The new version 0.3.0 takes a long time for quantization and eventually fails due to OOM

[okwinds]

Describe the bug
I used the sample code (W8A16) to quantize THUDM/glm-4-9b-chat-hf on an Nvidia 4090, and the entire process was very slow (nearly 24 hours), with extremely high memory usage, to the point where an Out of Memory (OOM) error occurred at the final step. When the OOM occurred, there were no obvious error messages, only displayed.

[1] 216936 killed python3 test_ct.py
WSL environment:
compressed-tensors 0.8.0
llmcompressor 0.3.0

Memory : 47 GB
Swap : 40 GB

Using the same example code and consistent environment, the versions were updated to compressed-tensors 0.7.0 and llmcompressor 0.2.0. The quantization process was completed smoothly, and it took only 2 hours.

Expected behavior
Hoping for performance improvements.

Environment

1. OS Ubuntu 22.04
2. Python version 3.11.9
3. CUDA Version 12.4.1

Errors
OOM
[1] 216936 killed python3 test_ct.py