423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-12-16 20:00:36 +08:00
Code Issues Packages Projects Releases Wiki Activity

Merge pull request #6078 from wtmlon/support-efficient-tokens-calculation

Browse Source 
support effective tokens calculation on sft/dpo

Former-commit-id: bd639a137e
This commit is contained in:
hoshi-hiyouga
2024-11-20 13:43:15 +08:00
committed by GitHub
parent 2ea2bc7c71 e27a0c3d53
commit 302e4e22bf
4 changed files with 37 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
src/llamafactory/train/dpo/workflow.py
View File

@@ -19,6 +19,7 @@ from typing import TYPE_CHECKING, List, Optional
from ...data import PairwiseDataCollatorWithPadding, get_dataset, get_template_and_fix_tokenizer
from ...extras.constants import IGNORE_INDEX
from ...extras.misc import cal_effective_tokens
from ...extras.ploting import plot_loss
from ...hparams import ModelArguments
from ...model import load_model, load_tokenizer
@@ -64,6 +65,12 @@ def run_dpo(
# Update arguments
training_args.remove_unused_columns = False # important for multimodal and pairwise dataset
effective_token_num = 0.0
if finetuning_args.include_effective_tokens_per_second:
for data in dataset_module["train_dataset"]:
effective_token_num += len(data["chosen_input_ids"])
effective_token_num += len(data["rejected_input_ids"])
# Initialize our Trainer
trainer = CustomDPOTrainer(
model=model,
@@ -79,6 +86,12 @@ def run_dpo(
# Training
if training_args.do_train:
train_result = trainer.train(resume_from_checkpoint=training_args.resume_from_checkpoint)
if finetuning_args.include_effective_tokens_per_second:
train_result.metrics["effective_tokens_per_sec"] = cal_effective_tokens(
effective_token_num, train_result.metrics["epoch"], train_result.metrics["train_runtime"]
)
trainer.save_model()
trainer.log_metrics("train", train_result.metrics)
trainer.save_metrics("train", train_result.metrics)