diff --git a/openllm-python/src/openllm/_runners.py b/openllm-python/src/openllm/_runners.py
index 5c43efe5..9b069f06 100644
--- a/openllm-python/src/openllm/_runners.py
+++ b/openllm-python/src/openllm/_runners.py
@@ -93,7 +93,7 @@ class CTranslateRunnable(bentoml.Runnable):
         spaces_between_special_tokens=False,
         clean_up_tokenization_spaces=True,  #
       )
-      yield GenerationOutput(
+      out = GenerationOutput(
         prompt_token_ids=prompt_token_ids,  #
         prompt='',
         finished=request_output.is_last,
@@ -109,7 +109,7 @@ class CTranslateRunnable(bentoml.Runnable):
           )
         ],
       ).model_dump_json()
-
+      yield bentoml.io.SSE(out).marshal()
 
 @registry
 class vLLMRunnable(bentoml.Runnable):
@@ -286,7 +286,7 @@ class PyTorchRunnable(bentoml.Runnable):
         if config['logprobs']:
           sample_logprobs.append({token: token_logprobs})
 
-        yield GenerationOutput(
+        out = GenerationOutput(
           prompt='',
           finished=False,
           outputs=[
@@ -303,13 +303,14 @@ class PyTorchRunnable(bentoml.Runnable):
           prompt_logprobs=prompt_logprobs if config['prompt_logprobs'] else None,
           request_id=request_id,
         ).model_dump_json()
+        yield bentoml.io.SSE(out).marshal()
         if stopped:
           break
       else:
         finish_reason = 'length'
       if stopped:
         finish_reason = 'stop'
-      yield GenerationOutput(
+      out = GenerationOutput(
         prompt='',
         finished=True,
         outputs=[
@@ -326,6 +327,7 @@ class PyTorchRunnable(bentoml.Runnable):
         prompt_logprobs=prompt_logprobs if config['prompt_logprobs'] else None,
         request_id=request_id,
       ).model_dump_json()
+      yield bentoml.io.SSE(out).marshal()
 
     # Clean
     del past_key_values, out