Weight teleportation cuts RL sampler sync pause to under 0.23s across the internet
gharik · x · 2026-11-03
The humansand team shares weight teleportation, tackling slow, expensive weight updates from trainers to samplers in large-scale RL.
- Sampler pause time under 0.23s, with 3s non-blocking compression + prepare; faster than existing intra-cluster methods and works over the internet ("maybe even space").
- Built on massively parallel GPU compression of weight deltas plus a heterogeneous decompression and patching pipeline.
- After these gains, the bottleneck becomes the GPU's decompression ASIC.
- Commentator: forget p2p RDMA — weight sync pauses are now counted in milliseconds.
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
- Amazon allegedly canceling GPU orders and closing accounts buying compute hardware; Tim Sweeney among victims — chemist_slime · 2026-11-03
- Northwood raises $375M Series C as a16z touts space's early-internet moment — a16z · 2026-11-03