Handle strict argument in load_weights for better model compatibility

[kamillobinski]

Some models within the MLX ecosystem, especially custom or older forks define load_weights(weights) without a strict keyword argument. With recent updates, utils.load_model() passes strict=True unconditionally which causes runtime errors when loading such models.

This PR inspects the load_weights signature at runtime using pythons inspect module and passes the strict parameter only if it’s explicitly supported.

This allows:

• full compatibility with legacy/custom models,
• smooth usage of updated utils across projects,
• forward compatibility with future models that may redefine the method.

No functionality is altered for models that already support strict - this simply makes the system more robust when loading models from various sources.

[Blaizzy]

Hey Kamil

Thanks!

Could you share an example of a model that fails with strict true?

[Blaizzy]

Because strict here is to check if all model weights match the parameters properly

https://ml-explore.github.io/mlx/build/html/python/nn/_autosummary/mlx.nn.Module.load_weights.html

[kamillobinski]

Hey! You are right, strict=True is valid per nn.Module.load_weights() and most models do support it. That said, some models override load_weights without carrying over the strict parameter. One example is:

# mlx-audio/tts/models/sesame/model.py

def load_weights(self, weights):
    self.model.load_weights(weights)

Calling utils.load_model(..., strict=True) for mlx-community/csm-1b-fp16 on that throws:

TypeError: Model.load_weights() got an unexpected keyword argument 'strict'

This patch guards that call using inspect to check if strict is accepted. It avoids runtime crashes and ensures downstream compatibility. A small, safe fallback.

Let me know if you'd prefer a different fix.

[lucasnewman]

I think a quick tweak to the Sesame implementation of load_weights() that passed through the strict parameter would be best — it was just an oversight from when I added it originally, and we definitely want the parameter validation in general.

[kamillobinski]

Thanks Lucas, thats fair. Just pushed an update that adds strict to the sesame models load_weights implementation.
I kept the inspect fallback in utils.load_model as a precaution, but feel free to remove it or let me know and i will revert those lines cleanly.

Appreciate the quick feedback!

[Blaizzy]

Thanks @kamillobinski and @lucasnewman!