llama.cpp
38355c6c
- CUDA: use registers instead of smem in topk-moe (#16647)
Go
Login via GitHub
Home
Pricing
FAQ
Install
Login
via GitHub
Commit
View On
GitHub
Commit
58 days ago
CUDA: use registers instead of smem in topk-moe (#16647) Uses the technique used in the vulkan PR #16641. Neat trick!
References
#16647 - CUDA: use registers instead of smem in topk-moe
Author
am17an
Parents
81387858
Loading