Show HN: UATC-Closed-loop VRAM control and dynamic data pruning for LLM training
Category: ai-ml
Tags: llm-training, control-theory, vram-management, edge-ai, pytorch
Score: 7.3/10 (Innovation: 8, Technical: 9, Documentation: 6, Utility: 6)
UATC applies control theory (Kalman filters, PID, Smith predictor) to dynamically manage VRAM and prune data during LLM training on edge GPUs, preventing OOM crashes. Its closed-loop approach adapts batch size, learning rate, and pruning in real-time, showing strong results on a T4 with Qwen2.5-1.5B. This is an innovative combination of control systems with deep learning training, but the project is early-stage with limited documentation and no license.
Target audience: ai researchers, ml engineers, edge compute developers
Repository: https://github.com/sajjaddoda72-design/UATC · Python · NOASSERTION · 2 stars
View on Hacker News