当前位置：首页 > news >正文

开源模型应用落地-glm模型小试-glm-4-9b-chat-vLLM集成（四）

news 文章来源：https://blog.csdn.net/qq839019311/article/details/143567391 2025/5/15 6:01:41

一、前言

GLM-4是智谱AI团队于2024年1月16日发布的基座大模型，旨在自动理解和规划用户的复杂指令，并能调用网页浏览器。其功能包括数据分析、图表创建、PPT生成等，支持128K的上下文窗口，使其在长文本处理和精度召回方面表现优异，且在中文对齐能力上超过GPT-4。与之前的GLM系列产品相比，GLM-4在各项性能上提高了60%，并且在指令跟随和多模态功能上有显著强化，适合于多种应用场景。尽管在某些领域仍逊于国际一流模型，GLM-4的中文处理能力使其在国内大模型中占据领先地位。该模型的研发历程自2020年始，经过多次迭代和改进，最终构建出这一高性能的AI系统。

在开源模型应用落地-glm模型小试-glm-4-9b-chat-快速体验（一）已经掌握了glm-4-9b-chat的基本入门。

在开源模型应用落地-glm模型小试-glm-4-9b-chat-批量推理（二）已经掌握了glm-4-9b-chat的批量推理。

在开源模型应用落地-glm模型小试-glm-4-9b-chat-Gradio集成（三）已经掌握了如何集成Gradio进行页面交互。

本篇将介绍如何集成vLLM进行推理加速。

二、术语

2.1.GLM-4-9B

是智谱 AI 推出的一个开源预训练模型，属于 GLM-4 系列。它于 2024 年 6 月 6 日发布，专为满足高效能语言理解和生成任务而设计，并支持最高 1M（约两百万字）的上下文输入。该模型拥有更强的基础能力，支持26种语言，并且在多模态能力上首次实现了显著进展。

GLM-4-9B的基础能力包括：

- 中英文综合性能提升 40%，在特别的中文对齐能力、指令遵从和工程代码等任务中显著增强

- 较 Llama 3 8B 的性能提升，尤其在数学问题解决和代码编写等复杂任务中表现优越

- 增强的函数调用能力，提升了 40% 的性能

- 支持多轮对话，还支持网页浏览、代码执行、自定义工具调用等高级功能，能够快速处理大量信息并给出高质量的回答

2.2.GLM-4-9B-Chat

是智谱 AI 在 GLM-4-9B 系列中推出的对话版本模型。它设计用于处理多轮对话，并具有一些高级功能，使其在自然语言处理任务中更加高效和灵活。

2.3.vLLM

vLLM是一个开源的大模型推理加速框架，通过PagedAttention高效地管理attention中缓存的张量，实现了比HuggingFace Transformers高14-24倍的吞吐量。

三、前置条件

3.1.基础环境及前置条件

1. 操作系统：centos7

2. NVIDIA Tesla V100 32GB CUDA Version: 12.2

3.最低硬件要求

3.2.下载模型

huggingface：

https://huggingface.co/THUDM/glm-4-9b-chat/tree/main

ModelScope：

魔搭社区

使用git-lfs方式下载示例：

3.3.创建虚拟环境

conda create --name glm4 python=3.10
conda activate glm4

3.4.安装依赖库

pip install torch>=2.5.0
pip install torchvision>=0.20.0
pip install transformers>=4.46.0
pip install huggingface-hub>=0.25.1
pip install sentencepiece>=0.2.0
pip install jinja2>=3.1.4
pip install pydantic>=2.9.2
pip install timm>=1.0.9
pip install tiktoken>=0.7.0
pip install numpy==1.26.4 
pip install accelerate>=1.0.1
pip install sentence_transformers>=3.1.1
pip install openai>=1.51.0
pip install einops>=0.8.0
pip install pillow>=10.4.0
pip install sse-starlette>=2.1.3
pip install bitsandbytes>=0.43.3# using with VLLM Framework
pip install vllm>=0.6.3

四、技术实现

4.1.vLLM服务端实现

# -*- coding: utf-8 -*-
import time
from asyncio.log import logger
import re
import uvicorn
import gc
import json
import torch
import random
import string
from vllm import SamplingParams, AsyncEngineArgs, AsyncLLMEngine
from fastapi import FastAPI, HTTPException, Response
from fastapi.middleware.cors import CORSMiddleware
from contextlib import asynccontextmanager
from typing import List, Literal, Optional, Union
from pydantic import BaseModel, Field
from transformers import AutoTokenizer, LogitsProcessor
from sse_starlette.sse import EventSourceResponseEventSourceResponse.DEFAULT_PING_INTERVAL = 1000MAX_MODEL_LENGTH = 8192@asynccontextmanager
async def lifespan(app: FastAPI):yieldif torch.cuda.is_available():torch.cuda.empty_cache()torch.cuda.ipc_collect()app = FastAPI(lifespan=lifespan)app.add_middleware(CORSMiddleware,allow_origins=["*"],allow_credentials=True,allow_methods=["*"],allow_headers=["*"],
)def generate_id(prefix: str, k=29) -> str:suffix = ''.join(random.choices(string.ascii_letters + string.digits, k=k))return f"{prefix}{suffix}"class ModelCard(BaseModel):id: str = ""object: str = "model"created: int = Field(default_factory=lambda: int(time.time()))owned_by: str = "owner"root: Optional[str] = Noneparent: Optional[str] = Nonepermission: Optional[list] = Noneclass ModelList(BaseModel):object: str = "list"data: List[ModelCard] = ["glm-4"]class FunctionCall(BaseModel):name: Optional[str] = Nonearguments: Optional[str] = Noneclass ChoiceDeltaToolCallFunction(BaseModel):name: Optional[str] = Nonearguments: Optional[str] = Noneclass UsageInfo(BaseModel):prompt_tokens: int = 0total_tokens: int = 0completion_tokens: Optional[int] = 0class ChatCompletionMessageToolCall(BaseModel):index: Optional[int] = 0id: Optional[str] = Nonefunction: FunctionCalltype: Optional[Literal["function"]] = 'function'class ChatMessage(BaseModel):# “function” 字段解释：# 使用较老的OpenAI API版本需要注意在这里添加 function 字段并在 process_messages函数中添加相应角色转换逻辑为 observationrole: Literal["user", "assistant", "system", "tool"]content: Optional[str] = Nonefunction_call: Optional[ChoiceDeltaToolCallFunction] = Nonetool_calls: Optional[List[ChatCompletionMessageToolCall]] = Noneclass DeltaMessage(BaseModel):role: Optional[Literal["user", "assistant", "system"]] = Nonecontent: Optional[str] = Nonefunction_call: Optional[ChoiceDeltaToolCallFunction] = Nonetool_calls: Optional[List[ChatCompletionMessageToolCall]] = Noneclass ChatCompletionResponseChoice(BaseModel):index: intmessage: ChatMessagefinish_reason: Literal["stop", "length", "tool_calls"]class ChatCompletionResponseStreamChoice(BaseModel):delta: DeltaMessagefinish_reason: Optional[Literal["stop", "length", "tool_calls"]]index: intclass ChatCompletionResponse(BaseModel):model: strid: Optional[str] = Field(default_factory=lambda: generate_id('chatcmpl-', 29))object: Literal["chat.completion", "chat.completion.chunk"]choices: List[Union[ChatCompletionResponseChoice, ChatCompletionResponseStreamChoice]]created: Optional[int] = Field(default_factory=lambda: int(time.time()))system_fingerprint: Optional[str] = Field(default_factory=lambda: generate_id('fp_', 9))usage: Optional[UsageInfo] = Noneclass ChatCompletionRequest(BaseModel):model: strmessages: List[ChatMessage]temperature: Optional[float] = 0.8top_p: Optional[float] = 0.8max_tokens: Optional[int] = Nonestream: Optional[bool] = Falsetools: Optional[Union[dict, List[dict]]] = Nonetool_choice: Optional[Union[str, dict]] = Nonerepetition_penalty: Optional[float] = 1.1class InvalidScoreLogitsProcessor(LogitsProcessor):def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:if torch.isnan(scores).any() or torch.isinf(scores).any():scores.zero_()scores[..., 5] = 5e4return scoresdef process_response(output: str, tools: dict | List[dict] = None, use_tool: bool = False) -> Union[str, dict]:lines = output.strip().split("\n")arguments_json = Nonespecial_tools = ["cogview", "simple_browser"]tools = {tool['function']['name'] for tool in tools} if tools else {}if len(lines) >= 2 and lines[1].startswith("{"):function_name = lines[0].strip()arguments = "\n".join(lines[1:]).strip()if function_name in tools or function_name in special_tools:try:arguments_json = json.loads(arguments)is_tool_call = Trueexcept json.JSONDecodeError:is_tool_call = function_name in special_toolsif is_tool_call and use_tool:content = {"name": function_name,"arguments": json.dumps(arguments_json if isinstance(arguments_json, dict) else arguments,ensure_ascii=False)}if function_name == "simple_browser":search_pattern = re.compile(r'search\("(.