FEAT / Optim: Add GaLore optimizer

docs/source/en/trainer.md

@@ -252,6 +252,98 @@ trainer = Trainer(..., args=training_args)
 
 NEFTune is disabled after training to restore the original embedding layer to avoid any unexpected behavior.
 
+## GaLore
+
+Gradient Low-Rank Projection (GaLore) is a memory-efficient low-rank training strategy that allows full-parameter learning but is more memory-efficient than common low-rank adaptation methods, such as LoRA.
+
+First make sure to install GaLore official repository:
+
+```bash
+pip install git+https://github.com/jiaweizzhao/GaLore
+```
+
+Then simply add one of `["galore_adamw", "galore_adafactor", "galore_adamw_8bit"]` in `optim` together with `optim_target_modules`, which should be a list of strings, corresponding to the target module names you want to adapt. Below is an end-to-end example script (make sure to `pip install trl datasets`):
+
+```python
+import torch
+import datasets
+import trl
+
+from transformers import TrainingArguments, AutoConfig, AutoTokenizer, AutoModelForCausalLM
+
+train_dataset = datasets.load_dataset('imdb', split='train')
+
+args = TrainingArguments(
+    output_dir="./test-galore",
+    max_steps=100,
+    per_device_train_batch_size=2,
+    optim="galore_adamw",
+    optim_target_modules=["attn", "mlp"]
+)
+
+model_id = "google/gemma-2b"
+
+config = AutoConfig.from_pretrained(model_id)
+
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+model = AutoModelForCausalLM.from_config(config).to(0)
+
+trainer = trl.SFTTrainer(
+    model=model, 
+    args=args,
+    train_dataset=train_dataset,
+    dataset_text_field='text',
+    max_seq_length=512,
+)
+
+trainer.train()
+```
+
+You can read more about the method in the [original repository](https://github.com/jiaweizzhao/GaLore) or the [paper](https://arxiv.org/abs/2403.03507).
+
+Currently you can only train Linear layers that are considered as GaLore layers and will use low-rank decomposition to be trained while remaining layers will be optimized in the conventional manner.
+
+Note it will take a bit of time before starting the training (~3 minutes for a 2B model on a NVIDIA A100), but training should go smoothly afterwards.
+
+You can also perform layer-wise optimization by post-pending the optimizer name with `layerwise` like below:
+
+```python
+import torch
+import datasets
+import trl
+
+from transformers import TrainingArguments, AutoConfig, AutoTokenizer, AutoModelForCausalLM
+
+train_dataset = datasets.load_dataset('imdb', split='train')
+
+args = TrainingArguments(
+    output_dir="./test-galore",
+    max_steps=100,
+    per_device_train_batch_size=2,
+    optim="galore_adamw_layerwise",
+    optim_target_modules=["attn", "mlp"]
+)
+
+model_id = "google/gemma-2b"
+
+config = AutoConfig.from_pretrained(model_id)
+
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+model = AutoModelForCausalLM.from_config(config).to(0)
+
+trainer = trl.SFTTrainer(
+    model=model, 
+    args=args,
+    train_dataset=train_dataset,
+    dataset_text_field='text',
+    max_seq_length=512,
+)
+
+trainer.train()
+```
+
+Note layerwise optimization 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.
+
 ## 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

@@ -70,6 +70,7 @@
     is_fsdp_available,
     is_ftfy_available,
     is_g2p_en_available,
+    is_galore_torch_available,
     is_ipex_available,
     is_jieba_available,
     is_jinja_available,
@@ -324,6 +325,14 @@ def require_bs4(test_case):
     return unittest.skipUnless(is_bs4_available(), "test requires BeautifulSoup4")(test_case)
 
 
+def require_galore_torch(test_case):
+    """
+    Decorator marking a test that requires Galore. These tests are skipped when Galore isn't installed.
+    https://github.com/jiaweizzhao/GaLore
+    """
+    return unittest.skipUnless(is_galore_torch_available(), "test requires Galore")(test_case)
+
+
 def require_cv2(test_case):
     """
     Decorator marking a test that requires OpenCV.

