diff --git a/multimodal/dashinfer_vlm/api_server/config.py b/multimodal/dashinfer_vlm/api_server/config.py
index 09d2dd9c9..ba34ff821 100644
--- a/multimodal/dashinfer_vlm/api_server/config.py
+++ b/multimodal/dashinfer_vlm/api_server/config.py
@@ -58,7 +58,7 @@ def add_context_args(parser):
         choices=["cuda"],
         help="device (Default: cuda)",
     )
-    group.add_argument("--max_length", type=int, default=32000, help="model max length")
+    group.add_argument("--max_length", type=int, default=128000, help="model max length")
     group.add_argument("--max_batch", type=int, default=128, help="max batch")
     group.add_argument(
         "--parallel_size",
diff --git a/multimodal/dashinfer_vlm/api_server/server.py b/multimodal/dashinfer_vlm/api_server/server.py
index 266a44ae7..288116dbb 100644
--- a/multimodal/dashinfer_vlm/api_server/server.py
+++ b/multimodal/dashinfer_vlm/api_server/server.py
@@ -117,12 +117,16 @@ def init():
     context.set("tokenizer", tokenizer)
     cuda_devices = [str(i) for i in range(tensor_parallel_size)]
     compute_unit = "CUDA:" + ",".join(cuda_devices)
+    if hasattr(config, "max_position_embeddings"):
+        engine_max_len = min(config.max_position_embeddings, context.get("max_length"))
+    else:
+        engine_max_len = context.get("max_length")
     # allspark model config
     as_model_config = allspark.AsModelConfig(
         model_name=model_name,
         model_path=as_graph_path,
         weights_path=as_weight_path,
-        engine_max_length=context.get("max_length"),
+        engine_max_length=engine_max_len,
         engine_max_batch=context.get("max_batch"),
         compute_unit=compute_unit,
         enable_prefix_cache=True if context.get("enable_prefix_cache") else False,
diff --git a/multimodal/dashinfer_vlm/vl_inference/runtime/qwen_vl.py b/multimodal/dashinfer_vlm/vl_inference/runtime/qwen_vl.py
index e7c6b8fc9..1071b1b34 100644
--- a/multimodal/dashinfer_vlm/vl_inference/runtime/qwen_vl.py
+++ b/multimodal/dashinfer_vlm/vl_inference/runtime/qwen_vl.py
@@ -267,8 +267,8 @@ def __init__(
             "QWEN2-AL": "audio_attention_mask",
             "GUMMY-AL": "audio_attention_mask",
         }
-        self.max_input_len = int(getenv("DS_LLM_MAX_IN_TOKENS", "20000"))
-        self.max_total_len = int(getenv("DS_LLM_MAX_TOKENS", "32000"))
+        self.max_total_len = as_config.engine_max_length
+        self.max_input_len = self.max_total_len - 1
 
     """
     支持input_tokens格式: