Weight teleportation cuts RL sampler pauses below 0.23s at internet speeds
gharik · x · 2026-11-03
humansand shares 'weight teleportation' for large-scale RL: sampler pause under 0.23s and 3s non-blocking compression+prepare, with the bottleneck pushed to the GPU's decompression ASIC (nvCOMP + Blackwell decompression engine). GPU-parallel compression of weight deltas plus a heterogeneous decompression/patching pipeline enables samplers anywhere; implementation shared with the community.
More from Infra
- Anthropic discloses $895.4B in contractual spending commitments ahead of IPO — Polymarket · 2026-11-03
- Polymarket: Anthropic discloses $895.4B in contractual spending commitments ahead of IPO — Polymarket · 2026-11-03
- RTX 5080 reportedly discontinued, store prices nearing $2,000 — PAstynome · 2026-11-03
- Inference startups hitting $100M+ ARR are just reselling prebooked compute, not real businesses — abhiadesai · 2026-11-03
- Stripe Data Pipeline adds real-time sync to Lakebase Postgres with seconds-level updates — matei_zaharia · 2026-11-03
- AI datacenter capex hits $1.2T in 2026 as growth peaks at 56% — the 'second derivative' worry explained — TiernanRayTech · 2026-11-03