[cuda] qdq support for fp8 #47
Conversation
Signed-off-by: yiliu30 <yi4.liu@intel.com>
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Summary of Changes
Hello @yiliu30, 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 foundational support for Quantize-Dequantize (QDQ) operations within the FP8 quantization scheme. The primary goal is to enable FP8 models to run by dequantizing weights and activations to a higher precision before computation, which broadens hardware compatibility and supports specific quantization workflows. A new environment variable controls this behavior, and the core changes involve modifying the computation path for FP8 linear layers.
Highlights
- New Environment Variable: I've introduced a new environment variable,
VLLM_W8A8_FP8_QDQ_MODE, which acts as a toggle to enable or disable the Quantize-Dequantize (QDQ) path for W8A8 FP8 quantization. This provides explicit control over this new functionality. - Expanded Hardware Compatibility: With the
VLLM_W8A8_FP8_QDQ_MODEenabled, the minimum required GPU compute capability for W8A8 FP8 operations has been lowered from SM89 (Lovelace) to SM80 (Ampere). This change allows FP8 models utilizing the QDQ approach to run on a broader range of NVIDIA GPUs. - QDQ Computation Path Implementation: A new computation path has been implemented within the W8A8 FP8 linear layers. When
VLLM_W8A8_FP8_QDQ_MODEis active, quantized inputs and weights are first dequantized to a higher precision before the matrix multiplication is performed. This ensures that the computation itself happens in a higher precision, aligning with the QDQ methodology. - Output Padding Adjustment: I've modified the output padding behavior for W8A8 FP8 linear layers. Specifically, when the QDQ mode is active, output padding is explicitly set to
None, which adjusts how the output tensors are handled in this specific quantization scenario.
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
-
Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩
There was a problem hiding this comment.
Code Review
This pull request introduces a Quantize-Dequantize (QDQ) mode for FP8, useful for emulation and debugging on hardware without native FP8 support. The implementation uses a new environment variable VLLM_W8A8_FP8_QDQ_MODE to conditionally dequantize tensors before matrix multiplication. There are a couple of high-severity issues related to leftover debugging code (breakpoint() calls) that must be removed, and a medium-severity suggestion to improve code clarity.
| # from disk in this case. Skip requantization in this case (since) | ||
| # we already are quantized with the single scale. | ||
| # * Sample Model: nm-testing/Phi-3-mini-128k-instruct-FP8 | ||
| # breakpoint() |
There was a problem hiding this comment.
This breakpoint() call is a leftover from debugging and should be removed.
| # TODO(luka) remove this parameter in favor of __init__ | ||
| use_per_token_if_dynamic: Optional[bool] = None | ||
| ) -> torch.Tensor: | ||
| # breakpoint() |
There was a problem hiding this comment.
This breakpoint() call is a debugging remnant and should be removed.
| if envs.VLLM_W8A8_FP8_QDQ_MODE: | ||
| pad_output = None |
There was a problem hiding this comment.
Using False instead of None for pad_output more explicitly conveys the intent to disable padding when VLLM_W8A8_FP8_QDQ_MODE is enabled.
| if envs.VLLM_W8A8_FP8_QDQ_MODE: | |
| pad_output = None | |
| if envs.VLLM_W8A8_FP8_QDQ_MODE: | |
| pad_output = False |
Essential Elements of an Effective PR Description Checklist
PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS ABOVE HAVE BEEN CONSIDERED.
Purpose
Test Plan
Test Result
BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing (anything written below this line will be removed by GitHub Actions)