[LiRunGuo]

Description

Fixes #3645.

PyTorch builds that auto-register CUDA graph RNG state expose
CUDAGraph.register_generator_state() as a deprecated no-op. Transformer
Engine currently treats method availability as proof that explicit
registration is required, so every forward, backward, and wgrad graph emits
warnings. Pipeline-parallel training multiplies that warning volume across
layers and microbatches, and the resulting stderr traffic can delay graph
capture enough to trip a distributed watchdog.

This change probes the installed implementation once per process by capturing a
tiny random operation and checking whether replay advances the default CUDA
generator without explicit registration. Builds that auto-register RNG state
advance across replays and skip the per-graph registration loop. Builds that
still require explicit registration, including the NGC 26.09 case noted in the
issue, retain the existing behavior. The probe restores the original generator
state. A unit test covers both branches.

Type of change

• Bug fix (non-breaking change which fixes an issue)

Changes

• Cache a capability result that distinguishes required graph RNG registration
  from PyTorch's deprecated no-op implementation.
• Preserve the explicit registration path for older or vendor-patched PyTorch
  builds that still need it.
• Add coverage for both API behaviors.

Validation

Validated on an NVIDIA RTX A6000 with PyTorch 2.14.1+cu130 and CUDA 13.0.

The affected PyTorch build emitted the expected C++ deprecation warning outside
Python's warning machinery:

Warning: CUDAGraph.register_generator_state() is deprecated, and will be
removed in a future PyTorch release. It is now a no-op and can be safely
removed from your code.

The behavioral probe correctly inferred that explicit registration is
unnecessary. RNG advanced across graph replays, and backward remained finite:

explicit_registration_required False
cached_result_matches True
dropout_replays_differ True
backward_grad_finite True
PASS

The focused capability test passed both legacy and auto-registration cases:

2 passed, 25 warnings in 4.87s

qa/L0_license/test.sh and the changed-file pre-commit checks also pass.

This draft was prepared with AI assistance. The human author must review the
complete diff and validation results before marking it ready for maintainer
review.