![\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250418233646-6802e20e1da94.png\")
<\/p>\n\u672c\u5730\u670d\u52a1\u5668\u90e8\u7f72\u5f00\u6e90\u5927\u6a21\u578b\u6709\u4e00\u4e2a\u524d\u63d0，\u5c31\u662f\u5f97\u6709 GPU \u663e\u5361\u8d44\u6e90，\u5728\u6211\u4e0b\u9762\u7684\u4f8b\u5b50\u4e2d\u6211\u79df\u7528\u4e86 autodl \u4e2d\u7684\u7b97\u529b\u8d44\u6e90，\u5177\u4f53\u662f\u79df\u7528\u4e86\u4e00\u5f20\u6d88\u8d39\u7ea7\u522b\u7684 RTX 3090 \u663e\u5361\u3002<\/p>\n
<\/p>\n
# \u5347\u7ea7pip
\npython -m pip install –upgrade pip
\n# \u66f4\u6362 pypi \u6e90\u52a0\u901f\u5e93\u7684\u5b89\u88c5
\npip config set global.index-url https:\/\/pypi.tuna.tsinghua.edu.cn\/simple<\/p>\n
pip install fastapi==0.104.1
\npip install uvicorn==0.24.0.post1
\npip install requests==2.25.1
\npip install modelscope==1.9.5
\npip install transformers==4.42.4
\npip install streamlit==1.24.0
\npip install sentencepiece==0.1.99
\npip install accelerate==0.24.1
\npip install tiktoken==0.7.0<\/p>\n
\u8fd9\u91cc\u8981\u6ce8\u610f transformers \u7684\u7248\u672c\u662f 4.42.4<\/p>\n
GLM-4-9B-Chat \u6a21\u578b\u5927\u5c0f\u4e3a 18 GB，\u4e0b\u8f7d\u6a21\u578b\u5927\u6982\u9700\u8981 10~20 \u5206\u949f\u3002<\/p>\n
\u7531\u4e8e\u540e\u9762\u6211\u4eec\u8981\u4f7f\u7528\u4e00\u4e2a\u5f00\u6e90\u7684 embedding \u6a21\u578b BAAI\/bge-base-zh-v1.5<\/p>\n
\u6240\u4ee5\u4f7f\u7528\u4ee5\u4e0b\u4ee3\u7801\u4e0b\u8f7d 2 \u4e2a\u6a21\u578b\u6587\u4ef6\u5230\u672c\u5730\u6587\u4ef6\u7cfb\u7edf：<\/p>\n
\u8fd0\u884c python download.py<\/p>\n
import torch
\nfrom modelscope import snapshot_download, AutoModel, AutoTokenizer
\nimport os
\nmodel_dir = snapshot_download('ZhipuAI\/glm-4-9b-chat', cache_dir='\/root\/autodl-tmp', revision='master')
\nembedding_model_dir = snapshot_download('BAAI\/bge-base-zh-v1.5', cache_dir='\/root\/autodl-tmp', revision='master')<\/p>\n
GLM \u5f00\u6e90\u6a21\u578b\u5b98\u65b9\u7ed9\u4e86\u4e00\u4e2a Demo \u65b9\u4fbf\u6211\u4eec\u505a\u6d4b\u8bd5，\u4ee5\u4e0b\u662f\u4ee3\u7801：<\/p>\n
\u8fd0\u884c python trans_cli_demo.py<\/p>\n
"""
\nThis script creates a CLI demo with transformers backend for the glm-4-9b model,
\nallowing users to interact with the model through a command-line interface.<\/p>\n
Usage:
\n– Run the script to start the CLI demo.
\n– Interact with the model by typing questions and receiving responses.<\/p>\n
Note: The script includes a modification to handle markdown to plain text conversion,
\nensuring that the CLI interface displays formatted text correctly.<\/p>\n
If you use flash attention, you should install the flash-attn and add attn_implementation="flash_attention_2" in model loading.
\n"""<\/p>\n
import os
\nimport torch
\nfrom threading import Thread
\nfrom transformers import AutoTokenizer, StoppingCriteria, StoppingCriteriaList, TextIteratorStreamer, AutoModelForCausalLM<\/p>\n
MODEL_PATH = os.environ.get('MODEL_PATH', '\/root\/autodl-tmp\/ZhipuAI\/glm-4-9b-chat')<\/p>\n
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH, trust_remote_code=True)<\/p>\n
model = AutoModelForCausalLM.from_pretrained(
\n MODEL_PATH,
\n trust_remote_code=True,
\n device_map="auto"
\n).eval()<\/p>\n
class StopOnTokens(StoppingCriteria):
\n def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
\n stop_ids = model.config.eos_token_id
\n for stop_id in stop_ids:
\n if input_ids[0][-1] == stop_id:
\n return True
\n return False<\/p>\n
if __name__ == "__main__":
\n history = []
\n max_length = 8192
\n top_p = 0.8
\n temperature = 0.6
\n stop = StopOnTokens()<\/p>\n
print("Welcome to the GLM-4-9B CLI chat. Type your messages below.")
