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

Feat: Allow usage of native Mistral FA when no sample_packing (#669)

Browse Source 
* Allow usage of native Mistral FA when no sample_packing

* fix: do not apply custom patch when sample_pack off

* chore: lint

* chore: pin transformer to v4.35.0.dev0

* fix: split sample_packing to separate test
This commit is contained in:
NanoCode012
2023-10-04 20:40:47 +09:00
committed by GitHub
parent 90e0d673f7
commit 697c50d408
4 changed files with 125 additions and 95 deletions
Download Patch File Download Diff File Expand all files Collapse all files

92
tests/e2e/test_mistral.py
View File

@@ -71,53 +71,6 @@ class TestMistral(unittest.TestCase):
train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
assert (Path(output_dir) / "adapter_model.bin").exists()
def test_lora_packing(self):
# pylint: disable=duplicate-code
output_dir = tempfile.mkdtemp()
cfg = DictDefault(
{
"base_model": "openaccess-ai-collective/tiny-mistral",
"base_model_config": "openaccess-ai-collective/tiny-mistral",
"flash_attention": True,
"sample_packing": True,
"sequence_len": 1024,
"load_in_8bit": True,
"adapter": "lora",
"lora_r": 32,
"lora_alpha": 64,
"lora_dropout": 0.05,
"lora_target_linear": True,
"val_set_size": 0.1,
"special_tokens": {
"unk_token": "<unk>",
"bos_token": "<s>",
"eos_token": "</s>",
},
"datasets": [
{
"path": "mhenrichsen/alpaca_2k_test",
"type": "alpaca",
},
],
"num_epochs": 2,
"micro_batch_size": 2,
"gradient_accumulation_steps": 1,
"output_dir": output_dir,
"learning_rate": 0.00001,
"optimizer": "adamw_torch",
"lr_scheduler": "cosine",
"max_steps": 20,
"save_steps": 10,
"eval_steps": 10,
}
)
normalize_config(cfg)
cli_args = TrainerCliArgs()
dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
assert (Path(output_dir) / "adapter_model.bin").exists()
def test_ft(self):
# pylint: disable=duplicate-code
output_dir = tempfile.mkdtemp()
@@ -161,48 +114,3 @@ class TestMistral(unittest.TestCase):
train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
assert (Path(output_dir) / "pytorch_model.bin").exists()
def test_ft_packing(self):
# pylint: disable=duplicate-code
output_dir = tempfile.mkdtemp()
cfg = DictDefault(
{
"base_model": "openaccess-ai-collective/tiny-mistral",
"base_model_config": "openaccess-ai-collective/tiny-mistral",
"flash_attention": True,
"sample_packing": True,
"sequence_len": 1024,
"val_set_size": 0.1,
"special_tokens": {
"unk_token": "<unk>",
"bos_token": "<s>",
"eos_token": "</s>",
},
"datasets": [
{
"path": "mhenrichsen/alpaca_2k_test",
"type": "alpaca",
},
],
"num_epochs": 2,
"micro_batch_size": 2,
"gradient_accumulation_steps": 1,
"output_dir": output_dir,
"learning_rate": 0.00001,
"optimizer": "adamw_torch",
"lr_scheduler": "cosine",
"max_steps": 20,
"save_steps": 10,
"eval_steps": 10,
}
)
if is_torch_bf16_gpu_available():
cfg.bf16 = True
else:
cfg.fp16 = True
normalize_config(cfg)
cli_args = TrainerCliArgs()
dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
assert (Path(output_dir) / "pytorch_model.bin").exists()