PR #25494 vulkan : dequant q8_0 KV once in coopmat1

vulkan : dequant q8_0 KV once in coopmat1 #25494

0cc4m merged 9 commits into ggml-org:master from Nathanw1014:vulkan-coopmat1-fa-dequant-transpose

vulkan : dequant q8_0 KV once in coopmat1

484ad9ba

Nathanw1014 requested a review 95 days ago

github-actions added Vulkan

github-actions added ggml

vulkan : fall back instead of aborting when FA scratch exceeds maxSto…

b8058343

vulkan : require KV-cache layout in FA dequant path

d7f6bce0

vulkan : skip FA dequant path on coopmat2

643e95e7

tests : add contiguously-allocated quant K/V FA tests

c3b3bb98

Nathanw1014 requested a review from

ggerganov 83 days ago

github-actions added testing

0cc4m commented on 2026-08-06

vulkan : trim comments

dd39addb

jeffbolznv commented on 2026-08-09

vulkan : tighten permutation checks for FA path

f9f99bf1

vulkan : set prealloc_x_need_sync after the FA dispatch

500c8876

0cc4m requested changes on 2026-08-15

Nathanw1014 requested a review from

0cc4m 57 days ago

rillomas commented on 2026-08-18

vulkan : exclude Intel Xe1 from FA dequant path

b890946f

Nathanw1014 force pushed from 59443262 to b890946f 54 days ago

0cc4m approved these changes on 2026-08-19

jeffbolznv approved these changes on 2026-08-19

0cc4m merged dc72703f into master 54 days ago

Reviewers

jeffbolznv

0cc4m

rillomas

ggerganov

Assignees

No one assigned

Labels

testing Vulkan ggml

Milestone

No milestone

llama.cpp vulkan : dequant q8_0 KV once in coopmat1 #25494 Merged

vulkan : dequant q8_0 KV once in coopmat1 #25494

llama.cpp
vulkan : dequant q8_0 KV once in coopmat1
#25494

Merged