diff --git a/auto_round/schemes.py b/auto_round/schemes.py index afbd35053..0ca7710e0 100644 --- a/auto_round/schemes.py +++ b/auto_round/schemes.py @@ -191,6 +191,18 @@ def is_preset_scheme(name: str) -> bool: MXFP8 = QuantizationScheme.from_dict( + { + "bits": 8, + "group_size": 32, + "data_type": "mx_fp", + "act_bits": 8, + "act_data_type": "mx_fp", + "act_group_size": 32, + "act_sym": True, + } +) + +MXFP8_RCEIL = QuantizationScheme.from_dict( { "bits": 8, "group_size": 32, @@ -202,6 +214,7 @@ def is_preset_scheme(name: str) -> bool: } ) + NVFP4 = QuantizationScheme.from_dict( { "bits": 4, @@ -265,6 +278,7 @@ def is_preset_scheme(name: str) -> bool: "MXFP4": MXFP4, "MXFP4_RCEIL": MXFP4_RCEIL, "MXFP8": MXFP8, + "MXFP8_RCEIL": MXFP8_RCEIL, "NVFP4": NVFP4, "FPW8A16": FPW8A16, "W2A16G64": W2A16G64,