Motivation

DeepSeek-V3-0324 gsm8k accuracy: 0.951

After this upgrade, the compilation speed of both NVCC and NVRTC is truly impressive! First, NVRTC takes about 1s per kernel to compile, while NVCC's compilation time has improved from the previous 4s per kernel to 1.2s~1.3s per kernel. Finally, we no longer have to endure long waits for precompilation!

Refer to deepseek-ai/DeepGEMM@d75b218
NVRTC may have performance loss with some cases and NVCC JIT speed is also 4x faster now. So I keep using NVCC here.

@zhyncs Dependency pipeline:

1. merge chore: upgrade deepgemm #6073
2. release new version of sgl-kernel and upload wheel to pip
3. update srt's sgl-kernel tag
4. merge this PR

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.