Add MAGI-1 library #1386
Add MAGI-1 library #1386
Conversation
jiahy0825
requested review from
SBrandeis,
gary149,
Wauplin,
julien-c,
pcuenca and
ngxson
as code owners
There was a problem hiding this comment.
Hi @jiahy0825 - Thanks for the PR, couple of quick notes:
- I see that instead of one model per repo, there are multiple different models in the same repo: https://huggingface.co/sand-ai/MAGI-1/tree/main/ckpt/magi
- Since all the model repos are diffusers repos, the download tracking should automatically work if these checkpoints were split in 3 repos.
Let me know if you have any questions.
Hi, thanks a lot for your reply! We would like to keep all these models in a single repo because our model pipeline is relatively complex — it involves multiple components like T5, VAE, and others. Having all related models organized within one repo would make it much easier for users to find and use them together. Therefore, we would prefer to track downloads through this repo via the PR if possible. Please let me know if you have any better suggestions. Thanks again for your help! |
There was a problem hiding this comment.
in this case I'd defer to @Wauplin if we can make an exception. In general, we do strongly recommend having separate repos for separate models!
There was a problem hiding this comment.
Sorry for the delayed review. Agree with @Vaibhavs10 on the fact that it's usually best to split the different variants of a models in separate repos. If you don't want to reupload all t5/vae weights to each separate repo, you can update the Magi library to load the weights from individual repos except for the t5/vae backbones that would have to be fetched from the "main" MAGI-1 repo. By doing this, you'll have separate download counts for each variants (24B, 24B_distill, 24B_distill_quant, 4.5B). Also it would make them more easily discoverable on the Hub for the users.
But note that spliting repos is only a strong recommendation, not a hard requirements. If for your needs you believe it makes more sense to keep everything together then, let's keep it like this! Only the download count rule has to be updated to count only "1" when a user instanciate the model.
| prettyLabel: "MAGI-1", | ||
| repoName: "MAGI-1", | ||
| repoUrl: "https://github.com/SandAI-org/MAGI-1", | ||
| countDownloads: `path_extension:"json"`, |
There was a problem hiding this comment.
Doing this will likely double-count many downloads since there are several .json files needed to instantiate a model. What we usually do is to track downloads on a single file that is essential for the model to be loaded correctly. Could https://huggingface.co/sand-ai/MAGI-1/blob/main/ckpt/vae/config.json be used for that for instance?
| @@ -508,6 +508,12 @@ export const MODEL_LIBRARIES_UI_ELEMENTS = { | |||
| repoName: "mindspore", | |||
| repoUrl: "https://github.com/mindspore-ai/mindspore", | |||
| }, | |||
| "magi-1": { | |||
There was a problem hiding this comment.
as @Vaibhavs10 mentioned, you'll need to set library_name: "magi-1" in your model card metadata to make this work
This PR adds MAGI-1 library