Adaptive VRAM governor keeps 130M model training for 1M steps on an 8GB GPU without OOM

uBazzyZ- · reddit · 2026-09-10

A developer repeatedly hit CUDA OOMs training small models on an RTX 5060 Ti (8GB), so they built MEM Orchestrator, a lightweight runtime control layer for PyTorch: it monitors VRAM pressure and temporarily downshifts micro-batch size and gradient accumulation before an OOM, then steps back up when memory clears.

Implementation details:

Results: a 130M model survived a 1M-step endurance run with zero crashes, and a 255M model kept training through injected +1.2GB memory spikes during a 50k-step FineWeb-Edu run. The author explicitly doesn't claim this solves OOM and is seeking feedback from CUDA allocator / DeepSpeed / distributed-training experts.

Original post →

More from Infra

Infra channel →