]> git.djapps.eu Git - pkg/ggml/sources/ggml/commit
HIP: Disable ROCWMMA fattn on CDNA when compiled against ROCWMMA 2.0.0 (llama/16221)
authoruvos <redacted>
Wed, 1 Oct 2025 21:09:25 +0000 (23:09 +0200)
committerGeorgi Gerganov <redacted>
Sun, 12 Oct 2025 04:57:25 +0000 (07:57 +0300)
commite82b073791dc363bac527226709c41990b6568ab
tree6281f12134bd85c5e484211c018f181060e644a5
parent2d972e8ea43b41dbfd4d7ff6aa429c74ee0f041e
HIP: Disable ROCWMMA fattn on CDNA when compiled against ROCWMMA 2.0.0 (llama/16221)

* HIP: Disable ROCWMMA fatt on CDNA when compiled against ROCWMMA 2.0.0

rocwmma 2.0.0 includes a bug in the code fakeing fp16 accumulation on CDNA

* CUDA: Fix volta condition in ggml_cuda_should_use_wmma_fattn
CMakeLists.txt
src/ggml-cuda/common.cuh
src/ggml-cuda/fattn-tile.cu
src/ggml-cuda/fattn-wmma-f16.cu
src/ggml-cuda/fattn-wmma-f16.cuh
src/ggml-cuda/fattn.cu
src/ggml-cuda/vendors/hip.h
src/ggml-hip/CMakeLists.txt