remove model_type, use model_arch

swift/llm/__init__.py

@@ -13,8 +13,8 @@
  from .train import sft_main, pt_main, rlhf_main
  from .argument import (EvalArguments, InferArguments, SftArguments, ExportArguments, DeployArguments, RLHFArguments,
  WebuiArguments, AppUIArguments)
- from .template import TEMPLATE_MAPPING, StopWords, get_template, TemplateType
- from .model import MODEL_MAPPING, ModelType, get_model_tokenizer, get_default_template_type, ConfigReader
+ from .template import TEMPLATE_MAPPING, Template, StopWords, get_template, TemplateType
+ from .model import MODEL_MAPPING, ModelType, get_model_tokenizer, get_default_template_type, HfConfigFactory
  from .dataset import (AlpacaPreprocessor, ClsPreprocessor, ComposePreprocessor, ConversationsPreprocessor,
  ListPreprocessor, PreprocessFunc, RenameColumnsPreprocessor, SmartPreprocessor,
  TextGenerationPreprocessor, DatasetName, DatasetLoader, HubDatasetLoader, LocalDatasetLoader,
@@ -41,7 +41,7 @@
  'RLHFArguments', 'AppUIArguments'
  ],
  'template': ['TEMPLATE_MAPPING', 'Template', 'StopWords', 'get_template', 'TemplateType'],
- 'model': ['MODEL_MAPPING', 'ModelType', 'get_model_tokenizer', 'get_default_template_type', 'ConfigReader'],
+ 'model': ['MODEL_MAPPING', 'ModelType', 'get_model_tokenizer', 'get_default_template_type', 'HfConfigFactory'],
  'dataset': [
  'AlpacaPreprocessor', 'ClsPreprocessor', 'ComposePreprocessor', 'ConversationsPreprocessor',
  'ListPreprocessor', 'PreprocessFunc', 'RenameColumnsPreprocessor', 'SmartPreprocessor',

swift/llm/argument/model_args.py

@@ -7,7 +7,8 @@
 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.llm import MODEL_KEYS_MAPPING, MODEL_MAPPING, ConfigReader
+from swift.llm import MODEL_KEYS_MAPPING, MODEL_MAPPING, RLHFArguments
+from swift.llm.model import fix_do_sample_warning
 from swift.utils import get_dist_setting, get_logger, use_hf_hub
 
 logger = get_logger()
@@ -35,12 +36,8 @@ def handle_do_sample(self) -> None:
  if self.temperature == 0:
  self.do_sample = False
  from swift.llm import InferArguments, SftArguments
- if self.do_sample is False and (isinstance(self, SftArguments) or
- (isinstance(self, InferArguments) and self.infer_backend == 'pt')):
- # fix warning
- self.temperature = 1.
- self.top_p = 1.
- self.top_k = 50
+ if (isinstance(self, SftArguments) or (isinstance(self, InferArguments) and self.infer_backend == 'pt')):
+ fix_do_sample_warning(self)
  logger.info('Due to do_sample=False, the following settings are applied: args.temperature: '
  f'{self.temperature}, args.top_p: {self.top_p}, args.top_k: {self.top_k}.')
 
@@ -86,19 +83,6 @@ def select_bnb(self) -> None:
  self.bnb_4bit_compute_dtype = bnb_4bit_compute_dtype
  self.load_in_4bit, self.load_in_8bit = load_in_4bit, load_in_8bit
 
- def is_quant_model(self: Union['SftArguments', 'InferArguments']) -> bool:
- """Judge if the current model has already been a quantized model"""
- # Check if the model is gptq, awq, aqlm model. Do not check for other quantization situations such as bnb.
- if self.model_type is not None:
- for k in ['int4', 'int8', 'awq', 'aqlm']:
- if k in self.model_type:
- return True
-
- model_path = self.model_id_or_path or self.resume_from_checkpoint or self.ckpt_dir
- bits = ConfigReader.read_config('quantization_config.bits', self.model_type, model_path, self.model_revision)
- if bits:
- return True
-
  def __post_init__(self: Union['SftArguments', 'InferArguments']):
  self.select_bnb()
 

swift/llm/dataset/dataset.py

@@ -177,9 +177,11 @@ class DatasetName:
  def get_dataset_name_list(cls) -> List[str]:
  res = []
  for k in cls.__dict__.keys():
- if k.startswith('__') or k == 'get_dataset_name_list':
+ if k.startswith('__'):
  continue
- res.append(cls.__dict__[k])
+ value = cls.__dict__[k]
+ if isinstance(value, str):
+ res.append(value)
  return res
 
