I created the ROCmFPX quant family for AMD Hardware. It runs certain LLM models faster on AMD Hardware. ROCmFP4 is smaller than normal 4 bit quants but cosplays as a 5 bit quant.