femboyisp/emry

Emry is a monitoring tool for machine learning training runs. When you train a model that takes hours or days to complete, you want to watch metrics like loss and learning rate without that monitoring code slowing down the training itself. Emry is designed around this constraint: the core emit() call is meant to take under 10 microseconds and never blocks the training loop. The setup is minimal. Your training loop calls emry.run() to start a named run, then calls run.emit() each step with whatever metrics you care about as keyword arguments. Emry handles the rest: it writes an append-only log file you can read with standard tools like jq or pandas, and if you are running in a terminal it brings up a live dashboard automatically. The dashboard shows a loss curve, phase markers, checkpoint annotations, and an optional overlay of a prior run so you can compare current performance against a baseline. Two dashboards are included. The terminal version (emry watch) runs in any standard terminal. The web version (emry web) serves a local browser dashboard that works fully air-gapped with no CDN dependency. Both have the same feature set. For cluster environments like SLURM, a sidecar mode lets the monitoring engine outlive the training process if the process crashes or is killed. Emry also automatically samples GPU utilization, memory, and temperature via nvidia-smi when a GPU is present, and can send a Slack or Discord alert if a metric goes NaN or infinity. The core engine is written in Rust for performance, the Python API wraps it via the maturin build tool. Installing it is a single pip install with no account or external service required. This is for machine learning researchers and engineers who run long training jobs and want lightweight, self-hosted observability that is easy to read both live and after the fact.