implementation detail · filed under optimization
All-to-all Gradient Exchange with Local FP32 Summation
An implementation detail that exchanges local gradients across ranks with all-to-all, then sums them locally in FP32.
- source
- 1
- models
- 2
- lab adopt it
- 1
- strongest
- used
How sources treat it
One count per evidence span, weakest treatment to strongest.
used 1
Documented in
Evidence
1 span quoted from the sources, strongest treatment first.
First, an all-to-all operation exchanges local gradients across ranks, and then each rank performs a local sum in FP32.
usedoptimizationin DeepSeek-V4DeepSeek
Filed alongside
Other methods under optimization :: training parallelism.
Expert ParallelismTensor ParallelismMoonEPCommunication-Computation OverlapPipeline ParallelismCache-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 ShardingSequence Parallelism for Activations