implementation detail · filed under model architecture
Two-Phase Block AttnRes Schedule
A Block AttnRes schedule with a batched inter-block pass followed by an online-softmax merge of intra-block partial sums.
Also called two-phase Block AttnRes kernel schedule, two-phase schedule.
- 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.
Block AttnRes [57] follows a two-phase schedule: a batched inter-block pass reads the cached block representations once per block, after which each layer folds in the intra-block partial sum through an online-softmax merge
coreinference servingin Kimi K3
Filed alongside
Other methods under model architecture :: normalization & residual.
Gated ResidualManifold-Constrained Hyper-ConnectionsAttention Residuals (AttnRes)GatedNormIdentity Hyper-ConnectionsRMSNormSingle-Pass mHCBlock AttnResHyper-ConnectionsIndependent Per-Branch NormalizationPer-Branch Scalar Write GateQK-ClipAll-branch Elementwise Read GateBounded Positive GatesData-Dependent Residual Read and Write OperatorsDynamic Gating of Residual Reads and WritesElementwise Data-Dependent Residual Read GateFour-Branch Residual StreamFour-Stream Hyper-Connection Combine UpdateFull Attention ResidualsLow-Rank Gated MixPost-Embedding RMSNormPre-LNSimplified AltUp