fix(layers): avoid duplicate enable_proxy_error kwarg in make_quant_linear

[Dhruv Bhadauriya (ThinkerDesigns)]

Fixes #196

Root cause: make_quant_linear() unpacks **layer_conf (loaded from saved model config) while also passing enable_proxy_error=False as an explicit kwarg. When the saved config contains enable_proxy_error (from models quantized with the algorithm branch), Python raises TypeError: got multiple values for keyword argument 'enable_proxy_error'.

Fix: Filter out enable_proxy_error from the unpacked config before passing it, so only the explicit kwarg is used.

[Copilot]

Pull request overview

This PR updates the layer-replacement path used during AutoModelForCausalLM.from_pretrained() to prevent a Python TypeError when loading models whose saved quantization config already includes enable_proxy_error.

Changes:

• Copy and filter layer_conf to remove enable_proxy_error before unpacking **conf into the quantized linear constructor.
• Preserve the explicit enable_proxy_error=False kwarg to ensure only one value is provided (fixing the duplicate-kwarg crash).

Suppressed comments (1)

vptq/layers/model_base.py:52

• This change adds compatibility logic for configs that contain enable_proxy_error, but there’s no regression test ensuring make_quant_linear() can handle a saved layer_conf that includes enable_proxy_error without raising TypeError. Adding a unit/regression test would prevent this from reappearing (e.g., construct a tiny module with a Linear, pass config_for_layers containing enable_proxy_error, and assert replacement succeeds).

            # Pop enable_proxy_error to avoid duplicate kwarg when saved config
            # (from algorithm branch) also contains it — see issue #196
            conf = dict(layer_conf)
            conf.pop("enable_proxy_error", None)
            new_module = target_layer(
                **conf,
                enable_proxy_error=False,

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.