[Graph Optimization] SOT+CUDAGraph support ERNIE4.5T VL 28B / 424B (#4645)

* 45TVL support sot+CUDAGraph

* mv unitest from ce_deploy 2 e2e

* add test_EB_VL_Lite_sot_serving

* rm useless line

* add openai_client

* fix unitest && reduce computing resources

Author: DrRyanHuang

benchmarks/benchmark_serving.py

@@ -982,7 +982,7 @@ def main(args: argparse.Namespace):
         if args.result_dir:
             file_name = os.path.join(args.result_dir, file_name)
         with open(file_name, "w", encoding="utf-8") as outfile:
-            json.dump(result_json, outfile)
+            json.dump(result_json, outfile, ensure_ascii=False)
         save_to_pytorch_benchmark_format(args, result_json, file_name)
 
 

fastdeploy/model_executor/models/ernie4_5_vl/ernie4_5_vl_moe.py

@@ -277,7 +277,7 @@ def load_state_dict(self, state_dict):
     def forward(self, hidden_states: paddle.Tensor, vl_moe_meta: VLMoEMeta):
         if self.num_shared_experts > 0:
             shared_experts_out = self.shared_experts(hidden_states)
-        hidden_states, vl_moe_meta.text_input, vl_moe_meta.image_input = text_image_gather_scatter(
+        hidden_states, text_input, image_input = text_image_gather_scatter(
             hidden_states,
             vl_moe_meta.text_input,
             vl_moe_meta.image_input,
@@ -286,8 +286,8 @@ def forward(self, hidden_states: paddle.Tensor, vl_moe_meta: VLMoEMeta):
             vl_moe_meta.image_index,
             True,
         )
-        text_out = self.text_fused_moe(vl_moe_meta.text_input)
-        image_out = self.image_fused_moe(vl_moe_meta.image_input)
+        text_out = self.text_fused_moe(text_input)
+        image_out = self.image_fused_moe(image_input)
         hidden_states, _, _ = text_image_gather_scatter(
             hidden_states,
             text_out,