1
2
3
4
5
6
7

from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained(
    model_dir,
    torch_dtype="auto",
    device_map="auto",
    local_files_only=True,
)

 1
 2
 3
 4
 5
 6
 7
 8
 9
10

# Python虚拟环境名：Qwen2.5，版本号：3.10
conda create -n Qwen2.5 python=3.10 -y

# 激活虚拟环境
conda activate Qwen2.5

# 下载Python依赖包
pip install torch
pip install "transformers>=4.43.1"
pip install "accelerate>=0.26.0"

1
2
3
4
5

# Git大文件系统
git lfs install

# 下载模型文件
git clone https://www.modelscope.cn/qwen/Qwen2.5-1.5B-Instruct.git Qwen2.5-1.5B-Instruct

1
2
3
4
5
6

# 切换到Git目录
cd Qwen2.5-1.5B-Instruct

# 中断继续下载
git lfs install
git lfs pull

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12

import os
from transformers import AutoConfig

# 加载配置文件
config_path = os.path.join('D:', os.path.sep, 'ModelSpace', 'Qwen2.5', 'Qwen2.5-1.5B-Instruct', 'config.json')
config = AutoConfig.from_pretrained(config_path)

# 查看词汇表大小，输出: 151936
print(config.vocab_size)

# 查看配置所有内容
print(config)

1
2
3
4
5
6
7
8

_import_structure = {
    # ....
    "models.auto": [
        # ....
        "AutoConfig",
        # ....
    ]
}

1
2
3
4
5

_import_structure = {
    # ...
    "configuration_auto": ["CONFIG_MAPPING", "MODEL_NAMES_MAPPING", "AutoConfig"],
    # ...
}

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11

CONFIG_MAPPING_NAMES = OrderedDict(
    [
        # ...
        ("qwen2", "Qwen2Config"),
        ("qwen2_audio", "Qwen2AudioConfig"),
        ("qwen2_audio_encoder", "Qwen2AudioEncoderConfig"),
        ("qwen2_moe", "Qwen2MoeConfig"),
        ("qwen2_vl", "Qwen2VLConfig"),
        # ...
    ]
)

1
2
3
4
5

class AutoModelForCausalLM(_BaseAutoModelClass):
    _model_mapping = MODEL_FOR_CAUSAL_LM_MAPPING


AutoModelForCausalLM = auto_class_update(AutoModelForCausalLM, head_doc="causal language modeling")

1
2
3
4
5
6
7
8

MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
    [
        # ...
        ("qwen2", "Qwen2ForCausalLM"),
        ("qwen2_moe", "Qwen2MoeForCausalLM"),
        # ...
    ]
)

1
2
3
4
5
6

model = AutoModelForCausalLM.from_pretrained(
    model_dir,
    torch_dtype="auto",
    device_map="auto",
    local_files_only=True,
)

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13

elif use_safetensors is not False and os.path.isfile(
    os.path.join(pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_NAME, variant))
):
    # Load from a safetensors checkpoint
    archive_file = os.path.join(
        pretrained_model_name_or_path, subfolder, _add_variant(SAFE_WEIGHTS_NAME, variant)
    )

# ...省略...

if is_local:
    logger.info(f"loading weights file {archive_file}")
    resolved_archive_file = archive_file

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20

if (
    is_safetensors_available()
    and isinstance(resolved_archive_file, str)
    and resolved_archive_file.endswith(".safetensors")
):
    with safe_open(resolved_archive_file, framework="pt") as f:
        metadata = f.metadata()

    if metadata.get("format") == "pt":
        pass

# ...省略...

from_pt = not (from_tf | from_flax)

# load pt weights early so that we know which dtype to init the model under
if from_pt:
    if not is_sharded and state_dict is None:
        # Time to load the checkpoint
        state_dict = load_state_dict(resolved_archive_file)

1
2
3
4
5
6

with ContextManagers(init_contexts):
    # Let's make sure we don't run the init function of buffer modules
    model = cls(config, *model_args, **model_kwargs)

# make sure we use the model's config since the __init__ call might have copied it
config = model.config

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44

elif from_pt:
    # restore default dtype
    if dtype_orig is not None:
        torch.set_default_dtype(dtype_orig)

    (
        model,
        missing_keys,
        unexpected_keys,
        mismatched_keys,
        offload_index,
        error_msgs,
    ) = cls._load_pretrained_model(
        model,
        state_dict,
        # ...省略...


# ...省略...

# Set model in evaluation mode to deactivate DropOut modules by default
model.eval()

# ...省略...

model.generation_config = GenerationConfig.from_pretrained(
    pretrained_model_name_or_path,
    cache_dir=cache_dir,
    force_download=force_download,
    resume_download=resume_download,
    proxies=proxies,
    local_files_only=local_files_only,
    token=token,
    revision=revision,
    subfolder=subfolder,
    _from_auto=from_auto_class,
    _from_pipeline=from_pipeline,
    **kwargs,
)

# ...省略...

if not is_fsdp_enabled() and not is_deepspeed_zero3_enabled():
    dispatch_model(model, **device_map_kwargs)