кɵɵѕнī

Enabling model parallelism (training 30b on 2x 3090s and beyond)

15

Does as it says on the tin. Now multi-gpu users can choose to use them for faster training (DDP) or bigger models (MP) This required a minor change to the...

load_in_8bit now respects 'balanced' device maps in multi-gpu environments

4

# What does this PR do? Fixes `max_memory` generation for 'auto' 'balanced' and 'balanced_low_0' `device_map`s for models being loaded in 8bit Fixes # (N/A) no issues found, but one guy...

The whisper models can transcribe and translate at the same time, if configured correctly. > The models were trained on either English-only data or multilingual data. The English-only models were...

First of all, this is a beautiful program. Thank you for making it. I think it would be useful to have a setting to use a GPU for local models....