add save & load support for NPU optimized model

[rnwang04]

Description

1. Why the change?

• fix support of NPU save & load func
• add save & load support for NPU optimized model

2. User API changes

• for non-optimized model

# save.py
from ipex_llm.transformers.npu_model import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True,
                                             load_in_low_bit="sym_int4",
                                             attn_implementation="eager")
model.save_low_bit("llama_low_bit_npu")

# load.py
model = AutoModelForCausalLM.load_low_bit("llama_low_bit_npu",
                                          attn_implementation="eager")

• for optimized model

# save.py
from ipex_llm.transformers.npu_model import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained(
    model_path,
    torch_dtype=torch.float16,
    trust_remote_code=True,
    attn_implementation="eager",
    load_in_low_bit="sym_int4",
    optimize_model=True,
    max_output_len=args.max_output_len,
    max_prompt_len=args.max_prompt_len,
    intra_pp=args.intra_pp,
    inter_pp=args.inter_pp,
    transpose_value_cache=not args.disable_transpose_value_cache,
)

model.save_low_bit("llama_low_bit_npu")

# load.py
model = AutoModelForCausalLM.load_low_bit("llama_low_bit_npu",
                                          attn_implementation="eager",
                                          torch_dtype=torch.float16,
                                          optimize_model=True,
                                          max_output_len=args.max_output_len,
                                          max_prompt_len=args.max_prompt_len,
                                          intra_pp=args.intra_pp,
                                          inter_pp=args.inter_pp,
                                          transpose_value_cache=not args.disable_transpose_value_cache)

Remain issues:

• for optimized model, save.py & load.py must be different scirpt, otherwise program will hang at loading stage .

4. How to test?

• Unit test: Please manually trigger the PR Validation here by inputting the PR number (e.g., 1234). And paste your action link here once it has been successfully finished.

[plusbang]

LGTM, maybe we could also add related example later.

[rnwang04]

LGTM, maybe we could also add related example later.

Yeah we can add the usage in our NPU examples in later PR.

[rnwang04]

PR validation is passed: https://github.com/intel-analytics/ipex-llm-workflow/actions/runs/10680656064