src/transformers/trainer.py

@@ -83,6 +83,8 @@
     DistributedTensorGatherer,
     IterableDatasetShard,
     LabelSmoother,
+    LayerWiseDummyOptimizer,
+    LayerWiseDummyScheduler,
     LengthGroupedSampler,
     SequentialDistributedSampler,
     distributed_broadcast_scalars,
@@ -111,6 +113,7 @@
     RemoveColumnsCollator,
     TrainerMemoryTracker,
     TrainOutput,
+    check_target_module_exists,
     default_compute_objective,
     denumpify_detensorize,
     enable_full_determinism,
@@ -140,6 +143,7 @@
     is_apex_available,
     is_bitsandbytes_available,
     is_datasets_available,
+    is_galore_torch_available,
     is_in_notebook,
     is_ipex_available,
     is_peft_available,
@@ -1009,7 +1013,17 @@ def create_optimizer(self):
                 },
             ]
 
-            optimizer_cls, optimizer_kwargs = Trainer.get_optimizer_cls_and_kwargs(self.args)
+            optimizer_cls, optimizer_kwargs = Trainer.get_optimizer_cls_and_kwargs(self.args, opt_model)
+
+            # Overwrite `params` in case it's created by `get_optimizer_cls_and_kwargs`
+            # e.g. for GaLoRe optimizer.
+            if "params" in optimizer_kwargs:
+                optimizer_grouped_parameters = optimizer_kwargs.pop("params")
+
+            # 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":
@@ -1032,7 +1046,9 @@ def create_optimizer(self):
         return self.optimizer
 
     @staticmethod
-    def get_optimizer_cls_and_kwargs(args: TrainingArguments) -> Tuple[Any, Any]:
+    def get_optimizer_cls_and_kwargs(
+        args: TrainingArguments, model: Optional[PreTrainedModel] = None
+    ) -> Tuple[Any, Any]:
         """
         Returns the optimizer class and optimizer parameters based on the training arguments.
 
