Profiling an Ensemble including BLS

[corentin87]

Hi everyone,

I’m working with an ensemble model that consists of two Python models and two TensorRT (TRT) models (detection and classification). One of the Python modules (module.py) invokes the TRT classification engine using BLS.

The pipeline looks like this:

Preprocessing → Detection → Post-processing w/ BLS call to Classification

From #982
, I understand that:

Model Analyzer only supports “joint” tuning for a pipeline when using a single top-level ensemble composed of models that are not themselves ensembles or using BLS.

Is this still the current limitation?

If so, are there any recommended workarounds for handling cases like mine?

Thanks in advance!

[corentin87]

I am currently, profiling by disabling BLS exec calls. I guess it's the best I can do at the moment. And hopefully the team can make that kind of profiling available in the future.