ggml-cuda: Add NVFP4 dp4a kernel #20644
IMbackK
merged
14 commits into
ggml-org:master
from
michaelw9999:nvfp4-dp4a
JohannesGaessler
requested changes
on 2026-03-22
Forced F32 path for NVFP4/Cublas and removed Fusion/TensorScale
fa79ea63
Removed stale code
53450f12
Renamed k to ne
7fd898be
Added check for dst_t to cuda_cast template for float
caa8fba0
am17an
commented
on 2026-03-24
am17an
commented
on 2026-03-24
am17an
commented
on 2026-03-24
Restored ggml_cuda_ue4m3_to_fp32, changed vecdot ints to int32ts
55acc41c
Simplified ggml_cuda_ue4m3_to_fp32
5a7e19b4
Removed NVFP4-MMQ block checks
82f0e6bb
Added CUDART/HIP Check and HIP/fp8 include
daf439b1
Added NVFP4 to Test-backend-ops
9ab7cf21
Added hip_fp8_e4m3 to __nv_fp8_e4m3 typedef
e30f0b3c
Restored last include to baseline
e728b2a4
Removed whitespace artifacts
8c8f368a
JohannesGaessler
approved these changes
on 2026-03-25
Update ggml/src/ggml-cuda/ggml-cuda.cu
0780545b
am17an
approved these changes
on 2026-03-25
try CI fix
af41687c
am17an
approved these changes
on 2026-03-26
IMbackK
approved these changes
on 2026-03-26
Reviewers
Assignees
No one assigned
Labels
testing
Nvidia GPU
python
ggml
Milestone
No milestone
Login to write a write a comment.
Login via GitHub