[test] add ut and bm for get_last_loc

[Alcanderian]

Motivation

#6676 (review)

performance no diff compared with torch compile

Correctness test passed!
get-last-loc-performance:
    batch_size  PyTorch  Triton
0          1.0    5.088   4.832
1          2.0    4.960   5.056
2          4.0    5.088   5.088
3          8.0    5.216   5.056
4         16.0    5.280   5.216
5         32.0    5.376   5.216
6         64.0    5.440   5.280
7        128.0    5.408   5.408
8        256.0    5.472   5.472
9        512.0    5.568   5.600
10      1024.0    5.696   5.568

Modifications

Checklist

• Format your code according to the Code Formatting with Pre-Commit.
• Add unit tests as outlined in the Running Unit Tests.
• Update documentation / docstrings / example tutorials as needed, according to Writing Documentation.
• Provide throughput / latency benchmark results and accuracy evaluation results as needed, according to Benchmark and Profiling and Accuracy Results.
• For reviewers: If you haven't made any contributions to this PR and are only assisting with merging the main branch, please remove yourself as a co-author when merging the PR.
• Please feel free to join our Slack channel at https://slack.sglang.ai to discuss your PR.

[gemini-code-assist]

Hello @Alcanderian, 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!

Summary of Changes

Hello team, gemini-code-assist here to provide a summary of this pull request. This PR, authored by @Alcanderian, focuses on adding comprehensive testing and performance benchmarking for a function related to retrieving the last location (get_last_loc). The primary motivation, as indicated by the title and description, is to ensure the correctness and evaluate the performance of a Triton-based implementation of this function compared to a PyTorch reference. The changes introduce a new file containing both a unit test to verify functional correctness and a benchmark script to measure execution time across various batch sizes.

Highlights

• Add Unit Test: A unit test (test_get_last_loc) is added to compare the output of the Triton implementation (get_last_loc_triton) against a PyTorch reference implementation (get_last_loc_torch) using torch.testing.assert_close.
• Add Performance Benchmark: A performance benchmark (get_benchmark) is introduced using triton.testing.perf_report to compare the latency of the Triton and PyTorch implementations of get_last_loc across a range of batch sizes (1 to 1024).
• New Benchmark Script: A new Python script (benchmark/kernels/scheduler_batch/benchmark_get_last_loc_triton.py) is added to house the unit test and performance benchmark for the get_last_loc function/kernel.

Changelog

• benchmark/kernels/scheduler_batch/benchmark_get_last_loc_triton.py
  • Added a PyTorch reference implementation for get_last_loc.
  • Added a Triton kernel (get_last_loc_kernel) and a Python wrapper (get_last_loc_triton) for the get_last_loc logic.
  • Implemented a unit test (test_get_last_loc) to verify the correctness of the Triton implementation.
  • Implemented a performance benchmark (get_benchmark) comparing PyTorch and Triton get_last_loc implementations.
  • Added a main execution block (if __name__ == "__main__") to run the test and benchmark.

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.

---

Code arrives, a new test,
Benchmark runs, put to the test.
PyTorch, Triton race,
Performance finds its place,
Correctness puts minds at rest.

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 PR introduces valuable unit tests and benchmarks for the get_last_loc function, comparing PyTorch and Triton implementations. The implementations appear correct, and the testing setup is solid. The benchmark results provided in the PR description are also informative.

I found one medium-severity issue related to a default path that should be addressed. Overall, good work on adding these tests and benchmarks!

Summary of Findings

• Incorrect Default Save Path: The run_benchmark function has a default save_path argument that seems to be for a different benchmark (write_req_to_token_pool instead of get_last_loc). This could lead to results being saved in an unexpected or incorrect location.
• Magic Numbers: The test and benchmark functions use several magic numbers for max_batch, max_context_len, etc. While acceptable for tests, consider defining them as constants if they are reused or if their meaning isn't immediately obvious. (Not commented inline due to review settings)
• Docstrings: Most functions lack docstrings. Adding them would improve code clarity and maintainability for future contributors. (Not commented inline due to review settings)

Merge Readiness

The PR is in good shape, but the medium-severity issue regarding the default save_path in run_benchmark should be addressed before merging to ensure clarity and prevent potential issues with benchmark result storage. After that, I believe this PR will be ready for merge. As an AI, I am not authorized to approve pull requests; please ensure further review and approval from other maintainers.