[CPU] Refine default config for the CPU backend #19539
Conversation
Signed-off-by: jiang1.li <[email protected]>
bigPYJ1151
requested review from
WoosukKwon,
alexm-redhat,
comaniac,
njhill,
robertgshaw2-redhat and
ywang96
as code owners
|
👋 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 🚀 |
![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 @bigPYJ1151, 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 the default configuration for the CPU backend. The primary goal is to enable torch.compile with the inductor backend by default for better performance, while simultaneously optimizing CI run times by disabling inductor for most tests. Additionally, it introduces more granular control over default sequence and batched token limits based on the usage context and sets specific defaults for the CPU platform.
Highlights
- CPU Backend Defaults: Enabled
torch.compilewith theinductorbackend by default for the CPU platform in V1 mode, moving away from the previously enforced eager mode. - CI Performance Optimization: Introduced an environment variable (
VLLM_CPU_CI_ENV) to disable theinductorbackend specifically within most CPU CI test runs to reduce compilation time and speed up tests. The benchmark test explicitly re-enables it. - Per-Context Sequence/Token Limits: Refactored the default
max_num_seqssetting to be a dictionary allowing different default values based on theUsageContext(e.g.,LLM_CLASS,OPENAI_API_SERVER). Added specific, lower default values formax_num_seqsandmax_num_batched_tokenstailored for the CPU backend. - Compilation Settings Refinement: Adjusted the global compilation settings context manager used during model warm-up to pass the VLLM config and make the
freezingsetting conditional onmax_autotune.
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 configureGemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩
bigPYJ1151
changed the title
There was a problem hiding this comment.
Code Review
This pull request refines the CPU backend's default PyTorch Compile (Inductor) configuration and introduces more granular control over max_num_seqs based on usage context. The changes aim to enable Inductor by default for CPU for better performance, while using eager mode in most CI tests to save time.
Key changes include enabling Inductor by default for CPU, allowing context-specific defaults for sequence and batch token limits, and updating Inductor settings in the cpu_model_runner.
A critical concern is the modification to torch._inductor.config.freezing handling, which might default to False when True could be required by CPU Inductor backends. Additionally, the removal of force_disable_caches and changes to custom_ops handling warrant verification to prevent regressions or unexpected behavior. The CI script changes also warrant review.
| def _set_global_compilation_settings(config: VllmConfig): | ||
| import torch._inductor.config | ||
|
|
||
| # Note: The CPPGEMM backend requires freezing parameters. | ||
| freezing_value = torch._inductor.config.freezing | ||
| torch._inductor.config.freezing = True | ||
| # Note: workaround for "ValueError: fast mode: can't pickle cyclic objects | ||
| # including object type dict" | ||
| force_disable_caches = torch._inductor.config.force_disable_caches | ||
| torch._inductor.config.force_disable_caches = True | ||
| yield | ||
| torch._inductor.config.freezing = freezing_value | ||
| torch._inductor.config.force_disable_caches = force_disable_caches | ||
| inductor_config = config.compilation_config.inductor_compile_config | ||
| try: | ||
| # Note: The MKLDNN and CPPGEMM backend requires freezing parameters. | ||
| freezing_value = torch._inductor.config.freezing | ||
| if inductor_config.get("max_autotune", False): | ||
| torch._inductor.config.freezing = True | ||
| yield | ||
| finally: | ||
| torch._inductor.config.freezing = freezing_value |
There was a problem hiding this comment.
The logic for setting torch._inductor.config.freezing has changed. It's now conditional on inductor_config.get("max_autotune", False). Verify if freezing=False is safe and intended for the default CPU Inductor path, or if freezing=True should be set by default for CPU, independently of max_autotune if required by the underlying Inductor backends.
| if os.environ.get("VLLM_CPU_CI_ENV", "0") != "0": | ||
| backend = "eager" | ||
| else: | ||
| backend = "inductor" |
There was a problem hiding this comment.
The line compilation_config.custom_ops += ["none"] was removed. Clarify the expected behavior of custom ops with this change, for both "eager" and "inductor" backends under CompilationLevel.DYNAMO_ONCE. Ensure custom ops are correctly handled.
| python3 -m vllm.entrypoints.openai.api_server --model facebook/opt-125m --dtype half & | ||
| timeout 600 bash -c 'until curl localhost:8000/v1/models; do sleep 1; done' || exit 1 | ||
| python3 benchmarks/benchmark_serving.py \ | ||
| VLLM_CPU_CI_ENV=0 python3 benchmarks/benchmark_serving.py \ |
There was a problem hiding this comment.
Setting VLLM_CPU_CI_ENV=0 here seems to override the environment variable set in the docker run command. Is this intentional?
| VLLM_CPU_CI_ENV=0 python3 benchmarks/benchmark_serving.py \ | |
| python3 benchmarks/benchmark_serving.py \ | |
| --backend vllm \ |
|
Hi @Isotr0py @DarkLight1337 Would you please help to take a review on this PR? Some minor changes of default CPU backend configs, the cpu-test has passed. Thanks :) |
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
3 participants
Essential Elements of an Effective PR Description Checklist
supported_models.mdandexamplesfor a new model.Purpose
default_max_num_seqsfor differentUsageContext. Settingdefault_max_num_seqsanddefault_max_num_batched_tokensfor the CPU backend.Test Plan
CI pipeline
Test Result
(Optional) Documentation Update