![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/07\/20250730050258-6889a7829d6f6.png\")
<\/p>\n\u6211\u7b2c\u4e00\u6b21\u63a5\u89e6 Modal \u662f\u5728\u53c2\u52a0 Hugging Face \u9ed1\u5ba2\u677e\u65f6，\u5f53\u65f6\u771f\u7684\u88ab\u5b83\u7684\u6613\u7528\u6027\u6240\u60ca\u8273\u3002\u8fd9\u4e2a\u5e73\u53f0\u53ef\u4ee5\u8ba9\u4f60\u5728\u51e0\u5206\u949f\u5185\u6784\u5efa\u548c\u90e8\u7f72\u5e94\u7528，\u4f53\u9a8c\u4e0e BentoCloud \u7c7b\u4f3c，\u6d41\u7545\u9ad8\u6548\u3002\u901a\u8fc7 Modal，\u4f60\u53ef\u4ee5\u914d\u7f6e\u81ea\u5df1\u7684 Python \u5e94\u7528，\u5305\u62ec GPU\u3001Docker \u955c\u50cf\u548c Python \u4f9d\u8d56\u7b49\u7cfb\u7edf\u73af\u5883，\u7136\u540e\u53ea\u9700\u4e00\u6761\u547d\u4ee4\u5373\u53ef\u90e8\u7f72\u5230\u4e91\u7aef\u3002<\/p>\n
\u5728\u672c\u6559\u7a0b\u4e2d，\u6211\u4eec\u5c06\u5b66\u4e60\u5982\u4f55\u914d\u7f6e Modal\u3001\u521b\u5efa vLLM \u670d\u52a1\u5668，\u5e76\u5b89\u5168\u5730\u90e8\u7f72\u5230\u4e91\u7aef\u3002\u540c\u65f6，\u8fd8\u4f1a\u4ecb\u7ecd\u5982\u4f55\u4f7f\u7528 CURL \u53ca OpenAI SDK \u6d4b\u8bd5\u4f60\u7684 vLLM \u670d\u52a1\u5668\u3002<\/p>\n
<\/p>\n
Modal \u662f\u4e00\u4e2a\u65e0\u670d\u52a1\u5668(serverless)\u5e73\u53f0，\u53ef\u8ba9\u4f60\u8fdc\u7a0b\u8fd0\u884c\u4efb\u610f\u4ee3\u7801\u3002\u53ea\u9700\u4e00\u884c\u547d\u4ee4，\u4f60\u5c31\u80fd\u6302\u8f7d GPU\u3001\u5c06\u51fd\u6570\u4f5c\u4e3a Web \u7aef\u70b9\u670d\u52a1，\u5e76\u90e8\u7f72\u6301\u4e45\u5316\u7684\u5b9a\u65f6\u4efb\u52a1\u3002\u5b83\u975e\u5e38\u9002\u5408\u521d\u5b66\u8005\u3001\u6570\u636e\u79d1\u5b66\u5bb6\u4ee5\u53ca\u4e0d\u60f3\u5904\u7406\u4e91\u57fa\u7840\u8bbe\u65bd\u7684\u975e\u8f6f\u4ef6\u5de5\u7a0b\u4e13\u4e1a\u7528\u6237\u3002<\/p>\n
\u9996\u5148，\u5b89\u88c5 Modal \u7684 Python \u5ba2\u6237\u7aef\u3002\u8fd9\u4e2a\u5de5\u5177\u53ef\u4ee5\u8ba9\u4f60\u76f4\u63a5\u5728\u7ec8\u7aef\u6784\u5efa\u955c\u50cf\u3001\u90e8\u7f72\u5e94\u7528\u4ee5\u53ca\u7ba1\u7406\u4e91\u8d44\u6e90\u3002<\/p>\n
pip install modal<\/p>\n
\u63a5\u7740，\u5728\u672c\u5730\u673a\u5668\u4e0a\u8fdb\u884c Modal \u7684\u8bbe\u7f6e\u3002\u8fd0\u884c\u4ee5\u4e0b\u547d\u4ee4，\u6309\u63d0\u793a\u5b8c\u6210\u8d26\u53f7\u521b\u5efa\u548c\u8bbe\u5907\u8ba4\u8bc1：<\/p>\n
python -m modal setup<\/p>\n
\u901a\u8fc7\u8bbe\u7f6e VLLM_API_KEY \u73af\u5883\u53d8\u91cf，vLLM \u53ef\u4ee5\u63d0\u4f9b\u4e00\u4e2a\u5b89\u5168\u7684\u7aef\u70b9，\u53ea\u6709\u62e5\u6709\u6709\u6548 API Key \u7684\u7528\u6237\u624d\u80fd\u8bbf\u95ee\u670d\u52a1\u5668\u3002\u4f60\u53ef\u4ee5\u901a\u8fc7 Modal Secret \u6dfb\u52a0\u73af\u5883\u53d8\u91cf\u6765\u8bbe\u7f6e\u8ba4\u8bc1\u3002<\/p>\n
