Optimizes GPU memory usage and automates recovery from CUDA out-of-memory errors during BayesFlow training.
The bayesflow-memory skill provides a specialized set of tools and best practices for managing VRAM constraints during complex BayesFlow simulations and training. It empowers developers to implement automated recovery strategies like batch-size backoff, utilize gradient checkpointing to trade compute for memory, and perform analytical memory estimation to prevent crashes before they happen. Whether you are dealing with limited hardware or scaling large models, this skill ensures stable training loops through advanced memory probing and efficient sampling techniques.
Key Features
010 GitHub stars
02Predictive memory probing to find optimal configurations pre-training
03Automated OOM detection and batch-size backoff recovery loops
04Detached sampling and chunked log-prob calculations for reduced peak memory
05Gradient checkpointing support for PyTorch-backed BayesFlow models
06Analytical VRAM estimation for hardware-aware hyperparameter optimization
Use Cases
01Training complex BayesFlow models on consumer GPUs with limited VRAM
02Pruning hyperparameter optimization trials based on predicted memory footprints
03Building resilient training pipelines that recover automatically from memory crashes
