llama.cpp
HIP: use hipBLAS for dense prefill on gfx900, keep MMQ for MoE
#24588

Merged

HIP: use hipBLAS for dense prefill on gfx900, keep MMQ for MoE #24588

JohannesGaessler merged 4 commits into ggml-org:master from DEV-DUFORD:gfx900-mmq

HIP: keep MMQ for gfx900 MoE and Q8_0, use hipBLAS for dense K-quants

25c55df1

HIP: tighten conditional block to be explicitly for gfx900

717920bf

HIP: Further simplified gfx900 conditional block

76fd3124

DEV-DUFORD requested a review 60 days ago

github-actions added Nvidia GPU

github-actions added ggml

sanmai commented on 2026-06-15

IMbackK requested changes on 2026-06-15

removed unnecessary comment

5c54430b

DEV-DUFORD requested a review from

IMbackK 58 days ago

github-actions added CUDA

IMbackK approved these changes on 2026-06-25

JohannesGaessler approved these changes on 2026-06-30

JohannesGaessler merged d9df1100 into master 43 days ago

Reviewers

JohannesGaessler

IMbackK

sanmai

Assignees

No one assigned

Labels

Nvidia GPU ggml CUDA

Milestone

No milestone