Huawei-backed ASP RL training scheduler accepted at NeurIPS, cuts iteration time up to 76%
机器之心 · wechat · 2026-11-07
ASP (Adaptive Scheduling Pipeline), developed by Shenzhen Big Data Research Institute with Huawei, decouples rollout generation from training in multi-instance asynchronous RL and adds queue-aware scheduling that eliminates padding, skewness, and sorting bubbles offline. It is reportedly the first LLM RL-training acceleration system built on domestic Ascend compute accepted at NeurIPS.
- On 64 Ascend 910B3 NPUs with Qwen3-30B-A3B/32B, iteration time drops 76% vs MindSpeed-RL and >60% vs verl; generation throughput reaches 58 tokens/s/NPU.
- Scaling to 3-4 generation nodes yields near-linear speedup; bounded staleness (3.3 training steps) shows no visible convergence degradation.
- Deployed on a 100-card Ascend cluster for telecom RL training, cutting the training cycle from 3+ months to 2 months; verl migration and open-sourcing planned.
More from Infra
- Batch APIs: the 50% LLM discount only pays off after model choice and prompt caching — arpit_bhayani · 2026-11-13
- RTX Pro 6000 bricks mid-Qwen run, owner jokes it "died of a broken heart" over Jensen news — wine_dark · 2026-11-13
- Bittensor's Chutes Handles 256 Concurrent Requests on a Single RTX 5090 at ~18k tok/s — markjeffrey · 2026-11-13
- Google expands TPU v8 lineup: 8tx cuts latency 2x, 8i Max doubles silicon per rack — Beth_Kindig · 2026-11-13
- Horace He to share two mental models for reasoning about inference at SGLang Summit — cHHillee · 2026-11-13
- Used A100 prices fall another 20-30% as sellers rush ahead of a collapse — Bascilian · 2026-11-13