PR #28102 CUDA/HIP: Flash Attention tuning (gfx1201)

CUDA/HIP: Flash Attention tuning (gfx1201) #28102

IMbackK merged 4 commits into ggml-org:master from pwilkin:fattn-wmma-rdna4-256

pwilkin requested a review from

ggerganov 18 days ago

pwilkin requested a review 18 days ago

github-actions added testing

github-actions added ggml

github-actions added CUDA

IMbackK assigned

IMbackK 18 days ago

pwilkin force pushed to d68f8769 17 days ago

IMbackK approved these changes on 2026-09-03

JohannesGaessler requested changes on 2026-09-05

JohannesGaessler assigned

JohannesGaessler 13 days ago

JohannesGaessler commented on 2026-09-06

pwilkin force pushed from f6264d8d to a3b28f93 12 days ago

JohannesGaessler commented on 2026-09-06

HIP: enable mma FA for head size 256 on RDNA4, tune configs

47e51920

HIP: prefer whole-tile FA grids over stream-k on AMD WMMA

d2bf8ad7

revise stream_k logic

cbefddc5

revise kernel selection logic

461f7c1e

JohannesGaessler force pushed from 1873e6e3 to 461f7c1e 9 days ago

JohannesGaessler approved these changes on 2026-09-09

IMbackK approved these changes on 2026-09-11

IMbackK merged 16378d93 into master 7 days ago

Reviewers

IMbackK

JohannesGaessler

ggerganov

Assignees

IMbackK

JohannesGaessler

Labels

testing ggml CUDA

Milestone

No milestone