TVM Team Open-Sources PithTrain: Agent-Native MoE Framework
青稞AI · wechat · 2026-07-07
Tianqi Chen's team has open-sourced PithTrain, a streamlined, agent-native MoE training framework. They introduced "agent-task efficiency (ATE)"—measuring the cost for coding agents to understand, operate, and scale training frameworks—along with the ATE-Bench evaluation. Built on four agent-native design principles, PithTrain significantly improves ATE while maintaining production-grade throughput: on the ATE-Bench, it reduces conversation turns by up to 62% and saves 64% of GPU time for the same tasks. The authors will host a live session in the Qingke Community on July 11.
More from Infra
- AI Infra Meme: While Countries Need TW-Scale Power for Datacenters, People Argue About Lightbulbs — mgostIH · 2026-10-27
- Legora runs 2,000 parallel agent evals on Modal Sandboxes for legal AI — blelbach · 2026-10-27
- Cloud release velocity scores: AWS 95, Azure 90, Google Cloud 82 — DavidLinthicum · 2026-10-27
- Selling a 5090 to scale local inference: one business weighs 5 hardware paths — Anarchaotic · 2026-10-27
- AWS tutorial: fleet-wide edge CV model lifecycle with SageMaker SDK v3, ~$0.07 per run — AWS ML Blog · 2026-10-27
- Qwen3.8-27B hits 40 tok/s on Jetson Thor with INT4 + MTP, accuracy beats FP16 — No-Dragonfly6246 · 2026-10-27