Repository navigation
DeepGEMM vLLM performance benchmarks - #1
Open
cleonard530 wants to merge 4 commits into
Open
cleonard530 wants to merge 4 commits into
cleonard530 wants to merge 4 commits into
Conversation
cleonard530
commented
Oct 1, 2026
Comment on lines
+16
to
+19
| fp8_fp4_gemm_nn (128x4096x7168) 0.027455 0.031249 +0.003795 +13.82% | ||
| fp8_fp4_gemm_nt (128x4096x7168) 0.027190 0.029886 +0.002696 +9.92% | ||
| fp8_fp4_gemm_tn (128x4096x7168) 0.028339 0.033678 +0.005339 +18.84% | ||
| fp8_fp4_gemm_tt (128x4096x7168) 0.028041 0.030935 +0.002894 +10.32% |
Owner
Author
There was a problem hiding this comment.
Here are the longer relative runtimes for the fp8_fp4_gemm methods
This was referenced Oct 1, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
DeepGEMM stable-ABI performance benchmarks
This PR adds two scripts comparing DeepGEMM
mainwith itsSTABLE_TORCH_LIBRARYmigration.benchmark_kernels.pymeasures 22 supported GEMM cases on an H200 using 1,000 direct calls. Most
results differed by no more than ±1.5%.
The fourDispatch overhead cut down drastically. Now they kernels are within +/- 1.5%fp8_fp4_gemm_{nn,nt,tn,tt}operations were 10–20% slower with the stable ABI, corresponding to
approximately 2–6 µs per call. This appears to be host-dispatch overhead
becoming visible for these relatively short kernels.
Direct-call results
benchmark_deepgemm_dispatch.pymeasures the four FP8/FP4 operations under a vLLM-like workload: 200 calls
per simulated forward pass, averaged across 20 repetitions.
Eager execution, while CUDA-graph replay eliminatedshowed similar stable-ABI regressions
the difference. There is still a small regression for
p8_fp4_gemm_tnon vllm (about 7%, 0.315596 ms extra per one 'forward' call).Results: NT, NN, TN, and TT.