Skip to content

NEST SSL model-support training-step tests produce NaN loss on synthetic batches #15810

Description

@chtruong814

Two NEST SSL model-support jobs are failing because the synthetic AudioNoiseBatch used by the generated training-step test produces a non-finite loss.

Failures from run https://github.com/NVIDIA-NeMo/NeMo/actions/runs/27729959354?pr=15802:

Both jobs pass model init and inference; only the direct synthetic training_step() check fails. The test constructs random audio, noise, and independent random noisy_audio, which may not be a valid training sample for these restored NEST SSL artifacts.

Suggested follow-up: replace the synthetic batch with a minimal valid sample for these models, or keep the generated model-support training-step check disabled for these two artifacts while preserving init and inference coverage.

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type
No fields configured for issues without a type.

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions