaMLP(self.config)
+        new_mlp.gate_proj.weight.data = w1
+        new_mlp.up_proj.weight.data = w2
+        new_mlp.down_proj.weight.data = self.swiglu.w3.weight.data
+
+        set_module_name(model, name, new_mlp)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:  # pylint: disable=invalid-name
+        return self.swiglu(x)
 
 
 # Disable the transformation of the attention mask in LlamaModel as the flash attention
@@ -114,6 +221,7 @@ def flashattn_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
     cu_seqlens: Optional[torch.Tensor] = None,
     max_seqlen: Optional[torch.Tensor] = None,
 ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
@@ -153,9 +261,14 @@ def flashattn_forward(
         value_states = torch.cat(value_states, dim=-1)
 
     else:
-        query_states = self.q_proj(hidden_states)
-        key_states = self.k_proj(hidden_states)
-        value_states = self.v_proj(hidden_states)
+        if isinstance(self, FusedAttention):
+            query_states, key_states, value_states = self.qkv_proj(hidden_states).split(
+                self.out_features, dim=-1
+            )
+        else:
+            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
@@ -208,6 +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(
@@ -217,7 +332,12 @@ def flashattn_forward(
         qkv = rearrange(qkv, "b s ... -> (b s) ...")
 
         output = flash_attn_varlen_qkvpacked_func(
-            qkv, cu_seqlens, max_seqlen, dropout_p=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:
@@ -240,7 +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,
         )
@@ -253,6 +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:
@@ -285,7 +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,
             )
@@ -491,6 +612,13 @@ def llama_model_forward(
             dtype=torch.bool,
             device=inputs_embeds.device,
         )
+        padding_mask = None
+    else:
+        if 0 in attention_mask:
+            padding_mask = attention_mask
+        else:
+            padding_mask = None
+
     attention_mask = (
         self._prepare_decoder_attention_mask(  # pylint: disable=protected-access
             attention_mask,
@@ -525,7 +653,9 @@ def llama_model_forward(
             def create_custom_forward(module):
                 def custom_forward(*inputs):
                     # None for past_key_value
-                    return module(*inputs)
+                    return module(
+                        *inputs,
+                    )
 
                 return custom_forward
 
@@ -534,9 +664,10 @@ def llama_model_forward(
                 hidden_states,
                 attention_mask,
                 position_ids,
-                None,
+                past_key_value,
                 output_attentions,
                 None,
+                padding_mask,
                 cu_seqlens,
                 max_seqlen,
             )
@@ -548,6 +679,7 @@ def llama_model_forward(
                 past_key_value=past_key_value,
                 output_attentions=output_attentions,
                 use_cache=use_cache,
+                padding_mask=padding_mask,
                 cu_seqlens=cu_seqlens,
                 max_seqlen=max_seqlen,
             )
@@ -586,15 +718,6 @@ class LlamaDecoderLayer(OriginalLlamaDecoderLayer):
     patched version of LlamaDecoderLayer to pass through the precalculated cu_seqlens
     """
 
-    def __init__(self, config: LlamaConfig):
-        super(LlamaDecoderLayer, self).__init__(config)
-        self.attn_dropout = None
-        self.mlp_dropout = None
-        if config.dropout_attn:
-            self.attn_dropout = GaussianDropout(p=config.dropout_attn)
-        if config.dropout_mlp:
-            self.mlp_dropout = GaussianDropout(p=config.dropout_mlp)
-
     def forward(
         self,
         hidden_states: torch.Tensor,
@@ -603,6 +726,7 @@ class LlamaDecoderLayer(OriginalLlamaDecoderLayer):
         past_key_value: Optional[Tuple[torch.Tensor]] = None,
         output_attentions: Optional[bool] = False,
         use_cache: Optional[bool] = False,
+        padding_mask: Optional[torch.LongTensor] = None,
         cu_seqlens: Optional[torch.Tensor] = None,
         max_seqlen: Optional[torch.Tensor] = None,
     ) -> Tuple[
@@ -635,19 +759,16 @@ class LlamaDecoderLayer(OriginalLlamaDecoderLayer):
             past_key_value=past_key_value,
             output_attentions=output_attentions,
             use_cache=use_cache,
+            padding_mask=padding_mask,
             cu_seqlens=cu_seqlens,
             max_seqlen=max_seqlen,
         )
-        if self.training and self.attn_dropout:
-            hidden_states = self.attn_dropout(hidden_states)
         hidden_states = residual + hidden_states
 
         # Fully Connected
         residual = hidden_states
         hidden_states = self.post_attention_layernorm(hidden_states)
         hidden_states = self.mlp(hidden_states)
-        if self.training and self.mlp_dropout:
-            hidden_states = self.mlp_dropout(hidden_states)
         hidden_states = residual + hidden_states
 
         outputs = (hidden_states,)

