Fixes #271 Issue

[LukaDarsalia]

Checklist

• Read CONTRIBUTING.md, and accept the CLA by including the provided snippet. We will not accept PR without this.
• Run pre-commit hook.
• If you changed Rust code, run cargo check, cargo clippy, cargo test.

PR Description

Fixes #271

Summary

This PR fixes a crash in the quantized PyTorch Moshi models when loaded on Windows/Linux due to meta tensors being passed into int8_vectorwise_quant.

The issue happens because replace_linear_with_qlinear() was called before real weights were loaded, resulting in meta tensors being quantized, which bitsandbytes does not support.

Fix Details

• In QLinear.__init__, we now check if the weight is on the meta device.
• If it is, we create dummy CB and SCB tensors with correct shapes and set a flag self.is_meta = True.
• We delay quantization until real weights are loaded.
• The forward pass checks the meta status and raises a clear error if used too early.
• Once weights are loaded, _check_meta_status will turn off meta mode and ensure scale tensors are in float32.

CLA

I, LukaDarsalia, confirm that I have read and understood the terms of the CLA of Kyutai-labs, as outlined in the repository's CONTRIBUTING.md, and I agree to be bound by these terms.