keep scale_weight on gpu to avoid needless casts

2025-09-02 01:04:16 +03:00
parent 82608009c8
commit 42d7180759
2 changed files with 4 additions and 5 deletions
@@ -1305,7 +1305,7 @@ class WanVideoModelLoader:
        if "fp8" in quantization:
            for k, v in sd.items():
                if k.endswith(".scale_weight"):
-                    scale_weights[k] = v.to(base_dtype)
+                    scale_weights[k] = v.to(device, base_dtype)
        
        if "fp8_e4m3fn" in quantization:
            weight_dtype = torch.float8_e4m3fn