llama.cpp
CUDA: Improve NVFP4 W4A4 activation quantization
#25730

Merged

CUDA: Improve NVFP4 W4A4 activation quantization #25730

ORippler merged 5 commits into ggml-org:master from ORippler:osimons/nvfp4_amax_quant

ORippler requested a review 25 days ago

github-actions added ggml

github-actions added CUDA

Squash history before conflict-resolution during rebase on master

46caff64

ORippler force pushed from 5d70e935 to 46caff64 25 days ago

compiler massaging to avoid unnecessary LDCs

8a6ca53c

am17an approved these changes on 2026-07-21

kvalues_mxfp4 -> kvalues_nvfp4 in quantize_mmq_nvfp4

b1ea6058

Always pass in src1_scale.ptr

598f69b6

Extract ggml_cuda_is_aligned helper

1394d6ee

taronaeo approved these changes on 2026-07-22

ORippler merged 1a064ab0 into master 18 days ago

ORippler deleted the osimons/nvfp4_amax_quant branch 18 days ago

Reviewers

taronaeo

am17an

Assignees

No one assigned

Labels

ggml CUDA

Milestone

No milestone