fix: zero-init workspace buffer for trtllm-gen fmha #1643
Conversation
![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 @yyihuang, 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 refines workspace buffer management for trtllm-gen fmha operations within the FlashInfer library. It ensures that trtllm-gen fmha receives a dedicated, zero-initialized workspace, which is crucial for maintaining correct internal state across kernel calls. This change improves the robustness and correctness of these specific attention mechanisms by separating their workspace requirements from general-purpose buffers.
Highlights
- Dedicated Workspace Buffer for trtllm-gen FMHA: Introduced a new, dedicated global workspace buffer (global_trtllm_gen_fmha_workspace_buffer) that is explicitly zero-initialized for use with trtllm-gen fmha operations. This ensures proper counter behavior across kernel executions.
- Workspace Buffer Initialization Change: The existing general global_workspace_buffer is now initialized with torch.empty instead of torch.zeros, as it no longer requires zero-initialization for its specific use cases.
- Test File Updates: Updated test_trtllm_gen_attention.py and test_trtllm_gen_mla.py to correctly utilize the newly introduced zero-initialized workspace buffer for trtllm-gen fmha and the general workspace buffer for other operations.
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 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. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.
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 correctly addresses the need for a zero-initialized workspace buffer for trtllm-gen fmha by introducing a separate, zero-initialized global buffer. The changes in the test files properly distinguish between the workspace for the reference implementation (now using torch.empty for efficiency) and the one for trtllm-gen. My main feedback is to refactor the duplicated buffer initialization logic into a shared helper function to improve code maintainability across the test files.
| global global_workspace_buffer, global_trtllm_gen_fmha_workspace_buffer | ||
| if global_workspace_buffer is None: | ||
| global_workspace_buffer = torch.zeros( | ||
| global_workspace_buffer = torch.empty( | ||
| workspace_size, dtype=torch.int8, device=GPU_DEVICE | ||
| ) | ||
| workspace_buffer = global_workspace_buffer | ||
| if global_trtllm_gen_fmha_workspace_buffer is None: | ||
| global_trtllm_gen_fmha_workspace_buffer = torch.zeros( | ||
| workspace_size, dtype=torch.int8, device=GPU_DEVICE | ||
| ) | ||
| workspace_buffer_ref = global_workspace_buffer | ||
| workspace_buffer = global_trtllm_gen_fmha_workspace_buffer |
There was a problem hiding this comment.
This block of code for initializing workspace buffers is repeated in test_trtllm_batch_decode (lines 514-524) and test_trtllm_gen_prefill_deepseek (lines 713-723) within this file. A similar block also exists in tests/test_trtllm_gen_mla.py.
To improve maintainability and reduce code duplication, consider extracting this logic into a shared helper function in a test utility file (e.g., tests/conftest.py or a new tests/utils.py) and importing it where needed.
A shared helper could look like this:
from typing import Tuple
def get_workspace_buffers(device: str) -> Tuple[torch.Tensor, torch.Tensor]:
"""Initializes and returns the workspace buffers."""
global global_workspace_buffer, global_trtllm_gen_fmha_workspace_buffer
if global_workspace_buffer is None:
global_workspace_buffer = torch.empty(
workspace_size, dtype=torch.int8, device=device
)
if global_trtllm_gen_fmha_workspace_buffer is None:
global_trtllm_gen_fmha_workspace_buffer = torch.zeros(
workspace_size, dtype=torch.int8, device=device
)
return global_trtllm_gen_fmha_workspace_buffer, global_workspace_bufferThen you can replace the repeated blocks in all test functions with a single call to this helper.
