remove torch install for now

install torch for tests
fix cuda version for autogptq index
set torch in requirements so that it installs properly
move gptq install around to work with github cicd

.github/CODE_OF_CONDUCT.md

@@ -0,0 +1,129 @@
+# Contributor Covenant Code of Conduct
+
+## Our Pledge
+
+We as members, contributors, and leaders pledge to make participation in our
+community a harassment-free experience for everyone, regardless of age, body
+size, visible or invisible disability, ethnicity, sex characteristics, gender
+identity and expression, level of experience, education, socio-economic status,
+nationality, personal appearance, race, religion, or sexual identity
+and orientation.
+
+We pledge to act and interact in ways that contribute to an open, welcoming,
+diverse, inclusive, and healthy community.
+
+## Our Standards
+
+Examples of behavior that contributes to a positive environment for our
+community include:
+
+* Demonstrating empathy and kindness toward other people
+* Being respectful of differing opinions, viewpoints, and experiences
+* Giving and gracefully accepting constructive feedback
+* Accepting responsibility and apologizing to those affected by our mistakes,
+  and learning from the experience
+* Focusing on what is best not just for us as individuals, but for the
+  overall community
+
+Examples of unacceptable behavior include:
+
+* The use of sexualized language or imagery, and sexual attention or
+  advances of any kind
+* Trolling, insulting or derogatory comments, and personal or political attacks
+* Public or private harassment
+* Publishing others' private information, such as a physical or email
+  address, without their explicit permission
+* Other conduct which could reasonably be considered inappropriate in a
+  professional setting
+
+## Enforcement Responsibilities
+
+Community leaders are responsible for clarifying and enforcing our standards of
+acceptable behavior and will take appropriate and fair corrective action in
+response to any behavior that they deem inappropriate, threatening, offensive,
+or harmful.
+
+Community leaders have the right and responsibility to remove, edit, or reject
+comments, commits, code, wiki edits, issues, and other contributions that are
+not aligned to this Code of Conduct, and will communicate reasons for moderation
+decisions when appropriate.
+
+## Scope
+
+This Code of Conduct applies within all community spaces, and also applies when
+an individual is officially representing the community in public spaces.
+Examples of representing our community include using an official e-mail address,
+posting via an official social media account, or acting as an appointed
+representative at an online or offline event.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be
+reported to the community leaders responsible for enforcement on Discord
+at https://discord.gg/QYF8QrtEUm
+
+All complaints will be reviewed and investigated promptly and fairly.
+
+All community leaders are obligated to respect the privacy and security of the
+reporter of any incident.
+
+## Enforcement Guidelines
+
+Community leaders will follow these Community Impact Guidelines in determining
+the consequences for any action they deem in violation of this Code of Conduct:
+
+### 1. Correction
+
+**Community Impact**: Use of inappropriate language or other behavior deemed
+unprofessional or unwelcome in the community.
+
+**Consequence**: A private, written warning from community leaders, providing
+clarity around the nature of the violation and an explanation of why the
+behavior was inappropriate. A public apology may be requested.
+
+### 2. Warning
+
+**Community Impact**: A violation through a single incident or series
+of actions.
+
+**Consequence**: A warning with consequences for continued behavior. No
+interaction with the people involved, including unsolicited interaction with
+those enforcing the Code of Conduct, for a specified period of time. This
+includes avoiding interactions in community spaces as well as external channels
+like social media. Violating these terms may lead to a temporary or
+permanent ban.
+
+### 3. Temporary Ban
+
+**Community Impact**: A serious violation of community standards, including
+sustained inappropriate behavior.
+
+**Consequence**: A temporary ban from any sort of interaction or public
+communication with the community for a specified period of time. No public or
+private interaction with the people involved, including unsolicited interaction
+with those enforcing the Code of Conduct, is allowed during this period.
+Violating these terms may lead to a permanent ban.
+
+### 4. Permanent Ban
+
+**Community Impact**: Demonstrating a pattern of violation of community
+standards, including sustained inappropriate behavior,  harassment of an
+individual, or aggression toward or disparagement of classes of individuals.
+
+**Consequence**: A permanent ban from any sort of public interaction within
+the community.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant][homepage],
+version 2.0, available at
+https://www.contributor-covenant.org/version/2/0/code_of_conduct.html.
+
+Community Impact Guidelines were inspired by [Mozilla's code of conduct
+enforcement ladder](https://github.com/mozilla/diversity).
+
+[homepage]: https://www.contributor-covenant.org
+
+For answers to common questions about this code of conduct, see the FAQ at
+https://www.contributor-covenant.org/faq. Translations are available at
+https://www.contributor-covenant.org/translations.

.github/CONTRIBUTING.md

@@ -0,0 +1,76 @@
+# Contributing to axolotl
+
+First of all, thank you for your interest in contributing to axolotl! We appreciate the time and effort you're willing to invest in making our project better. This document provides guidelines and information to make the contribution process as smooth as possible.
+
+## Table of Contents
+
+- [Code of Conduct](#code-of-conduct)
+- [Getting Started](#getting-started)
+- [How to Contribute](#how-to-contribute)
+  - [Reporting Bugs](#reporting-bugs)
+  - [Suggesting Enhancements](#suggesting-enhancements)
+  - [Submitting Pull Requests](#submitting-pull-requests)
+- [Style Guidelines](#style-guidelines)
+  - [Code Style](#code-style)
+  - [Commit Messages](#commit-messages)
+- [Additional Resources](#additional-resources)
+
+## Code of Conductcode
+
+All contributors are expected to adhere to our [Code of Conduct](CODE_OF_CONDUCT.md). Please read it before participating in the axolotl community.
+
+## Getting Started
+
+Bugs? Please check for open issue else create a new [Issue](https://github.com/OpenAccess-AI-Collective/axolotl/issues/new).
+
+PRs are **greatly welcome**!
+
+1. Fork the repository and clone it to your local machine.
+2. Set up the development environment by following the instructions in the [README.md](https://github.com/OpenAccess-AI-Collective/axolotl/tree/main/README.md) file.
+3. Explore the codebase, run tests, and verify that everything works as expected.
+
+Please run below to setup env
+```bash
+pip3 install -r requirements-dev.txt -r requirements-tests.txt
+pre-commit install
+
+# test
+pytest tests/
+```
+
+## How to Contribute
+
+### Reporting Bugs
+
+If you encounter a bug or issue while using axolotl, please open a new issue on the [GitHub Issues](https://github.com/OpenAccess-AI-Collective/axolotl/issues) page. Provide a clear and concise description of the problem, steps to reproduce it, and any relevant error messages or logs.
+
+### Suggesting Enhancements
+
+We welcome ideas for improvements and new features. To suggest an enhancement, open a new issue on the [GitHub Issues](https://github.com/OpenAccess-AI-Collective/axolotl/issues) page. Describe the enhancement in detail, explain the use case, and outline the benefits it would bring to the project.
+
+### Submitting Pull Requests
+
+1. Create a new branch for your feature or bugfix. Use a descriptive name like `feature/your-feature-name` or `fix/your-bugfix-name`.
+2. Make your changes, following the [Style Guidelines](#style-guidelines) below.
+3. Test your changes and ensure that they don't introduce new issues or break existing functionality.
+4. Commit your changes, following the [commit message guidelines](#commit-messages).
+5. Push your branch to your fork on GitHub.
+6. Open a new pull request against the `main` branch of the axolotl repository. Include a clear and concise description of your changes, referencing any related issues.
+
+## Style Guidelines
+
+### Code Style
+
+axolotl uses [{codestyle}]({URLofCodestyle}) as its code style guide. Please ensure that your code follows these guidelines.
+
+### Commit Messages
+
+Write clear and concise commit messages that briefly describe the changes made in each commit. Use the imperative mood and start with a capitalized verb, e.g., "Add new feature" or "Fix bug in function".
+
+## Additional Resources
+
+- [GitHub Help](https://help.github.com/)
+- [GitHub Pull Request Documentation](https://docs.github.com/en/github/collaborating-with-issues-and-pull-requests)
+- [{codestyle}]({URLofCodestyle})
+
+Thank you once again for your interest in contributing to axolotl. We look forward to collaborating with you and creating an even better project together!

.github/FUNDING.yml

@@ -0,0 +1,13 @@
+# These are supported funding model platforms
+
+github: OpenAccess-AI-Collective # Replace with up to 4 GitHub Sponsors-enabled usernames e.g., [user1, user2]
+patreon: # Replace with a single Patreon username
+open_collective: # Replace with a single Open Collective username
+ko_fi: # Replace with a single Ko-fi username
+tidelift: # Replace with a single Tidelift platform-name/package-name e.g., npm/babel
+community_bridge: # Replace with a single Community Bridge project-name e.g., cloud-foundry
+liberapay: # Replace with a single Liberapay username
+issuehunt: # Replace with a single IssueHunt username
+otechie: # Replace with a single Otechie username
+lfx_crowdfunding: # Replace with a single LFX Crowdfunding project-name e.g., cloud-foundry
+custom: # Replace with up to 4 custom sponsorship URLs e.g., ['link1', 'link2']

.github/ISSUE_TEMPLATE/bug-report.yaml

@@ -0,0 +1,105 @@
+name: Bug Report
+description: File a bug report
+labels: ["bug", "needs triage"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        ## Before you start
+        Please **make sure you are on the latest version.**
+        If you encountered the issue after you installed, updated, or reloaded, **please try restarting before reporting the bug**.
+
+  - type: checkboxes
+    id: no-duplicate-issues
+    attributes:
+      label: "Please check that this issue hasn't been reported before."
+      description: "The **Label filters** may help make your search more focussed."
+      options:
+        - label: "I searched previous [Bug Reports](https://github.com/OpenAccess-AI-Collective/axolotl/labels/bug) didn't find any similar reports."
+          required: true
+
+  - type: textarea
+    id: expected
+    attributes:
+      label: Expected Behavior
+      description: Tell us what **should** happen.
+    validations:
+      required: true
+
+  - type: textarea
+    id: what-happened
+    attributes:
+      label: Current behaviour
+      description: |
+        Tell us what happens instead of the expected behavior.
+        Provide stacktrace and/or screenshots.
+    validations:
+      required: true
+
+  - type: textarea
+    id: reproduce
+    attributes:
+      label: Steps to reproduce
+      description: |
+        Which exact steps can a developer take to reproduce the issue?
+        The more detail you provide, the easier it will be to narrow down and fix the bug.
+        Please paste in tasks and/or queries **as text, not screenshots**.
+      placeholder: |
+        Example of the level of detail needed to reproduce any bugs efficiently and reliably.
+        1. Go to the '...' page.
+        2. Click on the '...' button.
+        3. Scroll down to '...'.
+        4. Observe the error.
+    validations:
+      required: true
+
+  - type: textarea
+    id: possible-solution
+    attributes:
+      label: Possible solution
+      description: |
+        Not obligatory, but please suggest a fix or reason for the bug, if you have an idea.
+
+
+  - type: checkboxes
+    id: operating-systems
+    attributes:
+      label: Which Operating Systems are you using?
+      description: You may select more than one.
+      options:
+        - label: Linux
+        - label: macOS
+        - label: Windows
+
+  - type: input
+    id: Python-version
+    attributes:
+      label: Python Version
+      description: Which {Programming} version are you using?
+      placeholder: 3.10 / please change accordingly
+    validations:
+      required: true
+
+  - type: input
+    id: axolotl-branch-commit
+    attributes:
+      label: axolotl branch-commit
+      description: On which branch/commit are you?
+      placeholder: main/4d6490b
+    validations:
+      required: true
+
+  - type: checkboxes
+    id: acknowledgements
+    attributes:
+      label: 'Acknowledgements'
+      description: 'Please confirm the following:'
+      options:
+        - label: 'My issue title is concise, descriptive, and in title casing.'
+          required: true
+        - label: 'I have searched the existing issues to make sure this bug has not been reported yet.'
+          required: true
+        - label: 'I am using the latest version of axolotl.'
+          required: true
+        - label: 'I have provided enough information for the maintainers to reproduce and diagnose the issue.'
+          required: true

.github/ISSUE_TEMPLATE/config.yml

@@ -0,0 +1,7 @@
+blank_issues_enabled: false
+contact_links:
+  - name: Ask a question
+    url: https://github.com/OpenAccess-AI-Collective/axolotl/discussions/categories/q-a
+    about: Ask questions and discuss with other community members
+  - name: Discuss the Project in Discord
+    url: https://discord.gg/HhrNrHJPRb

.github/ISSUE_TEMPLATE/docs.yml

@@ -0,0 +1,46 @@
+name: Documentation Improvement / Clarity
+description: Make a suggestion to improve the project documentation.
+labels: ['needs triage', 'docs']
+body:
+  - type: markdown
+    attributes:
+      value: '## :book: Documentation :book:'
+  - type: markdown
+    attributes:
+      value: |
+        * Ask questions in [Discord](https://discord.gg/HhrNrHJPRb).
+        * Before you file an issue read the [Contributing guide](./CONTRIBUTING.md).
+        * Check to make sure someone hasn't already opened a [similar issue](https://github.com/OpenAccess-AI-Collective/axolotl/issues).
+  - type: textarea
+    attributes:
+      label: What piece of documentation is affected?
+      description: Please link to the article you'd like to see updated.
+    validations:
+      required: true
+  - type: textarea
+    attributes:
+      label: What part(s) of the article would you like to see updated?
+      description: |
+        - Give as much detail as you can to help us understand the change you want to see.
+        - Why should the docs be changed? What use cases does it support?
+        - What is the expected outcome?
+    validations:
+      required: true
+  - type: textarea
+    attributes:
+      label: Additional Information
+      description: Add any other context or screenshots about the feature request here.
+    validations:
+      required: false
+  - type: checkboxes
+    id: acknowledgements
+    attributes:
+      label: 'Acknowledgements'
+      description: 'Please confirm the following:'
+      options:
+        - label: 'My issue title is concise, descriptive, and in title casing.'
+          required: true
+        - label: 'I have searched the existing issues to make sure this feature has not been requested yet.'
+          required: true
+        - label: 'I have provided enough information for the maintainers to understand and evaluate this request.'
+          required: true

.github/ISSUE_TEMPLATE/feature-request.yaml

@@ -0,0 +1,63 @@
+name: Feature Request / Enhancement
+description: Suggest a new feature or feature enhancement for the project
+labels: ["enhancement", "needs triage"]
+body:
+  - type: checkboxes
+    id: no-duplicate-issues
+    attributes:
+      label: "⚠️ Please check that this feature request hasn't been suggested before."
+      description: "There are two locations for previous feature requests. Please search in both. Thank you. The **Label filters** may help make your search more focussed."
+      options:
+        - label: "I searched previous [Ideas in Discussions](https://github.com/OpenAccess-AI-Collective/axolotl/discussions/categories/ideas) didn't find any similar feature requests."
+          required: true
+        - label: "I searched previous [Issues](https://github.com/OpenAccess-AI-Collective/axolotl/labels/enhancement) didn't find any similar feature requests."
+          required: true
+
+  - type: textarea
+    id: feature-description
+    validations:
+      required: true
+    attributes:
+      label: "🔖 Feature description"
+      description: "A clear and concise description of what the feature request is."
+      placeholder: "You should add ..."
+
+  - type: textarea
+    id: solution
+    validations:
+      required: true
+    attributes:
+      label: "✔️ Solution"
+      description: "A clear and concise description of what you want to happen, and why."
+      placeholder: "In my use-case, ..."
+
+  - type: textarea
+    id: alternatives
+    validations:
+      required: false
+    attributes:
+      label: "❓ Alternatives"
+      description: "A clear and concise description of any alternative solutions or features you've considered."
+      placeholder: "I have considered ..."
+
+  - type: textarea
+    id: additional-context
+    validations:
+      required: false
+    attributes:
+      label: "📝 Additional Context"
+      description: "Add any other context or screenshots about the feature request here."
+      placeholder: "..."
+
+  - type: checkboxes
+    id: acknowledgements
+    attributes:
+      label: 'Acknowledgements'
+      description: 'Please confirm the following:'
+      options:
+        - label: 'My issue title is concise, descriptive, and in title casing.'
+          required: true
+        - label: 'I have searched the existing issues to make sure this feature has not been requested yet.'
+          required: true
+        - label: 'I have provided enough information for the maintainers to understand and evaluate this request.'
+          required: true

.github/PULL_REQUEST_TEMPLATE/pull_request_template_simple.md

@@ -0,0 +1,22 @@
+<!--- Provide a general summary of your changes in the Title above -->
+
+# Description
+
+<!--- Describe your changes in detail -->
+
+## Motivation and Context
+
+<!--- Why is this change required? What problem does it solve? -->
+<!--- If it fixes an open issue, please link to the issue here. -->
+
+## How has this been tested?
+
+<!--- Please describe in detail how you tested your changes. -->
+<!--- Include details of your testing environment, tests ran to see how -->
+<!--- your change affects other areas of the code, etc. -->
+
+## Screenshots (if appropriate)
+
+## Types of changes
+
+<!--- What types of changes does your code introduce? Put an `x` in all the boxes that apply: -->

.github/SECURITY.md

@@ -0,0 +1,9 @@
+# Security Policy
+
+## Supported Versions
+
+Due to the nature of the fast development that is happening in this project, only the latest released version can be supported.
+
+## Reporting a Vulnerability
+
+If you find a vulnerability, please contact us on  [Discord](https://discord.gg/xcu3ECkH9a) rather than creating a GitHub issue to allow us some time to fix it before it is a known vulnerability to others.

.github/SUPPORT.md

@@ -0,0 +1,10 @@
+# Support
+
+If you need help with this project or have questions, please:
+
+1. Check the documentation.
+2. Search the existing issues and pull requests.
+3. Create a new issue if your question is not answered or your problem is not solved.
+4. Have a look in the [Discord server](https://discord.gg/HhrNrHJPRb)
+
+Please note that this project is maintained by volunteers who have limited availability. We'll do our best to address your questions and concerns in a timely manner.

.github/workflows/main.yml

@@ -4,7 +4,6 @@ on:
   push:
     branches:
       - "main"
-      - "dev"
 
 jobs:
   build-axolotl:
@@ -14,21 +13,16 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: cu118
+          - cuda: 118
             cuda_version: 11.8.0
             python_version: "3.9"
             pytorch: 2.0.1
             axolotl_extras:
-          - cuda: cu118
+          - cuda: 118
             cuda_version: 11.8.0
             python_version: "3.10"
             pytorch: 2.0.1
             axolotl_extras:
-          - cuda: cu118
-            cuda_version: 11.8.0
-            python_version: "3.9"
-            pytorch: 2.0.1
-            axolotl_extras: gptq
     runs-on: self-hosted
     steps:
       - name: Checkout
@@ -50,10 +44,11 @@ jobs:
         with:
           context: .
           build-args: |
-            BASE_TAG=${{ github.ref_name }}-base-py${{ matrix.python_version }}-${{ matrix.cuda }}-${{ matrix.pytorch }}
+            BASE_TAG=${{ github.ref_name }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}
+            CUDA=${{ matrix.cuda }}
           file: ./docker/Dockerfile
           push: ${{ github.event_name != 'pull_request' }}
-          tags: ${{ steps.metadata.outputs.tags }}-py${{ matrix.python_version }}-${{ matrix.cuda }}-${{ 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 }}
           labels: ${{ steps.metadata.outputs.labels }}
   build-axolotl-runpod:
     needs: build-axolotl
@@ -72,11 +67,7 @@ jobs:
             python_version: "3.10"
             pytorch: 2.0.1
             axolotl_extras:
-          - cuda: 118
-            cuda_version: 11.8.0
-            python_version: "3.9"
-            pytorch: 2.0.1
-            axolotl_extras: gptq
+            is_latest: true
     runs-on: self-hosted
     steps:
       - name: Checkout
@@ -102,5 +93,7 @@ jobs:
             CUDA=${{ matrix.cuda }}
           file: ./docker/Dockerfile-runpod
           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 }}

README.md

@@ -1,10 +1,40 @@
 # Axolotl
 
+Axolotl is a tool designed to streamline the fine-tuning of various AI models, offering support for multiple configurations and architectures.
+
+<table>
+<tr>
+<td>
+
+## Table of Contents
+- [Introduction](#axolotl)
+- [Supported Features](#axolotl-supports)
+- [Quickstart](#quickstart-)
+- [Installation](#installation)
+  - [Docker Installation](#environment)
+  - [Conda/Pip venv Installation](#condapip-venv)
+  - [LambdaLabs Installation](#lambdalabs)
+- [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)
+  - [Inference](#inference)
+  - [Merge LORA to Base](#merge-lora-to-base)
+- [Common Errors](#common-errors-)
+- [Need Help?](#need-help-)
+- [Badge](#badge-)
+- [Community Showcase](#community-showcase)
+- [Contributing](#contributing-)
+
+</td>
+<td>
+
 <div align="center">
   <img src="image/axolotl.png" alt="axolotl" width="160">
   <div>
     <p>
-      <b>One repo to finetune them all! </b>
+      <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!!
@@ -14,27 +44,34 @@
   </div>
 </div>
 
+</td>
+</tr>
+</table>
+
 ## Axolotl supports
 
-|          | fp16/fp32 | lora | qlora | gptq | gptq w/ lora | gptq w/flash attn | flash attn | xformers attn |
-|----------|:----------|:-----|-------|------|:-------------|-------------------|------------|---------------|
-| llama    | ✅         | ✅    | ✅     | ✅    | ✅             | ✅                 | ✅          | ✅             |
-| Pythia   | ✅         | ✅    | ✅     | ❌    | ❓            | ❌                 | ❌          | ❓             |
-| cerebras | ✅         | ✅    | ✅     | ❌    | ❓            | ❌                 | ❌          | ✅             |
-| mpt      | ✅         | ❌    | ❓     | ❌    | ❓            | ❌                 | ❌          | ❓             |
-| falcon   | ✅         | ✅    | ✅     | ❌    | ❓            | ❌                 | ❌          | ✅             |
-| gpt-j    | ✅         | ✅    | ✅     | ❌    | ❓            | ❌                 | ❓          | ✅             |
-| XGen     | ✅         | ❓    | ✅     | ❓    | ❓            | ❓                 | ❓          | ✅
+|          | fp16/fp32 | lora | qlora | gptq | gptq w/flash attn | flash attn | xformers attn |
+|----------|:----------|:-----|-------|------|-------------------|------------|---------------|
+| llama    | ✅         | ✅    | ✅     | ✅             | ✅                 | ✅          | ✅             |
+| Pythia   | ✅         | ✅    | ✅     | ❌             | ❌                 | ❌          | ❓             |
+| cerebras | ✅         | ✅    | ✅     | ❌             | ❌                 | ❌          | ❓             |
+| mpt      | ✅         | ❌    | ❓     | ❌             | ❌                 | ❌          | ❓             |
+| falcon   | ✅         | ✅    | ✅     | ❌             | ❌                 | ❌          | ❓             |
+| gpt-j    | ✅         | ✅    | ✅     | ❌             | ❌                 | ❓          | ❓             |
+| XGen     | ✅         | ❓    | ✅     | ❓             | ❓                 | ❓          | ✅             |
 
 
 ## Quickstart ⚡
 
+Get started with Axolotl in just a few steps! This quickstart guide will walk you through setting up and running a basic fine-tuning task.
+
 **Requirements**: Python >=3.9 and Pytorch >=2.0.
 
 ```bash
 git clone https://github.com/OpenAccess-AI-Collective/axolotl
+cd axolotl
 
-pip3 install -e .
+pip3 install -e .[flash-attn]
 pip3 install -U git+https://github.com/huggingface/peft.git
 
 # finetune lora
@@ -63,7 +100,7 @@ accelerate launch scripts/finetune.py examples/openllama-3b/lora.yml \
   ```
 
 - Conda/Pip venv
-  1. Install python **3.9**
+  1. Install python >=**3.9**
 
   2. Install pytorch stable https://pytorch.org/get-started/locally/
 
@@ -116,9 +153,7 @@ accelerate launch scripts/finetune.py examples/openllama-3b/lora.yml \
 
   pip3 install -e . # change depend on needs
   pip3 install protobuf==3.20.3
-  pip3 install -U requests
-  pip3 install -U --ignore-installed psutil
-  pip3 install -U scipy
+  pip3 install -U --ignore-installed requests Pillow psutil scipy
   pip3 install git+https://github.com/huggingface/peft.git # not for gptq
   ```
 
@@ -130,13 +165,14 @@ accelerate launch scripts/finetune.py examples/openllama-3b/lora.yml \
 
 ### Dataset
 
+Axolotl supports a variety of dataset formats. Below are some of the formats you can use.
 Have dataset(s) in one of the following format (JSONL recommended):
 
 - `alpaca`: instruction; input(optional)
   ```json
   {"instruction": "...", "input": "...", "output": "..."}
   ```
-- `sharegpt:chat`: conversations
+- `sharegpt:chat`: conversations where `from` is `human`/`gpt`
   ```json
   {"conversations": [{"from": "...", "value": "..."}]}
   ```
@@ -221,10 +257,18 @@ Have dataset(s) in one of the following format (JSONL recommended):
   ```json
   {"conversations": [{"role": "...", "value": "..."}]}
   ```
+- `metharme`: instruction, adds additional eos tokens
+  ```json
+  {"prompt": "...", "generation": "..."}
+  ```
 - `sharegpt_simple.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
+  ```json
+  {"conversations": [{"from": "...", "value": "..."}]}
+  ```
 - `sharegpt_jokes`: creates a chat where bot is asked to tell a joke, then explain why the joke is funny
   ```json
   {"conversations": [{"title": "...", "text": "...", "explanation": "..."}]}
@@ -234,11 +278,29 @@ Have dataset(s) in one of the following format (JSONL recommended):
 
 #### How to add custom prompts
 
-  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>`.
+Using yaml. Example:
+```yaml
+datasets:
+  - path: repo
+    type:
+      system_prompt: ""
+      no_input_format: |-
+        User: {instruction}<|end_of_turn|>
+        Assistant:
+      format: |-
+        User: {instruction}
+        {input}<|end_of_turn|>
+        Assistant:
+```
 
-Optionally, download some datasets, see [data/README.md](data/README.md)
+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
 
 
 ### Config
@@ -266,11 +328,20 @@ See [examples](examples) for quick start. It is recommended to duplicate and mod
       name: enron_emails
       type: completion # format from earlier
 
+  # huggingface repo with multiple named configurations/subsets
+  datasets:
+    - path: bigcode/commitpackft
+      name:
+        - ruby
+        - python
+        - typescript
+      type: ... # unimplemented custom format
+
   # local
   datasets:
-    - path: json
-      data_files: data.jsonl # or json
-      type: alpaca # format from earlier
+    - path: data.jsonl # or json
+      ds_type: json # see other options below
+      type: alpaca
   ```
 
 - loading
@@ -322,6 +393,8 @@ tokenizer_type: AutoTokenizer
 trust_remote_code:
 # use_fast option for tokenizer loading from_pretrained, default to True
 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_token_embeddings_to_32x:
@@ -343,16 +416,39 @@ fp16: true
 # Use CUDA tf32
 tf32: true # require >=ampere
 
+# No AMP (automatic mixed precision)
+bfloat16: true # require >=ampere
+float16: 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
   - 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>
+    ds_type: # Optional[str] (json|arrow|parquet) defines the datatype when path is a file
     data_files: # path to source data files
     shards: # number of shards to split data into
     name: # name of dataset configuration to load
 
+  # custom user prompt
+  - path: repo
+    type:
+      # the below are defaults. only set what's needed.
+      system_prompt: ""
+      field_system: system
+      field_instruction: instruction
+      field_output: input
+
+      # customizable to be single line or multi-line
+      system_format: "{system}"
+      # 'format' can include {input}
+      format: |-
+        User: {instruction} {input}
+        Assistant:
+      # 'no_input_format' cannot include {input}
+      no_input_format: "{instruction} "
+
 # 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
@@ -360,10 +456,13 @@ dataset_prepared_path: data/last_run_prepared
 push_dataset_to_hub: # repo path
 # 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`
 hf_use_auth_token: # boolean
-# How much of the dataset to set aside as evaluation. 1 = 100%, 0.50 = 50%, etc
+# 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
 # Num shards for whole dataset
 dataset_shard_num:
@@ -373,9 +472,19 @@ dataset_shard_idx:
 # 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_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
+# 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'
+sample_packing:
+# 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
 adapter: lora
@@ -400,12 +509,19 @@ lora_modules_to_save:
 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
+
 # wandb configuration if you're using it
-wandb_mode:
-wandb_project:
+wandb_mode: # "offline" to save run metadata locally and not sync to the server, "disabled" to turn off wandb
+wandb_project: # your wandb project name
+wandb_entity: # a wandb Team name if using a Team
 wandb_watch:
-wandb_run_id:
-wandb_log_model: # 'checkpoint'
+wandb_run_id: # set the name 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
 output_dir: ./completed-model
@@ -417,16 +533,22 @@ eval_batch_size: 2
 num_epochs: 3
 warmup_steps: 100
 learning_rate: 0.00003
+lr_quadratic_warmup:
 logging_steps:
-save_steps:
-eval_steps:
+save_strategy: # set to `no` to skip checkpoint saves
+save_steps: # leave empty to save at each epoch
+eval_steps: # leave empty to eval at each epoch
+save_total_limit: # checkpoints saved at a time
+max_steps:
 
 # save model as safetensors (require safetensors package)
 save_safetensors:
 
 # whether to mask out or include the human's prompt from the training labels
 train_on_inputs: false
-# don't use this, leads to wonky training (according to someone on the internet)
+# 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
@@ -462,8 +584,8 @@ max_grad_norm:
 flash_optimum:
 # whether to use xformers attention patch https://github.com/facebookresearch/xformers:
 xformers_attention:
-# whether to use flash attention patch https://github.com/HazyResearch/flash-attention:
-flash_attention:  # require a100 for llama
+# whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:
+flash_attention:
 # whether to use scaled-dot-product attention
 # https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html
 sdp_attention:
@@ -472,6 +594,10 @@ landmark_attention:
 # xpos RoPE see https://github.com/kaiokendev/cutoff-len-is-context-len/blob/main/util/xpos_rope_llama_monkey_patch.py
 # llama only
 xpos_rope:
+# RoPE Scaling https://github.com/huggingface/transformers/pull/24653
+rope_scaling:
+  type: # linear | dynamic
+  factor: # float
 
 # resume from a specific checkpoint dir
 resume_from_checkpoint:
@@ -494,14 +620,14 @@ tokens:
 fsdp:
 fsdp_config:
 
-# Deepspeed
+# Deepspeed config path
 deepspeed:
 
 # Path to torch distx for optim 'adamw_anyprecision'
 torchdistx_path:
 
-# Set padding for data collator to 'longest'
-collator_pad_to_longest:
+# Set to HF dataset for type: 'completion' for streaming instead of pre-tokenize
+pretraining_dataset:
 
 # Debug mode
 debug:
@@ -519,10 +645,17 @@ strict:
 
 Run
 ```bash
-accelerate launch scripts/finetune.py configs/your_config.yml
+accelerate launch scripts/finetune.py your_config.yml
 ```
 
-#### Multi-GPU Config
+#### Multi-GPU
+
+You can optionally pre-tokenize dataset with the following before finetuning:
+```bash
+CUDA_VISIBLE_DEVICES="" accelerate ... --prepare_ds_only
+```
+
+##### Config
 
 - llama FSDP
 ```yaml
@@ -535,7 +668,22 @@ fsdp_config:
   fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
 ```
 
-- llama Deepspeed: append `ACCELERATE_USE_DEEPSPEED=true` in front of finetune command
+- llama Deepspeed
+```yaml
+deepspeed: deepspeed/zero3.json
+```
+
+##### Weights & Biases Logging
+
+- wandb options
+```yaml
+wandb_mode:
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+```
 
 ### Inference
 
@@ -571,7 +719,7 @@ CUDA_VISIBLE_DEVICES="" python3 scripts/finetune.py ...
 
 ## Common Errors 🧰
 
-> Cuda out of memory
+> 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
   - `micro_batch_size`
@@ -579,6 +727,12 @@ Please reduce any below
   - `gradient_accumulation_steps`
   - `sequence_len`
 
+> `failed (exitcode: -9)`
+
+Usually means your system has run out of system memory.
+Similarly, you should consider reducing the same settings as when you run out of VRAM.
+Additionally, look into upgrading your system RAM which should be simpler than GPU upgrades.
+
 > RuntimeError: expected scalar type Float but found Half
 
 Try set `fp16: true`
@@ -607,6 +761,8 @@ Building something cool with Axolotl? Consider adding a badge to your model card
 
 ## Community Showcase
 
+Check out some of the projects and models that have been built using Axolotl! Have a model you'd like to add to our Community Showcase? Open a PR with your model.
+
 Open Access AI Collective
 - [Minotaur 13b](https://huggingface.co/openaccess-ai-collective/minotaur-13b)
 - [Manticore 13b](https://huggingface.co/openaccess-ai-collective/manticore-13b)
@@ -617,7 +773,9 @@ PocketDoc Labs
 
 ## Contributing 🤝
 
-Bugs? Please check for open issue else create a new [Issue](https://github.com/OpenAccess-AI-Collective/axolotl/issues/new).
+Please read the [contributing guide](./.github/CONTRIBUTING.md)
+
+Bugs? Please check the [open issues](https://github.com/OpenAccess-AI-Collective/axolotl/issues/bug) else create a new Issue.
 
 PRs are **greatly welcome**!
 

data/README.md

@@ -1,24 +0,0 @@
-
-## Download some datasets
-```shell
-curl https://raw.githubusercontent.com/tloen/alpaca-lora/main/alpaca_data_gpt4.json -o data/raw/alpaca_data_gpt4.json
-curl https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json -L -o data/raw/vicuna_cleaned.json
-curl https://github.com/teknium1/GPTeacher/blob/main/Instruct/gpt4-instruct-similarity-0.6-dataset.json?raw=true -L -o data/raw/gpt4-instruct-similarity-0.6-dataset.json
-curl https://github.com/teknium1/GPTeacher/blob/main/Roleplay/roleplay-similarity_0.6-instruct-dataset.json?raw=true -L -o data/raw/roleplay-similarity_0.6-instruct-dataset.json
-```
-
-## Convert the JSON data files to JSONL.
-
-```shell
-python3 ./scripts/alpaca_json_to_jsonl.py --file data/alpaca_data_gpt4.json --output data/alpaca_data_gpt4.jsonl
-python3 ./scripts/alpaca_json_to_jsonl.py --file data/raw/vicuna_cleaned.json --output data/vicuna_cleaned.jsonl
-python3 ./scripts/alpaca_json_to_jsonl.py --file data/raw/roleplay-similarity_0.6-instruct-dataset.json --output data/roleplay-similarity_0.6-instruct-dataset.jsonl
-python3 ./scripts/alpaca_json_to_jsonl.py --file data/raw/gpt4-instruct-similarity-0.6-dataset.json --output data/gpt4-instruct-similarity-0.6-dataset.jsonl
-```
----
-
-Using JSONL makes it easier to subset the data if you want a smaller training set, i.e get 2000 random examples.
-
-```shell
-shuf -n2000 data/vicuna_cleaned.jsonl > data/vicuna_cleaned.subset0.jsonl
-```

data/raw/.gitignore

@@ -1 +0,0 @@
-**

deepspeed/zero2.json

@@ -0,0 +1,46 @@
+{
+    "zero_optimization": {
+      "stage": 2,
+      "offload_optimizer": {
+        "device": "cpu"
+      },
+      "contiguous_gradients": true,
+      "overlap_comm": true
+    },
+    "bf16": {
+      "enabled": "auto"
+    },
+    "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": [
+          0.9,
+          0.999
+        ],
+        "eps": 1e-8,
+        "weight_decay": "auto"
+      }
+    },
+    "scheduler": {
+      "type": "WarmupDecayLR",
+      "params": {
+        "warmup_min_lr": "auto",
+        "warmup_max_lr": "auto",
+        "warmup_num_steps": "auto",
+        "total_num_steps": "auto"
+      }
+    },
+    "train_batch_size": "auto",
+    "train_micro_batch_size_per_gpu": "auto",
+    "wall_clock_breakdown": false
+}

deepspeed/zero3.json

@@ -37,18 +37,18 @@
       "lr": "auto",
       "betas": [
         0.9,
-        0.999
+        0.95
       ],
       "eps": 1e-8,
       "weight_decay": "auto"
     }
   },
   "scheduler": {
-    "type": "OneCycle",
+    "type": "WarmupLR",
     "params": {
-      "cycle_min_lr": 0.00001,
-      "cycle_max_lr": 0.00003,
-      "cycle_first_step_size": 120
+      "warmup_min_lr": "auto",
+      "warmup_max_lr": "auto",
+      "warmup_num_steps": "auto"
     }
   },
   "train_batch_size": "auto",

docker/Dockerfile

@@ -11,14 +11,13 @@ RUN apt-get update && \
 
 WORKDIR /workspace
 
-RUN pip3 install --force-reinstall "peft @ git+https://github.com/huggingface/peft.git@main"
 RUN git clone --depth=1 https://github.com/OpenAccess-AI-Collective/axolotl.git
 # If AXOLOTL_EXTRAS is set, append it in brackets
 RUN cd axolotl && \
     if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        pip install -e .[$AXOLOTL_EXTRAS]; \
+        pip install -e .[flash-attn,gptq,$AXOLOTL_EXTRAS]; \
     else \
-        pip install -e .; \
+        pip install -e .[flash-attn,gptq]; \
     fi
 
 # fix so that git fetch/pull from remote works

docker/Dockerfile-base

@@ -31,26 +31,6 @@ 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
 
-
-FROM base-builder AS flash-attn-builder
-
-WORKDIR /workspace
-
-ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 9.0+PTX"
-
-RUN git clone https://github.com/Dao-AILab/flash-attention.git && \
-    cd flash-attention && \
-    git checkout v2.0.1  && \
-    python3 setup.py bdist_wheel && \
-    cd csrc/fused_dense_lib && \
-    python3 setup.py bdist_wheel && \
-    cd ../xentropy && \
-    python3 setup.py bdist_wheel && \
-    cd ../rotary && \
-    python3 setup.py bdist_wheel && \
-    cd ../layer_norm && \
-    python3 setup.py bdist_wheel
-
 FROM base-builder AS deepspeed-builder
 
 ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 9.0+PTX"
@@ -90,13 +70,8 @@ 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
-COPY --from=flash-attn-builder /workspace/flash-attention/dist/flash_attn-*.whl wheels
-COPY --from=flash-attn-builder /workspace/flash-attention/csrc/fused_dense_lib/dist/fused_dense_lib-*.whl wheels
-COPY --from=flash-attn-builder /workspace/flash-attention/csrc/xentropy/dist/xentropy_cuda_lib-*.whl wheels
-COPY --from=flash-attn-builder /workspace/flash-attention/csrc/rotary/dist/rotary_emb-*.whl wheels
-COPY --from=flash-attn-builder /workspace/flash-attention/csrc/layer_norm/dist/dropout_layer_norm-*.whl wheels
 
-RUN pip3 install wheels/deepspeed-*.whl wheels/flash_attn-*.whl wheels/fused_dense_lib-*.whl wheels/xentropy_cuda_lib-*.whl wheels/rotary_emb-*.whl wheels/dropout_layer_norm-*.whl
+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 && \

examples/cerebras/qlora.yml

@@ -23,6 +23,7 @@ lora_target_modules:
 lora_target_linear:
 lora_fan_in_fan_out:
 wandb_project:
+wandb_entity:
 wandb_watch:
 wandb_run_id:
 wandb_log_model:
@@ -35,7 +36,7 @@ torchdistx_path:
 lr_scheduler: cosine
 learning_rate: 0.0002
 train_on_inputs: false
-group_by_length: true
+group_by_length: false
 bf16: true
 fp16: false
 tf32: true

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

@@ -0,0 +1,67 @@
+base_model: codellama/CodeLlama-13b-hf
+base_model_config: codellama/CodeLlama-13b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./lora-out
+
+sequence_len: 100000
+sample_packing: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 3
+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
+eval_steps: 20
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

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

@@ -0,0 +1,69 @@
+base_model: codellama/CodeLlama-13b-hf
+base_model_config: codellama/CodeLlama-13b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+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: last_run_prepared
+val_set_size: 0.01
+output_dir: ./qlora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 100000
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules:
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 3
+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
+eval_steps: 20
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

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

@@ -0,0 +1,67 @@
+base_model: codellama/CodeLlama-34b-hf
+base_model_config: codellama/CodeLlama-34b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./lora-out
+
+sequence_len: 100000
+sample_packing: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 3
+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
+eval_steps: 20
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

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

@@ -0,0 +1,69 @@
+base_model: codellama/CodeLlama-34b-hf
+base_model_config: codellama/CodeLlama-34b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+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: last_run_prepared
+val_set_size: 0.01
+output_dir: ./qlora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 100000
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules:
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 3
+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
+eval_steps: 20
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

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

@@ -0,0 +1,67 @@
+base_model: codellama/CodeLlama-7b-hf
+base_model_config: codellama/CodeLlama-7b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./lora-out
+
+sequence_len: 100000
+sample_packing: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 3
+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
+eval_steps: 20
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

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

@@ -0,0 +1,69 @@
+base_model: codellama/CodeLlama-7b-hf
+base_model_config: codellama/CodeLlama-7b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: CodeLlamaTokenizer
+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: last_run_prepared
+val_set_size: 0.01
+output_dir: ./qlora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 100000
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules:
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 3
+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
+eval_steps: 20
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/code-llama/README.md

@@ -0,0 +1,22 @@
+# Overview
+
+This is an example of CodeLLaMA configuration for 7b, 13b and 34b.
+
+The 7b variant fits on any 24GB VRAM GPU and will take up about 17 GB of VRAM during training if using qlora and 20 GB if using lora. On a RTX 4090 it trains 3 epochs of the default dataset in about 15 minutes.
+
+The 13b variant will fit if you change these settings to these values:
+gradient_accumulation_steps: 2
+micro_batch_size: 1
+
+The 34b variant does not fit on 24GB of VRAM - you will need something with +40 gb VRAM that also supports flash attention v2 - A6000 or A100 are good choices.
+
+```shell
+accelerate launch scripts/finetune.py examples/code-llama/[MODEL_SIZE]/qlora.yml
+
+```
+or
+
+```shell
+accelerate launch scripts/finetune.py examples/code-llama/[MODEL_SIZE]/lora.yml
+
+```

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

@@ -24,6 +24,7 @@ lora_target_modules:
 lora_target_linear: true
 lora_fan_in_fan_out:
 wandb_project:
+wandb_entity:
 wandb_watch:
 wandb_run_id:
 wandb_log_model:

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

@@ -38,6 +38,7 @@ lora_target_linear: true
 lora_fan_in_fan_out:
 
 wandb_project:
+wandb_entity:
 wandb_watch:
 wandb_run_id:
 wandb_log_model:

examples/falcon/config-7b.yml

@@ -24,6 +24,7 @@ lora_target_modules:
 lora_target_linear: true
 lora_fan_in_fan_out:
 wandb_project:
+wandb_entity:
 wandb_watch:
 wandb_run_id:
 wandb_log_model:

examples/gptj/qlora.yml

@@ -20,6 +20,7 @@ lora_target_modules:
 lora_target_linear: true
 lora_fan_in_fan_out:
 wandb_project:
+wandb_entity:
 wandb_watch:
 wandb_run_id:
 wandb_log_model:
@@ -32,7 +33,7 @@ torchdistx_path:
 lr_scheduler: cosine
 learning_rate: 0.0001
 train_on_inputs: false
-group_by_length: true
+group_by_length: false
 bf16: true
 fp16: false
 tf32: true

examples/gptq-lora-7b/README.md

@@ -1,8 +0,0 @@
-# LLaMa 7B using LoRA
-
-This is a good place to start for beginners. This will run on an NVIDIA RTX4090 with no other changes needed.
-
-```shell
-accelerate launch scripts/finetune.py examples/gptq-lora-7b/config.yml
-
-```

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

@@ -1,62 +0,0 @@
-base_model: Neko-Institute-of-Science/LLaMA-7B-4bit-128g
-base_model_config: Neko-Institute-of-Science/LLaMA-7B-4bit-128g
-model_type: LlamaForCausalLM
-tokenizer_type: LlamaTokenizer
-trust_remote_code:
-load_in_8bit: true
-gptq: true
-datasets:
-  - path: vicgalle/alpaca-gpt4
-    type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.02
-adapter:
-lora_model_dir:
-sequence_len: 2048
-max_packed_sequence_len:
-lora_r: 8
-lora_alpha: 16
-lora_dropout: 0.05
-lora_target_modules:
-  - q_proj
-  - v_proj
-lora_fan_in_fan_out: false
-wandb_project: llama-7b-lora-int4
-wandb_watch:
-wandb_run_id:
-wandb_log_model:
-output_dir: ./llama-7b-lora-int4
-gradient_accumulation_steps: 1
-micro_batch_size: 1
-num_epochs: 3
-optimizer: adamw_bnb_8bit
-torchdistx_path:
-lr_scheduler: cosine
-learning_rate: 0.0000002
-train_on_inputs: false
-group_by_length: false
-fp16: true
-bf16: false
-tf32: true
-early_stopping_patience:
-resume_from_checkpoint:
-local_rank:
-logging_steps: 5
-xformers_attention:
-flash_attention:
-gradient_checkpointing: true
-gptq_groupsize: 128
-gptq_model_v1: false
-warmup_steps: 20
-eval_steps: 110
-save_steps: 660
-debug:
-deepspeed:
-weight_decay: 0.0001
-fsdp:
-fsdp_config:
-tokens:
-  pad_token: "[PAD]"
-  bos_token: "<s>"
-  eos_token: "</s>"
-  unk_token: "<unk>"

examples/jeopardy-bot/config.yml

@@ -18,6 +18,7 @@ lora_dropout:
 lora_target_modules:
 lora_fan_in_fan_out: false
 wandb_project:
+wandb_entity:
 wandb_watch:
 wandb_run_id:
 wandb_log_model:

examples/llama-2/gptq-lora.yml

@@ -0,0 +1,76 @@
+base_model: TheBloke/Llama-2-7B-GPTQ
+base_model_config: TheBloke/Llama-2-7B-GPTQ
+is_llama_derived_model: false
+gptq: true
+gptq_bits: 4
+model_type: AutoModelForCausalLM
+tokenizer_type: LlamaTokenizer
+tokenizer_use_fast: true
+tokenizer_legacy: true
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+push_dataset_to_hub:
+hf_use_auth_token: true
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+adapter: lora
+lora_model_dir:
+sequence_len: 4096
+sample_packing:
+lora_r: 8
+lora_alpha: 32
+lora_dropout: 0.05
+lora_target_modules:
+  - k_proj
+  - o_proj
+  - q_proj
+  - v_proj
+lora_target_linear:
+lora_fan_in_fan_out:
+wandb_project:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+output_dir: ./model-out
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 3
+optimizer: adamw_torch
+adam_beta2: 0.95
+adam_eps: 0.00001
+max_grad_norm: 1.0
+torchdistx_path:
+lr_scheduler: cosine
+lr_quadratic_warmup: true
+learning_rate: 0.000017
+train_on_inputs: false
+group_by_length: false
+bf16: false
+fp16: false
+float16: true
+tf32: true
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention:
+sdp_attention:
+flash_optimum:
+gptq_groupsize:
+gptq_model_v1:
+warmup_steps: 100
+eval_steps:
+save_steps:
+debug:
+deepspeed:
+weight_decay: 0.1
+special_tokens:
+  bos_token: "<s>"
+  eos_token: "</s>"
+  unk_token: "<unk>"

examples/llama-2/lora.yml

@@ -2,6 +2,7 @@ base_model: meta-llama/Llama-2-7b-hf
 base_model_config: meta-llama/Llama-2-7b-hf
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
+is_llama_derived_model: true
 
 load_in_8bit: true
 load_in_4bit: false
@@ -15,7 +16,7 @@ val_set_size: 0.01
 output_dir: ./lora-out
 
 sequence_len: 4096
-max_packed_sequence_len: 4096
+sample_packing: true
 
 adapter: lora
 lora_model_dir:
@@ -26,6 +27,7 @@ lora_target_linear: true
 lora_fan_in_fan_out:
 
 wandb_project:
+wandb_entity:
 wandb_watch:
 wandb_run_id:
 wandb_log_model:
@@ -38,7 +40,7 @@ lr_scheduler: cosine
 learning_rate: 0.0002
 
 train_on_inputs: false
-group_by_length: true
+group_by_length: false
 bf16: true
 fp16: false
 tf32: false
@@ -48,8 +50,8 @@ early_stopping_patience:
 resume_from_checkpoint:
 local_rank:
 logging_steps: 1
-xformers_attention: true
-flash_attention:
+xformers_attention:
+flash_attention: true
 
 warmup_steps: 10
 eval_steps: 20
@@ -63,4 +65,3 @@ special_tokens:
   bos_token: "<s>"
   eos_token: "</s>"
   unk_token: "<unk>"
-  pad_token: "<pad>"

examples/llama-2/qlora.yml

@@ -2,6 +2,7 @@ base_model: meta-llama/Llama-2-7b-hf
 base_model_config: meta-llama/Llama-2-7b-hf
 model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
+is_llama_derived_model: true
 
 load_in_8bit: false
 load_in_4bit: true
@@ -18,7 +19,8 @@ adapter: qlora
 lora_model_dir:
 
 sequence_len: 4096
-max_packed_sequence_len: 4096
+sample_packing: true
+
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
@@ -27,6 +29,7 @@ lora_target_linear: true
 lora_fan_in_fan_out:
 
 wandb_project:
+wandb_entity:
 wandb_watch:
 wandb_run_id:
 wandb_log_model:
@@ -39,7 +42,7 @@ lr_scheduler: cosine
 learning_rate: 0.0002
 
 train_on_inputs: false
-group_by_length: true
+group_by_length: false
 bf16: true
 fp16: false
 tf32: false
@@ -49,8 +52,8 @@ early_stopping_patience:
 resume_from_checkpoint:
 local_rank:
 logging_steps: 1
-xformers_attention: true
-flash_attention:
+xformers_attention:
+flash_attention: true
 
 warmup_steps: 10
 eval_steps: 20
@@ -64,4 +67,3 @@ special_tokens:
   bos_token: "<s>"
   eos_token: "</s>"
   unk_token: "<unk>"
-  pad_token: "<pad>"

examples/llama-2/relora.yml

@@ -0,0 +1,73 @@
+base_model: meta-llama/Llama-2-7b-hf
+base_model_config: meta-llama/Llama-2-7b-hf
+model_type: LlamaForCausalLM
+tokenizer_type: LlamaTokenizer
+is_llama_derived_model: true
+
+load_in_8bit: false
+load_in_4bit: true
+strict: false
+
+datasets:
+  - path: teknium/GPT4-LLM-Cleaned
+    type: alpaca
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./relora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 4096
+sample_packing: true
+
+lora_r: 8
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules:
+lora_target_linear: true
+lora_fan_in_fan_out:
+
+relora_steps: 150
+relora_warmup_steps: 10
+relora_cpu_offload: false
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_run_id:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 4
+num_epochs: 3
+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
+eval_steps: 20
+save_steps: 50
+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

@@ -20,6 +20,7 @@ lora_target_modules:
   - v_proj
 lora_fan_in_fan_out: false
 wandb_project: mpt-alpaca-7b
+wandb_entity:
 wandb_watch:
 wandb_run_id:
 wandb_log_model:

examples/openllama-3b/config.yml

@@ -22,6 +22,7 @@ lora_target_modules:
 lora_target_linear:
 lora_fan_in_fan_out:
 wandb_project:
+wandb_entity:
 wandb_watch:
 wandb_run_id:
 wandb_log_model:

examples/openllama-3b/lora.yml

@@ -28,6 +28,7 @@ lora_target_modules:
   - o_proj
 lora_fan_in_fan_out:
 wandb_project:
+wandb_entity:
 wandb_watch:
 wandb_run_id:
 wandb_log_model:

examples/openllama-3b/qlora.yml

@@ -22,6 +22,7 @@ lora_target_modules:
 lora_target_linear: true
 lora_fan_in_fan_out:
 wandb_project:
+wandb_entity:
 wandb_watch:
 wandb_run_id:
 wandb_log_model:
@@ -34,7 +35,7 @@ torchdistx_path:
 lr_scheduler: cosine
 learning_rate: 0.0002
 train_on_inputs: false
-group_by_length: true
+group_by_length: false
 bf16: true
 fp16: false
 tf32: true

examples/pythia-12b/config.yml

@@ -23,6 +23,7 @@ lora_target_modules:
 lora_target_linear: true
 lora_fan_in_fan_out: true  # pythia/GPTNeoX lora specific
 wandb_project:
+wandb_entity:
 wandb_watch:
 wandb_run_id:
 wandb_log_model:
@@ -46,4 +47,3 @@ local_rank:
 gradient_checkpointing: true
 fsdp:
 fsdp_config:
-collator_pad_to_longest: true

examples/pythia/lora.yml

@@ -17,6 +17,7 @@ lora_target_modules:
 lora_target_linear:
 lora_fan_in_fan_out: true  # pythia/GPTNeoX lora specific
 wandb_project:
+wandb_entity:
 wandb_watch:
 wandb_run_id:
 wandb_log_model:

examples/redpajama/config-3b.yml

@@ -21,6 +21,7 @@ lora_target_modules:
   - v_proj
 lora_fan_in_fan_out: false
 wandb_project: redpajama-alpaca-3b
+wandb_entity:
 wandb_watch:
 wandb_run_id:
 wandb_log_model:

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

@@ -20,6 +20,7 @@ lora_target_modules:
   - mlp_down
 lora_fan_in_fan_out:
 wandb_project: lora-replit
+wandb_entity:
 wandb_watch:
 wandb_run_id:
 wandb_log_model:

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

@@ -37,6 +37,7 @@ lora_target_linear: true
 lora_fan_in_fan_out:
 
 wandb_project:
+wandb_entity:
 wandb_watch:
 wandb_run_id:
 wandb_log_model:

requirements.txt

@@ -1,21 +1,31 @@
+--extra-index-url https://download.pytorch.org/whl/cu118
+--extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
+torch==2.0.1
+auto-gptq
+packaging
 peft @ git+https://github.com/huggingface/peft.git
 transformers @ git+https://github.com/huggingface/transformers.git
-bitsandbytes>=0.39.0
+bitsandbytes>=0.41.1
 accelerate @ git+https://github.com/huggingface/accelerate@2a289f6108e77a77a4efffb3f6316bc98538413b
 addict
 fire
-PyYAML==6.0
+PyYAML>=6.0
 datasets
-accelerate>=0.19.0
+flash-attn>=2.0.8
 sentencepiece
 wandb
 einops
 xformers
 optimum
 hf_transfer
+colorama
+numba
+numpy>=1.24.4
 # qlora things
 bert-score==0.3.13
 evaluate==0.4.0
 rouge-score==0.1.2
 scipy
 scikit-learn==1.2.2
+pynvml
+art

scripts/alpaca_json_to_jsonl.py

@@ -1,52 +0,0 @@
-"""Module to convert json file to jsonl"""
-
-import os
-import sys
-from pathlib import Path
-from typing import Optional, Union
-
-import fire
-
-from axolotl.convert import (
-    FileReader,
-    FileWriter,
-    JsonlSerializer,
-    JsonParser,
-    JsonToJsonlConverter,
-    StdoutWriter,
-)
-from axolotl.logging_config import configure_logging
-
-configure_logging()
-
-# add src to the pythonpath so we don't need to pip install this
-project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
-src_dir = os.path.join(project_root, "src")
-sys.path.insert(0, src_dir)
-
-
-def main(
-    file: Path,
-    output: Optional[Path] = None,
-    to_stdout: Optional[bool] = False,
-):
-    """
-    Convert a json file to jsonl
-    """
-
-    file_reader = FileReader()
-    writer: Union[StdoutWriter, FileWriter]
-    if to_stdout or output is None:
-        writer = StdoutWriter()
-    else:
-        writer = FileWriter(output)
-    json_parser = JsonParser()
-    jsonl_serializer = JsonlSerializer()
-
-    converter = JsonToJsonlConverter(file_reader, writer, json_parser, jsonl_serializer)
-
-    converter.convert(file, output)
-
-
-if __name__ == "__main__":
-    fire.Fire(main)

scripts/finetune.py

@@ -6,24 +6,28 @@ import os
 import random
 import signal
 import sys
+from dataclasses import dataclass, field
 from pathlib import Path
 from typing import Any, Dict, List, Optional, Union
 
 import fire
 import torch
+import transformers
 import yaml
 
 # add src to the pythonpath so we don't need to pip install this
+from art import text2art
 from optimum.bettertransformer import BetterTransformer
 from transformers import GenerationConfig, TextStreamer
 
 from axolotl.logging_config import configure_logging
-from axolotl.utils.data import load_prepare_datasets, load_pretraining_dataset
+from axolotl.utils.config import normalize_config, validate_config
+from axolotl.utils.data import prepare_dataset
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.models import load_model, load_tokenizer
+from axolotl.utils.distributed import is_main_process
+from axolotl.utils.models import load_model, load_model_config, load_tokenizer
 from axolotl.utils.tokenization import check_dataset_labels
 from axolotl.utils.trainer import setup_trainer
-from axolotl.utils.validation import validate_config
 from axolotl.utils.wandb import setup_wandb_env_vars
 
 project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
@@ -33,30 +37,31 @@ sys.path.insert(0, src_dir)
 configure_logging()
 LOG = logging.getLogger("axolotl.scripts")
 
-
-DEFAULT_DATASET_PREPARED_PATH = "last_run_prepared"
 os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
 
 
-def choose_device(cfg):
-    def get_device():
-        try:
-            if torch.cuda.is_available():
-                return f"cuda:{cfg.local_rank}"
+@dataclass
+class TrainerCliArgs:
+    """
+    dataclass representing the various non-training arguments
+    """
 
-            if torch.backends.mps.is_available():
-                return "mps"
+    debug: bool = field(default=False)
+    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)
 
-            raise SystemError("No CUDA/mps device found")
-        except Exception:  # pylint: disable=broad-exception-caught
-            return "cpu"
 
-    cfg.device = get_device()
-    if cfg.device_map != "auto":
-        if cfg.device.startswith("cuda"):
-            cfg.device_map = {"": cfg.local_rank}
-        else:
-            cfg.device_map = {"": cfg.device}
+def print_axolotl_text_art(suffix=None):
+    font = "nancyj"
+    ascii_text = "  axolotl"
+    if suffix:
+        ascii_text += f"  x  {suffix}"
+    ascii_art = text2art(" axolotl", font=font)
+    if is_main_process():
+        print(ascii_art)
 
 
 def get_multi_line_input() -> Optional[str]:
@@ -69,6 +74,8 @@ def get_multi_line_input() -> Optional[str]:
 
 
 def do_inference(cfg, model, tokenizer, prompter: Optional[str]):
+    if prompter == "None":
+        prompter = None
     default_tokens = {"unk_token": "<unk>", "bos_token": "<s>", "eos_token": "</s>"}
 
     for token, symbol in default_tokens.items():
@@ -90,6 +97,8 @@ def do_inference(cfg, model, tokenizer, prompter: Optional[str]):
             max_seq_len=255, mem_freq=50, top_k=5, max_cache_size=None
         )
 
+    model = model.to(cfg.device)
+
     while True:
         print("=" * 80)
         # support for multiline inputs
@@ -141,6 +150,10 @@ def choose_config(path: Path):
             "No YAML config files found in the specified directory. Are you using a .yml extension?"
         )
 
+    if len(yaml_files) == 1:
+        print(f"Using default YAML file '{yaml_files[0]}'")
+        return yaml_files[0]
+
     print("Choose a YAML file:")
     for idx, file in enumerate(yaml_files):
         print(f"{idx + 1}. {file}")
@@ -164,80 +177,20 @@ def check_not_in(list1: List[str], list2: Union[Dict[str, Any], List[str]]) -> b
 
 
 def train(
-    config: Path = Path("configs/"),
-    prepare_ds_only: bool = False,
-    **kwargs,
+    *,
+    cfg: DictDefault,
+    cli_args: TrainerCliArgs,
 ):
-    if Path(config).is_dir():
-        config = choose_config(config)
-
-    # load the config from the yaml file
-    with open(config, encoding="utf-8") as file:
-        cfg: DictDefault = DictDefault(yaml.safe_load(file))
-    # 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()
-    for k, _ in kwargs.items():
-        # if not strict, allow writing to cfg even if it's not in the yml already
-        if k in cfg_keys or not cfg.strict:
-            # handle booleans
-            if isinstance(cfg[k], bool):
-                cfg[k] = bool(kwargs[k])
-            else:
-                cfg[k] = kwargs[k]
-
-    validate_config(cfg)
-
-    # setup some derived config / hyperparams
-    cfg.gradient_accumulation_steps = cfg.gradient_accumulation_steps or (
-        cfg.batch_size // cfg.micro_batch_size
-    )
-    cfg.batch_size = (
-        cfg.batch_size or cfg.micro_batch_size * cfg.gradient_accumulation_steps
-    )
-    cfg.world_size = int(os.environ.get("WORLD_SIZE", 1))
-    cfg.local_rank = int(os.environ.get("LOCAL_RANK", 0))
-    choose_device(cfg)
-    cfg.ddp = cfg.ddp if cfg.ddp is not None else cfg.world_size != 1
-    if cfg.ddp:
-        cfg.device_map = {"": int(os.environ.get("LOCAL_RANK", 0))}
-        cfg.batch_size = cfg.batch_size * cfg.world_size
-
-    setup_wandb_env_vars(cfg)
-    if cfg.device == "mps":
-        cfg.load_in_8bit = False
-        cfg.tf32 = False
-        if cfg.bf16:
-            cfg.fp16 = True
-        cfg.bf16 = False
-
-    if cfg.tf32:
-        torch.backends.cuda.matmul.allow_tf32 = True
-
     # load the tokenizer first
-    tokenizer_config = cfg.tokenizer_config or cfg.base_model_config
-    LOG.info(f"loading tokenizer... {tokenizer_config}")
-    tokenizer = load_tokenizer(tokenizer_config, cfg.tokenizer_type, cfg)
+    LOG.info(f"loading tokenizer... {cfg.tokenizer_config or cfg.base_model_config}")
+    tokenizer = load_tokenizer(cfg)
 
-    if (
-        check_not_in(["shard", "merge_lora"], kwargs) and not cfg.inference
+    if not (
+        cli_args.shard or cli_args.merge_lora or cli_args.inference
     ):  # don't need to load dataset for these
-        if not cfg.pretraining_dataset:
-            train_dataset, eval_dataset = load_prepare_datasets(
-                tokenizer, cfg, DEFAULT_DATASET_PREPARED_PATH
-            )
-        else:
-            train_dataset = load_pretraining_dataset(
-                cfg.pretraining_dataset,
-                tokenizer,
-                max_tokens=cfg.sequence_len,
-                seed=cfg.seed,
-            )
-            # https://discuss.huggingface.co/t/how-to-use-huggingface-trainer-streaming-datasets-without-wrapping-it-with-torchdatas-iterablewrapper/25230
-            train_dataset = train_dataset.with_format("torch")
-            eval_dataset = None
+        train_dataset, eval_dataset, total_num_steps = prepare_dataset(cfg, tokenizer)
 
-    if cfg.debug or "debug" in kwargs:
+    if cli_args.debug or cfg.debug:
         LOG.info("check_dataset_labels...")
         check_dataset_labels(
             train_dataset.select(
@@ -246,47 +199,58 @@ def train(
             tokenizer,
         )
 
-    if prepare_ds_only:
+    if cli_args.prepare_ds_only:
         LOG.info("Finished preparing dataset. Exiting...")
         return
 
     # Load the model and tokenizer
-    LOG.info("loading model and peft_config...")
-    model, peft_config = load_model(
-        cfg.base_model,
-        cfg.base_model_config,
-        cfg.model_type,
-        tokenizer,
-        cfg,
-        adapter=cfg.adapter,
-    )
+    LOG.info("loading model and (optionally) peft_config...")
+    model, peft_config = load_model(cfg, tokenizer, inference=cli_args.inference)
 
-    if "merge_lora" in kwargs and cfg.adapter is not None:
+    safe_serialization = cfg.save_safetensors is True
+
+    if cli_args.merge_lora and cfg.adapter is not None:
         LOG.info("running merge of LoRA with base model")
         model = model.merge_and_unload()
         model.to(dtype=torch.float16)
 
         if cfg.local_rank == 0:
             LOG.info("saving merged model")
-            model.save_pretrained(str(Path(cfg.output_dir) / "merged"))
+            model.save_pretrained(
+                str(Path(cfg.output_dir) / "merged"),
+                safe_serialization=safe_serialization,
+            )
+            tokenizer.save_pretrained(str(Path(cfg.output_dir) / "merged"))
         return
 
-    if cfg.inference:
-        LOG.info("calling do_inference function")
-        prompter: Optional[str] = "AlpacaPrompter"
-        if "prompter" in kwargs:
-            if kwargs["prompter"] == "None":
-                prompter = None
-            else:
-                prompter = kwargs["prompter"]
-        do_inference(cfg, model, tokenizer, prompter=prompter)
+    if cli_args.inference:
+        LOG.debug("Running inference on model")
+        do_inference(cfg, model, tokenizer, prompter=cli_args.prompter)
         return
 
-    if "shard" in kwargs:
-        model.save_pretrained(cfg.output_dir)
+    if cli_args.shard:
+        LOG.debug("Re-saving model w/ sharding")
+        model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
         return
 
-    trainer = setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer)
+    if cfg.resume_from_checkpoint is None and cfg.auto_resume_from_checkpoints:
+        possible_checkpoints = [
+            str(cp) for cp in Path(cfg.output_dir).glob("checkpoint-*")
+        ]
+        if len(possible_checkpoints) > 0:
+            sorted_paths = sorted(
+                possible_checkpoints,
+                key=lambda path: int(path.split("-")[-1]),
+            )
+            cfg.resume_from_checkpoint = sorted_paths[-1]
+            LOG.info(
+                f"Using Auto-resume functionality to start with checkpoint at {cfg.resume_from_checkpoint}"
+            )
+    resume_from_checkpoint = cfg.resume_from_checkpoint
+
+    trainer = setup_trainer(
+        cfg, train_dataset, eval_dataset, model, tokenizer, total_num_steps
+    )
 
     model.config.use_cache = False
 
@@ -305,7 +269,7 @@ def train(
         def terminate_handler(_, __, model):
             if cfg.flash_optimum:
                 model = BetterTransformer.reverse(model)
-            model.save_pretrained(cfg.output_dir)
+            model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
             sys.exit(0)
 
         signal.signal(
@@ -315,23 +279,10 @@ def train(
     LOG.info("Starting trainer...")
     if cfg.group_by_length:
         LOG.info("hang tight... sorting dataset for group_by_length")
-    resume_from_checkpoint = cfg.resume_from_checkpoint
-    if cfg.resume_from_checkpoint is None and cfg.auto_resume_from_checkpoints:
-        possible_checkpoints = [
-            str(cp) for cp in Path(cfg.output_dir).glob("checkpoint-*")
-        ]
-        if len(possible_checkpoints) > 0:
-            sorted_paths = sorted(
-                possible_checkpoints,
-                key=lambda path: int(path.split("-")[-1]),
-            )
-            resume_from_checkpoint = sorted_paths[-1]
-            LOG.info(
-                f"Using Auto-resume functionality to start with checkpoint at {resume_from_checkpoint}"
-            )
 
     if not Path(cfg.output_dir).is_dir():
         os.makedirs(cfg.output_dir, exist_ok=True)
+    tokenizer.save_pretrained(cfg.output_dir)
     if cfg.flash_optimum:
         with torch.backends.cuda.sdp_kernel(
             enable_flash=True, enable_math=True, enable_mem_efficient=True
@@ -342,17 +293,69 @@ def train(
 
     LOG.info(f"Training Completed!!! Saving pre-trained model to {cfg.output_dir}")
 
+    if cfg.relora_steps:
+        if cfg.adapter == "lora" and not (cfg.load_in_4bit or cfg.load_in_8bit):
+            model = model.merge_and_unload()
+        else:
+            # final model weights have already been saved by `ReLoRACallback.on_train_end`
+            return
+
     # TODO do we need this fix? https://huggingface.co/docs/accelerate/usage_guides/fsdp#saving-and-loading
     # only save on rank 0, otherwise it corrupts output on multi-GPU when multiple processes attempt to write the same file
     if cfg.fsdp:
-        model.save_pretrained(cfg.output_dir)
+        trainer.save_model(cfg.output_dir)
     elif cfg.local_rank == 0:
         if cfg.flash_optimum:
             model = BetterTransformer.reverse(model)
-        model.save_pretrained(cfg.output_dir)
 
-    # trainer.save_model(cfg.output_dir)  # TODO this may be needed for deepspeed to work? need to review another time
+        model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
+
+
+def load_cfg(config: Path = Path("examples/"), **kwargs):
+    if Path(config).is_dir():
+        config = choose_config(config)
+
+    # load the config from the yaml file
+    with open(config, encoding="utf-8") as file:
+        cfg: DictDefault = DictDefault(yaml.safe_load(file))
+    # 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()
+    for k, _ in kwargs.items():
+        # if not strict, allow writing to cfg even if it's not in the yml already
+        if k in cfg_keys or not cfg.strict:
+            # handle booleans
+            if isinstance(cfg[k], bool):
+                cfg[k] = bool(kwargs[k])
+            else:
+                cfg[k] = kwargs[k]
+
+    model_config = load_model_config(cfg)
+
+    # figure out if the model is llama
+    cfg.is_llama_derived_model = (
+        (hasattr(model_config, "model_type") and model_config.model_type == "llama")
+        or cfg.is_llama_derived_model
+        or "llama" in cfg.base_model
+        or (cfg.model_type and "llama" in cfg.model_type.lower())
+    )
+    validate_config(cfg)
+
+    normalize_config(cfg)
+
+    setup_wandb_env_vars(cfg)
+    return cfg
+
+
+def do_train(config: Path = Path("examples/"), **kwargs):
+    print_axolotl_text_art()
+    parsed_cfg = load_cfg(config, **kwargs)
+    parser = transformers.HfArgumentParser((TrainerCliArgs))
+    parsed_cli_args, _ = parser.parse_args_into_dataclasses(
+        return_remaining_strings=True
+    )
+    train(cfg=parsed_cfg, cli_args=parsed_cli_args)
 
 
 if __name__ == "__main__":
-    fire.Fire(train)
+    fire.Fire(do_train)

setup.py

@@ -2,14 +2,27 @@
 
 from setuptools import find_packages, setup
 
-install_requires = []
-with open("./requirements.txt", encoding="utf-8") as requirements_file:
-    # don't include peft yet until we check the int4
-    # need to manually install peft for now...
-    reqs = [r.strip() for r in requirements_file.readlines() if "peft" not in r]
-    reqs = [r for r in reqs if r and r[0] != "#"]
-    for r in reqs:
-        install_requires.append(r)
+
+def parse_requirements():
+    _install_requires = []
+    _dependency_links = []
+    with open("./requirements.txt", encoding="utf-8") as requirements_file:
+        lines = [
+            r.strip() for r in requirements_file.readlines() if "auto-gptq" not in r
+        ]
+        for line in lines:
+            if line.startswith("--extra-index-url"):
+                # Handle custom index URLs
+                _, url = line.split()
+                _dependency_links.append(url)
+            elif "flash-attn" not in line and line and line[0] != "#":
+                # Handle standard packages
+                _install_requires.append(line)
+    return _install_requires, _dependency_links
+
+
+install_requires, dependency_links = parse_requirements()
+
 
 setup(
     name="axolotl",
@@ -18,15 +31,15 @@ setup(
     package_dir={"": "src"},
     packages=find_packages(),
     install_requires=install_requires,
+    dependency_links=dependency_links,
     extras_require={
         "gptq": [
-            "alpaca_lora_4bit @ git+https://github.com/winglian/alpaca_lora_4bit.git@setup_pip",
+            "auto-gptq",
         ],
-        "gptq_triton": [
-            "alpaca_lora_4bit[triton] @ git+https://github.com/winglian/alpaca_lora_4bit.git@setup_pip",
+        "flash-attn": [
+            "flash-attn==2.0.8",
         ],
         "extras": [
-            "flash-attn",
             "deepspeed",
         ],
     },

src/axolotl/datasets.py

@@ -5,7 +5,7 @@ import os
 from typing import List
 
 import torch
-from datasets import IterableDataset
+from datasets import Dataset, IterableDataset
 
 from .prompt_tokenizers import PromptTokenizingStrategy
 
@@ -18,9 +18,9 @@ from .prompt_tokenizers import PromptTokenizingStrategy
 LOG = logging.getLogger("axolotl")
 
 
-class TokenizedPromptDataset(IterableDataset):
+class TokenizedPromptDataset(Dataset):
     """
-    Iterable dataset that returns tokenized prompts from a stream of text files.
+    Dataset that returns tokenized prompts from a stream of text files.
         Args:
             prompt_tokenizer (PromptTokenizingStrategy): The prompt tokenizing method for proccessing the data.
             dataset (dataset.Dataset): Dataset with text files.
@@ -30,19 +30,18 @@ class TokenizedPromptDataset(IterableDataset):
         self,
         prompt_tokenizer: PromptTokenizingStrategy,
         dataset: IterableDataset,
+        **kwargs,
     ):
         self.prompt_tokenizer = prompt_tokenizer
-        self.dataset = dataset
+        super().__init__(self.process(dataset).data, **kwargs)
 
-    def __iter__(self):
-        features = self.dataset.features.keys()
-        num_proc = os.cpu_count()
-        return iter(
-            self.dataset.map(
-                self.prompt_tokenizer.tokenize_prompt,
-                num_proc=num_proc,
-                remove_columns=features,
-            )
+    def process(self, dataset):
+        features = dataset.features.keys()
+        num_proc = min(64, os.cpu_count())
+        return dataset.map(
+            self.prompt_tokenizer.tokenize_prompt,
+            num_proc=num_proc,
+            remove_columns=features,
         )
 
 
@@ -77,14 +76,21 @@ class ConstantLengthDataset(IterableDataset):
             self.tokens_dtype = torch.int64
 
     def __iter__(self):
-        buffer = {"input_ids": [], "attention_mask": [], "labels": []}
+        buffer = {
+            "input_ids": [],
+            "attention_mask": [],
+            "labels": [],
+            "position_ids": [],
+        }
         buffer_len = 0
         for dataset in self.datasets:
+            idx = 0
             iterator = iter(dataset)
             more_examples = True
             while more_examples:
                 try:
                     example = next(iterator)
+                    idx += 1
                 except StopIteration:
                     more_examples = False
                     example = None
@@ -106,6 +112,9 @@ class ConstantLengthDataset(IterableDataset):
                         attention_mask = torch.cat(buffer["attention_mask"], dim=-1)[
                             : self.seq_length
                         ]
+                        position_ids = torch.cat(buffer["position_ids"], dim=-1)[
+                            : self.seq_length
+                        ]
                         labels = torch.cat(buffer["labels"], dim=-1)[: self.seq_length]
                         if labels.size() == input_ids.size() and (
                             attention_mask.size() == input_ids.size()
@@ -114,6 +123,7 @@ class ConstantLengthDataset(IterableDataset):
                                 "input_ids": input_ids,
                                 "labels": labels,
                                 "attention_mask": attention_mask,
+                                "position_ids": position_ids,
                             }
                         else:
                             LOG.warning(
@@ -123,8 +133,10 @@ class ConstantLengthDataset(IterableDataset):
                         "input_ids": [],
                         "attention_mask": [],
                         "labels": [],
+                        "position_ids": [],
                     }
                     buffer_len = 0
+                    idx = 1
 
                 if example:
                     # FIXME
@@ -133,11 +145,6 @@ class ConstantLengthDataset(IterableDataset):
                         input_ids = example["input_ids"]
                         attention_mask = example["attention_mask"]
                         labels = example["labels"]
-                        if (
-                            buffer["input_ids"]
-                            and input_ids[0] == self.tokenizer.bos_token_id
-                        ):
-                            attention_mask[0] = 0
 
                         if add_concat_token:
                             input_ids.append(self.concat_token_id)
@@ -148,13 +155,17 @@ class ConstantLengthDataset(IterableDataset):
                             input_ids, dtype=self.tokens_dtype
                         )
                         attention_mask_with_concat = torch.tensor(
-                            attention_mask, dtype=self.tokens_dtype
+                            [idx * m for m in attention_mask], dtype=torch.int16
                         )
                         labels_with_concat = torch.tensor(
                             labels, dtype=self.tokens_dtype
                         )
+                        position_ids = torch.arange(
+                            len(input_ids), dtype=self.tokens_dtype
+                        )
 
                         buffer["input_ids"].append(input_ids_with_concat)
                         buffer["attention_mask"].append(attention_mask_with_concat)
                         buffer["labels"].append(labels_with_concat)
+                        buffer["position_ids"].append(position_ids)
                         buffer_len += len(input_ids)

src/axolotl/logging_config.py

@@ -1,16 +1,42 @@
-"""Logging configuration settings"""
+"""
+Common logging module for axolotl
+"""
 
 import os
 import sys
+from logging import Formatter
 from logging.config import dictConfig
 from typing import Any, Dict
 
+from colorama import Fore, Style, init
+
+
+class ColorfulFormatter(Formatter):
+    """
+    Formatter to add coloring to log messages by log type
+    """
+
+    COLORS = {
+        "WARNING": Fore.YELLOW,
+        "ERROR": Fore.RED,
+        "CRITICAL": Fore.RED + Style.BRIGHT,
+    }
+
+    def format(self, record):
+        log_message = super().format(record)
+        return self.COLORS.get(record.levelname, "") + log_message + Fore.RESET
+
+
 DEFAULT_LOGGING_CONFIG: Dict[str, Any] = {
     "version": 1,
     "formatters": {
         "simple": {
             "format": "[%(asctime)s] [%(levelname)s] [%(name)s.%(funcName)s:%(lineno)d] [PID:%(process)d] %(message)s",
         },
+        "colorful": {
+            "()": ColorfulFormatter,
+            "format": "[%(asctime)s] [%(levelname)s] [%(name)s.%(funcName)s:%(lineno)d] [PID:%(process)d] %(message)s",
+        },
     },
     "filters": {},
     "handlers": {
@@ -20,14 +46,25 @@ DEFAULT_LOGGING_CONFIG: Dict[str, Any] = {
             "filters": [],
             "stream": sys.stdout,
         },
+        "color_console": {
+            "class": "logging.StreamHandler",
+            "formatter": "colorful",
+            "filters": [],
+            "stream": sys.stdout,
+        },
     },
     "root": {"handlers": ["console"], "level": os.getenv("LOG_LEVEL", "INFO")},
     "loggers": {
-        "axolotl": {"handlers": ["console"], "level": "DEBUG", "propagate": False},
+        "axolotl": {
+            "handlers": ["color_console"],
+            "level": "DEBUG",
+            "propagate": False,
+        },
     },
 }
 
 
 def configure_logging():
     """Configure with default logging"""
+    init()  # Initialize colorama
     dictConfig(DEFAULT_LOGGING_CONFIG)

src/axolotl/monkeypatch/llama_attn_hijack_flash.py

@@ -2,122 +2,47 @@
 
 # copied from https://github.com/lm-sys/FastChat/blob/main/fastchat/train/llama_flash_attn_monkey_patch.py
 
-from typing import Optional, Tuple
+import warnings
+from typing import List, Optional, Tuple, Union
 
 import torch
+import torch.nn.functional as F
 import transformers
 from einops import rearrange
 from flash_attn.bert_padding import pad_input, unpad_input
-from flash_attn.flash_attn_interface import flash_attn_varlen_qkvpacked_func
-from transformers.models.llama.modeling_llama import apply_rotary_pos_emb
+from transformers.modeling_outputs import BaseModelOutputWithPast
+from transformers.models.llama.modeling_llama import (
+    LlamaDecoderLayer as OriginalLlamaDecoderLayer,
+)
+from transformers.models.llama.modeling_llama import apply_rotary_pos_emb, repeat_kv
 
+from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
 
-def forward(
-    self,
-    hidden_states: torch.Tensor,
-    attention_mask: Optional[torch.Tensor] = None,
-    position_ids: Optional[torch.Tensor] = None,
-    past_key_value: Optional[Tuple[torch.Tensor]] = None,
-    output_attentions: bool = False,
-    use_cache: bool = False,
-) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-    """Input shape: Batch x Time x Channel
-
-    attention_mask: [bsz, q_len]
-    """
-    # pylint: disable=duplicate-code
-    bsz, q_len, _ = hidden_states.size()
-
-    query_states = (
-        self.q_proj(hidden_states)
-        .view(bsz, q_len, self.num_heads, self.head_dim)
-        .transpose(1, 2)
+try:
+    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,
     )
-    key_states = (
-        self.k_proj(hidden_states)
-        .view(bsz, q_len, self.num_heads, self.head_dim)
-        .transpose(1, 2)
+except ImportError:
+    from flash_attn.flash_attn_interface import (
+        flash_attn_unpadded_kvpacked_func as flash_attn_varlen_kvpacked_func,
     )
-    value_states = (
-        self.v_proj(hidden_states)
-        .view(bsz, q_len, self.num_heads, self.head_dim)
-        .transpose(1, 2)
+    from flash_attn.flash_attn_interface import (
+        flash_attn_unpadded_qkvpacked_func as flash_attn_varlen_qkvpacked_func,
     )
-    # [bsz, q_len, nh, hd]
-    # [bsz, nh, q_len, hd]
 
-    kv_seq_len = key_states.shape[-2]
-    assert past_key_value is None, "past_key_value is not supported"
 
-    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
+def replace_llama_attn_with_flash_attn(packed: Optional[bool] = False):
+    transformers.models.llama.modeling_llama.LlamaModel._prepare_decoder_attention_mask = (  # pylint: disable=protected-access
+        _prepare_decoder_attention_mask
     )
-    # [bsz, nh, t, hd]
-    assert not output_attentions, "output_attentions is not supported"
-    assert not use_cache, "use_cache is not supported"
-
-    # Flash attention codes from
-    # https://github.com/HazyResearch/flash-attention/blob/main/flash_attn/flash_attention.py
-
-    # transform the data into the format required by flash attention
-    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]
-    # 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
-
-    if key_padding_mask is None:
-        qkv = rearrange(qkv, "b s ... -> (b s) ...")
-        max_s = q_len
-        cu_q_lens = torch.arange(
-            0,
-            (bsz + 1) * q_len,
-            step=q_len,
-            dtype=torch.int32,
-            device=qkv.device,
+    transformers.models.llama.modeling_llama.LlamaAttention.forward = flashattn_forward
+    if packed:
+        transformers.models.llama.modeling_llama.LlamaDecoderLayer = LlamaDecoderLayer
+        transformers.models.llama.modeling_llama.LlamaModel.forward = (
+            llama_model_forward
         )
-        output = flash_attn_varlen_qkvpacked_func(
-            qkv, cu_q_lens, max_s, 0.0, softmax_scale=None, causal=True
-        )
-        output = rearrange(output, "(b s) ... -> b s ...", b=bsz)
-    else:
-        nheads = qkv.shape[-2]
-
-        # pylint: disable=invalid-name
-        x = rearrange(qkv, "b s three h d -> b s (three h d)")
-        x_unpad, indices, cu_q_lens, max_s = unpad_input(x, key_padding_mask)
-        x_unpad = rearrange(
-            x_unpad,
-            "nnz (three h d) -> nnz three h d",
-            three=3,
-            h=nheads,
-        )
-        output_unpad = flash_attn_varlen_qkvpacked_func(
-            x_unpad,
-            cu_q_lens,
-            max_s,
-            0.0,
-            softmax_scale=None,
-            causal=True,
-        )
-        output = rearrange(
-            pad_input(
-                rearrange(output_unpad, "nnz h d -> nnz (h d)"),
-                indices,
-                bsz,
-                q_len,
-            ),
-            "b s (h d) -> b s h d",
-            h=nheads,
-        )
-    return (
-        self.o_proj(rearrange(output, "b s h d -> b s (h d)")),
-        None,
-        None,
-    )
 
 
 # Disable the transformation of the attention mask in LlamaModel as the flash attention
@@ -133,8 +58,541 @@ def _prepare_decoder_attention_mask(
     return attention_mask
 
 
-def replace_llama_attn_with_flash_attn():
-    transformers.models.llama.modeling_llama.LlamaModel._prepare_decoder_attention_mask = (  # pylint: disable=protected-access
-        _prepare_decoder_attention_mask
+def flashattn_forward(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.Tensor] = 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]]]:
+    """Input shape: Batch x Time x Channel
+
+    attention_mask: [bsz, q_len]
+    """
+    # pylint: disable=duplicate-code
+    bsz, q_len, _ = hidden_states.size()
+
+    if not hasattr(self, "pretraining_tp"):
+        self.pretraining_tp = 1
+
+    if self.pretraining_tp > 1:
+        key_value_slicing = (
+            self.num_key_value_heads * self.head_dim
+        ) // self.pretraining_tp
+        query_slices = self.q_proj.weight.split(
+            (self.num_heads * self.head_dim) // self.pretraining_tp, dim=0
+        )
+        key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
+        value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
+
+        query_states = [
+            F.linear(hidden_states, query_slices[i]) for i in range(self.pretraining_tp)
+        ]
+        query_states = torch.cat(query_states, dim=-1)
+
+        key_states = [
+            F.linear(hidden_states, key_slices[i]) for i in range(self.pretraining_tp)
+        ]
+        key_states = torch.cat(key_states, dim=-1)
+
+        value_states = [
+            F.linear(hidden_states, value_slices[i]) for i in range(self.pretraining_tp)
+        ]
+        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)
+
+    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)
+    # [bsz, q_len, nh, hd]
+    # [bsz, nh, q_len, hd]
+
+    kv_seq_len = key_states.shape[-2]
+    if past_key_value is not None:
+        kv_seq_len += past_key_value[0].shape[-2]
