ok Merge branch 'master' of github.com:modelscope/modelscope into release/1.19

modelscope/msdatasets/context/dataset_context_config.py

@@ -17,7 +17,8 @@ class DatasetContextConfig:
                  data_files: Union[str, Sequence[str],
                                    Mapping[str, Union[str, Sequence[str]]]],
                  download_mode: DownloadMode, cache_root_dir: str,
-                 use_streaming: bool, stream_batch_size: int, trust_remote_code: bool, **kwargs):
+                 use_streaming: bool, stream_batch_size: int,
+                 trust_remote_code: bool, **kwargs):
 
         self._download_config = None
         self._data_meta_config = None

modelscope/pipelines/nlp/llm_pipeline.py

@@ -7,7 +7,7 @@ from typing import Any, Callable, Dict, Generator, Iterator, List, Tuple, Union
 import json
 import numpy as np
 import torch
-from transformers import PreTrainedModel, PreTrainedTokenizer
+from transformers import AutoConfig, PreTrainedModel, PreTrainedTokenizer
 
 from modelscope import (AutoModelForCausalLM, AutoTokenizer, Pipeline,
                         snapshot_download)
@@ -170,9 +170,17 @@ class LLMPipeline(Pipeline, PipelineStreamingOutputMixin):
                  **kwargs):
         self.device_map = kwargs.pop('device_map', None)
         self.llm_framework = llm_framework
-        # TODO: qwen-int4 need 'cuda'/'auto' device_map.
-        if not self.device_map and 'qwen' in kwargs['model'].lower():
-            self.device_map = 'cuda'
+
+        if os.path.exists(kwargs['model']):
+            config = AutoConfig.from_pretrained(
+                kwargs['model'], trust_remote_code=True)
+            q_config = config.__dict__.get('quantization_config', None)
+            if q_config:
+                if q_config.get(
+                        'quant_method',
+                        'gptq') == 'gptq' and torch.cuda.device_count():
+                    self.device_map = 'cuda'
+
         self.torch_dtype = kwargs.pop('torch_dtype', None)
         self.ignore_file_pattern = kwargs.pop('ignore_file_pattern', None)
 

modelscope/utils/hf_util.py

@@ -6,6 +6,7 @@ from types import MethodType
 from typing import Dict, Literal, Optional, Union
 
 from transformers import AutoConfig as AutoConfigHF
+from transformers import AutoFeatureExtractor as AutoFeatureExtractorHF
 from transformers import AutoImageProcessor as AutoImageProcessorHF
 from transformers import AutoModel as AutoModelHF
 from transformers import AutoModelForCausalLM as AutoModelForCausalLMHF
@@ -14,6 +15,7 @@ from transformers import \
     AutoModelForSequenceClassification as AutoModelForSequenceClassificationHF
 from transformers import \
     AutoModelForTokenClassification as AutoModelForTokenClassificationHF
+from transformers import AutoProcessor as AutoProcessorHF
 from transformers import AutoTokenizer as AutoTokenizerHF
 from transformers import BatchFeature as BatchFeatureHF
 from transformers import BitsAndBytesConfig as BitsAndBytesConfigHF
@@ -161,6 +163,7 @@ def _patch_pretrained_class():
                                       ignore_file_pattern, **kwargs)
             return ori_get_config_dict(cls, model_dir, **kwargs)
 
+        PretrainedConfig.from_pretrained = from_pretrained
         PretrainedConfig.get_config_dict = get_config_dict
 
     def patch_model_base():
@@ -177,9 +180,54 @@ def _patch_pretrained_class():
 
         PreTrainedModel.from_pretrained = from_pretrained
 
+    def patch_image_processor_base():
+        """ Monkey patch AutoImageProcessorHF.from_pretrained to adapt to modelscope hub.
+        """
+        ori_from_pretrained = AutoImageProcessorHF.from_pretrained.__func__
+
+        @classmethod
+        def from_pretrained(cls, pretrained_model_name_or_path, *model_args,
+                            **kwargs):
+            model_dir = get_model_dir(pretrained_model_name_or_path, None,
+                                      **kwargs)
+            return ori_from_pretrained(cls, model_dir, *model_args, **kwargs)
+
+        AutoImageProcessorHF.from_pretrained = from_pretrained
+
+    def patch_auto_processor_base():
+        """ Monkey patch AutoProcessorHF.from_pretrained to adapt to modelscope hub.
+        """
+        ori_from_pretrained = AutoProcessorHF.from_pretrained.__func__
+
+        @classmethod
+        def from_pretrained(cls, pretrained_model_name_or_path, *model_args,
+                            **kwargs):
+            model_dir = get_model_dir(pretrained_model_name_or_path, None,
+                                      **kwargs)
+            return ori_from_pretrained(cls, model_dir, *model_args, **kwargs)
+
+        AutoProcessorHF.from_pretrained = from_pretrained
+
+    def patch_feature_extractor_base():
+        """ Monkey patch AutoFeatureExtractorHF.from_pretrained to adapt to modelscope hub.
+        """
+        ori_from_pretrained = AutoFeatureExtractorHF.from_pretrained.__func__
+
+        @classmethod
+        def from_pretrained(cls, pretrained_model_name_or_path, *model_args,
+                            **kwargs):
+            model_dir = get_model_dir(pretrained_model_name_or_path, None,
+                                      **kwargs)
+            return ori_from_pretrained(cls, model_dir, *model_args, **kwargs)
+
+        AutoFeatureExtractorHF.from_pretrained = from_pretrained
+
     patch_tokenizer_base()
     patch_config_base()
     patch_model_base()
+    patch_image_processor_base()
+    patch_auto_processor_base()
+    patch_feature_extractor_base()
 
 
 def patch_hub():

tests/msdatasets/test_general_datasets.py

@@ -121,7 +121,8 @@ class GeneralMsDatasetTest(unittest.TestCase):
                                                    f'please check the url: {py_script_url}'
 
             # Load the dataset
-            ds = MsDataset.load(py_script_file, subset_name='cola', split='train')
+            ds = MsDataset.load(
+                py_script_file, subset_name='cola', split='train')
             sample = next(iter(ds))
             logger.info(f'>>output of test_local_py_script:\n {sample}')
             assert sample