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/mistral_attn_hijack_flash.py

@@ -0,0 +1,643 @@
+"""Flash attention monkey patch for mistral model"""
+# pylint: disable=duplicate-code
+
+import logging
+from typing import List, Optional, Tuple, Union
+
+import torch
+import transformers
+from einops import rearrange
+from flash_attn.bert_padding import pad_input, unpad_input
+from flash_attn.flash_attn_interface import (  # pylint: disable=ungrouped-imports
+    flash_attn_kvpacked_func,
+    flash_attn_varlen_kvpacked_func,
+    flash_attn_varlen_qkvpacked_func,
+)
+from transformers.modeling_outputs import BaseModelOutputWithPast
+from transformers.models.mistral.modeling_mistral import (
+    MistralAttention as OriginalMistralAttention,
+)
+from transformers.models.mistral.modeling_mistral import (
+    MistralDecoderLayer as OriginalMistralDecoderLayer,
+)
+from transformers.models.mistral.modeling_mistral import apply_rotary_pos_emb, repeat_kv
+
+from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
+
+LOG = logging.getLogger("axolotl.monkeypatch.mistral")
+
+
+def replace_mistral_attn_with_flash_attn(
+    packed: Optional[bool] = False,
+):
+    transformers.models.mistral.modeling_mistral.MistralModel._prepare_decoder_attention_mask = (  # pylint: disable=protected-access
+        _prepare_decoder_attention_mask
+    )
+    transformers.models.mistral.modeling_mistral.MistralAttention.forward = (
+        flashattn_forward
+    )
+    if packed:
+        transformers.models.mistral.modeling_mistral.MistralDecoderLayer = (
+            MistralDecoderLayer
+        )
+        transformers.models.mistral.modeling_mistral.MistralModel.forward = (
+            mistral_model_forward
+        )
+
+
+@torch.jit.script
+def _make_sliding_window_causal_mask(
+    bsz: int,
+    tgt_len: int,
+    dtype: torch.dtype,
+    device: torch.device,
+    past_key_values_length: int = 0,
+    sliding_window: int = 4096,
+):
+    """
+    Make causal mask used for sliding window attention
+    """
+    tensor = torch.full(
+        (tgt_len, tgt_len),
+        fill_value=1,
+        device=device,
+    )
+    mask = torch.tril(tensor, diagonal=0)
+    # make the mask banded to account for sliding window
+    # NOTE: HF implementation is wrong as of 14-10-2023 for torch.triu, needs +1
+    mask = torch.triu(mask, diagonal=-sliding_window + 1)
+    mask = torch.log(mask).to(dtype)
+
+    if past_key_values_length > 0:
+        mask = torch.cat(
+            [
+                torch.zeros(
+                    tgt_len, past_key_values_length, dtype=dtype, device=device
+                ),
+                mask,
+            ],
+            dim=-1,
+        )
+    return mask[None, None, :, :].expand(
+        bsz, 1, tgt_len, tgt_len + past_key_values_length
+    )
+
+
+# Disable the transformation of the attention mask in LlamaModel as the flash attention
+# requires the attention mask to be the same as the key_padding_mask
+def _prepare_decoder_attention_mask(
+    self,
+    attention_mask,
+    input_shape,
+    inputs_embeds,
+    past_key_values_length,
+    sliding_window,
+):  # pylint: disable=unused-argument
+    # [bsz, seq_len]
+    if attention_mask is None:
+        return attention_mask
+
+    # NOTE: attention mask and sliding masks are only broadcastable in certain scenarios.
+    # Without attention_mask.shape[0] == 1, error will trigger after eval loss but only when wandb is enabled.
+    if input_shape[-1] > 1 and attention_mask.shape[0] == 1:
+        sliding_window_mask = _make_sliding_window_causal_mask(
+            bsz=input_shape[0],
+            tgt_len=input_shape[1],
+            dtype=inputs_embeds.dtype,
+            device=inputs_embeds.device,
+            past_key_values_length=past_key_values_length,
+            sliding_window=sliding_window,
+        )
+        attention_mask = attention_mask + sliding_window_mask
+    else:
+        LOG.info("skipping sliding window mask, not broadcastable with attention mask")
+
+    return attention_mask
+
+
+def flashattn_forward(
+    self: OriginalMistralAttention,
+    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: bool = False,
+    use_cache: bool = False,
+    cu_seqlens: Optional[torch.Tensor] = None,
+    max_seqlen: Optional[torch.Tensor] = None,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    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:
+        kv_seq_len += past_key_value[0].shape[-2]
+    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
+    )
+
+    use_sliding_windows = (
+        hasattr(self.config, "sliding_window") is not None
+        and kv_seq_len > self.config.sliding_window
+    )
+
+    if use_sliding_windows:
+        window_size = (self.config.sliding_window, self.config.sliding_window)
+    else:
+        window_size = (-1, -1)
+
+    if past_key_value is not None:
+        # Activate slicing cache only if the config has a value `sliding_windows` attribute
+        if (
+            hasattr(self.config, "sliding_window")
+            and kv_seq_len > self.config.sliding_window
+        ):
+            slicing_tokens = kv_seq_len - self.config.sliding_window
+
+            past_key = past_key_value[0]
+            past_value = past_key_value[1]
+
+            past_key = past_key[:, :, slicing_tokens:, :].contiguous()
+            past_value = past_value[:, :, slicing_tokens:, :].contiguous()
+
+            if past_key.shape[-2] != self.config.sliding_window - 1:
