PKU open-sources RayOrch, lineage-aware data-prep engine with up to 15.14x speedup
PekingUniversity · hf · 2026-09-28
- Peking University released RayOrch, an open-source programming model and distributed execution engine for foundation-model data preparation pipelines.
- Problem: pipelines expand heterogeneous documents/videos into parent-child record trees; existing systems either hide parallelism behind coarse-grained jobs or expose flat records, forcing apps to manage lineage and regrouping.
- RayOrch preserves parent-child relations throughout execution: programs declare ordered variable-cardinality expansions with matching gathers, validated by the compiler; the runtime tracks child membership, parents, ordinals, and terminal states; per-call FIFO ready queues batch children across parents; gathers reconstruct results from declared membership; typed parent-scoped failures suppress only that parent's undispatched siblings.
- Performance on NVIDIA H20 GPUs: 15.14x speedup scaling MinerU from 4 to 64 GPUs, 7.82x for a video pipeline from 8 to 64 GPUs; reduces end-to-end time by 13.1% vs Ray Data and 29.0% vs Daft on MinerU, and 16.0% vs Ray Data on Docling.
- Code: github.com/OpenDCAI/RayOrch
More from Infra
- Sarvam AI Launches Token Factory: India-Hosted Inference for Open Models — itsOmSarraf_ · 2026-11-16
- BIS Issues Interim Final Rule Addressing Remote Access to Advanced AI Chips — apples_jimmy · 2026-11-16
- Jailbroken PS4 runs a 9B open-source model at up to 43 tok/s by talking directly to its 2013-era GPU — AIFlow_ML · 2026-11-16
- Mobile AI Factory: agents auto-build on-device SLMs with 19x less GPU time — pmttyji · 2026-11-16
- Stratechery's Fictional Obituary: Imagining Jensen Huang's 2026 Exit From Nvidia — Gregory_C_Allen · 2026-11-16
- vLLM serves DeepSeek V4.1-Flash at 954K tokens/s per GPU, $38.96/GPU-hour on GB300 NVL72 — bookwormengr · 2026-11-16