fixed a sse stream issue

koboldcpp.py

@@ -70,7 +70,7 @@ dry_seq_break_max = 128
 extra_images_max = 4 # for kontext/qwen img
 
 # global vars
-KcppVersion = "1.109"
+KcppVersion = "1.109.1"
 showdebug = True
 kcpp_instance = None #global running instance
 global_memory = {"tunnel_url": "", "restart_target":"", "input_to_exit":False, "load_complete":False, "restart_override_config_target":""}
@@ -3678,6 +3678,8 @@ class KcppServerRequestHandler(http.server.SimpleHTTPRequestHandler):
                                     event_str = json.dumps({"token": tokenStr, "finish_reason":None})
                                     await self.send_kai_sse_event(event_str)
                                 tokenStr = "" # now the final finish reason can be sent alone
+                                if delta and 'role' in delta:
+                                    delta = {'role':delta["role"],'content':''}
                             if api_format == 4:  # if oai chat, set format to expected openai streaming response
                                 if streamDone and ("logprobs" in genparams and genparams["logprobs"]): # this is a hack that sends an extra message containing ALL the logprobs
                                     lastlogprobs = handle.last_logprobs()