Loading Model Safetensors

[Aliasthump]

Why is it that every time I open Zonos it downloads the 3.3GB safetensors? Why does it have to do it every time? Are they deleted after I close out? Is it just piling up on Model.safetensors and wasting HD space? Also, where is it downloading to? I want to see it on my hard drive.

[darkacorn]

i assume you use the docker ? that is probably a mistake as the hf cache path isnt mapped throw - docker files by default are stateless cross restarts

there should be either a volume mount or a -v to the local hf cache

[dillonroach]

@Aliasthump in whatever app you're using, go digging to where it instantiates the model and replace the load with something like this (it's essentially what the original code does anyway, it just does the hf hub download to cache first, then this):

        model_path = "/<path-to-folder>/Zonos-v0.1-transformer/model.safetensors"
        config_path = "/<path-to-folder>/Zonos-v0.1-transformer/config.json"
        self.model = Zonos.from_local(config_path=config_path, model_path=model_path, device="cuda")
        self.model.requires_grad_(False).eval()

Works the same for hybrid, just change the paths. I much prefer having the model files/locations under my own control and not let hub do its thing.
'self.model' here is just where I happen to stuff it, by the way, that can be replaced with whatever the app is looking for.

[MithrilMan]

Since I play a lot with different docker containers that uses huggingface cache, usually I map the hub to a specific docker host path like this example that uses relative path

 volumes:
      - ./../_VOLUMES/huggingface-home:/workspace/huggingface