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

fix optimizer reset for relora sft (#1414)

Browse Source 
* fix optimizer reset

* set states to reset for 8bit optimizers and handle quantile runtime error for embeddings

* fix relora test to check grad_norm

* use flash attn for relora and tweak hyperparams for test

* fix messages field for test dataset
This commit is contained in:
Wing Lian
2024-12-03 08:58:23 -05:00
committed by GitHub
parent 81ef3e45f7
commit 1ef70312ba
4 changed files with 64 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

56
tests/e2e/test_relora_llama.py
View File

@@ -7,13 +7,15 @@ import os
import unittest
from pathlib import Path
from tbparse import SummaryReader
from axolotl.cli import load_datasets
from axolotl.common.cli import TrainerCliArgs
from axolotl.train import train
from axolotl.utils.config import normalize_config
from axolotl.utils.dict import DictDefault
from .utils import with_temp_dir
from .utils import most_recent_subdir, with_temp_dir
LOG = logging.getLogger("axolotl.tests.e2e")
os.environ["WANDB_DISABLED"] = "true"
@@ -29,36 +31,48 @@ class TestReLoraLlama(unittest.TestCase):
# pylint: disable=duplicate-code
cfg = DictDefault(
{
"base_model": "JackFram/llama-68m",
"tokenizer_type": "LlamaTokenizer",
"sequence_len": 1024,
"base_model": "HuggingFaceTB/SmolLM2-135M",
"sequence_len": 2048,
"sample_packing": True,
"pad_to_sequence_len": True,
"flash_attention": True,
"load_in_8bit": True,
"adapter": "lora",
"lora_r": 32,
"lora_r": 8,
"lora_alpha": 16,
"lora_dropout": 0.05,
"lora_target_modules": ["q_proj", "v_proj"],
"relora_steps": 25,
"relora_warmup_steps": 5,
"relora_anneal_steps": 5,
"relora_steps": 100,
"relora_warmup_steps": 20,
"relora_anneal_steps": 10,
"relora_prune_ratio": 0.9,
"relora_cpu_offload": True,
"val_set_size": 0.0,
"special_tokens": {},
"special_tokens": {
"pad_token": "<|endoftext|>",
},
"chat_template": "chatml",
"datasets": [
{
"path": "mhenrichsen/alpaca_2k_test",
"type": "alpaca",
"path": "mlabonne/FineTome-100k",
"type": "chat_template",
"split": "train[:10%]",
"field_messages": "conversations",
"message_field_role": "from",
"message_field_content": "value",
},
],
"warmup_steps": 15,
"warmup_steps": 20,
"num_epochs": 2,
"max_steps": 51, # at least 2x relora_steps
"micro_batch_size": 4,
"max_steps": 205, # at least 2x relora_steps
"micro_batch_size": 2,
"gradient_accumulation_steps": 1,
"output_dir": temp_dir,
"learning_rate": 0.00001,
"optimizer": "adamw_torch",
"optimizer": "adamw_8bit",
"lr_scheduler": "cosine",
"save_safetensors": True,
"use_tensorboard": True,
}
)
normalize_config(cfg)
@@ -66,4 +80,14 @@ class TestReLoraLlama(unittest.TestCase):
dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
assert (Path(temp_dir) / "model.safetensors").exists()
assert (
Path(temp_dir) / "checkpoint-100/adapter/adapter_model.safetensors"
).exists()
assert (Path(temp_dir) / "checkpoint-100/relora/model.safetensors").exists()
tb_log_path = most_recent_subdir(temp_dir + "/runs")
event_file = os.path.join(tb_log_path, sorted(os.listdir(tb_log_path))[0])
reader = SummaryReader(event_file)
df = reader.scalars # pylint: disable=invalid-name
df = df[(df.tag == "train/grad_norm")] # pylint: disable=invalid-name
assert df.value.values[-1] < 0.2, "grad_norm is too high"