- Helm Chart Version:
2.0.4
- compatibility: This version targets EVA App v3.0.0 and above.
Performance Enhancements
- Convert MMGDino to an internal-only model in the App workflow:
MMGDino will still be loaded into TorchServe for direct calls and internal use, but will be hidden from GET /models responses and excluded from /sync-based dynamic camera scaling.
- Improved ensemble behavior in situations where only some models are available: Adjust voting requirements based on models that are actually registered and in READY state, and return an explicit
503 if the requested current_model cannot provide a reference score.
- Improvements to deployment and model archive configuration: Add automatic creation of
dummy.pt when creating a model archive, remove placeholder files from Docker build context, and fix Helm rendering to work correctly when using the Recreate deployment strategy.
- Serving stack and worker profile updates: Upgrade core runtime dependencies and adjust camera-to-worker profiles for
OmDet, MMGDino, and RT-DETRV2.
- RT-DETRV2 VitPose verification memory usage limit: Add configurable chunk processing to RT-DETRV2 VitPose verification to avoid GPU memory spikes in crowded scenes with a lot of people detection.
- Force local model artifact loading in handler: Only load model artifacts from the local MAR/model directory to avoid Hugging Face Hub hits or downloads during runtime.
Bug Fixes
- Camera registry removal modification: Modify that the App entry is actually removed when calling
DELETE /models/{model}/cameras/{app_id} by awaiting app_registry.remove().
- LLMDet/MMGDino compiled text backbone mask processing modification: To resolve the mask shape error in Transformers 5.x, the tokenizer
attention_mask is passed to the compiled BERT text backbone, and the existing text_self_attention_masks is maintained in the downstream encoder logic.
- Fix RT-DETRV2 non-person target handling: Skip RT-DETRV2/VitPose verification for non-person targets to avoid duplicating person verification work in multi-target requests.
- Fix OmDet backbone input size: Restores the OmDet backbone input size expected by the current Transformers runtime.