[PyTorch] Get `skip_fp8_weight_update` only in CUDA Graph Capturing (#1854)

only get skip_fp8_weight_update in fp8_graph_capturing
Signed-off-by: Xin Yao <xiny@nvidia.com>
Co-authored-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

transformer_engine/pytorch/module/grouped_linear.py

@@ -668,7 +668,10 @@ class GroupedLinear(TransformerEngineBaseModule):
         ), "GroupedLinear doesn't support input tensor in FP8."
         assert len(m_splits) == self.num_gemms, "Number of splits should match number of GEMMs."
 
+        if FP8GlobalStateManager.fp8_graph_capturing():
             skip_fp8_weight_update = FP8GlobalStateManager.get_skip_fp8_weight_update_tensor()
+        else:
+            skip_fp8_weight_update = None
         if skip_fp8_weight_update is not None:
             is_first_microbatch = False