Repository navigation
perf(kda): parallelize diagonal solves in Prefill - #1703
Merged
Merged
Conversation
shihaobai
marked this pull request as ready for review
November 11, 2026 14:04
shihaobai
force-pushed
the
bsh/glm-kda-split-triangular
branch
from
November 11, 2026 14:09
dae6c2d to
8c815ee
Compare
shihaobai
changed the title
Solve the four diagonal blocks independently and retain FP32 inverse tiles through merging. Select by chunk occupancy to bound scratch, with CUDA regression coverage for packed layouts, dtypes, precision modes and fallback behavior.
shihaobai
force-pushed
the
bsh/glm-kda-split-triangular
branch
from
November 11, 2026 15:24
4d70ee5 to
87deeb0
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Compute the four independent 16x16 diagonal inverses in parallel before merging each 64x64 inverse. KDA Prefill opts into the shared solver's 64-token branch, which supports both packed and dense batch layouts. Select splitting by the FP32 workspace budget and minimum input length; the workspace may use at most the FP32 input size.
H200, 8192 tokens (16 requests of 512 tokens), 64 heads: the solver takes 0.968 → 0.476 ms with 32 MiB additional workspace. A 64-token, 64-head solve takes 0.0347 → 0.0254 ms.
Validation:
--report-api-errors no, CUDA memory caching disabled).