ros/LLaMA-Factory
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix #4699

Browse Source 
slow tokenizer for yi models


Former-commit-id: 4d23a0bcda0c15a903a62eec72d14c584ce020dd
This commit is contained in:
hiyouga
2024-07-14 15:34:22 +08:00
parent 71e4404c0d
commit b92214f78b
2 changed files with 92 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
src/llamafactory/hparams/parser.py
View File

@@ -79,7 +79,11 @@ def _set_transformers_logging(log_level: Optional[int] = logging.INFO) -> None:
transformers.utils.logging.enable_explicit_format()
def _verify_model_args(model_args: "ModelArguments", finetuning_args: "FinetuningArguments") -> None:
def _verify_model_args(
model_args: "ModelArguments",
data_args: "DataArguments",
finetuning_args: "FinetuningArguments",
) -> None:
if model_args.adapter_name_or_path is not None and finetuning_args.finetuning_type != "lora":
raise ValueError("Adapter is only valid for the LoRA method.")
@@ -99,6 +103,10 @@ def _verify_model_args(model_args: "ModelArguments", finetuning_args: "Finetunin
if model_args.adapter_name_or_path is not None and len(model_args.adapter_name_or_path) != 1:
raise ValueError("Quantized model only accepts a single adapter. Merge them first.")
if data_args.template == "yi" and model_args.use_fast_tokenizer:
logger.warning("We should use slow tokenizer for the Yi models.")
model_args.use_fast_tokenizer = False
def _check_extra_dependencies(
model_args: "ModelArguments",
@@ -237,7 +245,7 @@ def get_train_args(args: Optional[Dict[str, Any]] = None) -> _TRAIN_CLS:
logger.warning("`neat_packing` requires `packing` is True. Change it to True.")
data_args.packing = True
_verify_model_args(model_args, finetuning_args)
_verify_model_args(model_args, data_args, finetuning_args)
_check_extra_dependencies(model_args, finetuning_args, training_args)
if (
@@ -361,7 +369,7 @@ def get_infer_args(args: Optional[Dict[str, Any]] = None) -> _INFER_CLS:
if finetuning_args.stage == "rm" and model_args.visual_inputs:
raise ValueError("Reward server does not support MLLM yet. Stay tuned.")
_verify_model_args(model_args, finetuning_args)
_verify_model_args(model_args, data_args, finetuning_args)
_check_extra_dependencies(model_args, finetuning_args)
if model_args.export_dir is not None and model_args.export_device == "cpu":
@@ -384,7 +392,7 @@ def get_eval_args(args: Optional[Dict[str, Any]] = None) -> _EVAL_CLS:
if model_args.infer_backend == "vllm":
raise ValueError("vLLM backend is only available for API, CLI and Web.")
_verify_model_args(model_args, finetuning_args)
_verify_model_args(model_args, data_args, finetuning_args)
_check_extra_dependencies(model_args, finetuning_args)
model_args.device_map = "auto"