More surgical fix for #10267

comfy/model_patcher.py

@@ -123,16 +123,26 @@ def move_weight_functions(m, device):
     return memory
 
 class LowVramPatch:
-    def __init__(self, key, patches):
+    def __init__(self, key, patches, convert_func=None, set_func=None):
         self.key = key
         self.patches = patches
+        self.convert_func = convert_func
+        self.set_func = set_func
+
     def __call__(self, weight):
+        if self.convert_func is not None:
+            weight = self.convert_func(weight.to(dtype=torch.float32, copy=True), inplace=True)
+
         intermediate_dtype = weight.dtype
-        if intermediate_dtype not in [torch.float32, torch.float16, torch.bfloat16]: #intermediate_dtype has to be one that is supported in math ops
+        if self.set_func is None and intermediate_dtype not in [torch.float32, torch.float16, torch.bfloat16]: #intermediate_dtype has to be one that is supported in math ops
             intermediate_dtype = torch.float32
             return comfy.float.stochastic_rounding(comfy.lora.calculate_weight(self.patches[self.key], weight.to(intermediate_dtype), self.key, intermediate_dtype=intermediate_dtype), weight.dtype, seed=string_to_seed(self.key))
 
-        return comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=intermediate_dtype)
+        out = comfy.lora.calculate_weight(self.patches[self.key], weight, self.key, intermediate_dtype=intermediate_dtype)
+        if self.set_func is not None:
+            return self.set_func(out, seed=string_to_seed(self.key), return_weight=True)
+        else:
+            return out
 
 def get_key_weight(model, key):
     set_func = None
@@ -657,13 +667,15 @@ def load(self, device_to=None, lowvram_model_memory=0, force_patch_weights=False
                         if force_patch_weights:
                             self.patch_weight_to_device(weight_key)
                         else:
-                            m.weight_function = [LowVramPatch(weight_key, self.patches)]
+                            _, set_func, convert_func = get_key_weight(self.model, weight_key)
+                            m.weight_function = [LowVramPatch(weight_key, self.patches, convert_func, set_func)]
                             patch_counter += 1
                     if bias_key in self.patches:
                         if force_patch_weights:
                             self.patch_weight_to_device(bias_key)
                         else:
-                            m.bias_function = [LowVramPatch(bias_key, self.patches)]
+                            _, set_func, convert_func = get_key_weight(self.model, bias_key)
+                            m.bias_function = [LowVramPatch(bias_key, self.patches, convert_func, set_func)]
                             patch_counter += 1
 
                     cast_weight = True
@@ -825,10 +837,12 @@ def partially_unload(self, device_to, memory_to_free=0):
                         module_mem += move_weight_functions(m, device_to)
                         if lowvram_possible:
                             if weight_key in self.patches:
-                                m.weight_function.append(LowVramPatch(weight_key, self.patches))
+                                _, set_func, convert_func = get_key_weight(self.model, weight_key)
+                                m.weight_function.append(LowVramPatch(weight_key, self.patches, convert_func, set_func))
                                 patch_counter += 1
                             if bias_key in self.patches:
-                                m.bias_function.append(LowVramPatch(bias_key, self.patches))
+                                _, set_func, convert_func = get_key_weight(self.model, bias_key)
+                                m.bias_function.append(LowVramPatch(bias_key, self.patches, convert_func, set_func))
                                 patch_counter += 1
                             cast_weight = True
 

comfy/ops.py

@@ -416,8 +416,10 @@ def convert_weight(self, weight, inplace=False, **kwargs):
                 else:
                     return weight * self.scale_weight.to(device=weight.device, dtype=weight.dtype)
 
-            def set_weight(self, weight, inplace_update=False, seed=None, **kwargs):
+            def set_weight(self, weight, inplace_update=False, seed=None, return_weight=False, **kwargs):
                 weight = comfy.float.stochastic_rounding(weight / self.scale_weight.to(device=weight.device, dtype=weight.dtype), self.weight.dtype, seed=seed)
+                if return_weight:
+                    return weight
                 if inplace_update:
                     self.weight.data.copy_(weight)
                 else: