llama.cpp merges native Blackwell NVFP4 support

This is a meaningful infrastructure step, not just another quantization tweak: llama.cpp is moving from "can load the format" toward actually exploiting Blackwell silicon the way NVIDIA intended. It’s still preliminary, but it should matter immediately to anyone chasing better throughput-per-watt on local or semi-local rigs.

• –The merge targets SM120 Blackwell GPUs, so the win is tied to newer NVIDIA hardware rather than a broad across-the-board speedup
• –Native NVFP4 support lowers the gap between model packaging and kernel support, which is why GGUF variants are already surfacing so quickly
• –For local inference users, this strengthens llama.cpp’s position as the first stop for bleeding-edge quant formats and vendor-specific hardware features
• –The "preliminary" label matters: expect rough edges, model-by-model quirks, and a period of rapid follow-up fixes
• –This is especially relevant for MoE and larger models where memory bandwidth and quantized math are the bottlenecks