Update optimum_encoder.py

optimum_encoder.py

@@ -10,7 +10,7 @@ from semantic_router.utils.logger import logger
 
 
 class OptimumEncoder(BaseEncoder):
-    name: str = "mixedbread-ai/mxbai-embed-large-v1"
+    name: str = "mixedbread-ai/mxbai-embed-large-v1/onnx/model_fp16.onnx"
     type: str = "huggingface"
     score_threshold: float = 0.5
     tokenizer_kwargs: Dict = {}
@@ -60,13 +60,14 @@ class OptimumEncoder(BaseEncoder):
 
         provider_options = {
             "trt_engine_cache_enable": True,
-            "trt_engine_cache_path": os.getenv('HF_HOME')
+            "trt_engine_cache_path": os.getenv('HF_HOME'),
+            "trt_fp16_enable": True
         }
 
         ort_model = ORTModelForFeatureExtraction.from_pretrained(
             self.name, 
-            use_io_binding=True,
-            provider=['TensorrtExecutionProvider'],
+            model_save_dir=os.getenv('HF_HOME'),
+            providers=['TensorrtExecutionProvider'],
             provider_options=provider_options,
             **self.model_kwargs
         )