+
+    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+    query_states, key_states = apply_rotary_pos_emb(
+        query_states, key_states, cos, sin, position_ids
     )
-    transformers.models.llama.modeling_llama.LlamaAttention.forward = forward
+    # [bsz, nh, t, hd]
+
+    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 output_attentions:
+        warnings.warn(
+            "Output attentions is not supported for patched `LlamaAttention`, returning `None` instead."
+        )
+
+    #
+    # flash-attn v2 start
+    #
+
+    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
+
+    if cu_seqlens is not None and max_seqlen is not None:
+        # 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, 0.0, softmax_scale=None, causal=True
+        )
+        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,
+            0.0,
+            softmax_scale=None,
+            causal=is_causal,
+        )
+        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),
+                causal=is_causal,
+            )
+        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,
+            )
+            output_unpad = flash_attn_varlen_kvpacked_func(
+                q_unpad,
+                kv_unpad,
+                cu_seqlens_q,
+                cu_seqlens_k,
+                max_seqlen_q,
+                max_seqlen_k,
+                0.0,
+                softmax_scale=None,
+                causal=is_causal,
+            )
+            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)")
+
+    #
+    # flash-attn v2 end
+    #
+
+    if self.pretraining_tp > 1:
+        attn_output = attn_output.split(self.hidden_size // self.pretraining_tp, dim=2)
+        o_proj_slices = self.o_proj.weight.split(
+            self.hidden_size // self.pretraining_tp, dim=1
+        )
+        attn_output = sum(
+            F.linear(attn_output[i], o_proj_slices[i])
+            for i in range(self.pretraining_tp)
+        )
+    else:
+        attn_output = self.o_proj(attn_output)
+
+    return attn_output, None, 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 llama_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,
+        )
+    )
+
+    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,
+                None,
+                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 LlamaDecoderLayer(OriginalLlamaDecoderLayer):
+    """
+    patched version of LlamaDecoderLayer 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/llama_attn_hijack_sdp.py

@@ -0,0 +1,140 @@
+"""
+Patched LlamaAttention to use torch.nn.functional.scaled_dot_product_attention
+"""
+
+import warnings
+from typing import Optional, Tuple
+
+import torch
+import torch.nn.functional as F
+import transformers.models.llama.modeling_llama
+from transformers.models.llama.modeling_llama import apply_rotary_pos_emb, repeat_kv
+
+
+def hijack_llama_sdp_attention():
+    transformers.models.llama.modeling_llama.LlamaAttention.forward = (
+        sdp_attention_forward
+    )
+
+
+def sdp_attention_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: bool = False,
+    use_cache: bool = False,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    # pylint: disable=duplicate-code
+    bsz, q_len, _ = hidden_states.size()
+
+    if not hasattr(self, "pretraining_tp"):
+        self.pretraining_tp = 1
+
+    if self.pretraining_tp > 1:
+        key_value_slicing = (
+            self.num_key_value_heads * self.head_dim
+        ) // self.pretraining_tp
+        query_slices = self.q_proj.weight.split(
+            (self.num_heads * self.head_dim) // self.pretraining_tp, dim=0
+        )
+        key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
+        value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
+
+        query_states = [
+            F.linear(hidden_states, query_slices[i]) for i in range(self.pretraining_tp)
+        ]
+        query_states = torch.cat(query_states, dim=-1)
+
+        key_states = [
+            F.linear(hidden_states, key_slices[i]) for i in range(self.pretraining_tp)
+        ]
+        key_states = torch.cat(key_states, dim=-1)
+
+        value_states = [
+            F.linear(hidden_states, value_slices[i]) for i in range(self.pretraining_tp)
+        ]
+        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)
+
+    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)
+    # [bsz, q_len, nh, hd]
+    # [bsz, nh, q_len, hd]
+
+    kv_seq_len = key_states.shape[-2]
+    if past_key_value is not None:
+        kv_seq_len += past_key_value[0].shape[-2]
+
+    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+    query_states, key_states = apply_rotary_pos_emb(
+        query_states, key_states, cos, sin, position_ids
+    )
+    # [bsz, nh, t, hd]
+
+    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 output_attentions:
+        warnings.warn(
+            "Output attentions is not supported for patched `LlamaAttention`, returning `None` instead."
+        )
+
+    #
+    # sdp-attn start
+    #
+
+    with torch.backends.cuda.sdp_kernel():
+        attn_output = torch.nn.functional.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask,
+            is_causal=False,
+        )
+
+    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()}"
+        )
+    attn_output = attn_output.transpose(1, 2)
+    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+    #
+    # sdp-attn end
+    #
+
+    if self.pretraining_tp > 1:
+        attn_output = attn_output.split(self.hidden_size // self.pretraining_tp, dim=2)
+        o_proj_slices = self.o_proj.weight.split(
+            self.hidden_size // self.pretraining_tp, dim=1
+        )
+        attn_output = sum(
+            F.linear(attn_output[i], o_proj_slices[i])
+            for i in range(self.pretraining_tp)
+        )
+    else:
+        attn_output = self.o_proj(attn_output)
+
+    return attn_output, None, past_key_value

src/axolotl/monkeypatch/llama_attn_hijack_xformers.py

@@ -3,12 +3,13 @@ Directly copied the code from https://raw.githubusercontent.com/oobabooga/text-g
 """
 
 import logging
-import math
+import warnings
 from typing import Optional, Tuple
 
 import torch
+import torch.nn.functional as F
 import transformers.models.llama.modeling_llama
-from torch import nn
+from transformers.models.llama.modeling_llama import apply_rotary_pos_emb, repeat_kv
 
 try:
     import xformers.ops
@@ -20,12 +21,6 @@ def hijack_llama_attention():
     transformers.models.llama.modeling_llama.LlamaAttention.forward = xformers_forward
 
 
-def hijack_llama_sdp_attention():
-    transformers.models.llama.modeling_llama.LlamaAttention.forward = (
-        sdp_attention_forward
-    )
-
-
 def xformers_forward(
     self,
     hidden_states: torch.Tensor,
@@ -38,30 +33,57 @@ def xformers_forward(
     # pylint: disable=duplicate-code
     bsz, q_len, _ = hidden_states.size()
 
-    query_states = (
-        self.q_proj(hidden_states)
-        .view(bsz, q_len, self.num_heads, self.head_dim)
-        .transpose(1, 2)
-    )
-    key_states = (
-        self.k_proj(hidden_states)
-        .view(bsz, q_len, self.num_heads, self.head_dim)
-        .transpose(1, 2)
-    )
-    value_states = (
-        self.v_proj(hidden_states)
-        .view(bsz, q_len, self.num_heads, self.head_dim)
-        .transpose(1, 2)
-    )
+    if not hasattr(self, "pretraining_tp"):
+        self.pretraining_tp = 1
+
+    if self.pretraining_tp > 1:
+        key_value_slicing = (
+            self.num_key_value_heads * self.head_dim
+        ) // self.pretraining_tp
+        query_slices = self.q_proj.weight.split(
+            (self.num_heads * self.head_dim) // self.pretraining_tp, dim=0
+        )
+        key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
+        value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)
+
+        query_states = [
+            F.linear(hidden_states, query_slices[i]) for i in range(self.pretraining_tp)
+        ]
+        query_states = torch.cat(query_states, dim=-1)
+
+        key_states = [
+            F.linear(hidden_states, key_slices[i]) for i in range(self.pretraining_tp)
+        ]
+        key_states = torch.cat(key_states, dim=-1)
+
+        value_states = [
+            F.linear(hidden_states, value_slices[i]) for i in range(self.pretraining_tp)
+        ]
+        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)
+
+    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)
+    # [bsz, q_len, nh, hd]
+    # [bsz, nh, q_len, hd]
 
     kv_seq_len = key_states.shape[-2]
     if past_key_value is not None:
         kv_seq_len += past_key_value[0].shape[-2]
+
     cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
-    (
-        query_states,
-        key_states,
-    ) = transformers.models.llama.modeling_llama.apply_rotary_pos_emb(
+    query_states, key_states = apply_rotary_pos_emb(
         query_states, key_states, cos, sin, position_ids
     )
     # [bsz, nh, t, hd]
@@ -73,162 +95,61 @@ def xformers_forward(
 
     past_key_value = (key_states, value_states) if use_cache else None
 
-    # We only apply xformers optimizations if we don't need to output the whole attention matrix
-    if not output_attentions:
-        query_states = query_states.transpose(1, 2)
-        key_states = key_states.transpose(1, 2)
-        value_states = value_states.transpose(1, 2)
+    # 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)
 
-        # This is a nasty hack. We know attention_mask in transformers is either LowerTriangular or all Zeros.
-        # We therefore check if one element in the upper triangular portion is zero. If it is, then the mask is all zeros.
-        if attention_mask is None or attention_mask[0, 0, 0, 1] == 0:
-            # input and output should be of form (bsz, q_len, num_heads, head_dim)
-            attn_output = xformers.ops.memory_efficient_attention(
-                query_states, key_states, value_states, attn_bias=None
-            )
-        else:
-            # input and output should be of form (bsz, q_len, num_heads, head_dim)
-            attn_output = xformers.ops.memory_efficient_attention(
-                query_states,
-                key_states,
-                value_states,
-                attn_bias=xformers.ops.LowerTriangularMask(),
-            )
-        attn_weights = None
+    if output_attentions:
+        warnings.warn(
+            "Output attentions is not supported for patched `LlamaAttention`, returning `None` instead."
+        )
+
+    #
+    # xformers-attn start
+    #
+
+    query_states = query_states.transpose(1, 2)
+    key_states = key_states.transpose(1, 2)
+    value_states = value_states.transpose(1, 2)
+
+    # This is a nasty hack. We know attention_mask in transformers is either LowerTriangular or all Zeros.
+    # We therefore check if one element in the upper triangular portion is zero. If it is, then the mask is all zeros.
+    if attention_mask is None or attention_mask[0, 0, 0, 1] == 0:
+        # input and output should be of form (bsz, q_len, num_heads, head_dim)
+        attn_output = xformers.ops.memory_efficient_attention(
+            query_states, key_states, value_states, attn_bias=None
+        )
     else:
-        attn_weights = torch.matmul(
-            query_states, key_states.transpose(2, 3)
-        ) / math.sqrt(self.head_dim)
+        # input and output should be of form (bsz, q_len, num_heads, head_dim)
+        attn_output = xformers.ops.memory_efficient_attention(
+            query_states,
+            key_states,
+            value_states,
+            # attn_bias=attention_mask,
+            attn_bias=xformers.ops.LowerTriangularMask(),
+        )
 
-        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
-            attn_weights = torch.max(
-                attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min)
-            )
-
-        # 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()}"
-            )
-
-        attn_output = attn_output.transpose(1, 2)
-
-    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
-    attn_output = self.o_proj(attn_output)
-    return attn_output, attn_weights, past_key_value
-
-
-def sdp_attention_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: bool = False,
-    use_cache: bool = False,
-) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
-    # pylint: disable=duplicate-code
-    bsz, q_len, _ = hidden_states.size()
-
-    query_states = (
-        self.q_proj(hidden_states)
-        .view(bsz, q_len, self.num_heads, self.head_dim)
-        .transpose(1, 2)
-    )
-    key_states = (
-        self.k_proj(hidden_states)
-        .view(bsz, q_len, self.num_heads, self.head_dim)
-        .transpose(1, 2)
-    )
-    value_states = (
-        self.v_proj(hidden_states)
-        .view(bsz, q_len, self.num_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,
-    ) = transformers.models.llama.modeling_llama.apply_rotary_pos_emb(
-        query_states, key_states, cos, sin, position_ids
-    )
-    # [bsz, nh, t, hd]
-
-    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
-
-    # We only apply sdp attention if we don't need to output the whole attention matrix
-    if not output_attentions:
-        with torch.backends.cuda.sdp_kernel():
-            attn_output = torch.nn.functional.scaled_dot_product_attention(
-                query_states,
-                key_states,
-                value_states,
-                attn_mask=attention_mask,
-                is_causal=False,
-            )
-            attn_weights = None
-    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
-            attn_weights = torch.max(
-                attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min)
-            )
-
-        # 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()}"
-            )
-
-    attn_output = attn_output.transpose(1, 2)
+    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 = attn_output.reshape(bsz, q_len, self.hidden_size)
 
-    attn_output = self.o_proj(attn_output)
+    #
+    # xformers-attn end
+    #
 
-    return attn_output, attn_weights, past_key_value
+    if self.pretraining_tp > 1:
+        attn_output = attn_output.split(self.hidden_size // self.pretraining_tp, dim=2)
+        o_proj_slices = self.o_proj.weight.split(
+            self.hidden_size // self.pretraining_tp, dim=1
+        )
+        attn_output = sum(
+            F.linear(attn_output[i], o_proj_slices[i])
+            for i in range(self.pretraining_tp)
+        )
+    else:
+        attn_output = self.o_proj(attn_output)
+
+    return attn_output, None, past_key_value

src/axolotl/monkeypatch/llama_expand_mask.py

@@ -0,0 +1,52 @@
+"""
+expands the binary attention mask per 3.2.2 of https://arxiv.org/pdf/2107.02027.pdf
+"""
+from typing import Optional
+
+import torch
+
+
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    This expansion handles packed sequences so that sequences share the same attention mask integer value
+    when they attend to each other within that sequence.
+    This expansion transforms the mask to lower triangular form to prevent future peeking.
+    """
+    bsz, src_len = mask.size()
+    tgt_len = tgt_len if tgt_len is not None else src_len
+
+    mask = mask.unsqueeze(1).unsqueeze(2)
+    mask = mask.expand(bsz, 1, tgt_len, src_len)
+
+    # Create a binary mask from the original mask where zeros remain zeros and all other values are set to one
+    binary_mask = torch.where(
+        mask != 0,
+        torch.tensor(1).to(dtype),
+        torch.tensor(0).to(dtype),
+    )
+
+    # Create a block-diagonal mask.
+    # we multiply by the binary mask so that 0's in the original mask are correctly excluded
+    zero_one_mask = torch.eq(mask, mask.transpose(-1, -2)).int() * binary_mask
+
+    # Now let's create a lower triangular mask of ones that will zero out the upper triangular part
+    lower_triangular_ones = torch.tril(torch.ones((tgt_len, src_len), dtype=dtype)).to(
+        mask.device
+    )
+
+    # Use the lower triangular mask to zero out the upper triangular part of the zero_one_mask
+    masked_zero_one_mask = zero_one_mask * lower_triangular_ones
+    inverted_mask = 1.0 - masked_zero_one_mask
+
+    return inverted_mask.masked_fill(
+        inverted_mask.to(torch.bool), torch.finfo(dtype).min
+    )
+
+
+def hijack_expand_mask():
+    import transformers
+
+    transformers.models.llama.modeling_llama._expand_mask = (  # pylint: disable=protected-access
+        _expand_mask
+    )

src/axolotl/monkeypatch/relora.py

@@ -0,0 +1,393 @@
+"""Implements the ReLoRA training procedure from https://arxiv.org/abs/2307.05695, minus the initial full fine-tune."""
+import glob
+import json
+import logging
+import os.path
+import shutil
+from pathlib import Path
+from typing import Dict, List, Sequence
+
+import bitsandbytes as bnb
+import peft
+import safetensors.torch as st
+import torch
+from huggingface_hub import snapshot_download
+from torch.optim.lr_scheduler import LRScheduler
+from torch.optim.optimizer import Optimizer
+from transformers import (
+    TrainerCallback,
+    TrainerControl,
+    TrainerState,
+    TrainingArguments,
+)
+from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR
+
+from axolotl.utils.dict import DictDefault
+from axolotl.utils.distributed import is_main_process
+
+LOG = logging.getLogger("axolotl.relora")
+
+
+def reset_optimizer(optimizer: torch.optim.Optimizer):
+    for group in optimizer.param_groups:
+        for param in group["params"]:
+            param_state = optimizer.state[param]
+            for key in param_state:
+                if "qmap" in key:
+                    continue
+
+                if key == "step" and isinstance(param_state[key], int):
+                    param_state[key] = 0
+                else:
+                    param_state[key] = torch.zeros_like(param_state[key])
+
+
+class ReLoRACallback(TrainerCallback):
+    """Callback to merge LoRA weights into the base model and save full-weight checkpoints"""
+
+    def __init__(self, cfg: DictDefault):
+        self.relora_steps = cfg.relora_steps
+        self.cpu_offload = cfg.relora_cpu_offload
+        self.quantized = cfg.load_in_4bit or cfg.load_in_8bit
+        self.last_full_model = cfg.base_model
+        self.resume_from_checkpoint = cfg.resume_from_checkpoint
+
+        if not os.path.exists(self.last_full_model):
+            self.last_full_model = str(Path(snapshot_download(cfg.base_model)))
+
+        assert os.path.exists(
+            self.last_full_model
+        ), "for ReLORA base_model must be a local path"
+
+        self.num_lora_restarts = 0
+        self.need_full_save = False
+
+    def on_train_begin(
+        self,
+        _args: TrainingArguments,
+        _state: TrainerState,
+        control: TrainerControl,
+        model: peft.LoraModel,
+        **_kwargs,
+    ):
+        if self.resume_from_checkpoint:
+            weight_path = os.path.join(self.resume_from_checkpoint, "relora")
+            if not os.path.exists(weight_path):
+                LOG.warning(
+                    "Resuming ReLoRA from checkpoint, but no full-weight save found"
+                )
+            else:
+                LOG.info(f"Loading adjusted base weights from {weight_path}")
+                load_weight_checkpoint(model, weight_path)
+        return control
+
+    def on_step_begin(
+        self,
+        args: TrainingArguments,
+        state: TrainerState,
+        control: TrainerControl,
+        model: peft.LoraModel,
+        optimizer: torch.optim.Optimizer,
+        **_kwargs,
+    ):
+        if state.global_step > 0 and state.global_step % self.relora_steps == 0:
+            checkpoint_folder = os.path.join(
+                args.output_dir,
+                f"{PREFIX_CHECKPOINT_DIR}-{state.global_step}",
+                "relora",
+            )
+
+            with torch.no_grad():
+                merge_and_save(
+                    model,
+                    self.last_full_model,
+                    checkpoint_folder,
+                    reinit=True,
+                    quantized=self.quantized,
+                    actually_save=is_main_process(),
+                    cpu_offload=self.cpu_offload,
+                )
+                reset_optimizer(optimizer)
+
+            if self.quantized:
+                self.last_full_model = checkpoint_folder
+            self.num_lora_restarts += 1
+
+        return control
+
+    def on_save(
+        self,
+        args: TrainingArguments,
+        state: TrainerState,
+        control: TrainerControl,
+        model: peft.LoraModel,
+        **_kwargs,
+    ):
+        checkpoint_folder = os.path.join(
+            args.output_dir, f"{PREFIX_CHECKPOINT_DIR}-{state.global_step}", "relora"
+        )
+        if (
+            state.global_step >= self.relora_steps
+            and state.global_step % self.relora_steps != 0
+        ):
+            if self.quantized:
+                if is_main_process() and self.last_full_model != checkpoint_folder:
+                    # ensure the latest full parameter save is in the latest checkpoint
+                    # folder, so that automatic pruning of checkpoints does not remove it
+                    LOG.info(f"moving last full parameter save to {checkpoint_folder}")
+                    os.makedirs(checkpoint_folder, exist_ok=True)
+                    chunks = glob.glob(
+                        f"{self.last_full_model}/model*.safetensors"
+                    ) + glob.glob(f"{self.last_full_model}/model*.index.json")
+                    for path in chunks:
+                        new_path = os.path.abspath(shutil.move(path, checkpoint_folder))
+                        try:
+                            os.symlink(new_path, path)
+                        except OSError:
+                            # probably on windows without permission to symlink
+                            pass
+
+                    self.last_full_model = checkpoint_folder
+            else:
+                model.model.save_pretrained(checkpoint_folder, safe_serialization=True)
+
+        return control
+
+    def on_log(
+        self,
+        _args: TrainingArguments,
+        _state: TrainerState,
+        control: TrainerControl,
+        logs: Dict[str, float],
+        **_kwargs,
+    ):
+        logs["num_lora_restarts"] = self.num_lora_restarts
+        return control
+
+    def on_train_end(
+        self,
+        args: TrainingArguments,
+        _state: TrainerState,
+        control: TrainerControl,
+        model: peft.LoraModel,
+        **_kwargs,
+    ):
+        if self.quantized:
+            # perform final merge and save
+            with torch.no_grad():
+                merge_and_save(
+                    model,
+                    self.last_full_model,
+                    args.output_dir,
+                    reinit=False,
+                    quantized=self.quantized,
+                    actually_save=is_main_process(),
+                    cpu_offload=self.cpu_offload,
+                )
+        # no need to save if unquantized, as finetune.py will call merge_and_unload()
+        return control
+
+
+class ReLoRAScheduler(LRScheduler):
+    """Wraps another scheduler to apply per-lora-restart learning rate warmups."""
+
+    def __init__(
+        self,
+        optimizer: Optimizer,
+        inner_schedule: LRScheduler,
+        relora_steps: int,
+        warmup_steps: int,
+        min_lr_scale: float = 0.001,
+    ) -> None:
+        self.inner_schedule = inner_schedule
+        self.relora_steps = relora_steps
+        self.warmup_steps = warmup_steps
+        self.min_lr_scale = min_lr_scale
+        super().__init__(optimizer, inner_schedule.last_epoch, inner_schedule.verbose)
+
+    def get_lr(self) -> float:
+        self.inner_schedule.last_epoch = self.last_epoch
+
+        original = self.inner_schedule.get_lr()
+        step = self.last_epoch
+        if step < self.relora_steps:
+            scale = 1
+        else:
+            cycle_t = min(1.0, (step % self.relora_steps) / self.warmup_steps)
+            scale = cycle_t * (1 - self.min_lr_scale) + self.min_lr_scale
+
+        if isinstance(original, Sequence):
+            return [lr * scale for lr in original]
+        return original * scale
+
+
+def sharded_paths(path: str, module_names: List[str]) -> Dict[str, str]:
+    model_name = "model.safetensors"
+    if not os.path.exists(str(Path(path) / model_name)) and not os.path.exists(
+        str(Path(path) / f"{model_name}.index.json")
+    ):
+        model_name = "pytorch_model.bin"
+
+    index_path = str(Path(path) / f"{model_name}.index.json")
+    if os.path.exists(index_path):
+        with open(index_path, "r", encoding="utf-8") as file:
+            data = json.load(file)
+        return data["weight_map"]
+    return {(module_name + ".weight"): model_name for module_name in module_names}
+
+
+def lora_delta_weight(layer: peft.tuners.lora.LoraLayer, device) -> torch.Tensor:
+    if isinstance(layer, (peft.tuners.lora.Linear8bitLt, peft.tuners.lora.Linear4bit)):
+        adapter = layer.active_adapter
+        return (
+            peft.utils.transpose(
+                layer.lora_B[adapter].weight.detach().to(device)
+                @ layer.lora_A[adapter].weight.detach().to(device),
+                getattr(layer, "fan_in_fan_out", False),
+            )
+            * layer.scaling[adapter]
+        )
+
+    return layer.get_delta_weight().to(device)
+
+
+def find_lora_modules(model: peft.LoraModel) -> Dict[str, peft.tuners.lora.LoraLayer]:
+    modules: Dict[str, peft.tuners.lora.LoraLayer] = {}
+
+    key_list = [key for key, _ in model.model.named_modules() if "lora" not in key]
+    for key in key_list:
+        try:
+            # pylint: disable=protected-access
+            _parent, target, _target_name = peft.utils._get_submodules(model.model, key)
+        except AttributeError:
+            continue
+
+        if isinstance(target, peft.tuners.lora.LoraLayer):
+            modules[key] = target
+
+    return modules
+
+
+def update_weights(
+    target: peft.tuners.lora.LoraLayer, new_weight: torch.Tensor, reinit: bool, device
+):
+    if reinit:
+        for adapter_name in target.lora_A:
+            target.reset_lora_parameters(adapter_name)
+        for adapter_name in target.lora_embedding_A:
+            target.reset_lora_parameters(adapter_name)
+
+    if isinstance(target, peft.tuners.lora.Linear4bit):
+        # This could be faster, but the quantization of Linear4bit weights occurs
+        # when the module is moved from cpu to gpu. Without meddling *too* deeply in
+        # PEFT's innards or maintaining a duplicate of that codepath, this is good
+        # enough for now.
+        target.weight.quant_state = None
+        target.weight.data = new_weight.cpu()
+        target.to(device)
+    elif isinstance(target, peft.tuners.lora.Linear8bitLt):
+        target.weight = bnb.nn.Int8Params(new_weight, requires_grad=False).to(device)
+    else:
+        target.weight.data = new_weight.to(device)
+
+
+def merge_and_save(
+    model: peft.LoraModel,
+    model_src: str,
+    model_dst: str,
+    reinit: bool = False,
+    quantized: bool = False,
+    cpu_offload: bool = False,
+    actually_save: bool = True,
+):
+    modules = find_lora_modules(model)
+
+    if not quantized:
+        for module_name, target in modules.items():
+            update = target.get_delta_weight(target.active_adapter).detach()
+            target.weight.data += update
+
+            if reinit:
+                for adapter_name in target.lora_A:
+                    target.reset_lora_parameters(adapter_name)
+                for adapter_name in target.lora_embedding_A:
+                    target.reset_lora_parameters(adapter_name)
+        return
+
+    os.makedirs(model_dst, exist_ok=True)
+    shard_paths = sharded_paths(model_src, modules.keys())
+    out_shard_paths = {}
+
+    unique_shards = list(set(shard_paths.values()))
+    for shard_path in unique_shards:
+        out_tensors = {}
+        if shard_path.endswith(".safetensors"):
+            in_tensors = st.load_file(str(Path(model_src) / shard_path))
+        else:
+            in_tensors = torch.load(Path(model_src) / shard_path)
+            if "state_dict" in in_tensors:
+                in_tensors = in_tensors["state_dict"]
+
+        for module_name, target in modules.items():
+            key = module_name + ".weight"
+            if key not in shard_paths or shard_paths[key] != shard_path:
+                continue
+
+            orig_weight = in_tensors[key]
+            old_dev = target.weight.device
+            math_dev = "cpu" if cpu_offload else old_dev
+
+            delta_weight = lora_delta_weight(target, math_dev)
+            new_weight = orig_weight.to(math_dev) + delta_weight
+            del delta_weight
+
+            if actually_save:
+                out_tensors[key] = new_weight.half().cpu()
+
+            update_weights(target, new_weight, reinit=reinit, device=old_dev)
+
+        if actually_save:
+            out_shard_name = shard_path
+            if out_shard_name.startswith("pytorch_model"):
+                out_shard_name = (
+                    out_shard_name.replace("pytorch_model", "model").rstrip(".bin")
+                    + ".safetensors"
+                )
+
