Repository navigation
Conversation
shihaobai
force-pushed
the
bsh/glm-mtp-fused-norm
branch
from
November 12, 2026 11:12
05e3762 to
c2e7e21
Compare
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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 the three GLM MTP input RMSNorms and concatenation before eh_proj into one Triton kernel. Preserve main-norm intermediate rounding and both in-place input updates. Use the fused path for 4096-wide BF16/FP16 CUDA inputs with the default norm warp count; retain the existing path otherwise.
Validated on H100 and retested after rebasing onto main db192ec: 46 GPU tests pass, covering rounding, strided buffers, CUDA graph replay and fallback. All 18 actual-checkpoint projection comparisons are byte-exact (1–8192 tokens). Norm+concat is 2.17–2.38x faster under CUDA graph at 1–64 tokens; including eh_proj, the gain is 1.18–1.19x. The 8192-token eager consumer is 1.12x faster. No end-to-end TPM claim.