| global global_workspace_buffer, global_trtllm_gen_fmha_workspace_buffer | |
| if global_workspace_buffer is None: | |
| global_workspace_buffer = torch.zeros( | |
| global_workspace_buffer = torch.empty( | |
| workspace_size, dtype=torch.int8, device=GPU_DEVICE | |
| ) | |
| workspace_buffer = global_workspace_buffer | |
| if global_trtllm_gen_fmha_workspace_buffer is None: | |
| global_trtllm_gen_fmha_workspace_buffer = torch.zeros( | |
| workspace_size, dtype=torch.int8, device=GPU_DEVICE | |
| ) | |
| workspace_buffer_ref = global_workspace_buffer | |
| workspace_buffer = global_trtllm_gen_fmha_workspace_buffer | |
| workspace_buffer, workspace_buffer_ref = get_workspace_buffers(GPU_DEVICE) |
| global global_workspace_buffer, global_trtllm_gen_fmha_workspace_buffer | ||
| if global_workspace_buffer is None: | ||
| global_workspace_buffer = torch.zeros( | ||
| global_workspace_buffer = torch.empty( | ||
| workspace_size, dtype=torch.int8, device=device | ||
| ) | ||
| workspace_buffer = global_workspace_buffer | ||
| if global_trtllm_gen_fmha_workspace_buffer is None: | ||
| global_trtllm_gen_fmha_workspace_buffer = torch.zeros( | ||
| workspace_size, dtype=torch.int8, device=device | ||
| ) | ||
| workspace_buffer = global_trtllm_gen_fmha_workspace_buffer | ||
| workspace_buffer_ref = global_workspace_buffer |
There was a problem hiding this comment.
This workspace buffer initialization logic is duplicated from tests/test_trtllm_gen_attention.py. To avoid code duplication across test files and improve maintainability, consider creating a shared helper function in a test utility file (e.g., tests/conftest.py or a new tests/utils.py) and importing it in both files.
A shared helper could look like this:
from typing import Tuple
def get_workspace_buffers(device: str) -> Tuple[torch.Tensor, torch.Tensor]:
"""Initializes and returns the workspace buffers."""
global global_workspace_buffer, global_trtllm_gen_fmha_workspace_buffer
if global_workspace_buffer is None:
global_workspace_buffer = torch.empty(
workspace_size, dtype=torch.int8, device=device
)
if global_trtllm_gen_fmha_workspace_buffer is None:
global_trtllm_gen_fmha_workspace_buffer = torch.zeros(
workspace_size, dtype=torch.int8, device=device
)
return global_trtllm_gen_fmha_workspace_buffer, global_workspace_bufferAfter creating the shared helper, you can replace this block with a single call.
| global global_workspace_buffer, global_trtllm_gen_fmha_workspace_buffer | |
| if global_workspace_buffer is None: | |
| global_workspace_buffer = torch.zeros( | |
| global_workspace_buffer = torch.empty( | |
| workspace_size, dtype=torch.int8, device=device | |
| ) | |
| workspace_buffer = global_workspace_buffer | |
| if global_trtllm_gen_fmha_workspace_buffer is None: | |
| global_trtllm_gen_fmha_workspace_buffer = torch.zeros( | |
| workspace_size, dtype=torch.int8, device=device | |
| ) | |
| workspace_buffer = global_trtllm_gen_fmha_workspace_buffer | |
| workspace_buffer_ref = global_workspace_buffer | |
| workspace_buffer, workspace_buffer_ref = get_workspace_buffers(device) |
| workspace_size, dtype=torch.int8, device=GPU_DEVICE | ||
| ) | ||
| workspace_buffer = global_workspace_buffer | ||
| if global_trtllm_gen_fmha_workspace_buffer is None: |
There was a problem hiding this comment.
this only make it zero the first time, later execution won't enter this branch, is it expected?
And is it possible to reuse with ref workspace buffer?
There was a problem hiding this comment.
Yes it's expected.
No. trtllm-gen kernel and fi kernel should re-use individual workspace as fi kernel does not require zero-init workspace.
3 participants
📌 Description
trtllm-gen fmha should take exclusive zero-init workspace buffer to ensure zero counter cross kernel executions. Update test file as an example.
🔍 Related Issues
🚀 Pull Request Checklist
Thank you for contributing to FlashInfer! Before we review your pull request, please make sure the following items are complete.
✅ Pre-commit Checks
pre-commitby runningpip install pre-commit(or used your preferred method).pre-commit install.pre-commit run --all-filesand fixed any reported issues.🧪 Tests
unittest, etc.).Reviewer Notes