FP4 MoE Kernels on Blackwell Give Identical Bits
RightNow AI says its FP4 MoE kernels on Blackwell give identical logits at batch sizes of 1 and 64 on a 594B-parameter model across 8 B300 GPUs.
RightNow AI says its FP4 MoE kernels on Blackwell give identical logits at batch sizes of 1 and 64 on a 594B-parameter model across 8 B300 GPUs.