cuda: increase matmul rows/block from 8 to 16#266
Open
arpesenti wants to merge 1 commit into
Open
Conversation
Doubles rows/block in Q8_0, HC-expand, grouped matmul kernels and MoE gate_up decode kernels, improving GPU occupancy. Benchmark (DGX Spark GB10, q2, 128 gen tokens): 2048 ctx: 14.4 → 18.1 t/s (+25.7%) 4096 ctx: 14.9 → 19.1 t/s (+28.0%) 8192 ctx: 14.7 → 18.7 t/s (+27.5%) 32768 ctx: 13.5 → 16.8 t/s (+24.4%) 65536 ctx: 12.5 → 15.3 t/s (+22.7%) Prefill unchanged. Regression tests pass.
|
Any progress on merging this PR? |
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
3 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
CUDA: increase matmul rows/block from 8 to 16
Doubles the number of output rows processed per block in the warp-level Q8_0, HC-expand, grouped
matmul and MoE gate_up decode kernels, improving GPU occupancy on architectures with high CUDA core
counts. Prefill is unaffected.
Benchmark (DGX Spark GB10, q2, 128 gen tokens)
Prefill unchanged across all frontiers (within ±1 t/s noise).
Changes
tests/cuda_long_context_smoke.c(missingcomp_kv_f16arg)Verification
make cuda-spark— clean buildmake cuda-regression— passed