EVA Vision v1.3.3 (Apr 30, 2026)

Helm-based TorchServe configuration flow organized: Helm ConfigMap's config.properties.template has been mounted to the /app/eva_ts/config.properties.template path, so that a runtime configuration file is created based on the deployed template.
Helm's model and worker settings are reflected at runtime: The load model, number of workers, batch size, and batch delay time set in Helm values.yaml have been improved to be reflected in the generated config.properties.
Added TorchServe runtime control options: Added number_of_netty_threads and worker_retry_timeout_sec entries to the TorchServe settings template rendered in Helm.
Improved Kubernetes service defaults: Changed the default service type from NodePort to ClusterIP, and modified nodePort to be rendered only when the service type is NodePort.
Adjusted basic serving resources and model parameters: For more stable runtime configuration, the default memory request/limit was increased to 32Gi, the number of workers for some models was adjusted to 1, and the placement settings of RT-DETRV2 were changed.

Fixed an issue where Helm values settings were not reflected in TorchServe runtime settings: Fixed an issue where model and worker settings in Helm values.yaml could be ignored due to a mismatch between the deployed settings file path and the runtime settings creation flow.
Modified torch.compile behavior for dynamic input in RT-DETRV2: Applied the dynamic=True option when compiling RT-DETRV2 to support various input sizes more reliably.
Fix missing IG model archive dependency: Included interest_zone.py in IG model archive settings to enable interest-zone logic at runtime.