[NeuralChat] support full parameters finetuning (#824)

* support full parameters finetuning.

Author: lkk12014402

intel_extension_for_transformers/llm/finetuning/finetuning.py

@@ -284,6 +284,15 @@ def finetune(self):
  raise NotImplementedError(
  f"Unsupported bits {finetune_args.bits}, only support 4 and 8 now."
  )
+ if finetune_args.full_finetune:
+ raise ValueError(
+ f"qlora and full_finetune can't be True at the same time."
+ )
+ elif finetune_args.full_finetune:
+ if finetune_args.bits not in [16, 32]:
+ raise ValueError(
+ f"full finetune only support 16 and 32 bits."
+ )
 
  config = self.load_model_config(self.model_args)
  if config.architectures[0].endswith("ForCausalLM") \
@@ -482,48 +491,50 @@ def concatenate_data(dataset, max_seq_length):
  )
 
  if training_args.do_train:
- # PEFT settings
- if finetune_args.peft == "lora":
- if finetune_args.lora_all_linear:
- target_modules = self.find_all_linear_names(model)
- else:
- target_modules = finetune_args.lora_target_modules
-
- peft_config = LoraConfig(
- r=finetune_args.lora_rank,
- lora_alpha=finetune_args.lora_alpha,
- lora_dropout=finetune_args.lora_dropout,
- target_modules=target_modules,
- bias="none",
- task_type=TaskType.CAUSAL_LM,
- )
- elif finetune_args.peft == "llama_adapter":
- peft_config = AdaptionPromptConfig(
- adapter_layers=finetune_args.adapter_layers,
- adapter_len=finetune_args.adapter_len,
- task_type="CAUSAL_LM",
- )
- elif finetune_args.peft == "ptun":
- peft_config = PromptEncoderConfig(
- num_virtual_tokens=finetune_args.num_virtual_tokens,
- encoder_hidden_size=finetune_args.ptun_hidden_size,
- task_type="CAUSAL_LM",
- )
- elif finetune_args.peft == "prefix":
- peft_config = PrefixTuningConfig(
- num_virtual_tokens=finetune_args.num_virtual_tokens,
- task_type="CAUSAL_LM",
- )
- elif finetune_args.peft == "prompt":
- peft_config = PromptTuningConfig(
- num_virtual_tokens=finetune_args.num_virtual_tokens,
- task_type="CAUSAL_LM",
- )
+ if not finetune_args.full_finetune:
+ # PEFT settings
+ if finetune_args.peft == "lora":
+ if finetune_args.lora_all_linear:
+ target_modules = self.find_all_linear_names(model)
+ else:
+ target_modules = finetune_args.lora_target_modules
+
+ peft_config = LoraConfig(
+ r=finetune_args.lora_rank,
+ lora_alpha=finetune_args.lora_alpha,
+ lora_dropout=finetune_args.lora_dropout,
+ target_modules=target_modules,
+ bias="none",
+ task_type=TaskType.CAUSAL_LM,
+ )
+ elif finetune_args.peft == "llama_adapter":
+ peft_config = AdaptionPromptConfig(
+ adapter_layers=finetune_args.adapter_layers,
+ adapter_len=finetune_args.adapter_len,
+ task_type="CAUSAL_LM",
+ )
+ elif finetune_args.peft == "ptun":
+ peft_config = PromptEncoderConfig(
+ num_virtual_tokens=finetune_args.num_virtual_tokens,
+ encoder_hidden_size=finetune_args.ptun_hidden_size,
+ task_type="CAUSAL_LM",
+ )
+ elif finetune_args.peft == "prefix":
+ peft_config = PrefixTuningConfig(
+ num_virtual_tokens=finetune_args.num_virtual_tokens,
+ task_type="CAUSAL_LM",
+ )
+ elif finetune_args.peft == "prompt":
+ peft_config = PromptTuningConfig(
+ num_virtual_tokens=finetune_args.num_virtual_tokens,
+ task_type="CAUSAL_LM",
+ )
+
+ model = get_peft_model(model, peft_config)
+ model.print_trainable_parameters()
 
- model = get_peft_model(model, peft_config)
  if model_dtype == torch.bfloat16:
  model = model.to(model_dtype)
- model.print_trainable_parameters()
 
  if finetune_args.device != 'hpu':
  # Initialize our Trainer
@@ -806,24 +817,33 @@ def preprocess_logits_for_metrics(logits, labels):
  else:
  raise ValueError("Must provide model_name_or_path to load a pretrained Seq2SeqLM model.")
 
