Explanation of data location on Workflows

[avickars]

📚 The doc issue

I didn't see anything relating to this in the doc, so I figured Ill post the question here. I am wondering where data is stored during a workflow? For instance, in my case we are looking to execute a series of models on an image (i.e. a PGIE -> SGIE....), and we are wondering where is say the image that is sent to torchserve is stored? In our ideal world we'd just move the image to the GPU once during say the pre-processing step of the workflow and then use that same instance of the image in both the PGIE and SGIE. However, I can see that we have to decode the image during the pre-processing of both the PGIE and the SGIE. So Im just looking to understand how that image would flow in the workflow (i.e. is it copied, where is it stored etc).

Any help would be greatly appreciated!

Suggest a potential alternative/fix

No response

[agunapal]

@avickars I wad discussing with @maaquib regarding this. Currently, the preprocessing step is not shared by the different elements of the DAG in the workflow. It's not optimal but optimizing this to a zero copy solution is not planned for the near future.

[agunapal]

Closing since there is no follow up comment. Please re-open if required