Upload handler.py

handler.py

@@ -179,7 +179,7 @@ def load_pipeline_opt(repo_id: str, dtype: torch.dtype) -> Any:
 
 def load_pipeline_para(repo_id: str, dtype: torch.dtype) -> Any:
     weight = int8_dynamic_activation_int4_weight() if IS_4BIT else int8_dynamic_activation_int8_weight()
-    pipe = FluxPipeline.from_pretrained(repo_id, torch_dtype=dtype)
+    pipe = FluxPipeline.from_pretrained(repo_id, torch_dtype=dtype).to("cuda")
     if IS_PARA:
         if IS_MGPU:
             from para_attn.context_parallel import init_context_parallel_mesh