transformers

Keep MXFP4 weights quantized on XPU when use_kernels is set

#47923

Merged