Skip to content

MiniMax-H3 / ggml-cuda: cuDNN fused attention on Ampere and newer (B200 max render 15.4 to 11.3 s) - #3

Open
oobabooga wants to merge 9 commits into
perf/h3-gguf-speed-v3from
perf/h3-cudnn-attn
Open

oobabooga wants to merge 9 commits into
perf/h3-gguf-speed-v3from
perf/h3-cudnn-attn

Enhance your code review process with GitHub Actions

GitHub Actions make it easy to automate all your software workflows, now with world-class CI/CD.
Build, test, and deploy your code right from GitHub. Learn more about GitHub Actions.

Linux, macOS, Windows, and containers
Linux, macOS, Windows, and containers
Matrix builds
Matrix builds
Any language
Any language
Live logs
Live logs
Built-in secret store
Built-in secret store
Multi-container testing
Multi-container testing