+                raise ValueError(
+                    f"past key much have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got"
+                    f" {past_key.shape}"
+                )
+
+            past_key_value = (past_key, past_value) if use_cache else None
+
+        if past_key_value is not None:
+            key_states = torch.cat([past_key_value[0], key_states], dim=2)
+            value_states = torch.cat([past_key_value[1], value_states], dim=2)
+
+    past_key_value = (key_states, value_states) if use_cache else None
+
+    # 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 self.training:
+        # during training q,k,v always have same seqlen
+        assert key_states.shape == query_states.shape
+        is_causal = True
+    else:
+        # turn off FA causal mask after first inference autoregressive iteration
+        # 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(
+            [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) ...")
+
+        output = flash_attn_varlen_qkvpacked_func(
+            qkv,
+            cu_seqlens,
+            max_seqlen,
+            dropout_p=dropout_rate,
+            softmax_scale=None,
+            causal=True,
+            window_size=window_size,
+        )
+        output = rearrange(output, "(b s) ... -> b s ...", b=bsz)
+    elif query_states.shape == key_states.shape:
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+        qkv_unpad, cu_seqlens_q, max_seqlen_q, _, output_pad_fn = generate_qkv(
+            query_states,
+            key_states,
+            value_states,
+            qkvpacked=True,
+            # We have disabled _prepare_decoder_attention_mask in LlamaModel
+            # the attention_mask should be the same as the key_padding_mask
+            key_padding_mask=attention_mask,
+            query_padding_mask=attention_mask[:, -query_states.size(1) :]
+            if attention_mask is not None
+            else None,
+        )
+        output_unpad = flash_attn_varlen_qkvpacked_func(
+            qkv_unpad,
+            cu_seqlens_q,
+            max_seqlen_q,
+            dropout_p=dropout_rate,
+            softmax_scale=None,
+            causal=is_causal,
+            window_size=window_size,
+        )
+        output = output_pad_fn(output_unpad)
+    else:
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+        if attention_mask is None or attention_mask.all().item():
+            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,
+            )
+        else:
+            (  # pylint: disable=unbalanced-tuple-unpacking
+                q_unpad,
+                kv_unpad,
+                cu_seqlens_q,
+                cu_seqlens_k,
+                max_seqlen_q,
+                max_seqlen_k,
+                _,
+                _,
+                output_pad_fn,
+            ) = generate_qkv(
+                query_states,
+                key_states,
+                value_states,
+                kvpacked=True,
+                key_padding_mask=attention_mask,
+                query_padding_mask=attention_mask[:, -query_states.size(1) :]
+                if attention_mask is not None
+                else None,
+            )
+            if q_unpad.dtype != kv_unpad.dtype:
+                kv_unpad = kv_unpad.to(q_unpad.dtype)
+            output_unpad = flash_attn_varlen_kvpacked_func(
+                q_unpad,
+                kv_unpad,
+                cu_seqlens_q,
+                cu_seqlens_k,
+                max_seqlen_q,
+                max_seqlen_k,
+                dropout_p=dropout_rate,
+                softmax_scale=None,
+                causal=is_causal,
+                window_size=window_size,
+            )
+            output = output_pad_fn(output_unpad)
+
+    attn_output = output
+    if attn_output.size() != (bsz, q_len, self.num_heads, self.head_dim):
+        raise ValueError(
+            f"`attn_output` should be of size {(bsz, q_len, self.num_heads, self.head_dim)}, but is"
+            f" {attn_output.size()}"
+        )
+    attn_output = rearrange(attn_output, "b s h d -> b s (h d)")
+
+    attn_output = self.o_proj(attn_output)
+
+    if not output_attentions:
+        attn_weights = None
+
+    return attn_output, attn_weights, past_key_value
+
+
+# based on https://github.com/Dao-AILab/flash-attention/blob/364a5b/tests/test_flash_attn.py#L38
+def generate_qkv(
+    q,
+    k,
+    v,
+    query_padding_mask=None,
+    key_padding_mask=None,
+    kvpacked=False,
+    qkvpacked=False,
+):  # pylint: disable=invalid-name,unnecessary-lambda-assignment
+    """
+    Arguments:
+        q: (batch_size, seqlen_q, nheads, d)
+        k: (batch_size, seqlen_k, nheads_k, d)
+        v: (batch_size, seqlen_k, nheads_k, d)
+        query_padding_mask: (batch_size, seqlen), bool
+        key_padding_mask: (batch_size, seqlen), bool
+    """
+    assert not (kvpacked and qkvpacked)
+    batch_size, seqlen_q, nheads, d = q.shape
+    _, seqlen_k, nheads_k, _ = k.shape
+    assert k.shape == (batch_size, seqlen_k, nheads_k, d)
+    assert v.shape == (batch_size, seqlen_k, nheads_k, d)
+
+    if query_padding_mask is not None:
+        q_unpad, indices_q, cu_seqlens_q, max_seqlen_q = unpad_input(
+            q, query_padding_mask
+        )
+
+        output_pad_fn = lambda output_unpad: pad_input(  # noqa: E731
+            output_unpad, indices_q, batch_size, seqlen_q
+        )
+
+    else:
