diff --git a/fastdeploy/engine/engine.py b/fastdeploy/engine/engine.py
index 1d6103176..c508f4ee5 100644
--- a/fastdeploy/engine/engine.py
+++ b/fastdeploy/engine/engine.py
@@ -377,7 +377,7 @@ class LLMEngine:
             "PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION": "python",
             "FLAGS_use_append_attn": 1,
             "NCCL_ALGO": "Ring",
-            "FLAGS_max_partition_size": int(os.getenv("FLAGS_max_partition_size", 32768)),
+            "FLAGS_max_partition_size": int(os.getenv("FLAGS_max_partition_size", 1024)),
             "FLAGS_hardamard_moe_block_size": int(os.getenv("FLAGS_hardamard_moe_block_size", 128)),
             "FLAGS_hardamard_use_diagonal_block_matrix": int(
                 os.getenv("FLAGS_hardamard_use_diagonal_block_matrix", 0)
diff --git a/fastdeploy/model_executor/layers/attention/append_attn_backend.py b/fastdeploy/model_executor/layers/attention/append_attn_backend.py
index 551e19e59..029764c63 100644
--- a/fastdeploy/model_executor/layers/attention/append_attn_backend.py
+++ b/fastdeploy/model_executor/layers/attention/append_attn_backend.py
@@ -113,7 +113,7 @@ class AppendAttentionBackend(AttentionBackend):
         self.group_size: int = self.num_heads // self.kv_num_heads
         self.head_dim: int = fd_config.model_config.head_dim
         self.num_layers: int = fd_config.model_config.num_hidden_layers
-        self.max_partition_size: int = int(os.getenv("FLAGS_max_partition_size", 32768))
+        self.max_partition_size: int = int(os.getenv("FLAGS_max_partition_size", 1024))
         self.encoder_block_shape_q: int = encoder_block_shape_q
         self.decoder_block_shape_q: int = decoder_block_shape_q