Bhavanatumma/ta2614748 (Azure#1294)

* [bhavanatumma/ta2614748] - import model modification for OSS models * [bhavanatumma/ta2614748] - Converting to MLFlow OSS transformers flavor * [bhavanatumma/ta2614748] - Converting to MLFlow OSS transformers flavor * [bhavanatumma/ta2614748] - Converting to MLFlow OSS transformers flavor * [bhavanatumma/ta2614748] - Converting to MLFlow OSS transformers flavor - lint fix * [bhavanatumma/ta2614748] - Converting to MLFlow OSS transformers flavor - v2, adding conda list approach * [bhavanatumma/ta2614748] - lint fix * [bhavanatumma/ta2614748] - review comments * [bhavanatumma/ta2614748] - current progress * [bhavanatumma/ta2614748] - current progress v2 * Merge branch 'main' of https://github.com/Azure/azureml-assets into bhavanatumma/ta2614748 # Conflicts: Resolved # assets/common/components/mlflow_model_local_validation/spec.yaml # assets/training/model_management/components/convert_model_to_mlflow/spec.yaml # assets/training/model_management/components/import_model/spec.yaml * [bhavanatumma/ta2614748] - ref point * [bhavanatumma/ta2614748] - ref point * [bhavanatumma/ta2614748] - ref point v2 * [bhavanatumma/ta2614748] - added OSS and HF transformers support variable - model flavor * [bhavanatumma/ta2614748] - lint fixes * [bhavanatumma/ta2614748] - test fixes * [bhavanatumma/ta2614748] - test fixes * [bhavanatumma/ta2614748] - test fixes * [bhavanatumma/ta2614748] - review comments and adding Whisper model * [bhavanatumma/ta2614748] - review comments and adding Whisper model * [bhavanatumma/ta2614748] - changing version * [bhavanatumma/ta2614748] - changing versions * [bhavanatumma/ta2614748] - changing signatures * [bhavanatumma/ta2614748] - adding new env * [bhavanatumma/ta2614748] - adding new task text2text * [bhavanatumma/ta2614748] - adding new task text2text * [bhavanatumma/ta2614748] - supporting t2t tasks * [bhavanatumma/ta2614748] - whisper changes * [bhavanatumma/ta2614748] - whisper changes + convertors * [bhavanatumma/ta2614748] - fixing tests * [bhavanatumma/summarization-oss] - version change * [bhavanatumma/ta2614748] - lint fix * [bhavanatumma/ta2614748] - signature fix * [bhavanatumma/ta2614748] - signature fix * pl --------- Co-authored-by: bhavanatumma <[email protected]>
anushree1808 · Mar 7, 2024 · a65574a · a65574a
1 parent 1ff1c9d
commit a65574a
Show file tree

Hide file tree

Showing 11 changed files with 290 additions and 87 deletions.
diff --git a/assets/common/components/mlflow_model_local_validation/spec.yaml b/assets/common/components/mlflow_model_local_validation/spec.yaml
@@ -1,7 +1,7 @@
 $schema: https://azuremlschemas.azureedge.net/latest/commandComponent.schema.json
 
 name: mlflow_model_local_validation
-version: 0.0.6
+version: 0.0.7
 type: command
 
 is_deterministic: True
@@ -17,6 +17,7 @@ command: >-
   --model-path ${{inputs.model_path}}
   $[[--test-data-path ${{inputs.test_data_path}}]]
   $[[--column-rename-map "${{inputs.column_rename_map}}"]]
+  $[[--task-name ${{inputs.task_name}}]]
   --output-model-path ${{outputs.mlflow_model_folder}}
 
 inputs:
@@ -35,7 +36,34 @@ inputs:
     optional: true
     description: |
       Provide mapping of dataset column names that should be renamed before inferencing.
-      eg: col1:ren1; col2:ren2; col3:ren3 
+      eg: col1:ren1; col2:ren2; col3:ren3
+
+  task_name:
+    description: A Hugging face task on which model was trained on
+    enum:
+      - fill-mask
+      - token-classification
+      - question-answering
+      - summarization
+      - text-generation
+      - text2text-generation
+      - text-classification
+      - translation
+      - image-classification
+      - image-classification-multilabel
+      - image-object-detection
+      - image-instance-segmentation
+      - image-to-text
+      - text-to-image
+      - text-to-image-inpainting
+      - image-text-to-text
+      - image-to-image
+      - zero-shot-image-classification
+      - mask-generation
+      - video-multi-object-tracking
+      - visual-question-answering
+    optional: true
+    type: string
 
 outputs:
   mlflow_model_folder:

diff --git a/assets/common/src/scripts/validate_model.py b/assets/common/src/scripts/validate_model.py
@@ -6,6 +6,7 @@
 # Script uses minimal required dependency to process data and load mlflow model.
 # Please refrain from adding any additional dependencies that could cause failure while running the script.
 
