vllm_omni.config.composable_parallel.aggregation ¶
Closed aggregation-pattern hierarchy for declarative parallel strategies.
An aggregation pattern describes how the per-worker results of one mesh axis are recombined into a single logical result (e.g. take one tensor-parallel rank's output, union disjoint data-parallel shards). The hierarchy is closed: subclasses may only be defined in this module.
AGGREGATION_PATTERN_VARIANTS module-attribute ¶
AGGREGATION_PATTERN_VARIANTS: tuple[
type[AggregationPattern], ...
] = (
TakeRank,
Union,
GatherDim,
AllGather,
StitchSpatial,
StitchPipeline,
Combine,
FanInByStage,
)
AggregationConflictError ¶
Bases: ValueError
Raised when two aggregation inputs conflict for the same request id.
AggregationPattern dataclass ¶
AllGather dataclass ¶
Bases: AggregationPattern
All ranks materialize the same aggregate result across the axis.
Combine dataclass ¶
Bases: AggregationPattern
Custom reducer combination (CFG). Adapter resolves reducer_id.
FanInByStage dataclass ¶
Bases: AggregationPattern
Fan-in results routed by stage (stage replicas).
GatherDim dataclass ¶
Bases: AggregationPattern
Gather along a dimension and materialize on one designated rank.
Default designated rank is axis-rank 0 unless adapter overrides.
StitchPipeline dataclass ¶
StitchSpatial dataclass ¶
TakeRank dataclass ¶
Bases: AggregationPattern
Select one rank's result (TP — all ranks agree).
Union dataclass ¶
Bases: AggregationPattern
Union of disjoint replica results (DP, HSDP).