fix(asr): catch RuntimeError when checking CUDA graphs conditional-node support

[mittalpk]

Fixes #15905

What does this PR do ?

maybe_enable_cuda_graphs() crashes instead of falling back gracefully when cuda-python is importable but there's no real NVIDIA driver to talk to (e.g. ROCm/AMD builds of torch).

Collection: asr

Changelog

• nemo/collections/asr/parts/submodules/transducer_decoding/label_looping_base.py: added RuntimeError to the exception types caught around check_cuda_python_cuda_graphs_conditional_nodes_supported() in maybe_enable_cuda_graphs().

Root cause

On systems where torch.cuda.is_available() returns True (e.g. ROCm/AMD builds of torch, which shim the CUDA API) but no real NVIDIA driver/libcuda.so is present, check_cuda_python_cuda_graphs_conditional_nodes_supported()'s call to cuda.cuDriverGetVersion() raises a raw RuntimeError ("Failed to dlopen libcuda.so.1") from inside the cuda-python native extension — before the function's own EnvironmentError/ImportError checks even run. maybe_enable_cuda_graphs() only caught (ImportError, ModuleNotFoundError, EnvironmentError), so this RuntimeError propagated uncaught and crashed decoding instead of falling back to CudaGraphsMode.NO_WHILE_LOOPS as intended.

Testing

Added tests/collections/asr/decoding/test_label_looping_cuda_graphs_fallback.py (3 tests, no real GPU/ROCm hardware needed — mocks check_cuda_python_cuda_graphs_conditional_nodes_supported directly):

• falls back to NO_WHILE_LOOPS when the check raises RuntimeError (the new case)
• still falls back on the previously-handled error types (EnvironmentError)
• still enables FULL_GRAPH mode when the check succeeds

Verified the new RuntimeError test fails against the pre-fix code and passes against the fix. Ran the full tests/collections/asr/decoding/ + tests/collections/common/test_optional_cuda_graphs.py suites (283 passed). One unrelated pre-existing failure (test_ctc_decoding.py::test_batched_decoding_labels) reproduces identically with and without this change — confirmed by running it both in isolation and as part of the full suite on unmodified main, so it's not something this PR introduces.

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you add or update any necessary documentation? (not applicable — no user-facing behavior/API changed, only an internal fallback's exception handling)
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc) — no

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.