search

Modalities / modalities

A framework for training multimodal foundation models.
MIT License
39 stars 3 forks source link

Improving throughput for PackedDataGenerator #107

Closed le1nux closed 3 months ago

le1nux commented 3 months ago

On DGX the PackedDataGenerator was quite slow with throughput ranging between 30 - 300 samples/second.

The PR adds queuing, batching and rate limiting, resulting in more than 20.000 samples / second.