[Bug] Fix batch invariant in torch 2.10 (#30907)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
Signed-off-by: Wentao Ye <44945378+yewentao256@users.noreply.github.com>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>

vllm/model_executor/layers/batch_invariant.py

@@ -933,8 +933,6 @@ def enable_batch_invariant_mode():
     _batch_invariant_MODE = True
     _batch_invariant_LIB = torch.library.Library("aten", "IMPL")
 
-    # Batch invariant matmuls are no longer needed after cublas overrides
-    if not is_torch_equal_or_newer("2.10.0.dev"):
     if (
         current_platform.is_device_capability_family(100)
         or current_platform.is_device_capability(80)
@@ -949,9 +947,7 @@ def enable_batch_invariant_mode():
     else:
         # Only source of batch invariance for Hopper is split-k, can disable through
         # cuBLAS workspace config
-            _original_cublas_workspace_cfg = os.environ.get(
-                "CUBLAS_WORKSPACE_CONFIG", None
-            )
+        _original_cublas_workspace_cfg = os.environ.get("CUBLAS_WORKSPACE_CONFIG", None)
         _original_cublaslt_workspace_size = os.environ.get(
             "CUBLASLT_WORKSPACE_SIZE", None
         )