Refined NeuralChat finetuning config

intel_extension_for_transformers/llm/finetuning/finetuning.py

@@ -54,15 +54,15 @@
 import importlib.util
 from transformers.utils.import_utils import is_optimum_available
 from .data_utils import preprocess_dataset, ALPACA_PROMPT_DICT
-from intel_extension_for_transformers.neural_chat.config import FinetuningConfig
+from intel_extension_for_transformers.neural_chat.config import BaseFinetuningConfig
 
 
 def is_optimum_habana_available():
     return is_optimum_available() and importlib.util.find_spec("optimum.habana") != None
 
 
 class Finetuning:
-    def __init__(self, finetuning_config: FinetuningConfig):
+    def __init__(self, finetuning_config: BaseFinetuningConfig):
         self.model_args, self.data_args, self.training_args, self.finetune_args = (
             finetuning_config.model_args, 
             finetuning_config.data_args, 

intel_extension_for_transformers/neural_chat/README.md

@@ -122,6 +122,26 @@ We provide multiple plugins to augment the chatbot on top of LLM inference. Our
 
 Finetune the pretrained large language model (LLM) with the instruction-following dataset for creating the customized chatbot is very easy for NeuralChat.
 
+### Finetuning for Text Generation Task
+
+**command line experience**
+
+```shell
+neuralchat finetune --base_model "meta-llama/Llama-2-7b-chat-hf" --config pipeline/finetuning/config/finetuning.yaml
+```
+
+
+**Python API experience**
+
+```python
+>>> from intel_extension_for_transformers.neural_chat import TextGenerationFinetuningConfig
+>>> from intel_extension_for_transformers.neural_chat import finetune_model
+>>> finetune_cfg = TextGenerationFinetuningConfig()
+>>> finetuned_model = finetune_model(finetune_cfg)
+```
+
+### Finetuning for Summarization Task
+
 **command line experience**
 
 ```shell
@@ -132,12 +152,65 @@ neuralchat finetune --base_model "meta-llama/Llama-2-7b-chat-hf" --config pipeli
 **Python API experience**
 
