[Pipeline-Parallelism][TF32] Disable TF32 for Pipeline-Parallel numerical checks (#1382)

* check in

* fancy context style
Co-authored-by: Masaki Kozuki <mkozuki@nvidia.com>

tests/L0/run_transformer/test_pipeline_parallel_fwd_bwd.py

@@ -6,6 +6,7 @@ import unittest
 
 import torch
 from torch.testing._internal import common_utils
+from torch.testing._internal import common_cuda
 
 logging.getLogger("torch").setLevel(logging.WARNING)
 
@@ -138,6 +139,7 @@ class PipelineParallelForwardBackwardTestBase:
 
             pp_utils.update_num_microbatches(0)
             
+            with common_cuda.tf32_off():
                 loss = fwd_bwd_func(
                     testing_utils.fwd_step_func,
                     batch,