MobileNetV5

[rwightman]

Switching to RMS Norm

Fixes for correct export

Tinkering with 'mobiletnetv5' details, fixing some issues with msfa

A few tweaks and comments to example MNV5 impl

Update RmsNorm2d modules to use own 2d eager kernel instead of torch rms_norm w/ permute

Fix propagation of act_layer to RmsNormAct*, use ConvNormAct for stem instead of just Conv2d

Fixes from weights conversion

Plumbing norm_layer through to MultiQueryAttention2d

impl forward_features for Transformers compatibility

Adding forward_* APIs to MobileNetV5Encoder

cleanup

cleanup, model entrypt rename

Large redundant with 300m

Update input size for configs

Fix stem conv layer name

fix: always norm in MSFA

Always call final MSFA norm layer

Remove some FIXME, fix MSFA docstring. Remove use_layer_scale and rely on values == None, not used currently in any case.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[JulienMaille]

Cool, do you have any plans to pretrain it?

[rwightman]

@JulienMaille possibly yes, I may do something at a 'base' or slighty smaller size as a reference.

More near term, I will bring over the image encoder only weights as timm models for fine-tune. Currently the model def is just being utilized as the encoder for the full weights as loaded in transformers. There wasn't time to co-ordinate the other bits before release.

[rwightman]

It should be pointed out that the '300m' size is the only official google size as used in gemma 3n. The base definition there is my own scale down that I used for testing, validation of the architecture. I did some initial epochs of pretrain, etc in sanity checks. Though I might tweak / change that model def before any final weights are trained on my end.