llama.cpp
CUDA: use MMVF for thin f16/bf16 mul_mat at small batch size
#29633
Merged
Merged
CUDA: use MMVF for thin f16/bf16 mul_mat at small batch size #29633
ynankani
merged
2 commits into
ggml-org:master
from
ynankani:ynankani/cuda_bf16_mmvf_small_batch
JohannesGaessler
requested changes
on 2026-09-29
CUDA: use MMVF for thin f16/bf16 mul_mat at small batch size
bbbb1d0a
JohannesGaessler
commented
on 2026-09-30
adjust kernel selection logic
ebae2624
JohannesGaessler
approved these changes
on 2026-09-30
am17an
approved these changes
on 2026-10-05
Login to write a write a comment.
Login via GitHub