Repository navigation
OpenDPD 2.4.2 reduces hosted Studio memory residency by returning unused guest pages to the host and avoiding duplicate disk caches. The VM keeps its 8 GiB capacity, four vCPUs, 6 GiB API service limit and GPU resources. Waveform synthesis, precision, numerical metrics, training algorithms and acceptance thresholds are unchanged.
Production validation: idle VM RSS fell from 4,886.22 to 1,145.32 MiB after the public GPU and browser workflows (76.56% lower). All three synthetic 1 GiB allocation/release cycles returned approximately 1 GiB to the host within the 10-second measurement wait. Host cgroup usage fell from 8,465.59 to 1,133.36 MiB. Residency varies with active workloads; the guest capacity was not reduced.
Validation passed: 1,256 targeted backend tests, 483 frontend tests, required CI, strict documentation, production builds and hashed runtime audits. A 983,040-sample NR waveform has exactly the same SHA256 as 2.4.1. Public browser checks passed NR generation/download integrity, session isolation, Virtual PA/dataset creation, and all four real CUDA tasks: PA training, PA evaluation, DPD training and DPD inference. Desktop and phone checks passed; the checked Studio views had no serious or critical Axe violations and no JavaScript errors. The public default NR model generated and previewed 614,400 samples successfully. Test sessions and containers were cleaned up.
Runtime locks update filelock to 4.2.1 and setuptools to 84.1.0 with verified artifact hashes. API, GPU image, PyPI wheel and Pages frontend are bound to the reviewed release source 970d6968a2947e77fc378b65efa2f2694b914cae. The drained VM restart completed successfully; prior source, runtime, image and unit remain available for rollback.
Production validation record · Release notes · Hosted Studio