+        q_unpad = rearrange(q, "b s h d -> (b s) h d")
+        cu_seqlens_q = torch.arange(
+            0,
+            (batch_size + 1) * seqlen_q,
+            step=seqlen_q,
+            dtype=torch.int32,
+            device=q_unpad.device,
+        )
+        max_seqlen_q = seqlen_q
+
+        output_pad_fn = lambda output_unpad: rearrange(  # noqa: E731
+            output_unpad, "(b s) h d -> b s h d", b=batch_size
+        )
+
+    if key_padding_mask is not None:
+        k_unpad, _, cu_seqlens_k, max_seqlen_k = unpad_input(k, key_padding_mask)
+        v_unpad, _, _, _ = unpad_input(v, key_padding_mask)
+    else:
+        k_unpad = rearrange(k, "b s h d -> (b s) h d")
+        v_unpad = rearrange(v, "b s h d -> (b s) h d")
+        cu_seqlens_k = torch.arange(
+            0,
+            (batch_size + 1) * seqlen_k,
+            step=seqlen_k,
+            dtype=torch.int32,
+            device=k_unpad.device,
+        )
+        max_seqlen_k = seqlen_k
+
+    if qkvpacked:
+        assert nheads == nheads_k
+        qkv_unpad = torch.stack([q_unpad, k_unpad, v_unpad], dim=1)
+        qkv = torch.stack([q, k, v], dim=2)
+        return (qkv_unpad, cu_seqlens_q, max_seqlen_q, qkv, output_pad_fn)
+
+    if kvpacked:
+        kv_unpad = torch.stack([k_unpad, v_unpad], dim=1)
+        kv = torch.stack([k, v], dim=2)
+        return (
+            q_unpad,
+            kv_unpad,
+            cu_seqlens_q,
+            cu_seqlens_k,
+            max_seqlen_q,
+            max_seqlen_k,
+            q,
+            kv,
+            output_pad_fn,
+        )
+
+    return (
+        q_unpad,
+        k_unpad,
+        v_unpad,
+        cu_seqlens_q,
+        cu_seqlens_k,
+        max_seqlen_q,
+        max_seqlen_k,
+        q,
+        k,
+        v,
+        output_pad_fn,
+    )
+
+
+def mistral_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,
+    return_dict: Optional[bool] = None,
+) -> Union[Tuple, BaseModelOutputWithPast]:
+    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
+    )
+    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"
+        )
+
+    seq_length_with_past = seq_length
+    past_key_values_length = 0
+
+    if past_key_values is not None:
+        past_key_values_length = past_key_values[0][0].shape[2]
+        seq_length_with_past = seq_length_with_past + past_key_values_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)
+    # embed positions
+    if attention_mask is None:
+        attention_mask = torch.ones(
+            (batch_size, seq_length_with_past),
+            dtype=torch.bool,
+            device=inputs_embeds.device,
+        )
+    attention_mask = (
+        self._prepare_decoder_attention_mask(  # pylint: disable=protected-access
+            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:
+            transformers.logger.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
+    next_decoder_cache = () if use_cache else None
+
+    for idx, decoder_layer in enumerate(self.layers):
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+        if self.gradient_checkpointing and self.training:
+
+            def create_custom_forward(module):
+                def custom_forward(*inputs):
+                    # None for past_key_value
+                    return module(*inputs)
+
+                return custom_forward
+
+            layer_outputs = torch.utils.checkpoint.checkpoint(
+                create_custom_forward(decoder_layer),
+                hidden_states,
+                attention_mask,
+                position_ids,
+                past_key_value,
+                output_attentions,
+                None,
+                cu_seqlens,
+                max_seqlen,
+            )
+        else:
+            layer_outputs = decoder_layer(
+                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 = 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],)
+
+    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 = next_decoder_cache if use_cache else None
+    if not return_dict:
+        return tuple(
+            v
+            for v in [hidden_states, next_cache, all_hidden_states, all_self_attns]
+            if v is not None
+        )
+    return BaseModelOutputWithPast(
+        last_hidden_state=hidden_states,
+        past_key_values=next_cache,
+        hidden_states=all_hidden_states,
+        attentions=all_self_attns,
+    )
+
+
+class MistralDecoderLayer(OriginalMistralDecoderLayer):
+    """
+    patched version of MistralDecoderLayer to pass through the precalculated cu_seqlens
+    """
+
+    def 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,
+        use_cache: Optional[bool] = False,
+        cu_seqlens: Optional[torch.Tensor] = None,
+        max_seqlen: Optional[torch.Tensor] = None,
+    ) -> Tuple[
+        torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]
+    ]:
+        """
+        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, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            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`).
+            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+            cu_seqlens (`torch.Tensor`, *optional*) cumulative sequence len when packing
+        """
+
+        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 = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if use_cache:
+            outputs += (present_key_value,)
+
+        return outputs

