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

support batch infer in vllm

Browse Source 
Former-commit-id: 3ef5ed3b9a44eed2f7e3ff221dfc343d0a97c0b5
This commit is contained in:
hiyouga
2024-12-04 13:50:00 +00:00
parent 53edd62f8b
commit c1768cfb14
29 changed files with 148 additions and 407 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
src/llamafactory/train/dpo/workflow.py
View File

@@ -19,7 +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.misc import calculate_tps
from ...extras.ploting import plot_loss
from ...hparams import ModelArguments
from ...model import load_model, load_tokenizer
@@ -65,12 +65,6 @@ 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,
@@ -86,13 +80,12 @@ def run_dpo(
# Training
if training_args.do_train:
train_result = trainer.train(resume_from_checkpoint=training_args.resume_from_checkpoint)
trainer.save_model()
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"]
train_result.metrics["effective_tokens_per_sec"] = calculate_tps(
dataset_module["train_dataset"], train_result.metrics, stage="rm"
)
trainer.save_model()
trainer.log_metrics("train", train_result.metrics)
trainer.save_metrics("train", train_result.metrics)
trainer.save_state()