feat: compatible with hf_pipeline (#1221)

compatible with hf_pipeline

modelscope/__init__.py

@@ -56,7 +56,7 @@ if TYPE_CHECKING:
             AutoModelForPreTraining, AutoModelForTextEncoding,
             AutoImageProcessor, BatchFeature, Qwen2VLForConditionalGeneration,
             T5EncoderModel, Qwen2_5_VLForConditionalGeneration, LlamaModel,
-            LlamaPreTrainedModel, LlamaForCausalLM)
+            LlamaPreTrainedModel, LlamaForCausalLM, hf_pipeline)
     else:
         print(
             'transformer is not installed, please install it if you want to use related modules'

modelscope/hub/check_model.py

@@ -71,6 +71,10 @@ def check_local_model_is_latest(
             headers=snapshot_header,
             use_cookies=cookies,
         )
+        model_cache = None
+        # download via non-git method
+        if not os.path.exists(os.path.join(model_root_path, '.git')):
+            model_cache = ModelFileSystemCache(model_root_path)
         for model_file in model_files:
             if model_file['Type'] == 'tree':
                 continue

modelscope/hub/git.py

@@ -46,6 +46,7 @@ class GitCommandWrapper(metaclass=Singleton):
         git_env = os.environ.copy()
         git_env['GIT_TERMINAL_PROMPT'] = '0'
         command = [self.git_path, *args]
+        command = [item for item in command if item]
         response = subprocess.run(
             command,
             stdout=subprocess.PIPE,

modelscope/pipelines/builder.py

@@ -10,6 +10,8 @@ from modelscope.utils.config import ConfigDict, check_config
 from modelscope.utils.constant import (DEFAULT_MODEL_REVISION, Invoke, Tasks,
                                        ThirdParty)
 from modelscope.utils.hub import read_config
+from modelscope.utils.import_utils import is_transformers_available
+from modelscope.utils.logger import get_logger
 from modelscope.utils.plugins import (register_modelhub_repo,
                                       register_plugins_repo)
 from modelscope.utils.registry import Registry, build_from_cfg
@@ -17,6 +19,7 @@ from .base import Pipeline
 from .util import is_official_hub_path
 
 PIPELINES = Registry('pipelines')
+logger = get_logger()
 
 
 def normalize_model_input(model,
@@ -72,7 +75,7 @@ def pipeline(task: str = None,
              config_file: str = None,
              pipeline_name: str = None,
              framework: str = None,
-             device: str = 'gpu',
+             device: str = None,
              model_revision: Optional[str] = DEFAULT_MODEL_REVISION,
              ignore_file_pattern: List[str] = None,
              **kwargs) -> Pipeline:
@@ -109,6 +112,7 @@ def pipeline(task: str = None,
     if task is None and pipeline_name is None:
         raise ValueError('task or pipeline_name is required')
 
+    pipeline_props = None
     if pipeline_name is None:
         # get default pipeline for this task
         if isinstance(model, str) \
@@ -157,8 +161,11 @@ def pipeline(task: str = None,
                 if pipeline_name:
                     pipeline_props = {'type': pipeline_name}
                 else:
-                    check_config(cfg)
+                    try:
-                    pipeline_props = cfg.pipeline
+                        check_config(cfg)
+                        pipeline_props = cfg.pipeline
+                    except AssertionError as e:
+                        logger.info(str(e))
 
         elif model is not None:
             # get pipeline info from Model object
@@ -166,9 +173,13 @@ def pipeline(task: str = None,
             if not hasattr(first_model, 'pipeline'):
                 # model is instantiated by user, we should parse config again
                 cfg = read_config(first_model.model_dir)
-                check_config(cfg)
+                try:
-                first_model.pipeline = cfg.pipeline
+                    check_config(cfg)
-            pipeline_props = first_model.pipeline
+                    first_model.pipeline = cfg.pipeline
+                except AssertionError as e:
+                    logger.info(str(e))
+            if first_model.__dict__.get('pipeline'):
+                pipeline_props = first_model.pipeline
         else:
             pipeline_name, default_model_repo = get_default_pipeline_info(task)
             model = normalize_model_input(default_model_repo, model_revision)
@@ -176,6 +187,23 @@ def pipeline(task: str = None,
     else:
         pipeline_props = {'type': pipeline_name}
 
