Add vllm awq loading logic

[ACupofAir]

Description

• We use an environment variable to get the group size. This is because within ipex-llm, we cannot get the quantization config.
• This pr is apply main's Add vllm awq loading logic #11950 to ipex-llm-mainline branch.

[ACupofAir]

Have test on docker image vllm-ipex-054:0903. There is no exception in the output.
Results:

1. chatglm3-6b 1 card:
   

2. llama2-13b 2 cards:
   

[gc-fu]

Have test on docker image vllm-ipex-054:0903. There is no exception in the output. Results:

1. chatglm3-6b 1 card:
   
2. llama2-13b 2 cards:
   

The result of chatglm3-6b seems wired for me, it might be caused by not setting export BIGDL_LLM_SDP_IGNORE_MASK=0.

Please set this environment variable and test again. Also, please adding one awq test result into the thread. For instance: https://huggingface.co/TheBloke/Llama-2-7B-Chat-AWQ

[ACupofAir]

Have test on docker image vllm-ipex-054:0903. There is no exception in the output. Results:

1. chatglm3-6b 1 card:
   
2. llama2-13b 2 cards:
   

The result of chatglm3-6b seems wired for me, it might be caused by not setting export BIGDL_LLM_SDP_IGNORE_MASK=0.

Please set this environment variable and test again. Also, please adding one awq test result into the thread. For instance: https://huggingface.co/TheBloke/Llama-2-7B-Chat-AWQ

All result is reasonable now.

1. verification for awq model(llama2-7b-chat-awq)
   
2. result for chatglm3-6b after export BIGDL_LLM_SDP_IGNORE_MASK=0
   

Attention: to run the awq model need apply this pr(analytics-zoo/vllm#29) for vllm 0.5.4

[gc-fu]

LGTM