6 months since intro of NVFP4, and it's basically still a myth
#4
by
zenmagnets
- opened
NVFP4 still just a technical paper. Shame that Nvidia hasn't gotten around to make sure their own GPUs can actually support it. Makes me want to sell stock. Am I wrong?
@zenmagnets Agreed, recent models are just not getting official NVFP4s (like qwen NEXT). Seems its also suboptimal inference on vLLM on DGX Spark which has the specialised blackwell kernels. Ive heard its a software problem that is being resolved...but theres no way to monitor progress of this so it feels like a myth.