Support GPTQ on ChatGLM2-6B

[YIYANGCAI]

Type of Change

bug fix

Description

ChatGLM's source code, whose forwarding functions take several positional arguments, besides fixed "hidden_states". However, current GPTQ API only collects keyword arguments automatically. Thus, this pr is to add positional arguments automatic collection.

Expected Behavior & Potential Risk

Execute GPTQ quantization on ChatGLM2-6B automatically.

How has this PR been tested?

PreCI

Dependency Change?

no

[xin3he]

Is it possible that we use args and kwargs insteand of seperating hidden_state from them? I think it would be more robust.

neural-compressor/neural_compressor/adaptor/torch_utils/awq.py

Line 457 in 7472254

for args, kwargs in zip(self.total_block_args, self.total_block_kwargs):

[YIYANGCAI]

Is it possible that we use args and kwargs insteand of seperating hidden_state from them? I think it would be more robust.

neural-compressor/neural_compressor/adaptor/torch_utils/awq.py

Line 457 in 7472254

for args, kwargs in zip(self.total_block_args, self.total_block_kwargs):

Hi Xin, this is a good point, I will try it on multiple models.