00cda8cc70
* data loading refactor (wip) * updates * progress * pytest * pytest fix * lint * zero_first -> filelock, more simplifications * small simplification * import change * nit * lint * simplify dedup * couldnt resist * review comments WIP * continued wip * minor changes * fix; remove contrived test * further refactor * set default seed in pydantic config * lint * continued simplication * lint * renaming and nits * filelock tests * fix * fix * lint * remove nullable arg * remove unnecessary code * moving dataset save fn to shared module * remove debug print * matching var naming * fn name change * coderabbit comments * naming nit * fix test
143 lines
4.3 KiB
Python
143 lines
4.3 KiB
Python
"""
|
|
Simple end-to-end test for Cut Cross Entropy integration
|
|
"""
|
|
|
|
import pytest
|
|
|
|
from axolotl.common.datasets import load_datasets
|
|
from axolotl.train import train
|
|
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 ..utils import check_model_output_exists
|
|
|
|
# pylint: disable=duplicate-code
|
|
|
|
|
|
@pytest.fixture()
|
|
def min_cfg(temp_dir):
|
|
return {
|
|
"base_model": "HuggingFaceTB/SmolLM2-135M",
|
|
"plugins": [
|
|
"axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin",
|
|
],
|
|
"cut_cross_entropy": True,
|
|
"sequence_len": 1024,
|
|
"val_set_size": 0.02,
|
|
"special_tokens": {
|
|
"pad_token": "<|endoftext|>",
|
|
},
|
|
"datasets": [
|
|
{
|
|
"path": "mhenrichsen/alpaca_2k_test",
|
|
"type": "alpaca",
|
|
},
|
|
],
|
|
"num_epochs": 1,
|
|
"micro_batch_size": 8,
|
|
"gradient_accumulation_steps": 1,
|
|
"learning_rate": 0.00001,
|
|
"optimizer": "adamw_torch_fused",
|
|
"output_dir": temp_dir,
|
|
"lr_scheduler": "cosine",
|
|
"save_safetensors": True,
|
|
"max_steps": 10,
|
|
"bf16": "auto",
|
|
}
|
|
|
|
|
|
class TestCutCrossEntropyIntegration:
|
|
"""
|
|
e2e tests for cut_cross_entropy integration with Axolotl
|
|
"""
|
|
|
|
# pylint: disable=redefined-outer-name
|
|
def test_llama_w_cce(self, min_cfg, temp_dir):
|
|
cfg = DictDefault(min_cfg)
|
|
cfg = validate_config(cfg)
|
|
prepare_plugins(cfg)
|
|
normalize_config(cfg)
|
|
dataset_meta = load_datasets(cfg=cfg)
|
|
|
|
major, minor, _ = get_pytorch_version()
|
|
if (major, minor) < (2, 4):
|
|
with pytest.raises(ImportError):
|
|
train(cfg=cfg, dataset_meta=dataset_meta)
|
|
else:
|
|
train(cfg=cfg, dataset_meta=dataset_meta)
|
|
check_model_output_exists(temp_dir, cfg)
|
|
|
|
# pylint: disable=redefined-outer-name
|
|
def test_qwen2_w_cce(self, temp_dir):
|
|
cfg = DictDefault(
|
|
{
|
|
"base_model": "Qwen/Qwen2.5-0.5B",
|
|
"plugins": [
|
|
"axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin",
|
|
],
|
|
"cut_cross_entropy": True,
|
|
"sequence_len": 1024,
|
|
"val_set_size": 0.02,
|
|
"special_tokens": {
|
|
"pad_token": "<|endoftext|>",
|
|
},
|
|
"datasets": [
|
|
{
|
|
"path": "mhenrichsen/alpaca_2k_test",
|
|
"type": "alpaca",
|
|
},
|
|
],
|
|
"num_epochs": 1,
|
|
"micro_batch_size": 4,
|
|
"gradient_accumulation_steps": 1,
|
|
"learning_rate": 0.00001,
|
|
"optimizer": "adamw_torch_fused",
|
|
"output_dir": temp_dir,
|
|
"lr_scheduler": "cosine",
|
|
"save_safetensors": True,
|
|
"max_steps": 10,
|
|
"bf16": "auto",
|
|
}
|
|
)
|
|
cfg = validate_config(cfg)
|
|
prepare_plugins(cfg)
|
|
normalize_config(cfg)
|
|
dataset_meta = load_datasets(cfg=cfg)
|
|
|
|
major, minor, _ = get_pytorch_version()
|
|
if (major, minor) < (2, 4):
|
|
with pytest.raises(ImportError):
|
|
train(cfg=cfg, dataset_meta=dataset_meta)
|
|
else:
|
|
train(cfg=cfg, dataset_meta=dataset_meta)
|
|
check_model_output_exists(temp_dir, cfg)
|
|
|
|
@pytest.mark.parametrize(
|
|
"attention_type",
|
|
[
|
|
"flash_attention",
|
|
"sdp_attention",
|
|
# "xformers_attention",
|
|
],
|
|
)
|
|
def test_llama_w_cce_and_attention(self, min_cfg, temp_dir, attention_type):
|
|
cfg = DictDefault(
|
|
min_cfg
|
|
| {
|
|
attention_type: True,
|
|
}
|
|
)
|
|
cfg = validate_config(cfg)
|
|
prepare_plugins(cfg)
|
|
normalize_config(cfg)
|
|
dataset_meta = load_datasets(cfg=cfg)
|
|
|
|
major, minor, _ = get_pytorch_version()
|
|
if (major, minor) < (2, 4):
|
|
with pytest.raises(ImportError):
|
|
train(cfg=cfg, dataset_meta=dataset_meta)
|
|
else:
|
|
train(cfg=cfg, dataset_meta=dataset_meta)
|
|
check_model_output_exists(temp_dir, cfg)
|