@@ -1170,6 +1186,122 @@ def get_optimizer_cls_and_kwargs(args: TrainingArguments) -> Tuple[Any, Any]:
             optimizer_cls = torch.optim.Adagrad
         elif args.optim == OptimizerNames.RMSPROP:
             optimizer_cls = torch.optim.RMSprop
+        elif args.optim in [
+            OptimizerNames.GALORE_ADAMW,
+            OptimizerNames.GALORE_ADAMW_8BIT,
+            OptimizerNames.GALORE_ADAFACTOR,
+            OptimizerNames.GALORE_ADAMW_LAYERWISE,
+            OptimizerNames.GALORE_ADAMW_8BIT_LAYERWISE,
+            OptimizerNames.GALORE_ADAFACTOR_LAYERWISE,
+        ]:
+            if not is_galore_torch_available():
+                raise ImportError(
+                    "You need to install `galore_torch` in order to use GaLore optimizers"
+                    " install it with `pip install git+https://github.com/jiaweizzhao/GaLore`"
+                )
+            from galore_torch import GaLoreAdafactor, GaLoreAdamW, GaLoreAdamW8bit
+
+            is_layerwise = args.optim.lower().endswith("layerwise")
+            if is_layerwise and args.parallel_mode == ParallelMode.DISTRIBUTED:
+                raise NotImplementedError("GaLore does not support DDP at this time")
+
+            optimizer_mapping = {
+                OptimizerNames.GALORE_ADAMW: GaLoreAdamW,
+                OptimizerNames.GALORE_ADAMW_8BIT: GaLoreAdamW8bit,
+                OptimizerNames.GALORE_ADAFACTOR: GaLoreAdafactor,
+                OptimizerNames.GALORE_ADAMW_LAYERWISE: GaLoreAdamW,
+                OptimizerNames.GALORE_ADAMW_8BIT_LAYERWISE: GaLoreAdamW8bit,
+                OptimizerNames.GALORE_ADAFACTOR_LAYERWISE: GaLoreAdafactor,
+            }
+
+            optimizer_cls = optimizer_mapping[args.optim]
+
+            if args.optim_target_modules is None:
+                raise ValueError(
+                    "You need to define a `optim_target_modules` in order to properly use GaLoRe optimizers"
+                )
+
+            if not isinstance(args.optim_target_modules, (list, str)):
+                raise ValueError(
+                    f"`optim_target_modules` has to be a list of strings, a string corresponding to a regex, or a specific module or 'all-linear', you passed {args.optim_target_modules}"
+                )
+
+            if model is None:
+                raise ValueError("You need to pass a model in order to correctly initialize a GaLore optimizer.")
+
+            logger.warning(
+                "Activated GaLoRE fine-tuning, depending on your model size and hardware, the training might take a while before starting. Please be patient !"
+            )
+
+            all_linear = (
+                isinstance(args.optim_target_modules, str)
+                and args.optim_target_modules.replace("_", "-") == "all-linear"
+            )
+
+            galore_params = []
+            galore_params_names = []
+            for module_name, module in model.named_modules():
+                if not isinstance(module, nn.Linear):
+                    continue
+
+                if not check_target_module_exists(args.optim_target_modules, module_name) and not all_linear:
+                    continue
+
+                galore_params.append(module.weight)
+                galore_params_names.append(module_name + ".weight")
+
+            if len(galore_params) == 0:
+                raise ValueError(
+                    f"None of the target modules were found! ({args.optim_target_modules}). Please make sure to pass a valid `target_modules`."
+                )
+
+            non_galore_params = [p for n, p in model.named_parameters() if n not in galore_params_names]
+
+            galore_optim_kwargs = {
+                "rank": optim_args.pop("rank", 128),
+                "update_proj_gap": optim_args.pop("update_proj_gap", 200),
+                "scale": optim_args.pop("scale", 0.25),
+                "proj_type": optim_args.pop("proj_type", "std"),
+            }
+
+            # The default args are from the official repository: https://github.com/jiaweizzhao/GaLore
+            param_groups = [
+                {"params": non_galore_params},
+                {"params": galore_params, **galore_optim_kwargs},
+            ]
+
+            if is_layerwise:
+                # For layer-wise optimizers, the optimization step is done through post accumulation
+                # gradient hooks. The trick is to first attach these hooks to the model parameters then
+                # create a dummy optimizer that will perform no-ops in the Trainer.
+                # See the original implementation or the nice implementation from @hiyouga
+                # here: https://github.com/hiyouga/LLaMA-Factory/commit/8664262cde3919e10eaecbd66e8c5d356856362e#diff-ebe08ab14496dfb9e06075f0fdd36799ef6d1535cc4dd4715b74c4e3e06fe3ba
+                if args.gradient_accumulation_steps != 1:
+                    raise ValueError("Layerwise GaLoRE optimizer do not support gradient accumulation !")
+
+                optimizer_dict = {}
+                for param in non_galore_params:
+                    param_groups = [{"params": [param]}]
+                    optimizer_dict[param] = optimizer_cls(param_groups, **optimizer_kwargs)
+                for param in galore_params:
+                    param_groups = [{"params": [param], **galore_optim_kwargs}]
+                    optimizer_dict[param] = optimizer_cls(param_groups, **optimizer_kwargs)
+
+                def optimizer_hook(param):
+                    if param.grad is not None:
+                        optimizer_dict[param].step()
+                        optimizer_dict[param].zero_grad()
+
+                for param in model.parameters():
+                    param.register_post_accumulate_grad_hook(optimizer_hook)
+
+                optimizer_cls = LayerWiseDummyOptimizer
+                optimizer_kwargs.update({"optimizer_dict": optimizer_dict})
+
+            optimizer_kwargs.update({"params": param_groups})
+
+            if args.optim == OptimizerNames.GALORE_ADAFACTOR:
+                optimizer_kwargs.update({"scale_parameter": False, "relative_step": False})
         else:
             raise ValueError(f"Trainer cannot instantiate unsupported optimizer: {args.optim}")
         return optimizer_cls, optimizer_kwargs
@@ -1182,6 +1314,30 @@ def create_scheduler(self, num_training_steps: int, optimizer: torch.optim.Optim
         Args:
             num_training_steps (int): The number of training steps to do.
         """
+        if optimizer is not None and isinstance(optimizer, LayerWiseDummyOptimizer):
+            optimizer_dict = optimizer.optimizer_dict
+            scheduler_dict = {}
+
+            for param in optimizer_dict.keys():
+                scheduler_dict[param] = get_scheduler(
+                    self.args.lr_scheduler_type,
+                    optimizer=optimizer_dict[param],
+                    num_warmup_steps=self.args.get_warmup_steps(num_training_steps) * 2,
+                    num_training_steps=num_training_steps * 2,
+                )
+
+            def scheduler_hook(param):
+                # Since the optimizer hook has been already attached we only need to
+                # attach the scheduler hook
+                if param.grad is not None:
+                    scheduler_dict[param].step()
+
+            for param in optimizer_dict.keys():
+                param.register_post_accumulate_grad_hook(scheduler_hook)
+
+            self._created_lr_scheduler = True
+            self.lr_scheduler = LayerWiseDummyScheduler()
+
         if self.lr_scheduler is None:
             self.lr_scheduler = get_scheduler(
                 self.args.lr_scheduler_type,

