From 5cefa64a60d18f8e3b17c14ea68a869450f80639 Mon Sep 17 00:00:00 2001 From: Disty0 Date: Wed, 11 Jun 2025 20:58:54 +0300 Subject: [PATCH] SDNQ update accepted dtypes --- modules/sdnq/__init__.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/modules/sdnq/__init__.py b/modules/sdnq/__init__.py index ef62c432f..64134723a 100644 --- a/modules/sdnq/__init__.py +++ b/modules/sdnq/__init__.py @@ -381,7 +381,7 @@ class SDNQConfig(QuantizationConfigMixin): Args: weights_dtype (`str`, *optional*, defaults to `"int8"`): The target dtype for the weights after quantization. Supported values are: - ("int8", "int6", "int5", "int4", "int3", "int2", "uint8", "uint6", "uint5", "uint4", "uint3", "uint2", "uint1", "bool", "float8_e4m3fn", "float8_e4m3fnuz", "float8_e5m2", "float8_e5m2fnuz") + ("int8", "int7", "int6", "int5", "int4", "int3", "int2", "uint8", "uint7", "uint6", "uint5", "uint4", "uint3", "uint2", "uint1", "bool", "float8_e4m3fn", "float8_e4m3fnuz", "float8_e5m2", "float8_e5m2fnuz") modules_to_not_convert (`list`, *optional*, default to `None`): The list of modules to not quantize, useful for quantizing models that explicitly require to have some modules left in their original precision (e.g. Whisper encoder, Llava encoder, Mixtral gate layers). @@ -411,7 +411,7 @@ class SDNQConfig(QuantizationConfigMixin): r""" Safety checker that arguments are correct """ - accepted_weights = ["int8", "int6", "int5", "int4", "int3", "int2", "uint8", "uint6", "uint5", "uint4", "uint3", "uint2", "uint1", "bool", "float8_e4m3fn", "float8_e4m3fnuz", "float8_e5m2", "float8_e5m2fnuz"] + accepted_weights = ["int8", "int7", "int6", "int5", "int4", "int3", "int2", "uint8", "uint7", "uint6", "uint5", "uint4", "uint3", "uint2", "uint1", "bool", "float8_e4m3fn", "float8_e4m3fnuz", "float8_e5m2", "float8_e5m2fnuz"] if self.weights_dtype not in accepted_weights: raise ValueError(f"Only support weights in {accepted_weights} but found {self.weights_dtype}") if not isinstance(self.modules_to_not_convert, list):