🎈 perf: improve embedding perform

Author: hcd233

src/engine/embedding/sentence_transformer.py

@@ -1,3 +1,5 @@
+import asyncio
+import concurrent.futures
 from typing import List
 
 from cachetools import LRUCache
@@ -22,6 +24,16 @@ def from_config(cls, config: EmbeddingConfig) -> "EmbeddingEngine":
 
         logger.success(f"[EmbeddingEngine] load model from {config.path}")
         return cls(model=model, cache=cache, **config.model_dump())
+    
+    @property
+    def executor(self) -> concurrent.futures.ThreadPoolExecutor:
+        """延迟初始化线程池执行器"""
+        if self._executor is None:
+            self._executor = concurrent.futures.ThreadPoolExecutor(
+                max_workers=4,  # 单个模型使用一个worker避免GPU内存冲突
+                thread_name_prefix=f"gpu_reranker_{self.alias}"
+            )
+        return self._executor
 
     def invoke(self, sentences: List[str]) -> List[EmbeddingResult]:
         not_cached_ids = []
@@ -60,6 +72,15 @@ def invoke(self, sentences: List[str]) -> List[EmbeddingResult]:
             embeddings = no_cached_embeds
 
         return [EmbeddingResult(embedding=embedding, index=i, object="embedding") for i, embedding in enumerate(embeddings)]
+    
+    async def async_invoke(self, sentences: List[str]) -> List[EmbeddingResult]:
+        loop = asyncio.get_event_loop()
+        return await loop.run_in_executor(
+            self.executor,
+            self.invoke,
+            sentences
+        )
+
 
     def stream(self, sentences: List[str]) -> List[EmbeddingResult]:
-        raise NotImplementedError(f"{self.__class__.__name__} does not implement `stream` method")
+        raise NotImplementedError(f"{self.__class__.__name__} does not implement `stream` method")