Disable Cascade Attention for Batch Invariance (#32561)

Signed-off-by: frankwang28 <frank.wbb@hotmail.com>
Signed-off-by: Frank Wang <41319051+frankwang28@users.noreply.github.com>
Co-authored-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>

vllm/config/vllm.py

@@ -959,6 +959,18 @@ class VllmConfig:
                     "when cudagraph_mode piecewise cudagraphs is used, "
                     f"cudagraph_mode={self.compilation_config.cudagraph_mode}"
                 )
+        from vllm.model_executor.layers.batch_invariant import vllm_is_batch_invariant
+
+        if (
+            self.model_config
+            and vllm_is_batch_invariant()
+            and not self.model_config.disable_cascade_attn
+        ):
+            self.model_config.disable_cascade_attn = True
+            logger.warning_once(
+                "Disabling cascade attention when VLLM_BATCH_INVARIANT is enabled.",
+                scope="local",
+            )
 
         if self.parallel_config.use_ubatching:
             a2a_backend = self.parallel_config.all2all_backend