\n while True:
\n user_input = input("\\\\nYou: ")
\n if user_input.lower() in ["exit", "quit"]:
\n break
\n history.append([user_input, ""])<\/p>\n
messages = []
\n for idx, (user_msg, model_msg) in enumerate(history):
\n if idx == len(history) – 1 and not model_msg:
\n messages.append({"role": "user", "content": user_msg})
\n break
\n if user_msg:
\n messages.append({"role": "user", "content": user_msg})
\n if model_msg:
\n messages.append({"role": "assistant", "content": model_msg})
\n model_inputs = tokenizer.apply_chat_template(
\n messages,
\n add_generation_prompt=True,
\n tokenize=True,
\n return_tensors="pt"
\n ).to(model.device)
\n streamer = TextIteratorStreamer(
\n tokenizer=tokenizer,
\n timeout=60,
\n skip_prompt=True,
\n skip_special_tokens=True
\n )
\n generate_kwargs = {
\n "input_ids": model_inputs,
\n "streamer": streamer,
\n "max_new_tokens": max_length,
\n "do_sample": False, # \u6539\u4e3a False
\n "top_p": top_p,
\n "temperature": temperature,
\n "stopping_criteria": StoppingCriteriaList([stop]),
\n "repetition_penalty": 1.2,
\n "eos_token_id": model.config.eos_token_id,
\n }
\n try:
\n t = Thread(target=model.generate, kwargs=generate_kwargs)
\n t.start()
\n print("GLM-4:", end="", flush=True)
\n for new_token in streamer:
\n if new_token:
\n print(new_token, end="", flush=True)
\n history[-1][1] += new_token
\n except Exception as e:
\n print(f"An error occurred: {e}")
\n print(f"Error type: {type(e)}")
\n import traceback
\n traceback.print_exc()<\/p>\n
history[-1][1] = history[-1][1].strip()<\/p>\n
\u6ce8\u610f\u4ee5\u4e0a\u4ee3\u7801\u548c GLM \u5b98\u65b9\u63d0\u4f9b\u7684\u53ef\u80fd\u4e0d\u592a\u4e00\u6837，\u56e0\u4e3a\u5b98\u65b9\u7684\u6709\u7684\u62a5\u9519，\u6240\u4ee5\u6211\u7565\u4e3a\u4fee\u6539\u4e86\u4e00\u4e0b\u3002<\/p>\n
\u76f4\u63a5\u8fd0\u884c trans_cli_demo.py \u5c31\u53ef\u4ee5\u548c\u6a21\u578b\u4ea4\u4e92\u4e86<\/p>\n
![\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250418233646-6802e20e97c16.png\")
<\/p>\n
\u8fd0\u884c\u4ee5\u4e0b\u4ee3\u7801\u521b\u5efa\u5e76\u542f\u52a8 Api \u670d\u52a1：<\/p>\n
\u8fd0\u884c python api.py<\/p>\n
from fastapi import FastAPI, Request
\nfrom transformers import AutoTokenizer, AutoModelForCausalLM
\nimport uvicorn
\nimport json
\nimport datetime
\nimport torch<\/p>\n
# \u8bbe\u7f6e\u8bbe\u5907\u53c2\u6570
\nDEVICE = "cuda" # \u4f7f\u7528CUDA
\nDEVICE_ID = "0" # CUDA\u8bbe\u5907ID，\u5982\u679c\u672a\u8bbe\u7f6e\u5219\u4e3a\u7a7a
\nCUDA_DEVICE = f"{DEVICE}:{DEVICE_ID}" if DEVICE_ID else DEVICE # \u7ec4\u5408CUDA\u8bbe\u5907\u4fe1\u606f<\/p>\n
# \u6e05\u7406GPU\u5185\u5b58\u51fd\u6570
\ndef torch_gc():
\n if torch.cuda.is_available(): # \u68c0\u67e5\u662f\u5426\u53ef\u7528CUDA
\n with torch.cuda.device(CUDA_DEVICE): # \u6307\u5b9aCUDA\u8bbe\u5907
\n torch.cuda.empty_cache() # \u6e05\u7a7aCUDA\u7f13\u5b58
\n torch.cuda.ipc_collect() # \u6536\u96c6CUDA\u5185\u5b58\u788e\u7247<\/p>\n
# \u521b\u5efaFastAPI\u5e94\u7528
\napp = FastAPI()<\/p>\n
# \u5904\u7406POST\u8bf7\u6c42\u7684\u7aef\u70b9
\n@app.post("\/")
\nasync def create_item(request: Request):
\n global model, tokenizer # \u58f0\u660e\u5168\u5c40\u53d8\u91cf\u4ee5\u4fbf\u5728\u51fd\u6570\u5185\u90e8\u4f7f\u7528\u6a21\u578b\u548c\u5206\u8bcd\u5668
\n json_post_raw = await request.json() # \u83b7\u53d6POST\u8bf7\u6c42\u7684JSON\u6570\u636e
