fix(quantization): nvfp4_quantize(backend='cuda') silently corrupts scale factors when global_scale is not float32 - #3497

fix(quantization): nvfp4_quantize(backend='cuda') silently corrupts scale factors when global_scale is not float32#3497

bkryu merged 2 commits into

flashinfer-ai:mainflashinfer-ai/flashinfer:mainfrom

bkryu:nvfp4_quant_cuda_fixbkryu/flashinfer:nvfp4_quant_cuda_fixCopy head branch name to clipboard