src/transformers/trainer_pt_utils.py

@@ -34,6 +34,7 @@
 import torch
 import torch.distributed as dist
 from torch import nn
+from torch.optim.lr_scheduler import LRScheduler
 from torch.utils.data import Dataset, IterableDataset, RandomSampler, Sampler
 from torch.utils.data.distributed import DistributedSampler
 
@@ -1226,3 +1227,47 @@ def from_json_file(cls, json_file):
 
     def to_dict(self):
         return copy.deepcopy(self.__dict__)
+
+
+class LayerWiseDummyOptimizer(torch.optim.Optimizer):
+    """
+    For Layer-wise optimizers such as GaLoRE optimizer, the optimization
+    step is already done through the post gradient hooks. Therefore
+    the trick is to create a dummy optimizer that can take arbitrary
+    args and kwargs and return a no-op during training.
+
+    Initial idea from @hiyouga in LLaMA-Factory:
+    https://github.com/hiyouga/LLaMA-Factory/commit/8664262cde3919e10eaecbd66e8c5d356856362e#diff-ebe08ab14496dfb9e06075f0fdd36799ef6d1535cc4dd4715b74c4e3e06fe3ba
+    """
+
+    def __init__(self, optimizer_dict=None, *args, **kwargs):
+        dummy_tensor = torch.randn(1, 1)
+        self.optimizer_dict = optimizer_dict
+        super().__init__([dummy_tensor], {"lr": 1e-03})
+
+    def zero_grad(self, set_to_none: bool = True) -> None:
+        pass
+
+    def step(self, closure=None) -> Optional[float]:
+        pass
+
+
+class LayerWiseDummyScheduler(LRScheduler):
+    """
+    For Layer-wise optimizers such as GaLoRE optimizer, the optimization and scheduling step
+    are already done through the post gradient hooks. Therefore
+    the trick is to create a dummy scheduler that can take arbitrary
+    args and kwargs and return a no-op during training.
+    """
+
+    def __init__(self, *args, **kwargs):
+        optimizer = LayerWiseDummyOptimizer()
+        last_epoch = -1
+        verbose = False
+        super().__init__(optimizer, last_epoch, verbose)
+
+    def get_lr(self):
+        return [group["lr"] for group in self.optimizer.param_groups]
+
+    def _get_closed_form_lr(self):
+        return self.base_lrs