llama.cpp
CUDA: experimental native mxfp4 support for blackwell
#17906

Merged

CUDA: experimental native mxfp4 support for blackwell #17906

am17an merged 15 commits into ggml-org:master from am17an:mxfp4

am17an requested a review from

JohannesGaessler 19 days ago

github-actions added Nvidia GPU

github-actions added ggml

am17an marked this pull request as draft 19 days ago

CISC commented on 2025-12-10

am17an force pushed from 16e8a11d to 9dde464c 19 days ago

JohannesGaessler commented on 2025-12-10

JohannesGaessler commented on 2025-12-11

am17an force pushed from b978da31 to a1672f62 18 days ago

am17an force pushed from 870c9d72 to 61c41a0d 17 days ago

mediouni-m commented on 2025-12-14

am17an force pushed from 8acc50ba to 83f62fb9 12 days ago

am17an requested a review from

JohannesGaessler 12 days ago

JohannesGaessler commented on 2025-12-17

am17an force pushed from f5e46e55 to 40b7df86 12 days ago

am17an force pushed from e4c2ac39 to 43c262b9 12 days ago

am17an force pushed from 43c262b9 to bdde3284 12 days ago

am17an force pushed from bdde3284 to 5dfa63da 11 days ago

am17an marked this pull request as ready for review 11 days ago

am17an requested a review from

JohannesGaessler 10 days ago

ORippler commented on 2025-12-19

CUDA: experimental native mxfp4 support for blackwell

4bc93a6c

optimize load_tiles

38648a4f

optimize quantize_mxfp4

f8d20c59

cleanup

ae617ac0

first pass review: formatting

b0e3c620

use interleaved layout for mma

9ebe043c

mmq: add assert for size

d7edade3

use __nv_fp4x4_e2m1

97460714

use iter_k as 512, cleanup

d6cb832b

Use 1200 as blackwell instead of 1000

3fca15c2

address review comments

41329d35

mmq: fix stride

b3919910

am17an force pushed from 5dfa63da to c5799d69 9 days ago

am17an force pushed from c5799d69 to 092f8612 9 days ago

am17an changed the title ~~CUDA: experimental native mxfp4 support for blackwell [WIP]~~ CUDA: experimental native mxfp4 support for blackwell 8 days ago

quantize.cu: use reference impl of e8m0 scale

404054c1

am17an force pushed from 092f8612 to 404054c1 8 days ago

JohannesGaessler commented on 2025-12-22

address review comments

5d3780d4

JohannesGaessler approved these changes on 2025-12-24

JohannesGaessler commented on 2025-12-24

add 120f-virtual + minor fixes

dc04da57

am17an merged c8a2417d into master 5 days ago

am17an deleted the mxfp4 branch 5 days ago

Reviewers

JohannesGaessler

ggerganov

CISC

ORippler

woachk

mediouni-m

Assignees

No one assigned

Labels

Nvidia GPU ggml

Milestone

No milestone

llama.cpp CUDA: experimental native mxfp4 support for blackwell #17906 Merged

CUDA: experimental native mxfp4 support for blackwell #17906

llama.cpp
CUDA: experimental native mxfp4 support for blackwell
#17906

Merged