Save error when quanting qwen2 with group_size=-1 in main branch

[zzh-www]

When reload quanted model, i found bias, qweight all are zero tensor.
You can run examples/quantization/basic_usage_wikitext2.py by changing code like this to reproduce it:

# use qwen2 model
pretrained_model_id = "Qwen2.5-3B-Instruct"
quantized_model_id = "Qwen2.5-3B-Instruct-int4-gptq-by-wikitext-test"

quantize_config = QuantizeConfig(
  bits=4,
  group_size=-1, # change group_size from 128 to -1 for reproducing
)

[Qubitium]

Thanks for reporting. This should be a packing error caused by recent refractors. Also, none of our ci tests checked for group_size=-1 😅

[zzh-www]

Using groupsize=-1 causes very little loss on some tasks, and the performance will be better. Hopefully, this can be fixed soon.❤️

[Qubitium]

@zzh-www Fixed. Please confirm.