fix and add doc

python/tvm/contrib/cutlass/build.py

@@ -312,6 +312,12 @@ def tune_cutlass_kernels(
         Wheter or not use slower but very accurate (compared to tf32) 3xtf32 mode for
         fp32 inputs on tensorcore.
 
+    split_k_slices : list of int
+        Split factor candidates for split-K GEMM. If split-K > 1, the GEMM K-loop is computed in
+        parallel accross split-K blocks, and a seperate global reduction kernel is launched to
+        accumulate partial reductions. The profiler will pick the best split-k factor from the
+        given candidate list. Note that the larger split-K factor requires a larger workspace.
+
     profile_all_alignments : bool
         When True, profile all kernal variants with smaller alignments than the largest possible.
 

tests/python/relay/test_op_grad_level2.py

@@ -268,13 +268,12 @@ def verify_conv2d_backward_weight(
 
 
 def test_conv2d_backward_weight():
-    # verify_conv2d_backward_weight((2, 8, 32, 32), (2, 4, 32, 32), (3, 3), (1, 1), (1, 1))
-    # verify_conv2d_backward_weight((2, 16, 15, 15), (2, 3, 32, 32), (3, 3), (2, 2), (0, 0))
+    verify_conv2d_backward_weight((2, 8, 32, 32), (2, 4, 32, 32), (3, 3), (1, 1), (1, 1))
+    verify_conv2d_backward_weight((2, 16, 15, 15), (2, 3, 32, 32), (3, 3), (2, 2), (0, 0))
     verify_conv2d_backward_weight(
         (1, 16, 32, 32), (1, 16, 32, 32), (3, 3), (1, 1), (1, 1), groups=16, out_channels=16
     )
 
 
 if __name__ == "__main__":
-    # pytest.main([__file__])
-    test_conv2d_backward_weight()
+    pytest.main([__file__])