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

Act offload lora fix (#2928) [skip ci]

Browse Source 
* fix activation offloading with lora

* update w e2e test

* add docs for error
This commit is contained in:
Wing Lian
2025-07-24 16:10:04 -04:00
committed by GitHub
parent 1407aac779
commit 0ff2f172ef
5 changed files with 275 additions and 83 deletions
Download Patch File Download Diff File Expand all files Collapse all files

83
tests/e2e/test_activation_offloading.py Normal file
View File

@@ -0,0 +1,83 @@
"""
E2E tests for activation offloading
"""
import pytest
from axolotl.common.datasets import load_datasets
from axolotl.train import train
from axolotl.utils.config import normalize_config, validate_config
from axolotl.utils.dict import DictDefault
from .utils import check_model_output_exists
# pylint: disable=duplicate-code
class TestActivationOffloading:
"""
E2E test cases for activation offloading
"""
@pytest.mark.parametrize(
"adapter",
["lora", "qlora", None],
)
def test_activation_offloading(
self,
temp_dir,
adapter,
):
# pylint: disable=duplicate-code
cfg = DictDefault(
{
"base_model": "HuggingFaceTB/SmolLM2-135M",
"sequence_len": 1024,
"val_set_size": 0.0,
"special_tokens": {
"pad_token": "<|endoftext|>",
"eos_token": "<|im_end|>",
},
"datasets": [
{
"chat_template": "chatml",
"path": "mlabonne/FineTome-100k",
"type": "chat_template",
"split": "train[:10%]",
"field_messages": "conversations",
"message_field_role": "from",
"message_field_content": "value",
},
],
"num_epochs": 1,
"max_steps": 2,
"micro_batch_size": 1,
"gradient_accumulation_steps": 1,
"output_dir": temp_dir,
"learning_rate": 0.00001,
"optimizer": "adamw_torch_8bit",
"lr_scheduler": "cosine",
"flash_attention": True,
"sample_packing": True,
"bf16": "auto",
"save_safetensors": True,
"gradient_checkpointing": True,
"activation_offloading": True,
"save_first_step": False,
"lora_r": 8,
"lora_alpha": 16,
"lora_target_linear": True,
}
)
if adapter == "lora":
cfg["adapter"] = "lora"
if adapter == "qlora":
cfg["adapter"] = "qlora"
cfg["load_in_4bit"] = True
cfg = validate_config(cfg)
normalize_config(cfg)
dataset_meta = load_datasets(cfg=cfg)
train(cfg=cfg, dataset_meta=dataset_meta)
check_model_output_exists(temp_dir, cfg)

56
tests/utils/schemas/validation/test_activation_offloading.py
View File

@@ -21,62 +21,6 @@ class TestActivationOffloading:
assert cfg.gradient_checkpointing is True
assert cfg.activation_offloading is True
def test_gc_converts_offload_w_lora(self, min_base_cfg):
cfg = (
DictDefault(
gradient_checkpointing="offload",
adapter="lora",
)
| min_base_cfg
)
cfg = validate_config(cfg)
assert cfg.gradient_checkpointing is True
assert cfg.activation_offloading == "legacy"
def test_gc_converts_offload_w_qlora(self, min_base_cfg):
cfg = (
DictDefault(
gradient_checkpointing="offload",
adapter="qlora",
load_in_4bit=True,
)
| min_base_cfg
)
cfg = validate_config(cfg)
assert cfg.gradient_checkpointing is True
assert cfg.activation_offloading == "legacy"
def test_ac_impl_changes_w_lora(self, min_base_cfg):
cfg = (
DictDefault(
gradient_checkpointing=True,
activation_offloading=True,
adapter="lora",
)
| min_base_cfg
)
cfg = validate_config(cfg)
assert cfg.gradient_checkpointing is True
assert cfg.activation_offloading == "legacy"
def test_ac_impl_changes_w_qlora(self, min_base_cfg):
cfg = (
DictDefault(
gradient_checkpointing=True,
activation_offloading=True,
adapter="qlora",
load_in_4bit=True,
)
| min_base_cfg
)
cfg = validate_config(cfg)
assert cfg.gradient_checkpointing is True
assert cfg.activation_offloading == "legacy"
def test_ac_offload_impl_noop_wo_adapter(self, min_base_cfg):
cfg = (
DictDefault(