eIQ AI Toolkit 2.0 (Fairy)¶

Backend¶

Redis keys in optimization pipeline were not unique and did not allow two pipelines to run at the same time
Incorrect vela profiling inference time after failed run
Prevented hangs during parallel profiling on a single board by failing concurrent runs early

Acceleration ratio in profiling runs
Add Pydoclint - docstring linter
List of kernel kinds for Neutron 3.0.1 release.
OnnxOpVersionConversion pass
Model explorer graph conversion
OnnxPeepholeOptimizer pass
Enable to set number of warmup and regular inference runs for on-device profiling, add statistics for warmup and regular inferences, add total and init memory footprint to on-device profiling result
Profiling detection of Neutron models

Remove properties ‘inputs’, ‘outputs’, ‘input_shape’, ‘output_shape’, ‘order’ from structure ‘Node’ of profiled model graph and edit all parsers according to it
Update MCU profiling input file schema
Updated Neutron simulated profiling parsing, option to dump statistics to JSON file was removed from converter.
Pinned numpy to 2.3.5
Migrated to eiq-olive 0.9.0
Redesigned resource download endpoints to two-step utilizing pre-signed URLs

Artifact table in optimization detail
General statistics cards to profiling detail
Chip in the list of models for optimization pipeline input that shows if model was converted for vela or neutron
Functionality to save optimization pipeline configurations as templates
Enable to set warmup and regular inference number in on-device profiling form
Warnings for possibly incompatible model selection in profiling
Make table rows clickable
Spinner on download button
Model Explorer
Model detail page
Dataset detail page
Show Profiling tag for Neutron profiling models