Modalities / modalities

Modalities, a PyTorch-native framework for distributed and reproducible foundation model training.
MIT License
61 stars 5 forks source link

Calculate total_steps for a scheduler dynamically. #210

Open lllAlexanderlll opened 2 months ago

lllAlexanderlll commented 2 months ago

Feature request

Calculate total_steps for a scheduler dynamically.

Motivation

Currently, a scheduler's total_steps must be specified manually. Each time the network topology changes, we need to adapt the value. For example, I used 2 GPUs during debugging, but within the tests, I wanted to use only a single. Manually adapting the config is then needed, which is somewhat hacky.