diff --git a/fastdeploy/engine/engine.py b/fastdeploy/engine/engine.py
index 01f28819f..d550a3701 100644
--- a/fastdeploy/engine/engine.py
+++ b/fastdeploy/engine/engine.py
@@ -483,9 +483,6 @@ class LLMEngine:
             if self.cfg.scheduler_config.splitwise_role == "prefill":
                 variables["FLAGS_fmt_write_cache_completed_signal"] = 1
 
-        if self.cfg.model_config.enable_mm:
-            variables["FLAGS_max_partition_size"] = 1024
-
         command_prefix = ""
         for k, v in variables.items():
             command_prefix += f"{k}={v} "
diff --git a/fastdeploy/model_executor/layers/attention/append_attn_backend.py b/fastdeploy/model_executor/layers/attention/append_attn_backend.py
index 346251a30..14562c3f7 100644
--- a/fastdeploy/model_executor/layers/attention/append_attn_backend.py
+++ b/fastdeploy/model_executor/layers/attention/append_attn_backend.py
@@ -148,6 +148,9 @@ class AppendAttentionBackend(AttentionBackend):
         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", 1024))
+        # split kv still has bug in speculative decoding
+        if self.speculative_method is not None:
+            self.max_partition_size = self.max_seq_len
         self.encoder_block_shape_q: int = encoder_block_shape_q
         self.decoder_block_shape_q: int = decoder_block_shape_q