|
| 1 | +# Copyright (c) ModelScope Contributors. All rights reserved. |
| 2 | +""" |
| 3 | +Proxy utilities for forwarding requests to internal services. |
| 4 | +
|
| 5 | +This module provides HTTP proxy functionality to route requests from the Tinker server |
| 6 | +to appropriate model or sampler services based on base_model routing. |
| 7 | +""" |
| 8 | + |
| 9 | +from __future__ import annotations |
| 10 | + |
| 11 | +import httpx |
| 12 | +import os |
| 13 | +from fastapi import Request, Response |
| 14 | +from typing import Any |
| 15 | + |
| 16 | +from twinkle.utils.logger import get_logger |
| 17 | + |
| 18 | +logger = get_logger() |
| 19 | + |
| 20 | + |
| 21 | +class ServiceProxy: |
| 22 | + """HTTP proxy for routing requests to internal model and sampler services. |
| 23 | +
|
| 24 | + This proxy handles: |
| 25 | + 1. URL construction using localhost to avoid external routing loops |
| 26 | + 2. Header forwarding with appropriate cleanup |
| 27 | + 3. Debug logging for troubleshooting |
| 28 | + 4. Error handling and response forwarding |
| 29 | + """ |
| 30 | + |
| 31 | + def __init__( |
| 32 | + self, |
| 33 | + http_options: dict[str, Any] | None = None, |
| 34 | + route_prefix: str = '/api/v1', |
| 35 | + ): |
| 36 | + """Initialize the service proxy. |
| 37 | +
|
| 38 | + Args: |
| 39 | + http_options: HTTP server options (host, port) for internal routing |
| 40 | + route_prefix: URL prefix for routing (default: '/api/v1') |
| 41 | + """ |
| 42 | + self.http_options = http_options or {} |
| 43 | + self.route_prefix = route_prefix |
| 44 | + # Disable proxy for internal requests to avoid routing through external proxies |
| 45 | + self.client = httpx.AsyncClient(timeout=None, trust_env=False) |
| 46 | + |
| 47 | + def _build_target_url(self, service_type: str, base_model: str, endpoint: str) -> str: |
| 48 | + """Build the target URL for internal service routing. |
| 49 | +
|
| 50 | + Constructs URLs using localhost to avoid extra external hops. |
| 51 | + When requests come from www.modelscope.com/twinkle, we proxy to |
| 52 | + localhost:port directly instead of back to modelscope.com. |
| 53 | +
|
| 54 | + Args: |
| 55 | + service_type: Either 'model' or 'sampler' |
| 56 | + base_model: The base model name for routing |
| 57 | + endpoint: The target endpoint name |
| 58 | +
|
| 59 | + Returns: |
| 60 | + Complete target URL for the internal service |
| 61 | + """ |
| 62 | + prefix = self.route_prefix.rstrip('/') if self.route_prefix else '' |
| 63 | + host = self.http_options.get('host', 'localhost') |
| 64 | + port = self.http_options.get('port', 8000) |
| 65 | + |
| 66 | + # Use localhost for internal routing |
| 67 | + if host == '0.0.0.0': |
| 68 | + host = 'localhost' |
| 69 | + |
| 70 | + base_url = f'http://{host}:{port}' |
| 71 | + return f'{base_url}{prefix}/{service_type}/{base_model}/{endpoint}' |
| 72 | + |
| 73 | + def _prepare_headers(self, request_headers) -> dict[str, str]: |
| 74 | + """Prepare headers for proxying by removing problematic headers. |
| 75 | +
|
| 76 | + Args: |
| 77 | + request_headers: Original request headers (case-insensitive from FastAPI) |
| 78 | +
|
| 79 | + Returns: |
| 80 | + Cleaned headers safe for proxying |
| 81 | + """ |
| 82 | + logger.debug('prepare_headers request_headers=%s', request_headers) |
| 83 | + # Convert to dict while preserving case-insensitive lookups for special headers |
| 84 | + headers = dict(request_headers) |
| 85 | + # Remove headers that should not be forwarded |
| 86 | + headers.pop('host', None) |
| 87 | + headers.pop('content-length', None) |
| 88 | + # Add serve_multiplexed_model_id for sticky sessions if present |
| 89 | + # Use case-insensitive lookup from original request_headers |
| 90 | + request_id = request_headers.get('X-Ray-Serve-Request-Id') |
