User tests PrismaQuant NVFP4 models on RTX Pro 4500

A user benchmarks Qwen3.6-27B models quantized with the new PrismaQuant method on an RTX Pro 4500 via Oculink to resolve tool call errors encountered with previous NVFP4 implementations.

The testing compares PrismaSCOUT (~5.31 bits, ~20 GB) and PrismaAURA (~5.5 bits, ~23 GB) against other quantization methods using vLLM 0.24.

PrismaQuant selects the optimal format for each linear layer to maximize model capabilities at specific bit depths, though it currently only supports Blackwell architecture in vLLM.

KLD results indicate PrismaAURA has lower divergence from BF16 (0.0342) compared to PrismaSCOUT (0.055), while maintaining higher precision and larger weight size.