- # PEFT settings
- if finetune_args.peft == "lora":
- if finetune_args.lora_all_linear:
- target_modules = self.find_all_linear_names(model)
- else:
- target_modules = finetune_args.lora_target_modules
- peft_config = LoraConfig(
- r=finetune_args.lora_rank,
- lora_alpha=finetune_args.lora_alpha,
- lora_dropout=finetune_args.lora_dropout,
- target_modules=target_modules,
- bias="none",
- task_type=TaskType.SEQ_2_SEQ_LM,
+ if finetune_args.qlora:
+ model = prepare_model_for_kbit_training(
+ model, use_gradient_checkpointing=training_args.gradient_checkpointing
  )
 
- # model = prepare_model_for_int8_training(model)
- model = get_peft_model(model, peft_config)
- model.print_trainable_parameters()
+ if not finetune_args.full_finetune:
+ # PEFT settings
+ if finetune_args.peft == "lora":
+ if finetune_args.lora_all_linear:
+ target_modules = self.find_all_linear_names(model)
+ else:
+ target_modules = finetune_args.lora_target_modules
+ peft_config = LoraConfig(
+ r=finetune_args.lora_rank,
+ lora_alpha=finetune_args.lora_alpha,
+ lora_dropout=finetune_args.lora_dropout,
+ target_modules=target_modules,
+ bias="none",
+ task_type=TaskType.SEQ_2_SEQ_LM,
+ )
+
+ # model = prepare_model_for_int8_training(model)
+ model = get_peft_model(model, peft_config)
+ model.print_trainable_parameters()
+
+ if model_dtype == torch.bfloat16:
+ model = model.to(model_dtype)
 
  if training_args.do_eval and not training_args.do_train:
  config = PeftConfig.from_pretrained(model_args.model_name_or_path)
@@ -839,9 +859,26 @@ def preprocess_logits_for_metrics(logits, labels):
  label_pad_token_id=label_pad_token_id,
  pad_to_multiple_of=8)
 
- # Create Trainer instance
- trainer = Seq2SeqTrainer(
+
+ if finetune_args.device != 'hpu':
+ # Create Trainer instance
+ trainer = Seq2SeqTrainer(
+ model=model,
+ args=training_args,
+ data_collator=data_collator,
+ train_dataset=train_dataset if training_args.do_train else None,
+ eval_dataset=eval_dataset if training_args.do_eval else None,
+ compute_metrics=compute_metrics,
+ preprocess_logits_for_metrics=preprocess_logits_for_metrics if training_args.do_eval else None,
+ )
+ else:
+ from optimum.habana import GaudiConfig, GaudiSeq2SeqTrainer # pylint: disable=E0611 E0401
+ gaudi_config = GaudiConfig()
+ gaudi_config.use_fused_adam = True
+ gaudi_config.use_fused_clip_norm = True
+ trainer = GaudiSeq2SeqTrainer(
  model=model,
+ gaudi_config=gaudi_config,
  args=training_args,
  data_collator=data_collator,
  train_dataset=train_dataset if training_args.do_train else None,
@@ -850,6 +887,7 @@ def preprocess_logits_for_metrics(logits, labels):
  preprocess_logits_for_metrics=preprocess_logits_for_metrics if training_args.do_eval else None,
  )
 
+
  # Training
  if training_args.do_train:
  checkpoint = None

intel_extension_for_transformers/neural_chat/config.py

@@ -342,6 +342,10 @@ class FinetuningArguments:
  default=4,
  metadata={"help": "How many bits to use."}
  )
+ full_finetune: bool = field(
+ default=False,
+ metadata={"help": "Finetune the entire model without adapters."}
+ )
 
 @dataclass
 class TTSDatasetArguments:

intel_extension_for_transformers/neural_chat/tests/nightly/finetuning/test_finetuning.py

@@ -49,7 +49,7 @@ def tearDownClass(self):
  shutil.rmtree('./tmp', ignore_errors=True)
  os.remove(test_data_file)
 
- def test_finetune_clm(self):
+ def test_finetune_clm_lora(self):
  model_args = ModelArguments(model_name_or_path="facebook/opt-125m")
  data_args = DataArguments(train_file=test_data_file)
  training_args = TrainingArguments(
@@ -85,7 +85,70 @@ def test_finetune_clm_qlora(self):
  )
  finetune_model(finetune_cfg)
 