| 91 | + if request_id is not None: |
| 92 | + headers['serve_multiplexed_model_id'] = request_id |
| 93 | + return headers |
| 94 | + |
| 95 | + async def proxy_request( |
| 96 | + self, |
| 97 | + request: Request, |
| 98 | + endpoint: str, |
| 99 | + base_model: str, |
| 100 | + service_type: str, |
| 101 | + ) -> Response: |
| 102 | + """Generic proxy method to forward requests to model or sampler services. |
| 103 | +
|
| 104 | + This method consolidates the common proxy logic for both model and sampler endpoints. |
| 105 | +
|
| 106 | + Args: |
| 107 | + request: The incoming FastAPI request |
| 108 | + endpoint: The target endpoint name (e.g., 'create_model', 'asample') |
| 109 | + base_model: The base model name for routing |
| 110 | + service_type: Either 'model' or 'sampler' to determine the target service |
| 111 | +
|
| 112 | + Returns: |
| 113 | + Proxied response from the target service |
| 114 | + """ |
| 115 | + body_bytes = await request.body() |
| 116 | + target_url = self._build_target_url(service_type, base_model, endpoint) |
| 117 | + # Pass original request.headers (case-insensitive) instead of dict conversion |
| 118 | + headers = self._prepare_headers(request.headers) |
| 119 | + |
| 120 | + try: |
| 121 | + # Debug logging for troubleshooting proxy issues |
| 122 | + logger.debug( |
| 123 | + 'proxy_request service=%s endpoint=%s target_url=%s request_id=%s', |
| 124 | + service_type, |
| 125 | + endpoint, |
| 126 | + target_url, |
| 127 | + headers.get('serve_multiplexed_model_id'), |
| 128 | + ) |
| 129 | + |
| 130 | + # Forward the request to the target service |
| 131 | + response = await self.client.request( |
| 132 | + method=request.method, |
| 133 | + url=target_url, |
| 134 | + content=body_bytes, |
| 135 | + headers=headers, |
| 136 | + params=request.query_params, |
| 137 | + ) |
| 138 | + |
| 139 | + # Debug logging for response |
| 140 | + logger.debug( |
| 141 | + 'proxy_response status=%s body_preview=%s', |
| 142 | + response.status_code, |
| 143 | + response.text[:200], |
| 144 | + ) |
| 145 | + |
| 146 | + return Response( |
| 147 | + content=response.content, |
| 148 | + status_code=response.status_code, |
| 149 | + headers=dict(response.headers), |
| 150 | + media_type=response.headers.get('content-type'), |
| 151 | + ) |
| 152 | + except Exception as e: |
| 153 | + logger.error('Proxy error: %s', str(e), exc_info=True) |
| 154 | + return Response(content=f'Proxy Error: {str(e)}', status_code=502) |
| 155 | + |
| 156 | + async def proxy_to_model(self, request: Request, endpoint: str, base_model: str) -> Response: |
| 157 | + """Proxy request to model endpoint. |
| 158 | +
|
| 159 | + Routes the request to the appropriate model deployment based on base_model. |
| 160 | +
|
| 161 | + Args: |
| 162 | + request: The incoming FastAPI request |
| 163 | + endpoint: The target endpoint name (e.g., 'create_model', 'forward') |
| 164 | + base_model: The base model name for routing |
| 165 | +
|
| 166 | + Returns: |
| 167 | + Proxied response from the model service |
| 168 | + """ |
| 169 | + return await self.proxy_request(request, endpoint, base_model, 'model') |
| 170 | + |
| 171 | + async def proxy_to_sampler(self, request: Request, endpoint: str, base_model: str) -> Response: |
| 172 | + """Proxy request to sampler endpoint. |
| 173 | +
|
| 174 | + Routes the request to the appropriate sampler deployment based on base_model. |
| 175 | +
|
| 176 | + Args: |
| 177 | + request: The incoming FastAPI request |
| 178 | + endpoint: The target endpoint name (e.g., 'asample') |
| 179 | + base_model: The base model name for routing |
| 180 | +
|
| 181 | + Returns: |
| 182 | + Proxied response from the sampler service |
| 183 | + """ |
| 184 | + return await self.proxy_request(request, endpoint, base_model, 'sampler') |
0 commit comments