+            for module_name in in_tensors:
+                if module_name not in out_tensors:
+                    out_tensors[module_name] = in_tensors[module_name].half()
+                out_shard_paths[module_name] = out_shard_name
+
+            shard_fn = str(Path(model_dst) / out_shard_name)
+            LOG.info(f"saving tensors to {shard_fn}")
+            st.save_file(out_tensors, shard_fn, metadata={"format": "pt"})
+
+        del in_tensors
+        del out_tensors
+        torch.cuda.empty_cache()
+
+    if actually_save and len(unique_shards) > 1:
+        with open(
+            str(Path(model_dst, "model.safetensors.index.json")), "w", encoding="utf-8"
+        ) as file:
+            json.dump({"metadata": {}, "weight_map": out_shard_paths}, file)
+
+
+def load_weight_checkpoint(model: peft.LoraModel, checkpoint_path: str):
+    modules = find_lora_modules(model)
+    shard_paths = sharded_paths(checkpoint_path, modules.keys())
+    unique_shards = list(set(shard_paths.values()))
+
+    for shard_path in unique_shards:
+        tensors = st.load_file(os.path.join(checkpoint_path, shard_path))
+
+        for module_name, target in modules.items():
+            key = module_name + ".weight"
+            if key not in shard_paths or shard_paths[key] != shard_path:
+                continue
+
+            new_weight = tensors[key]
+            update_weights(
+                target, new_weight, reinit=False, device=target.weight.device
+            )

src/axolotl/monkeypatch/utils.py

@@ -0,0 +1,103 @@
+"""
+Shared utils for the monkeypatches
+"""
+import torch
+
+
+def get_cu_seqlens(attn_mask):
+    """generate a cumulative sequence length mask for flash attention using attn mask"""
+    if len(attn_mask.shape) == 1:
+        attn_mask = attn_mask.unsqueeze(0)
+
+    device = attn_mask.device
+    results = []
+    max_seq_lens = []
+
+    for row in attn_mask:
+        # Exclude zeros to avoid adding their positions to the mask
+        t_non_zeros = row[row != 0]
+        # Find where the sequence number changes (including the first position)
+        seq_change = torch.cat(
+            [
+                torch.tensor([1], dtype=torch.int32, device=device),
+                t_non_zeros[1:] != t_non_zeros[:-1],
+            ]
+        )
+        # Get the indices where the sequence changes
+        change_indices = torch.cat(
+            [
+                (seq_change == 1).nonzero(as_tuple=True)[0],
+                torch.tensor([len(t_non_zeros)], dtype=torch.int32, device=device),
+            ]
+        )
+        # Calculate the sequence lengths
+        seq_lengths = change_indices[1:] - change_indices[:-1]
+        # Calculate the length of the final sequence or padding
+        final_seq_length = len(row) - change_indices[-1]
+        # Append the length of the final sequence or padding to seq_lengths
+        if final_seq_length.item():
+            seq_lengths = torch.cat(
+                [
+                    seq_lengths,
+                    torch.tensor(
+                        [final_seq_length.item()], dtype=torch.int32, device=device
+                    ),
+                ]
+            )
+        # Calculate the cumulative sequence lengths
+        cu_seqlens = torch.cat(
+            [torch.tensor([0], dtype=torch.int32, device=device), seq_lengths.cumsum(0)]
+        )
+        max_seq_len = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
+        results.append(cu_seqlens)
+        max_seq_lens.append(max_seq_len)
+
+    return torch.stack(results).to(dtype=torch.int32), torch.stack(max_seq_lens)
+
+
+def get_cu_seqlens_from_pos_ids(position_ids):
+    """generate a cumulative sequence length mask for flash attention using pos ids"""
+    if len(position_ids.shape) == 1:
+        position_ids = position_ids.unsqueeze(0)
+
+    device = position_ids.device
+    results = []
+    max_seq_lens = []
+
+    for row in position_ids:
+        # Count the number of consecutive zeros from the right side
+        padding_length = (row == 0).int().flip(dims=[0]).cumprod(dim=0).sum().item()
+
+        # Adjust the row to exclude padding
+        adjusted_row = row[:-padding_length] if padding_length else row.clone()
+
+        # Find where the position resets to 0 (indicating a new sequence)
+        seq_starts = torch.cat(
+            [
+                torch.tensor([True], dtype=torch.bool, device=device),
+                adjusted_row[1:] == 0,
+            ]
+        )
+        # Get the indices where the sequence starts
+        start_indices = torch.cat(
+            [
+                (seq_starts).nonzero(as_tuple=True)[0],
+                torch.tensor([len(adjusted_row)], dtype=torch.int32, device=device),
+            ]
+        )
+        # Calculate the sequence lengths
+        seq_lengths = start_indices[1:] - start_indices[:-1]
+        # Calculate the cumulative sequence lengths
+        cu_seqlens = torch.cat(
+            [torch.tensor([0], dtype=torch.int32, device=device), seq_lengths.cumsum(0)]
+        )
+        # Append the padding length to the cumulative sequence lengths
+        if padding_length:
+            cu_seqlens = torch.cat(
+                [cu_seqlens, torch.tensor([len(row)], dtype=torch.int32, device=device)]
+            )
+        max_seq_len = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
+        results.append(cu_seqlens)
+        max_seq_lens.append(max_seq_len)
+
+    return torch.stack(results).to(dtype=torch.int32), torch.stack(max_seq_lens)

src/axolotl/prompt_strategies/__init__.py

@@ -2,8 +2,10 @@
 
 import importlib
 
+from axolotl.prompt_strategies.user_defined import UserDefinedDatasetConfig
 
-def load(strategy, tokenizer, cfg):
+
+def load(strategy, tokenizer, cfg, ds_cfg):
     try:
         load_fn = "load"
         if strategy.split(".")[-1].startswith("load_"):
@@ -11,6 +13,9 @@ def load(strategy, tokenizer, cfg):
             strategy = ".".join(strategy.split(".")[:-1])
         mod = importlib.import_module(f".{strategy}", "axolotl.prompt_strategies")
         func = getattr(mod, load_fn)
-        return func(tokenizer, cfg)
+        load_kwargs = {}
+        if strategy == "user_defined":
+            load_kwargs["ds_cfg"] = UserDefinedDatasetConfig(**ds_cfg)
+        return func(tokenizer, cfg, **load_kwargs)
     except Exception:  # pylint: disable=broad-exception-caught
         return None

src/axolotl/prompt_strategies/alpaca_w_system.py

@@ -57,6 +57,8 @@ class SystemDataPrompter(AlpacaPrompter):
     Alpaca Style Prompter that uses system prompts from the dataset
     """
 
+    system_format: str = "### System:\n{system}\n\n"
+
     def build_prompt_w_system(
         self,
         system: str,
@@ -66,7 +68,11 @@ class SystemDataPrompter(AlpacaPrompter):
     ) -> Generator[str, None, None]:
         # returns the full prompt from instruction and optional input
         # if a label (=response, =output) is provided, it's also appended.
-        formatted_sys_prompt = f"### System:\n{system}\n\n" if system else ""
+        formatted_sys_prompt = (
+            self.system_format.format(system=system)
+            if system and self.system_format
+            else ""
+        )
         if input:
             res = formatted_sys_prompt + self.turn_format.format(
                 instruction=instruction, input=input
@@ -86,12 +92,21 @@ class OpenOrcaSystemDataPrompter(SystemDataPrompter):
     """
 
     def match_prompt_style(self):
+        # pylint: disable=duplicate-code
         if self.prompt_style == PromptStyle.INSTRUCT.value:
-            self.turn_format = "### User:\n{instruction}\n\n### Additional Context:\n{input}\n\n### Assistant:\n"
-            self.turn_no_input_format = "### User:\n{instruction}\n\n### Assistant:\n"
+            self.turn_format = "### Human:\n{instruction}\n### Additional Context:\n{input}\n### Assistant:\n"
+            self.turn_no_input_format = "### Human:\n{instruction}\n### Assistant:\n"
+            self.system_format = "### System:\n{system}\n"
         if self.prompt_style == PromptStyle.CHAT.value:
             self.turn_format = "USER: {instruction}\n{input}\nASSISTANT:"
             self.turn_no_input_format = "USER: {instruction}\nASSISTANT:"
+            self.system_format = "SYSTEM: {system}\n"
+        if self.prompt_style == PromptStyle.CHATML.value:
+            self.turn_format = "<|im_start|>user\n{instruction}\n{input}<|im_end|>\n<|im_start|>assistant\n"
+            self.turn_no_input_format = (
+                "<|im_start|>user\n{instruction}<|im_end|>\n<|im_start|>assistant\n"
+            )
+            self.system_format = "<|im_start|>system\n{system}<|im_end|>\n"
 
 
 class OpenOrcaPromptTokenizingStrategy(InstructionWSystemPromptTokenizingStrategy):
@@ -137,3 +152,12 @@ def load_open_orca(tokenizer, cfg):
         cfg.train_on_inputs,
         cfg.sequence_len,
     )
+
+
+def load_open_orca_chatml(tokenizer, cfg):
+    return OpenOrcaPromptTokenizingStrategy(
+        OpenOrcaSystemDataPrompter(PromptStyle.CHATML.value),
+        tokenizer,
+        cfg.train_on_inputs,
+        cfg.sequence_len,
+    )

src/axolotl/prompt_strategies/llama2_chat.py

@@ -0,0 +1,205 @@
+"""
+Prompt Strategy for finetuning Llama2 chat models
+see also https://github.com/facebookresearch/llama/blob/6c7fe276574e78057f917549435a2554000a876d/llama/generation.py#L213 for ma reference implementation.
+
+This implementation is based on the Vicuna PR and the fastchat repo, see also:
+https://github.com/lm-sys/FastChat/blob/cdd7730686cb1bf9ae2b768ee171bdf7d1ff04f3/fastchat/conversation.py#L847
+
+Use dataset type: "llama2_chat" in conig.yml to use this prompt style.
+
+E.g. in the config.yml:
+```
+datasets:
+  - path: llama_finetune_train.jsonl
+    type: llama2_chat
+```
+
+The dataset itself should look like this:
+```
+{'conversations':[{"from": "human", "value": "Who are you?"}, {"from": "gpt", "value": "I am Vicuna"},...]}
+```
+in a jsonl file. The first message should be from the human, the second from gpt.
+For a custom system message, the first "from" can be "system" (followed by alternating "human" and "gpt" turns).
+
+Important: Don't use "special_tokens:" in your config.yml if you are not sure what you are doing!
+"""
+
+import logging
+from dataclasses import dataclass, field
+from typing import Generator, List, Sequence
+
+from axolotl.prompt_tokenizers import PromptTokenizingStrategy
+from axolotl.prompters import IGNORE_TOKEN_ID, SHAREGPT_ASSERTION_FAILED_ROLE
+
+
+@dataclass
+class Llama2ChatConversation:
+    """A class that manages prompt templates and keeps all conversation history.
+    copied from https://github.com/lm-sys/FastChat/blob/main/fastchat/conversation.py"""
+
+    name: str = "llama2"
+    # The system prompt
+    system: str = (
+        "[INST] <<SYS>>\nYou are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. "
+        "Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. "
+        "Please ensure that your responses are socially unbiased and positive in nature.\n\n"
+        "If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. "
+        "If you don't know the answer to a question, please don't share false information.\n<</SYS>>\n\n"
+    )
+    roles: Sequence[str] = ("[INST]", "[/INST]")
+    messages: List[List[str]] = field(default_factory=list)
+    offset: int = 0
+    sep = " "
+    sep2 = " </s><s>"
+    stop_token_ids = [2]
+
+    def get_prompt(self) -> str:
+        """Get the prompt for generation."""
+        seps = [self.sep, self.sep2]
+        ret = ""
+        for i, (role, message) in enumerate(self.messages):
+            if (i == len(self.messages) - 1) and (role == self.roles[0]):
+                # last message is from user (due to length),
+                #  return prompt without it for training
+                return ret
+            if i == 0:
+                ret += self.system + message.strip()
+            else:
+                ret += role + " " + message.strip() + seps[i % 2]
+        return ret
+
+    def append_message(self, role: str, message: str):
+        """Append a new message."""
+        self.messages.append([role, message])
+
+
+class LLama2ChatTokenizingStrategy(PromptTokenizingStrategy):
+    """
+    Tokenizing strategy for ShareGPT prompts.
+    adapted from https://github.com/lm-sys/FastChat/blob/main/fastchat/train/train.py
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.sequence_len = 4096
+        self.tokenizer.add_special_tokens({"pad_token": "<pad>"})
+        # https://huggingface.co/meta-llama/Llama-2-7b-chat-hf/blob/main/added_tokens.json
+
+    def tokenize_prompt(self, prompt):
+        conv = next(self.prompter.build_prompt(prompt))
+        conversation_str = conv.get_prompt()
+
+        # Tokenize conversations
+        input_ids = self.tokenizer(
+            conversation_str,
+            return_tensors="pt",
+            padding="max_length",
+            max_length=self.sequence_len,
+            truncation=True,
+        ).input_ids[0]
+        target = input_ids.clone()
+
+        # Mask targets. Only compute loss on the assistant outputs.
+        sep = conv.roles[1]
+
+        total_len = int(target.ne(self.tokenizer.pad_token_id).sum())
+
+        turns = conversation_str.split(conv.sep2)
+        cur_len = 1
+        target[:cur_len] = IGNORE_TOKEN_ID
+        for turn in turns:
+            if turn == "":
+                break
+            turn_len = len(self.tokenizer(turn).input_ids)
+
+            parts = turn.split(sep)
+            if len(parts) != 2:
+                break
+            parts[0] += sep
+            # "-1" is hardcoded for the LLaMA tokenizer to make the offset correct.
+            instruction_len = len(self.tokenizer(parts[0]).input_ids) - 1
+
+            # Ignore the user instructions
+            target[cur_len - 1 : cur_len + instruction_len] = IGNORE_TOKEN_ID
+            cur_len += turn_len + 2  # due to length of role token
+
+        target[cur_len:] = IGNORE_TOKEN_ID
+
+        if cur_len < self.sequence_len:
+            if cur_len != total_len:
+                target[:] = IGNORE_TOKEN_ID
+                logging.warning(
+                    f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}."
+                    f" (ignored)"
+                )
+
+        attention_mask = input_ids.ne(self.tokenizer.pad_token_id).tolist()
+        input_ids = input_ids.tolist()
+        target = target.tolist()
+        # this is a fix for the tokenizer which tokenizes [ differently with eos tokens and
+        # follows the original llama implementation
+        for i in range(2, total_len - 2):
+            if input_ids[i] == 29961:
+                input_ids[i] = 518
+            if target[i] == 29961:
+                target[i] = 518
+        return {
+            "input_ids": input_ids,
+            "labels": target,
+            "attention_mask": attention_mask,
+        }
+
+
+class Llama2ChatPrompter:  # pylint: disable=too-few-public-methods
+    """
+    A prompter that generates prompts for Llama2 models.
+    """
+
+    system_prompt = (
+        "[INST] <<SYS>>\nYou are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. "
+        "Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. "
+        "Please ensure that your responses are socially unbiased and positive in nature.\n\n"
+        "If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. "
+        "If you don't know the answer to a question, please don't share false information.\n<</SYS>>\n\n"
+    )
+
+    def build_prompt(self, source) -> Generator[Llama2ChatConversation, None, None]:
+        # see https://github.com/lm-sys/FastChat/blob/da0641e567cf93756b0978ab5a6b092e96f06240/fastchat/train/train.py#L78
+        source = source["conversations"]  # fix data structure for datasets
+
+        # if system prompt provided, use it
+        if source[0]["from"] == "system":
+            system = f"[INST] <<SYS>>\n{source[0]['value']}\n<</SYS>>\n\n"
+            source = source[1:]
+        else:
+            system = self.system_prompt
+
+        conv = Llama2ChatConversation(system=system)
+
+        if len(source) < 2:
+            # If there isn't a back and forth conversation, ignore it
+            # also happens on the data splitting leaving empty conversations
+            raise IndexError
+
+        roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
+
+        if roles[source[0]["from"]] != conv.roles[0]:
+            # Skip the first one if it is not from human
+            source = source[1:]
+
+        conv.messages = []  # pylint: disable=R0801
+        for j, sentence in enumerate(source):
+            role = roles[sentence["from"]]
+            assert role == conv.roles[j % 2], SHAREGPT_ASSERTION_FAILED_ROLE
+            if sentence["value"]:
+                conv.append_message(role, sentence["value"])
+        yield conv
+
+
+def load(tokenizer, cfg) -> LLama2ChatTokenizingStrategy:
+    return LLama2ChatTokenizingStrategy(
+        Llama2ChatPrompter(),
+        tokenizer,
+        cfg.train_on_inputs,
+        cfg.sequence_len,
+    )

src/axolotl/prompt_strategies/metharme.py

@@ -0,0 +1,76 @@
+"""Module containing the MetharmenPromptTokenizingStrategy and MetharmePrompter class"""
+
+import logging
+from typing import Tuple
+
+from axolotl.prompt_tokenizers import InstructionPromptTokenizingStrategy
+from axolotl.prompters import AlpacaPrompter
+
+LOG = logging.getLogger("axolotl")
+
+IGNORE_TOKEN_ID = -100
+
+# pylint: disable=duplicate-code
+
+
+class MetharmePromptTokenizingStrategy(InstructionPromptTokenizingStrategy):
+    """
+    Tokenizing strategy for the Metharme models
+    """
+
+    def parse_instruction_fields(self, prompt) -> Tuple[str, str, str]:
+        return (prompt["prompt"], "", prompt["generation"])
+
+    def _tokenize(
+        self,
+        prompt: str,
+        add_eos_token: bool = True,
+        strip_bos_token: bool = False,
+        num_eos_tokens: int = 3,
+    ):
+        result = self.tokenizer(
+            prompt,
+            truncation=True,
+            max_length=self.sequence_len,
+            padding=False,
+            return_tensors=None,
+        )
+        if len(result["input_ids"]) == 0:
+            LOG.warning("Tokenizer result is empty. You may want to audit your dataset")
+        # If there's already an EOS token there, subtract from the number added
+        if result["input_ids"][-1] == self.tokenizer.eos_token_id:
+            num_eos_tokens -= 1
+
+        if num_eos_tokens > 0 and add_eos_token and len(result["input_ids"]) > 0:
+            for _ in range(num_eos_tokens):
+                if len(result["input_ids"]) < self.sequence_len:
+                    result["input_ids"].append(self.tokenizer.eos_token_id)
+                    result["attention_mask"].append(1)
+
+        if result["input_ids"][0] == self.tokenizer.bos_token_id and strip_bos_token:
+            result["input_ids"] = result["input_ids"][1:]
+            result["attention_mask"] = result["attention_mask"][1:]
+
+        result["labels"] = result["input_ids"].copy()
+        return result
+
+
+class MetharmePrompter(AlpacaPrompter):
+    """
+    Prompter for the Metharme models.
+    """
+
+    system_prompt = ""
+    system_no_input_prompt = ""
+    system_format = ""
+    turn_format = "{instruction}"
+    turn_no_input_format = "{instruction}"
+
+    def __init__(self, *args, **kwargs):  # pylint: disable=super-init-not-called
+        pass
+
+
+def load(tokenizer, cfg):
+    return MetharmePromptTokenizingStrategy(
+        MetharmePrompter(), tokenizer, cfg.train_on_inputs, cfg.sequence_len
+    )

src/axolotl/prompt_strategies/orcamini.py

@@ -0,0 +1,46 @@
+"""
+Prompt Strategy for finetuning Orca Mini (v2) models
+see also https://huggingface.co/psmathur/orca_mini_v2_7b for more information
+
+Use dataset type: orcamini in conig.yml to use this prompt style.
+
+Compared to the alpaca_w_system.open_orca dataset type,
+this one specifies the system prompt with "### System:".
+
+Not suited/tested for multiple-turn conversations without further adjustments.
+"""
+from typing import Generator, Union
+
+from axolotl.prompt_strategies.alpaca_w_system import OpenOrcaPromptTokenizingStrategy
+from axolotl.prompters import AlpacaPrompter
+
+
+class OrcaMiniPrompter(AlpacaPrompter):
+    """Adjusted Prompter for Orca Mini (v2) datasets"""
+
+    def match_prompt_style(self):
+        self.turn_no_input_format = (
+            "### System:\n{system}\n\n### User:\n{instruction}\n\n### Response:\n"
+        )
+
+    def build_prompt_w_system(
+        self,
+        system: str,
+        instruction: str,
+        output: Union[None, str] = None,
+    ) -> Generator[str, None, None]:
+        # returns the full prompt from instruction and optional input
+        # if a label (=response, =output) is provided, it's also appended.
+        res = self.turn_no_input_format.format(system=system, instruction=instruction)
+        if output:
+            res = f"{res}{output}"
+        yield res
+
+
+def load(tokenizer, cfg):
+    return OpenOrcaPromptTokenizingStrategy(
+        OrcaMiniPrompter(),
+        tokenizer,
+        cfg.train_on_inputs,
+        cfg.sequence_len,
+    )

src/axolotl/prompt_strategies/user_defined.py

@@ -0,0 +1,98 @@
+"""
+User Defined prompts with configuration from the YML config
+"""
+
+from dataclasses import dataclass
+from functools import partial
+from typing import Optional, Tuple
+
+from axolotl.prompt_strategies.alpaca_w_system import (
+    InstructionWSystemPromptTokenizingStrategy,
+    SystemDataPrompter,
+)
+
+
+@dataclass
+class UserDefinedDatasetConfig:
+    """
+    dataclass configuration representing a userdefined dataset type
+    """
+
+    system_prompt: str = ""
+    field_system: str = "system"
+    field_instruction: str = "instruction"
+    field_input: str = "input"
+    field_output: str = "output"
+    format: str = "{instruction} {input} "
+    no_input_format: str = "{instruction} "
+    system_format: str = "{system}"
+
+    def __getitem__(self, item):
+        return getattr(self, item)
+
+
+class UserDefinedPromptTokenizationStrategy(InstructionWSystemPromptTokenizingStrategy):
+    """
+    Prompt Tokenization Strategy for user defined prompts
+    """
+
+
+def load(tokenizer, cfg, ds_cfg: Optional[UserDefinedDatasetConfig] = None):
+    if not ds_cfg:
+        raise ValueError("Missing dataset prompt configuration")
+
+    system_prompt = ""
+    if ds_cfg.system_prompt:
+        system_prompt = ds_cfg.system_prompt
+
+    def parse_instruction_fields(
+        field_instruction,
+        field_input,
+        field_output,
+        field_system,
+        system_prompt,
+        prompt,
+    ) -> Tuple[str, str, str, str]:
+        return (
+            prompt[field_instruction],
+            prompt[field_input] if field_input in prompt else "",
+            prompt[field_output] if field_output in prompt else "",
+            prompt[field_system] if field_system in prompt else system_prompt,
+        )
+
+    turn_format = ds_cfg.format
+    turn_no_input_format = ds_cfg.no_input_format
+    system_format = ds_cfg.system_format
+
+    class UserDefinedPrompter(SystemDataPrompter):
+        """
+        Prompter for user defined prompts
+        """
+
+        def match_prompt_style(self):
+            self.turn_format = turn_format
+            self.turn_no_input_format = turn_no_input_format
+            self.system_format = system_format
+
+    prompter = UserDefinedPrompter()
+
+    strat = UserDefinedPromptTokenizationStrategy(
+        prompter,
+        tokenizer,
+        cfg.train_on_inputs,
+        cfg.sequence_len,
+    )
+
+    setattr(
+        strat,
+        "parse_instruction_fields",
+        partial(
+            parse_instruction_fields,
+            ds_cfg.field_instruction,
+            ds_cfg.field_input,
+            ds_cfg.field_output,
+            ds_cfg.field_system,
+            system_prompt,
+        ),
+    )
+    return strat

src/axolotl/prompt_tokenizers.py

@@ -13,7 +13,7 @@ from axolotl.prompters import IGNORE_TOKEN_ID
 LOG = logging.getLogger("axolotl")
 
 IGNORE_INDEX = -100
-LLAMA_DEFAULT_PAD_TOKEN = "[PAD]"  # nosec
+LLAMA_DEFAULT_PAD_TOKEN = "<pad>"  # nosec
 LLAMA_DEFAULT_EOS_TOKEN = "</s>"  # nosec
 LLAMA_DEFAULT_BOS_TOKEN = "<s>"  # nosec
 LLAMA_DEFAULT_UNK_TOKEN = "<unk>"  # nosec
@@ -74,15 +74,22 @@ class PromptTokenizingStrategy(abc.ABC):
             padding=False,
             return_tensors=None,
         )
+        if len(result["input_ids"]) == 0:
+            LOG.warning("Tokenizer result is empty. You may want to audit your dataset")
         if (
-            result["input_ids"][-1] != self.tokenizer.eos_token_id
+            len(result["input_ids"]) > 0
+            and result["input_ids"][-1] != self.tokenizer.eos_token_id
             and len(result["input_ids"]) < self.sequence_len
             and add_eos_token
         ):
             result["input_ids"].append(self.tokenizer.eos_token_id)
             result["attention_mask"].append(1)
 
-        if result["input_ids"][0] == self.tokenizer.bos_token_id and strip_bos_token:
+        if (
+            len(result["input_ids"]) > 0
+            and result["input_ids"][0] == self.tokenizer.bos_token_id
+            and strip_bos_token
+        ):
             result["input_ids"] = result["input_ids"][1:]
             result["attention_mask"] = result["attention_mask"][1:]
 

src/axolotl/prompters.py

@@ -16,6 +16,7 @@ class PromptStyle(Enum):
 
     INSTRUCT = "instruct"
     CHAT = "chat"
+    CHATML = "chatml"
 
 
 class AlpacaPrompter:
@@ -25,6 +26,7 @@ class AlpacaPrompter:
 
     system_prompt = "Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request.\n\n"
     system_no_input_prompt = "Below is an instruction that describes a task. Write a response that appropriately completes the request.\n\n"
+    system_format: str = "{system}"
     turn_format: str
     turn_no_input_format: str
     prompt_style: Optional[PromptStyle] = None
@@ -34,14 +36,23 @@ class AlpacaPrompter:
         self.match_prompt_style()
 
     def match_prompt_style(self):
+        # pylint: disable=duplicate-code
         if self.prompt_style == PromptStyle.INSTRUCT.value:
             self.turn_format = "### Instruction:\n{instruction}\n\n### Input:\n{input}\n\n### Response:\n"
             self.turn_no_input_format = (
                 "### Instruction:\n{instruction}\n\n### Response:\n"
             )
+            self.system_format = "### System:\n{system}\n\n"
         if self.prompt_style == PromptStyle.CHAT.value:
             self.turn_format = "USER: {instruction}\n{input}\nASSISTANT:"
             self.turn_no_input_format = "USER: {instruction}\nASSISTANT:"
+            self.system_format = "SYSTEM: {system}\n"
+        if self.prompt_style == PromptStyle.CHATML.value:
+            self.turn_format = "<|im_start|>user\n{instruction}\n{input}<|im_end|>\n<|im_start|>assistant\n"
+            self.turn_no_input_format = (
+                "<|im_start|>user\n{instruction}<|im_end|>\n<|im_start|>assistant\n"
+            )
+            self.system_format = "<|im_start|>system\n{system}<|im_end|>\n"
 
     def build_prompt(
         self,
@@ -52,13 +63,17 @@ class AlpacaPrompter:
         # returns the full prompt from instruction and optional input
         # if a label (=response, =output) is provided, it's also appended.
         if input:
-            res = self.system_prompt + self.turn_format.format(
-                instruction=instruction, input=input
-            )
+            res = (
+                self.system_format.format(system=self.system_prompt)
+                if self.system_prompt
+                else ""
+            ) + self.turn_format.format(instruction=instruction, input=input)
         else:
-            res = self.system_no_input_prompt + self.turn_no_input_format.format(
-                instruction=instruction
-            )
+            res = (
+                self.system_format.format(system=self.system_no_input_prompt)
+                if self.system_prompt
+                else ""
+            ) + self.turn_no_input_format.format(instruction=instruction)
         if output:
             res = f"{res}{output}"
         yield res
@@ -260,6 +275,11 @@ class Conversation:
         self.messages.append([role, message])
 
 
+SHAREGPT_ASSERTION_FAILED_ROLE = (
+    "Role did not alternate between turns (gpt and human). Please check your data."
+)
+
+
 class ShareGPTPrompter:  # pylint: disable=too-few-public-methods
     """
     A prompter that generates prompts for the ShareGPT
@@ -296,7 +316,9 @@ class ShareGPTPrompter:  # pylint: disable=too-few-public-methods
         if len(source) < 2:
             # If there isn't a back and forth conversation, ignore it
             # also happens on the data splitting leaving empty conversations
-            raise IndexError
+            raise IndexError(
+                f"A conversation entry has less than 2 messages :\n{source}"
+            )
 
         conv = self._conversation.copy()
         roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
@@ -316,7 +338,7 @@ class ShareGPTPrompter:  # pylint: disable=too-few-public-methods
         conv.messages = []
         for j, sentence in enumerate(source):
             role = roles[sentence["from"]]
-            assert role == conv.roles[j % 2]