 

swift/llm/infer/lmdeploy.py

@@ -21,7 +21,7 @@
 
 from swift.llm import InferArguments, InferTemplate
 from swift.llm.infer.base import InferFramework
-from swift.llm.model import ConfigReader
+from swift.llm.model import HfConfigFactory
 from swift.llm.model.model import get_model_tokenizer
 from swift.llm.template.template import Template, get_template
 from swift.utils import get_logger, get_seed
@@ -411,7 +411,7 @@ def get_lmdeploy_engine(
  lmdeploy_engine.is_multimodal = is_multimodal
  lmdeploy_engine.hf_tokenizer = tokenizer
  lmdeploy_engine.model_config = model_config
- lmdeploy_engine.max_model_len = ConfigReader.get_max_model_len(model_config)
+ lmdeploy_engine.max_model_len = HfConfigFactory.get_max_model_len(model_config)
 
  generation_config_path = os.path.join(model_dir, 'generation_config.json')
  if os.path.isfile(generation_config_path):

swift/llm/infer/transformers.py

@@ -15,10 +15,9 @@
 from transformers.utils import is_torch_npu_available
 
 from swift import get_logger
-from swift.llm import DeployArguments, InferArguments, StopWords, Template, get_model_tokenizer, get_template
+from swift.llm import (DeployArguments, HfConfigFactory, InferArguments, StopWords, Template, get_model_tokenizer,
+ get_template, to_device)
 from swift.llm.dataset.utils import safe_tokenizer_decode
-from swift.llm.model import ConfigReader
-from swift.llm.model.utils import to_device
 from swift.llm.template.base import StopWordsCriteria
 from swift.llm.utils import Messages, set_generation_config
 from swift.plugin.tuner import Tuner, extra_tuners
@@ -493,7 +492,7 @@ def _prepare_inputs(model: PreTrainedModel,
  generation_config.bos_token_id = tokenizer.bos_token_id
  if generation_config.max_new_tokens is not None:
  generation_config.max_length = 20 # fix max_length, max_new_tokens warning
- max_length = ConfigReader.get_max_model_len(model.config)
+ max_length = HfConfigFactory.get_max_model_len(model.config)
  if max_length and token_len + generation_config.max_new_tokens > max_length:
  generation_config.max_new_tokens = max_length - token_len
  if generation_config.max_new_tokens <= 0:

swift/llm/model/__init__.py

@@ -1,3 +1,14 @@
-from .config import ConfigReader
-from .loader import safe_snapshot_download
-from .model import MODEL_MAPPING, ModelType, get_default_template_type, get_model_tokenizer
+from .constant import LLMModelType, MLLMModelType, ModelType
+from .register import (MODEL_MAPPING, Model, ModelGroup, TemplateGroup, fix_do_sample_warning,
+ get_default_template_type, get_model_tokenizer)
+from .utils import HfConfigFactory, safe_snapshot_download
+
+
+def _register_files():
+ from . import qwen
+ from . import llama
+ # TODO
+ # from . import model
+
+
+_register_files()

swift/llm/model/constant.py

@@ -0,0 +1,62 @@
+from typing import List
+
+
+class LLMModelType:
+ qwen = 'qwen'
+ modelscope_agent = 'modelscope_agent'
+ qwen2 = 'qwen2'
+ qwen2_5 = 'qwen2_5'
+ qwen2_moe = 'qwen2_moe'
+
+ chatglm2 = 'chatglm2'
+ chatglm3 = 'chatglm3'
+ glm4 = 'glm4'
+
+ llama2 = 'llama2'
+ llama3 = 'llama3'
+ llama3_1 = 'llama3_1'
+ llama3_2 = 'llama3_2'
+ reflection_llama3_1 = 'reflection_llama3_1'
+ chinese_llama2 = 'chinese_llama2'
+ chinese_alpaca2 = 'chinese_alpaca2'
+ llama3_chinese = 'llama3_chinese'
+
+ longwriter_glm4 = 'longwriter_glm4'
+ longwriter_llama3_1 = 'longwriter_llama3_1'
+
+ atom = 'atom'
+
+ codefuse_qwen = 'codefuse_qwen'
+
+
+class MLLMModelType:
+ qwen_vl = 'qwen_vl'
+ qwen_audio = 'qwen_audio'
+ qwen2_vl = 'qwen2_vl'
+ qwen2_audio = 'qwen2_audio'
+
+ glm4v = 'glm4v'
+ llama3_2_vision = 'llama3_2_vision'
+ llama3_1_omni = 'llama3_1_omni'
+ idefics3_llama3 = 'idefics3_llama3'
+
+ llava1_5 = 'llava1_5'
+ llava1_6_mistral = 'llava1_6_mistral'
+ llava1_6_vicuna = 'llava1_6_vicuna'
+ llava1_6_yi = 'llava1_6_yi'
+ llava1_6_llama3_1 = 'llava1_6_llama3_1'
+ llava_next = 'llava_next'
+
+
+class ModelType(LLMModelType, MLLMModelType):
+
+ @classmethod
+ def get_model_name_list(cls) -> List[str]:
+ res = []
+ for k in cls.__dict__.keys():
+ if k.startswith('__'):
+ continue
+ value = cls.__dict__[k]
+ if isinstance(value, str):
+ res.append(value)
+ return res