What’s the state of native NVFP4 kernel support on GB10? AFAIK vLLM still falls back to weight-only FP4 via Marlin.
Is there an official commitment and timeline for first-class native NVFP4 support for GB10/SM121, or is it just “we’re committed to the best experience”?
Depends on what you mean. W4A4 is basically non existent. W4a16 is better now. B12x was recently merged into flashinfer so that’s a viable path.
If you want better support you’ll usually have to rely on non official forks and images
This is incorrect. Far from non-existent, NVFP4 native kernels exist and are supported in current Flashinfer under vLLM.
As of yet they do not outpace Marlin Int4, but the gap has been substantially narrowed to the point that they are certainly usable, and further optimizations like b12x are experimental or in flight.