Support generation config in ORTModel (#651)

* support generation config

* add can_generate method in ORTModelForConditionalGeneration

* trigger actions

* fix typog

* rollback

optimum/onnxruntime/modeling_decoder.py

@@ -21,7 +21,7 @@
 
 import numpy as np
 import torch
-from transformers import AutoModelForCausalLM
+from transformers import AutoModelForCausalLM, GenerationConfig
 from transformers.file_utils import add_start_docstrings_to_model_forward
 from transformers.modeling_outputs import CausalLMOutputWithCrossAttentions
 
@@ -339,6 +339,7 @@ def __init__(
         use_io_binding: Optional[bool] = None,
         model_save_dir: Optional[Union[str, Path, TemporaryDirectory]] = None,
         preprocessors: Optional[List] = None,
+        generation_config: Optional[GenerationConfig] = None,
         **kwargs
     ):
         """
@@ -357,6 +358,9 @@ def __init__(
                 The directory under which the model exported to ONNX was saved.
             preprocessors (`Optional[List]`, defaults to `None`):
                 The list of the preprocessors (tokenizer, processor, feature_extractor) to save alongside the ORTModel.
+            generation_config (`Optional[GenerationConfig]`, defaults to `None`):
+                The generation configuration used by default when calling `generate()`.
+                Refer to https://huggingface.co/docs/transformers/main/en/main_classes/text_generation#transformers.GenerationMixin.generate.
         """
         self.shared_attributes_init(
             decoder_session,
@@ -400,6 +404,10 @@ def show_deprecated_argument(arg_name):
             self.decoder_with_past_model_path = Path(decoder_with_past_session._model_path)
             self.decoder_with_past_model_name = self.decoder_with_past_model_path.name
 
+        if generation_config is None:
+            generation_config = GenerationConfig.from_model_config(config)
+        self.generation_config = generation_config
+
     @staticmethod
     def load_model(
         decoder_path: Union[str, Path],
@@ -617,13 +625,28 @@ def _from_pretrained(
         if model_save_dir is None:
             model_save_dir = new_model_save_dir
 
+        generation_config = None
+        try:
+            generation_config = GenerationConfig.from_pretrained(
+                model_id,
+                cache_dir=cache_dir,
+                force_download=force_download,
+                local_files_only=local_files_only,
+                use_auth_token=use_auth_token,
+                revision=revision,
+                subfolder=subfolder,
+            )
+        except OSError:
+            logger.info("Generation config file not found, using a generation config created from the model config.")
+
         return cls(
             model[0],
             config,
             decoder_with_past_session=model[1],
             use_io_binding=use_io_binding,
             model_save_dir=model_save_dir,
             preprocessors=preprocessors,
+            generation_config=generation_config,
         )
 
     @classmethod
@@ -777,3 +800,7 @@ def _reorder_cache(past: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor) ->
             tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past)
             for layer_past in past
         )
+
+    def can_generate(self):
+        """Returns True to validate the check that the model using `GenerationMixin.generate()` can indeed generate."""
+        return True

optimum/onnxruntime/modeling_seq2seq.py

@@ -17,7 +17,6 @@
 """
 
 import logging
-import re
 import shutil
 from abc import ABC, abstractmethod
 from pathlib import Path
@@ -26,7 +25,7 @@
 
 import numpy as np
 import torch
-from transformers import AutoModelForSeq2SeqLM, AutoModelForSpeechSeq2Seq
+from transformers import AutoModelForSeq2SeqLM, AutoModelForSpeechSeq2Seq, GenerationConfig
 from transformers.file_utils import add_start_docstrings_to_model_forward
 from transformers.modeling_outputs import BaseModelOutput, Seq2SeqLMOutput
 
@@ -728,6 +727,7 @@ def __init__(
         use_io_binding: Optional[bool] = None,
         model_save_dir: Optional[Union[str, Path, TemporaryDirectory]] = None,
         preprocessors: Optional[List] = None,
+        generation_config: Optional[GenerationConfig] = None,
         **kwargs,
     ):
         """
@@ -748,6 +748,9 @@ def __init__(
                 The directory under which the model exported to ONNX was saved.
             preprocessors (`Optional[List]`, defaults to `None`):
                 The list of the preprocessors (tokenizer, processor, feature_extractor) to save alongside the ORTModel.
+            generation_config (`Optional[GenerationConfig]`, defaults to `None`):
+                The generation configuration used by default when calling `generate()`.
+                Refer to https://huggingface.co/docs/transformers/main/en/main_classes/text_generation#transformers.GenerationMixin.generate.
         """
         # TODO: remove at version 2.0
         def show_deprecated_argument(arg_name):
@@ -804,6 +807,10 @@ def show_deprecated_argument(arg_name):
             self.decoder_with_past_model_path = Path(decoder_with_past_session._model_path)
             self.decoder_with_past_model_name = self.decoder_with_past_model_path.name
 
+        if generation_config is None:
+            generation_config = GenerationConfig.from_model_config(config)
+        self.generation_config = generation_config
+
     @abstractmethod
     def _initialize_encoder(
         self,
@@ -1066,13 +1073,28 @@ def _from_pretrained(
         if model_save_dir is None:
             model_save_dir = new_model_save_dir
 
+        generation_config = None
+        try:
+            generation_config = GenerationConfig.from_pretrained(
+                model_id,
+                cache_dir=cache_dir,
+                force_download=force_download,
+                local_files_only=local_files_only,
+                use_auth_token=use_auth_token,
+                revision=revision,
+                subfolder=subfolder,
+            )
+        except OSError:
+            logger.info("Generation config file not found, using a generation config created from the model config.")
+
         return cls(
             *inference_sessions[:2],
             config,
             decoder_with_past_session=inference_sessions[2],
             use_io_binding=use_io_binding,
             model_save_dir=model_save_dir,
             preprocessors=preprocessors,
+            generation_config=generation_config,
         )
 
     @classmethod
@@ -1170,6 +1192,12 @@ def to(self, device: Union[torch.device, str, int]):
 
         return self
 
+    def can_generate(self):
+        logger.warning(
+            "ORTModelForConditionalGeneration is an abstract class and is not meant to be used for generation. Please use ORTModelForSeq2SeqLM or ORTModelForSpeechSeq2Seq."
+        )
+        return False
+
 
 class ORTModelForSeq2SeqLM(ORTModelForConditionalGeneration, GenerationMixin):
     """
@@ -1278,6 +1306,10 @@ def _reorder_cache(past, beam_idx) -> Tuple[Tuple[torch.FloatTensor]]:
             )
         return reordered_past
 
+    def can_generate(self):
+        """Returns True to validate the check that the model using `GenerationMixin.generate()` can indeed generate."""
+        return True
+
 
 class ORTModelForSpeechSeq2Seq(ORTModelForConditionalGeneration, GenerationMixin):
     """
@@ -1390,3 +1422,7 @@ def _reorder_cache(past, beam_idx) -> Tuple[Tuple[torch.FloatTensor]]:
                 tuple(past_state.index_select(0, beam_idx) for past_state in layer_past[:2]) + layer_past[2:],
             )
         return reordered_past
+
+    def can_generate(self):
+        """Returns True to validate the check that the model using `GenerationMixin.generate()` can indeed generate."""
+        return True