\n json_post = json.dumps(json_post_raw) # \u5c06JSON\u6570\u636e\u8f6c\u6362\u4e3a\u5b57\u7b26\u4e32
\n json_post_list = json.loads(json_post) # \u5c06\u5b57\u7b26\u4e32\u8f6c\u6362\u4e3aPython\u5bf9\u8c61
\n prompt = json_post_list.get('prompt') # \u83b7\u53d6\u8bf7\u6c42\u4e2d\u7684\u63d0\u793a
\n history = json_post_list.get('history') # \u83b7\u53d6\u8bf7\u6c42\u4e2d\u7684\u5386\u53f2\u8bb0\u5f55
\n max_length = json_post_list.get('max_length', 2048) # \u83b7\u53d6\u8bf7\u6c42\u4e2d\u7684\u6700\u5927\u957f\u5ea6
\n top_p = json_post_list.get('top_p', 0.7) # \u83b7\u53d6\u8bf7\u6c42\u4e2d\u7684top_p\u53c2\u6570
\n temperature = json_post_list.get('temperature', 0.95) # \u83b7\u53d6\u8bf7\u6c42\u4e2d\u7684\u6e29\u5ea6\u53c2\u6570<\/p>\n
# \u51c6\u5907\u8f93\u5165
\n messages = []
\n if history:
\n for h in history:
\n messages.append({"role": "user", "content": h[0]})
\n messages.append({"role": "assistant", "content": h[1]})
\n messages.append({"role": "user", "content": prompt})<\/p>\n
input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt").to(model.device)<\/p>\n
# \u751f\u6210\u56de\u590d
\n with torch.no_grad():
\n outputs = model.generate(
\n input_ids,
\n max_new_tokens=max_length,
\n do_sample=True,
\n top_p=top_p,
\n temperature=temperature,
\n )<\/p>\n
response = tokenizer.decode(outputs[0][input_ids.shape[1]:], skip_special_tokens=True)<\/p>\n
now = datetime.datetime.now() # \u83b7\u53d6\u5f53\u524d\u65f6\u95f4
\n time = now.strftime("%Y-%m-%d %H:%M:%S") # \u683c\u5f0f\u5316\u65f6\u95f4\u4e3a\u5b57\u7b26\u4e32
\n # \u6784\u5efa\u54cd\u5e94JSON
\n answer = {
\n "response": response,
\n "history": history + [[prompt, response]],
\n "status": 200,
\n "time": time
\n }
\n # \u6784\u5efa\u65e5\u5fd7\u4fe1\u606f
\n log = "[" + time + "] " + '", prompt:"' + prompt + '", response:"' + repr(response) + '"'
\n print(log) # \u6253\u5370\u65e5\u5fd7
\n torch_gc() # \u6267\u884cGPU\u5185\u5b58\u6e05\u7406
\n return answer # \u8fd4\u56de\u54cd\u5e94<\/p>\n
# \u4e3b\u51fd\u6570\u5165\u53e3
\nif __name__ == '__main__':
\n # \u52a0\u8f7d\u9884\u8bad\u7ec3\u7684\u5206\u8bcd\u5668\u548c\u6a21\u578b
\n tokenizer = AutoTokenizer.from_pretrained("\/root\/autodl-tmp\/ZhipuAI\/glm-4-9b-chat", trust_remote_code=True)
\n model = AutoModelForCausalLM.from_pretrained(
\n "\/root\/autodl-tmp\/ZhipuAI\/glm-4-9b-chat",
\n torch_dtype=torch.bfloat16,
\n trust_remote_code=True,
\n device_map="auto",
\n )
\n model.eval() # \u8bbe\u7f6e\u6a21\u578b\u4e3a\u8bc4\u4f30\u6a21\u5f0f
\n # \u542f\u52a8FastAPI\u5e94\u7528
\n # \u75286006\u7aef\u53e3\u53ef\u4ee5\u5c06autodl\u7684\u7aef\u53e3\u6620\u5c04\u5230\u672c\u5730，\u4ece\u800c\u5728\u672c\u5730\u4f7f\u7528api
\n uvicorn.run(app, host='0.0.0.0', port=6006, workers=1) # \u5728\u6307\u5b9a\u7aef\u53e3\u548c\u4e3b\u673a\u4e0a\u542f\u52a8\u5e94\u7528<\/p>\n
\u6d4b\u8bd5\u670d\u52a1<\/p>\n
curl -X POST "http:\/\/127.0.0.1:6006" \\\\
\n -H 'Content-Type: application\/json' \\\\
\n -d '{"prompt": "\u4f60\u597d", "history": []}'<\/p>\n
\u5229\u7528 FastApi \u540c\u6837\u53ef\u4ee5\u6d4b\u8bd5\u6a21\u578b\u7684\u8c03\u7528\u548c\u4ea4\u4e92\u3002<\/p>\n
![\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250418233647-6802e20f045da.png\")
<\/p>\n
\u6ce8\u610f，\u4ee5\u4e0a\u4ee3\u7801\u4f60\u53ef\u80fd\u4f1a\u5728\u7f51\u7edc\u4e0a\u627e\u5230\u7c7b\u4f3c\u7684，\u6211\u5728\u6700\u5f00\u59cb\u4f7f\u7528\u90a3\u4e9b\u4ee3\u7801\u7684\u65f6\u5019\u62a5\u5404\u79cd\u9519，\u539f\u56e0\u5927\u6982\u5305\u62ec\u6a21\u578b\u548c\u4ee3\u7801\u7248\u672c\u4e0d\u517c\u5bb9，\u7ec4\u4ef6\u5e93\u7248\u672c\u95ee\u9898\u7b49\u3002\u6240\u4ee5\u4ee5\u4e0a\u4ee3\u7801\u662f\u7ecf\u8fc7\u6211\u7684\u4fee\u6539\u4e4b\u540e\u53ef\u8fd0\u884c\u7684\u4ee3\u7801<\/p>\n
\u5728\u4e4b\u524d\u7684\u6587\u7ae0\u4e2d\u6211\u4eec\u901a\u8fc7 Ollama \u5728\u7b14\u8bb0\u672c\u7535\u8111\u4e0a\u90e8\u7f72\u8fc7\u5927\u6a21\u578b，\u901a\u8fc7\u5927\u6a21\u578b\u4ea7\u54c1\u7684 API \u8c03\u7528\u8fc7\u5927\u6a21\u578b ，\u552f\u72ec\u6ca1\u6709\u5728\u670d\u52a1\u5668\u4e0a\u79c1\u6709\u5316\u90e8\u7f72\u4e00\u4e2a\u5927\u6a21\u578b\u3002<\/p>\n
\u524d\u6587\u6211\u4eec\u5df2\u7ecf\u5728\u670d\u52a1\u5668\u4e0a\u90e8\u7f72\u597d\u4e86\u5927\u6a21\u578b glm-4-9b-chat \u8fd9\u662f\u4e00\u4e2a\u62e5\u6709 90 \u4ebf\u53c2\u6570\u7684\u6a21\u578b\u3002\u4e0b\u9762\u6211\u4eec\u4ecb\u7ecd\u5982\u4f55\u5728 llamaindex \u4e2d\u8c03\u7528\u5b83\u3002<\/p>\n
\u5f88\u7b80\u5355，\u9996\u5148\u6211\u4eec\u8fd8\u662f\u5148\u81ea\u5b9a\u4e49\u4e00\u4e2aLLM ，\u53c2\u8003\u4ee5\u4e0b\u4ee3\u7801：<\/p>\n
import logging
\nfrom typing import Any, List, Optional
\nfrom llama_index.core.llms import (
\n CustomLLM,
\n CompletionResponse,
\n CompletionResponseGen,
\n LLMMetadata,
\n)
\nfrom llama_index.core.llms.callbacks import llm_completion_callback
\nfrom transformers import AutoTokenizer, AutoModelForCausalLM
\nimport torch<\/p>\n
# \u8bbe\u7f6e\u65e5\u5fd7
\nlogging.basicConfig(level=logging.DEBUG)
\nlogger = logging.getLogger(__name__)<\/p>\n
class LocalGLM4(CustomLLM):<\/p>\n
context_window: int = 8192 # \u9ed8\u8ba4\u4e0a\u4e0b\u6587\u7a97\u53e3\u5927\u5c0f
\n num_output: int = 2048 # \u9ed8\u8ba4\u8f93\u51fa\u7684token\u6570\u91cf
\n model_name: str = "glm-4-9b-chat" # \u6a21\u578b\u540d\u79f0
\n tokenizer: object = None # \u5206\u8bcd\u5668
\n model: object = None # \u6a21\u578b<\/p>\n
def __init__(self, pretrained_model_name_or_path: str):
\n super().__init__()<\/p>\n
# GPU\u65b9\u5f0f\u52a0\u8f7d\u6a21\u578b
\n self.tokenizer = AutoTokenizer.from_pretrained(
\n pretrained_model_name_or_path, trust_remote_code=True
\n )
\n self.model = AutoModelForCausalLM.from_pretrained(
\n pretrained_model_name_or_path,
\n torch_dtype=torch.float16, # \u6216\u8005\u4f7f\u7528 torch.bfloat16
\n low_cpu_mem_usage=True,
\n trust_remote_code=True,
\n device_map="auto",
\n )<\/p>\n
# CPU\u65b9\u5f0f\u52a0\u8f7d\u6a21\u578b
\n # self.tokenizer = AutoTokenizer.from_pretrained(pretrained_model_name_or_path, device_map="cpu", trust_remote_code=True)
\n # self.model = AutoModelForCausalLM.from_pretrained(pretrained_model_name_or_path, device_map="cpu", trust_remote_code=True)
\n # self.model = self.model.float()<\/p>\n
# \u5c1d\u8bd5\u83b7\u53d6\u6a21\u578b\u7684\u5b9e\u9645\u4e0a\u4e0b\u6587\u7a97\u53e3\u5927\u5c0f
\n if hasattr(self.model.config, 'seq_length'):
\n self.context_window = self.model.config.seq_length
\n elif hasattr(self.model.config, 'max_position_embeddings'):
\n self.context_window = self.model.config.max_position_embeddings
\n logger.info(f"Using context window size: {self.context_window}")<\/p>\n
@property
\n def metadata(self) -> LLMMetadata:
\n """Get LLM metadata."""
\n # \u5f97\u5230LLM\u7684\u5143\u6570\u636e
\n return LLMMetadata(
\n context_window=self.context_window,
\n num_output=self.num_output,
\n model_name=self.model_name,
\n )<\/p>\n
@llm_completion_callback()
\n def complete(self, prompt: str, **kwargs: Any) -> CompletionResponse:
\n # \u5b8c\u6210\u51fd\u6570
\n print("\u5b8c\u6210\u51fd\u6570")<\/p>\n
inputs = self.tokenizer.encode(prompt, return_tensors="pt").cuda() # GPU\u65b9\u5f0f
\n # inputs = self.tokenizer.encode(prompt, return_tensors='pt') # CPU\u65b9\u5f0f
\n outputs = self.model.generate(inputs, max_length=self.num_output)
\n response = self.tokenizer.decode(outputs[0])
\n return CompletionResponse(text=response)<\/p>\n
@llm_completion_callback()
\n def stream_complete(self, prompt: str, **kwargs: Any) -> CompletionResponseGen:
\n # \u6d41\u5f0f\u5b8c\u6210\u51fd\u6570
\n print("\u6d41\u5f0f\u5b8c\u6210\u51fd\u6570")<\/p>\n
inputs = self.tokenizer.encode(prompt, return_tensors="pt").cuda() # GPU\u65b9\u5f0f
\n # inputs = self.tokenizer.encode(prompt, return_tensors='pt') # CPU\u65b9\u5f0f
\n outputs = self.model.generate(inputs, max_length=self.num_output)
\n response = self.tokenizer.decode(outputs[0])
\n for token in response:
\n yield CompletionResponse(text=token, delta=token)<\/p>\n
\u5269\u4e0b\u7684\u6b65\u9aa4\u8ddf\u4e4b\u524d\u7684\u8c03\u7528\u65b9\u5f0f\u3001\u4ee3\u7801\u7f16\u7a0b\u6a21\u578b\u51e0\u4e4e\u6ca1\u6709\u4efb\u4f55\u533a\u522b：<\/p>\n
embed_model_path = "\/root\/autodl-tmp\/BAAI\/bge-base-zh-v1.5"
\n pretrained_model_name_or_path = r"\/root\/autodl-tmp\/ZhipuAI\/glm-4-9b-chat"<\/p>\n
# \u8bbe\u7f6eLLM\u548c\u5d4c\u5165\u6a21\u578b
\n logger.info("Setting up LLM and embedding model")
\n Settings.llm = LocalGLM4(pretrained_model_name_or_path)
\n Settings.embed_model = HuggingFaceEmbedding(
\n model_name=f"{embed_model_path}", device="cuda"
\n )<\/p>\n
# \u4ece\u6307\u5b9a\u76ee\u5f55\u52a0\u8f7d\u6587\u6863\u6570\u636e
\n logger.info("Loading documents")
\n documents = SimpleDirectoryReader(input_files=[".\/data\/sample.txt"]).load_data()<\/p>\n
# \u521b\u5efa\u7d22\u5f15\u548c\u67e5\u8be2\u5f15\u64ce
\n logger.info("Creating index and query engine")
\n index = VectorStoreIndex.from_documents(documents)
\n query_engine = index.as_query_engine(streaming=False)<\/p>\n
# \u6267\u884c\u67e5\u8be2
\n logger.info("Executing query")
\n response = query_engine.query(query)<\/p>\n
# \u5904\u7406\u5e76\u8f93\u51fa\u54cd\u5e94
\n if hasattr(response, "response_gen"):
\n # \u6d41\u5f0f\u8f93\u51fa
\n for text in response.response_gen:
\n print(text, end="", flush=True)
\n sys.stdout.flush() # \u786e\u4fdd\u7acb\u5373\u8f93\u51fa
\n else:
\n # \u975e\u6d41\u5f0f\u8f93\u51fa
\n print(response.response, end="", flush=True)<\/p>\n
\u76f8\u5173\u4ee3\u7801\u53ef\u4ee5\u5728\u8fd9\u91cc\u67e5\u770b：github.com\/xiaobox\/lla\u2026<\/p>\n
\u5229\u7528\u79df\u7528\u7684 GPU \u8d44\u6e90\u90e8\u7f72\u4e86\u5f00\u6e90\u5927\u6a21\u578b glm-4-9b-chat ，\u901a\u8fc7\u719f\u6089\u90e8\u7f72\u65b9\u5f0f\u548c\u6d41\u7a0b，\u4f60\u53ef\u4ee5\u7167\u732b\u753b\u864e\u90e8\u7f72\u5176\u4ed6\u5f00\u6e90\u6a21\u578b\u3002\u63a5\u7740\u6211\u4eec\u5c06\u4e4b\u524d RAG \u9879\u76ee\u4e2d\u5bf9LLM\u7684\u8c03\u7528\u6539\u4e3a\u670d\u52a1\u5668\u90e8\u7f72\u7684\u672c\u5730\u5f00\u6e90\u6a21\u578b，\u5b9e\u73b0\u4e86\u6a21\u578b\u548c\u8c03\u7528\u7684\u79c1\u6709\u5316\u3002\u5e0c\u671b\u8fd9\u7bc7\u6587\u7ae0\u80fd\u591f\u5e2e\u52a9\u5230\u6709\u7c7b\u4f3c\u9700\u6c42\u7684\u670b\u53cb\u3002<\/p>\n
\u5927\u6a21\u578b\u65f6\u4ee3，\u706b\u7206\u51fa\u5708\u7684LLM\u5927\u6a21\u578b\u8ba9\u7a0b\u5e8f\u5458\u4eec\u5f00\u59cb\u91cd\u65b0\u8bc4\u4f30\u81ea\u5df1\u7684\u672c\u9886\u3002 \u201cAI\u4f1a\u53d6\u4ee3\u90a3\u4e9b\u884c\u4e1a？\u201d\u201c\u8c01\u7684\u996d\u7897\u53c8\u5c06\u4e0d\u4fdd\u4e86？\u201d\u7b49\u95ee\u9898\u70ed\u8bae\u4e0d\u65ad\u3002<\/p>\n
\u4e8b\u5b9e\u4e0a，\u62a2\u4f60\u996d\u7897\u7684\u4e0d\u662fAI，\u800c\u662f\u4f1a\u5229\u7528AI\u7684\u4eba\u3002<\/p>\n
\u7ee7\u79d1\u5927\u8baf\u98de\u3001\u963f\u91cc\u3001\u534e\u4e3a\u7b49\u5de8\u5934\u516c\u53f8\u53d1\u5e03AI\u4ea7\u54c1\u540e，\u5f88\u591a\u4e2d\u5c0f\u4f01\u4e1a\u4e5f\u9646\u7eed\u8fdb\u573a！\u8d85\u9ad8\u5e74\u85aa，\u6316\u6398AI\u5927\u6a21\u578b\u4eba\u624d！ \u5982\u4eca\u5927\u5382\u8001\u677f\u4eec，\u4e5f\u66f4\u503e\u5411\u4e8e\u4f1aAI\u7684\u4eba，\u666e\u901a\u7a0b\u5e8f\u5458，\u8fd8\u6709\u5e94\u5bf9\u7684\u673a\u4f1a\u5417？<\/p>\n
\u4e0d\u5982\u6210\u4e3a\u300c\u638c\u63e1AI\u5de5\u5177\u7684\u6280\u672f\u4eba\u300d，\u6bd5\u7adfAI\u65f6\u4ee3，\u8c01\u5148\u5c1d\u8bd5，\u8c01\u5c31\u80fd\u5360\u5f97\u5148\u673a！<\/p>\n
\u4f46\u662fLLM\u76f8\u5173\u7684\u5185\u5bb9\u5f88\u591a，\u73b0\u5728\u7f51\u4e0a\u7684\u8001\u8bfe\u7a0b\u8001\u6559\u6750\u5173\u4e8eLLM\u53c8\u592a\u5c11\u3002\u6240\u4ee5\u73b0\u5728\u5c0f\u767d\u5165\u95e8\u5c31\u53ea\u80fd\u9760\u81ea\u5b66，\u5b66\u4e60\u6210\u672c\u548c\u95e8\u69db\u5f88\u9ad8\u3002<\/p>\n
\u9488\u5bf9\u6240\u6709\u81ea\u5b66\u9047\u5230\u56f0\u96be\u7684\u540c\u5b66\u4eec，\u6211\u5e2e\u5927\u5bb6\u7cfb\u7edf\u68b3\u7406\u5927\u6a21\u578b\u5b66\u4e60\u8109\u7edc，\u5c06\u8fd9\u4efd LLM\u5927\u6a21\u578b\u8d44\u6599 \u5206\u4eab\u51fa\u6765：\u5305\u62ecLLM\u5927\u6a21\u578b\u4e66\u7c4d\u3001640\u5957\u5927\u6a21\u578b\u884c\u4e1a\u62a5\u544a\u3001LLM\u5927\u6a21\u578b\u5b66\u4e60\u89c6\u9891\u3001LLM\u5927\u6a21\u578b\u5b66\u4e60\u8def\u7ebf\u3001\u5f00\u6e90\u5927\u6a21\u578b\u5b66\u4e60\u6559\u7a0b\u7b49, 😝\u6709\u9700\u8981\u7684\u5c0f\u4f19\u4f34，\u53ef\u4ee5 \u626b\u63cf\u4e0b\u65b9\u4e8c\u7ef4\u7801\u9886\u53d6🆓\u2193\u2193\u2193<\/p>\n
👉CSDN\u5927\u793c\u5305<\/font>🎁：\u5168\u7f51\u6700\u5168\u300aLLM\u5927\u6a21\u578b\u5165\u95e8+\u8fdb\u9636\u5b66\u4e60\u8d44\u6e90\u5305\u300b\u514d\u8d39\u5206\u4eab（\u5b89\u5168\u94fe\u63a5，\u653e\u5fc3\u70b9\u51fb）<\/font><\/b>👈<\/p>\n \u200b AI\u5927\u6a21\u578b\u5df2\u7ecf\u6210\u4e3a\u4e86\u5f53\u4eca\u79d1\u6280\u9886\u57df\u7684\u4e00\u5927\u70ed\u70b9，\u90a3\u4ee5\u4e0b\u8fd9\u4e9b\u5927\u6a21\u578b\u4e66\u7c4d\u5c31\u662f\u975e\u5e38\u4e0d\u9519\u7684\u5b66\u4e60\u8d44\u6e90\u3002<\/p>\n \u8fd9\u5957\u5305\u542b640\u4efd\u62a5\u544a\u7684\u5408\u96c6，\u6db5\u76d6\u4e86\u5927\u6a21\u578b\u7684\u7406\u8bba\u7814\u7a76\u3001\u6280\u672f\u5b9e\u73b0\u3001\u884c\u4e1a\u5e94\u7528\u7b49\u591a\u4e2a\u65b9\u9762\u3002\u65e0\u8bba\u60a8\u662f\u79d1\u7814\u4eba\u5458\u3001\u5de5\u7a0b\u5e08，\u8fd8\u662f\u5bf9AI\u5927\u6a21\u578b\u611f\u5174\u8da3\u7684\u7231\u597d\u8005，\u8fd9\u5957\u62a5\u544a\u5408\u96c6\u90fd\u5c06\u4e3a\u60a8\u63d0\u4f9b\u5b9d\u8d35\u7684\u4fe1\u606f\u548c\u542f\u793a\u3002(\u51e0\u4e4e\u6db5\u76d6\u6240\u6709\u884c\u4e1a)<\/p>\n \u76ee\u6807：\u4e86\u89e3AI\u5927\u6a21\u578b\u7684\u57fa\u672c\u6982\u5ff5\u3001\u53d1\u5c55\u5386\u7a0b\u548c\u6838\u5fc3\u539f\u7406\u3002<\/p>\n<\/li>\n \u5185\u5bb9：<\/p>\n \u76ee\u6807：\u638c\u63e1AI\u5927\u6a21\u578bAPI\u7684\u4f7f\u7528\u548c\u5f00\u53d1，\u4ee5\u53ca\u76f8\u5173\u7684\u7f16\u7a0b\u6280\u80fd\u3002<\/p>\n<\/li>\n \u5185\u5bb9：<\/p>\n \u76ee\u6807：\u6df1\u5165\u7406\u89e3AI\u5927\u6a21\u578b\u7684\u5e94\u7528\u67b6\u6784，\u5e76\u80fd\u591f\u8fdb\u884c\u79c1\u6709\u5316\u90e8\u7f72\u3002<\/p>\n<\/li>\n \u5185\u5bb9：<\/p>\n \u76ee\u6807：\u638c\u63e1\u591a\u79cdAI\u5927\u6a21\u578b\u7684\u79c1\u6709\u5316\u90e8\u7f72，\u5305\u62ec\u591a\u6a21\u6001\u548c\u7279\u5b9a\u9886\u57df\u6a21\u578b\u3002<\/p>\n<\/li>\n \u5185\u5bb9：<\/p>\n \u8fd9\u4efd LLM\u5927\u6a21\u578b\u8d44\u6599 \u5305\u62ecLLM\u5927\u6a21\u578b\u4e66\u7c4d\u3001640\u5957\u5927\u6a21\u578b\u884c\u4e1a\u62a5\u544a\u3001LLM\u5927\u6a21\u578b\u5b66\u4e60\u89c6\u9891\u3001LLM\u5927\u6a21\u578b\u5b66\u4e60\u8def\u7ebf\u3001\u5f00\u6e90\u5927\u6a21\u578b\u5b66\u4e60\u6559\u7a0b\u7b49, 😝\u6709\u9700\u8981\u7684\u5c0f\u4f19\u4f34，\u53ef\u4ee5 \u626b\u63cf\u4e0b\u65b9\u4e8c\u7ef4\u7801\u9886\u53d6🆓\u2193\u2193\u2193<\/p>\n 👉CSDN\u5927\u793c\u5305<\/font>🎁：\u5168\u7f51\u6700\u5168\u300aLLM\u5927\u6a21\u578b\u5165\u95e8+\u8fdb\u9636\u5b66\u4e60\u8d44\u6e90\u5305\u300b\u514d\u8d39\u5206\u4eab（\u5b89\u5168\u94fe\u63a5，\u653e\u5fc3\u70b9\u51fb）<\/font><\/b>👈<\/p>\n \u200b \u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb964\u6b21\uff0c\u70b9\u8d5e16\u6b21\uff0c\u6536\u85cf20\u6b21\u3002\u672c\u5730\u670d\u52a1\u5668\u90e8\u7f72\u5f00\u6e90\u5927\u6a21\u578b\u6709\u4e00\u4e2a\u524d\u63d0\uff0c\u5c31\u662f\u5f97\u6709 GPU \u663e\u5361\u8d44\u6e90\uff0c\u5728\u6211\u4e0b\u9762\u7684\u4f8b\u5b50\u4e2d\u6211\u79df\u7528\u4e86 autodl \u4e2d\u7684\u7b97\u529b\u8d44\u6e90\uff0c\u5177\u4f53\u662f\u79df\u7528\u4e86\u4e00\u5f20\u6d88\u8d39\u7ea7\u522b\u7684 RTX 3090 \u663e\u5361\u3002_\u5927\u6a21\u578b\u5728\u670d\u52a1\u5668\u90e8\u7f72<\/p>\n","protected":false},"author":2,"featured_media":23359,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[1789,175,347,1788,50,132,1790],"topic":[],"class_list":["post-23367","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-chatglm","tag-chatgpt","tag-llama","tag-llama3","tag-50","tag-132","tag-1790"],"yoast_head":"\n![](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250418233647-6802e20f5411c.png\")
<\/p>\n\u4e00\u3001LLM\u5927\u6a21\u578b\u7ecf\u5178\u4e66\u7c4d<\/h3>\n
![\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250418233647-6802e20f8148e.png\")
<\/p>\n\u4e8c\u3001640\u5957LLM\u5927\u6a21\u578b\u62a5\u544a\u5408\u96c6<\/h3>\n
![\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250418233647-6802e20fd8c93.png\")
<\/p>\n\u4e09\u3001LLM\u5927\u6a21\u578b\u7cfb\u5217\u89c6\u9891\u6559\u7a0b<\/h3>\n
![\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250418233648-6802e21077344.png\")
<\/p>\n\u56db\u3001LLM\u5927\u6a21\u578b\u5f00\u6e90\u6559\u7a0b（LLaLA\/Meta\/chatglm\/chatgpt）<\/h4>\n
![\"\u5728\u8fd9\u91cc\u63d2\u5165\u56fe\u7247\u63cf\u8ff0\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250418233649-6802e211467a4.png\")
<\/p>\nLLM\u5927\u6a21\u578b\u5b66\u4e60\u8def\u7ebf \u2193<\/h2>\n
\u9636\u6bb51：AI\u5927\u6a21\u578b\u65f6\u4ee3\u7684\u57fa\u7840\u7406\u89e3<\/h4>\n
\n
\n
\u9636\u6bb52：AI\u5927\u6a21\u578bAPI\u5e94\u7528\u5f00\u53d1\u5de5\u7a0b<\/h4>\n
\n
\n
\u9636\u6bb53：AI\u5927\u6a21\u578b\u5e94\u7528\u67b6\u6784\u5b9e\u8df5<\/h4>\n
\n
\n
\u9636\u6bb54：AI\u5927\u6a21\u578b\u79c1\u6709\u5316\u90e8\u7f72<\/h4>\n
\n
\n
<\/p>\n","protected":false},"excerpt":{"rendered":"