Open-source "weight teleportation" syncs RL weights in <0.23s without RDMA
gharik · x · 2026-11-03
- The humansand team open-sourced weight teleportation, tackling slow and costly weight syncs from trainers to samplers in large-scale RL.
- Sampler pause time is <0.23s, with 3s non-blocking compression/prepare; the bottleneck shifts to the GPU's decompression ASIC.
- Key techniques: massively parallel GPU compression of weight deltas plus a heterogeneous decompression-and-patching pipeline, outperforming intra-cluster methods and enabling samplers anywhere.
More from Infra
- ISTA (GSQ-RCO) vs Unsloth quant beef heats up as users ask why ISTA won't publish KLD — Queasy_Back2588 · 2026-11-03
- Has anyone benchmarked int8 models before and after the RTX 20-40 series unlock? — Dave-C · 2026-11-03
- B300/GB300 Blackwell financing deals smell like bogus multi-layer SPV schemes, warns insider — TheZachMueller · 2026-11-03
- First look at Google Cloud Run standalone instances: SSH-able microVMs — rseroter · 2026-11-03
- Northwood raises $375M Series C as a16z touts space's early-internet moment — a16z · 2026-11-03
- Intel ships 14A PDK 0.9, a make-or-break milestone for its foundry ambitions — ryanshrout · 2026-11-03