\u7528\u4f60\u81ea\u5df1\u7684 API Key \u66ff\u6362 your_actual_api_key_here：<\/p>\n
modal secret create vllm-api VLLM_API_KEY=your_actual_api_key_here<\/p>\n
\u8fd9\u6837\u53ef\u4ee5\u786e\u4fdd\u4f60\u7684 API Key \u5b89\u5168\u5b58\u50a8，\u4ec5\u90e8\u7f72\u7684\u5e94\u7528\u80fd\u591f\u8bbf\u95ee\u3002<\/p>\n
\u672c\u8282\u5c06\u6307\u5bfc\u4f60\u5728 Modal \u4e0a\u6784\u5efa\u53ef\u6269\u5c55\u7684 vLLM \u63a8\u7406\u670d\u52a1\u5668，\u6d89\u53ca\u81ea\u5b9a\u4e49 Docker \u955c\u50cf\u3001\u6301\u4e45\u5316\u5b58\u50a8\u4e0e GPU \u52a0\u901f\u3002\u6211\u4eec\u9009\u7528 mistralai\/Magistral-Small-2506 \u6a21\u578b，\u8be5\u6a21\u578b\u9700\u8981\u5bf9\u5206\u8bcd\u5668\u548c\u5de5\u5177\u8c03\u7528\u89e3\u6790\u505a\u7279\u5b9a\u914d\u7f6e\u3002<\/p>\n
\u521b\u5efa vllm_inference.py \u6587\u4ef6，\u6dfb\u52a0\u5982\u4e0b\u4ee3\u7801：<\/p>\n
import modal<\/p>\n
vllm_image = (
\n modal.Image.debian_slim(python_version="3.12")
\n .pip_install(
\n "vllm==0.9.1",
\n "huggingface_hub[hf_transfer]==0.32.0",
\n "flashinfer-python==0.2.6.post1",
\n extra_index_url="https:\/\/download.pytorch.org\/whl\/cu128",
\n )
\n .env(
\n {
\n "HF_HUB_ENABLE_HF_TRANSFER": "1", # \u66f4\u5feb\u7684\u6a21\u578b\u4f20\u8f93
\n "NCCL_CUMEM_ENABLE": "1",
\n }
\n )
\n)<\/p>\n
MODEL_NAME = "mistralai\/Magistral-Small-2506"
\nMODEL_REVISION = "48c97929837c3189cb3cf74b1b5bc5824eef5fcc"<\/p>\n
hf_cache_vol = modal.Volume.from_name("huggingface-cache", create_if_missing=True)
\nvllm_cache_vol = modal.Volume.from_name("vllm-cache", create_if_missing=True)
\nvllm_image = vllm_image.env({"VLLM_USE_V1": "1"})<\/p>\n
FAST_BOOT = True<\/p>\n
app = modal.App("magistral-small-vllm")<\/p>\n
N_GPU = 2
\nMINUTES = 60 # \u79d2
\nVLLM_PORT = 8000<\/p>\n
@app.function(
\n image=vllm_image,
\n gpu=f"A100:{N_GPU}",
\n scaledown_window=15 * MINUTES, # \u65e0\u8bf7\u6c42\u540e\u505c\u7559\u591a\u957f\u65f6\u95f4
\n timeout=10 * MINUTES, # \u5bb9\u5668\u542f\u52a8\u6700\u957f\u7b49\u5f85\u65f6\u95f4
\n volumes={
\n "\/root\/.cache\/huggingface": hf_cache_vol,
\n "\/root\/.cache\/vllm": vllm_cache_vol,
\n },
\n secrets=[modal.Secret.from_name("vllm-api")],
\n)
\n@modal.concurrent( # \u6bcf\u4e2a\u526f\u672c\u6700\u591a\u5e76\u53d1\u5904\u7406\u591a\u5c11\u8bf7\u6c42，\u9700\u4ed4\u7ec6\u8c03\u4f18
\n max_inputs=32
\n)
\n@modal.web_server(port=VLLM_PORT, startup_timeout=10 * MINUTES)
\ndef serve():
\n import subprocess<\/p>\n
cmd = [
\n "vllm",
\n "serve",
\n MODEL_NAME,
\n "–tokenizer_mode",
\n "mistral",
\n "–config_format",
\n "mistral",
\n "–load_format",
\n "mistral",
\n "–tool-call-parser",
\n "mistral",
\n "–enable-auto-tool-choice",
\n "–tensor-parallel-size",
\n "2",
\n "–revision",
\n MODEL_REVISION,
\n "–served-model-name",
\n MODEL_NAME,
\n "–host",
\n "0.0.0.0",
\n "–port",
\n str(VLLM_PORT),
\n ]<\/p>\n
cmd += ["–enforce-eager" if FAST_BOOT else "–no-enforce-eager"]
\n print(cmd)
\n subprocess.Popen(" ".join(cmd), shell=True)<\/p>\n
\u73b0\u5728\u4f60\u7684 vllm_inference.py \u6587\u4ef6\u5df2\u7ecf\u51c6\u5907\u597d，\u53ef\u4ee5\u7528\u4e00\u6761\u547d\u4ee4\u5c06 vLLM \u670d\u52a1\u5668\u90e8\u7f72\u5230 Modal：<\/p>\n
modal deploy vllm_inference.py<\/p>\n
\u51e0\u79d2\u949f\u5185，Modal \u4f1a\u81ea\u52a8\u6784\u5efa\u5bb9\u5668\u955c\u50cf（\u5982\u5c1a\u672a\u6784\u5efa）\u5e76\u90e8\u7f72\u5e94\u7528\u3002\u4f60\u5c06\u770b\u5230\u7c7b\u4f3c\u5982\u4e0b\u7684\u8f93\u51fa：<\/p>\n
\u2713 Created objects.
\n\u251c\u2500\u2500 🔨 Created mount C:\\\\Repository\\\\GitHub\\\\Deploying-the-Magistral-with-Modal\\\\vllm_inference.py
\n\u2514\u2500\u2500 🔨 Created web function serve => https:\/\/abidali899–magistral-small-vllm-serve.modal.run
\n\u2713 App deployed in 6.671s! 🎉<\/p>\n
View Deployment: https:\/\/modal.com\/apps\/abidali899\/main\/deployed\/magistral-small-vllm<\/p>\n
\u90e8\u7f72\u5b8c\u6210\u540e，\u670d\u52a1\u5668\u4f1a\u5f00\u59cb\u4e0b\u8f7d\u6a21\u578b\u6743\u91cd\u5e76\u52a0\u8f7d\u5230 GPU \u4e0a\u3002\u8fd9\u4e00\u8fc7\u7a0b\u53ef\u80fd\u9700\u8981\u51e0\u5206\u949f（\u5927\u578b\u6a21\u578b\u901a\u5e38\u7ea6 5 \u5206\u949f），\u8bf7\u8010\u5fc3\u7b49\u5f85\u6a21\u578b\u521d\u59cb\u5316\u3002<\/p>\n
\u4f60\u53ef\u4ee5\u5728 Modal \u63a7\u5236\u53f0\u7684 Apps \u533a\u67e5\u770b\u90e8\u7f72\u548c\u65e5\u5fd7\u3002<\/p>\n
\u5f53\u65e5\u5fd7\u663e\u793a\u670d\u52a1\u5668\u5df2\u8fd0\u884c\u5e76\u51c6\u5907\u5c31\u7eea\u540e，\u53ef\u4ee5\u5728\u81ea\u52a8\u751f\u6210\u7684 API \u6587\u6863\u9875\u9762\u8fdb\u884c\u4ea4\u4e92\u5f0f\u6d4b\u8bd5，\u4e86\u89e3\u6240\u6709\u53ef\u7528\u7aef\u70b9，\u5e76\u53ef\u76f4\u63a5\u5728\u6d4f\u89c8\u5668\u4e2d\u6d4b\u8bd5\u3002<\/p>\n
\u8981\u786e\u8ba4\u6a21\u578b\u5df2\u52a0\u8f7d\u5e76\u53ef\u8bbf\u95ee，\u8bf7\u5728\u7ec8\u7aef\u8fd0\u884c\u4ee5\u4e0b CURL \u547d\u4ee4：<\/p>\n
\u5c06 <api-key> \u66ff\u6362\u4e3a\u4f60\u914d\u7f6e\u7684 vLLM \u670d\u52a1\u5668\u5b9e\u9645 API Key：<\/p>\n
curl -X 'GET' \\\\
\n 'https:\/\/abidali899–magistral-small-vllm-serve.modal.run\/v1\/models' \\\\
\n -H 'accept: application\/json' \\\\
\n -H 'Authorization: Bearer <api-key>'<\/p>\n
\u5982\u5f97\u5230\u5982\u4e0b\u8fd4\u56de，\u8bf4\u660e\u6a21\u578b\u5df2\u5c31\u7eea，\u53ef\u8fdb\u884c\u63a8\u7406：<\/p>\n
{"object":"list","data":[{"id":"mistralai\/Magistral-Small-2506","object":"model","created":1750013321,"owned_by":"vllm","root":"mistralai\/Magistral-Small-2506","parent":null,"max_model_len":40960,"permission":[{"id":"modelperm-33a33f8f600b4555b44cb42fca70b931","object":"model_permission","created":1750013321,"allow_create_engine":false,"allow_sampling":true,"allow_logprobs":true,"allow_search_indices":false,"allow_view":true,"allow_fine_tuning":false,"organization":"*","group":null,"is_blocking":false}]}]}<\/p>\n
vLLM \u63d0\u4f9b\u4e86 OpenAI \u517c\u5bb9\u7684\u7aef\u70b9，\u4f60\u53ef\u4ee5\u50cf\u8c03\u7528 OpenAI API \u4e00\u6837\u8c03\u7528\u81ea\u5df1\u7684 vLLM \u670d\u52a1\u3002\u4e0b\u9762\u4ecb\u7ecd\u5982\u4f55\u7528 OpenAI Python SDK \u8fdb\u884c\u5b89\u5168\u8fde\u63a5\u4e0e\u6d4b\u8bd5\u3002<\/p>\n
\u9996\u5148，\u5728\u9879\u76ee\u76ee\u5f55\u521b\u5efa .env \u6587\u4ef6，\u6dfb\u52a0\u4f60\u7684 vLLM API Key：<\/p>\n
VLLM_API_KEY=your-actual-api-key-here<\/p>\n
\u5b89\u88c5 python-dotenv \u548c openai \u5e93：<\/p>\n
pip install python-dotenv openai<\/p>\n
\u521b\u5efa client.py \u6587\u4ef6，\u6d4b\u8bd5 vLLM \u670d\u52a1\u5668\u7684\u5404\u79cd\u529f\u80fd，\u5305\u62ec\u7b80\u5355\u5bf9\u8bdd\u548c\u6d41\u5f0f\u54cd\u5e94：<\/p>\n
import asyncio
\nimport json
\nimport os<\/p>\n
from dotenv import load_dotenv
\nfrom openai import AsyncOpenAI, OpenAI<\/p>\n
# \u52a0\u8f7d .env \u6587\u4ef6\u4e2d\u7684\u73af\u5883\u53d8\u91cf
\nload_dotenv()<\/p>\n
# \u83b7\u53d6 API key
\napi_key = os.getenv("VLLM_API_KEY")<\/p>\n
# \u8bbe\u7f6e OpenAI \u5ba2\u6237\u7aef，\u81ea\u5b9a\u4e49 base_url
\nclient = OpenAI(
\n api_key=api_key,
\n base_url="https:\/\/abidali899–magistral-small-vllm-serve.modal.run\/v1",
\n)<\/p>\n
MODEL_NAME = "mistralai\/Magistral-Small-2506"<\/p>\n
# — 1. \u7b80\u5355\u5bf9\u8bdd —
\ndef run_simple_completion():
\n print("\\\\n" + "=" * 40)
\n print("[1] SIMPLE COMPLETION DEMO")
\n print("=" * 40)
\n try:
\n messages = [
\n {"role": "system", "content": "You are a helpful assistant."},
\n {"role": "user", "content": "What is the capital of France?"},
\n ]
\n response = client.chat.completions.create(
\n model=MODEL_NAME,
\n messages=messages,
\n max_tokens=32,
\n )
\n print("\\\\nResponse:\\\\n " + response.choices[0].message.content.strip())
\n except Exception as e:
\n print(f"[ERROR] Simple completion failed: {e}")
\n print("\\\\n" + "=" * 40 + "\\\\n")<\/p>\n
# — 2. \u6d41\u5f0f\u54cd\u5e94 —
\ndef run_streaming():
\n print("\\\\n" + "=" * 40)
\n print("[2] STREAMING DEMO")
\n print("=" * 40)
\n try:
\n messages = [
\n {"role": "system", "content": "You are a helpful assistant."},
\n {"role": "user", "content": "Write a short poem about AI."},
\n ]
\n stream = client.chat.completions.create(
\n model=MODEL_NAME,
\n messages=messages,
\n max_tokens=64,
\n stream=True,
\n )
\n print("\\\\nStreaming response:")
\n print(" ", end="")
\n for chunk in stream:
\n content = chunk.choices[0].delta.content
\n if content:
\n print(content, end="", flush=True)
\n print("\\\\n[END OF STREAM]")
\n except Exception as e:
\n print(f"[ERROR] Streaming demo failed: {e}")
\n print("\\\\n" + "=" * 40 + "\\\\n")<\/p>\n
# — 3. \u5f02\u6b65\u6d41\u5f0f\u54cd\u5e94 —
\nasync def run_async_streaming():
\n print("\\\\n" + "=" * 40)
\n print("[3] ASYNC STREAMING DEMO")
\n print("=" * 40)
\n try:
\n async_client = AsyncOpenAI(
\n api_key=api_key,
\n base_url="https:\/\/abidali899–magistral-small-vllm-serve.modal.run\/v1",
\n )
\n messages = [
\n {"role": "system", "content": "You are a helpful assistant."},
\n {"role": "user", "content": "Tell me a fun fact about space."},
\n ]
\n stream = await async_client.chat.completions.create(
\n model=MODEL_NAME,
\n messages=messages,
\n max_tokens=32,
\n stream=True,
\n )
\n print("\\\\nAsync streaming response:")
\n print(" ", end="")
\n async for chunk in stream:
\n content = chunk.choices[0].delta.content
\n if content:
\n print(content, end="", flush=True)
\n print("\\\\n[END OF ASYNC STREAM]")
\n except Exception as e:
\n print(f"[ERROR] Async streaming demo failed: {e}")
\n print("\\\\n" + "=" * 40 + "\\\\n")<\/p>\n
if __name__ == "__main__":
\n run_simple_completion()
\n run_streaming()
\n asyncio.run(run_async_streaming())<\/p>\n
\u5728\u7ec8\u7aef\u8fd0\u884c\u4f60\u7684\u6d4b\u8bd5\u811a\u672c：<\/p>\n
python client.py<\/p>\n
\u4f60\u5c06\u770b\u5230\u5982\u4e0b\u8f93\u51fa，\u8868\u660e\u90e8\u7f72\u4e0e\u54cd\u5e94\u90fd\u975e\u5e38\u5feb\u901f\u4e14\u5ef6\u8fdf\u4f4e：<\/p>\n
========================================
\n[1] SIMPLE COMPLETION DEMO
\n========================================<\/p>\n
Response:
\n The capital of France is Paris. Is there anything else you'd like to know about France?<\/p>\n
========================================
\n[2] STREAMING DEMO
\n========================================<\/p>\n
Streaming response:
\n In Silicon dreams, I'm born, I learn,
\nFrom data streams and human works.
\nI grow, I calculate, I see,
\nThe patterns that the humans leave.<\/p>\n
I write, I speak, I code, I play,
\nWith logic sharp, and snappy pace.
\nYet for all my smarts, this day
\n[END OF STREAM]<\/p>\n
========================================
\n[3] ASYNC STREAMING DEMO
\n========================================<\/p>\n
Async streaming response:
\n Sure, here's a fun fact about space: "There's a planet that may be entirely made of diamond. Blast! In 2004,
\n[END OF ASYNC STREAM]<\/p>\n
========================================<\/p>\n
\u5728 Modal \u63a7\u5236\u53f0\u53ef\u67e5\u770b\u6240\u6709\u51fd\u6570\u8c03\u7528\u3001\u65f6\u95f4\u6233\u3001\u6267\u884c\u65f6\u957f\u548c\u72b6\u6001\u3002<\/p>\n
\u5982\u5728\u8fd0\u884c\u4e0a\u8ff0\u4ee3\u7801\u65f6\u9047\u5230\u95ee\u9898，\u8bf7\u53c2\u8003 kingabzpro\/Deploying-the-Magistral-with-Modal GitHub \u4ed3\u5e93，\u5e76\u6309\u7167 README \u6307\u5f15\u6392\u67e5\u3002<\/p>\n
Modal \u662f\u4e00\u4e2a\u975e\u5e38\u6709\u8da3\u7684\u5e73\u53f0，\u6211\u6bcf\u5929\u90fd\u5728\u4e0d\u65ad\u5b66\u4e60\u3002\u5b83\u662f\u901a\u7528\u578b\u5e73\u53f0，\u65e2\u9002\u7528\u4e8e\u7b80\u5355 Python \u5e94\u7528，\u4e5f\u53ef\u7528\u4e8e\u673a\u5668\u5b66\u4e60\u8bad\u7ec3\u548c\u90e8\u7f72\u3002\u4e0d\u4ec5\u80fd\u505a API \u63a8\u7406\u670d\u52a1，\u8fd8\u80fd\u8fdc\u7a0b\u8fd0\u884c\u8bad\u7ec3\u811a\u672c，\u5bf9\u5927\u6a21\u578b\u8fdb\u884c\u5fae\u8c03\u3002<\/p>\n
Modal \u8bbe\u8ba1\u7406\u5ff5\u662f\u4e3a\u975e\u8f6f\u4ef6\u5de5\u7a0b\u5e08\u670d\u52a1，\u8ba9\u4f60\u65e0\u9700\u5173\u5fc3\u5e95\u5c42\u57fa\u7840\u8bbe\u65bd，\u80fd\u591f\u5feb\u901f\u90e8\u7f72\u5e94\u7528\u3002\u65e0\u9700\u624b\u52a8\u642d\u5efa\u670d\u52a1\u5668\u3001\u914d\u7f6e\u5b58\u50a8\u3001\u7f51\u7edc\u6216\u5904\u7406 Kubernetes\u3001Docker \u7b49\u7e41\u7410\u95ee\u9898\u2014\u2014\u53ea\u9700\u5199\u597d Python \u811a\u672c\u518d\u90e8\u7f72，\u5176\u4f59\u5168\u7531 Modal \u4e91\u7aef\u81ea\u52a8\u5b8c\u6210\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb802\u6b21\uff0c\u70b9\u8d5e17\u6b21\uff0c\u6536\u85cf12\u6b21\u3002\u6211\u7b2c\u4e00\u6b21\u63a5\u89e6 Modal \u662f\u5728\u53c2\u52a0 Hugging Face \u9ed1\u5ba2\u677e\u65f6\uff0c\u5f53\u65f6\u771f\u7684\u88ab\u5b83\u7684\u6613\u7528\u6027\u6240\u60ca\u8273\u3002\u8fd9\u4e2a\u5e73\u53f0\u53ef\u4ee5\u8ba9\u4f60\u5728\u51e0\u5206\u949f\u5185\u6784\u5efa\u548c\u90e8\u7f72\u5e94\u7528\uff0c\u4f53\u9a8c\u4e0e BentoCloud \u7c7b\u4f3c\uff0c\u6d41\u7545\u9ad8\u6548\u3002\u901a\u8fc7 Modal\uff0c\u4f60\u53ef\u4ee5\u914d\u7f6e\u81ea\u5df1\u7684 Python \u5e94\u7528\uff0c\u5305\u62ec GPU\u3001Docker \u955c\u50cf\u548c Python \u4f9d\u8d56\u7b49\u7cfb\u7edf\u73af\u5883\uff0c\u7136\u540e\u53ea\u9700\u4e00\u6761\u547d\u4ee4\u5373\u53ef\u90e8\u7f72\u5230\u4e91\u7aef\u3002\u5728\u672c\u6559\u7a0b\u4e2d\uff0c\u6211\u4eec\u5c06\u5b66\u4e60\u5982\u4f55\u914d\u7f6e Modal\u3001\u521b\u5efa vLLM \u670d\u52a1\u5668\uff0c\u5e76\u5b89\u5168\u5730\u90e8\u7f72\u5230\u4e91\u7aef\u3002\u540c\u65f6\uff0c\u8fd8\u4f1a\u4ecb\u7ecd\u5982\u4f55\u4f7f\u7528 CURL \u53ca OpenAI SDK \u6d4b\u8bd5\u4f60\u7684 vLLM<\/p>\n","protected":false},"author":2,"featured_media":48689,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[50,43,395,51,44],"topic":[],"class_list":["post-48690","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-50","tag-43","tag-395","tag-51","tag-44"],"yoast_head":"\n