tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 3 Packages Projects Releases Wiki Activity

use smaller pretrained models for ci

Browse Source
This commit is contained in:
Wing Lian
2026-04-23 13:51:01 +00:00
parent 1bf65c500e
commit 431888c1de
24 changed files with 614 additions and 205 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
tests/e2e/multigpu/test_tp.py
View File

@@ -8,7 +8,7 @@ from accelerate.test_utils import execute_subprocess_async, get_torch_dist_uniqu
from axolotl.utils.dict import DictDefault
from tests.e2e.utils import check_tensorboard, require_torch_2_7_0
from tests.e2e.utils import check_tensorboard_loss_decreased, require_torch_2_7_0
class TestTensorParallel:
@@ -21,7 +21,7 @@ class TestTensorParallel:
def test_fft_sft(self, temp_dir):
cfg = DictDefault(
{
"base_model": "Qwen/Qwen2.5-0.5B",
"base_model": "axolotl-ai-co/tiny-qwen2-129m",
"sequence_len": 2048,
"val_set_size": 0.01,
"datasets": [
@@ -63,6 +63,6 @@ class TestTensorParallel:
]
)
check_tensorboard(
temp_dir + "/runs", "train/train_loss", 1.0, "Train Loss (%s) is too high"
check_tensorboard_loss_decreased(
temp_dir + "/runs", max_initial=5.0, max_final=4.7
)