Update app.py

app.py

@@ -31,6 +31,7 @@ client = httpx.AsyncClient(
 async def forward_request(
     request: Request,
     path: str,
+    body_bytes: Optional[bytes] = None,
     stream: bool = False
 ):
     """Forward request to target endpoint"""
@@ -38,8 +39,9 @@ async def forward_request(
     # Build target URL
     target_url = f"{TARGET_BASE_URL.rstrip('/')}/{path.lstrip('/')}"
     
-    # Get request body
-    body = await request.body()
+    # Get request body if not provided
+    if body_bytes is None:
+        body_bytes = await request.body()
     
     # Prepare headers
     headers = dict(request.headers)
@@ -49,26 +51,14 @@ async def forward_request(
     
     # Forward the request with original headers (including the user's API key)
     try:
-        if stream:
-            # For streaming, we return a streaming response
-            response = await client.request(
-                method=request.method,
-                url=target_url,
-                headers=headers,
-                content=body,
-                follow_redirects=True,
-            )
-            return response
-        else:
-            # For non-streaming, just get the response
-            response = await client.request(
-                method=request.method,
-                url=target_url,
-                headers=headers,
-                content=body,
-                follow_redirects=True,
-            )
-            return response
+        response = await client.request(
+            method=request.method,
+            url=target_url,
+            headers=headers,
+            content=body_bytes,
+            follow_redirects=True,
+        )
+        return response
     except httpx.TimeoutException:
         raise HTTPException(status_code=504, detail="Gateway Timeout")
     except httpx.RequestError as e:
@@ -86,24 +76,20 @@ async def stream_response(response: httpx.Response):
 async def proxy_all(path: str, request: Request):
     """Main proxy endpoint - catches all requests"""
     
-    # Special handling for streaming completions
+    # Read the body once
+    body_bytes = await request.body()
+    
+    # Check if streaming is requested (only for POST requests to completion endpoints)
     is_streaming = False
-    if path.endswith("completions") or path.endswith("chat/completions"):
-        # Check if streaming is requested
-        if request.method == "POST":
-            body = await request.body()
-            try:
-                data = json.loads(body)
-                is_streaming = data.get("stream", False)
-            except:
-                pass
-            # Re-create request scope - we need to re-read body
-            # Create a new request with the same body
-            new_request = Request(request.scope, receive=request.receive)
-            # Forward with streaming if needed
-            response = await forward_request(new_request, path, stream=is_streaming)
-    else:
-        response = await forward_request(request, path, stream=False)
+    if request.method == "POST" and (path.endswith("completions") or path.endswith("chat/completions")):
+        try:
+            data = json.loads(body_bytes)
+            is_streaming = data.get("stream", False)
+        except:
+            pass
+    
+    # Forward the request
+    response = await forward_request(request, path, body_bytes, stream=is_streaming)
     
     # Handle streaming responses
     if is_streaming:
@@ -140,9 +126,9 @@ async def root():
     }
 
 @app.get("/v1/models")
-async def list_models():
+async def list_models(request: Request):
     """Proxy for models list"""
-    response = await forward_request(Request({"method": "GET", "url": "/v1/models"}), "v1/models")
+    response = await forward_request(request, "v1/models")
     content = await response.aread()
     return Response(
         content=content,