Companion examples for the PR that routes the MTP drafter's mhc_post through the
torch reference on sm_121. These are the exact scripts/commands used to reproduce
the crash, prove the fix without touching the plugin, measure decode rate, and run a
long-context recall test.
Environment (matched to requirements-serve.lock.txt): vLLM 0.23.0, torch
2.11.0+cu130, tilelang 0.1.9, apache-tvm-ffi 0.1.9; NVIDIA GB10 (Grace
Blackwell, sm_121), 128 GB unified memory; CUDA 13.0 / driver 580.x; DGX-OS