llama.cpp
b91f7de7
- vulkan: add flipped FA path, each subgroup owns 16 Q rows
Go
Login via GitHub
Home
Pricing
FAQ
Install
Login
via GitHub
Commit
View On
GitHub
Commit
7 days ago
vulkan: add flipped FA path, each subgroup owns 16 Q rows
References
#31323 - vulkan: optimize prompt processing Flash Attention for RDNA3 and RDNA4
Author
0cc4m
Committer
0cc4m
Parents
cdf118a7
Loading