Add the support of variable dtypes of params and grads in capturable Fused Adam.

[ptrendx]

Description

Added support for the types of parameters and gradients to differ in FusedAdam(capturable=True) to get it to parity with non-capturable mode.

Fixes #3358

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

[ptrendx]

/te-ci

[greptile-apps]

Greptile Summary

The PR brings capturable FusedAdam to parity with its non-capturable mode by dispatching parameter and gradient element types independently.

• Splits capturable Adam kernel templates into separate parameter and gradient types.
• Validates parameter and gradient dtype homogeneity independently before launch.
• Adds CUDA graph numerical coverage for FP32 parameters with BF16 gradients and BF16 parameters with FP32 gradients and master weights.
• Corrects the capturable-master tensor-list diagnostic from four lists to five.

Confidence Score: 5/5

The PR appears safe to merge, with the mixed-dtype dispatch guarded by existing list validation and covered by capturable numerical tests.

The changed kernels consistently use independently dispatched parameter and gradient types, preserve the optimizer’s moment and master-weight contracts, and introduce no accepted correctness or security failure.

Important Files Changed

Filename	Overview
transformer_engine/common/multi_tensor/adam.cu	Independently dispatches parameter and gradient types in both capturable Adam kernels while retaining homogeneous-list validation and correct typed pointer arithmetic.
tests/pytorch/test_fused_optimizer.py	Adds focused CUDA graph comparisons against non-capturable FusedAdam for both supported mixed parameter/gradient dtype directions.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A["FusedAdam capturable step"] --> B["Parameter-dtype bucket"]
  B --> C["Validate homogeneous parameter dtype"]
  B --> D["Validate homogeneous gradient dtype"]
  C --> E["Dispatch PARAM_T"]
  D --> F["Dispatch GRAD_T"]
  E --> G["Capturable Adam kernel"]
  F --> G
  G --> H["Update parameters and optimizer state"]

Loading

_{Reviews (1): Last reviewed commit: "[pre-commit.ci] auto fixes from pre-comm..." | Re-trigger Greptile}

[vthumbe1503]

LGTM