llama.cpp
Optimized HIP Q2_0 dot-product path for gfx1201 by replacing HIP's byte permutation with native permutation instructions.
#26753
Merged
Merged
Optimized HIP Q2_0 dot-product path for gfx1201 by replacing HIP's byte permutation with native permutation instructions. #26753
ggerganov
merged
7 commits into
ggml-org:master
from
LunalFresh:hip-gfx1201-q2-perm
thelittlefireman
commented
on 2026-08-08
hip/gfx1201: optimize q2_0 vec_dot_q2_0_q8_1 with native amdgcn perm
c43d584b
Broadened HIP's Q2_0 perm optimization
857d3378
Remove redundant HIP perm availability guard
ae5a832a
Optimize HIP Q2_0 MMQ unpack with native perm
8963530a
cuda: label HIP preprocessor guard
48338513
cuda: label HIP preprocessor guard
480d7b37
Restore MMQ tile index handling
f66bcd34
IMbackK
approved these changes
on 2026-08-29
Reviewers
Assignees
Labels
ggml
merge ready
CUDA
Milestone
No milestone
Login to write a write a comment.
Login via GitHub