Oct 19 15 days ago

1. 17:55

   W8A8 expert kernel (FP8 x FP8 tensor cores, K-major weights at load): 11.4 ms vs 17.7 bf16 per call; A/B results

   saaivignesh20
   Verified signature

   ab8889c
2. 16:38

   FP8 routed experts for MAGI-2 Preview

   saaivignesh20
   Verified signature

   21fe007