HIP: add fattn-mma-f16 for RDNA4 #18481
JohannesGaessler
merged
22 commits into
ggml-org:master
from
fattn_for_rdna4
finish VQ mma
01c68426
flash_attn_ext_f16_iter
97be8006
KQ_rowsum
4a91e14d
correct exp
c0c1cd13
fix scale error
e05eb9ff
fix softmax scale
c12f0b80
fix softmax scale
55ea70cd
enable fattn on cpu side
73abf644
fix random error
4f883935
zhanghuiyulo
changed the title 
disable fattn-mma-f16 on rdna3
65850036
fix wrong col for rdna
90c02e34
use identity mat to transpose
7e09adb7
resolve conflicts
b66845e3
Merge branch 'master' into fattn_for_rdna4
103141ff
JohannesGaessler
commented
on 2026-01-07
basic tuning for DeepSeek-R1-Distill-Qwen-1.5B
41fd92ba
IMbackK
commented
on 2026-01-08
fix volta compile error
c4489fa2
align rdna4 policy for fattn
498c847c
adjust fattn policy
2a666f5b
adjust kernel selection logic
7acbc334
update as the review comments
3d948d29
keep fattn-wmma logic
093c5b7e
adjust kernel selection logic
6de57828
JohannesGaessler
approved these changes
on 2026-01-12
Reviewers
Assignees
No one assigned
Labels
Nvidia GPU
ggml
Milestone
No milestone
Login to write a write a comment.
Login via GitHub