How SGLang decides the weight format Four rows, each a checkpoint and a quantization flag. An FP8 checkpoint with no flag runs as fp8. A modelopt checkpoint with modelopt_fp4 is a known compatible pair and runs. A GPTQ checkpoint with no flag is converted at runtime and runs on the Marlin kernel. A GPTQ checkpoint with the flag gptq on a CUDA build skips the conversion and fails at startup with Unknown quantization method. CHECKPOINT QUANT_METHOD --QUANTIZATION RESULT AT STARTUP fp8 weights on disk not set your command line fp8 read from the checkpoint modelopt weights on disk modelopt_fp4 your command line modelopt_fp4 known compatible pair gptq weights on disk not set your command line gptq_marlin converted at runtime gptq weights on disk gptq your command line startup error Unknown quantization method