1d91d905c9
* remove deprecated wandb env var * remove os.environ wandb setting; unused loggers * remove os.environ wandb setting; unused loggers
90 lines
2.7 KiB
Python
90 lines
2.7 KiB
Python
"""
|
|
E2E tests for multigpu lora tinyllama
|
|
"""
|
|
|
|
from pathlib import Path
|
|
|
|
import pytest
|
|
import yaml
|
|
from accelerate.test_utils import execute_subprocess_async
|
|
from huggingface_hub import snapshot_download
|
|
from transformers.testing_utils import get_torch_dist_unique_port
|
|
from transformers.utils import is_torch_bf16_gpu_available
|
|
|
|
from axolotl.utils.dict import DictDefault
|
|
|
|
from tests.e2e.utils import check_tensorboard, require_torch_2_6_0
|
|
|
|
AXOLOTL_ROOT = Path(__file__).parent.parent.parent.parent
|
|
|
|
|
|
@pytest.fixture(scope="session", autouse=True)
|
|
def download_model():
|
|
# download the model
|
|
snapshot_download("HuggingFaceTB/SmolLM2-135M")
|
|
|
|
|
|
class TestPackedFlex:
|
|
"""
|
|
Test case for Packed training of llama models
|
|
"""
|
|
|
|
@require_torch_2_6_0
|
|
def test_loss_llama(self, temp_dir):
|
|
# pylint: disable=duplicate-code
|
|
cfg = DictDefault(
|
|
{
|
|
"base_model": "HuggingFaceTB/SmolLM2-135M",
|
|
"sequence_len": 1024,
|
|
"sample_packing": True,
|
|
"flex_attention": True,
|
|
"val_set_size": 0.0,
|
|
"special_tokens": {
|
|
"pad_token": "<|endoftext|>",
|
|
},
|
|
"datasets": [
|
|
{
|
|
"path": "tatsu-lab/alpaca",
|
|
"type": "alpaca",
|
|
"split": "train[:10%]",
|
|
},
|
|
],
|
|
"num_epochs": 1,
|
|
"micro_batch_size": 2,
|
|
"gradient_accumulation_steps": 2,
|
|
"gradient_checkpointing": True,
|
|
"output_dir": temp_dir,
|
|
"learning_rate": 0.00001,
|
|
"optimizer": "adamw_torch_fused",
|
|
"lr_scheduler": "cosine",
|
|
"max_steps": 2,
|
|
"use_tensorboard": True,
|
|
"save_strategy": "no",
|
|
}
|
|
)
|
|
if is_torch_bf16_gpu_available():
|
|
cfg.bf16 = True
|
|
else:
|
|
cfg.fp16 = True
|
|
|
|
# write cfg to yaml file
|
|
Path(temp_dir).mkdir(parents=True, exist_ok=True)
|
|
with open(Path(temp_dir) / "config.yaml", "w", encoding="utf-8") as fout:
|
|
fout.write(yaml.dump(cfg.to_dict(), Dumper=yaml.Dumper))
|
|
|
|
execute_subprocess_async(
|
|
[
|
|
"axolotl",
|
|
"train",
|
|
str(Path(temp_dir) / "config.yaml"),
|
|
"--num-processes",
|
|
"2",
|
|
"--main-process-port",
|
|
f"{get_torch_dist_unique_port()}",
|
|
]
|
|
)
|
|
|
|
check_tensorboard(
|
|
temp_dir + "/runs", "train/train_loss", 2.0, "Train Loss is too high"
|
|
)
|