423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-08-22 22:02:51 +08:00
Code Issues Packages Projects Releases Wiki Activity

Update cal_lr.py

Browse Source 
Former-commit-id: 5619e76dc5aa22bd77da80c33a6e52adb4f517d5
This commit is contained in:
hiyouga 2023-11-14 21:13:01 +08:00
parent b45fb2b3da
commit 426c2fc340
1 changed files with 3 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
tests/cal_lr.py
View File

@ -15,8 +15,8 @@ from llmtuner.extras.constants import IGNORE_INDEX
from llmtuner.tuner.core import get_train_args, load_model_and_tokenizer
BASE_LR = 3e-4
BASE_BS = 4_000_000
BASE_LR = 3e-4 # 1.5e-4 for 30B-70B models
BASE_BS = 4_000_000 # from llama paper
def calculate_lr(
@ -54,7 +54,7 @@ def calculate_lr(
batch_max_len = cutoff_len * batch_size # max tokens in a batch
valid_ratio = valid_tokens / total_tokens
batch_valid_len = batch_max_len * valid_ratio
lr = BASE_LR * math.sqrt(batch_valid_len / BASE_BS)
lr = BASE_LR * math.sqrt(batch_valid_len / BASE_BS) # lr ~ sqrt(batch_size)
lr = lr / 6.0 if is_mistral else lr
print("Optimal learning rate is {:.2e} for valid ratio% {:.2f} and effective batch size {:.2f}".format(
lr, valid_ratio * 100, batch_valid_len