+            assert role == conv.roles[j % 2], SHAREGPT_ASSERTION_FAILED_ROLE
             conv.append_message(role, sentence["value"])
 
         for part in conv.get_prompt():

src/axolotl/utils/bench.py

@@ -0,0 +1,43 @@
+"""Benchmarking and measurement utilities"""
+
+import pynvml
+import torch
+
+
+def gpu_memory_usage(device=0):
+    return torch.cuda.memory_allocated(device) / 1024.0**3
+
+
+def gpu_memory_usage_all(device=0):
+    usage = torch.cuda.memory_allocated(device) / 1024.0**3
+    reserved = torch.cuda.memory_reserved(device) / 1024.0**3
+    smi = gpu_memory_usage_smi(device)
+    return usage, reserved - usage, max(0, smi - reserved)
+
+
+def gpu_memory_usage_smi(device=0):
+    if isinstance(device, torch.device):
+        device = device.index
+    if isinstance(device, str) and device.startswith("cuda:"):
+        device = int(device[5:])
+
+    pynvml.nvmlInit()
+    handle = pynvml.nvmlDeviceGetHandleByIndex(device)
+    info = pynvml.nvmlDeviceGetMemoryInfo(handle)
+    return info.used / 1024.0**3
+
+
+def log_gpu_memory_usage(log, msg, device):
+    if not torch.cuda.is_available():
+        return (0, 0, 0)
+
+    usage, cache, misc = gpu_memory_usage_all(device)
+    extras = []
+    if cache > 0:
+        extras.append(f"+{cache:.03f}GB cache")
+    if misc > 0:
+        extras.append(f"+{misc:.03f}GB misc")
+    log.info(
+        f"GPU memory usage {msg}: {usage:.03f}GB ({', '.join(extras)})", stacklevel=2
+    )
+    return usage, cache, misc

src/axolotl/utils/callbacks.py

@@ -1,5 +1,6 @@
 """Callbacks for Trainer class"""
 
+import logging
 import os
 
 from optimum.bettertransformer import BetterTransformer
@@ -11,6 +12,10 @@ from transformers import (
 )
 from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR, IntervalStrategy
 
+from axolotl.utils.bench import log_gpu_memory_usage
+
+LOG = logging.getLogger("axolotl.callbacks")
+
 
 class SavePeftModelCallback(TrainerCallback):  # pylint: disable=too-few-public-methods
     """Callback to save the PEFT adapter"""
@@ -28,7 +33,9 @@ class SavePeftModelCallback(TrainerCallback):  # pylint: disable=too-few-public-
         )
 
         peft_model_path = os.path.join(checkpoint_folder, "adapter_model")
-        kwargs["model"].save_pretrained(peft_model_path)
+        kwargs["model"].save_pretrained(
+            peft_model_path, save_safetensors=args.save_safetensors
+        )
 
         return control
 
@@ -67,3 +74,25 @@ class SaveBetterTransformerModelCallback(
             # the trainer will raise an exception since it can't save a BetterTransformer wrapped model
             control.should_save = False
         return control
+
+
+class GPUStatsCallback(
+    TrainerCallback
+):  # pylint: disable=too-few-public-methods disable=unused-argument
+    """Callback to track GPU utilization"""
+
+    def __init__(self, cfg):
+        self.cfg = cfg
+        self.logged = False
+
+    def on_step_end(
+        self,
+        args: TrainingArguments,
+        state: TrainerState,
+        control: TrainerControl,
+        **kwargs,
+    ):
+        if not self.logged and state.global_step > 1:
+            log_gpu_memory_usage(LOG, "while training", self.cfg.device)
+            self.logged = True
+        return control

src/axolotl/utils/collators.py

@@ -0,0 +1,121 @@
+"""
+DataCollator for axolotl to pad labels and position_ids for packed sequences
+"""
+from dataclasses import dataclass
+from typing import Any, Optional, Union
+
+import numpy as np
+from transformers import PreTrainedTokenizerBase
+from transformers.utils import PaddingStrategy
+
+
+@dataclass
+class DataCollatorForSeq2Seq:
+    """
+    Data collator that will dynamically pad the inputs received, as well as the labels and position_ids
+
+    Args:
+        tokenizer ([`PreTrainedTokenizer`] or [`PreTrainedTokenizerFast`]):
+            The tokenizer used for encoding the data.
+        model ([`PreTrainedModel`]):
+            The model that is being trained. If set and has the *prepare_decoder_input_ids_from_labels*, use it to
+            prepare the *decoder_input_ids*
+
+            This is useful when using *label_smoothing* to avoid calculating loss twice.
+        padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `True`):
+            Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
+            among:
+
+            - `True` or `'longest'` (default): Pad to the longest sequence in the batch (or no padding if only a single
+              sequence is provided).
+            - `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum
+              acceptable input length for the model if that argument is not provided.
+            - `False` or `'do_not_pad'`: No padding (i.e., can output a batch with sequences of different lengths).
+        max_length (`int`, *optional*):
+            Maximum length of the returned list and optionally padding length (see above).
+        pad_to_multiple_of (`int`, *optional*):
+            If set will pad the sequence to a multiple of the provided value.
+
+            This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
+            7.5 (Volta).
+        label_pad_token_id (`int`, *optional*, defaults to -100):
+            The id to use when padding the labels (-100 will be automatically ignored by PyTorch loss functions).
+        return_tensors (`str`):
+            The type of Tensor to return. Allowable values are "np", "pt" and "tf".
+    """
+
+    tokenizer: PreTrainedTokenizerBase
+    model: Optional[Any] = None
+    padding: Union[bool, str, PaddingStrategy] = True
+    max_length: Optional[int] = None
+    pad_to_multiple_of: Optional[int] = None
+    label_pad_token_id: int = -100
+    position_pad_token_id: int = 0
+    return_tensors: str = "pt"
+
+    def __call__(self, features, return_tensors=None):
+        labels = None
+        if return_tensors is None:
+            return_tensors = self.return_tensors
+
+        for feature_name, pad_token_id in [
+            ("labels", self.label_pad_token_id),
+            ("position_ids", self.position_pad_token_id),
+        ]:
+            feat = (
+                [feature[feature_name] for feature in features]
+                if feature_name in features[0].keys()
+                else None
+            )
+            labels = feat if feat and feature_name == "labels" else labels
+            # We have to pad the labels before calling `tokenizer.pad` as this method won't pad them and needs them of the
+            # same length to return tensors.
+            if feat is not None:
+                max_feature_length = max(len(l) for l in feat)  # noqa: E741
+                if self.pad_to_multiple_of is not None:
+                    max_feature_length = (
+                        (max_feature_length + self.pad_to_multiple_of - 1)
+                        // self.pad_to_multiple_of
+                        * self.pad_to_multiple_of
+                    )
+
+                padding_side = self.tokenizer.padding_side
+                for feature in features:
+                    remainder = [pad_token_id] * (
+                        max_feature_length - len(feature[feature_name])
+                    )
+                    if isinstance(feature[feature_name], list):
+                        feature[feature_name] = (
+                            feature[feature_name] + remainder
+                            if padding_side == "right"
+                            else remainder + feature[feature_name]
+                        )
+                    elif padding_side == "right":
+                        feature[feature_name] = np.concatenate(
+                            [feature[feature_name], remainder]
+                        ).astype(np.int64)
+                    else:
+                        feature[feature_name] = np.concatenate(
+                            [remainder, feature[feature_name]]
+                        ).astype(np.int64)
+
+        features = self.tokenizer.pad(
+            features,
+            padding=self.padding,
+            max_length=self.max_length,
+            pad_to_multiple_of=self.pad_to_multiple_of,
+            return_tensors=return_tensors,
+        )
+
+        # prepare decoder_input_ids
+        if (
+            labels is not None
+            and self.model is not None
+            and hasattr(self.model, "prepare_decoder_input_ids_from_labels")
+        ):
+            decoder_input_ids = self.model.prepare_decoder_input_ids_from_labels(
+                labels=features["labels"]
+            )
+            features["decoder_input_ids"] = decoder_input_ids
+
+        return features

src/axolotl/utils/config.py

@@ -1,13 +1,91 @@
-"""Module for validating config files"""
+"""Module for working with config dicts"""
 
 import logging
+import os
 
 import torch
 
+from axolotl.utils.bench import log_gpu_memory_usage
+
 LOG = logging.getLogger("axolotl")
 
 
+def choose_device(cfg):
+    def get_device():
+        try:
+            if torch.cuda.is_available():
+                return f"cuda:{cfg.local_rank}"
+
+            if torch.backends.mps.is_available():
+                return "mps"
+
+            raise SystemError("No CUDA/mps device found")
+        except Exception:  # pylint: disable=broad-exception-caught
+            return "cpu"
+
+    cfg.device = get_device()
+    if cfg.device_map != "auto":
+        if cfg.device.startswith("cuda"):
+            cfg.device_map = {"": cfg.local_rank}
+        else:
+            cfg.device_map = {"": cfg.device}
+
+    # in `accelerate launch`, we need to not pass through any device map and let
+    # accelerate figure out which parts of the model to put on which gpu
+    accelerate_vars = [var for var in os.environ if var.startswith("ACCELERATE_USE_")]
+    if accelerate_vars:
+        cfg.device_map = None
+
+
+def normalize_config(cfg):
+    # setup some derived config / hyperparams
+    cfg.gradient_accumulation_steps = cfg.gradient_accumulation_steps or (
+        cfg.batch_size // cfg.micro_batch_size
+    )
+    cfg.batch_size = (
+        cfg.batch_size or cfg.micro_batch_size * cfg.gradient_accumulation_steps
+    )
+    cfg.world_size = int(os.environ.get("WORLD_SIZE", 1))
+    cfg.local_rank = int(os.environ.get("LOCAL_RANK", 0))
+    choose_device(cfg)
+    cfg.ddp = cfg.ddp if cfg.ddp is not None else cfg.world_size != 1
+    if cfg.ddp:
+        cfg.device_map = {"": int(os.environ.get("LOCAL_RANK", 0))}
+        cfg.batch_size = cfg.batch_size * cfg.world_size
+
+    if cfg.device == "mps":
+        cfg.load_in_8bit = False
+        cfg.tf32 = False
+        if cfg.bf16:
+            cfg.fp16 = True
+        cfg.bf16 = False
+    else:
+        torch.backends.cuda.matmul.allow_tf32 = cfg.tf32 or False
+
+    if cfg.bf16 or cfg.bfloat16:
+        cfg.torch_dtype = torch.bfloat16
+    elif cfg.load_in_8bit or cfg.fp16 or cfg.float16:
+        cfg.torch_dtype = torch.float16
+    else:
+        cfg.torch_dtype = torch.float32
+
+    log_gpu_memory_usage(LOG, "baseline", cfg.device)
+
+
 def validate_config(cfg):
+    if cfg.max_packed_sequence_len and cfg.sample_packing:
+        raise ValueError(
+            "please set only one of max_packed_sequence_len (deprecated soon) or sample_packing"
+        )
+    if cfg.max_packed_sequence_len:
+        LOG.warning(
+            str(
+                PendingDeprecationWarning(
+                    "max_packed_sequence_len will be deprecated in favor of sample_packing"
+                )
+            )
+        )
+
     if cfg.gradient_accumulation_steps and cfg.batch_size:
         raise ValueError(
             "please set only one of gradient_accumulation_steps or batch_size"
@@ -19,9 +97,7 @@ def validate_config(cfg):
             "To calculate the equivalent gradient_accumulation_steps, divide batch_size / micro_batch_size / number of gpus.",
         )
     if cfg.load_4bit:
-        raise ValueError(
-            "cfg.load_4bit parameter has been deprecated and replaced by cfg.gptq"
-        )
+        raise ValueError("cfg.load_4bit parameter has been deprecated")
 
     if cfg.adapter == "qlora":
         if cfg.merge_lora:
@@ -48,6 +124,19 @@ def validate_config(cfg):
     if not cfg.load_in_8bit and cfg.adapter == "lora":
         LOG.warning("We recommend setting `load_in_8bit: true` for LORA finetuning")
 
+    if cfg.relora_steps:
+        if cfg.adapter not in ("lora", "qlora"):
+            raise ValueError("cfg.adapter must be lora or qlora to use ReLoRA")
+
+        if cfg.fsdp:
+            raise ValueError("fsdp not supported with ReLoRA")
+
+        if cfg.deepspeed:
+            raise ValueError("deepspeed not supported with ReLoRA")
+
+        if cfg.lr_scheduler == "one_cycle":
+            raise ValueError("ReLoRA is not compatible with the one_cycle scheduler")
+
     if cfg.trust_remote_code:
         LOG.warning(
             "`trust_remote_code` is set to true. Please make sure that you reviewed the remote code/model."
@@ -76,7 +165,7 @@ def validate_config(cfg):
                 "You should probably set bfloat16 or float16 to true to "
                 "load the model in float16 for BetterTransformers"
             )
-        if int(torch.__version__.split(".")[0]) < 2:
+        if int(torch.__version__.split(".", maxsplit=1)[0]) < 2:
             LOG.warning("torch>=2.0.0 required")
             raise ValueError(
                 f"flash_optimum for BetterTransformers may not be used with {torch.__version__}"
@@ -97,6 +186,24 @@ def validate_config(cfg):
             "push_to_hub_model_id is deprecated. Please use hub_model_id instead."
         )
 
+    if cfg.gptq and cfg.model_revision:
+        raise ValueError(
+            "model_revision is not supported for GPTQ models. "
+            + "Please download the model from HuggingFace Hub manually for correct branch, "
+            + "point to its path, and remove model_revision from the config."
+        )
+
+    if cfg.sample_packing and cfg.sdp_attention:
+        # incompatible due to bug w/ accelerate causing 0.0 loss when using llama2
+        raise ValueError(
+            "sample_packing not compatible with sdp_attention. Use flash_attention"
+        )
+
+    if cfg.sample_packing and cfg.xformers_attention:
+        raise ValueError(
+            "sample_packing not compatible with xformers_attention. Use flash_attention"
+        )
+
     # TODO
     # MPT 7b
     # https://github.com/facebookresearch/bitsandbytes/issues/25

src/axolotl/utils/data.py

@@ -1,13 +1,19 @@
 """Module containing data utilities"""
 import functools
-import itertools
+import hashlib
 import logging
 from hashlib import md5
 from pathlib import Path
-from typing import List, Tuple, Union
+from typing import Tuple, Union
 
 import torch
-from datasets import Dataset, DatasetDict, load_dataset, load_from_disk
+from datasets import (
+    Dataset,
+    DatasetDict,
+    concatenate_datasets,
+    load_dataset,
+    load_from_disk,
+)
 from huggingface_hub import hf_hub_download
 from transformers import PreTrainedTokenizerBase
 
@@ -35,8 +41,46 @@ from axolotl.prompters import (
     ShareGPTPrompter,
     SummarizeTLDRPrompter,
 )
+from axolotl.utils.dict import DictDefault
+from axolotl.utils.distributed import is_main_process, zero_first
+from axolotl.utils.trainer import (
+    calculate_total_num_steps,
+    process_datasets_for_packing,
+)
 
 LOG = logging.getLogger("axolotl")
+DEFAULT_DATASET_PREPARED_PATH = "last_run_prepared"
+
+
+def prepare_dataset(cfg, tokenizer):
+    if not cfg.pretraining_dataset:
+        with zero_first(is_main_process()):
+            train_dataset, eval_dataset = load_prepare_datasets(
+                tokenizer, cfg, DEFAULT_DATASET_PREPARED_PATH
+            )
+    else:
+        train_dataset = load_pretraining_dataset(
+            cfg.pretraining_dataset,
+            tokenizer,
+            max_tokens=cfg.sequence_len,
+            seed=cfg.seed or 42,
+        )
+        # https://discuss.huggingface.co/t/how-to-use-huggingface-trainer-streaming-datasets-without-wrapping-it-with-torchdatas-iterablewrapper/25230
+        train_dataset = train_dataset.with_format("torch")
+        eval_dataset = None
+
+    with zero_first(is_main_process()):
+        train_dataset, eval_dataset = process_datasets_for_packing(
+            cfg, train_dataset, eval_dataset
+        )
+    if cfg.max_steps:
+        total_num_steps = min(
+            calculate_total_num_steps(cfg, train_dataset, tokenizer), cfg.max_steps
+        )
+        LOG.info(f"Maximum number of steps set at {total_num_steps}")
+    else:
+        total_num_steps = calculate_total_num_steps(cfg, train_dataset, tokenizer)
+    return train_dataset, eval_dataset, total_num_steps
 
 
 def load_tokenized_prepared_datasets(
@@ -90,8 +134,17 @@ def load_tokenized_prepared_datasets(
             seed = 42
 
         datasets = []
+
+        def for_d_in_datasets(dataset_configs):
+            for dataset in dataset_configs:
+                if dataset.name and isinstance(dataset.name, list):
+                    for name in dataset.name:
+                        yield DictDefault({**dataset, "name": name})
+                else:
+                    yield dataset
+
         # pylint: disable=invalid-name
-        for d in cfg.datasets:
+        for d in for_d_in_datasets(cfg.datasets):
             ds: Union[Dataset, DatasetDict] = None
             ds_from_hub = False
             try:
@@ -109,6 +162,7 @@ def load_tokenized_prepared_datasets(
             local_path = Path(d.path)
             if local_path.exists():
                 if local_path.is_dir():
+                    # TODO dirs with arrow or parquet files could be loaded with `load_from_disk`
                     ds = load_dataset(
                         d.path,
                         name=d.name,
@@ -117,8 +171,15 @@ def load_tokenized_prepared_datasets(
                         split=None,
                     )
                 elif local_path.is_file():
+                    ds_type = "json"
+                    if d.ds_type:
+                        ds_type = d.ds_type
+                    elif ".parquet" in d.path:
+                        ds_type = "parquet"
+                    elif ".arrow" in d.path:
+                        ds_type = "arrow"
                     ds = load_dataset(
-                        "json",
+                        ds_type,
                         name=d.name,
                         data_files=d.path,
                         streaming=False,
@@ -155,13 +216,27 @@ def load_tokenized_prepared_datasets(
                     )
                 else:
                     ds = ds.shuffle(seed=seed).shard(num_shards=d.shards, index=0)
+
+            d_base_type = d_prompt_style = None
             d_type = d.type
-            d_type_split = d_type.split(":")
-            d_base_type = d_type_split[0]
-            d_prompt_style = d_type_split[1] if len(d_type_split) > 1 else None
+            if isinstance(d_type, str):
+                d_type_split = d_type.split(":")
+                d_base_type = d_type_split[0]
+                d_prompt_style = d_type_split[1] if len(d_type_split) > 1 else None
             if "train" in ds:
                 ds = ds["train"]
-            if ds_strategy := load(d.type, tokenizer, cfg):
+            if (
+                "input_ids" in ds.features
+                and "attention_mask" in ds.features
+                and "labels" in ds.features
+            ):
+                # dataset is already tokenized, just drop it straight in
+                datasets.append(ds)
+            elif isinstance(d.type, DictDefault):
+                ds_strategy = load("user_defined", tokenizer, cfg, d.type.to_dict())
+                ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
+                datasets.append(ds_wrapper)
+            elif ds_strategy := load(d.type, tokenizer, cfg, d):
                 ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
                 datasets.append(ds_wrapper)
             elif d_base_type == "alpaca":
@@ -262,20 +337,12 @@ def load_tokenized_prepared_datasets(
                 raise ValueError(
                     f"unhandled prompt tokenization strategy: {d.type} {suffix}"
                 )
-        LOG.info("tokenizing, merging, and shuffling master dataset")
+        LOG.info("merging datasets")
+        dataset = concatenate_datasets(datasets)
 
-        samples: List[int] = []
-        chunk_size = 1000
-        for d in datasets:
-            d_iter = iter(d)
-            while True:
-                chunk = list(itertools.islice(d_iter, chunk_size))
-                if not chunk:
-                    break
-                samples.extend(chunk)
-
-        LOG.info("shuffle")
-        dataset = Dataset.from_list(samples).shuffle(seed=seed)
+        if len(datasets) > 1:
+            LOG.info("shuffle merged datasets")
+            dataset = dataset.shuffle(seed=seed)
         if cfg.local_rank == 0:
             LOG.info(f"Saving merged prepared dataset to disk... {prepared_ds_path}")
             dataset.save_to_disk(prepared_ds_path)
@@ -374,11 +441,12 @@ def load_prepare_datasets(
             dataset = Dataset.from_list(list(constant_len_dataset))
 
             # filter out bad data
+            # TODO convert to dataset.filter(...)
             dataset = Dataset.from_list(
                 [
                     d
                     for d in dataset
-                    if len(d["input_ids"]) < cfg.sequence_len
+                    if len(d["input_ids"]) <= cfg.sequence_len
                     and len(d["input_ids"]) > 0
                     and len(d["input_ids"]) == len(d["attention_mask"])
                     and len(d["input_ids"]) == len(d["labels"])
@@ -413,7 +481,41 @@ def load_prepare_datasets(
         )
 
     if cfg.val_set_size:
-        dataset = dataset.train_test_split(test_size=cfg.val_set_size, shuffle=False)
+        # ensure we end up with the same fingerprint by doing rank0 first and being able to cache
+        to_hash_train = (
+            dataset._fingerprint  # pylint: disable=protected-access
+            + "|"
+            + str(cfg.val_set_size)
+            + "|"
+            + "train"
+            + "|"
+            + str(cfg.seed or 42)
+        )
+        to_hash_test = (
+            dataset._fingerprint  # pylint: disable=protected-access
+            + "|"
+            + str(cfg.val_set_size)
+            + "|"
+            + "test"
+            + "|"
+            + str(cfg.seed or 42)
+        )
+        train_fingerprint = hashlib.md5(
+            to_hash_train.encode(), usedforsecurity=False
+        ).hexdigest()
+        test_fingerprint = hashlib.md5(
+            to_hash_test.encode(), usedforsecurity=False
+        ).hexdigest()
+
+        with zero_first(is_main_process()):
+            dataset = dataset.train_test_split(
+                test_size=cfg.val_set_size,
+                shuffle=False,
+                seed=cfg.seed or 42,
+                train_new_fingerprint=train_fingerprint,
+                test_new_fingerprint=test_fingerprint,
+            )
+
         train_dataset = dataset["train"]
         eval_dataset = dataset["test"]
     else:

src/axolotl/utils/dataloader.py

@@ -0,0 +1,300 @@
+# pylint: skip-file
+import hashlib
+import itertools
+import logging
+import math
+from typing import Any, Callable, List, Union
+
+import numba
+import numpy as np
+from torch.utils.data import DistributedSampler, Sampler
+
+LOG = logging.getLogger("axolotl.utils.dataloader")
+
+
+@numba.njit
+def ffd_check(a: np.ndarray, c: int, n: int):
+    # First-fit-decreasing bin packing
+    # Check if a[] could fit in n bins with capacity c
+    # https://en.wikipedia.org/wiki/First-fit-decreasing_bin_packing
+
+    a = np.sort(a)[::-1]
+    bins = np.full((n,), c, dtype=a.dtype)
+    for size in a:
+        not_found = True
+        for idx in range(n):
+            if bins[idx] >= size:
+                bins[idx] -= size
+                not_found = False
+                break
+
+        if not_found:
+            return False
+
+    return True
+
+
+@numba.njit
+def ffd_with_result(a: np.ndarray, c: int, start_index: int):
+    # First-fit-decreasing bin packing (with result return)
+
+    indices = np.argsort(a)[::-1]
+    a = a[indices]
+
+    bins: List[Any] = []
+    bins_result: List[Any] = []
+    for a_id, size in enumerate(a):
+        add_new = True
+        for idx in range(len(bins)):
+            if bins[idx] >= size:
+                bins[idx] -= size
+                bins_result[idx].append(indices[a_id] + start_index)
+                add_new = False
+                break
+
+        if add_new:
+            bins.append(c - size)
+            bins_result.append([indices[a_id] + start_index])
+
+    return bins_result, len(a)
+
+
+@numba.njit
+def allocate(
+    lengths: np.ndarray, lengths_cumsum: np.ndarray, rank: int, c: int, n: int
+):
+    """
+    :param lengths: array of lengths of each sample
+    :param lengths_cumsum: cumulative sum of consecutive lengths
+    :param rank: rank for this process
+    :param c: length of tokens per batch
+    :param n: number of ranks
+    :return:
+    """
+    # Dynamic batch allocator, similar to Multifit
+    # https://en.wikipedia.org/wiki/Multifit_algorithm
+    # ~99.5% efficiency on OpenChat training set (12 * 2048 ctx len)
+
+    s = 0
+    start_index = 0
+    result = []
+    result_totseqs = []
+
+    while True:
+        # binary search [left, right)
+        left = 1
+        right = 1 + np.searchsorted(lengths_cumsum[start_index:], s + c * n, "right")
+
+        while right - left > 1:
+            mid = (left + right) // 2
+            if ffd_check(lengths[start_index : start_index + mid], c, n):
+                left = mid
+            else:
+                right = mid
+
+        # use length left
+        batch, tot_seqs = ffd_with_result(
+            lengths[start_index : start_index + left], c, start_index
+        )
+        if len(batch) < n:
+            break
+
+        start_index += left
+        s = lengths_cumsum[start_index - 1]
+
+        # add local rank
+        result.append(batch[rank])
+        # add total seqs for all ranks
+        result_totseqs.append(tot_seqs)
+        # yield batch[rank], tot_seqs, s, len(result) * c * n
+    return result, result_totseqs, s, len(result) * c * n
+
+
+def chunk(iterable, n):
+    """
+    Chunk data into tuples of length n
+    """
+    # batched('ABCDEFG', 3) --> ABC DEF G
+    if n < 1:
+        raise ValueError("n must be at least one")
+    it = iter(iterable)
+    while batch := tuple(itertools.islice(it, n)):
+        yield batch
+
+
+def hash_indices(lst: List[int]) -> str:
+    # Convert the list of integers to a string representation
+    concatenated = ",".join(map(str, lst))
+
+    # Generate the hash
+    sha256 = hashlib.sha256()
+    sha256.update(concatenated.encode())
+
+    return sha256.hexdigest()
+
+
+class MultipackDistributedDataloader:
+    """Unpadded data loading using Multipack.
+    Adapted from https://github.com/imoneoi/openchat/blob/v3_fix_mle_loss/ochat/training_deepspeed/multipack_dataloader.py
+    Approximate (at most ~1.22x) the optimal solution of the identical-machines scheduling problem, which is NP-hard.
+    """
+
+    def __init__(
+        self,
+        dataset: Any,
+        collate_fn: Callable,
+        seq_max_length: int = 2048,
+        batch_size: int = 1,
+        sampler: Union[Sampler, DistributedSampler] = None,
+        packing_efficiency_estimate: float = 1.0,
+        sample_packing_seq_len_multiplier: int = 1,
+        device_count: int = 1,
+    ):
+        # Dataset
+        self.dataset = dataset
+        self.lengths = (
+            dataset.data.column("position_ids")
+            .to_pandas()
+            .apply(lambda x: x[-1] + 1)
+            .values
+        )
+        assert isinstance(self.lengths, np.ndarray)
+        assert batch_size % sample_packing_seq_len_multiplier == 0
+        assert batch_size >= sample_packing_seq_len_multiplier
+        self.sampler = sampler
+        self.batch_size = batch_size
+        self.sample_packing_seq_len_multiplier = sample_packing_seq_len_multiplier
+        self.seq_max_length = seq_max_length
+        self.batch_max_length = batch_size * seq_max_length
+        self.collate_fn = collate_fn
+
+        self.num_replicas = 1
+        self.rank = 0
+
+        # statistics
+        self.eff_total_used = 0
+        self.eff_total_slots = 0
+        self.packing_efficiency_estimate = packing_efficiency_estimate or 1.0
+        self.device_count = device_count
+
+    def generate_batches(self, set_stats=False):
+        LOG.info("generating packed batches")
+        if self.sampler:
+            indices = [idx for idx in self.sampler]
+        else:
+            indices = range(0, len(self.dataset))
+
+        LOG.info(hash_indices(indices))
+        lengths = self.lengths[indices]
+        lengths_cumsum = np.cumsum(lengths)
+
+        batches, totseqs, total_used, total_slots = allocate(
+            lengths=lengths,
+            lengths_cumsum=lengths_cumsum,
+            rank=self.rank,
+            # c=self.batch_max_length,
+            c=self.seq_max_length * self.sample_packing_seq_len_multiplier,
+            n=self.num_replicas,
+        )
+
+        batches = [[indices[b_idx] for b_idx in batch] for batch in batches]
