Upgrade to CUTLASS 4.2.1 (#18372)

jasl · web-flow · commit ddf7bce8c7ff · 2025-10-17T14:28:59.000-04:00
* Upgrade to CUTLASS 4.2.1

* Fix test: mbarrier.try_wait requires .target sm_90 or higher
diff --git a/3rdparty/cutlass b/3rdparty/cutlass
@@ -1 +1 @@
-Subproject commit b2dd65dc864e09688245b316ac46c4a6cd07e15c
+Subproject commit f3fde58372d33e9a5650ba7b80fc48b3b49d40c8
diff --git a/tests/python/tir-base/test_tir_ptx_cp_async.py b/tests/python/tir-base/test_tir_ptx_cp_async.py
@@ -95,7 +95,7 @@ def ptx_cp_async_barrier(
             B[tx, i] = A_shared[tx, i]
 
 
-@pytest.mark.xfail(reason="temp skip test due to cuda env update")
+@tvm.testing.requires_cuda_compute_version(9)
 def test_ptx_cp_async_barrier():
     f = ptx_cp_async_barrier
 
diff --git a/tests/python/tir-transform/test_tir_transform_inject_ptx_async_copy.py b/tests/python/tir-transform/test_tir_transform_inject_ptx_async_copy.py
@@ -214,8 +214,7 @@ def ptx_global_to_shared_copy_fp32x1_barrier(
             B[tx, i] = A_shared[tx, i]
 
 
-@pytest.mark.xfail(reason="temp skip test due to cuda env update")
-@tvm.testing.requires_cuda
+@tvm.testing.requires_cuda_compute_version(9)
 def test_inject_async_copy_barrier():
     dtype = "float32"
     vec_size = 1
