FEAT / Trainer: LOMO optimizer support

docs/source/en/trainer.md

@@ -382,6 +382,56 @@ trainer.train()
 
 Note layerwise optimization is a bit experimental and does not support DDP (Distributed Data Parallel), thus you can run the training script only on a single GPU. Please see [this appropriate section](https://github.com/jiaweizzhao/GaLore?tab=readme-ov-file#train-7b-model-with-a-single-gpu-with-24gb-memory) for more details. Other features such as gradient clipping, DeepSpeed, etc might not be supported out of the box. Please [raise an issue on GitHub](https://github.com/huggingface/transformers/issues) if you encounter such issue.
 
+## LOMO optimizer
+
+The LOMO optimizers have been introduced in [Full Parameter Fine-Tuning for Large Language Models with Limited Resources](https://hf.co/papers/2306.09782) and [AdaLomo: Low-memory Optimization with Adaptive Learning Rate](https://hf.co/papers/2310.10195). 
+They both consist of an efficient full-parameter fine-tuning method. These optimizers fuse the gradient computation and the parameter update in one step to reduce memory usage. Supported optimizers for LOMO are `"lomo"` and `"adalomo"`. First either install LOMO from pypi `pip install lomo-optim` or install it from source with `pip install git+https://github.com/OpenLMLab/LOMO.git`. 
+
+<Tip>
+
+According to the authors, it is recommended to use `AdaLomo` without `grad_norm` to get better performance and higher throughput.
+
+</Tip>
+
+Below is a simple script to demonstrate how to fine-tune [google/gemma-2b](https://huggingface.co/google/gemma-2b) on IMDB dataset in full precision:
+
+```python
+import torch
+import datasets
+from transformers import TrainingArguments, AutoTokenizer, AutoModelForCausalLM
+import trl
+
+train_dataset = datasets.load_dataset('imdb', split='train')
+
+args = TrainingArguments(
+    output_dir="./test-lomo",
+    max_steps=1000,
+    per_device_train_batch_size=4,
+    optim="adalomo",
+    gradient_checkpointing=True,
+    logging_strategy="steps",
+    logging_steps=1,
+    learning_rate=2e-6,
+    save_strategy="no",
+    run_name="lomo-imdb",
+)
+
+model_id = "google/gemma-2b"
+
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+model = AutoModelForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True).to(0)
+
+trainer = trl.SFTTrainer(
+    model=model, 
+    args=args,
+    train_dataset=train_dataset,
+    dataset_text_field='text',
+    max_seq_length=1024,
+)
+
+trainer.train()
+```
+
 ## Accelerate and Trainer
 
 The [`Trainer`] class is powered by [Accelerate](https://hf.co/docs/accelerate), a library for easily training PyTorch models in distributed environments with support for integrations such as [FullyShardedDataParallel (FSDP)](https://pytorch.org/blog/introducing-pytorch-fully-sharded-data-parallel-api/) and [DeepSpeed](https://www.deepspeed.ai/).

src/transformers/testing_utils.py

@@ -82,6 +82,7 @@
     is_keras_nlp_available,
     is_levenshtein_available,
     is_librosa_available,
+    is_lomo_available,
     is_natten_available,
     is_nltk_available,
     is_onnx_available,
@@ -338,6 +339,14 @@ def require_galore_torch(test_case):
     return unittest.skipUnless(is_galore_torch_available(), "test requires GaLore")(test_case)
 
 
+def require_lomo(test_case):
+    """
+    Decorator marking a test that requires LOMO. These tests are skipped when LOMO-optim isn't installed.
+    https://github.com/OpenLMLab/LOMO
+    """
+    return unittest.skipUnless(is_lomo_available(), "test requires LOMO")(test_case)
+
+
 def require_cv2(test_case):
     """
     Decorator marking a test that requires OpenCV.

src/transformers/trainer.py

@@ -153,6 +153,7 @@
     is_galore_torch_available,
     is_in_notebook,
     is_ipex_available,
+    is_lomo_available,
     is_peft_available,
     is_safetensors_available,
     is_sagemaker_dp_enabled,
@@ -1059,12 +1060,18 @@ def create_optimizer(self):
             if "params" in optimizer_kwargs:
                 optimizer_grouped_parameters = optimizer_kwargs.pop("params")
 
+            # Overwrite `model` in case it's created by `get_optimizer_cls_and_kwargs`
+            # e.g. for LOMO optimizer.
+            if "model" in optimizer_kwargs:
+                optimizer_grouped_parameters = optimizer_kwargs.pop("model")
+
             # For layer-wise dummy optimizers we overwrite optimizer_grouped_parameters with `optimizer_dict`
             # to avoid arguments conflicts.
             if "optimizer_dict" in optimizer_kwargs:
                 optimizer_grouped_parameters = optimizer_kwargs.pop("optimizer_dict")
 
             self.optimizer = optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs)
+
             if optimizer_cls.__name__ == "Adam8bit":
                 import bitsandbytes
 
@@ -1382,6 +1389,26 @@ def optimizer_hook(param):
 
             if args.optim == OptimizerNames.GALORE_ADAFACTOR:
                 optimizer_kwargs.update({"scale_parameter": False, "relative_step": False})
+        elif args.optim in [OptimizerNames.LOMO, OptimizerNames.ADALOMO]:
+            if not is_lomo_available():
+                raise ImportError(
+                    "You need to install `lomo_optim` in order to use LOMO optimizers"
+                    " install it with `pip install lomo-optim`"
+                )
+            if not is_accelerate_available("0.30.0"):
+                raise ImportError("You need to have `accelerate>=0.30.0` to be able to use LOMO optimizers")
+
+            if model is None:
+                raise ValueError("You need to pass a `model` in order to correctly initialize a LOMO optimizer.")
+
+            from lomo_optim import AdaLomo, Lomo
+
+            if "ada" in args.optim:
+                optimizer_cls = AdaLomo
+            else:
+                optimizer_cls = Lomo
+
+            optimizer_kwargs.update({"model": model})
         else:
             raise ValueError(f"Trainer cannot instantiate unsupported optimizer: {args.optim}")
         return optimizer_cls, optimizer_kwargs
@@ -2045,6 +2072,9 @@ def _inner_training_loop(
                 model, self.optimizer, self.lr_scheduler = self.accelerator.prepare(
                     self.model, self.optimizer, self.lr_scheduler
                 )
+        elif self.args.optim in [OptimizerNames.LOMO, OptimizerNames.ADALOMO]:
+            # In this case we are in DDP + LOMO, which should be supported
+            self.optimizer = self.accelerator.prepare(self.optimizer)
 
         if self.is_fsdp_enabled:
             self.model = self.model_wrapped = model
@@ -2143,7 +2173,6 @@ def _inner_training_loop(
         self._globalstep_last_logged = self.state.global_step
         model.zero_grad()
         grad_norm: Optional[float] = None
-
         self.control = self.callback_handler.on_train_begin(args, self.state, self.control)
 
         total_batched_samples = 0
@@ -2275,8 +2304,8 @@ def _inner_training_loop(
                         else:
                             grad_norm = _grad_norm
 
-                    # Optimizer step
                     self.optimizer.step()
+
                     optimizer_was_run = not self.accelerator.optimizer_step_was_skipped
                     if optimizer_was_run:
                         # Delay optimizer scheduling until metrics are generated
@@ -3229,7 +3258,6 @@ def training_step(self, model: nn.Module, inputs: Dict[str, Union[torch.Tensor,
         """
         model.train()
         inputs = self._prepare_inputs(inputs)
-
         if is_sagemaker_mp_enabled():
             loss_mb = smp_forward_backward(model, inputs, self.args.gradient_accumulation_steps)
             return loss_mb.reduce_mean().detach().to(self.args.device)
@@ -3240,14 +3268,20 @@ def training_step(self, model: nn.Module, inputs: Dict[str, Union[torch.Tensor,
         del inputs
         torch.cuda.empty_cache()
 
