[TPU] Add TPU specific var VLLM_TPU_MOST_MODEL_LEN

[Chenyaaang]

Purpose

Add TPU specific environment variable: VLLM_TPU_MOST_MODEL_LEN. This is used to pass in the request length for most of requests, comparing with the existing variablemax_model_len for the length of longest requests. This will benefit when most_model_len is much longer than most_model_len, and a large portion of requests fall inside most_model_len, such as 1% requests are 32k for most_model_len and 99% requests are 2k for most_model_len, in the following 2 ways:

1. When max_model_len is very large, attention kernel usually doesn't reach its best performance (because it falls outside the tuned shape), by using most-model-len, we can easily see perf gain
2. When max_model_len and max_num_reqs are both large, it's very likely to cause SMEM OOM, and we have to increase the page size above the pre-set limit 256 (which also harms kernel performance). By introducing most-model-len, we keep the max-num-seqs when the requests' lengths are within most-model-len, and decrease max-num-seqs for the extra long outlier requests.

The change is kept within tpu_model_runner, with no change in scheduler, so it doesn't interfere other systems.

In tpu_model_runner.py, we initialize num_reqs_most_model_len and num_reqs_max_model_len for most_model_len and max_model_len condition respectively based on the max number without SMEM OOM. The scheduler's output is then routed to either max_model_len or most_model_len depending on the request lengths. The requests might be executed in more than one batch, and the final result will be concated together. To avoid recompilation, we pre-compiled the backbone with both max_model_len and most_model_len.

Test

The change is tested and compared to baseline, at commit dac8cc4. We started the server of llama3-8B, with max-model-len = 32k, max-num-reqs = 128, and send benchmark requests with input-len = 1800, output-len=128. We compare the compilation time and throughput with and without using most-model-len = 2048.

	Backbone compilation time (sec)	Total init time (sec)	throughput
w/o most-model-len	76.06	177.76	5.75
w/ most-model-len	120.36	222.49	7.46

The result shows a ~30% perf gain and the increase of compilation time is acceptable.

[github-actions]

👋 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 fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

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 ready label to the PR or enable auto-merge.

🚀

[gemini-code-assist]

Summary of Changes

Hello @Chenyaaang, 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 a key optimization for vLLM's TPU backend by enabling a dual-length model processing strategy. By differentiating between typical and maximum request lengths, the system can achieve higher throughput and more efficient memory utilization, particularly for common workloads. The changes are confined to the TPU-specific components, ensuring targeted improvements without affecting other parts of the system.

Highlights

• New Environment Variable for TPU Optimization: Introduced VLLM_TPU_MOST_MODEL_LEN as a new optional environment variable. This allows users to specify a typical or 'most common' request length for TPU workloads, which is distinct from the max_model_len that handles the longest possible requests. This distinction is crucial for optimizing performance and memory usage.
• Dynamic Request Batching and Execution: Implemented a sophisticated request processing mechanism within tpu_model_runner.py. Requests are now dynamically batched and routed based on their length: shorter requests (within most_model_len) can be processed more efficiently, potentially in larger batches, while longer requests (up to max_model_len) are handled separately. This allows for optimal resource allocation and throughput.
• Enhanced SMEM OOM Prevention: Added a new utility function, PallasAttentionBackend.get_max_num_seqs, to precisely calculate the maximum number of sequences that can be processed without encountering Shared Memory (SMEM) Out-Of-Memory errors. This calculation is now used to determine optimal batch sizes for both most_model_len and max_model_len scenarios, replacing a previous, less granular approach.
• Optimized Model Pre-compilation: Updated the model backbone pre-compilation routines (_precompile_backbone and profile_run) to ensure that the model is compiled for both the max_model_len and, if specified, the most_model_len configurations. This pre-compilation prevents costly runtime recompilations when switching between different request length profiles, contributing to overall performance gains.
• Refactored Input Preparation for TPU: The _prepare_inputs method in tpu_model_runner.py has been significantly refactored to handle the new dual-length strategy. It now intelligently prepares input tensors (like block_tables, query_start_loc, and seq_lens) based on whether the current batch of requests aligns with the most_model_len or max_model_len profile, and can split a single scheduler output into multiple execution batches.

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 configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

The pull request introduces a new TPU-specific environment variable VLLM_TPU_MOST_MODEL_LEN to optimize performance by handling requests with lengths within most_model_len differently from outlier requests exceeding it. The changes are localized within tpu_model_runner.py and envs.py, with no impact on other systems. The code includes modifications to initialize and utilize num_reqs_most_model_len and num_reqs_max_model_len based on the new environment variable, and adjustments to the input preparation and model execution to accommodate the new logic. I have provided some suggestions for improved documentation and comments.

[yaochengji]

Thanks for the contribution! Left a few comments.

[yaochengji]

LGTM, thanks!