Only convert layers that actually have scaled weights...

fp8_optimization.py

@@ -110,7 +110,7 @@ def convert_fp8_scaled_linear(module, sd, original_dtype, params_to_keep={}, pat
                     lora = (lora_diffs, lora_strengths)
                     setattr(submodule, "lora", lora)
 
-            if isinstance(submodule, nn.Linear) and (has_scale or has_fp8_weight):
+            if isinstance(submodule, nn.Linear) and (has_scale and has_fp8_weight):
                 original_forward = submodule.forward
                 setattr(submodule, "original_forward", original_forward)
                 setattr(submodule, "forward", lambda input, m=submodule: fp8_scaled_linear_forward(m, original_dtype, input))

nodes_model_loading.py

@@ -651,9 +651,11 @@ class WanVideoModelLoader:
                             quantization = "fp8_e5m2"
                             break
 
+        if "scaled_fp8" in sd and quantization != "fp8_e4m3fn_scaled":
+            raise ValueError("The model is a scaled fp8 model, please set quantization to 'fp8_e4m3fn_scaled'")
+
         if merge_loras and "scaled" in quantization:
             raise ValueError("scaled models currently do not support merging LoRAs, please disable merging or use a non-scaled model")
-            
 
         if "vace_blocks.0.after_proj.weight" in sd and not "patch_embedding.weight" in sd:
             raise ValueError("You are attempting to load a VACE module as a WanVideo model, instead you should use the vace_model input and matching T2V base model")