Integrate LightLLM into serve worker (#2888)

Author: zeyugao

docs/lightllm_integration.md

@@ -0,0 +1,18 @@
+# LightLLM Integration
+You can use [LightLLM](https://github.com/ModelTC/lightllm) as an optimized worker implementation in FastChat.
+It offers advanced continuous batching and a much higher (~10x) throughput.
+See the supported models [here](https://github.com/ModelTC/lightllm?tab=readme-ov-file#supported-model-list).
+
+## Instructions
+1. Please refer to the [Get started](https://github.com/ModelTC/lightllm?tab=readme-ov-file#get-started) to install LightLLM. Or use [Pre-built image](https://github.com/ModelTC/lightllm?tab=readme-ov-file#container)
+
+2. When you launch a model worker, replace the normal worker (`fastchat.serve.model_worker`) with the LightLLM worker (`fastchat.serve.lightllm_worker`). All other commands such as controller, gradio web server, and OpenAI API server are kept the same. Refer to [--max_total_token_num](https://github.com/ModelTC/lightllm/blob/4a9824b6b248f4561584b8a48ae126a0c8f5b000/docs/ApiServerArgs.md?plain=1#L23) to understand how to calcuate the `--max_total_token_num` argument.
+   ```
+   python3 -m fastchat.serve.lightllm_worker --model-path lmsys/vicuna-7b-v1.5 --tokenizer_mode "auto" --max_total_token_num 154000
+   ```
+
+   If you what to use quantized weight and kv cache for inference, try
+
+   ```
+   python3 -m fastchat.serve.lightllm_worker --model-path lmsys/vicuna-7b-v1.5 --tokenizer_mode "auto" --max_total_token_num 154000 --mode triton_int8weight triton_int8kv
+   ```

fastchat/serve/base_model_worker.py

@@ -126,18 +126,18 @@ def send_heart_beat(self):
             self.register_to_controller()
 
     def get_queue_length(self):
-        if (
-            self.semaphore is None
-            or self.semaphore._value is None
-            or self.semaphore._waiters is None
-        ):
+        if self.semaphore is None:
             return 0
         else:
-            return (
-                self.limit_worker_concurrency
-                - self.semaphore._value
-                + len(self.semaphore._waiters)
+            sempahore_value = (
+                self.semaphore._value
+                if self.semaphore._value is not None
+                else self.limit_worker_concurrency
             )
+            waiter_count = (
+                0 if self.semaphore._waiters is None else len(self.semaphore._waiters)
+            )
+            return self.limit_worker_concurrency - sempahore_value + waiter_count
 
     def get_status(self):
         return {