diff --git a/miles/rollout/session/session_server.py b/miles/rollout/session/session_server.py
index 07b90417ca..6e48c10b35 100644
--- a/miles/rollout/session/session_server.py
+++ b/miles/rollout/session/session_server.py
@@ -15,7 +15,6 @@
 import setproctitle
 import uvicorn
 from fastapi import FastAPI, Request
-from fastapi.responses import JSONResponse
 from starlette.responses import Response
 
 from miles.rollout.session.sessions import get_worker_stats, setup_session_routes
@@ -95,25 +94,21 @@ async def do_proxy(
     def build_proxy_response(self, result: dict) -> Response:
         content = result["response_body"]
         status_code = result["status_code"]
-        # Strip framing headers so JSONResponse / Response recompute them
-        # from the actual rendered body. Forwarding upstream's content-length
-        # verbatim breaks uvicorn h11 with "Too much data for declared
-        # Content-Length" whenever our re-serialization differs in even one
-        # byte. Mirrors the strip already done on the request path in do_proxy.
-        # Also strip "server": uvicorn adds its own Server header; passing
-        # the upstream one through produces two Server headers, which strict
-        # HTTP parsers (aiohttp/llhttp via litellm) reject as malformed.
+
         headers = {
             k: v
             for k, v in result["headers"].items()
             if k.lower() not in ("content-length", "transfer-encoding", "server")
         }
-        content_type = headers.get("content-type", "")
-        try:
-            data = orjson.loads(content)
-            return JSONResponse(content=data, status_code=status_code, headers=headers)
-        except (orjson.JSONDecodeError, UnicodeDecodeError):
-            return Response(content=content, status_code=status_code, headers=headers, media_type=content_type)
+
+        content_type = headers.get("content-type")
+
+        return Response(
+            content=content,
+            status_code=status_code,
+            headers=headers,
+            media_type=content_type,
+        )
 
 
 async def _stats_logger_loop(worker_port, interval_seconds: float = 30.0):
diff --git a/miles/rollout/session/sessions.py b/miles/rollout/session/sessions.py
index ae6c866361..aee310ad56 100644
--- a/miles/rollout/session/sessions.py
+++ b/miles/rollout/session/sessions.py
@@ -296,7 +296,7 @@ async def chat_completions(request: Request, session_id: str):
                 else:
                     return backend.build_proxy_response(result)
 
-            response = orjson.loads(result["response_body"])
+            response = await asyncio.to_thread(orjson.loads, result["response_body"])
 
             choice = response.get("choices", [{}])[0]