Skip to content

CUDA: make DeepSeek-V4-Pro correct on the indexed-attention path (top_k 512→1024) + enable decode LUT gate for in_dim>4096#478

Open
slackarea wants to merge 2 commits into
antirez:mainfrom
vcnngr:pro-cuda-fixes
Open

CUDA: make DeepSeek-V4-Pro correct on the indexed-attention path (top_k 512→1024) + enable decode LUT gate for in_dim>4096#478
slackarea wants to merge 2 commits into
antirez:mainfrom
vcnngr:pro-cuda-fixes

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