llama.cpp
CUDA/HIP: matrix-core (MFMA) lightning indexer for CDNA2 (gfx90a)
#29050

Merged

CUDA/HIP: matrix-core (MFMA) lightning indexer for CDNA2 (gfx90a) #29050

JohannesGaessler merged 8 commits into ggml-org:master from jadenmach2:rocm/cdna2-mfma-indexer

jadenmach2 requested a review 51 days ago

github-actions added ggml

github-actions added CUDA

IMbackK assigned

IMbackK 51 days ago

IMbackK force pushed 39 days ago

JohannesGaessler commented on 2026-10-06

jadenmach2 force pushed to c305ad49 32 days ago

jadenmach2 force pushed from c305ad49 to f4daef53 30 days ago

CUDA/HIP: matrix-core (MFMA) lightning indexer for CDNA2 (gfx90a)

b4c56870

Addressing reviewers comments, mma.cuh and physical warp size

4ebd6f00

CUDA/HIP: single mma.cuh matrix-core lightning indexer ::CDNA::

3a11d5b7

Support RDNA3

925561e8

CUDA/HIP: stream heads through the mma lightning indexer kernel

861aaa28

enable for CUDA

bcd5ac99

CUDA: Remove wmma lightning-indexer, also run head 32 through mma kernel

90ed669d

IMbackK force pushed from 45f60312 to 90ed669d 13 days ago

CUDA: Nv mma dosent transpose accumulator

2f4899d6

IMbackK approved these changes on 2026-10-27

JohannesGaessler approved these changes on 2026-11-05

JohannesGaessler merged 15e50990 into master 2 days ago

Reviewers

JohannesGaessler

IMbackK

Assignees

IMbackK

Labels

ggml CUDA

Milestone

No milestone