nemo_rl.data_plane.schema#
Shared constants and type aliases for the data-plane meta contract.
Module Contents#
Functions#
Return |
|
Return |
Data#
API#
- nemo_rl.data_plane.schema.Layout#
None
- nemo_rl.data_plane.schema.MICRO_BATCH_INDICES#
‘micro_batch_indices’
- nemo_rl.data_plane.schema.MICRO_BATCH_LENGTHS#
‘micro_batch_lengths’
- nemo_rl.data_plane.schema.ELEM_COUNTS_PER_GB#
‘elem_counts_per_gb’
- nemo_rl.data_plane.schema.GLOBAL_FORWARD_PAD_SEQLEN#
‘global_forward_pad_seqlen’
- nemo_rl.data_plane.schema.ROLLOUT_METRICS#
‘rollout_metrics’
- nemo_rl.data_plane.schema.GROUP_ID_TAG#
‘group_id’
- nemo_rl.data_plane.schema.PER_TOKEN_STAGING_FIELDS#
(‘token_ids_delta’, ‘token_mask_delta’, ‘generation_logprobs_delta’)
- nemo_rl.data_plane.schema.INPUT_IDS#
‘input_ids’
- nemo_rl.data_plane.schema.INPUT_LENGTHS#
‘input_lengths’
- nemo_rl.data_plane.schema.SAMPLE_MASK#
‘sample_mask’
- nemo_rl.data_plane.schema.MASK_SAMPLE#
‘mask_sample’
- nemo_rl.data_plane.schema.TRUNCATED#
‘truncated’
- nemo_rl.data_plane.schema.META_IDX#
‘meta_idx’
- nemo_rl.data_plane.schema.INVALID_TOOL_CALL_MASK#
‘invalid_tool_call_mask’
- nemo_rl.data_plane.schema.MALFORMED_THINKING_MASK#
‘malformed_thinking_mask’
- nemo_rl.data_plane.schema.DP_TRAIN_FIELDS#
(‘input_ids’, ‘input_lengths’, ‘generation_logprobs’, ‘prev_logprobs’, ‘reference_policy_logprobs’, …
- nemo_rl.data_plane.schema.OPD_FULL_HIDDEN_STATES_FIELD#
‘teacher_full_hidden_states’
- nemo_rl.data_plane.schema.OPD_FULL_LOGITS_FIELD#
‘teacher_full_logits’
- nemo_rl.data_plane.schema.OPD_FULL_FIELDS#
()
- nemo_rl.data_plane.schema.OPD_FULL_TEACHER_INDEX_FIELD#
‘opd_full_teacher_index’
- nemo_rl.data_plane.schema.SC_ROLLOUT_SCHEMA_FIELDS#
()
- nemo_rl.data_plane.schema.LP_SEED_FIELDS#
(‘input_ids’, ‘input_lengths’, ‘token_mask’, ‘sample_mask’)
- nemo_rl.data_plane.schema.TEACHER_LP_FIELDS#
()
- nemo_rl.data_plane.schema.PPO_VALUE_FIELDS#
(‘values’, ‘returns’)
- nemo_rl.data_plane.schema.DP_VALUE_TRAIN_FIELDS#
(‘input_ids’, ‘input_lengths’, ‘token_mask’, ‘sample_mask’)
- nemo_rl.data_plane.schema.VALUE_SEED_FIELDS#
None
- nemo_rl.data_plane.schema.DP_CALIB_INPUT_FIELDS#
()
- nemo_rl.data_plane.schema.ROUTED_EXPERTS_FIELD#
‘routed_experts’
- nemo_rl.data_plane.schema.ROUTED_LEN_FIELD#
‘routed_len’
- nemo_rl.data_plane.schema.ROUTED_EXPERTS_ENCODING_FIELD#
‘routed_experts_encoding’
- nemo_rl.data_plane.schema.ROUTED_EXTRAS_METADATA_FIELD#
‘extras_metadata_json’
- nemo_rl.data_plane.schema.ROUTE_ENCODING_NONE#
0
- nemo_rl.data_plane.schema.ROUTE_ENCODING_ENVELOPE#
1
- nemo_rl.data_plane.schema.ROUTE_ENCODING_LIST#
2
- nemo_rl.data_plane.schema.ROUTE_PLAN_TAG#
‘route_assembly_plan’
- nemo_rl.data_plane.schema.ROUTE_PASSTHROUGH_FLAG#
‘route_passthrough’
- nemo_rl.data_plane.schema.fields_with_optional_routed_experts(
- fields: Sequence[str],
- *,
- enabled: bool,
Return
fieldsplus routed experts when router replay is enabled.
- nemo_rl.data_plane.schema.fields_with_optional_opd_full(
- fields: Sequence[str],
- *,
- field: Optional[str],
- teacher_index_field: Optional[str] = None,
Return
fieldsplus the full-vocabulary MOPD teacher payload column(s).Added only when the run configures them: a GRPO run requesting a column nobody wrote would error on read, hence not folded into
DP_TRAIN_FIELDS.- Parameters:
fields – Base field list.
field – Payload column name, or
Nonewhen opd_full is off.teacher_index_field – Per-sample teacher-identity column name (hidden- state path only, see
OPD_FULL_TEACHER_INDEX_FIELD), orNonewhen opd_full is off or using the logits payload.
- Returns:
The field list, with the payload/index columns appended when applicable.