+?)"\s*,\s*recency_days\s*=\s*(\d+)\)')match = search_pattern.match(arguments)if match:content["arguments"] = json.dumps({"query": match.group(1),"recency_days": int(match.group(2))}, ensure_ascii=False)elif function_name == "cogview":content["arguments"] = json.dumps({"prompt": arguments}, ensure_ascii=False)return contentreturn output.strip()@torch.inference_mode()
async def generate_stream_glm4(params):messages = params["messages"]tools = params["tools"]tool_choice = params["tool_choice"]temperature = float(params.get("temperature", 1.0))repetition_penalty = float(params.get("repetition_penalty", 1.0))top_p = float(params.get("top_p", 1.0))max_new_tokens = int(params.get("max_tokens", 8192))messages = process_messages(messages, tools=tools, tool_choice=tool_choice)inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)params_dict = {"n": 1,"best_of": 1,"presence_penalty": 1.0,"frequency_penalty": 0.0,"temperature": temperature,"top_p": top_p,"top_k": -1,"repetition_penalty": repetition_penalty,"stop_token_ids": [151329, 151336, 151338],"ignore_eos": False,"max_tokens": max_new_tokens,"logprobs": None,"prompt_logprobs": None,"skip_special_tokens": True,}sampling_params = SamplingParams(**params_dict)async for output in engine.generate(prompt=inputs, sampling_params=sampling_params, request_id=f"{time.time()}"):output_len = len(output.outputs[0].token_ids)input_len = len(output.prompt_token_ids)ret = {"text": output.outputs[0].text,"usage": {"prompt_tokens": input_len,"completion_tokens": output_len,"total_tokens": output_len + input_len},"finish_reason": output.outputs[0].finish_reason,}yield retgc.collect()torch.cuda.empty_cache()def process_messages(messages, tools=None, tool_choice="none"):_messages = messagesprocessed_messages = []msg_has_sys = Falsedef filter_tools(tool_choice, tools):function_name = tool_choice.get('function', {}).get('name', None)if not function_name:return []filtered_tools = [tool for tool in toolsif tool.get('function', {}).get('name') == function_name]return filtered_toolsif tool_choice != "none":if isinstance(tool_choice, dict):tools = filter_tools(tool_choice, tools)if tools:processed_messages.append({"role": "system","content": None,"tools": tools})msg_has_sys = Trueif isinstance(tool_choice, dict) and tools:processed_messages.append({"role": "assistant","metadata": tool_choice["function"]["name"],"content": ""})for m in _messages:role, content, func_call = m.role, m.content, m.function_calltool_calls = getattr(m, 'tool_calls', None)if role == "function":processed_messages.append({"role": "observation","content": content})elif role == "tool":processed_messages.append({"role": "observation","content": content,"function_call": True})elif role == "assistant":if tool_calls:for tool_call in tool_calls:processed_messages.append({"role": "assistant","metadata": tool_call.function.name,"content": tool_call.function.arguments})else:for response in content.split("\n"):if "\n" in response:metadata, sub_content = response.split("\n", maxsplit=1)else:metadata, sub_content = "", responseprocessed_messages.append({"role": role,"metadata": metadata,"content": sub_content.strip()})else:if role == "system" and msg_has_sys:msg_has_sys = Falsecontinueprocessed_messages.append({"role": role, "content": content})if not tools or tool_choice == "none":for m in _messages:if m.role == 'system':processed_messages.insert(0, {"role": m.role, "content": m.content})breakreturn processed_messages@app.get("/health")
async def health() -> Response:"""Health check."""return Response(status_code=200)@app.get("/v1/models", response_model=ModelList)
async def list_models():model_card = ModelCard(id="glm-4")return ModelList(data=[model_card])@app.post("/v1/chat/completions", response_model=ChatCompletionResponse)
async def create_chat_completion(request: ChatCompletionRequest):if len(request.messages) < 1 or request.messages[-1].role == "assistant":raise HTTPException(status_code=400, detail="Invalid request")gen_params = dict(messages=request.messages,temperature=request.temperature,top_p=request.top_p,max_tokens=request.max_tokens or 1024,echo=False,stream=request.stream,repetition_penalty=request.repetition_penalty,tools=request.tools,tool_choice=request.tool_choice,)logger.debug(f"==== request ====\n{gen_params}")if request.stream:predict_stream_generator = predict_stream(request.model, gen_params)output = await anext(predict_stream_generator)if output:return EventSourceResponse(predict_stream_generator, media_type="text/event-stream")logger.debug(f"First result output：\n{output}")function_call = Noneif output and request.tools:try:function_call = process_response(output, request.tools, use_tool=True)except:logger.warning("Failed to parse tool call")if isinstance(function_call, dict):function_call = ChoiceDeltaToolCallFunction(**function_call)generate = parse_output_text(request.model, output, function_call=function_call)return EventSourceResponse(generate, media_type="text/event-stream")else:return EventSourceResponse(predict_stream_generator, media_type="text/event-stream")response = ""async for response in generate_stream_glm4(gen_params):passif response["text"].startswith("\n"):response["text"] = response["text"][1:]response["text"] = response["text"].strip()usage = UsageInfo()function_call, finish_reason = None, "stop"tool_calls = Noneif request.tools:try:function_call = process_response(response["text"], request.tools, use_tool=True)except Exception as e:logger.warning(f"Failed to parse tool call: {e}")if isinstance(function_call, dict):finish_reason = "tool_calls"function_call_response = ChoiceDeltaToolCallFunction(**function_call)function_call_instance = FunctionCall(name=function_call_response.name,arguments=function_call_response.arguments)tool_calls = [ChatCompletionMessageToolCall(id=generate_id('call_', 24),function=function_call_instance,type="function")]message = ChatMessage(role="assistant",content=None if tool_calls else response["text"],function_call=None,tool_calls=tool_calls,)logger.debug(f"==== message ====\n{message}")choice_data = ChatCompletionResponseChoice(index=0,message=message,finish_reason=finish_reason,)task_usage = UsageInfo.model_validate(response["usage"])for usage_key, usage_value in task_usage.model_dump().items():setattr(usage, usage_key, getattr(usage, usage_key) + usage_value)return ChatCompletionResponse(model=request.model,choices=[choice_data],object="chat.completion",usage=usage)async def predict_stream(model_id, gen_params):output = ""is_function_call = Falsehas_send_first_chunk = Falsecreated_time = int(time.time())function_name = Noneresponse_id = generate_id('chatcmpl-', 29)system_fingerprint = generate_id('fp_', 9)tools = {tool['function']['name'] for tool in gen_params['tools']} if gen_params['tools'] else {}delta_text = ""async for new_response in generate_stream_glm4(gen_params):decoded_unicode = new_response["text"]delta_text += decoded_unicode[len(output):]output = decoded_unicodelines = output.strip().split("\n")# 检查是否为工具# 这是一个简单的工具比较函数，不能保证拦截所有非工具输出的结果，比如参数未对齐等特殊情况。##TODO 如果你希望做更多处理，可以在这里进行逻辑完善。if not is_function_call and len(lines) >= 2:first_line = lines[0].strip()if first_line in tools:is_function_call = Truefunction_name = first_linedelta_text = lines[1]# 工具调用返回if is_function_call:if not has_send_first_chunk:function_call = {"name": function_name, "arguments": ""}tool_call = ChatCompletionMessageToolCall(index=0,id=generate_id('call_', 24),function=FunctionCall(**function_call),type="function")message = DeltaMessage(content=None,role="assistant",function_call=None,tool_calls=[tool_call])choice_data = ChatCompletionResponseStreamChoice(index=0,delta=message,finish_reason=None)chunk = ChatCompletionResponse(model=model_id,id=response_id,choices=[choice_data],created=created_time,system_fingerprint=system_fingerprint,object="chat.completion.chunk")yield ""yield chunk.model_dump_json(exclude_unset=True)has_send_first_chunk = Truefunction_call = {"name": None, "arguments": delta_text}delta_text = ""tool_call = ChatCompletionMessageToolCall(index=0,id=None,function=FunctionCall(**function_call),type="function")message = DeltaMessage(content=None,role=None,function_call=None,tool_calls=[tool_call])choice_data = ChatCompletionResponseStreamChoice(index=0,delta=message,finish_reason=None)chunk = ChatCompletionResponse(model=model_id,id=response_id,choices=[choice_data],created=created_time,system_fingerprint=system_fingerprint,object="chat.completion.chunk")yield chunk.model_dump_json(exclude_unset=True)# 用户请求了 Function Call 但是框架还没确定是否为Function Callelif (gen_params["tools"] and gen_params["tool_choice"] != "none") or is_function_call:continue# 常规返回else:finish_reason = new_response.get("finish_reason", None)if not has_send_first_chunk:message = DeltaMessage(content="",role="assistant",function_call=None,)choice_data = ChatCompletionResponseStreamChoice(index=0,delta=message,finish_reason=finish_reason)chunk = ChatCompletionResponse(model=model_id,id=response_id,choices=[choice_data],created=created_time,system_fingerprint=system_fingerprint,object="chat.completion.chunk")yield chunk.model_dump_json(exclude_unset=True)has_send_first_chunk = Truemessage = DeltaMessage(content=delta_text,role="assistant",function_call=None,)delta_text = ""choice_data = ChatCompletionResponseStreamChoice(index=0,delta=message,finish_reason=finish_reason)chunk = ChatCompletionResponse(model=model_id,id=response_id,choices=[choice_data],created=created_time,system_fingerprint=system_fingerprint,object="chat.completion.chunk")yield chunk.model_dump_json(exclude_unset=True)# 工具调用需要额外返回一个字段以对齐 OpenAI 接口if is_function_call:yield ChatCompletionResponse(model=model_id,id=response_id,system_fingerprint=system_fingerprint,choices=[ChatCompletionResponseStreamChoice(index=0,delta=DeltaMessage(content=None,role=None,function_call=None,),finish_reason="tool_calls")],created=created_time,object="chat.completion.chunk",usage=None).model_dump_json(exclude_unset=True)elif delta_text != "":message = DeltaMessage(content="",role="assistant",function_call=None,)choice_data = ChatCompletionResponseStreamChoice(index=0,delta=message,finish_reason=None)chunk = ChatCompletionResponse(model=model_id,id=response_id,choices=[choice_data],created=created_time,system_fingerprint=system_fingerprint,object="chat.completion.chunk")yield chunk.model_dump_json(exclude_unset=True)finish_reason = 'stop'message = DeltaMessage(content=delta_text,role="assistant",function_call=None,)delta_text = ""choice_data = ChatCompletionResponseStreamChoice(index=0,delta=message,finish_reason=finish_reason)chunk = ChatCompletionResponse(model=model_id,id=response_id,choices=[choice_data],created=created_time,system_fingerprint=system_fingerprint,object="chat.completion.chunk")yield chunk.model_dump_json(exclude_unset=True)yield '[DONE]'else:yield '[DONE]'async def parse_output_text(model_id: str, value: str, function_call: ChoiceDeltaToolCallFunction = None):delta = DeltaMessage(role="assistant", content=value)if function_call is not None:delta.function_call = function_callchoice_data = ChatCompletionResponseStreamChoice(index=0,delta=delta,finish_reason=None)chunk = ChatCompletionResponse(model=model_id,choices=[choice_data],object="chat.completion.chunk")yield "{}".format(chunk.model_dump_json(exclude_unset=True))yield '[DONE]'if __name__ == "__main__":MODEL_PATH = "/data/model/glm-4-9b-chat"tensor_parallel_size = 1tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, trust_remote_code=True)engine_args = AsyncEngineArgs(model=MODEL_PATH,tokenizer=MODEL_PATH,# 如果你有多张显卡，可以在这里设置成你的显卡数量tensor_parallel_size=tensor_parallel_size,dtype=torch.float16,trust_remote_code=True,# 占用显存的比例，请根据你的显卡显存大小设置合适的值，例如，如果你的显卡有80G，您只想使用24G，请按照24/80=0.3设置gpu_memory_utilization=0.9,enforce_eager=True,worker_use_ray=False,disable_log_requests=True,max_model_len=MAX_MODEL_LENGTH,)engine = AsyncLLMEngine.from_engine_args(engine_args)uvicorn.run(app, host='0.0.0.0', port=8000, workers=1)

