use fast stream instead of torch.cuda.current_stream in llama 4 shared experts overlap

[b8zhong]

Don't use torch.cuda.current_stream, use get_current_device_stream_fast(), just like #12524

python3 -m sglang.launch_server \
  --model-path=/opt/dlami/nvme/models/Llama-4-Scout-17B-16E-Instruct/ \
  --tp=8 \
  --trust-remote-code \
  --mem-fraction-static=0.7 \
  --context-length=131072 \
  --kv-cache-dtype=fp8_e4m3 \
  --attention-backend=fa3 \
  --model-loader-extra-config '{"enable_multithread_load": true, "num_threads": 8}'

python3 -m sglang.bench_serving --backend sglang --num-prompts 64 --dataset-name random --random-input-len 1024 --random-output-len 1024 --random-range-ratio 1 --max-concurrency=8 --flush-cache

Metric	Before	After	Δ	Gain
Request Throughput (req/s)	0.89	0.89	0.00	0.0%
Input Token Throughput (tok/s)	911.59	911.67	+0.08	+0.01%
Output Token Throughput (tok/s)	911.59	911.67	+0.08	+0.01%
Total Token Throughput (tok/s)	1823.19	1823.35	+0.16	+0.01%
Mean E2E Latency (ms)	8983.28	8983.05	−0.23	−0.0%
Median E2E Latency (ms)	8958.14	8965.86	+7.72	+0.1%
Mean TTFT (ms)	206.48	198.57	−7.91	−3.8%
Median TTFT (ms)	198.58	199.17	+0.59	+0.3%
P99 TTFT (ms)	288.02	214.91	−73.11	−25.4%
Mean ITL (ms)	8.58	8.59	+0.01	+0.1%
P95 ITL (ms)	8.90	8.90	0.00	0.0%
P99 ITL (ms)	9.23	9.18	−0.05	−0.5%

Generally, it doesn't make such a great difference.

root@ip-10-40-0-228:/sgl-workspace/sglang# python3 benchmark/gsm8k/bench_sglang.py --num-shots 8 --num-questions 1319 --parallel 500
/usr/local/lib/python3.12/dist-packages/torch/cuda/__init__.py:63: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
  import pynvml  # type: ignore[import]
100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1319/1319 [00:31<00:00, 42.38it/s]
Accuracy: 0.920
Invalid: 0.000
Latency: 31.533 s
Output throughput: 4312.445 token/s

[gemini-code-assist]

Summary of Changes

Hello @b8zhong, 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 CUDA stream management within the Llama 4 model's shared experts overlap implementation. By switching to a more optimized method for retrieving the current device stream, the change aims to enhance the efficiency of GPU operations and synchronization. Although the general performance metrics show minor differences, the improvement in tail latency for the first token suggests a positive impact on user experience for certain workloads.

Highlights

• Stream Optimization: Replaced self.device_module.current_stream() with get_current_device_stream_fast() for more efficient CUDA stream handling in the Llama 4 model's shared experts overlap logic.
• Performance Impact: While overall throughput metrics remained largely unchanged, the P99 Time To First Token (TTFT) saw a significant improvement of 25.4%, indicating better responsiveness for tail latencies.

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 introduces a performance optimization by replacing torch.cuda.current_stream() with a faster, cached version get_current_device_stream_fast() within the Llama 4 shared experts overlap implementation. This change is consistent with similar optimizations made elsewhere in the codebase and aims to reduce the overhead of CUDA API calls. The implementation appears correct and the change is well-contained. My review did not identify any issues.