 ```python
->>> from intel_extension_for_transformers.neural_chat import FinetuningConfig
+>>> from intel_extension_for_transformers.neural_chat import SummarizationFinetuningConfig
 >>> from intel_extension_for_transformers.neural_chat import finetune_model
->>> finetune_cfg = FinetuningConfig()
+>>> finetune_cfg = SummarizationFinetuningConfig()
 >>> finetuned_model = finetune_model(finetune_cfg)
 ```
 
+### Finetuning for Code Generation Task
+
+**command line experience**
+
+```shell
+neuralchat finetune --base_model "meta-llama/Llama-2-7b-chat-hf" --config pipeline/finetuning/config/finetuning.yaml
+```
+
+
+**Python API experience**
+
+```python
+>>> from intel_extension_for_transformers.neural_chat import CodeGenerationFinetuningConfig
+>>> from intel_extension_for_transformers.neural_chat import finetune_model
+>>> finetune_cfg = CodeGenerationFinetuningConfig()
+>>> finetuned_model = finetune_model(finetune_cfg)
+```
+
+### Finetuning for Text-to-Speech(TTS) Task
+
+**command line experience**
+
+```shell
+neuralchat finetune --base_model "meta-llama/Llama-2-7b-chat-hf" --config pipeline/finetuning/config/finetuning.yaml
+```
+
+
+**Python API experience**
+
+```python
+>>> from intel_extension_for_transformers.neural_chat import TTSFinetuningConfig
+>>> from intel_extension_for_transformers.neural_chat import finetune_model
+>>> finetune_cfg = TTSFinetuningConfig()
+>>> finetuned_model = finetune_model(finetune_cfg)
+```
+
+### Inference with Finetuned Model
+
+By default, Parameter-Efficient Fine-Tuning (PEFT) methods are used to accelerate the finetuning process, and to reduce the finetuning cost as well. Below shows the way to load the finetuned model of such and inference with it.
+
+**Python API experience**
+
+```python
+>>> from intel_extension_for_transformers.neural_chat import build_chatbot
+>>> from intel_extension_for_transformers.neural_chat.config import PipelineConfig, LoadingModelConfig
+>>> chatbot = build_chatbot(
+  PipelineConfig(
+    loading_config=LoadingModelConfig(peft_path="/path/to/peft_model")
+  )
+)
+>>> response = chatbot.predict("Tell me about Intel Xeon Scalable Processors.")
+```
+
 ## Quantization
 
 NeuralChat provides three quantization approaches respectively (PostTrainingDynamic, PostTrainingStatic, QuantAwareTraining) based on [Intel® Neural Compressor](https://github.com/intel/neural-compressor).

intel_extension_for_transformers/neural_chat/__init__.py

@@ -17,7 +17,12 @@
 
 from .config import PipelineConfig
 from .config import GenerationConfig
-from .config import FinetuningConfig
+from .config import (
+    TextGenerationFinetuningConfig,
+    SummarizationFinetuningConfig,
+    CodeGenerationFinetuningConfig,
+    TTSFinetuningConfig
+)
 from .config import OptimizationConfig
 from .chatbot import build_chatbot
 from .chatbot import finetune_model

intel_extension_for_transformers/neural_chat/chatbot.py

@@ -20,7 +20,7 @@
 from intel_extension_for_transformers.llm.quantization.optimization import Optimization
 from .config import PipelineConfig
 from .config import OptimizationConfig
-from .config import FinetuningConfig
+from .config import BaseFinetuningConfig
 from .plugins import is_plugin_enabled, get_plugin_instance, get_registered_plugins
 from .config import DeviceOptions
 from .models.base_model import get_model_adapter
@@ -89,14 +89,14 @@ def build_chatbot(config: PipelineConfig=None):
 
     return adapter
 
-def finetune_model(config: FinetuningConfig):
+def finetune_model(config: BaseFinetuningConfig):
     """Finetune the model based on the provided configuration.
 
     Args:
-        config (FinetuningConfig): Configuration for finetuning the model.
+        config (BaseFinetuningConfig): Configuration for finetuning the model.
     """
 
-    assert config is not None, "FinetuningConfig is needed for finetuning."
+    assert config is not None, "BaseFinetuningConfig is needed for finetuning."
     finetuning = Finetuning(config)
     finetuning.finetune()
 

intel_extension_for_transformers/neural_chat/config.py

@@ -312,12 +312,36 @@ class FinetuningArguments:
     )
 
 @dataclass
-class FinetuningConfig:
+class TTSDatasetArguments:
+    audio_paths: Optional[str] = field(default=None, metadata={"help": "The path of audios."})
+    gender: Optional[str] = field(default=None, metadata={"help": "Gender."})
+    language: Optional[str] = field(default="English", metadata={"help": "Language."})
+
+@dataclass
+class TTSModelArguments:
+    step: int = field(default=0, metadata={"help": "TTS model step."})
+    warmup_step: int = field(default=0, metadata={"help": "TTS model warmup step."})
+    learning_rate: float = field(default=5e-5, metadata={"help": "Learning rate."})
+
+@dataclass
+class BaseFinetuningConfig:
     model_args: ModelArguments
     data_args: DataArguments
     training_args: TrainingArguments
     finetune_args: FinetuningArguments
 
+TextGenerationFinetuningConfig = BaseFinetuningConfig
+
+SummarizationFinetuningConfig = BaseFinetuningConfig
+
+CodeGenerationFinetuningConfig = BaseFinetuningConfig
+
+@dataclass
+class TTSFinetuningConfig(BaseFinetuningConfig):
+    training_args: TrainingArguments
+    dataset_args: TTSDatasetArguments
+    model_args: TTSModelArguments
+
 @dataclass
 class GenerationConfig:
     device: str = "cpu"

intel_extension_for_transformers/neural_chat/examples/instruction_tuning/finetune_clm.py

@@ -22,10 +22,10 @@
     ModelArguments,
     DataArguments,
     FinetuningArguments,
-    FinetuningConfig,
+    TextGenerationFinetuningConfig,
 )
 from intel_extension_for_transformers.neural_chat.chatbot import finetune_model
-from intel_extension_for_transformers.neural_chat.pipeline.finetuning.finetuning import is_optimum_habana_available
+from intel_extension_for_transformers.llm.finetuning.finetuning import is_optimum_habana_available
 
 def main():
     # See all possible arguments in src/transformers/training_args.py
@@ -56,7 +56,7 @@ def main():
             finetune_args,
         ) = parser.parse_args_into_dataclasses()
 
-    finetune_cfg = FinetuningConfig(
+    finetune_cfg = TextGenerationFinetuningConfig(
         model_args=model_args,
         data_args=data_args,
         training_args=training_args,

intel_extension_for_transformers/neural_chat/examples/instruction_tuning/finetune_seq2seq.py

@@ -22,7 +22,7 @@
     ModelArguments,
     DataArguments,
     FinetuningArguments,
-    FinetuningConfig,
+    TextGenerationFinetuningConfig,
 )
 from intel_extension_for_transformers.neural_chat.chatbot import finetune_model
 
@@ -48,7 +48,7 @@ def main():
             finetune_args,
         ) = parser.parse_args_into_dataclasses()
 
-    finetune_cfg = FinetuningConfig(
+    finetune_cfg = TextGenerationFinetuningConfig(
         model_args=model_args,
         data_args=data_args,
         training_args=training_args,

intel_extension_for_transformers/neural_chat/tests/finetuning/test_finetuning.py

@@ -23,7 +23,7 @@
     ModelArguments,
     DataArguments,
     FinetuningArguments,
-    FinetuningConfig,
+    TextGenerationFinetuningConfig,
 )
 from intel_extension_for_transformers.neural_chat.chatbot import finetune_model
 
@@ -57,7 +57,7 @@ def test_finetune_clm(self):
             overwrite_output_dir=True
         )
         finetune_args = FinetuningArguments()
-        finetune_cfg = FinetuningConfig(
+        finetune_cfg = TextGenerationFinetuningConfig(
             model_args=model_args,
             data_args=data_args,
             training_args=training_args,
@@ -75,7 +75,7 @@ def test_finetune_seq2seq(self):
             overwrite_output_dir=True
         )
         finetune_args = FinetuningArguments()
-        finetune_cfg = FinetuningConfig(
+        finetune_cfg = TextGenerationFinetuningConfig(
             model_args=model_args,
             data_args=data_args,
             training_args=training_args,