- def test_finetune_seq2seq(self):
+ def test_finetune_clm_full_finetuning(self):
+ model_args = ModelArguments(model_name_or_path="facebook/opt-125m")
+ data_args = DataArguments(train_file=test_data_file)
+ training_args = TrainingArguments(
+ output_dir='./tmp',
+ do_train=True,
+ max_steps=3,
+ overwrite_output_dir=True
+ )
+ finetune_args = FinetuningArguments(device='cpu', full_finetune=True,
+ bits=16, do_lm_eval=False)
+ finetune_cfg = TextGenerationFinetuningConfig(
+ model_args=model_args,
+ data_args=data_args,
+ training_args=training_args,
+ finetune_args=finetune_args,
+ )
+ finetune_model(finetune_cfg)
+
+ def test_finetune_clm_value1(self):
+ model_args = ModelArguments(model_name_or_path="facebook/opt-125m")
+ data_args = DataArguments(train_file=test_data_file)
+ training_args = TrainingArguments(
+ output_dir='./tmp',
+ do_train=True,
+ max_steps=3,
+ overwrite_output_dir=True
+ )
+ finetune_args = FinetuningArguments(device='cpu', full_finetune=True, do_lm_eval=False)
+ finetune_cfg = TextGenerationFinetuningConfig(
+ model_args=model_args,
+ data_args=data_args,
+ training_args=training_args,
+ finetune_args=finetune_args,
+ )
+ # finetune_model(finetune_cfg)
+ try:
+ finetune_model(finetune_cfg)
+ except ValueError:
+ print("code pass")
+
+ def test_finetune_clm_value2(self):
+ model_args = ModelArguments(model_name_or_path="facebook/opt-125m")
+ data_args = DataArguments(train_file=test_data_file)
+ training_args = TrainingArguments(
+ output_dir='./tmp',
+ do_train=True,
+ max_steps=3,
+ overwrite_output_dir=True
+ )
+ finetune_args = FinetuningArguments(device='cpu', qlora=True,
+ full_finetune=True, do_lm_eval=False)
+ finetune_cfg = TextGenerationFinetuningConfig(
+ model_args=model_args,
+ data_args=data_args,
+ training_args=training_args,
+ finetune_args=finetune_args,
+ )
+ try:
+ finetune_model(finetune_cfg)
+ except ValueError:
+ print("code pass")
+
+ def test_finetune_seq2seq_lora(self):
  model_args = ModelArguments(model_name_or_path="google/flan-t5-small")
  data_args = DataArguments(train_file=test_data_file)
  training_args = Seq2SeqTrainingArguments(
@@ -103,5 +166,42 @@ def test_finetune_seq2seq(self):
  )
  finetune_model(finetune_cfg)
 
+ def test_finetune_seq2seq_qlora(self):
+ model_args = ModelArguments(model_name_or_path="google/flan-t5-small")
+ data_args = DataArguments(train_file=test_data_file)
+ training_args = Seq2SeqTrainingArguments(
+ output_dir='./tmp',
+ do_train=True,
+ max_steps=3,
+ overwrite_output_dir=True
+ )
+ finetune_args = FinetuningArguments(device='cpu', qlora=True)
+ finetune_cfg = TextGenerationFinetuningConfig(
+ model_args=model_args,
+ data_args=data_args,
+ training_args=training_args,
+ finetune_args=finetune_args,
+ )
+ finetune_model(finetune_cfg)
+
+ def test_finetune_seq2seq_full_finetuning(self):
+ model_args = ModelArguments(model_name_or_path="google/flan-t5-small")
+ data_args = DataArguments(train_file=test_data_file)
+ training_args = Seq2SeqTrainingArguments(
+ output_dir='./tmp',
+ do_train=True,
+ max_steps=3,
+ overwrite_output_dir=True
+ )
+ finetune_args = FinetuningArguments(device='cpu', full_finetune=True,
+ bits=16, do_lm_eval=False)
+ finetune_cfg = TextGenerationFinetuningConfig(
+ model_args=model_args,
+ data_args=data_args,
+ training_args=training_args,
+ finetune_args=finetune_args,
+ )
+ finetune_model(finetune_cfg)
+
 if __name__ == "__main__":
- unittest.main()
+ unittest.main()