+        kwargs = {}
+
+        # For LOMO optimizers you need to explicitly use the learnign rate
+        if self.args.optim in [OptimizerNames.LOMO, OptimizerNames.ADALOMO]:
+            kwargs["learning_rate"] = self._get_learning_rate()
+
         if self.args.n_gpu > 1:
             loss = loss.mean()  # mean() to average on multi-gpu parallel training
 
         if self.use_apex:
             with amp.scale_loss(loss, self.optimizer) as scaled_loss:
                 scaled_loss.backward()
         else:
-            self.accelerator.backward(loss)
+            self.accelerator.backward(loss, **kwargs)
 
         return loss.detach() / self.args.gradient_accumulation_steps
 

src/transformers/training_args.py

@@ -171,6 +171,8 @@ class OptimizerNames(ExplicitEnum):
     GALORE_ADAMW_LAYERWISE = "galore_adamw_layerwise"
     GALORE_ADAMW_8BIT_LAYERWISE = "galore_adamw_8bit_layerwise"
     GALORE_ADAFACTOR_LAYERWISE = "galore_adafactor_layerwise"
+    LOMO = "lomo"
+    ADALOMO = "adalomo"
 
 
 # Sometimes users will pass in a `str` repr of a dict in the CLI

src/transformers/utils/__init__.py

@@ -141,6 +141,7 @@
     is_keras_nlp_available,
     is_levenshtein_available,
     is_librosa_available,
+    is_lomo_available,
     is_mlx_available,
     is_natten_available,
     is_ninja_available,

src/transformers/utils/import_utils.py

@@ -99,6 +99,7 @@ def _is_package_available(pkg_name: str, return_version: bool = False) -> Union[
 _bitsandbytes_available = _is_package_available("bitsandbytes")
 _eetq_available = _is_package_available("eetq")
 _galore_torch_available = _is_package_available("galore_torch")
+_lomo_available = _is_package_available("lomo_optim")
 # `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")
@@ -328,6 +329,10 @@ def is_galore_torch_available():
     return _galore_torch_available
 
 
+def is_lomo_available():
+    return _lomo_available
+
+
 def is_pyctcdecode_available():
     return _pyctcdecode_available
 

tests/trainer/test_trainer.py

@@ -63,6 +63,7 @@
     require_deepspeed,
     require_galore_torch,
     require_intel_extension_for_pytorch,
+    require_lomo,
     require_optuna,
     require_peft,
     require_ray,
@@ -1229,6 +1230,49 @@ def test_dataloader_without_dataset(self):
             trainer.train()
             trainer.evaluate()
 
+    @require_lomo
+    @require_torch_gpu
+    def test_lomo(self):
+        config = LlamaConfig(vocab_size=100, hidden_size=32, num_hidden_layers=3, num_attention_heads=4)
+        tiny_llama = LlamaForCausalLM(config)
+
+        previous_params = {n: p.clone() for n, p in tiny_llama.named_parameters()}
+
+        x = torch.randint(0, 100, (128,))
+        train_dataset = RepeatDataset(x)
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            # Trainer without inf/nan filter
+            args = TrainingArguments(tmpdir, learning_rate=1e-2, logging_steps=5, optim="lomo", max_steps=20)
+            trainer = Trainer(tiny_llama, args, train_dataset=train_dataset)
+
+            # Check this works
+            _ = trainer.train()
+
+        for name, param in tiny_llama.named_parameters():
+            self.assertFalse(torch.allclose(param, previous_params[name].to(param.device), rtol=1e-12, atol=1e-12))
+
+    @require_lomo
+    @require_torch_gpu
+    def test_adalomo(self):
+        config = LlamaConfig(vocab_size=100, hidden_size=32, num_hidden_layers=3, num_attention_heads=4)
+        tiny_llama = LlamaForCausalLM(config)
+        x = torch.randint(0, 100, (128,))
+        train_dataset = RepeatDataset(x)
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            # Trainer without inf/nan filter
+            args = TrainingArguments(
+                tmpdir,
+                learning_rate=1e-9,
+                logging_steps=5,
+                optim="adalomo",
+            )
+            trainer = Trainer(tiny_llama, args, train_dataset=train_dataset)
+
+            # Check this works
+            _ = trainer.train()
+
     def test_galore_matched_modules(self):
         regex_patterns = [r".*.attn.*", r".*.mlp.*"]