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/stablelm_attn_hijack_flash.py

@@ -0,0 +1,415 @@
+# coding=utf-8
+# Copyright 2023 Stability AI, EleutherAI, and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# This code is based off the following work:
+# https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/modeling_llama.py
+# https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt_neox/modeling_gpt_neox.py
+""" PyTorch StableLM Epoch model. """
+import importlib
+import math
+from typing import Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from accelerate import init_empty_weights
+from einops import rearrange
+from flash_attn.flash_attn_interface import (  # pylint: disable=ungrouped-imports
+    flash_attn_varlen_qkvpacked_func,
+)
+from torch import nn
+from transformers import AutoConfig, AutoModelForCausalLM
+from transformers.modeling_outputs import BaseModelOutputWithPast
+from transformers.utils import logging
+
+from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
+
+logger = logging.get_logger(__name__)
+
+
+def replace_stablelm_attn_with_flash_attn(model_name="stabilityai/stablelm-3b-4e1t"):
+    # this is a wonky hack to get the remotely loaded module
+    model_config = AutoConfig.from_pretrained(model_name, trust_remote_code=True)
+    # we need to load the model here in order for modeling_stablelm_epoch to be available
+    with init_empty_weights():
+        AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)
+    module_name = model_config.__class__.__module__.replace(
+        ".configuration_stablelm_epoch", ".modeling_stablelm_epoch"
+    )
+    modeling_stablelm = importlib.import_module(module_name)
+    modeling_stablelm.Attention.forward = (  # pylint: disable=protected-access
+        flashattn_attn
+    )
+    modeling_stablelm.StableLMEpochModel.forward = (  # pylint: disable=protected-access
+        stablelm_model_forward
+    )
+    modeling_stablelm.DecoderLayer.forward = (  # pylint: disable=protected-access
+        decoder_layer_forward
+    )
+
+
+def rotate_half(x: torch.Tensor):
+    """Rotates half the hidden dims of the input."""
+    # pylint: disable=invalid-name
+    x1, x2 = torch.chunk(x, 2, dim=-1)
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
+    # The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
+    # pylint: disable=invalid-name
+    cos = cos.squeeze(1).squeeze(0)  # [seq_len, dim]
+    sin = sin.squeeze(1).squeeze(0)  # [seq_len, dim]
+    cos = cos[position_ids].unsqueeze(1)  # [batch_size, 1, seq_len, dim]
+    sin = sin[position_ids].unsqueeze(1)  # [batch_size, 1, seq_len, dim]
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(
+        batch, num_key_value_heads, n_rep, slen, head_dim
+    )
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def flashattn_attn(
+    self,
+    hidden_states: torch.FloatTensor,
+    attention_mask: torch.FloatTensor,
+    position_ids: torch.LongTensor,
+    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    output_attentions: Optional[bool] = False,  # pylint: disable=unused-argument
+    use_cache: Optional[bool] = False,
+    cu_seqlens: Optional[torch.Tensor] = None,
+    max_seqlen: Optional[torch.Tensor] = None,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    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)
+
+    query_rot = query_states[..., : self.rotary_ndims]
+    query_pass = query_states[..., self.rotary_ndims :]
+    key_rot = key_states[..., : self.rotary_ndims]
+    key_pass = key_states[..., self.rotary_ndims :]
+
+    kv_seq_len = key_states.shape[-2]
+    if past_key_value is not None:
+        kv_seq_len += past_key_value[0].shape[-2]
+    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+    query_states, key_states = apply_rotary_pos_emb(
+        query_rot, key_rot, cos, sin, position_ids
+    )
+
+    # [batch_size, num_heads, seq_len, head_dim]
+    query_states = torch.cat((query_states, query_pass), dim=-1)
+    key_states = torch.cat((key_states, key_pass), dim=-1)
+
+    if past_key_value is not None:
+        # Reuse k, v, self_attention
+        key_states = torch.cat((past_key_value[0], key_states), dim=2)
+        value_states = torch.cat((past_key_value[1], value_states), dim=2)
+
+    past_key_value = (key_states, value_states) if use_cache else None
+
+    # 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) ...")
+        softmax_scale = None
+
+        output = flash_attn_varlen_qkvpacked_func(
+            qkv, cu_seqlens, max_seqlen, 0.0, softmax_scale=softmax_scale, causal=True
+        )
+
+        attn_output = rearrange(output, "(b s) ... -> b s ...", b=bsz)
+        attn_output = rearrange(attn_output, "b s h d -> b s (h d)")
+    else:
+        attn_weights = torch.matmul(
+            query_states, key_states.transpose(2, 3)
+        ) / math.sqrt(self.head_dim)
