tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 2 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

34
tests/e2e/integrations/test_cut_cross_entropy.py
View File

@@ -10,7 +10,10 @@ from axolotl.utils import get_pytorch_version
from axolotl.utils.config import normalize_config, prepare_plugins, validate_config
from axolotl.utils.dict import DictDefault
from tests.e2e.utils import check_model_output_exists
from tests.e2e.utils import (
check_model_output_exists,
check_tensorboard_loss_decreased,
)
@pytest.fixture()
@@ -42,6 +45,7 @@ def min_cfg(temp_dir):
"max_steps": 10,
"bf16": "auto",
"save_first_step": False,
"use_tensorboard": True,
}
@@ -64,11 +68,18 @@ class TestCutCrossEntropyIntegration:
else:
train(cfg=cfg, dataset_meta=dataset_meta)
check_model_output_exists(temp_dir, cfg)
check_tensorboard_loss_decreased(
temp_dir + "/runs",
initial_window=5,
final_window=5,
max_initial=5.0,
max_final=4.7,
)
def test_qwen2_w_cce(self, temp_dir):
cfg = DictDefault(
{
"base_model": "Qwen/Qwen2.5-0.5B",
"base_model": "axolotl-ai-co/tiny-qwen2-129m",
"plugins": [
"axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin",
],
@@ -87,13 +98,14 @@ class TestCutCrossEntropyIntegration:
"num_epochs": 1,
"micro_batch_size": 4,
"gradient_accumulation_steps": 1,
"learning_rate": 0.00001,
"learning_rate": 2e-4,
"optimizer": "adamw_torch_fused",
"output_dir": temp_dir,
"lr_scheduler": "cosine",
"max_steps": 10,
"max_steps": 50,
"bf16": "auto",
"save_first_step": False,
"use_tensorboard": True,
}
)
cfg = validate_config(cfg)
@@ -108,6 +120,13 @@ class TestCutCrossEntropyIntegration:
else:
train(cfg=cfg, dataset_meta=dataset_meta)
check_model_output_exists(temp_dir, cfg)
check_tensorboard_loss_decreased(
temp_dir + "/runs",
initial_window=5,
final_window=5,
max_initial=5.0,
max_final=4.7,
)
@pytest.mark.parametrize(
"attention_type",
@@ -136,3 +155,10 @@ class TestCutCrossEntropyIntegration:
else:
train(cfg=cfg, dataset_meta=dataset_meta)
check_model_output_exists(temp_dir, cfg)
check_tensorboard_loss_decreased(
temp_dir + "/runs",
initial_window=5,
final_window=5,
max_initial=5.0,
max_final=4.7,
)