[PyTorch] Remove is_first_microbatch setting after cudagraph warmup

[buptzyb]

Description

Reset is_first_microbatch to True is unnecessary, because later we'll set it back to False: https://github.com/NVIDIA/TransformerEngine/blob/release_v2.12/transformer_engine/pytorch/module/layernorm_mlp.py#L2037-L2042

is_first_microbatch has two functions: the first is to control fp8 cast-transpose behavior, which is overtaken by the skip_fp8_weight_update tensor in graph replay. The second is to control whether the fused wgrad should clear main_grad buffer before making the accumulation. We should always capture the no-clear version in case the user captures one graph for all microbatches.

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

Please list the changes introduced in this PR:

• Change A
• Change B

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[greptile-apps]

Greptile Summary

Removes unnecessary code that was resetting is_first_microbatch to True after CUDA graph warmup iterations.

The removed code was originally added for Megatron-Core requirements to prevent warmup from altering control flow. However, this reset is actually unnecessary and potentially harmful:

• Why unnecessary: During graph replay, skip_fp8_weight_update tensor controls FP8 weight updates, and modules automatically set is_first_microbatch=False when this tensor is present (e.g., layernorm_mlp.py:2053-2054)
• Why harmful: Resetting to True forces capturing the "clear" version of gradient accumulation (for first microbatch only), but for graphs reused across all microbatches, the "no-clear" version is needed to properly accumulate gradients

The two functions of is_first_microbatch are:

1. Control FP8 cast-transpose behavior → now handled by skip_fp8_weight_update during graph replay
2. Control fused wgrad clearing of main_grad buffer → should capture no-clear version for multi-microbatch usage

This fix ensures correct gradient accumulation when one CUDA graph is captured and replayed for multiple microbatches.

Confidence Score: 5/5

• This PR is safe to merge with no identified risks
• The change removes unnecessary code that was actually causing incorrect behavior. The PR author provides clear technical justification, the logic aligns with how skip_fp8_weight_update controls behavior during graph replay, and the removed code was interfering with proper gradient accumulation in multi-microbatch scenarios. The change is small (5 lines), well-understood, and fixes a bug rather than introducing new functionality.
• No files require special attention

Important Files Changed

Filename	Overview
transformer_engine/pytorch/graph.py	Removed unnecessary reset of is_first_microbatch after warmup, fixing cudagraph capture for multi-microbatch scenarios

_{Last reviewed commit: 25af05a}

[greptile-apps]

_{1 file reviewed, no comments}

_{Edit Code Review Agent Settings | Greptile}