Welcome! Type "help" for available commands.
$
Summary
Pinned SGLang recipe serving the stock RadixArk Qwen3.8-27B-NVFP4 checkpoint with a calibrated NVFP4 DFlash2 drafter (D=16) on one NVIDIA DGX Spark (GB10).
Measured median whole-request throughput of 71.50 tok/s across 27 frozen payloads, a 25.19% increase over the prior BF16 DFlash2 recipe at D=8. Includes download, build, serve, and benchmark scripts with explicit rollback paths to D=8, BF16 drafter, or no-spec mode.