Hugging face model handler #3 (#38696)

Author: derrickaw

.github/trigger_files/beam_PostCommit_Yaml_Xlang_Direct.json

@@ -1,4 +1,4 @@
 {
   "comment": "Modify this file in a trivial way to cause this test suite to run",
-  "revision": 2
+  "revision": 3
 }

sdks/python/apache_beam/yaml/tests/runinference_huggingface.yaml

@@ -0,0 +1,62 @@
+# Licensed to the Apache Software Foundation (ASF) under one or more
+# contributor license agreements.  See the NOTICE file distributed with
+# this work for additional information regarding copyright ownership.
+# The ASF licenses this file to You under the Apache License, Version 2.0
+# (the "License"); you may not use this file except in compliance with
+# the License.  You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+pipelines:
+  - pipeline:
+      type: chain
+      transforms:
+        - type: Create
+          config:
+            elements:
+              - text: "I love Apache Beam!"
+              - text: "I hate this error."
+        - type: RunInference
+          config:
+            model_handler:
+              type: "HuggingFacePipeline"
+              config:
+                task: "text-classification"
+                inference_fn:
+                  callable: |
+                    def real_inference(batch, pipeline, inference_args):
+                      predictions = pipeline(batch, **inference_args) 
+                      
+                      # If it's a single dictionary (batch size of 1), wrap it in a list
+                      if isinstance(predictions, dict):
+                        predictions = [predictions]
+                      
+                      return {
+                        'label': [p['label'] for p in predictions],
+                        'score': [p['score'] for p in predictions]
+                      }
+                preprocess:
+                  callable: 'lambda x: x.text'
+        - type: MapToFields
+          config:
+            language: python
+            fields:
+              text: text
+              sentiment:
+                callable: 'lambda x: x.inference.inference["label"]'
+        - type: AssertEqual
+          config:
+            elements:
+              - text: "I love Apache Beam!"
+                sentiment: "POSITIVE"
+              - text: "I hate this error."
+                sentiment: "NEGATIVE"
+
+    options:
+      yaml_experimental_features: ['ML']

…apache_beam/yaml/tests/runinference.yaml …am/yaml/tests/runinference_vertexai.yamlsdks/python/apache_beam/yaml/tests/runinference.yaml renamed to sdks/python/apache_beam/yaml/tests/runinference_vertexai.yaml sdks/python/apache_beam/yaml/tests/runinference.yaml renamed to sdks/python/apache_beam/yaml/tests/runinference_vertexai.yaml

@@ -282,6 +282,55 @@ def inference_output_type(self):
                                           ('model_id', Optional[str])])
 
 
+@ModelHandlerProvider.register_handler_type('HuggingFacePipeline')
+class HuggingFacePipelineProvider(ModelHandlerProvider):
+  def __init__(
+      self,
+      task: Optional[str] = None,
+      model: Optional[str] = None,
+      preprocess: Optional[dict[str, str]] = None,
+      postprocess: Optional[dict[str, str]] = None,
+      device: Optional[Any] = None,
+      inference_fn: Optional[dict[str, str]] = None,
+      load_pipeline_args: Optional[dict[str, Any]] = None,
+      **kwargs):
+    try:
+      from apache_beam.ml.inference.huggingface_inference import HuggingFacePipelineModelHandler
+    except ImportError:
+      raise ValueError(
+          'Unable to import HuggingFacePipelineModelHandler. Please '
+          'install transformers dependencies.')
+
+    kwargs = {k: v for k, v in kwargs.items() if not k.startswith('_')}
+
+    inference_fn_obj = self.parse_processing_transform(
+        inference_fn, 'inference_fn') if inference_fn else None
+
+    handler_kwargs = {}
+    if inference_fn_obj:
+      handler_kwargs['inference_fn'] = inference_fn_obj
+
+    _handler = HuggingFacePipelineModelHandler(
+        task=task,
+        model=model,
+        device=device,
+        load_pipeline_args=load_pipeline_args,
+        **handler_kwargs,
+        **kwargs)
+
+    super().__init__(_handler, preprocess, postprocess)
+
+  @staticmethod
+  def validate(config):
+    if not config or (not config.get('task') and not config.get('model')):
+      raise ValueError(
+          "HuggingFacePipeline requires either 'task' or "
+          "'model' to be specified.")
+
+  def inference_output_type(self):
+    return Any
+
+
 @beam.ptransform.ptransform_fn
 def run_inference(
     pcoll,

sdks/python/apache_beam/yaml/yaml_ml.py

@@ -654,7 +654,8 @@ def get_portability_package_data():
           'transformers': [
               'transformers>=4.28.0,<4.56.0',
               'tensorflow>=2.12.0',
-              'torch>=1.9.0'
+              # Avoid torch 2.12.0+ which fails to run unit tests with segfault
+              'torch>=1.9.0,<2.12.0'
           ],
           'ml_cpu': [
               'tensorflow>=2.12.0',