diff --git a/atom/plugin/vllm/attention_backend/attention_gdn.py b/atom/plugin/vllm/attention_backend/attention_gdn.py
index b6158a086..a47f16431 100644
--- a/atom/plugin/vllm/attention_backend/attention_gdn.py
+++ b/atom/plugin/vllm/attention_backend/attention_gdn.py
@@ -395,8 +395,8 @@ def forward(
                     query=query_non_spec,
                     key=key_non_spec,
                     value=value_non_spec,
-                    a=a,
-                    b=b,
+                    a=a.unsqueeze(1),
+                    b=b.unsqueeze(1),
                     dt_bias=self.dt_bias,
                     A_log=self.A_log,
                     indices=non_spec_state_indices_tensor,
diff --git a/atom/utils/envs.py b/atom/utils/envs.py
index fb487bc10..8403947a1 100644
--- a/atom/utils/envs.py
+++ b/atom/utils/envs.py
@@ -89,7 +89,7 @@
     "ATOM_USE_CUSTOM_ALL_GATHER": lambda: (
         os.getenv("ATOM_USE_CUSTOM_ALL_GATHER", "1").lower() == "1"
     ),
-    "ATOM_USE_FLYDSL_GDR": lambda: os.getenv("ATOM_USE_FLYDSL_GDR", "0").lower() == "1",
+    "ATOM_USE_FLYDSL_GDR": lambda: os.getenv("ATOM_USE_FLYDSL_GDR", "1").lower() == "1",
     # --- MoE (DeepSeek-style shared experts) ---
     # Dual-stream MoE only when num_tokens <= threshold; 0 disables dual-stream registration.
     "ATOM_DUAL_STREAM_MOE_TOKEN_THRESHOLD": lambda: int(