eIQ AI Toolkit 2.1 (Gargoyle)

Backend

Added

  • Input and output name detection for TFLite and ONNX models

  • Database migration mechanism

  • Patch of model file for “ERROR” models

  • API Client auto-generation

  • Pyrefly and ty type checkers

  • Add kernels mapping for Neutron 3.1.2

  • Add on-device profiling support for FRDM platforms

  • Logging every REST API call and Celery task execution to MongoDB

  • Add logging of optimization and profiling jobs for AI Hub telemetry

Changed

  • Option to upload single file *.py model

  • Get specific error message when model upload fails during model check phase

  • Tensor arena size name and units to bytes

  • Celery shared task function’s positional arguments changed to keyword arguments for better usage of tasks in websocket messages

  • Enable download of uploaded model with status “ERROR”

  • Device info endpoint response extended with neutron firmware version

  • Updated dependencies to latest versions

  • Updated Python version to 3.13

  • Update filename of downloaded artifact to be relevant

Fixed

  • Raise exception when device is not available

  • Round per-layer execution time to 4 decimal places for i.MX95 NPU profiling

  • Fix on-device profiling for models with spaces in name

GUI

Added

  • Enable PyTorch single *.py file upload

  • Option to cancel running optimizations

  • Error message if optimization pipeline task failed

  • Click through to model detail

  • Show input and output tables for Onnx and TFLite models in Model Detail page

  • Move “Add model” nested menu button on top of the model table.

  • Move “Add dataset” nested menu button on top of the dataset table.

  • Add searchbars on top of model and dataset tables.

  • Move profiling menu buttons out of nested menu.

  • Profiling and Optimization overview pages

  • Display neutron version mismatch warning before on-device profiling

  • Display Neutron version tag in model detail

  • Training materials links

Changed

  • Refactored Error pages to handle all errors correctly

  • Unified platform names across the app

  • Websocket messages structure changed for args to kwargs

  • Enable download and patch of uploaded model with ERROR status

  • Routes for optimizations and profiling