+
+        # statistics
+        if set_stats:
+            self.eff_total_used += total_used
+            self.eff_total_slots += total_slots
+
+        return batches, totseqs
+
+    def __iter__(self):
+        if hasattr(self.sampler, "set_epoch"):
+            new_epoch = self.sampler.epoch + 1
+            self.sampler.set_epoch(new_epoch)
+            LOG.info(f"calling sampler.set_epoch({new_epoch})")
+        all_batches, _ = self.generate_batches(set_stats=True)
+        features = self.dataset.features.keys()
+        len_remaining = self._len_est()
+        for batches in chunk(
+            all_batches, self.batch_size // self.sample_packing_seq_len_multiplier
+        ):
+            chunked_data = []
+            attn_mask_cum_idx = 0
+            for batch in batches:
+                concatenated = {}
+                batched_data = [self.dataset[batch_idx] for batch_idx in batch]
+                for feature in features:
+                    if feature == "attention_mask":
+                        arrays = [
+                            (attn_mask_cum_idx + idx + 1) * np.array(item[feature])
+                            for idx, item in enumerate(batched_data)
+                            if feature in item
+                        ]
+                        attn_mask_cum_idx += len(batched_data)
+                        concatenated[feature] = np.concatenate(arrays)
+                    else:
+                        arrays = [
+                            np.array(item[feature])
+                            for item in batched_data
+                            if feature in item
+                        ]
+                        concatenated[feature] = np.concatenate(arrays)
+                chunked_data.append(concatenated)
+            yield self.collate_fn(chunked_data)
+            len_remaining -= 1
+            if not len_remaining:
+                return
+        # yield a no-op for cases where we don't have any data left to pack
+        for i in range(0, len_remaining):
+            yield self.collate_fn(
+                [
+                    {
+                        "input_ids": [0],
+                        "labels": [-100],
+                        "attention_mask": [True],
+                        "position_ids": [0],
+                    }
+                ]
+            )
+
+    def _len_est(self):
+        lengths_sum = np.sum(self.lengths)
+        lengths_sum_per_device = lengths_sum // self.device_count
+        LOG.info(
+            f"packing_efficiency_estimate: {self.packing_efficiency_estimate} "
+            f"total_num_tokens per device: {lengths_sum_per_device}"
+        )
+
+        # shave off 1% + 1 for dealing with variance in packing from random sampler to sampler
+        return (
+            math.floor(
+                0.99
+                * lengths_sum_per_device
+                / self.packing_efficiency_estimate
+                // self.seq_max_length
+                // self.batch_size
+            )
+            - 1
+        )
+
+    def __len__(self):
+        # this doesn't return the actual length b/c with distributed samplers, not all dataloaders get
+        # the same share of total tokens
+        # if not self.eff_total_used:
+        #     batches, _ = self.generate_batches(set_stats=True)
+        # LOG.info(
+        #     f"packing_efficiency_estimate: {self.packing_efficiency_estimate} "
+        #     f"actual packing efficiency: {self.efficiency()}"
+        # )
+        return max(1, self._len_est())
+
+    def len_w_stats(self):
+        if not self.eff_total_used:
+            batches, _ = self.generate_batches(set_stats=True)
+        LOG.info(
+            f"packing_efficiency_estimate: {self.packing_efficiency_estimate} "
+            f"actual packing efficiency: {self.efficiency()}"
+        )
+        return max(1, self._len_est())
+
+    def efficiency(self):
+        return self.eff_total_used / self.eff_total_slots

src/axolotl/utils/dict.py

@@ -10,3 +10,6 @@ class DictDefault(Dict):
 
     def __missing__(self, key):
         return None
+
+    def __or__(self, other):
+        return DictDefault(super().__or__(other))

src/axolotl/utils/distributed.py

@@ -0,0 +1,55 @@
+"""
+utility helpers for distributed checks
+"""
+from contextlib import contextmanager
+
+import torch.distributed as dist
+from accelerate import Accelerator
+
+accelerate = None  # pylint: disable=invalid-name
+
+
+def load_accelerate():
+    global accelerate  # pylint: disable=global-statement
+    accelerate = Accelerator()
+
+
+def is_distributed():
+    """
+    Check if distributed training is initialized.
+    """
+    global accelerate  # pylint: disable=global-statement
+    if not accelerate:
+        accelerate = Accelerator()
+    return dist.is_available() and dist.is_initialized()
+
+
+def barrier():
+    """
+    Acts as a barrier to wait for all processes. This ensures that all processes
+    reach the barrier before proceeding further.
+    """
+    if is_distributed():
+        dist.barrier()
+
+
+def is_main_process():
+    """
+    Check if the current process is the main process.
+    If not in distributed mode, always return True.
+    """
+    if not is_distributed():
+        return True
+    return dist.get_rank() == 0
+
+
+@contextmanager
+def zero_first(is_main):
+    """
+    runs the wrapped context so that rank 0 runs first before other ranks
+    """
+    if not is_main:  # other ranks wait first
+        barrier()
+    yield
+    if is_main:  # then rank 0 waits after it has run the context
+        barrier()

src/axolotl/utils/models.py

@@ -4,65 +4,79 @@
 import logging
 import math
 import os
-from pathlib import Path
-from typing import TYPE_CHECKING, Optional, Tuple  # noqa: F401
+from typing import Optional, Tuple  # noqa: F401
 
 import bitsandbytes as bnb
 import torch
 import transformers
 from optimum.bettertransformer import BetterTransformer
+from peft import PeftConfig, prepare_model_for_kbit_training
 from transformers import (  # noqa: F401
     AutoConfig,
     AutoModelForCausalLM,
     AutoTokenizer,
     BitsAndBytesConfig,
+    GPTQConfig,
     LlamaConfig,
     PreTrainedModel,
     PreTrainedTokenizerBase,
 )
 
-from axolotl.prompt_tokenizers import LLAMA_DEFAULT_PAD_TOKEN
+from axolotl.prompt_tokenizers import LLAMA_DEFAULT_EOS_TOKEN
+from axolotl.utils.bench import log_gpu_memory_usage
+from axolotl.utils.dict import DictDefault
 
 LOG = logging.getLogger("axolotl")
 
-if TYPE_CHECKING:
-    from peft import PeftConfig  # noqa: F401
 
-    from axolotl.utils.dict import DictDefault  # noqa: F401
+def load_model_config(cfg):
+    model_config_name = cfg.base_model_config or cfg.base_model
+    trust_remote_code: bool = False or cfg.trust_remote_code
+    return AutoConfig.from_pretrained(
+        model_config_name, trust_remote_code=trust_remote_code
+    )
 
 
-def load_tokenizer(
-    tokenizer_config,
-    tokenizer_type,
-    cfg,
-):
+def load_tokenizer(cfg):
+    tokenizer_kwargs = {}
     use_fast = True  # this is the default
+
     if cfg.tokenizer_use_fast is not None:
         use_fast = cfg.tokenizer_use_fast
-    if tokenizer_type:
-        tokenizer = getattr(transformers, tokenizer_type).from_pretrained(
-            tokenizer_config,
-            trust_remote_code=cfg.trust_remote_code or False,
-            use_fast=use_fast,
-        )
-    else:
-        tokenizer = AutoTokenizer.from_pretrained(
-            tokenizer_config,
-            trust_remote_code=cfg.trust_remote_code or False,
-            use_fast=use_fast,
-        )
+    if cfg.tokenizer_legacy is not None:
+        # True is the default w/ https://github.com/huggingface/transformers/pull/25224
+        tokenizer_kwargs["legacy"] = cfg.tokenizer_legacy
+
+    tokenizer_cls = AutoTokenizer
+    if cfg.tokenizer_type:
+        tokenizer_cls = getattr(transformers, cfg.tokenizer_type)
+
+    tokenizer_config = cfg.tokenizer_config or cfg.base_model_config
+    tokenizer = tokenizer_cls.from_pretrained(
+        tokenizer_config,
+        trust_remote_code=cfg.trust_remote_code or False,
+        use_fast=use_fast,
+        **tokenizer_kwargs,
+    )
+
+    if (
+        tokenizer.__class__.__name__
+        in [
+            "LlamaTokenizer",
+            "LlamaTokenizerFast",
+            "CodeLlamaTokenizer",
+        ]
+        and hasattr(tokenizer, "pad_token")
+        and not tokenizer.pad_token
+    ):
+        # set a pad_token, but use eos_token so we don't add a new token
+        tokenizer.pad_token = LLAMA_DEFAULT_EOS_TOKEN
 
     LOG.debug(f"EOS: {tokenizer.eos_token_id} / {tokenizer.eos_token}")
     LOG.debug(f"BOS: {tokenizer.bos_token_id} / {tokenizer.bos_token}")
     LOG.debug(f"PAD: {tokenizer.pad_token_id} / {tokenizer.pad_token}")
     LOG.debug(f"UNK: {tokenizer.unk_token_id} / {tokenizer.unk_token}")
 
-    if tokenizer.__class__.__name__ in [
-        "LlamaTokenizer",
-        "LlamaTokenizerFast",
-    ]:
-        tokenizer.pad_token = LLAMA_DEFAULT_PAD_TOKEN
-
     if tokenizer.__class__.__name__ == "GPTNeoXTokenizerFast":
         tokenizer.add_special_tokens({"pad_token": "[PAD]"})
         os.environ["TOKENIZERS_PARALLELISM"] = "false"
@@ -77,27 +91,28 @@ def load_tokenizer(
 
 
 def load_model(
-    base_model, base_model_config, model_type, tokenizer, cfg, adapter="lora"
-):
-    # type: (str, str, str, PreTrainedTokenizerBase, DictDefault, Optional[str]) -> Tuple[PreTrainedModel, Optional[PeftConfig]]
+    cfg: DictDefault,
+    tokenizer: PreTrainedTokenizerBase,
+    inference: bool = False,
+) -> Tuple[PreTrainedModel, Optional[PeftConfig]]:
     """
-    Load a model from a base model and a model type.
+    Load a model for a given configuration and tokenizer.
     """
+    base_model = cfg.base_model
+    base_model_config = cfg.base_model_config
+    model_type = cfg.model_type
 
     # TODO refactor as a kwarg
     load_in_8bit = cfg.load_in_8bit
-    cfg.is_llama_derived_model = "llama" in base_model or (
-        cfg.model_type and "llama" in cfg.model_type.lower()
-    )
 
     if cfg.is_llama_derived_model and cfg.flash_attention:
-        if cfg.device not in ["mps", "cpu"] and not cfg.inference:
+        if cfg.device not in ["mps", "cpu"] and not inference:
             from axolotl.monkeypatch.llama_attn_hijack_flash import (
                 replace_llama_attn_with_flash_attn,
             )
 
             LOG.info("patching with flash attention")
-            replace_llama_attn_with_flash_attn()
+            replace_llama_attn_with_flash_attn(packed=cfg.sample_packing)
     elif cfg.is_llama_derived_model and cfg.xformers_attention:
         from axolotl.monkeypatch.llama_attn_hijack_xformers import (
             hijack_llama_attention,
@@ -106,9 +121,7 @@ def load_model(
         LOG.info("patching with xformers attention")
         hijack_llama_attention()
     elif cfg.is_llama_derived_model and cfg.sdp_attention:
-        from axolotl.monkeypatch.llama_attn_hijack_xformers import (
-            hijack_llama_sdp_attention,
-        )
+        from axolotl.monkeypatch.llama_attn_hijack_sdp import hijack_llama_sdp_attention
 
         LOG.info("patching with sdp attention")
         hijack_llama_sdp_attention()
@@ -132,97 +145,52 @@ def load_model(
         LOG.info("patching with xpos rope")
         replace_llama_rope_with_xpos_rope()
 
-    if cfg.bf16 or cfg.bfloat16:
-        torch_dtype = torch.bfloat16
-    elif cfg.load_in_8bit or cfg.fp16 or cfg.float16:
-        torch_dtype = torch.float16
-    else:
-        torch_dtype = torch.float32
-    try:
-        if cfg.gptq:
-            from alpaca_lora_4bit.monkeypatch.peft_tuners_lora_monkey_patch import (
-                replace_peft_model_with_int4_lora_model,
-            )
-
-            replace_peft_model_with_int4_lora_model()
-    except Exception as err:
-        LOG.exception(err)
-        raise err
-
-    if not cfg.gptq and (
-        (cfg.adapter == "lora" and load_in_8bit)
-        or (cfg.adapter == "qlora" and cfg.load_in_4bit)
+    if (
+        cfg.is_llama_derived_model
+        and (cfg.max_packed_sequence_len or cfg.sample_packing)
+        and not inference
     ):
-        try:
-            from peft import prepare_model_for_kbit_training
-        except ImportError:
-            # For backward compatibility
-            from peft import (
-                prepare_model_for_int8_training as prepare_model_for_kbit_training,
-            )
+        from axolotl.monkeypatch.llama_expand_mask import hijack_expand_mask
+
+        LOG.info("patching _expand_mask")
+        hijack_expand_mask()
 
     model_kwargs = {}
     if cfg.model_revision:
         model_kwargs["revision"] = cfg.model_revision
+    if cfg.gptq:
+        # TODO we should figure out how read the models config.json first
+        model_kwargs["quantization_config"] = GPTQConfig(
+            bits=cfg.gptq_bits,
+            disable_exllama=True,
+        )
     if cfg.adapter == "qlora" and cfg.load_in_4bit:
         model_kwargs["quantization_config"] = BitsAndBytesConfig(
             load_in_4bit=True,
             llm_int8_threshold=6.0,
             llm_int8_has_fp16_weight=False,
-            bnb_4bit_compute_dtype=torch_dtype,
+            bnb_4bit_compute_dtype=cfg.torch_dtype,
             bnb_4bit_use_double_quant=True,
             bnb_4bit_quant_type="nf4",
         )
     try:
-        if cfg.gptq and cfg.is_llama_derived_model:
-            from alpaca_lora_4bit.autograd_4bit import load_llama_model_4bit_low_ram
-            from huggingface_hub import snapshot_download
-
-            try:
-                snapshot_download_kwargs = {}
-                if cfg.base_model_ignore_patterns:
-                    snapshot_download_kwargs[
-                        "ignore_patterns"
-                    ] = cfg.base_model_ignore_patterns
-                cache_model_path = Path(
-                    snapshot_download(base_model, **snapshot_download_kwargs)
-                )
-                files = (
-                    list(cache_model_path.glob("*.pt"))
-                    + list(cache_model_path.glob("*.safetensors"))
-                    + list(cache_model_path.glob("*.bin"))
-                )
-                if len(files) > 0:
-                    model_path = str(files[0])
-                else:
-                    LOG.warning(
-                        "unable to find a cached model file, this will likely fail..."
-                    )
-                    model_path = str(cache_model_path)
-            except Exception:  # pylint: disable=broad-exception-caught
-                model_path = cfg.base_model
-            model, _ = load_llama_model_4bit_low_ram(
-                base_model_config if base_model_config else base_model,
-                model_path,
-                device_map=cfg.device_map,
-                half=cfg.fp16,
-                groupsize=cfg.gptq_groupsize if cfg.gptq_groupsize else -1,
-                is_v1_model=cfg.gptq_model_v1
-                if cfg.gptq_model_v1 is not None
-                else True,
-            )
-            load_in_8bit = False
-        elif cfg.is_llama_derived_model and not cfg.trust_remote_code:
+        if cfg.is_llama_derived_model and not cfg.trust_remote_code and not cfg.gptq:
             from transformers import LlamaForCausalLM
 
-            config = LlamaConfig.from_pretrained(base_model_config)
+            config_kwargs = {}
+            if cfg.rope_scaling:
+                config_kwargs["rope_scaling"] = cfg.rope_scaling
+            config = LlamaConfig.from_pretrained(
+                base_model_config,
+                **config_kwargs,
+            )
             model = LlamaForCausalLM.from_pretrained(
                 base_model,
                 config=config,
+                device_map=cfg.device_map,
                 load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
                 load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
-                torch_dtype=torch_dtype,
-                device_map="auto" if cfg.world_size == 1 else cfg.device_map,
+                torch_dtype=cfg.torch_dtype,
                 **model_kwargs,
             )
         # elif model_type == "GPTNeoXForCausalLM" and cfg.flash_attention:
@@ -252,15 +220,24 @@ def load_model(
         #     )
         #     model.train() # sets to train instead of eval mode
         elif model_type and not cfg.trust_remote_code:
-            model = getattr(transformers, model_type).from_pretrained(
-                base_model,
-                load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
-                load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
-                torch_dtype=torch_dtype,
-                device_map=cfg.device_map,
-                trust_remote_code=cfg.trust_remote_code or False,
-                **model_kwargs,
-            )
+            if cfg.gptq:
+                model = AutoModelForCausalLM.from_pretrained(
+                    base_model,
+                    device_map=cfg.device_map,
+                    torch_dtype=cfg.torch_dtype,
+                    trust_remote_code=cfg.trust_remote_code or False,
+                    **model_kwargs,
+                )
+            else:
+                model = getattr(transformers, model_type).from_pretrained(
+                    base_model,
+                    device_map=cfg.device_map,
+                    load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
+                    load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
+                    torch_dtype=cfg.torch_dtype,
+                    trust_remote_code=cfg.trust_remote_code or False,
+                    **model_kwargs,
+                )
         else:
             config = AutoConfig.from_pretrained(
                 base_model,
@@ -285,10 +262,10 @@ def load_model(
             model = AutoModelForCausalLM.from_pretrained(
                 base_model,
                 config=config,
+                device_map=cfg.device_map,
                 load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
                 load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
-                torch_dtype=torch_dtype,
-                device_map=cfg.device_map,
+                torch_dtype=cfg.torch_dtype,
                 trust_remote_code=cfg.trust_remote_code or False,
                 **model_kwargs,
             )
@@ -299,10 +276,10 @@ def load_model(
         LOG.exception(err)
         model = AutoModelForCausalLM.from_pretrained(
             base_model,
+            device_map=cfg.device_map,
             load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,
             load_in_4bit=cfg.load_in_4bit and cfg.adapter is not None,
-            torch_dtype=torch_dtype,
-            device_map=cfg.device_map,
+            torch_dtype=cfg.torch_dtype,
             trust_remote_code=cfg.trust_remote_code or False,
             **model_kwargs,
         )
@@ -324,46 +301,49 @@ def load_model(
         )
         model.config.max_position_embeddings = cfg.sequence_len
 
-    if not cfg.gptq and (
-        (cfg.adapter == "lora" and load_in_8bit)
-        or (cfg.adapter == "qlora" and cfg.load_in_4bit)
+    if model.device.type == "cuda":
+        log_gpu_memory_usage(LOG, "after model load", model.device)
+
+    # make sure these are fp32 per Ramesh et al. (2021)
+    for name, module in model.named_modules():
+        if "norm" in name:
+            module.to(torch.float32)
+        if "lm_head" in name or "embed_tokens" in name:
+            if hasattr(module, "weight"):
+                module.to(torch.float32)
+
+    needs_fa2_dtype = cfg.adapter or cfg.fsdp
+    if (cfg.adapter == "lora" and load_in_8bit) or (
+        cfg.adapter == "qlora" and cfg.load_in_4bit
     ):
         LOG.info("converting PEFT model w/ prepare_model_for_kbit_training")
+        if cfg.gradient_checkpointing:
+            model.gradient_checkpointing_enable()
         model = prepare_model_for_kbit_training(
             model, use_gradient_checkpointing=cfg.gradient_checkpointing
         )
+        needs_fa2_dtype = True
 
-        # LlamaRMSNorm layers are in fp32 after kbit_training, so we need to
-        # convert them back to fp16/bf16 for flash-attn compatibility.
-        if cfg.flash_attention and cfg.is_llama_derived_model:
-            for name, module in model.named_modules():
-                if "norm" in name:
-                    module.to(torch_dtype)
-                if "lm_head" in name or "embed_tokens" in name:
-                    if hasattr(module, "weight"):
-                        module.to(torch_dtype)
+    # LlamaRMSNorm layers are in fp32 after kbit_training or full finetune, so we need to
+    # convert them back to fp16/bf16 for flash-attn compatibility.
+    if needs_fa2_dtype and (cfg.flash_attention and cfg.is_llama_derived_model):
+        LOG.info("converting modules to %s for flash attention", cfg.torch_dtype)
+        for name, module in model.named_modules():
+            if "norm" in name:
+                module.to(cfg.torch_dtype)
+            if "lm_head" in name or "embed_tokens" in name:
+                if hasattr(module, "weight"):
+                    module.to(cfg.torch_dtype)
 
-    model, lora_config = load_adapter(model, cfg, adapter)
+    model, lora_config = load_adapter(model, cfg, cfg.adapter)
 
     if cfg.ddp and not load_in_8bit:
         model.to(f"cuda:{cfg.local_rank}")
 
-    if cfg.gptq:
-        # Scales to half
-        LOG.info("Fitting 4bit scales and zeros to half")
-        for _, module in model.named_modules():
-            if "Autograd4bitQuantLinear" in str(type(module)) or "Linear4bitLt" in str(
-                type(module)
-            ):
-                if hasattr(module, "is_v1_model") and module.is_v1_model:
-                    module.zeros = module.zeros.half()
-                module.scales = module.scales.half()
-                module.bias = module.bias.half()
-
     if (
         torch.cuda.device_count() > 1
         and int(os.getenv("WORLD_SIZE", "1")) > 1
-        and (cfg.gptq or cfg.load_in_4bit)
+        and (cfg.load_in_4bit)
     ):
         # llama is PROBABLY model parallelizable, but the default isn't that it is
         # so let's only set it for the 4bit, see
@@ -382,17 +362,22 @@ def load_model(
     if cfg.flash_optimum:
         model = BetterTransformer.transform(model)
 
+    if cfg.adapter is not None:
+        log_gpu_memory_usage(LOG, "after adapters", model.device)
+
     # TODO resume_from_checkpoint handling
     return model, lora_config
 
 
-def load_adapter(model, cfg, adapter):
-    # type: (PreTrainedModel, DictDefault, Optional[str]) -> Tuple[PreTrainedModel, Optional[PeftConfig]]
+def load_adapter(model, cfg, adapter, inference=False):
+    # type: (PreTrainedModel, DictDefault, Optional[str], bool) -> Tuple[PreTrainedModel, Optional[PeftConfig]]
 
     if adapter is None:
         return model, None
+    if hasattr(model, "enable_input_require_grads"):
+        model.enable_input_require_grads()
     if adapter in ["lora", "qlora"]:
-        return load_lora(model, cfg)
+        return load_lora(model, cfg, inference=inference)
     if adapter == "llama-adapter":
         return load_llama_adapter(model, cfg)
 
@@ -410,7 +395,7 @@ def load_llama_adapter(model, cfg):
     )
 
     if cfg.lora_model_dir:
-        LOG.info("Loading pretained LORA")
+        LOG.debug("Loading pretained PEFT - llama_adapter")
         model = PeftModel.from_pretrained(
             model,
             cfg.lora_model_dir,
@@ -424,12 +409,8 @@ def load_llama_adapter(model, cfg):
     return model, peft_config
 
 
-def find_all_linear_names(bits, model):
-    cls = (
-        bnb.nn.Linear4bit
-        if bits == 4
-        else (bnb.nn.Linear8bitLt if bits == 8 else torch.nn.Linear)
-    )
+def find_all_linear_names(model):
+    cls = (bnb.nn.Linear4bit, bnb.nn.Linear8bitLt, torch.nn.Linear)
     lora_module_names = set()
     for name, module in model.named_modules():
         if isinstance(module, cls):
@@ -442,21 +423,15 @@ def find_all_linear_names(bits, model):
     return list(lora_module_names)
 
 
-def load_lora(model, cfg):
-    # type: (PreTrainedModel, DictDefault) -> Tuple[PreTrainedModel, Optional[PeftConfig]]
+def load_lora(model, cfg, inference=False):
+    # type: (PreTrainedModel, DictDefault, bool) -> Tuple[PreTrainedModel, Optional[PeftConfig]]
 
     from peft import LoraConfig, PeftModel, get_peft_model
 
     lora_target_modules = list(cfg.lora_target_modules or [])
 
     if cfg.lora_target_linear:
-        bits = None
-        if cfg.load_in_4bit:
-            bits = 4
-        elif cfg.load_in_8bit:
-            bits = 8
-
-        linear_names = find_all_linear_names(bits, model)
+        linear_names = find_all_linear_names(model)
         LOG.info(f"found linear modules: {repr(linear_names)}")
         lora_target_modules = list(set(lora_target_modules + linear_names))
 
@@ -472,10 +447,11 @@ def load_lora(model, cfg):
     )
 
     if cfg.lora_model_dir:
+        LOG.debug("Loading pretained PEFT - LoRA")
         model = PeftModel.from_pretrained(
             model,
             cfg.lora_model_dir,
-            is_trainable=not cfg.inference,
+            is_trainable=(not inference),
         )
     else:
         model = get_peft_model(model, lora_config)

src/axolotl/utils/trainer.py

@@ -1,34 +1,98 @@
 """Module containing the Trainer class and related functions"""
-
 import importlib
 import logging
 import math
 import os
 import sys
+from contextlib import contextmanager
 from dataclasses import dataclass, field
+from functools import partial
 from pathlib import Path
-from typing import Optional
+from typing import Optional, Union
 
-import bitsandbytes as bnb
+import numpy as np
 import torch.cuda
-import transformers
-from torch import nn
+from datasets import Dataset, set_caching_enabled
 from torch.optim.lr_scheduler import OneCycleLR
+from torch.utils.data import DataLoader, DistributedSampler, RandomSampler
 from transformers import EarlyStoppingCallback, Trainer, TrainingArguments
-from transformers.trainer_pt_utils import get_parameter_names
+from transformers.trainer_pt_utils import SequentialDistributedSampler
 
+from axolotl.monkeypatch.relora import ReLoRACallback, ReLoRAScheduler
 from axolotl.utils.callbacks import (
+    GPUStatsCallback,
     SaveBetterTransformerModelCallback,
     SavePeftModelCallback,
 )
-from axolotl.utils.schedulers import (
-    InterpolatingLogScheduler,
-    get_cosine_schedule_with_quadratic_warmup,
-)
+from axolotl.utils.collators import DataCollatorForSeq2Seq
+from axolotl.utils.dataloader import MultipackDistributedDataloader
+from axolotl.utils.schedulers import get_cosine_schedule_with_quadratic_warmup
 
 LOG = logging.getLogger("axolotl")
 
 
+@torch.jit.script
+def weighted_cross_entropy(
+    logits: torch.Tensor, labels: torch.Tensor, weights: torch.Tensor
+):
+    # Flatten the logits, labels, and weights tensors
+    logits = logits.view(
+        -1, logits.size(-1)
+    )  # logits becomes of shape [batch_size*sequence_length, vocab_size]
+    labels = labels.view(-1)  # labels becomes of shape [batch_size*sequence_length]
+    weights = weights.view(-1)  # weights becomes of shape [batch_size*sequence_length]
+
+    # Compute the unweighted cross entropy loss
+    losses = torch.nn.functional.cross_entropy(logits, labels, reduction="none")
+
+    # Apply the weights to the losses and compute their sum
+    return (weights * losses).sum()
+
+
+@torch.jit.script
+def create_weighted_mask(labels: torch.Tensor):
+    # Check if the tensor is 2D. If not, unsqueeze it to make it 2D
+    if len(labels.shape) == 1:
+        labels = labels.unsqueeze(0)
+
+    weights = torch.zeros_like(labels).float()
+    for i in range(labels.shape[0]):
+        mask = labels[i] != -100
+
+        # Create a tensor to track group ids
+        group_ids = torch.zeros_like(labels[i]).int()
+        curr_group_id = 0
+
+        for j in range(1, len(labels[i])):
+            if mask[j] and not mask[j - 1]:  # switch from masked to unmasked label
+                curr_group_id += 1  # start new group
+            group_ids[j] = (
+                curr_group_id if mask[j] else 0
+            )  # assign group id if unmasked label
+
+        # Count only unmasked labels in each group
+        group_counts = torch.bincount(group_ids[mask])
+
+        mask_weights = torch.zeros_like(labels[i]).float()
+        mask_weights[mask] = 1.0 / group_counts[group_ids[mask]]
+
+        weights[i] = mask_weights
+
+    return weights.squeeze()  # squeeze the output to match the input dimension
+
+
+def trainer_weighted_loss(model_output, labels, shift_labels=True):
+    logits = (
+        model_output["logits"] if isinstance(model_output, dict) else model_output[0]
+    )
+    if shift_labels:
+        logits = logits[..., :-1, :].contiguous()
+        labels = labels[..., 1:].contiguous()
+
+    weights = create_weighted_mask(labels)
+    return weighted_cross_entropy(logits, labels, weights)
+
+
 @dataclass
 class AxolotlTrainingArguments(TrainingArguments):
     """
@@ -39,6 +103,30 @@ class AxolotlTrainingArguments(TrainingArguments):
         default=False,
         metadata={"help": "Use quadratic warmup for cosine scheduling."},
     )
