Avoid downloading unnecessary files for AutoConfig

[gitlawr]

To Reproduce

Run the following script

from modelscope import AutoConfig

ignore_file_pattern = [
    # The default
    r'\w+\.bin',
    r'\w+\.safetensors',
    r'\w+\.pth',
    r'\w+\.pt',
    r'\w+\.h5',
    r'\w+\.ckpt',
    # Additional
    r'\w+\.zip',
    r'\w+\.onnx',
]
AutoConfig.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct", ignore_file_pattern=ignore_file_pattern)

Result

  File "/opt/homebrew/lib/python3.11/site-packages/modelscope/utils/hf_util/patcher.py", line 233, in from_pretrained
    model_dir = get_model_dir(
                ^^^^^^^^^^^^^^
TypeError: modelscope.utils.hf_util.patcher._patch_pretrained_class.<locals>.get_model_dir() got multiple values for keyword argument 'ignore_file_pattern'

Notes:
It works in modelscope v1.20.1
It does not work in modelscope v1.24.1

Model hub related: @liuyhwangyh @tastelikefeet @wangxingjun778

[tastelikefeet]

Fixed by #1303
Currently this arg is an inner argument which does not support from the user side (to keep the same argument list with the one of transformers), so you can remove the ignore_file_pattern and retry

[gitlawr]

If it's not cusomizable and has the same parameters as transformers, can it have the same behavior as transformers? The problem I'm facing is that the modelscope lib downloads unnecessary files.

Examples:

######## CASE 1
from transformers import AutoConfig

# transformers downloads nothing
model = "nvidia/stt_kk_ru_fastconformer_hybrid_large"
AutoConfig.from_pretrained(model)

from modelscope import AutoConfig

# modelscope downloads the 'stt_kk_ru_fastconformer_hybrid_large.nemo' 458MB file.
model = "nv-community/stt_kk_ru_fastconformer_hybrid_large"
AutoConfig.from_pretrained(model)

######## CASE 2
from transformers import AutoConfig

# transformers downloads:
# - block_config.py
# - configuration_decilm.py
# - transformers_4_44_2__modeling_rope_utils.py
# - config.json
# - transformers_4_44_2__configuration_llama.py
model = "nvidia/Llama-3_1-Nemotron-Ultra-253B-v1"
AutoConfig.from_pretrained(model)

from modelscope import AutoConfig

# modelscope downloads:
# BIAS.md                                                                special_tokens_map.json
# EXPLAINABILITY.md                                                      tokenizer.json
# PRIVACY.md                                                             tokenizer_config.json
# README.md                                                              tokenizer_name.txt
# SAFETY&SECURITY.md                                                     training_flowchart.png
# __init__.py                                                            transformers_4_44_2__activations.py
# accuracy_plot.png                                                      transformers_4_44_2__cache_utils.py
# block_config.py                                                        transformers_4_44_2__configuration_llama.py
# config.json                                                            transformers_4_44_2__modeling_attn_mask_utils.py
# configuration.json                                                     transformers_4_44_2__modeling_flash_attention_utils_backward_compat.py
# configuration_decilm.py                                                transformers_4_44_2__modeling_outputs.py
# modeling_decilm.py                                                     transformers_4_44_2__modeling_rope_utils.py
# nemo_common.json                                                       transformers_4_44_2__pytorch_utils.py
# nemo_model_config.yaml                                                 variable_cache.py
model = "nv-community/Llama-3_1-Nemotron-Ultra-253B-v1"
AutoConfig.from_pretrained(model)

[tastelikefeet]

If it's not cusomizable and has the same parameters as transformers, can it have the same behavior as transformers? The problem I'm facing is that the modelscope lib downloads unnecessary files.

Examples:

######## CASE 1
from transformers import AutoConfig

# transformers downloads nothing
model = "nvidia/stt_kk_ru_fastconformer_hybrid_large"
AutoConfig.from_pretrained(model)

from modelscope import AutoConfig

# modelscope downloads the 'stt_kk_ru_fastconformer_hybrid_large.nemo' 458MB file.
model = "nv-community/stt_kk_ru_fastconformer_hybrid_large"
AutoConfig.from_pretrained(model)

######## CASE 2
from transformers import AutoConfig

# transformers downloads:
# - block_config.py
# - configuration_decilm.py
# - transformers_4_44_2__modeling_rope_utils.py
# - config.json
# - transformers_4_44_2__configuration_llama.py
model = "nvidia/Llama-3_1-Nemotron-Ultra-253B-v1"
AutoConfig.from_pretrained(model)

from modelscope import AutoConfig

# modelscope downloads:
# BIAS.md                                                                special_tokens_map.json
# EXPLAINABILITY.md                                                      tokenizer.json
# PRIVACY.md                                                             tokenizer_config.json
# README.md                                                              tokenizer_name.txt
# SAFETY&SECURITY.md                                                     training_flowchart.png
# __init__.py                                                            transformers_4_44_2__activations.py
# accuracy_plot.png                                                      transformers_4_44_2__cache_utils.py
# block_config.py                                                        transformers_4_44_2__configuration_llama.py
# config.json                                                            transformers_4_44_2__modeling_attn_mask_utils.py
# configuration.json                                                     transformers_4_44_2__modeling_flash_attention_utils_backward_compat.py
# configuration_decilm.py                                                transformers_4_44_2__modeling_outputs.py
# modeling_decilm.py                                                     transformers_4_44_2__modeling_rope_utils.py
# nemo_common.json                                                       transformers_4_44_2__pytorch_utils.py
# nemo_model_config.yaml                                                 variable_cache.py
model = "nv-community/Llama-3_1-Nemotron-Ultra-253B-v1"
AutoConfig.from_pretrained(model)

Ok, I understand your problem, we will update the code.