Hand-written ROCm kernels push Qwen3.6 2.16x past vLLM on AMD MI350X
Netra bet against the NVIDIA default and published raw AMDGCN kernels for MI350X. Eight GPUs serve Qwen3.6-35B-A3B-FP8 at 81,331 output tokens/s (78,498 mean), 2.16x vLLM and 3.43x base SGLang. The honest takeaway: once the kernels got fast, the bottlenecks moved to scheduling, routing, and HTTP serialization, not compute.