🐈 Task: Ersilia Gen Models Revision

[Abellegese]

Summary

A task to revise Ersilia Generative model for speed, consistency and bug, etc...

Objective(s)

Done?	Model ID	Summary	Repo URL	Status	Speed / Perf	Notes / Issues	Optimization
✅	eos6ost	Needs code push for CPU version upgrade; All REINVENT models got almost same speed	https://github.com/ersilia-os/eos6ost	pushed	—	—
✅	eos57bx	Needs code push for CPU version upgrade; All REINVENT models got almost same speed	https://github.com/ersilia-os/eos57bx	—	All REINVENT models got almost same speed	—	—
✅	eos2401	Converted to CPU; very slow on CPU	https://github.com/ersilia-os/eos2401	pushed	very slow (CPU)	—	torch; can be optimized
✅	eos694w	Needs code push for CPU version upgrade; All REINVENT models got almost same speed	https://github.com/ersilia-os/eos694w	—	almost same speed	—	—
✅	eos633t	Slow as reinvent; example input issue fixed	https://github.com/ersilia-os/eos633t	pushed	as slow as reinvent	3rd example input wasn’t working → removed; run output incorrect → fixed	—
✅	eos8fma	Works and fast	https://github.com/ersilia-os/eos8fma	—	good; large batch can use multiprocessing	Suggest Ersilia-pack multiprocessing for scale	—
✅	eos4q1a	Works and fast	https://github.com/ersilia-os/eos4q1a	—	fast	—	—
✅	eos9taz	Scaffold decorator; a bit faster than reinvent	https://github.com/ersilia-os/eos9taz	—	very slow (~90s for 3 inputs)	Bottleneck is TF model (not IO)	can’t optimize much (TF backend)
✅	eos4f8y	Low resource generator; in-progress	https://github.com/ersilia-os/eos4f8y	in-progress	—	Needed ersilia-pack fix for advanced shell cmds; created correct run_input.csv/run_output.csv/run_columns.csv. It still requires model retraining which I need explanation how to deal with it.	—
✅	eos8vud	In-progress	https://github.com/ersilia-os/eos8vud	in-progress	Fast, I fixed dependency issues, code cleanup and testing.	code isnt pushed	—
✅	eos69e6	In progress	https://github.com/ersilia-os/eos69e6	in-progress	—	—	—
⬜	eos8bhe	In maintenance	https://github.com/ersilia-os/eos8bhe	maintenance	—	—	—
⬜	eos4qda	In maintenance	https://github.com/ersilia-os/eos4qda	maintenance	—	—	—

Documentation

No response

[GemmaTuron]

Hi @Abellegese,

Thanks for the clarity, I'm doing a revision of the models, here a few comments:

• eos8vud is extremely slow in my computer, and it produces Nans. I am not sure this is the expected behaviour based on your comments, I've left an issue in the model
• eos6ost is failing the image test workflow (re-running now)
• eos4qda: what is the plan for this model?
• eos69e6, you mention working on this model but I see the aMD64 image failing in the tests still

I am now testing the reinvent models

[GemmaTuron]

I have successfully tested:

• eos57bx
• eos694w
• eos8fma
• eos633t
• eos4q1a
• eos9taz
• eos4qda

Models that do not work for me:

• eos8vud: all Nan's, issue opened: https://github.com/ersilia-os/eos8vud
• eos2401: fails at fetch time, issue opened: Model fetching fails eos2401#5
• eos6ost: fails the test deep, issue opened: Test image fails eos6ost#7
• eos69e6: the deep check fails in my workstation: Model fails deep test eos69e6#4

[Abellegese]

@GemmaTuron I checked the following models with the new Ersilia push

• eos8uvd works (this previously fails because we are using randomly generated example input, I explained that already)
• eos2401 works