vLLM patch with hand-written SM120 SASS kernels combining 2-bit MoE experts and an FP4 delta cache to recover precision, enabling frontier MoE models to run on consumer Blackwell GPUs.