Add NVTX ranges to FP8 amax AR and grad output preprocessing (NVIDIA#1530)

Add NVTX ranges

Signed-off-by: Jaemin Choi <jaeminc@nvidia.com>
Co-authored-by: Jaemin Choi <jaeminc@nvidia.com>
Co-authored-by: Tim Moon <4406448+timmoon10@users.noreply.github.com>

Author: minitu

transformer_engine/pytorch/module/layernorm_linear.py

@@ -522,6 +522,7 @@ def backward(
 
             if ctx.grad_output_quantizer is not None:
                 ctx.grad_output_quantizer.set_usage(rowwise=True, columnwise=True)
+            nvtx_range_push(f"{nvtx_label}.grad_output_preprocess")
             (
                 grad_output,
                 grad_bias,
@@ -531,6 +532,7 @@ def backward(
                 ctx.parallel_mode == "row",
                 ctx.grad_output_quantizer,
             )
+            nvtx_range_pop(f"{nvtx_label}.grad_output_preprocess")
 
             # Prepare GEMM input
             # Note: Perform tensor-parallel communication if needed
@@ -747,7 +749,9 @@ def backward(
             wgrad = None
 
         if ctx.reduce_and_update_bwd_fp8_tensors and not is_graph_capturing():
+            nvtx_range_push(f"{nvtx_label}.reduce_and_update_fp8_tensors")
             FP8GlobalStateManager.reduce_and_update_fp8_tensors(forward=False)
+            nvtx_range_pop(f"{nvtx_label}.reduce_and_update_fp8_tensors")
 
         # Scatter fp8 weight buffers
         # if ctx.fp8 and not isinstance(weight, QuantizedTensor):

transformer_engine/pytorch/module/linear.py

@@ -427,6 +427,7 @@ def backward(ctx, grad_output: torch.Tensor) -> Tuple[Union[torch.Tensor, None],
             # Note: Cast to expected dtype and perform tensor-parallel communication
             if ctx.grad_output_quantizer is not None:
                 ctx.grad_output_quantizer.set_usage(rowwise=True, columnwise=True)
+            nvtx_range_push(f"{nvtx_label}.grad_output_preprocess")
             (
                 grad_output,
                 grad_bias,
@@ -436,6 +437,7 @@ def backward(ctx, grad_output: torch.Tensor) -> Tuple[Union[torch.Tensor, None],
                 ctx.parallel_mode == "row",
                 ctx.grad_output_quantizer,
             )
+            nvtx_range_pop(f"{nvtx_label}.grad_output_preprocess")
 
             # Prepare input tensor
             # Note: Perform tensor-parallel communication if needed
@@ -623,7 +625,9 @@ def backward(ctx, grad_output: torch.Tensor) -> Tuple[Union[torch.Tensor, None],
             wgrad = None
 
         if ctx.reduce_and_update_bwd_fp8_tensors and not is_graph_capturing():
+            nvtx_range_push(f"{nvtx_label}.reduce_and_update_fp8_tensors")
             FP8GlobalStateManager.reduce_and_update_fp8_tensors(forward=False)
+            nvtx_range_pop(f"{nvtx_label}.reduce_and_update_fp8_tensors")
 
         # Scatter fp8 weight buffers
         if ctx.fp8 and not isinstance(weight, QuantizedTensor):