zai-org · vinzid · Jul 22, 2025
diff --git a/inference/README.md b/inference/README.md
@@ -102,6 +102,27 @@ python vllm_cli_demo.py  # LLM Such as GLM-4-9B-0414
 vllm serve THUDM/GLM-4-9B-0414 --tensor_parallel_size 2
 ```
 
+### Use glm-4 to build an OpenAI-compatible service
+
+Start the server:
+
+```shell
+python glm4_server.py THUDM/GLM-4-9B-0414
+```
+
+Client request:
+
+```shell
+curl -X POST http://127.0.0.1:8000/v1/chat/completions \
+-H 'Content-Type: application/json' \
+-d \
+"{ \
+    \"messages\": [ \
+        {\"role\": \"user\", \"content\": \"Who are you?\"} \
+    ] \
+}"
+```
+
 ### Use glm-4v to build an OpenAI-compatible service
 
 Start the server:

diff --git a/inference/README_zh.md b/inference/README_zh.md
@@ -102,6 +102,27 @@ python vllm_cli_demo.py # LLM Such as GLM-4-9B-0414
 vllm serve THUDM/GLM-4-9B-0414 --tensor_parallel_size 2
 ```
 
+### 使用 glm-4 构建 OpenAI 服务
+
+启动服务端
+
+```shell
+python glm4_server.py THUDM/GLM-4-9B-0414
+```
+
+客户端请求：
+
+```shell
+curl -X POST http://127.0.0.1:8000/v1/chat/completions \
+-H 'Content-Type: application/json' \
+-d \
+"{ \
+    \"messages\": [ \
+        {\"role\": \"user\", \"content\": \"Who are you?\"} \
+    ] \
+}"
+```
+
 ### 使用 glm-4v 构建 OpenAI 服务
 
 启动服务端

diff --git a/inference/glm4_server.py b/inference/glm4_server.py
@@ -0,0 +1,99 @@
+import sys
+from threading import Thread
+from typing import List, Literal, Optional
+
+import torch
+import uvicorn
+from datetime import datetime
+from fastapi import FastAPI
+from pydantic import BaseModel
+from transformers import AutoModelForCausalLM, AutoTokenizer, TextIteratorStreamer
+
+
+app = FastAPI()
+
+
+class MessageInput(BaseModel):
+    role: Literal["user", "assistant", "system"]
+    content: str
+    name: Optional[str] = None
+
+
+class MessageOutput(BaseModel):
+    role: Literal["assistant"]
+    content: str = None
+    name: Optional[str] = None
+
+
+class Choice(BaseModel):
+    message: MessageOutput
+
+
+class Request(BaseModel):
+    messages: List[MessageInput]
+    temperature: Optional[float] = 0.8
+    top_p: Optional[float] = 0.8
+    max_tokens: Optional[int] = 128000
+    repetition_penalty: Optional[float] = 1.0
+
+
+class Response(BaseModel):
+    model: str
+    choices: List[Choice]
+
+
+@app.post("/v1/chat/completions", response_model=Response)
+async def create_chat_completion(request: Request):
+    global model, tokenizer
+
+    print(datetime.now())
+    print("\033[91m--received_request\033[0m", request)
+    messages = [message.model_dump() for message in request.messages]
+    model_inputs = tokenizer.apply_chat_template(
+        messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt"
+    ).to(model.device)
+    streamer = TextIteratorStreamer(tokenizer=tokenizer, timeout=60, skip_prompt=True, skip_special_tokens=True)
+    generate_kwargs = {
+        "input_ids": model_inputs["input_ids"],
+        "attention_mask": model_inputs["attention_mask"],
+        "streamer": streamer,
+        "max_new_tokens": request.max_tokens,
+        "do_sample": True,
+        "top_p": request.top_p,
+        "temperature": request.temperature if request.temperature > 0 else 0.8,
+        "repetition_penalty": request.repetition_penalty,
+        "eos_token_id": model.config.eos_token_id,
+    }
+    thread = Thread(target=model.generate, kwargs=generate_kwargs)
+    thread.start()
+
+    result = ""
+    for new_token in streamer:
+        result += new_token
+    print(datetime.now())
+    print("\033[91m--generated_text\033[0m", result)
+
+    message = MessageOutput(
+        role="assistant",
+        content=result,
+    )
+    choice = Choice(
+        message=message,
+    )
+    response = Response(model=sys.argv[1].split("/")[-1].lower(), choices=[choice])
+    return response
+
+
+torch.cuda.empty_cache()
+
+if __name__ == "__main__":
+    MODEL_PATH = sys.argv[1]
+
+    tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, trust_remote_code=True)
+    model = AutoModelForCausalLM.from_pretrained(
+        MODEL_PATH,
+        torch_dtype=torch.bfloat16,
+        device_map="auto",
+    ).eval()
+
+    uvicorn.run(app, host="0.0.0.0", port=8000, workers=1)