tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

E2e device cuda (#575)

Browse Source 
* use torch.cuda.current_device() instead of local_rank

* ignore NVML errors for gpu stats

* llama lora packing e2e tests
This commit is contained in:
Wing Lian
2023-09-14 22:49:27 -04:00
committed by GitHub
parent 9218ebecd2
commit 24146733db
4 changed files with 52 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

42
tests/e2e/test_lora_llama.py
View File

@@ -78,3 +78,45 @@ class TestLoraLlama(unittest.TestCase):
dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
def test_lora_packing(self):
cfg = DictDefault(
{
"base_model": "JackFram/llama-68m",
"base_model_config": "JackFram/llama-68m",
"tokenizer_type": "LlamaTokenizer",
"sequence_len": 1024,
"sample_packing": True,
"flash_attention": True,
"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": 8,
"gradient_accumulation_steps": 1,
"output_dir": tempfile.mkdtemp(),
"learning_rate": 0.00001,
"optimizer": "adamw_torch",
"lr_scheduler": "cosine",
}
)
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)