+
+        if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
+            raise ValueError(
+                f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
+                f" {attn_weights.size()}"
+            )
+
+        if attention_mask is not None:
+            if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
+                raise ValueError(
+                    f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
+                )
+            attn_weights = attn_weights + attention_mask
+
+        # Upcast attention to fp32
+        attn_weights = nn.functional.softmax(
+            attn_weights, dim=-1, dtype=torch.float32
+        ).to(query_states.dtype)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        # Merge heads
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+    # Final linear projection
+    attn_output = self.o_proj(attn_output)
+
+    return attn_output, None, past_key_value
+
+
+def decoder_layer_forward(
+    self,
+    hidden_states: Optional[torch.FloatTensor],
+    attention_mask: Optional[torch.FloatTensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_value: Optional[Tuple[torch.Tensor]] = None,
+    output_attentions: Optional[bool] = False,
+    use_cache: Optional[bool] = False,
+    cu_seqlens: Optional[torch.Tensor] = None,
+    max_seqlen: Optional[torch.Tensor] = None,
+) -> Union[
+    Tuple[torch.Tensor], Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]]
+]:
+    # pylint: disable=duplicate-code
+    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 = self.mlp(hidden_states)
+    hidden_states = residual + hidden_states
+
+    outputs = (hidden_states,)
+
+    if output_attentions:
+        outputs += (self_attn_weights,)
+
+    if use_cache:
+        outputs += (present_key_value,)
+
+    return outputs
+
+
+def stablelm_model_forward(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    attention_mask: Optional[torch.FloatTensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[Tuple[Tuple[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,
+    return_dict: Optional[bool] = None,
+) -> Union[Tuple, BaseModelOutputWithPast]:
+    # pylint: disable=duplicate-code
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    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"
+        )
+
+    seq_length_with_past = seq_length
+    past_key_values_length = 0
+
+    if past_key_values is not None:
+        past_key_values_length = past_key_values[0][0].shape[2]
+        seq_length_with_past = seq_length_with_past + past_key_values_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)
+    # Embed positions
+    if attention_mask is None:
+        attention_mask = torch.ones(
+            (batch_size, seq_length_with_past),
+            dtype=torch.bool,
+            device=inputs_embeds.device,
+        )
+    attention_mask = (
+        self._prepare_decoder_attention_mask(  # pylint: disable=protected-access
+            attention_mask,
+            (batch_size, seq_length),
+            inputs_embeds,
+            past_key_values_length,
+        )
+    )
+
+    hidden_states = inputs_embeds
+
+    if self.gradient_checkpointing and self.training:
+        if use_cache:
+            logger.warning(
+                "`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
+    next_decoder_cache = () if use_cache else None
+
+    for idx, decoder_layer in enumerate(self.layers):
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+        if self.gradient_checkpointing and self.training:
+
+            def create_custom_forward(module):
+                def custom_forward(*inputs):
+                    # None for past_key_value
+                    return module(*inputs)
+
+                return custom_forward
+
+            layer_outputs = torch.utils.checkpoint.checkpoint(
+                create_custom_forward(decoder_layer),
+                hidden_states,
+                attention_mask,
+                position_ids,
+                past_key_value,
+                output_attentions,
+                None,
+                cu_seqlens,
+                max_seqlen,
+            )
+        else:
+            layer_outputs = decoder_layer(
+                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 = 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],)
+
+    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 = next_decoder_cache if use_cache else None
+    if not return_dict:
+        return tuple(
+            v
+            for v in [hidden_states, next_cache, all_hidden_states, all_self_attns]
+            if v is not None
+        )
+    return BaseModelOutputWithPast(
+        last_hidden_state=hidden_states,
+        past_key_values=next_cache,
+        hidden_states=all_hidden_states,
+        attentions=all_self_attns,
+    )

