Search papers, labs, and topics across Lattice.
Pluralis Research
4
0
6
Approximate synchronization in Factored Gossip DiLoCo enables non-blocking communication that enhances compute utilization and resilience in distributed training.
Training billion-parameter Transformers can be stabilized by controlling curvature through a novel architecture warm-up strategy, leading to smoother convergence and fewer instabilities.
HyperVis reveals that leveraging hyperbolic geometry can significantly enhance compositional reasoning in VLMs, outperforming traditional discrete approaches.
Training LLMs across 176 untrusted workers is now possible without computation duplication, thanks to a new verification mechanism that detects corrupted inter-stage communication in pipeline parallelism.
