Disable FA3 processor - CUDA kernel mismatch on ZeroGPU (toggle USE_FA3 to re-enable)

app.py

@@ -37,11 +37,15 @@ pipe = QwenImageEditPlusPipeline.from_pretrained(
     torch_dtype=dtype,
 ).to(device)
 
-try:
-    pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())
-    print("Flash Attention 3 Processor set successfully.")
-except Exception as e:
-    print(f"Warning: Could not set FA3 processor: {e}")
+USE_FA3 = False  # Set True to re-enable Flash Attention 3 (needs compatible GPU kernel)
+if USE_FA3:
+    try:
+        pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3())
+        print("Flash Attention 3 Processor set successfully.")
+    except Exception as e:
+        print(f"Warning: Could not set FA3 processor: {e}")
+else:
+    print("FA3 disabled — using default attention (ZeroGPU compat).")
 
 EXAMPLES_CONFIG = [
     {