diff --git a/src/engine_args.py b/src/engine_args.py
index 45e50d1..3131172 100644
--- a/src/engine_args.py
+++ b/src/engine_args.py
@@ -4,6 +4,7 @@ import logging
 from torch.cuda import device_count
 from vllm import AsyncEngineArgs
 from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
+from src.utils import convert_limit_mm_per_prompt
 
 RENAME_ARGS_MAP = {
     "MODEL_NAME": "model",
@@ -89,7 +90,8 @@ DEFAULT_ARGS = {
     "qlora_adapter_name_or_path": os.getenv('QLORA_ADAPTER_NAME_OR_PATH', None),
     "disable_logprobs_during_spec_decoding": os.getenv('DISABLE_LOGPROBS_DURING_SPEC_DECODING', None),
     "otlp_traces_endpoint": os.getenv('OTLP_TRACES_ENDPOINT', None),
-    "use_v2_block_manager": os.getenv('USE_V2_BLOCK_MANAGER', 'true')
+    "use_v2_block_manager": os.getenv('USE_V2_BLOCK_MANAGER', 'true'),
+    "limit_mm_per_prompt": convert_limit_mm_per_prompt(os.getenv('LIMIT_MM_PER_PROMPT', "image=1"))
 }
 
 def match_vllm_args(args):
diff --git a/src/utils.py b/src/utils.py
index bfc8ce9..85af1e8 100644
--- a/src/utils.py
+++ b/src/utils.py
@@ -15,6 +15,10 @@ except ImportError:
 
 logging.basicConfig(level=logging.INFO)
 
+def convert_limit_mm_per_prompt(input_string: str):
+    key, value = input_string.split('=')
+    return {key: int(value)}
+
 def count_physical_cores():
     with open('/proc/cpuinfo') as f:
         content = f.readlines()