4.2.vLLM服务端启动

(glm4) [root@gpu test]# python -u glm_server.py 
WARNING 11-06 12:11:19 config.py:1668] Casting torch.bfloat16 to torch.float16.
WARNING 11-06 12:11:23 config.py:395] To see benefits of async output processing, enable CUDA graph. Since, enforce-eager is enabled, async output processor cannot be used
INFO 11-06 12:11:23 llm_engine.py:237] Initializing an LLM engine (v0.6.3.post1) with config: model='/data/model/glm-4-9b-chat', speculative_config=None, tokenizer='/data/model/glm-4-9b-chat', skip_tokenizer_init=False, tokenizer_mode=auto, revision=None, override_neuron_config=None, rope_scaling=None, rope_theta=None, tokenizer_revision=None, trust_remote_code=True, dtype=torch.float16, max_seq_len=8192, download_dir=None, load_format=LoadFormat.AUTO, tensor_parallel_size=1, pipeline_parallel_size=1, disable_custom_all_reduce=False, quantization=None, enforce_eager=True, kv_cache_dtype=auto, quantization_param_path=None, device_config=cuda, decoding_config=DecodingConfig(guided_decoding_backend='outlines'), observability_config=ObservabilityConfig(otlp_traces_endpoint=None, collect_model_forward_time=False, collect_model_execute_time=False), seed=0, served_model_name=/data/model/glm-4-9b-chat, num_scheduler_steps=1, chunked_prefill_enabled=False multi_step_stream_outputs=True, enable_prefix_caching=False, use_async_output_proc=False, use_cached_outputs=False, mm_processor_kwargs=None)
WARNING 11-06 12:11:24 tokenizer.py:169] Using a slow tokenizer. This might cause a significant slowdown. Consider using a fast tokenizer instead.
INFO 11-06 12:11:24 selector.py:224] Cannot use FlashAttention-2 backend for Volta and Turing GPUs.
INFO 11-06 12:11:24 selector.py:115] Using XFormers backend.
/usr/local/miniconda3/envs/glm4/lib/python3.10/site-packages/xformers/ops/fmha/flash.py:211: FutureWarning: `torch.library.impl_abstract` was renamed to `torch.library.register_fake`. Please use that instead; we will remove `torch.library.impl_abstract` in a future version of PyTorch.@torch.library.impl_abstract("xformers_flash::flash_fwd")
/usr/local/miniconda3/envs/glm4/lib/python3.10/site-packages/xformers/ops/fmha/flash.py:344: FutureWarning: `torch.library.impl_abstract` was renamed to `torch.library.register_fake`. Please use that instead; we will remove `torch.library.impl_abstract` in a future version of PyTorch.@torch.library.impl_abstract("xformers_flash::flash_bwd")
INFO 11-06 12:11:25 model_runner.py:1056] Starting to load model /data/model/glm-4-9b-chat...
INFO 11-06 12:11:25 selector.py:224] Cannot use FlashAttention-2 backend for Volta and Turing GPUs.
INFO 11-06 12:11:25 selector.py:115] Using XFormers backend.
Loading safetensors checkpoint shards:   0% Completed | 0/10 [00:00<?, ?it/s]
Loading safetensors checkpoint shards:  10% Completed | 1/10 [00:00<00:08,  1.01it/s]
Loading safetensors checkpoint shards:  20% Completed | 2/10 [00:01<00:07,  1.13it/s]
Loading safetensors checkpoint shards:  30% Completed | 3/10 [00:02<00:06,  1.14it/s]
Loading safetensors checkpoint shards:  40% Completed | 4/10 [00:03<00:05,  1.15it/s]
Loading safetensors checkpoint shards:  50% Completed | 5/10 [00:04<00:04,  1.18it/s]
Loading safetensors checkpoint shards:  60% Completed | 6/10 [00:05<00:03,  1.08it/s]
Loading safetensors checkpoint shards:  70% Completed | 7/10 [00:06<00:02,  1.07it/s]
Loading safetensors checkpoint shards:  80% Completed | 8/10 [00:07<00:01,  1.13it/s]
Loading safetensors checkpoint shards:  90% Completed | 9/10 [00:08<00:00,  1.10it/s]
Loading safetensors checkpoint shards: 100% Completed | 10/10 [00:08<00:00,  1.10it/s]
Loading safetensors checkpoint shards: 100% Completed | 10/10 [00:08<00:00,  1.11it/s]INFO 11-06 12:11:35 model_runner.py:1067] Loading model weights took 17.5635 GB
INFO 11-06 12:11:37 gpu_executor.py:122] # GPU blocks: 12600, # CPU blocks: 6553
INFO 11-06 12:11:37 gpu_executor.py:126] Maximum concurrency for 8192 tokens per request: 24.61x
INFO:     Started server process [1627618]
INFO:     Waiting for application startup.
INFO:     Application startup complete.
INFO:     Uvicorn running on http://0.0.0.0:8000 (Press CTRL+C to quit)

4.3.客户端实现

# -*- coding: utf-8 -*-
from openai import OpenAIbase_url = "http://127.0.0.1:8000/v1/"
client = OpenAI(api_key="EMPTY", base_url=base_url)
MODEL_PATH = "/data/model/glm-4-9b-chat"def chat(use_stream=False):messages = [{"role": "system","content": "你是一名专业的导游。",},{"role": "user","content": "请推荐一些广州特色的景点？",}]response = client.chat.completions.create(model=MODEL_PATH,messages=messages,stream=use_stream,max_tokens=8192,temperature=0.4,presence_penalty=1.2,top_p=0.9,)if response:if use_stream:for chunk in response:msg = chunk.choices[0].delta.contentprint(msg,end='',flush=True)else:print(response)else:print("Error:", response.status_code)if __name__ == "__main__":chat(use_stream=True)

4.4.客户端调用

(glm4) [root@gpu test]# python -u glm_client.py 当然可以！广州是中国广东省的省会，历史悠久，文化底蕴深厚，同时也是一座现代化的大都市。以下是一些广州的特色景点推荐：1. **白云山** - 广州著名的风景区，有“羊城第一秀”之称。山上空气清新，景色优美，是登山和观赏广州市区全景的好地方。2. **珠江夜游** - 乘坐游船在珠江上欣赏两岸的夜景，可以看到广州塔、海心沙等著名地标，以及璀璨的灯光秀。3. **长隆旅游度假区** - 包括长隆野生动物世界、长隆水上乐园、长隆国际大马戏等多个主题公园，适合家庭游玩。4. **陈家祠** - 又称陈氏书院，是一座典型的岭南传统建筑，以其精美的木雕、石雕和砖雕闻名。5. **越秀公园** - 公园内有五羊雕像，是广州的象征之一。还有中山纪念碑、镇海楼等历史遗迹。6. **北京路步行街** - 这里集合了购物、餐饮、娱乐于一体，是一条充满活力的商业街区。7. **上下九步行街** - 这条古老的街道以骑楼建筑为特色，两旁有许多老字号商店和小吃店，是体验广州传统文化的好去处。8. **广州塔（小蛮腰）** - 作为广州的地标性建筑，游客可以从这里俯瞰整个城市的壮丽景观。9. **南越王宫博物馆** - 展示了两千多年前南越国的历史文化，馆内有一座复原的宫殿模型。10. **荔湾湖公园** - 一个集自然风光与人文景观于一体的公园，湖水清澈，环境宜人。11. **广州动物园** - 是中国最大的城市动物园之一，拥有多种珍稀动物。12. **广州艺术博物院** - 收藏了大量珍贵的艺术品和历史文物，是了解广东乃至华南地区文化艺术的重要场所。这些景点不仅展示了广州的自然美景，也体现了其丰富的文化遗产和现代都市的风貌。希望您在广州旅行时能有一个愉快的体验！