fix(gptq): fix Hessian computation, variable-length sequence support, and layer output type handling

[sunnyxiaohu]

Summary

Fix critical bugs in the GPTQ quantization pipeline that cause incorrect quantization results, especially for MoE models and variable-length calibration data.

Problems

1. Hessian matrix corruption for MoE experts — add_batch() uses parameterless squeeze() which collapses [1, dim] to [dim] when an expert receives only 1 routed token, causing nsamples to be incorrectly accumulated as feature_dim instead of 1.

2. Variable-length sequence incompatibility — Catcher pre-allocates a fixed-size tensor [nsamples, seq_length, hidden_size], requiring all samples to have identical seq_len. Shorter sequences get zero-padded (introducing Hessian noise) and longer sequences are silently truncated.

3. Layer output type mismatch — Unconditional layer(...)[0] assumes tuple output, but some decoder layers return a plain tensor. [0] then incorrectly indexes the batch dimension.

4. ignore_layers exact match fails for MoE — Nested module names like mlp.experts.0.gate_proj cannot match the ignore pattern gate_proj with exact equality.

5. _make_quant AttributeError on non-standard Linear — Modules like TopKRouter lack in_features/out_features/bias attributes, causing crashes during weight replacement.

6. g_idx generation uses slow Python list comprehension — Replaced with vectorized tensor operations.

Changes

File	Fix
gptq_module.py	Remove dangerous squeeze(), fix add_batch() reshape logic, vectorize g_idx
catcher.py	Rewrite to dynamic list storage with per-sample kwargs and max_seq_length VRAM guard
gptq.py	Add _extract_hidden_states() helper, per-sample forward loop, substring ignore_layers matching
helper_layer.py	Use getattr(linear, "bias", None) for non-standard Linear modules

Testing

• Verified on Qwen3-30B-A3B (MoE, variable expert routing)
• Verified on standard dense models (Qwen3-4B)
• No regression on existing quantization quality metrics

[yghstill]

@sunnyxiaohu
Please per-commit code formatting:

pip3 install pre-commit black isort flake8
cd AngelSlim
pre-commit install

[sunnyxiaohu]

@sunnyxiaohu Please per-commit code formatting:

pip3 install pre-commit black isort flake8
cd AngelSlim
pre-commit install

fixed