[TPU][Bugfix] fix kv cache padding #20048
Conversation
Signed-off-by: Chengji Yao <chengjiyao@google.com>
|
👋 Hi! Thank you for contributing to the vLLM project. 💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels. Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging. To run CI, PR reviewers can either: Add 🚀 |
There was a problem hiding this comment.
Summary of Changes
Hello @yaochengji, 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 addresses a critical bug in VLLM's TPU backend where the KV cache memory allocation did not correctly account for head size padding. Previously, this oversight could lead to out-of-memory (OOM) errors. The changes ensure that the usable memory for the KV cache is appropriately adjusted based on the padded head size, thereby preventing crashes and improving stability on TPU.
Highlights
- TPU KV Cache Memory Bugfix: Fixed a critical bug in VLLM's TPU backend where the KV cache memory allocation did not correctly account for
head_sizepadding. This oversight previously led to out-of-memory (OOM) errors, which are now prevented by ensuring sufficient memory is reserved. - Refactored KV Cache Shape Calculation: The
get_kv_cache_shapefunction invllm/v1/attention/backends/pallas.pywas simplified. It now solely returns the padded head size and no longer attempts to adjustnum_blocksor log warnings, as this logic was insufficient and has been relocated. - Corrected KV Cache Memory Sizing: In
vllm/v1/worker/tpu_worker.py, thedetermine_available_memoryfunction now explicitly calculates thepadded_head_sizeand proportionally reduces thetpu_kv_cache_bytes(usable memory for KV cache) if padding occurs. This ensures sufficient memory is reserved to accommodate the padded head size, preventing OOMs. The warning for head size padding is also now issued from this location.
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
The pull request addresses a bug related to KV cache padding on TPUs. The changes involve modifying the get_kv_cache_shape function in vllm/v1/attention/backends/pallas.py and adjusting the usable memory size calculation in vllm/v1/worker/tpu_worker.py to account for padded head sizes. The changes look correct and necessary for the proper functioning of the code on TPUs.
|
@yaochengji , the pr description mentioned "we still need to modify the value in KVCacheConfig as well.", but I don't see it's changed in the pr. Do you mean modify the |
Yes. |
Signed-off-by: Chengji Yao <chengjiyao@google.com>
Signed-off-by: Chengji Yao <chengjiyao@google.com>
Signed-off-by: Chengji Yao <chengjiyao@google.com>
Signed-off-by: Chengji Yao <chengjiyao@google.com> Signed-off-by: Will Eaton <weaton@redhat.com>
Signed-off-by: Chengji Yao <chengjiyao@google.com>
Signed-off-by: Chengji Yao <chengjiyao@google.com>
Signed-off-by: Chengji Yao <chengjiyao@google.com> Signed-off-by: avigny <47987522+avigny@users.noreply.github.com>
Signed-off-by: Chengji Yao <chengjiyao@google.com>
Purpose
Fix the bug when padding head_size for TPU. Previously block_num is only changed by
get_kv_cache_shape. It's not enough since we still need to modify the value inKVCacheConfigas well.Test Plan
Test Result
Passed. (It will crash without the patch)