llama.cpp
vulkan : copy strided f16 KV once on AMD so reads spread across memory channels
#27703

Merged

vulkan : copy strided f16 KV once on AMD so reads spread across memory channels #27703

ggerganov merged 8 commits into ggml-org:master from Nathanw1014:vulkan-fa-f16-kv-contig-v2

Nathanw1014 requested a review from

ggerganov 81 days ago

Nathanw1014 requested a review 81 days ago

github-actions added testing

github-actions added Vulkan

github-actions added ggml

0cc4m approved these changes on 2026-08-26

Nathanw1014 force pushed from 4d9e62fa to 1781f451 54 days ago

0cc4m commented on 2026-09-22

vulkan : copy strided f16 KV once so reads spread across memory channels

24c1920b

tests : cover the f16 KV copy to scratch

5a718c4a

vulkan : gate the f16 KV scratch copy to AMD

f4a08880

vulkan : reuse the generic copy shader for the f16 KV scratch fill

22b50343

vulkan : extend the strided KV scratch copy to bf16

18c8eae6

vulkan : clean up comments

a5141ac1

consolidate copy shaders

60f13f77

0cc4m force pushed from 1781f451 to 60f13f77 31 days ago

disable on RDNA4

3fdbb713

0cc4m approved these changes on 2026-10-14

0cc4m added merge ready

ggerganov merged 733f57a9 into master 30 days ago

Reviewers

0cc4m

jeffbolznv

ggerganov

Assignees

No one assigned

Labels

testing Vulkan ggml merge ready

Milestone

No milestone