ADOPT optimizer integration (#2032) [skip ci]

* adopt integration

* stuff

* doc and test for ADOPT

* rearrangement

* fixed formatting

* hacking pre-commit

* chore: lint

* update module doc for adopt optimizer

* remove un-necessary example yaml for adopt optimizer

* skip test adopt if torch<2.5.1

* formatting

* use version.parse

* specifies required torch version for adopt_adamw

---------

Co-authored-by: sunny <sunnyliu19981005@gmail.com>
Co-authored-by: Wing Lian <wing@axolotl.ai>

tests/e2e/test_optimizers.py

@@ -13,7 +13,7 @@ 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 require_torch_2_5_1, with_temp_dir
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
@@ -65,3 +65,46 @@ class TestCustomOptimizers(unittest.TestCase):
 
         train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
         assert (Path(temp_dir) / "adapter_model.bin").exists()
+
+    @with_temp_dir
+    @require_torch_2_5_1
+    def test_adopt_adamw(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "JackFram/llama-68m",
+                "tokenizer_type": "LlamaTokenizer",
+                "sequence_len": 1024,
+                "load_in_8bit": True,
+                "adapter": "lora",
+                "lora_r": 8,
+                "lora_alpha": 16,
+                "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": 1,
+                "micro_batch_size": 8,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adopt_adamw",
+                "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)
+        assert (Path(temp_dir) / "adapter_model.bin").exists()

tests/e2e/utils.py

@@ -6,11 +6,13 @@ import shutil
 import tempfile
 import unittest
 from functools import wraps
-from importlib.metadata import version
 from pathlib import Path
 
 import torch
 
+# from importlib.metadata import version
+from packaging import version
+
 
 def with_temp_dir(test_func):
     @wraps(test_func)
@@ -43,12 +45,24 @@ def require_torch_2_3_1(test_case):
     """
 
     def is_min_2_3_1():
-        torch_version = version("torch")
-        return torch_version >= "2.3.1"
+        torch_version = version.parse(torch.__version__)
+        return torch_version >= version.parse("2.3.1")
 
     return unittest.skipUnless(is_min_2_3_1(), "test torch 2.3.1")(test_case)
 
 
+def require_torch_2_5_1(test_case):
+    """
+    Decorator marking a test that requires torch >= 2.3.1
+    """
+
+    def is_min_2_5_1():
+        torch_version = version.parse(torch.__version__)
+        return torch_version >= version.parse("2.5.1")
+
+    return unittest.skipUnless(is_min_2_5_1(), "test torch 2.5.1")(test_case)
+
+
 def is_hopper():
     compute_capability = torch.cuda.get_device_capability()
     return compute_capability == (9, 0)