pytorch/serve

Confused about Cumulative Inference Duration vs. PredictionTime

Open

#1 698 ouverte le 20 juin 2022

Voir sur GitHub
 (3 commentaires) (0 réactions) (0 assignés)Java (3 844 stars) (790 forks)batch import
help wantedquestion

Description

📚 The doc issue

I am running a model on TorchServe and I am trying to see how long it takes for inference. If I use logging and view the logs, then I can see there is something called PredictionTime: image

However, if I use the Metrics API, then I got something called "Cumulative Inference Duration" image

And in terms of values those 2 are very different. So I am not sure which one should I use to measure the total inference time for my requests?

Btw, there is also something else called HandlerTime in the logs image

What does it mean? Where can I find related information about what are the meanings of these metrics?

Thanks,

Suggest a potential alternative/fix

No response

Guide contributeur

Stack technique
pythonpytorch
Domaine
documentation
Type d'issue
documentation
DifficultéDifficulté estimée pour un nouveau contributeur, de 1 pour un très petit changement à 5 pour un travail expert.
2
Temps estiméFourchette de temps approximative pour investiguer, implémenter, tester et préparer une pull request.
1-3 hours
Statut d'activitéDisponibilité apparente de l'issue : fraîche, active, ancienne, bloquée ou en attente d'un mainteneur.
needs maintainer response
ClartéClarté avec laquelle l'issue explique le changement attendu, les critères d'acceptation et la prochaine étape.
mostly clear
Prérequis
Aucun
Accessibilité débutantScore de 1 à 100 estimant l'accessibilité de cette issue pour un premier contributeur.
40
Direction de recherche
The user is confused about the difference between PredictionTime, Cumulative Inference Duration, and HandlerTime metrics. To clarify, one should look at the TorchServe metrics documentation (e.g., docs/metrics.md) and the source code where these metrics are logged (e.g., ts/metrics.py or ts/handler.py). A maintainer should explain which metric corresponds to request level inference time and how they relate to each other.