423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-12-17 20:30:36 +08:00
Code Issues Packages Projects Releases Wiki Activity

support DPO training (2305.18290)

Browse Source
This commit is contained in:
hiyouga
2023-08-11 03:02:53 +08:00
parent 685dae4eff
commit 3ec4351cfd
34 changed files with 513 additions and 1027300 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
src/llmtuner/tuner/core/loader.py
View File

@@ -34,7 +34,7 @@ check_min_version("4.29.1")
require_version("datasets>=2.12.0", "To fix: pip install datasets>=2.12.0")
require_version("accelerate>=0.21.0", "To fix: pip install accelerate>=0.21.0")
require_version("peft>=0.4.0", "To fix: pip install peft>=0.4.0")
require_version("trl>=0.4.7", "To fix: pip install trl>=0.4.7")
require_version("trl>=0.5.0", "To fix: pip install trl>=0.5.0")
def load_model_and_tokenizer(
@@ -52,9 +52,6 @@ def load_model_and_tokenizer(
logger.warning("Checkpoint is not found at evaluation, load the original model.")
finetuning_args = FinetuningArguments(finetuning_type="none")
assert stage in ["pt", "sft"] or finetuning_args.finetuning_type == "lora", \
"RM and PPO training can only be performed with the LoRA method."
config_kwargs = {
"trust_remote_code": True,
"cache_dir": model_args.cache_dir,
@@ -132,8 +129,6 @@ def load_model_and_tokenizer(
})
if stage == "ppo": # load reward model
assert is_trainable, "PPO stage cannot be performed at evaluation."
assert model_args.reward_model is not None, "Reward model is necessary for PPO training."
logger.info("Load reward model from {}".format(model_args.reward_model))
model.pretrained_model.load_adapter(model_args.reward_model, "reward", is_trainable=False)
assert load_valuehead_params(model, model_args.reward_model), "Reward model is not correctly loaded."