Implementation:Vllm project Vllm Connections

Knowledge Sources	vllm
Domains	HTTP, Infrastructure
Last Updated	2026-02-08 00:00 GMT

Overview

Provides HTTP connection utilities for fetching remote resources with both synchronous and asynchronous client support.

Description

This module defines the HTTPConnection class that wraps requests.Session (synchronous) and aiohttp.ClientSession (asynchronous) for making HTTP requests. It provides convenience methods for fetching bytes, text, JSON, and downloading files, with URL validation, custom vLLM user-agent headers, timeout control, and optional session reuse for connection pooling. A module-level global_http_connection singleton is provided for shared use across vLLM.

Usage

Use this module when vLLM needs to download remote model weights, configuration files, tokenizer assets, or any other HTTP-accessible resources. The global instance provides a centralized, reusable HTTP client with proper error handling and consistent request headers.

Code Reference

Source Location

• Repository: vllm
• File: vllm/connections.py
• Lines: 1-189

Signature

class HTTPConnection:
    def __init__(self, *, reuse_client: bool = True) -> None: ...
    def get_sync_client(self) -> requests.Session: ...
    async def get_async_client(self) -> aiohttp.ClientSession: ...
    def get_response(self, url: str, *, stream: bool = False,
                     timeout: float | None = None, ...) -> requests.Response: ...
    async def get_async_response(self, url: str, *, timeout: float | None = None,
                                  ...) -> aiohttp.ClientResponse: ...
    def get_bytes(self, url: str, ...) -> bytes: ...
    async def async_get_bytes(self, url: str, ...) -> bytes: ...
    def get_text(self, url: str, ...) -> str: ...
    async def async_get_text(self, url: str, ...) -> str: ...
    def get_json(self, url: str, ...) -> str: ...
    async def async_get_json(self, url: str, ...) -> str: ...
    def download_file(self, url: str, save_path: Path, ...) -> Path: ...
    async def async_download_file(self, url: str, save_path: Path, ...) -> Path: ...

global_http_connection = HTTPConnection()

Import

from vllm.connections import HTTPConnection, global_http_connection

I/O Contract

Inputs

Name	Type	Required	Description
url	str	Yes	HTTP or HTTPS URL to fetch from
reuse_client	bool	No	Whether to reuse HTTP session across requests (default True)
stream	bool	No	Enable streaming response for large downloads (default False)
timeout	float or None	No	Request timeout in seconds (default None for no timeout)
extra_headers	Mapping[str, str] or None	No	Additional HTTP headers to include
allow_redirects	bool	No	Whether to follow HTTP redirects (default True)
save_path	Path	download_file only	Local file path for saving downloaded content
chunk_size	int	No	Chunk size in bytes for file downloads (default 128)

Outputs

Name	Type	Description
bytes	bytes	Raw response content (get_bytes / async_get_bytes)
text	str	Response body as text (get_text / async_get_text)
json	str	Parsed JSON response (get_json / async_get_json)
path	Path	Path to downloaded file (download_file / async_download_file)

Usage Examples

from vllm.connections import global_http_connection

# Synchronous: fetch JSON configuration from a remote URL
config = global_http_connection.get_json(
    "https://huggingface.co/meta-llama/Llama-2-7b/resolve/main/config.json",
    timeout=30.0,
)

# Synchronous: download a file to disk
from pathlib import Path
saved = global_http_connection.download_file(
    "https://example.com/model.safetensors",
    save_path=Path("/tmp/model.safetensors"),
    timeout=300.0,
    chunk_size=8192,
)

# Asynchronous: fetch text content
import asyncio
async def fetch():
    text = await global_http_connection.async_get_text(
        "https://example.com/tokenizer.json"
    )
    return text

Related Pages

• Environment:Vllm_project_Vllm_Python