[MoE] DeepSeek-V3/R1

[kylesayrs]

Purpose

• Support DeepSeek-V3 and R1
• Update MoE examples to reflect the current state of MoE models
• Share information about sequential onloading and deepseekv3 in readme

Fixes

• Fixes [DeepSeek-V3/R1] Anyone had success quantizing DeepSeek-V3 using llm-compressor? #1482
• Fixes Has anyone successfully quantinize Deepseek-R1 to w8a8? #1274
• Fixes [Question] Has anyone successfully quantinize Deepseek-V3 to int4-w4a16? #1203

Changes

• Add readme blurb and sequential onloading and deepseek r1
• Add example for R1
• Add a prepare_for_calibration method which replaces the MoE module with a module which calibrates all experts with all tokens (but still gates expert outputs as the model would normally)
  • In the future we can make this method more configurable to support
    • Sending all tokens to all experts
    • Using inference-time activations vs train-time activations

Testing

• Ran deepseek r1 example to completion

[gemini-code-assist]

Summary of Changes

Hello @kylesayrs, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request introduces initial support for the DeepSeekV3 model, specifically focusing on enabling its quantization using the GPTQ algorithm. It includes a new example script that showcases the end-to-end process of quantizing DeepSeekV3 to W4A16 and adds a custom module to correctly handle the model's Mixture-of-Experts (MoE) layers during the quantization process. This addresses issue #1482.

Highlights

• DeepSeekV3 Quantization Example: Added a new example script (examples/quantization_w4a16/deepseekv3_example.py) demonstrating how to quantize the DeepSeekV3 model to W4A16 using the GPTQ modifier. This script includes steps for loading the model, preparing the calibration dataset, configuring the GPTQ recipe, applying quantization, saving the compressed model, and performing a sample generation test.
• DeepSeekV3 MoE Handling: Introduced a new utility class DeepseekV3MoELinears in src/llmcompressor/utils/module.py specifically designed to handle the unique structure of the DeepSeekV3 MoE layer during quantization. The example script uses the module_bfs utility to replace the standard DeepseekV3MoE module with this new class before applying quantization.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in issue comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist is currently in preview and may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments to provide feedback.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configureGemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request adds an example for quantizing DeepSeekV3 models and introduces a DeepseekV3MoELinears utility class. The review focuses on improving the example's robustness and enhancing documentation in the new utility class.

[kylesayrs]

Splitting by DeepseekV3Attention and DeepseekV3MLP yields 14971 subgraphs
Memory usage is ~2331MiB
At about ~12s per expert, ~48 hours (one A100)

This means a single A100 (80gb) can, in theory, parallelize ~34x, cutting the runtime to about ~1.4 hours

[kylesayrs]

Another promising approach to reduce runtime, which I scoped out with @anmarques, would be to implement the following:

1. Implement the option to dispatch a sequential target across N-many GPUs (where N is how many are available). This dispatch would occur before calibration
2. Implement async gptq quantization (each quantization step kicks off an async thread which operates on the same device as the module & hessian)

Implementing these two features would allow a user to specify a sequential target (such as a decoder layer), and as long as one layer + hessians fits across their N gpus, all quantization operations would be fully parallelized.

This would enable maximal parallelization (excluding parallelizing calibration, which is more honorus and less beneficial than quant parallel). In theory, you could quantize deepseekv3 in 20 minutes across 4 A100s.

[ashgold]

Is this approach equally applicable to AWQ?

[kylesayrs]

@ashgold Yes, a similar technique could be applied for AWQ, where the smoothing of [qkv + gate/up] or [moe experts] could be done in parallel within an onloaded decoder layer.

[kylesayrs]

This branch is good to go, just requires validating each of the example scripts
EDIT: Good to go

[brian-dellabetta]

one nit on documentation, otherwise LGTM