nemo_rl.utils.multimodal_payload_metrics

Module Contents

Functions

_tensor_nbytes
_value_nbytes	Estimate data bytes, optionally counting shared leaves only once.
_value_segment_count	Count media leaves, optionally counting shared objects only once.
_typed_content_media_nbytes	Count media embedded in typed vLLM content without counting prompt text.
_typed_content_media_segment_count	Count media leaves embedded in typed content without counting text.
protocol5_serialized_nbytes	Return cloudpickle protocol-5 frame plus out-of-band buffer bytes.
collect_multimodal_payload_metrics	Measure one exact Ray argument without scanning when disabled.
collect_sharded_multimodal_payload_metrics	Aggregate metrics over the exact unique per-DP-shard Ray arguments.
merge_multimodal_payload_metrics	Aggregate payload measurements collected during one logging interval.
drain_multimodal_payload_metrics	Return and clear payload measurements recorded in this process.
print_multimodal_payload_metrics	Record metrics for the logger and print a stable, scrapeable line.

Data

_PENDING_PAYLOAD_METRICS
_PENDING_PAYLOAD_METRICS_LOCK

API

nemo_rl.utils.multimodal_payload_metrics._PENDING_PAYLOAD_METRICS: dict[str, int | float]

None

nemo_rl.utils.multimodal_payload_metrics._PENDING_PAYLOAD_METRICS_LOCK

‘Lock(…)’

nemo_rl.utils.multimodal_payload_metrics._tensor_nbytes(value: torch.Tensor | None) → int

nemo_rl.utils.multimodal_payload_metrics._value_nbytes(value: Any, seen: set[int] | None = None) → int

Estimate data bytes, optionally counting shared leaves only once.

nemo_rl.utils.multimodal_payload_metrics._value_segment_count(

value: Any,

seen: set[int] | None = None,

) → int

Count media leaves, optionally counting shared objects only once.

nemo_rl.utils.multimodal_payload_metrics._typed_content_media_nbytes(

value: Any,

seen: set[int] | None = None,

) → int

Count media embedded in typed vLLM content without counting prompt text.

nemo_rl.utils.multimodal_payload_metrics._typed_content_media_segment_count(

value: Any,

seen: set[int] | None = None,

) → int

Count media leaves embedded in typed content without counting text.

nemo_rl.utils.multimodal_payload_metrics.protocol5_serialized_nbytes(value: Any) → int

Return cloudpickle protocol-5 frame plus out-of-band buffer bytes.

nemo_rl.utils.multimodal_payload_metrics.collect_multimodal_payload_metrics(

data: Any,

boundary: str,

*,

enabled: bool,

) → dict[str, int | float]

Measure one exact Ray argument without scanning when disabled.

nemo_rl.utils.multimodal_payload_metrics.collect_sharded_multimodal_payload_metrics(

shards: collections.abc.Sequence[collections.abc.Mapping[str, Any]],

boundary: str,

*,

enabled: bool,

) → dict[str, int | float]

Aggregate metrics over the exact unique per-DP-shard Ray arguments.

nemo_rl.utils.multimodal_payload_metrics.merge_multimodal_payload_metrics(

metric_sets: collections.abc.Sequence[collections.abc.Mapping[str, int | float]],

) → dict[str, int | float]

Aggregate payload measurements collected during one logging interval.

Byte and segment counts are summed because repeated calls represent distinct Ray transfers. Per-call maxima and shard counts retain their maximum value, and physical-to-logical ratios are recomputed from the aggregated byte totals instead of averaging ratios.

nemo_rl.utils.multimodal_payload_metrics.drain_multimodal_payload_metrics() → dict[str, int | float]

Return and clear payload measurements recorded in this process.

nemo_rl.utils.multimodal_payload_metrics.print_multimodal_payload_metrics(

metrics: collections.abc.Mapping[str, int | float],

) → None

Record metrics for the logger and print a stable, scrapeable line.