diff --git a/ppdiffusers/deploy/export_model.py b/ppdiffusers/deploy/export_model.py
index 230fbd447..583bf6fe0 100644
--- a/ppdiffusers/deploy/export_model.py
+++ b/ppdiffusers/deploy/export_model.py
@@ -42,6 +42,10 @@ def convert_ppdiffusers_pipeline_to_fastdeploy_pipeline(
     pipeline = StableDiffusionPipeline.from_pretrained(
         model_path, unet=unet_model, safety_checker=None, feature_extractor=None
     )
+    # make sure we disable xformers
+    pipeline.disable_xformers_memory_efficient_attention()
+    pipeline.unet.set_default_attn_processor()
+    pipeline.vae.set_default_attn_processor()
     output_path = Path(output_path)
     # calculate latent's H and W
     latent_height = height // 8 if height is not None else None
diff --git a/ppdiffusers/deploy/infer.py b/ppdiffusers/deploy/infer.py
index 3b3e1bb63..d90dd7165 100644
--- a/ppdiffusers/deploy/infer.py
+++ b/ppdiffusers/deploy/infer.py
@@ -161,10 +161,13 @@ def create_paddle_inference_runtime(
     disable_paddle_pass=[],
     paddle_stream=None,
     workspace=None,
+    show_log=False
 ):
     assert not use_fp16 or not use_bf16, "use_fp16 and use_bf16 are mutually exclusive"
     option = fd.RuntimeOption()
     option.use_paddle_backend()
+    if show_log:
+        option.enable_paddle_log_info()
     if device_id == -1:
         option.use_cpu()
     else:
@@ -378,6 +381,15 @@ def main(args):
     elif args.backend == "paddle" or args.backend == "paddle_tensorrt":
         args.use_trt = args.backend == "paddle_tensorrt"
         runtime_options = dict(
+            unet=create_paddle_inference_runtime(
+                use_trt=args.use_trt,
+                dynamic_shape=unet_dynamic_shape,
+                use_fp16=args.use_fp16,
+                use_bf16=args.use_bf16,
+                device_id=args.device_id,
+                paddle_stream=paddle_stream,
+                workspace=10*1024*1024*1024
+            ),
             text_encoder=create_paddle_inference_runtime(
                 use_trt=args.use_trt,
                 dynamic_shape=text_encoder_dynamic_shape,
@@ -386,6 +398,7 @@ def main(args):
                 device_id=args.device_id,
                 disable_paddle_trt_ops=["arg_max", "range", "lookup_table_v2"],
                 paddle_stream=paddle_stream,
+                workspace=10*1024*1024*1024
             ),
             vae_encoder=create_paddle_inference_runtime(
                 use_trt=args.use_trt,
@@ -394,6 +407,7 @@ def main(args):
                 use_bf16=args.use_bf16,
                 device_id=args.device_id,
                 paddle_stream=paddle_stream,
+                workspace=10*1024*1024*1024
             ),
             vae_decoder=create_paddle_inference_runtime(
                 use_trt=args.use_trt,
@@ -402,14 +416,7 @@ def main(args):
                 use_bf16=args.use_bf16,
                 device_id=args.device_id,
                 paddle_stream=paddle_stream,
-            ),
-            unet=create_paddle_inference_runtime(
-                use_trt=args.use_trt,
-                dynamic_shape=unet_dynamic_shape,
-                use_fp16=args.use_fp16,
-                use_bf16=args.use_bf16,
-                device_id=args.device_id,
-                paddle_stream=paddle_stream,
+                workspace=10*1024*1024*1024
             ),
         )
     pipe = FastDeployStableDiffusionMegaPipeline.from_pretrained(
diff --git a/ppdiffusers/ppdiffusers/pipelines/stable_diffusion/pipeline_fastdeploy_stable_diffusion.py b/ppdiffusers/ppdiffusers/pipelines/stable_diffusion/pipeline_fastdeploy_stable_diffusion.py
index 319346858..a29bb65ee 100644
--- a/ppdiffusers/ppdiffusers/pipelines/stable_diffusion/pipeline_fastdeploy_stable_diffusion.py
+++ b/ppdiffusers/ppdiffusers/pipelines/stable_diffusion/pipeline_fastdeploy_stable_diffusion.py
@@ -272,7 +272,7 @@ def __call__(
 
                 unet_inputs = dict(
                     sample=latent_model_input,
-                    timestep=t,
+                    timestep=paddle.to_tensor(t).reshape([1,]),
                     encoder_hidden_states=prompt_embeds,
                     infer_op=infer_op_dict.get("unet", None),
                     output_shape=latent_model_input.shape,