[GH-ISSUE #18] Precompile all kernels (no jit) #15

New issue

Closed

opened 2026-10-26 21:10:31 +00:00 by bsakaguchi · 2 comments

bsakaguchi commented

2026-10-26 21:10:31 +00:00

Owner

Originally created by @YLGH on GitHub (Feb 26, 2025).
Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/18

Hi, thanks for the amazing library!

Just wondering, is there a way to disable jit by precompiling all the necessary shapes? Or is it recommended to just some "warmup" script to compile all the kernels? What's a good list of M,N,K values to do so?

Originally created by @YLGH on GitHub (Feb 26, 2025). Original GitHub issue: https://github.com/deepseek-ai/DeepGEMM/issues/18 Hi, thanks for the amazing library! Just wondering, is there a way to disable jit by precompiling all the necessary shapes? Or is it recommended to just some "warmup" script to compile all the kernels? What's a good list of M,N,K values to do so?

bsakaguchi closed this issue

2026-10-26 21:10:31 +00:00

bsakaguchi commented

2026-10-26 21:10:32 +00:00

Author

Owner

@jason-huang03 commented on GitHub (Feb 26, 2025):

Not possible, the N and K are template arguments.

@jason-huang03 commented on GitHub (Feb 26, 2025): Not possible, the N and K are template arguments.

bsakaguchi commented

2026-10-26 21:10:32 +00:00

Author

Owner

@LyricZhao commented on GitHub (Feb 27, 2025):

For your own model (if the shapes are fixed), you can simply modify the test scripts to test all the shapes your model used. And later for model training, make sure that you are using the same cache dir (DG_CACHE_DIR).

@LyricZhao commented on GitHub (Feb 27, 2025): For your own model (if the shapes are fixed), you can simply modify the test scripts to test all the shapes your model used. And later for model training, make sure that you are using the same cache dir (`DG_CACHE_DIR`).

bsakaguchi referenced this issue

2026-10-26 21:18:16 +00:00

[PR #469] [CLOSED] Fix cross-expert output overwrites in SM90 masked GEMM #472