Update app.py

app.py

@@ -61,8 +61,8 @@ def init_predictor():
         model_id,
         # subfolder="transformer",
         torch_dtype=torch.bfloat16,
-        device="cpu",
-    ).to("cpu").eval()
+        device="cuda",
+    ).to("cuda").eval()
     quantize_(text_encoder, float8_weight_only(), device="cpu")
     text_encoder.to("cpu")
     torch.cuda.empty_cache()