+    if not pipeline_props and is_transformers_available():
+        try:
+            from modelscope.utils.hf_util import hf_pipeline
+            return hf_pipeline(
+                task=task,
+                model=model,
+                framework=framework,
+                device=device,
+                **kwargs)
+        except Exception as e:
+            logger.error(
+                'We couldn\'t find a suitable pipeline from ms, so we tried to load it using the transformers pipeline,'
+                ' but that also failed.')
+            raise e
+
+    if not device:
+        device = 'gpu'
     pipeline_props['model'] = model
     pipeline_props['device'] = device
     cfg = ConfigDict(pipeline_props)

modelscope/utils/hf_util/__init__.py

@@ -1,2 +1,3 @@
 from .auto_class import *
 from .patcher import patch_context, patch_hub, unpatch_hub
+from .pipeline_builder import hf_pipeline

modelscope/utils/hf_util/patcher.py

@@ -27,7 +27,8 @@ def get_all_imported_modules():
     transformers_include_names = [
         'Auto.*', 'T5.*', 'BitsAndBytesConfig', 'GenerationConfig', 'Awq.*',
         'GPTQ.*', 'BatchFeature', 'Qwen.*', 'Llama.*', 'PretrainedConfig',
-        'PreTrainedTokenizer', 'PreTrainedModel', 'PreTrainedTokenizerFast'
+        'PreTrainedTokenizer', 'PreTrainedModel', 'PreTrainedTokenizerFast',
+        'Pipeline'
     ]
     peft_include_names = ['.*PeftModel.*', '.*Config']
     diffusers_include_names = ['^(?!TF|Flax).*Pipeline$']
@@ -252,6 +253,44 @@ def _patch_pretrained_class(all_imported_modules, wrap=False):
                     model_dir, *model_args, **kwargs)
                 return module_obj
 
+            def save_pretrained(
+                self,
+                save_directory: Union[str, os.PathLike],
+                safe_serialization: bool = True,
+                **kwargs,
+            ):
+                push_to_hub = kwargs.pop('push_to_hub', False)
+                if push_to_hub:
+                    from modelscope.hub.push_to_hub import push_to_hub
+                    from modelscope.hub.api import HubApi
+                    from modelscope.hub.repository import Repository
+
+                    token = kwargs.get('token')
+                    commit_message = kwargs.pop('commit_message', None)
+                    repo_name = kwargs.pop(
+                        'repo_id',
+                        save_directory.split(os.path.sep)[-1])
+
+                    api = HubApi()
+                    api.login(token)
+                    api.create_repo(repo_name)
+                    # clone the repo
+                    Repository(save_directory, repo_name)
+
+                super().save_pretrained(
+                    save_directory=save_directory,
+                    safe_serialization=safe_serialization,
+                    push_to_hub=False,
+                    **kwargs)
+
+                # Class members may be unpatched, so push_to_hub is done separately here
+                if push_to_hub:
+                    push_to_hub(
+                        repo_name=repo_name,
+                        output_dir=save_directory,
+                        commit_message=commit_message,
+                        token=token)
+
         if not hasattr(module_class, 'from_pretrained'):
             del ClassWrapper.from_pretrained
         else:
@@ -266,6 +305,9 @@ def _patch_pretrained_class(all_imported_modules, wrap=False):
         if not hasattr(module_class, 'get_config_dict'):
             del ClassWrapper.get_config_dict
 
+        if not hasattr(module_class, 'save_pretrained'):
+            del ClassWrapper.save_pretrained
+
         ClassWrapper.__name__ = module_class.__name__
         ClassWrapper.__qualname__ = module_class.__qualname__
         return ClassWrapper
@@ -289,12 +331,16 @@ def _patch_pretrained_class(all_imported_modules, wrap=False):
             has_from_pretrained = hasattr(var, 'from_pretrained')
             has_get_peft_type = hasattr(var, '_get_peft_type')
             has_get_config_dict = hasattr(var, 'get_config_dict')
+            has_save_pretrained = hasattr(var, 'save_pretrained')
         except:  # noqa
             continue
 
-        if wrap:
+        # save_pretrained is not a classmethod and cannot be overridden by replacing
+        # the class method. It requires replacing the class object method.
+        if wrap or ('pipeline' in name.lower() and has_save_pretrained):
             try:
-                if not has_from_pretrained and not has_get_config_dict and not has_get_peft_type:
+                if (not has_from_pretrained and not has_get_config_dict
+                        and not has_get_peft_type and not has_save_pretrained):
                     all_available_modules.append(var)
                 else:
                     all_available_modules.append(

modelscope/utils/hf_util/pipeline_builder.py

@@ -0,0 +1,54 @@
+import os
+from typing import Optional, Union
+
+import torch
+from transformers import Pipeline as PipelineHF
+from transformers import PreTrainedModel, TFPreTrainedModel, pipeline
+from transformers.pipelines import check_task, get_task
+
+from modelscope.hub import snapshot_download
+from modelscope.utils.hf_util.patcher import _patch_pretrained_class, patch_hub
+
+
+def _get_hf_device(device):
+    if isinstance(device, str):
+        device_name = device.lower()
+        eles = device_name.split(':')
+        if eles[0] == 'gpu':
+            eles = ['cuda'] + eles[1:]
+            device = ''.join(eles)
+    return device
+
+
+def _get_hf_pipeline_class(task, model):
+    if not task:
+        task = get_task(model)
+    normalized_task, targeted_task, task_options = check_task(task)
+    pipeline_class = targeted_task['impl']
+    pipeline_class = _patch_pretrained_class([pipeline_class])[0]
+    return pipeline_class
+
+
+def hf_pipeline(
+    task: str = None,
+    model: Optional[Union[str, 'PreTrainedModel', 'TFPreTrainedModel']] = None,
+    framework: Optional[str] = None,
+    device: Optional[Union[int, str, 'torch.device']] = None,
+    **kwargs,
+) -> PipelineHF:
+    if isinstance(model, str):
+        if not os.path.exists(model):
+            model = snapshot_download(model)
+
+    framework = 'pt' if framework == 'pytorch' else framework
+
+    device = _get_hf_device(device)
+    pipeline_class = _get_hf_pipeline_class(task, model)
+
+    return pipeline(
+        task=task,
+        model=model,
+        framework=framework,
+        device=device,
+        pipeline_class=pipeline_class,
+        **kwargs)

tests/utils/test_hf_util.py

@@ -40,6 +40,14 @@ class HFUtilTest(unittest.TestCase):
         with open(self.test_file2, 'w') as f:
             f.write('{}')
 
+        self.pipeline_qa_context = r"""
+            Extractive Question Answering is the task of extracting an answer from a text given a question. An example
+            of a question answering dataset is the SQuAD dataset, which is entirely based on that task. If you would
+            like to fine-tune a model on a SQuAD task, you may leverage the
+            examples/pytorch/question-answering/run_squad.py script.
+            """
+        self.pipeline_qa_question = 'What is a good example of a question answering dataset?'
+
     def tearDown(self):
         logger.info('TearDown')
         shutil.rmtree(self.model_dir, ignore_errors=True)
@@ -235,6 +243,59 @@ class HFUtilTest(unittest.TestCase):
                 'Qwen/Qwen1.5-0.5B-Chat', trust_remote_code=True)
             model.push_to_hub(self.create_model_name)
 
+    @unittest.skipUnless(test_level() >= 1, 'skip test in current test level')
+    def test_pipeline_model_id(self):
+        from modelscope import pipeline
+        model_id = 'damotestx/distilbert-base-cased-distilled-squad'
+        qa = pipeline('question-answering', model=model_id)
+        assert qa(
+            question=self.pipeline_qa_question,
+            context=self.pipeline_qa_context)
+
+    @unittest.skipUnless(test_level() >= 1, 'skip test in current test level')
+    def test_pipeline_auto_model(self):
+        from modelscope import pipeline, AutoModelForQuestionAnswering, AutoTokenizer
+        model_id = 'damotestx/distilbert-base-cased-distilled-squad'
+        model = AutoModelForQuestionAnswering.from_pretrained(model_id)
+        tokenizer = AutoTokenizer.from_pretrained(model_id)
+        qa = pipeline('question-answering', model=model, tokenizer=tokenizer)
+        assert qa(
+            question=self.pipeline_qa_question,
+            context=self.pipeline_qa_context)
+
+    @unittest.skipUnless(test_level() >= 1, 'skip test in current test level')
+    def test_pipeline_save_pretrained(self):
+        from modelscope import pipeline
+        model_id = 'damotestx/distilbert-base-cased-distilled-squad'
+
+        pipe_ori = pipeline('question-answering', model=model_id)
+
+        result_ori = pipe_ori(
+            question=self.pipeline_qa_question,
+            context=self.pipeline_qa_context)
+
+        # save_pretrained
+        repo_id = self.create_model_name
+        save_dir = './tmp_test_hf_pipeline'
+        try:
+            os.system(f'rm -rf {save_dir}')
+            self.api.delete_model(repo_id)
+            # wait for delete repo
+            import time
+            time.sleep(5)
+        except Exception:
+            # if repo not exists
+            pass
+        pipe_ori.save_pretrained(save_dir, push_to_hub=True, repo_id=repo_id)
+
+        # load from saved
+        pipe_new = pipeline('question-answering', model=repo_id)
+        result_new = pipe_new(
+            question=self.pipeline_qa_question,
+            context=self.pipeline_qa_context)
+
+        assert result_new == result_ori
+
 
 if __name__ == '__main__':
     unittest.main()