v4.4 tag release update. (#3032)

examples/python/CuTeDSL/blackwell/dense_gemm_alpha_beta_persistent.py

@@ -29,13 +29,11 @@
 import argparse
 from typing import Optional, Tuple, Type, Union
 
-import torch
 import cuda.bindings.driver as cuda
 
 import cutlass
 import cutlass.cute as cute
 import cutlass.cute.testing as testing
-import cutlass.torch as cutlass_torch
 import cutlass.utils as utils
 import cutlass.pipeline as pipeline
 from cutlass.pipeline import pipeline_init_arrive, pipeline_init_wait
@@ -1176,7 +1174,10 @@ class SM100PersistentDenseGemmAlphaBetaKernel:
                         tSR_sC[(None, None, None, c_pipeline_consumer_state.index)],
                         tSR_rC,
                     )
-                    cute.arch.fence_proxy("async.shared", space="cta")
+                    cute.arch.fence_proxy(
+                        "async.shared",
+                        space="cta",
+                    )
                     c_pipeline.consumer_release(c_pipeline_consumer_state)
 
                     # Advance pipeline states
@@ -1203,7 +1204,10 @@ class SM100PersistentDenseGemmAlphaBetaKernel:
                         tiled_copy_r2s, tRS_rD, tRS_sD[(None, None, None, d_buffer)]
                     )
                     # Fence and barrier to make sure shared memory store is visible to TMA store
-                    cute.arch.fence_proxy("async.shared", space="cta")
+                    cute.arch.fence_proxy(
+                        "async.shared",
+                        space="cta",
+                    )
                     epilog_sync_barrier.arrive_and_wait()
 
                     #
@@ -1875,6 +1879,9 @@ class SM100PersistentDenseGemmAlphaBetaKernel:
 
 
 def create_tensors(l, m, n, k, a_major, b_major, cd_major, ab_dtype, c_dtype, d_dtype):
+    import torch
+    import cutlass.torch as cutlass_torch
+
     torch.manual_seed(1111)
 
     a_torch_cpu = cutlass_torch.matrix(l, m, k, a_major == "m", ab_dtype)
@@ -1991,6 +1998,9 @@ def compare(
     beta,
     tolerance,
 ):
+    import torch
+    import cutlass.torch as cutlass_torch
+
     # Copy gpu result back
     kernel_result = d_torch_gpu.cpu()
 
@@ -2057,6 +2067,8 @@ def run_dense_gemm(
     # Unpack parameters
     m, n, k, l = mnkl
 
+    import torch
+
     if not torch.cuda.is_available():
         raise RuntimeError("GPU is required to run this example!")