Closed mbrenon closed 1 month ago
I ran inference on the example decoder models with this branch and everything looked good. Also converted tiny_llama with no issues. Would be great to get an official thumbs up from someone who has worked with the Stable Diffusion model.
@yichunk could you help review the changes to SD code? thanks!
Actually, I see we need another change for the T5 encoder/decoder models, similar to the other decoder-only models. Please add
head_dim=64
To the AttentionConfig
in get_model_config_t5() function in t5.py
. Thanks @mbrenon!
@yichunk could you help approve this PR? thanks!
Closing in favor of #120
This is a requirement for the upcoming OpenELM models.
Note that head_dim is also moved down to the attention config: this parameter is related to attention so it has no reason to live under the main model config.
Please review carefully to make sure I'm not breaking anything here :)
BUG=https://b.corp.google.com/issues/352478939