grail(@grail_ai):PULSE made weight sync 100x faster. That turned the trainer itself into the bottleneck. @erfan_mhi just fixed that too. Grail's GRPO trainer is now 1.8x faster on a single B200: 27% to 47% MFU, epoch time nearly halved. Decentralized post-training is converging on centralized speed.

grail

@grail_ai

Verifiable Post-Training for LLMs - an order of @covenant_ai

加入 August 2025

4 正在關注 1.1K 粉絲

grail@grail_ai

2026.03.18 21:41

PULSE made weight sync 100x faster. That turned the trainer itself into the bottleneck. @erfan_mhi just fixed that too. Grail's GRPO trainer is now 1.8x faster on a single B200: 27% to 47% MFU, epoch time nearly halved. Decentralized post-training is converging on centralized speed.