Expose timeout for stremer for fastchat worker

python/llm/src/ipex_llm/serving/fastchat/ipex_llm_worker.py

@@ -72,6 +72,7 @@ def __init__(
         load_low_bit_model: bool = False,
         stream_interval: int = 4,
         benchmark: str = "true",
+        streamer_timeout: int = 60,
     ):
         super().__init__(
             controller_addr,
@@ -83,6 +84,8 @@ def __init__(
             conv_template,
         )
 
+        self.streamer_timeout = streamer_timeout
+
         self.load_in_low_bit = load_in_low_bit
         self.load_low_bit_model = load_low_bit_model
         logger.info(
@@ -323,7 +326,7 @@ def generate_stream_gate(self, params):
         # Use TextIteratorStreamer for streaming output
         streamer = TextIteratorStreamer(
             tokenizer=self.tokenizer,
-            timeout=60,
+            timeout=self.streamer_timeout,
             skip_prompt=True,
             skip_special_tokens=True,
         )