Kernels · decode path

Kernel fusion on the V4.1 decode path

One decode layer of DeepSeek-V4.1-Flash (an indexer layer), simplified: one block = one kernel launch; communication and aux streams omitted; widths illustrative.

Kernel names follow vLLM main (Oct 2026); PR numbers in the step text.

kernel fused kernel changed in this step side CUDA stream baseline length