Update server.py

server.py

@@ -652,10 +652,10 @@ async def completions(request: Request):
                     continue
                 buf.extend(chunk_bytes)
                 if len(buf) >= threshold:
-                    yield bytes(buf)
+                    yield b"data: " + bytes(buf)
                     buf.clear()
             if buf:
-                yield bytes(buf)
+                yield b"data: " + bytes(buf)
             yield b"data: [DONE]\n\n"
         else:
             text = await service.generate(req)