implementation detail · filed under optimization
Split fused gradients before orthogonalization
Separating fused gradients or parameters into sub-matrices, orthogonalizing the parts independently, and restoring the original layout where specified.
Also called splitting of fused parameters.
- sources
- 2
- model
- 1
- lab adopt it
- 1
- strongest
- core
How sources treat it
One count per evidence span, weakest treatment to strongest.
used 2core 1
Documented in
Evidence
3 spans quoted from the sources, strongest treatment first.
We therefore split the fused gradient before orthogonalization, run NS on each sub-matrix independently, and gather the results back into the original layout before applying the update.
coreoptimizationin Qwen3.8-NextQwen
Fused parameters are split before orthogonalization
usedunclearin Qwen3.8-Flash-NextQwen
Filed alongside
Other methods under optimization :: optimizer.
MuonPer-Head MuonAdamWCategory-specific assignment of Muon and AdamWNesterov momentumSinkhorn-balanced updateAdam without weight decay for the N-gram embedding tableAdamW for attention and GDN output gatesAdamW for gated-residual low-rank projectionsAdamW for input embeddings and output headAdamW for the MoE routerAsynchronous Micro-Group pipelineCanzonaCUDA graph capture of the optimizer stepEight-step Newton–Schulz iterationHybrid Muon and AdamW parameter-group optimizer assignmentHybrid Newton–Schulz iterationsHybrid optimization with Muon and AdamMoonlight-style learning-rate scalingMuon orthogonalization accuracy refinementMuon restricted to two-dimensional linear-map weightsMuon SplitMuownPeer-to-peer shard retrieval for Muon orthogonalization