A high-performance inference engine for AI models
Commits on main
Faster pre-M5 prefill GEMM: f32 dequant staged as half, batch-only tile rule (#1061)
ea52b58 CCYeh committed 17 days ago
fix(hanashi): preserve hyphens in Qwen tool names (#1062)
3fc1548 xianxian.zhang committed 17 days ago
fix(download-manager): preserve partial download file and progress on error (#1051)
dbe194c onenewcode committed 18 days ago
Add MoE router, expert GEMV and combine kernels (#1000)
798803b Shahzeb Jadoon committed 19 days ago