Repository navigation
perf(moe): fuse GLM prefill SwiGLU and UE8M0 quantization - #1709
Merged
Merged
Conversation
…tion Drop gemm_out_a as soon as each activation path finishes reading it, and spell out the FP8 autotune key instead of unpacking a conditional dict. Co-authored-by: Cursor <cursoragent@cursor.com>
Note that the fused SwiGLU path is only for the Hopper GLM expert shape, and only for chunks of 896 to 32768 rows. Co-authored-by: Cursor <cursoragent@cursor.com>
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Fuse SwiGLU and FP8 quantization in the Hopper GLM expanded-MoE prefill path, preserving main's UE8M0 scales and intermediate rounding. Unsupported shapes, UE8M0-off execution and autotuner warmup retain the existing path.
Validation against main dc53088 on H100: 238 GPU tests pass; seven real-expert consumer cases are byte-exact, including a 32K chunk plus a fallback tail. With main's default kernel configs, the expert consumer is 1.22–1.26x faster at 896–32768 expanded rows (W13, activation/quantization, W2 and gather; excludes DeepEP communication).