423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-12-19 05:10:35 +08:00
Code Issues Packages Projects Releases Wiki Activity

change to right-padding, update reward score #803

Browse Source
This commit is contained in:
hiyouga
2023-09-08 20:04:31 +08:00
parent 8aaaa132d4
commit 8ea32e4046
15 changed files with 97 additions and 59 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/llmtuner/tuner/core/loader.py
View File

@@ -68,7 +68,7 @@ def load_model_and_tokenizer(
tokenizer = AutoTokenizer.from_pretrained(
model_args.model_name_or_path,
use_fast=model_args.use_fast_tokenizer,
padding_side=model_args.padding_side,
padding_side="right", # training with left-padded tensors in fp16 precision may cause overflow
**config_kwargs
)

6
src/llmtuner/tuner/core/parser.py
View File

@@ -96,6 +96,9 @@ def get_train_args(
# Check arguments (do not check finetuning_args since it may be loaded from checkpoints)
data_args.init_for_training()
if general_args.stage != "pt" and data_args.template is None:
raise ValueError("Please specify which `template` to use.")
if general_args.stage != "sft" and training_args.predict_with_generate:
raise ValueError("`predict_with_generate` cannot be set as True except SFT.")
@@ -221,6 +224,9 @@ def get_infer_args(
]:
model_args, data_args, finetuning_args, generating_args = parse_infer_args(args)
if data_args.template is None:
raise ValueError("Please specify which `template` to use.")
if model_args.quantization_bit is not None and finetuning_args.finetuning_type != "lora":
raise ValueError("Quantization is only compatible with the LoRA method.")

31
src/llmtuner/tuner/core/trainer.py
View File

@@ -44,26 +44,37 @@ class PeftModelMixin:
output_dir = output_dir if output_dir is not None else self.args.output_dir
os.makedirs(output_dir, exist_ok=True)
logger.info(f"Saving model checkpoint to {output_dir}")
model = self.model
model_unwrapped = unwrap_model(model)
model = unwrap_model(self.model)
if isinstance(model, PreTrainedModelWrapper):
# Custom state dict: https://github.com/lvwerra/trl/blob/v0.4.7/trl/models/modeling_value_head.py#L200
if isinstance(model_unwrapped, PreTrainedModelWrapper):
# Custom state dict: https://github.com/lvwerra/trl/blob/v0.7.1/trl/models/modeling_value_head.py#L200
model_state_dict = state_dict or model.state_dict()
v_head_state_dict = {
name.replace("v_head.", ""): model_state_dict[name].cpu().clone().detach()
for name in model_state_dict.keys() if name.startswith("v_head.")
}
torch.save(v_head_state_dict, os.path.join(output_dir, VALUE_HEAD_FILE_NAME))
model = model.pretrained_model
model = model_unwrapped.pretrained_model
model_unwrapped = unwrap_model(model)
state_dict = state_dict or get_state_dict(model)
if isinstance(model, (PeftModel, PreTrainedModel)):
model.config.use_cache = True
model.save_pretrained(output_dir, state_dict=state_dict, safe_serialization=self.args.save_safetensors)
model.config.use_cache = False
if not isinstance(model, (PeftModel, PreTrainedModel)):
if isinstance(model_unwrapped, (PeftModel, PreTrainedModel)):
model_unwrapped.config.use_cache = True
model_unwrapped.save_pretrained(
output_dir, state_dict=state_dict, safe_serialization=self.args.save_safetensors
)
model_unwrapped.config.use_cache = False
else:
logger.info("Trainer.model is not a `PreTrainedModel`, only saving its state dict.")
torch.save(state_dict, os.path.join(output_dir, WEIGHTS_NAME))
else:
torch.save(state_dict, os.path.join(output_dir, WEIGHTS_NAME))
model.config.use_cache = True
model.save_pretrained(
output_dir, state_dict=state_dict, safe_serialization=self.args.save_safetensors
)
model.config.use_cache = False
if self.finetuning_args.finetuning_type == "full" and self.tokenizer is not None:
try: