Skip to content

Conversation

@jeffbolznv
Copy link
Collaborator

Stacked on #17443.

This FA size came from #17012, and was getting a bunch of register spilling. Perf on RTX 4070 with coopmat1/2 disabled.:

before

FLASH_ATTN_EXT(hsk=72,hsv=72,nh=16,nr23=[1,1],kv=5776,nb=5776,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]):                   6 runs - 167320.33 us/run - 153.73 GFLOP/run - 918.79 GFLOPS

after

FLASH_ATTN_EXT(hsk=72,hsv=72,nh=16,nr23=[1,1],kv=5776,nb=5776,mask=0,sinks=0,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]):                  34 runs - 30001.44 us/run - 153.73 GFLOP/run -   5.12 TFLOPS

@github-actions github-actions bot added testing Everything test related Vulkan Issues specific to the Vulkan backend ggml changes relating to the ggml tensor library for machine learning labels Nov 23, 2025
@0cc4m 0cc4m merged commit d414db0 into ggml-org:master Nov 25, 2025
71 of 74 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

ggml changes relating to the ggml tensor library for machine learning testing Everything test related Vulkan Issues specific to the Vulkan backend

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants