SDNQ add dequantize model

modules/sdnq/dequantizer.py

@@ -103,6 +103,23 @@ def re_quantize_matmul_packed_int_symmetric(weight: torch.ByteTensor, scale: tor
     return re_quantize_matmul_symmetric(unpack_int_symetric(weight, shape, weights_dtype, dtype=scale.dtype), scale, result_shape, svd_up=svd_up, svd_down=svd_down)
 
 
+def dequantize_sdnq_model(model):
+    if hasattr(model, "sdnq_dequantizer"):
+        model.weight = torch.nn.Parameter(model.sdnq_dequantizer(model.weight, model.scale, model.zero_point, model.svd_up, model.svd_down))
+        del model.sdnq_dequantizer, model.scale, model.zero_point, model.svd_up, model.svd_down
+        return model
+    has_children = list(model.children())
+    if not has_children:
+        return model
+    for module in model.children():
+        if hasattr(module, "sdnq_dequantizer"):
+            module.weight = torch.nn.Parameter(module.sdnq_dequantizer(module.weight, module.scale, module.zero_point, module.svd_up, module.svd_down))
+            del module.sdnq_dequantizer, module.scale, module.zero_point, module.svd_up, module.svd_down
+        else:
+            module = dequantize_sdnq_model(module)
+    return model
+
+
 class AsymmetricWeightsDequantizer(torch.nn.Module):
     def __init__(
         self,

modules/sdnq/quantizer.py

@@ -13,7 +13,7 @@ from diffusers.utils import get_module_from_name
 from modules import devices, shared
 
 from .common import dtype_dict, use_tensorwise_fp8_matmul, allowed_types, conv_types, conv_transpose_types, use_contiguous_mm
-from .dequantizer import dequantizer_dict
+from .dequantizer import dequantizer_dict, dequantize_sdnq_model
 from .forward import get_forward_func
 
 
@@ -713,6 +713,16 @@ class SDNQQuantizer(DiffusersQuantizer, HfQuantizer):
         """
         return dtype
 
+    def _dequantize(self, model):
+        model = dequantize_sdnq_model(model)
+        if hasattr(model, "quantization_method"):
+            del model.quantization_method
+        if hasattr(model, "quantization_config"):
+            del model.quantization_config
+        if hasattr(model, "config") and hasattr(model.config, "quantization_config"):
+            del model.config.quantization_config
+        return model
+
     def is_serializable(self, *args, **kwargs) -> bool:  # pylint: disable=unused-argument, invalid-overridden-method
         return True