+import os
 import argparse
 import json
 import mlflow
@@ -20,6 +21,7 @@
 KV_COLON_SEP = ":"
 ITEM_COMMA_SEP = ","
 ITEM_SEMI_COLON_SEP = ";"
+SUMMARIZATION = 'summarization'
 
 
 def get_dict_from_comma_separated_str(dict_str: str, item_sep: str, kv_sep: str) -> Dict:
@@ -65,7 +67,7 @@ def _load_and_prepare_data(test_data_path: Path, mlmodel: Dict, col_rename_map:
     elif ext == ".csv":
         data = pd.read_csv(test_data_path)
     else:
-        raise Exception("Unsupported file type {ext}")
+        raise Exception(f"Unsupported file type {ext}")
 
     # translations
     if col_rename_map:
@@ -91,12 +93,15 @@ def _load_and_prepare_data(test_data_path: Path, mlmodel: Dict, col_rename_map:
     return data
 
 
-def _load_and_infer_model(model_dir, data):
+def _load_and_infer_model(model_dir, data, task_name):
     if data is None:
         print("Data not shared. Could not infer the loaded model")
         return
 
     try:
+        if task_name == SUMMARIZATION:
+            os.environ["MLFLOW_HUGGINGFACE_USE_DEVICE_MAP"] = "False"
+
         model = mlflow.pyfunc.load_model(str(model_dir))
     except Exception as e:
         raise Exception(f"Error in loading mlflow model: {e}")
@@ -114,6 +119,7 @@ def _get_parser():
     parser.add_argument("--model-path", type=Path, required=True, help="Model input path")
     parser.add_argument("--test-data-path", type=Path, required=False, help="Test dataset path")
     parser.add_argument("--column-rename-map", type=str, required=False, help="Column rename map as string")
+    parser.add_argument("--task-name", type=str, required=False, help="Hugging Face task type")
     return parser
 
 
@@ -124,6 +130,7 @@ def _get_parser():
     model_dir: Path = args.model_path
     test_data_path: Path = args.test_data_path
     col_rename_map_str: str = args.column_rename_map
+    task_name: str = args.task_name
 
     mlmodel_file_path = model_dir / MLMODEL_FILE_NAME
     conda_env_file_path = model_dir / CONDA_YAML_FILE_NAME
@@ -144,6 +151,7 @@ def _get_parser():
             mlmodel=mlmodel_dict,
             col_rename_map=col_rename_map,
         ),
+        task_name=task_name,
     )
 
     print("Local validation completed")
diff --git a/assets/training/model_management/components/convert_model_to_mlflow/spec.yaml b/assets/training/model_management/components/convert_model_to_mlflow/spec.yaml
@@ -1,7 +1,7 @@
 $schema: https://azuremlschemas.azureedge.net/latest/commandComponent.schema.json
 
