Tag
Achieved DeepSeek-V4-Flash MTP speculative decoding on 2× RTX PRO 6000 with a 38% throughput increase by fixing a mis-routed quantization format issue.