nvalchemi.hooks.ScalarSnapshot#
- class nvalchemi.hooks.ScalarSnapshot(*, stage, scalars, timestamp_s=<factory>, elapsed_s=None, event_count=None, step_count=None, batch_count=None, epoch_step_count=None, epoch=None, global_rank=0, messages=())[source]#
Scalar reporting payload for one hook event.
- Parameters:
stage (str)
scalars (dict[str, float])
timestamp_s (float)
elapsed_s (float | None)
event_count (int | None)
step_count (int | None)
batch_count (int | None)
epoch_step_count (int | None)
epoch (int | None)
global_rank (int)
messages (tuple[ReporterMessage, ...])
- stage#
Hook stage name associated with the snapshot.
- Type:
str
- scalars#
Flat scalar mapping using slash-separated semantic keys.
- Type:
dict[str, float]
- timestamp_s#
Wall-clock timestamp from
time.time().- Type:
float
- elapsed_s#
Seconds since the reporting state was created, when available.
- Type:
float | None
- event_count#
Reporting orchestrator event count, when available.
- Type:
int | None
- step_count#
Workflow step count from the hook context, when available.
- Type:
int | None
- batch_count#
Training batch count from the hook context, when available.
- Type:
int | None
- epoch_step_count#
Training epoch-local batch count from the hook context, when available.
- Type:
int | None
- epoch#
Training epoch from the hook context, when available.
- Type:
int | None
- global_rank#
Distributed rank from the hook context.
- Type:
int
- messages#
Recent reporting messages captured from the shared reporting state.
- Type:
tuple[ReporterMessage, …]