Add quantize/dequantize for mxfp8 and nvfp4#2688

Merged

awni merged 11 commits intomainfrom

mxfp8_and_nvfp4

Oct 28, 2025

Commits on Oct 27, 2025

Add quantize/dequantize slow path for mxfp8 and nvfp4
Awni Hannun
committed
fast cuda kernel for mx/nv quantization
Awni Hannun
committed
fallback for cuda < 12.8 (#2697)
Awni Hannun
committed
format (#2700)
Awni Hannun
committed
fix (#2701)
Awni Hannun
committed
metal kernels
Awni Hannun
committed
docs
Awni Hannun
committed
fix jit
Awni Hannun
committed

Commits on Oct 28, 2025

add default bits and group sizes
Awni Hannun
committed
improve quant docs
Awni Hannun
committed
fix output type of mxfp4 matmuls
awni
committed