+    sample_packing: bool = field(
+        default=False,
+        metadata={"help": "Use sample packing for efficient training."},
+    )
+    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"},
+    )
 
 
 class AxolotlTrainer(Trainer):
@@ -76,6 +164,77 @@ class AxolotlTrainer(Trainer):
                 return super().create_scheduler(num_training_steps, optimizer)
         return self.lr_scheduler
 
+    def _get_train_sampler(self) -> Optional[torch.utils.data.Sampler]:
+        if self.args.world_size > 1 and self.args.sample_packing:
+            return DistributedSampler(
+                self.train_dataset,
+                num_replicas=self.args.world_size,
+                rank=self.args.process_index,
+                seed=self.args.seed,
+            )
+        return super()._get_train_sampler()
+
+    def _get_eval_sampler(
+        self, eval_dataset: Dataset
+    ) -> Optional[torch.utils.data.Sampler]:
+        if self.args.world_size > 1 and self.args.sample_packing:
+            return SequentialDistributedSampler(
+                eval_dataset,
+                num_replicas=self.args.world_size,
+                rank=self.args.process_index,
+                batch_size=self.args.per_device_eval_batch_size,
+            )
+        return super()._get_eval_sampler(eval_dataset)
+
+    def get_train_dataloader(self) -> Union[DataLoader, MultipackDistributedDataloader]:
+        if self.args.sample_packing:
+            train_sampler = self._get_train_sampler()
+            return self.accelerator.prepare(
+                MultipackDistributedDataloader(
+                    self.train_dataset,
+                    batch_size=self._train_batch_size,
+                    seq_max_length=self.args.max_seq_length,
+                    collate_fn=self.data_collator,
+                    sampler=train_sampler,
+                    packing_efficiency_estimate=self.args.sample_packing_efficiency,
+                    sample_packing_seq_len_multiplier=self.args.sample_packing_seq_len_multiplier,
+                    device_count=int(os.environ.get("WORLD_SIZE", 1)),
+                )
+            )
+        return super().get_train_dataloader()
+
+    def get_eval_dataloader(
+        self, eval_dataset: Optional[Dataset] = None
+    ) -> Union[DataLoader, MultipackDistributedDataloader]:
+        if self.args.sample_packing:
+            eval_dataset = (
+                eval_dataset if eval_dataset is not None else self.eval_dataset
+            )
+
+            eval_sampler = self._get_eval_sampler(eval_dataset)
+            return self.accelerator.prepare(
+                MultipackDistributedDataloader(
+                    eval_dataset,
+                    batch_size=self.args.eval_batch_size,
+                    seq_max_length=self.args.max_seq_length,
+                    collate_fn=self.data_collator,
+                    sampler=eval_sampler,
+                    packing_efficiency_estimate=self.args.sample_packing_efficiency,
+                    sample_packing_seq_len_multiplier=self.args.eval_batch_size,
+                    device_count=int(os.environ.get("WORLD_SIZE", 1)),
+                )
+            )
+        return super().get_eval_dataloader(eval_dataset)
+
+    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)
+
 
 class OneCycleLRSchedulerTrainer(AxolotlTrainer):
     """
@@ -106,10 +265,163 @@ class OneCycleLRSchedulerTrainer(AxolotlTrainer):
         return self.lr_scheduler
 
 
-def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
-    total_num_steps = int(
-        math.ceil(len(train_dataset) * cfg.num_epochs / cfg.batch_size)
-    )
+class ReLoRATrainer(AxolotlTrainer):
+    """
+    Trainer subclass that uses the OneCycleLR scheduler
+    """
+
+    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
+
+
+def add_position_ids(sample):
+    sample["position_ids"] = torch.arange(len(sample["input_ids"]))
+    return sample
+
+
+def drop_long_seq(sample, sequence_len=2048):
+    return len(sample["input_ids"]) <= sequence_len
+
+
+@contextmanager
+def disable_datasets_caching():
+    try:
+        set_caching_enabled(False)
+        yield
+    finally:
+        set_caching_enabled(True)
+
+
+def process_datasets_for_packing(cfg, train_dataset, eval_dataset):
+    drop_long = partial(drop_long_seq, sequence_len=cfg.sequence_len)
+    train_dataset = train_dataset.filter(drop_long, num_proc=os.cpu_count())
+    if eval_dataset:
+        eval_dataset = eval_dataset.filter(drop_long, num_proc=os.cpu_count())
+
+    if cfg.sample_packing:
+        train_dataset = train_dataset.map(add_position_ids, num_proc=os.cpu_count())
+        if eval_dataset:
+            eval_dataset = eval_dataset.map(add_position_ids, num_proc=os.cpu_count())
+    return train_dataset, eval_dataset
+
+
+def calculate_total_num_steps(cfg, train_dataset, tokenizer):
+    if cfg.sample_packing:
+        # we have to drop anything longer then sequence len otherwise
+        # flash attention with position ids fails
+        if not cfg.total_num_tokens:
+            LOG.info("calculating total_num_tokens")
+            total_num_tokens = np.sum(
+                train_dataset.data.column("input_ids")
+                .to_pandas()
+                .apply(lambda x: len(x))  # pylint: disable=unnecessary-lambda
+                .values
+            )
+            LOG.info(f"📝 UPDATE CONFIG WITH: `total_num_tokens: {total_num_tokens}`")
+            cfg.total_num_tokens = total_num_tokens
+
+        if cfg.sample_packing_eff_est:
+            total_num_steps = (
+                # match count to len est in dataloader
+                (
+                    math.floor(
+                        0.99
+                        * cfg.total_num_tokens
+                        / cfg.sample_packing_eff_est
+                        / cfg.sequence_len
+                        // cfg.batch_size
+                        // int(os.environ.get("WORLD_SIZE", 1))
+                    )
+                    - 1
+                )
+                * cfg.num_epochs
+            )
+            LOG.info(
+                f"total_num_tokens: {cfg.total_num_tokens}, total_num_steps: {total_num_steps}"
+            )
+        else:
+            sampler = RandomSampler(train_dataset)
+            data_loader = MultipackDistributedDataloader(
+                train_dataset,
+                batch_size=cfg.micro_batch_size,
+                seq_max_length=cfg.max_packed_sequence_len or cfg.sequence_len,
+                collate_fn=DataCollatorForSeq2Seq(
+                    tokenizer,
+                    return_tensors="pt",
+                    padding="longest",
+                ),
+                sampler=sampler,
+                packing_efficiency_estimate=cfg.sample_packing_eff_est,
+                sample_packing_seq_len_multiplier=cfg.micro_batch_size,
+                device_count=int(os.environ.get("WORLD_SIZE", 1)),
+            )
+            data_loader_len = data_loader.len_w_stats()
+            actual_eff = data_loader.efficiency()
+            LOG.info(f"data_loader_len: {data_loader_len}")
+            total_num_steps = int(
+                math.floor(
+                    data_loader_len
+                    * cfg.micro_batch_size
+                    * cfg.num_epochs
+                    // cfg.batch_size
+                )
+            )
+            LOG.info(
+                f"📝 UPDATE CONFIG WITH: `sample_packing_eff_est: {math.ceil(actual_eff * 100.0) / 100.0}`"
+            )
+            cfg.sample_packing_eff_est = math.ceil(actual_eff * 100.0) / 100.0
+    else:
+        total_num_steps = int(
+            math.ceil(len(train_dataset) * cfg.num_epochs / cfg.batch_size)
+        )
+    LOG.info(f"total_num_steps: {total_num_steps}")
+    return total_num_steps
+
+
+def setup_fsdp_envs(cfg):
+    os.environ["ACCELERATE_USE_FSDP"] = "true"
+    if cfg.fsdp_config.fsdp_offload_params:
+        os.environ["FSDP_OFFLOAD_PARAMS"] = "true"
+    if cfg.fsdp_config.fsdp_sync_module_states:
+        os.environ["FSDP_SYNC_MODULE_STATES"] = "true"
+    if cfg.fsdp_config.fsdp_state_dict_type:
+        os.environ["FSDP_STATE_DICT_TYPE"] = cfg.fsdp_config.fsdp_state_dict_type
+    if cfg.fsdp_config.fsdp_transformer_layer_cls_to_wrap:
+        os.environ[
+            "FSDP_TRANSFORMER_CLS_TO_WRAP"
+        ] = cfg.fsdp_config.fsdp_transformer_layer_cls_to_wrap
+
+
+def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_steps):
+    if cfg.fsdp:
+        setup_fsdp_envs(cfg)
+    elif cfg.deepspeed:
+        os.environ["ACCELERATE_USE_DEEPSPEED"] = "true"
+
     warmup_steps = (
         cfg.warmup_steps
         if cfg.warmup_steps is not None
@@ -135,43 +447,19 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
         training_arguments_kwargs["seed"] = cfg.seed
 
     if cfg.gradient_checkpointing:
-        if cfg.gptq:
-            from alpaca_lora_4bit.gradient_checkpointing import (
-                apply_gradient_checkpointing,
-            )
-
-            gradient_checkpointing_ratio = (
-                cfg.gradient_checkpointing_ratio
-                if cfg.gradient_checkpointing_ratio
-                else 1.0
-            )
-            apply_gradient_checkpointing(
-                model, checkpoint_ratio=gradient_checkpointing_ratio
-            )
-        else:
-            training_arguments_kwargs[
-                "gradient_checkpointing"
-            ] = cfg.gradient_checkpointing
+        training_arguments_kwargs["gradient_checkpointing"] = cfg.gradient_checkpointing
     if cfg.fsdp:
         training_arguments_kwargs["fsdp"] = cfg.fsdp
         if cfg.fsdp_config:
             training_arguments_kwargs["fsdp_config"] = dict(cfg.fsdp_config)
 
+    # deepspeed
+    if cfg.deepspeed:
+        training_arguments_kwargs["deepspeed"] = cfg.deepspeed
+
     if cfg.lr_quadratic_warmup is not None:
         training_arguments_kwargs["lr_quadratic_warmup"] = cfg.lr_quadratic_warmup
 
-    # deepspeed
-    if (
-        os.environ.get("ACCELERATE_USE_DEEPSPEED") == "true"
-        and torch.cuda.device_count() > 1
-    ):
-        if cfg.deepspeed:
-            training_arguments_kwargs["deepspeed"] = cfg.deepspeed
-        else:
-            # make a guess here
-            # TODO search Path("./") for one
-            training_arguments_kwargs["deepspeed"] = "./ds_config.json"
-
     if cfg.adam_beta1:
         training_arguments_kwargs["adam_beta1"] = cfg.adam_beta1
     if cfg.adam_beta2:
@@ -186,10 +474,36 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
         training_arguments_kwargs["push_to_hub"] = True
         training_arguments_kwargs["hub_private_repo"] = True
 
+        if cfg.hub_strategy:
+            training_arguments_kwargs["hub_strategy"] = cfg.hub_strategy
+
     if cfg.save_safetensors:
         training_arguments_kwargs["save_safetensors"] = cfg.save_safetensors
 
+    if cfg.sample_packing_eff_est:
+        training_arguments_kwargs[
+            "sample_packing_efficiency"
+        ] = cfg.sample_packing_eff_est
+
+    if cfg.val_set_size == 0:
+        training_arguments_kwargs["evaluation_strategy"] = "no"
+    elif cfg.eval_steps:
+        training_arguments_kwargs["evaluation_strategy"] = "steps"
+        training_arguments_kwargs["eval_steps"] = cfg.eval_steps
+    else:
+        # we have an eval set, but no steps defined, use epoch
+        training_arguments_kwargs["evaluation_strategy"] = "epoch"
+
+    if cfg.save_strategy:
+        training_arguments_kwargs["save_strategy"] = cfg.save_strategy
+    else:
+        training_arguments_kwargs["save_strategy"] = (
+            "steps" if cfg.save_steps else "epoch"
+        )
+
     training_args = AxolotlTrainingArguments(  # pylint: disable=unexpected-keyword-arg
+        max_steps=total_num_steps if cfg.max_steps else -1,
+        max_seq_length=cfg.sequence_len,
         per_device_train_batch_size=cfg.micro_batch_size,
         per_device_eval_batch_size=cfg.eval_batch_size
         if cfg.eval_batch_size is not None
@@ -198,12 +512,9 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
         eval_accumulation_steps=cfg.gradient_accumulation_steps,
         num_train_epochs=cfg.num_epochs,
         learning_rate=cfg.learning_rate,
-        evaluation_strategy="steps" if cfg.val_set_size > 0 else "no",
-        save_strategy="steps" if cfg.save_steps else "epoch",
-        eval_steps=cfg.eval_steps if cfg.val_set_size > 0 else None,
         save_steps=cfg.save_steps,
         output_dir=cfg.output_dir,
-        save_total_limit=3,
+        save_total_limit=cfg.save_total_limit if cfg.save_total_limit else 4,
         load_best_model_at_end=(
             cfg.load_best_model_at_end is not False
             and cfg.val_set_size > 0
@@ -221,6 +532,10 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
         if cfg.lr_scheduler and cfg.lr_scheduler not in ("one_cycle", "log_sweep")
         else "cosine",
         weight_decay=cfg.weight_decay if cfg.weight_decay is not None else 0.0,
+        sample_packing=cfg.sample_packing if cfg.sample_packing else False,
+        sample_packing_seq_len_multiplier=cfg.micro_batch_size,
+        relora_steps=cfg.relora_steps,
+        relora_warmup_steps=cfg.relora_warmup_steps,
         **training_arguments_kwargs,
     )
 
@@ -230,68 +545,13 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
         if Path(cfg.torchdistx_path).exists():
             sys.path.append(cfg.torchdistx_path)
             importlib.import_module("torchdistx")
-    if (
-        cfg.optimizer == "adamw_bnb_8bit"
-        and not cfg.gptq
-        and "deepspeed" not in training_arguments_kwargs
-        and not cfg.fsdp
-    ):
-        decay_parameters = get_parameter_names(model, [nn.LayerNorm])
-        decay_parameters = [name for name in decay_parameters if "bias" not in name]
-        optimizer_grouped_parameters = [
-            {
-                "params": [
-                    p
-                    for n, p in model.named_parameters()
-                    if (n in decay_parameters and p.requires_grad)
-                ],
-                "weight_decay": training_args.weight_decay,
-            },
-            {
-                "params": [
-                    p
-                    for n, p in model.named_parameters()
-                    if (n not in decay_parameters and p.requires_grad)
-                ],
-                "weight_decay": 0.0,
-            },
-        ]
-
-        optimizer = bnb.optim.Adam8bit(
-            optimizer_grouped_parameters,
-            betas=(training_args.adam_beta1, training_args.adam_beta2),
-            eps=training_args.adam_epsilon,
-            lr=training_args.learning_rate,
-        )
-
-        if cfg.lr_scheduler == "one_cycle":
-            lr_scheduler_kwargs = (
-                cfg.lr_scheduler_kwargs if cfg.lr_scheduler_kwargs else {}
-            )
-            lr_scheduler = OneCycleLR(
-                optimizer,
-                cfg.learning_rate,
-                total_steps=total_num_steps,
-                epochs=cfg.num_epochs,
-                div_factor=cfg.lr_div_factor if cfg.lr_div_factor else 6,
-                **lr_scheduler_kwargs,
-            )
-        elif cfg.lr_scheduler == "log_sweep":
-            lr_scheduler = InterpolatingLogScheduler(
-                optimizer,
-                cfg.warmup_steps,
-                cfg.log_sweep_min_lr if cfg.log_sweep_min_lr else 1e-10,
-                cfg.log_sweep_max_lr if cfg.log_sweep_max_lr else 10,
-            )
-        else:
-            lr_scheduler = transformers.get_cosine_schedule_with_warmup(
-                optimizer,
-                training_args.warmup_steps,
-                total_num_steps,
-            )
-        trainer_kwargs["optimizers"] = (optimizer, lr_scheduler)
 
     callbacks = []
+    callbacks.append(GPUStatsCallback(cfg))
+
+    if cfg.relora_steps:
+        callbacks.append(ReLoRACallback(cfg))
+
     # TODO on_save callback to sync checkpoints to GCP/AWS in background
     if cfg.early_stopping_patience:
         early_stop_cb = EarlyStoppingCallback(
@@ -309,16 +569,18 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
         callbacks.append(SaveBetterTransformerModelCallback)
 
     data_collator_kwargs = {
-        "padding": True,
+        "padding": True,  # True/"longest" is the default
     }
-    if cfg.collator_pad_to_longest:
-        data_collator_kwargs["padding"] = "longest"
+    if cfg.pad_to_sequence_len:
+        data_collator_kwargs["pad_to_multiple_of"] = 64 * math.ceil(
+            cfg.sequence_len / 64
+        )
     else:
-        data_collator_kwargs["pad_to_multiple_of"] = 8
+        # 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
 
     if cfg.is_llama_derived_model and cfg.landmark_attention:
-        from functools import partial
-
         from axolotl.monkeypatch.llama_landmark_attn import (
             add_mem_tokens,
             get_mem_id,
@@ -336,17 +598,17 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
                 num_proc=32,
             )
 
-    trainer_cls = (
-        OneCycleLRSchedulerTrainer
-        if cfg.lr_scheduler == "one_cycle" and (cfg.fsdp or cfg.adapter == "qlora")
-        else AxolotlTrainer
-    )
+    trainer_cls = AxolotlTrainer
+    if cfg.lr_scheduler == "one_cycle" and (cfg.fsdp or cfg.adapter == "qlora"):
+        trainer_cls = OneCycleLRSchedulerTrainer
+    elif cfg.relora_steps:
+        trainer_cls = ReLoRATrainer
     trainer = trainer_cls(
         model=model,
         train_dataset=train_dataset,
         eval_dataset=eval_dataset,
         args=training_args,
-        data_collator=transformers.DataCollatorForSeq2Seq(
+        data_collator=DataCollatorForSeq2Seq(
             tokenizer,
             return_tensors="pt",
             **data_collator_kwargs,

src/axolotl/utils/wandb.py

@@ -9,6 +9,8 @@ def setup_wandb_env_vars(cfg):
     elif cfg.wandb_project and len(cfg.wandb_project) > 0:
         os.environ["WANDB_PROJECT"] = cfg.wandb_project
         cfg.use_wandb = True
+        if cfg.wandb_entity and len(cfg.wandb_entity) > 0:
+            os.environ["WANDB_ENTITY"] = cfg.wandb_entity
         if cfg.wandb_watch and len(cfg.wandb_watch) > 0:
             os.environ["WANDB_WATCH"] = cfg.wandb_watch
         if cfg.wandb_log_model and len(cfg.wandb_log_model) > 0:

tests/monkeypatch/test_llama_attn_hijack_flash.py

@@ -0,0 +1,30 @@
+"""
+Unit tests for the monkeypatch utils
+"""
+import unittest
+
+import torch
+
+from axolotl.monkeypatch.utils import get_cu_seqlens, get_cu_seqlens_from_pos_ids
+
+
+class TestMonkeyPatchUtils(unittest.TestCase):
+    """
+    Unit test class for monkeypatch utils
+    """
+
+    def test_get_cu_seqlens_1d(self):
+        attn_mask = torch.tensor([[1, 1, 1, 1, 2, 2, 2, 3, 3, 3, 3, 3, 4, 4, 0, 0]])
+        target_res = torch.tensor([0, 4, 7, 12, 14, 16], dtype=torch.int32)
+        self.assertTrue(torch.allclose(get_cu_seqlens(attn_mask)[0], target_res))
+
+    def test_get_cu_seqlens_from_pos_ids_1d(self):
+        position_ids = torch.tensor([[0, 1, 2, 3, 0, 1, 2, 0, 1, 2, 3, 4, 0, 1, 0, 0]])
+        target_res = torch.tensor([0, 4, 7, 12, 14, 16], dtype=torch.int32)
+        self.assertTrue(
+            torch.allclose(get_cu_seqlens_from_pos_ids(position_ids)[0], target_res)
+        )
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_dict.py

@@ -72,6 +72,13 @@ class DictDefaultTest(unittest.TestCase):
 
         assert cfg.random_key is None, "DictDefault should return None for missing keys"
 
+    def test_dict_or(self):
+        cfg = DictDefault({}) | DictDefault({})
+
+        assert (
+            cfg.random_key is None
+        ), "DictDefault should return None for missing keys after | operation"
+
     def test_dict_nested_missingparentkey(self):
         """
         Due to subclassing Dict, DictDefault will error if we try to access a nested key whose parent key does not exist.

tests/test_expand_mask.py

@@ -0,0 +1,44 @@
+"""
+Unit tests for the monkey patch for expand mask to handle packed sequences
+"""
+import unittest
+
+import torch
+
+from axolotl.monkeypatch.llama_expand_mask import _expand_mask
+
+
+class TestExpandMask(unittest.TestCase):
+    """
+    Test class for attention mask expansion for packed sequences
+    """
+
+    def test_output(self):
+        mask = torch.tensor([[1, 1, 1, 2], [2, 3, 3, 0]])
+        dtype = torch.float32
+        expected_output = torch.tensor(
+            [
+                [
+                    [
+                        [0.0000e00, -3.4028e38, -3.4028e38, -3.4028e38],
+                        [0.0000e00, 0.0000e00, -3.4028e38, -3.4028e38],
+                        [0.0000e00, 0.0000e00, 0.0000e00, -3.4028e38],
+                        [-3.4028e38, -3.4028e38, -3.4028e38, 0.0000e00],
+                    ]
+                ],
+                [
+                    [
+                        [0.0000e00, -3.4028e38, -3.4028e38, -3.4028e38],
+                        [-3.4028e38, 0.0000e00, -3.4028e38, -3.4028e38],
+                        [-3.4028e38, 0.0000e00, 0.0000e00, -3.4028e38],
+                        [-3.4028e38, -3.4028e38, -3.4028e38, -3.4028e38],
+                    ]
+                ],
+            ]
+        )
+        # Check that the output matches the expected output
+        self.assertTrue(torch.allclose(_expand_mask(mask, dtype), expected_output))
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_packed_dataset.py

@@ -27,7 +27,7 @@ class TestPacking(unittest.TestCase):
             }
         )
 
-    def test_resets_attention(self):
+    def test_increments_attention(self):
         prompter = AlpacaPrompter("chat")
         strat = AlpacaPromptTokenizingStrategy(
             prompter,
@@ -55,10 +55,14 @@ class TestPacking(unittest.TestCase):
         # first example doesn't have mask reset
         assert example["input_ids"][0] == self.tokenizer.bos_token_id
         assert example["attention_mask"][0] == 1
+        assert example["position_ids"][0] == 0
+        assert example["position_ids"][1] == 1
 
         # but subsequent one does
         assert example["input_ids"][next_bos_index] == self.tokenizer.bos_token_id
-        assert example["attention_mask"][next_bos_index] == 0
+        assert example["attention_mask"][next_bos_index] == 2
+        assert example["position_ids"][next_bos_index] == 0
+        assert example["position_ids"][next_bos_index + 1] == 1
 
 
 if __name__ == "__main__":

tests/test_prompt_tokenizers.py

@@ -4,13 +4,17 @@ import logging
 import unittest
 from pathlib import Path
 
-from transformers import AutoTokenizer
+from transformers import AutoTokenizer, LlamaTokenizer
 
 from axolotl.prompt_strategies.alpaca_chat import NoSystemPrompter
 from axolotl.prompt_strategies.alpaca_w_system import (
     InstructionWSystemPromptTokenizingStrategy,
     SystemDataPrompter,
 )
+from axolotl.prompt_strategies.llama2_chat import (
+    Llama2ChatPrompter,
+    LLama2ChatTokenizingStrategy,
+)
 from axolotl.prompt_tokenizers import (
     AlpacaPromptTokenizingStrategy,
     ShareGPTPromptTokenizingStrategy,
@@ -130,9 +134,95 @@ class InstructionWSystemPromptTokenizingStrategyTest(unittest.TestCase):
             "output": "Hi! How can I help?",
         }
         example = strat.tokenize_prompt(sample)
-        assert example["input_ids"][0:4] == [1, 835, 2184, 29901]  # "<s>### System:"
-        assert example["input_ids"][5:7] == [1509, 20118]  # "use cot"
-        assert example["input_ids"][9] == 11889  # USER
+        assert example["input_ids"][0:5] == [
+            1,
+            28962,
+            1254,
+            12665,
+            29901,
+        ]  # "<s>SYSTEM:"
+        assert example["input_ids"][5:7] == [671, 20118]  # " use cot"
+        assert example["input_ids"][8] == 11889  # USER
+
+
+class Llama2ChatTokenizationTest(unittest.TestCase):
+    """
+    Test class for prompt tokenization strategies with sys prompt from the dataset
+    """
+
+    def setUp(self) -> None:
+        # pylint: disable=duplicate-code
+        self.tokenizer = LlamaTokenizer.from_pretrained("NousResearch/Llama-2-7b-hf")
+        # woraround because official Meta repos are not open
+
+    def test_llama2_chat_integration(self):
+        with open(
+            Path(__file__).parent / "fixtures/conversation.json", encoding="utf-8"
+        ) as fin:
+            data = fin.read()
+            conversation = json.loads(data)
+        with open(
+            Path(__file__).parent / "fixtures/conversation.tokenized_llama2chat.json",
+            encoding="utf-8",
+        ) as fin:
+            data = fin.read()
+            tokenized_conversation = json.loads(data)
+        prompter = Llama2ChatPrompter()
+        strat = LLama2ChatTokenizingStrategy(
+            prompter,
+            self.tokenizer,
+            False,
+            4096,
+        )
+        example = strat.tokenize_prompt(conversation)
+        for fields in ["input_ids", "attention_mask", "labels"]:
+            self.assertEqual(len(example[fields]), len(tokenized_conversation[fields]))
+            self.assertEqual(example[fields], tokenized_conversation[fields])
+
+    def compare_with_transformers_integration(self):
+        # this needs transformers >= v4.31.0
+        from transformers.models.llama.tokenization_llama import B_SYS, E_SYS
+        from transformers.pipelines.conversational import Conversation
+
+        # from transformers.models.llama.tokenization_llama import DEFAULT_SYSTEM_PROMPT
+        # broken as of 23/7/20
+        # see https://github.com/huggingface/transformers/pull/24935
+        # pylint: disable=C0103
+        DEFAULT_SYSTEM_PROMPT = """\
+You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature.
+
+If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information."""
+        with open(
+            Path(__file__).parent / "fixtures/conversation.json", encoding="utf-8"
+        ) as fin:
+            data = fin.read()
+            conversation = json.loads(data)
+        with open(
+            Path(__file__).parent / "fixtures/conversation.tokenized_llama2chat.json",
+            encoding="utf-8",
+        ) as fin:
+            data = fin.read()
+            tokenized_conversation = json.loads(data)
+
+        user_input = []
+        answers = []
+        for msg in conversation["conversations"]:
+            if msg["from"] == "human":
+                user_input.append(msg["value"])
+            else:
+                answers.append(msg["value"])
+        hf_conf = Conversation(
+            text=user_input[-1],
+            past_user_inputs=[B_SYS + DEFAULT_SYSTEM_PROMPT + E_SYS + user_input[0]]
+            + user_input[1:-1],
+            generated_responses=answers,
+        )
+        # pylint: disable=W0212
+        hf_tokens = self.tokenizer._build_conversation_input_ids(hf_conf)
+
+        self.assertEqual(
+            hf_tokens, tokenized_conversation["input_ids"][: len(hf_tokens)]
+        )
 
 
 if __name__ == "__main__":

tests/test_prompters.py

@@ -70,7 +70,7 @@ class AlpacaPrompterTest(unittest.TestCase):
             )
         )
         assert "use cot" in res
-        assert res.startswith("### System:")
+        assert res.startswith("SYSTEM:")
         assert "### Instruction:" not in res
         assert "### Input:" not in res
         assert "alpacas" in res

tests/test_tokenizers.py

@@ -13,17 +13,22 @@ class TestTokenizers(unittest.TestCase):
     """
 
     def test_default_use_fast(self):
-        cfg = DictDefault({})
-        tokenizer = load_tokenizer("huggyllama/llama-7b", None, cfg)
+        cfg = DictDefault(
+            {
+                "tokenizer_config": "huggyllama/llama-7b",
+            }
+        )
+        tokenizer = load_tokenizer(cfg)
         assert "Fast" in tokenizer.__class__.__name__
 
     def test_dont_use_fast(self):
         cfg = DictDefault(
             {
+                "tokenizer_config": "huggyllama/llama-7b",
                 "tokenizer_use_fast": False,
             }
         )
-        tokenizer = load_tokenizer("huggyllama/llama-7b", None, cfg)
+        tokenizer = load_tokenizer(cfg)
         assert "Fast" not in tokenizer.__class__.__name__
 
 

tests/test_validation.py

@@ -6,8 +6,8 @@ from typing import Optional
 
 import pytest
 
+from axolotl.utils.config import validate_config
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.validation import validate_config
 
 
 class ValidationTest(unittest.TestCase):
@@ -313,3 +313,27 @@ class ValidationTest(unittest.TestCase):
         )
 
         validate_config(cfg)
+
+    def test_packing(self):
+        cfg = DictDefault(
+            {
+                "max_packed_sequence_len": 2048,
+            }
+        )
+        with self._caplog.at_level(logging.WARNING):
+            validate_config(cfg)
+            assert any(
+                "max_packed_sequence_len will be deprecated in favor of sample_packing"
+                in record.message
+                for record in self._caplog.records
+            )
+
+        cfg = DictDefault(
+            {
+                "max_packed_sequence_len": 2048,
+                "sample_packing": True,
+            }
+        )
+        regex_exp = r".*set only one of max_packed_sequence_len \(deprecated soon\) or sample_packing.*"
+        with pytest.raises(ValueError, match=regex_exp):
+            validate_config(cfg)