update

swift/llm/dataset/preprocess/core.py

@@ -4,7 +4,6 @@
 from functools import partial
 from typing import Any, Callable, Dict, List, Literal, Optional, Set, Union
 
-import numpy as np
 from datasets import Dataset as HfDataset
 from datasets import IterableDataset as HfIterableDataset
 

swift/llm/model/model_meta.py

@@ -1,5 +1,15 @@
+import itertools
 from dataclasses import dataclass, field
-from typing import Optional, List
+from typing import Callable, List, Optional, Tuple, TypeVar, Union
+
+from transformers import PreTrainedModel, PreTrainedTokenizerBase
+from transformers.utils.versions import require_version
+
+from swift.utils import get_logger
+
+logger = get_logger()
+GetModelTokenizerFunction = Callable[..., Tuple[Optional[PreTrainedModel], PreTrainedTokenizerBase]]
+
 
 @dataclass
 class Model:
@@ -52,6 +62,7 @@ class ModelMeta:
     support_megatron: bool = False
 
     def get_matched_model_groups(self, model_dir: str) -> List[ModelGroup]:
+        from .utils import HfConfigFactory
         model_name = HfConfigFactory._get_model_name(model_dir).lower()
         res = []
         seen = set()

swift/llm/model/register.py

@@ -14,16 +14,15 @@
                           PreTrainedModel, PreTrainedTokenizerBase)
 from transformers.integrations import is_deepspeed_zero3_enabled
 from transformers.utils import is_torch_bf16_gpu_available, is_torch_cuda_available, is_torch_npu_available
-from transformers.utils.versions import require_version
 
 from swift.utils import get_dist_setting, get_logger, is_ddp_plus_mp, is_dist, is_unsloth_available, use_torchacc
+from .model_meta import ModelMeta
 from .utils import AttnImpl, HfConfigFactory, safe_snapshot_download
 
 MODEL_MAPPING: Dict[str, Dict[str, Any]] = {}
 
 ARCH_MAPPING: Optional[Dict[str, Dict[str, List[str]]]] = None
 
-GetModelTokenizerFunction = Callable[..., Tuple[Optional[PreTrainedModel], PreTrainedTokenizerBase]]
 logger = get_logger()
 
 
@@ -237,13 +236,11 @@ def get_model_tokenizer(
     if download_model is None:
         download_model = load_model
     # download config.json
-    model_dir = safe_snapshot_download(
-                model_id_or_path, revision=revision, download_model=False, use_hf=use_hf)
+    model_dir = safe_snapshot_download(model_id_or_path, revision=revision, download_model=False, use_hf=use_hf)
     model_info = HfConfigFactory.get_model_info(model_dir)
 
     if download_model:
-        safe_snapshot_download(
-            model_id_or_path, revision=revision, download_model=download_model, use_hf=use_hf)
+        safe_snapshot_download(model_id_or_path, revision=revision, download_model=download_model, use_hf=use_hf)
 
     if not use_torchacc() and device_map is None:
         device_map = get_default_device_map()

swift/llm/model/utils.py

@@ -15,9 +15,11 @@
 
 from swift.hub import HFHub, MSHub, default_hub
 from swift.utils import deep_getattr, get_logger, is_dist, is_dist_ta, safe_ddp_context
+from .model_meta import ModelMeta
 
 logger = get_logger()
 
+
 @dataclass
 class ModelInfo:
     model_meta: ModelMeta
@@ -28,6 +30,7 @@ class ModelInfo:
     quant_bits: int
     config: PretrainedConfig
 
+
 class HfConfigFactory:
     """This class is used to read config from config.json(maybe params.json also)"""