[Feature request]: Interrogate model file

[MrDrMcCoy]

Currently, there is a lot of guesswork in getting models to work. Users have to fend for themselves in many cases to determine which text encoders, VAE, etc are supported or included, and sometimes even the quantization level is obscured until you get the log output from a working run.

The feature I would like to see here is something like --model-info, which will print every available metadata of what a model is, has, and needs to work.

[leejet]

When you’re not yet familiar with the model architecture, the best way to start is by looking at the examples — there are corresponding examples for flux, sd3, wan, and qwen-image in the dos directory.

[MrDrMcCoy]

I think there's some misunderstanding here. The docs you refer to are for usage of known model files, and do little to cover what a model file is structurally comprised of. I am looking for a utility function that will examine an improperly documented model file and provide as much detail as possible for what it is made of. For example:

• What model family does it belong to?
• What text encoders, etc does it require?
• What beyond the base model has been embedded in it?
• What is its quantization level?

This will directly assist with usage and allow users to more easily search for what they need to know if they need to dig deeper. As it stands today, a standalone safetensors or gguf file is entirely opaque.

[leejet]

When you obtain a model's weights from the internet, such as from Civitai or Hugging Face, you should also be able to find on the website which model architecture these weights belong to.

[MrDrMcCoy]

That is true, but many of those models don't say what they include or what they need for text encoders, etc to be used. Is the ability to examine a model file really useless to everyone but me?