FEAT / Trainer: Add adamw 4bit optimizer

src/transformers/trainer.py

@@ -163,6 +163,7 @@
     is_torch_neuroncore_available,
     is_torch_npu_available,
     is_torch_xla_available,
+    is_torchao_available,
     logging,
     strtobool,
 )
@@ -1434,6 +1435,23 @@ def optimizer_hook(param):
                 optimizer_cls = Lomo
 
             optimizer_kwargs.update({"model": model})
+        elif args.optim == OptimizerNames.ADAMW_TORCH_4BIT:
+            if not is_torchao_available() or version.parse(importlib.metadata.version("torchao")) <= version.parse(
+                "0.3.1"
+            ):
+                raise ImportError(
+                    "You need to have `torchao>0.3.1` in order to use torch 4-bit optimizers. "
+                    "Install it with `pip install https://github.com/pytorch/ao.git`"
+                )
+            if not version.parse(importlib.metadata.version("torch")) < version.parse("2.3"):
+                raise ImportError(
+                    "You need to have `torch>2.3` in order to use torch 4-bit optimizers. "
+                    "Install it with `pip install --upgrade torch`"
+                )
+            from torchao.prototype.low_bit_optim import AdamW4bit
+
+            optimizer_cls = AdamW4bit
+            optimizer_kwargs.update(adam_kwargs)
         else:
             raise ValueError(f"Trainer cannot instantiate unsupported optimizer: {args.optim}")
         return optimizer_cls, optimizer_kwargs

src/transformers/training_args.py

@@ -153,6 +153,7 @@ class OptimizerNames(ExplicitEnum):
     ADAMW_APEX_FUSED = "adamw_apex_fused"
     ADAFACTOR = "adafactor"
     ADAMW_ANYPRECISION = "adamw_anyprecision"
+    ADAMW_TORCH_4BIT = "adamw_torch_4bit"
     SGD = "sgd"
     ADAGRAD = "adagrad"
     ADAMW_BNB = "adamw_bnb_8bit"

src/transformers/utils/__init__.py

@@ -206,6 +206,7 @@
     is_torch_tpu_available,
     is_torch_xla_available,
     is_torch_xpu_available,
+    is_torchao_available,
     is_torchaudio_available,
     is_torchdistx_available,
     is_torchdynamo_available,

src/transformers/utils/import_utils.py

@@ -100,6 +100,7 @@ def _is_package_available(pkg_name: str, return_version: bool = False) -> Union[
 _eetq_available = _is_package_available("eetq")
 _galore_torch_available = _is_package_available("galore_torch")
 _lomo_available = _is_package_available("lomo_optim")
+_torchao_available = _is_package_available("torchao")
 # `importlib.metadata.version` doesn't work with `bs4` but `beautifulsoup4`. For `importlib.util.find_spec`, reversed.
 _bs4_available = importlib.util.find_spec("bs4") is not None
 _coloredlogs_available = _is_package_available("coloredlogs")
@@ -337,6 +338,10 @@ def is_torchvision_available():
     return _torchvision_available
 
 
+def is_torchao_available():
+    return _torchao_available
+
+
 def is_galore_torch_available():
     return _galore_torch_available
 

tests/trainer/test_trainer.py

@@ -98,6 +98,7 @@
     is_apex_available,
     is_bitsandbytes_available,
     is_safetensors_available,
+    is_torchao_available,
     is_torchdistx_available,
 )
 from transformers.utils.hp_naming import TrialShortNamer
@@ -4174,6 +4175,16 @@ def hp_name(trial):
                 dict(default_adam_kwargs, **default_anyprecision_kwargs),
             )
         )
+    if is_torchao_available():
+        import torchao
+
+        optim_test_params.append(
+            (
+                TrainingArguments(optim=OptimizerNames.ADAMW_TORCH_4BIT, output_dir="None"),
+                torchao.prototype.low_bit_optim.AdamW4bit,
+                default_adam_kwargs,
+            )
+        )
 
 
 @require_torch