diff --git a/benchmark/benchmark_serving.py b/benchmark/benchmark_serving.py
index d9250a0e..bad79ee6 100644
--- a/benchmark/benchmark_serving.py
+++ b/benchmark/benchmark_serving.py
@@ -102,7 +102,7 @@ async def query_model_vllm(prompt, verbose, ip_ports):
                 if verbose:
                     print('Done')
 
-                output = await resp.json()
+                output = await resp.json(content_type='text/plain')
                 # necessary for latency calc
                 output['response_len'] = expected_response_len
                 if verbose and 'generated_text' in output:
diff --git a/tests/e2e_test/test_e2e.py b/tests/e2e_test/test_e2e.py
index cc6b404d..3e928611 100644
--- a/tests/e2e_test/test_e2e.py
+++ b/tests/e2e_test/test_e2e.py
@@ -36,7 +36,7 @@ async def get_llumnix_response(prompt, sampling_params, ip_ports):
 
     async with aiohttp.ClientSession(timeout=timeout) as session:
         async with session.post(f'http://{ip_ports}/generate', json=request) as resp:
-            output = await resp.json()
+            output = await resp.json('text/plain')
             return output
 
 prompts = [