src/axolotl/monkeypatch/utils.py

@@ -101,3 +101,16 @@ def get_cu_seqlens_from_pos_ids(position_ids):
         max_seq_lens.append(max_seq_len)
 
     return torch.stack(results).to(dtype=torch.int32), torch.stack(max_seq_lens)
+
+
+def set_module_name(model, name, value):
+    if "." in name:
+        parent_name = name.rsplit(".", 1)[0]
+        child_name = name[len(parent_name) + 1 :]
+        parent = model.get_submodule(parent_name)
+    else:
+        parent_name = ""
+        parent = model
+        child_name = name
+
+    setattr(parent, child_name, value)

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

src/axolotl/prompt_strategies/alpaca_chat.py

@@ -1,6 +1,6 @@
-"""Module containing the AlpacaQAPromptTokenizingStrategy class"""
+"""Module for Alpaca prompt strategy classes"""
 
-from typing import Tuple
+from typing import Any, Dict, Optional, Tuple
 
 from axolotl.prompt_tokenizers import (
     AlpacaPromptTokenizingStrategy,
@@ -9,9 +9,13 @@ from axolotl.prompt_tokenizers import (
 from axolotl.prompters import AlpacaPrompter, PromptStyle, UnpromptedPrompter
 
 
-def load(tokenizer, cfg):
+def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
+    prompt_style = PromptStyle.CHAT.value
+    if ds_cfg and "conversation" in ds_cfg:
+        prompt_style = ds_cfg["conversation"]
+
     return AlpacaPromptTokenizingStrategy(
-        AlpacaPrompter(PromptStyle.CHAT.value),
+        AlpacaPrompter(prompt_style),
         tokenizer,
         cfg.train_on_inputs,
         cfg.sequence_len,

src/axolotl/prompt_strategies/context_qa.py

@@ -24,6 +24,15 @@ def load(tokenizer, cfg):
     )
 
 
+def load_v2(tokenizer, cfg):
+    return ContextQaV2PromptTokenizingStrategy(
+        ContextV2Prompter(),
+        tokenizer,
+        cfg.train_on_inputs,
+        cfg.sequence_len,
+    )
+
+
 class AlpacaContextPrompter(AlpacaPrompter):
     """
     Customized system prompted for concise QA
@@ -50,6 +59,38 @@ class AlpacaContextPromptTokenizingStrategy(InstructionPromptTokenizingStrategy)
         )
 
 
+class ContextQaV2PromptTokenizingStrategy(InstructionPromptTokenizingStrategy):
+    """
+    Tokenization Strategy to combine in-context article with a question and answer
+    """
+
+    def parse_instruction_fields(self, prompt) -> Tuple[str, str, str]:
+        return (
+            "Context: "
+            + prompt["context"]
+            + "\nQuestion: "
+            + prompt["question"]
+            + "\n",
+            "",
+            "Answer: " + prompt["answer"],
+        )
+
+
+class ContextV2Prompter(AlpacaPrompter):
+    """
+    Customized system prompted for concise QA
+    """
+
+    system_prompt = ""
+    system_no_input_prompt = ""
+
+    def match_prompt_style(self):
+        # pylint: disable=duplicate-code
+        self.turn_format = "{instruction}\n{input}"
+        self.turn_no_input_format = "{instruction}"
+        self.system_format = "{system}"
+
+
 class AlpacaMissingInfoContextPromptTokenizingStrategy(
     InstructionPromptTokenizingStrategy
 ):

src/axolotl/prompt_strategies/llama2_chat.py

@@ -81,8 +81,9 @@ class LLama2ChatTokenizingStrategy(PromptTokenizingStrategy):
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        self.sequence_len = 4096
-        self.tokenizer.add_special_tokens({"pad_token": "<pad>"})
+        self.tokenizer.add_special_tokens(
+            {"pad_token": getattr(self.tokenizer, "pad_token", "<pad>")}
+        )
         # https://huggingface.co/meta-llama/Llama-2-7b-chat-hf/blob/main/added_tokens.json
 
     def tokenize_prompt(self, prompt):

