Fix prefill latency performance drop of bench serving

[gaopengff]

Motivation

When use prompt as input for bench serving, the server will get incorrect input length as its encode operation added special tokens. This may cause bad performance of prefill latency. We should truncate it.
Old:

input_ids(1024) -> client decode -> prompt -> server  encode -> input_ids(1025)

New:

input_ids(1024) -> truncate -> input_ids(1023) -> client decode -> prompt -> server encode -> input_ids(1024)

[gemini-code-assist]

Summary of Changes

Hello @gaopengff, 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 resolves a performance regression observed in the prefill latency of benchmark serving. The root cause was an discrepancy in input length calculation: when prompts were used, the server's tokenizer would add special tokens, increasing the effective input length beyond what was intended. This change implements a crucial truncation step that adjusts the input_lens by accounting for these special tokens, thereby ensuring accurate input lengths and restoring optimal prefill latency performance.

Highlights

• Performance Fix: Addresses a performance drop in prefill latency during benchmark serving, specifically when prompts are used as input.
• Input Length Correction: Corrects an issue where the server's encoding operation added special tokens, leading to an incorrect and inflated input length.
• Truncation Logic: Introduces a truncation step for input_lens to subtract the number of special tokens that would be added by the tokenizer, ensuring the final encoded input length matches the intended length.

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 by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

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 pull request 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

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

This pull request correctly addresses a performance issue in benchmark serving caused by incorrect input lengths from server-side tokenization. The fix to truncate input lengths based on the number of special tokens is appropriate. I've provided one suggestion to further optimize the implementation by using a vectorized NumPy operation, which will improve performance, especially with a large number of prompts.

[gemini-code-assist]

For better performance and code clarity, you can replace the loop with a vectorized NumPy operation. This is more efficient, especially for a large num_prompts.

Additionally, the int() cast is redundant as tokenizer.num_special_tokens_to_add() already returns an integer.

        num_special_tokens = tokenizer.num_special_tokens_to_add()
        input_lens = np.maximum(0, input_lens - num_special_tokens)

[mingfeima]

@gaopengff explain that vllm benchmark uses the same logic and link the code for it. That will be more straight forward.

[gaopengff]

Vllm uses the same logic at https://github.com/vllm-project/vllm/blob/v0.12.0/vllm/benchmarks/datasets.py#L484, which subtract num_special_tokens to get real_input_len.

[mingfeima]

@zhyncs could you please find someone to review this one? a fix for benchmark scripts~