PR #16884 CUDA: fuse rope + set_rows

CUDA: fuse rope + set_rows #16884

am17an merged 5 commits into ggml-org:master from am17an:cuda-add-rope-fusion

am17an requested a review from

CISC 291 days ago

am17an requested a review from

slaren 291 days ago

github-actions added Nvidia GPU

github-actions added ggml

am17an commented on 2025-10-31

am17an requested a review from

JohannesGaessler 291 days ago

am17an commented on 2025-10-31

am17an force pushed 291 days ago

ORippler commented on 2025-10-31

JohannesGaessler commented on 2025-11-02

CUDA: add fused rope

acfd03d3

move k forward_expand up

b3761df3

create helper function instead of re-using params

67b6580b

make assert statement more in line with comment

c7c3b9f1

rope_norm: coalesced writes to global mem

67624935

am17an force pushed to 67624935 279 days ago

am17an requested a review from

JohannesGaessler 279 days ago

JohannesGaessler approved these changes on 2025-11-12

am17an merged a90eb94c into master 278 days ago

am17an deleted the cuda-add-rope-fusion branch 278 days ago

Reviewers

JohannesGaessler

ORippler

slaren

ggerganov

CISC

Assignees

No one assigned

Labels

Nvidia GPU ggml

Milestone

No milestone