src/axolotl/prompt_strategies/sharegpt.py

@@ -0,0 +1,142 @@
+"""Module containing the SimpleShareGPTPromptTokenizingStrategy class"""
+from typing import Any, Dict, Optional
+
+from fastchat.conversation import Conversation, SeparatorStyle, register_conv_template
+
+from axolotl.prompt_tokenizers import ShareGPTPromptTokenizingStrategy
+from axolotl.prompters import ShareGPTPrompterV2
+
+register_conv_template(
+    Conversation(
+        name="chatml",
+        system_template="<|im_start|>system\n{system_message}",
+        system_message="You are a helpful assistant.",
+        roles=["<|im_start|>user", "<|im_start|>assistant"],
+        sep_style=SeparatorStyle.CHATML,
+        sep="<|im_end|>",
+    )
+)
+
+
+def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
+    conversation = (
+        ds_cfg["conversation"] if ds_cfg and "conversation" in ds_cfg else None
+    )
+    field_human = ds_cfg["field_human"] if ds_cfg and "field_human" in ds_cfg else None
+    field_model = ds_cfg["field_model"] if ds_cfg and "field_model" in ds_cfg else None
+    strategy = SimpleShareGPTPromptTokenizingStrategy(
+        ShareGPTPrompterV2(
+            conversation=conversation,
+            role_key_model=field_model,
+            role_key_human=field_human,
+        ),
+        tokenizer,
+        cfg.train_on_inputs,
+        cfg.sequence_len,
+    )
+    if ds_cfg and "strict" in ds_cfg:
+        strategy.strict = ds_cfg["strict"]
+    return strategy
+
+
+def load_ultrachat(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None):
+    conversation = (
+        ds_cfg["conversation"] if ds_cfg and "conversation" in ds_cfg else None
+    )
+    strategy = UltrachatShareGPTPromptTokenizingStrategy(
+        ShareGPTPrompterV2(
+            conversation=conversation,
+        ),
+        tokenizer,
+        cfg.train_on_inputs,
+        cfg.sequence_len,
+    )
+    if ds_cfg and "strict" in ds_cfg:
+        strategy.strict = ds_cfg["strict"]
+    return strategy
+
+
+def load_role(tokenizer, cfg):
+    return SimpleRoleShareGPTPromptTokenizingStrategy(
+        ShareGPTPrompterV2(),
+        tokenizer,
+        cfg.train_on_inputs,
+        cfg.sequence_len,
+    )
+
+
+def load_guanaco(tokenizer, cfg):
+    return GuanacoShareGPTPromptTokenizingStrategy(
+        ShareGPTPrompterV2(),
+        tokenizer,
+        cfg.train_on_inputs,
+        cfg.sequence_len,
+    )
+
+
+class SimpleShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
+    """
+    basic sharegpt strategy to grab conversations from the sample row
+    """
+
+    _strict = True
+
+    @property
+    def strict(self):
+        return self._strict
+
+    @strict.setter
+    def strict(self, strict):
+        self._strict = strict
+
+    def get_conversation_thread(self, prompt):
+        conversations = prompt["conversations"]
+        if self.strict:
+            return conversations
+        # remap roles - allow for assistant turn
+        role_map = {"human": "human", "assistant": "gpt", "gpt": "gpt"}
+        turns = [
+            {"from": role_map[t["from"]], "value": t["value"]} for t in conversations
+        ]
+        return turns
+
+
+class SimpleRoleShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
+    """
+    basic sharegpt strategy to grab conversations from the sample row, but uses role instead of from
+    """
+
+    def get_conversation_thread(self, prompt):
+        conversations = prompt["conversations"]
+        # remap role: prompter/assistant, text: ... => from: human/gpt, value: ...
+        turns = [{"from": t["role"], "value": t["value"]} for t in conversations]
+        return turns
+
+
+class GuanacoShareGPTPromptTokenizingStrategy(ShareGPTPromptTokenizingStrategy):
+    """
+    sharegpt strategy that remaps oasst data to sharegpt format
+    """
+
+    def get_conversation_thread(self, prompt):
+        conversations = prompt["conversations"]
+        # remap role: prompter/assistant, text: ... => from: human/gpt, value: ...
+        role_map = {"prompter": "human", "assistant": "gpt"}
+        turns = [
+            {"from": role_map[t["role"]], "value": t["text"]} for t in conversations
+        ]
+        return turns
+
+
+class UltrachatShareGPTPromptTokenizingStrategy(SimpleShareGPTPromptTokenizingStrategy):
+    """
+    sharegpt strategy that remaps ultrachat data to sharegpt format
+    """
+
+    def get_conversation_thread(self, prompt):
+        conversations = prompt["messages"]
+        role_map = {"user": "human", "assistant": "gpt"}
+        turns = [
+            {"from": role_map[t["role"]], "value": t["content"]} for t in conversations
+        ]
+        return turns