specific method · filed under optimization
Sequence Parallelism for Activations
An activation-parallel method that decomposes a tensor-parallel all-reduce into reduce-scatter and all-gather with an intra-block kernel between them.
Also called sequence parallelism (SP) for activations.
- source
- 1
- model
- 1
- lab adopt it
- 1
- strongest
- core
How sources treat it
One count per evidence span, weakest treatment to strongest.
core 1
Documented in
Evidence
1 span quoted from the sources, strongest treatment first.
We therefore adopt sequence parallelism (SP) for activations: the TP all-reduce is decomposed into a reduce-scatter and an all-gather, with the intra-block kernel inserted between the two collectives
coreinference servingin Kimi K3
Filed alongside
Other methods under optimization :: training parallelism.
Expert ParallelismTensor ParallelismMoonEPCommunication-Computation OverlapPipeline ParallelismAll-to-all Gradient Exchange with Local FP32 SummationCache-Based Pipeline CommunicationContext ParallelismData Replica Reduction over the Data Center NetworkData-Weighted Data ParallelismDynamic Context Parallelism for Large Multimodal SamplesFully Balanced Expert-Parallel TrainingGPU Planning Kernel for Redundant Expert MigrationHybrid ZeRO Bucket Assignment for MuonKDA Context ParallelismKnapsack-Based Balanced Assignment of Dense Parameter MatricesLoad-Balanced Image ShardingModified DualPipe 1F1B Pipeline Overlap for mHCNS-FLOP-Balanced Static Parameter PartitioningPipeline Payload ExtensionsPipeline ZeRO-2 Gradient Sharding with CPU Offloadingpipeline-bubble scheduling of ViT computationRedundant-Expert Capacity ReservationSConv-Aware Tensor-Parallel Sharding