 name: convert_model_to_mlflow
-version: 0.0.21
+version: 0.0.22
 type: command
 
 is_deterministic: True
@@ -19,7 +19,7 @@ command: |
   pip_pkg_str="${pip_pkgs[*]}"
   if [[ -n "$pip_pkg_str" ]]; then echo "Installing $pip_pkg_str"; pip install $pip_pkg_str; echo "pip installation completed. For any installation error please check above logs"; fi;
   echo "Running model conversion ... "
-  python -u run_model_preprocess.py $[[--model-id ${{inputs.model_id}}]] $[[--task-name ${{inputs.task_name}}]] $[[--model-download-metadata ${{inputs.model_download_metadata}}]] $[[--license-file-path ${{inputs.license_file_path}}]] $[[--hf-config-args "${{inputs.hf_config_args}}"]] $[[--hf-tokenizer-args "${{inputs.hf_tokenizer_args}}"]] $[[--hf-model-args "${{inputs.hf_model_args}}"]] $[[--hf-pipeline-args "${{inputs.hf_pipeline_args}}"]] $[[--hf-config-class ${{inputs.hf_config_class}}]] $[[--hf-model-class ${{inputs.hf_model_class}}]] $[[--hf-tokenizer-class ${{inputs.hf_tokenizer_class}}]] $[[--hf-use-experimental-features ${{inputs.hf_use_experimental_features}}]] $[[--extra-pip-requirements "${{inputs.extra_pip_requirements}}"]] $[[--inference-base-image "${{inputs.inference_base_image}}"]] --model-framework ${{inputs.model_framework}} --model-path ${{inputs.model_path}} --mlflow-model-output-dir ${{outputs.mlflow_model_folder}} --model-import-job-path ${{outputs.model_import_job_path}}
+  python -u run_model_preprocess.py $[[--model-id ${{inputs.model_id}}]] $[[--task-name ${{inputs.task_name}}]] $[[--model-download-metadata ${{inputs.model_download_metadata}}]] $[[--license-file-path ${{inputs.license_file_path}}]] $[[--hf-config-args "${{inputs.hf_config_args}}"]] $[[--hf-tokenizer-args "${{inputs.hf_tokenizer_args}}"]] $[[--hf-model-args "${{inputs.hf_model_args}}"]] $[[--hf-pipeline-args "${{inputs.hf_pipeline_args}}"]] $[[--hf-config-class ${{inputs.hf_config_class}}]] $[[--hf-model-class ${{inputs.hf_model_class}}]] $[[--hf-tokenizer-class ${{inputs.hf_tokenizer_class}}]] $[[--hf-use-experimental-features ${{inputs.hf_use_experimental_features}}]] $[[--extra-pip-requirements "${{inputs.extra_pip_requirements}}"]] $[[--inference-base-image "${{inputs.inference_base_image}}"]] --model-framework ${{inputs.model_framework}} --model-path ${{inputs.model_path}} --mlflow-model-output-dir ${{outputs.mlflow_model_folder}} --model-import-job-path ${{outputs.model_import_job_path}} --model-flavor ${{inputs.model_flavor}}
   echo "Completed model conversion ... "
 
 inputs:
@@ -28,6 +28,15 @@ inputs:
     description: Huggingface model id (https://huggingface.co/<model_id>). A required parameter for Huggingface model framework. Can be provided as input here or in model_download_metadata JSON file.
     optional: true
 
+  model_flavor:
+    type: string
+    enum:
+      - HFTransformersV2
+      - OSS
+    default: HFTransformersV2
+    optional: false
+    description: Flavor of MLFlow to which the model is converted to.
+
   model_framework:
     type: string
     enum:
@@ -42,12 +51,12 @@ inputs:
   task_name:
     type: string
     enum:
-      - text-classification
       - fill-mask
       - token-classification
       - question-answering
       - summarization
       - text-generation
+      - text2text-generation
       - text-classification
       - translation
       - image-classification

diff --git a/assets/training/model_management/components/import_model/spec.yaml b/assets/training/model_management/components/import_model/spec.yaml
@@ -4,7 +4,7 @@ type: pipeline
 name: import_model
 display_name: Import model
 description: Import a model into a workspace or a registry
-version: 0.0.25
+version: 0.0.26
 
 # Pipeline inputs
 inputs:
@@ -38,6 +38,15 @@ inputs:
     type: string
     description: A valid model id for the model source selected. For example you can specify `bert-base-uncased` for importing HuggingFace bert base uncased model. Please specify the complete URL if **GIT** or **AzureBlob** is selected in `model_source`
 
+  model_flavor:
+    type: string
+    enum:
+      - HFTransformersV2
+      - OSS
+    default: HFTransformersV2
+    optional: false
+    description: Flavor of MLFlow to which model the model is converted to.
+
   token:
     type: string
     description: If set use it to access the private models or authenticate the user. For example, user can get the token for HF private model by creating account in Huggingface, accept the condition for models that needs to be downloaded and create access token from browser. For more details please visit - https://huggingface.co/docs/hub/security-tokens
@@ -52,18 +61,27 @@ inputs:
   task_name:
     description: A Hugging face task on which model was trained on
     enum:
-    - text-classification
-    - fill-mask
-    - token-classification
-    - question-answering
-    - summarization
-    - text-generation
-    - text-classification
-    - translation
-    - image-classification
-    - text-to-image
-    - zero-shot-image-classification
-    - video-multi-object-tracking
+      - fill-mask
+      - token-classification
+      - question-answering
+      - summarization
+      - text-generation
+      - text2text-generation
+      - text-classification
+      - translation
+      - image-classification
+      - image-classification-multilabel
+      - image-object-detection
+      - image-instance-segmentation
+      - image-to-text
+      - text-to-image
+      - text-to-image-inpainting
+      - image-text-to-text
+      - image-to-image
+      - zero-shot-image-classification
+      - mask-generation
+      - video-multi-object-tracking
+      - visual-question-answering
     optional: true
     type: string
 
