Fix magicoder model tokenizer issue and remove codegen streaming redu…

…ndant end format (#1086)

* Fix magicoder tokenizer issue and streaming redundant end format

Signed-off-by: lvliang-intel <liang1.lv@intel.com>

intel_extension_for_transformers/neural_chat/models/model_utils.py

@@ -566,7 +566,8 @@ def load_model(
         logging.error(f"Exception: {e}")
         raise ValueError(f"load_model: an unexpected error occurred, {e}")
 
-    if re.search("llama", model.config.architectures[0], re.IGNORECASE):
+    if re.search("llama", model.config.architectures[0], re.IGNORECASE) and \
+       not re.search("magicoder", model_name, re.IGNORECASE):
         # unwind broken decapoda-research config
         model.generation_config.pad_token_id = 0
         model.generation_config.bos_token_id = 1

intel_extension_for_transformers/neural_chat/server/multi_cpu_server.py

@@ -250,31 +250,13 @@ async def chat_completion_endpoint(request: ChatCompletionRequest):
             if attr == "stream":
                 continue
             setattr(gen_config, attr, value)
-        buffered_texts = ""
         if request.stream:
             generator, _ = chatbot.predict_stream(query=request.prompt, config=gen_config)
             if not isinstance(generator, types.GeneratorType):
                 generator = (generator,)
             def stream_generator():
-                nonlocal buffered_texts
                 for output in generator:
-                    if isinstance(output, str):
-                        chunks = output.split()
-                        for chunk in chunks:
-                            ret = {
-                                "text": chunk,
-                                "error_code": 0,
-                            }
-                            buffered_texts += chunk + ' '
-                            yield json.dumps(ret).encode() + b"\0"
-                    else:
-                        ret = {
-                            "text": output,
-                            "error_code": 0,
-                        }
-                        buffered_texts += output + ' '
-                        yield json.dumps(ret).encode() + b"\0"
-                yield f"data: [DONE]\n\n"
+                    yield output + "\0"
             return StreamingResponse(stream_generator(), media_type="text/event-stream")
         else:
             response = chatbot.predict(query=request.prompt, config=gen_config)

intel_extension_for_transformers/neural_chat/server/multi_hpu_server.py

@@ -289,7 +289,6 @@ async def chat_completion_endpoint(request: ChatCompletionRequest):
             def stream_generator():
                 for output in generator:
                     yield output + "\0"
-                yield f"data: [DONE]\n\n"
             return StreamingResponse(stream_generator(), media_type="text/event-stream")
         else:
             response = chatbot.predict(query=request.prompt, config=gen_config)