@@ -182,7 +200,7 @@ outputs:
 
 jobs:
   validation_trigger_import:
-    component: azureml:validation_trigger_import:0.0.4
+    component: azureml:validation_trigger_import:0.0.5
     compute: ${{parent.inputs.compute}}
     resources:
       instance_type: '${{parent.inputs.instance_type}}'
@@ -194,6 +212,7 @@ jobs:
       model_source: ${{parent.inputs.model_source}}
       model_id: ${{parent.inputs.model_id}}
       model_version: ${{parent.inputs.model_version}}
+      model_flavor: ${{parent.inputs.model_flavor}}
       model_description: ${{parent.inputs.model_description}}
       model_metadata: ${{parent.inputs.model_metadata}}
       registry_name: ${{parent.inputs.registry_name}}
@@ -234,12 +253,13 @@ jobs:
         type: uri_folder
 
   convert_model_to_mlflow:
-    component: azureml:convert_model_to_mlflow:0.0.21
+    component: azureml:convert_model_to_mlflow:0.0.22
     compute: ${{parent.inputs.compute}}
     resources:
       instance_type: '${{parent.inputs.instance_type}}'
     inputs:
       task_name: ${{parent.inputs.task_name}}
+      model_flavor: ${{parent.inputs.model_flavor}}
       license_file_path: ${{parent.inputs.license_file_path}}
       model_download_metadata: ${{parent.jobs.download_model.outputs.model_download_metadata}}
       model_path: ${{parent.jobs.download_model.outputs.model_output}}
@@ -259,14 +279,15 @@ jobs:
         type: uri_file
 
   mlflow_model_local_validation:
-    component: azureml:mlflow_model_local_validation:0.0.6
+    component: azureml:mlflow_model_local_validation:0.0.7
     compute: ${{parent.inputs.compute}}
     resources:
       instance_type: '${{parent.inputs.instance_type}}'
     inputs:
       model_path: ${{parent.jobs.convert_model_to_mlflow.outputs.mlflow_model_folder}}
       test_data_path: ${{parent.inputs.local_validation_test_data}}
       column_rename_map: ${{parent.inputs.local_validation_column_rename_map}}
+      task_name: ${{parent.inputs.task_name}}
     outputs:
       mlflow_model_folder: ${{parent.outputs.mlflow_model_folder}}
 

diff --git a/assets/training/model_management/components/validation_trigger_import/spec.yaml b/assets/training/model_management/components/validation_trigger_import/spec.yaml
@@ -4,7 +4,7 @@ type: command
 name: validation_trigger_import
 display_name: Validation Trigger Import model
 description: Component for enabling validation of import pipeline.
-version: 0.0.4
+version: 0.0.5
 
 # Pipeline inputs
 inputs:
@@ -38,6 +38,15 @@ inputs:
     type: string
     description: A valid model id for the model source selected. For example you can specify `bert-base-uncased` for importing HuggingFace bert base uncased model. Please specify the complete URL if **GIT** or **AzureBlob** is selected in `model_source`
 
+  model_flavor:
+    type: string
+    enum:
+      - HFTransformersV2
+      - OSS
+    default: HFTransformersV2
+    optional: false
+    description: Flavor of MLFlow to which model the model is converted to.
+
   ## Inputs for the MlFLow conversion
   license_file_path:
     type: uri_file

diff --git a/assets/training/model_management/src/azureml/model/mgmt/processors/transformers/config.py b/assets/training/model_management/src/azureml/model/mgmt/processors/transformers/config.py
@@ -69,6 +69,7 @@ class SupportedNLPTasks(_CustomEnum):
     TEXT_GENERATION = "text-generation"
     TEXT_CLASSIFICATION = "text-classification"
     TRANSLATION = "translation"
+    TEXT2TEXT_GENERATION = "text2text-generation"
 
 
 class SupportedASRModelFamily(_CustomEnum):
@@ -88,6 +89,7 @@ class SupportedTasks(_CustomEnum):
     TEXT_GENERATION = "text-generation"
     TEXT_CLASSIFICATION = "text-classification"
     TRANSLATION = "translation"
+    TEXT2TEXT_GENERATION = "text2text-generation"
     # Vision tasks
     IMAGE_CLASSIFICATION = "image-classification"
     # ASR