FEAT / Optim: Add GaLore optimizer

docs/source/en/trainer.md

@@ -252,6 +252,57 @@ 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 `galore_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",
+    galore_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).
+
+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.
+
 ## 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

@@ -1241,6 +1241,9 @@ def get_optimizer_cls_and_kwargs(
             ]
 
             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

tests/trainer/test_trainer.py

@@ -60,6 +60,7 @@
     require_accelerate,
     require_bitsandbytes,
     require_deepspeed,
+    require_galore_torch,
     require_intel_extension_for_pytorch,
     require_optuna,
     require_peft,
@@ -114,6 +115,8 @@
         GPT2Config,
         GPT2LMHeadModel,
         LineByLineTextDataset,
+        LlamaConfig,
+        LlamaForCausalLM,
         PreTrainedModel,
         Trainer,
         TrainerState,
@@ -1069,6 +1072,69 @@ def test_dataloader_without_dataset(self):
         trainer.train()
         trainer.evaluate()
 
+    @require_galore_torch
+    def test_galore(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="galore_adamw",
+                galore_target_modules=["attn", "mlp"],
+            )
+            trainer = Trainer(tiny_llama, args, train_dataset=train_dataset)
+
+            # Check this works
+            _ = trainer.train()
+
+    @require_galore_torch
+    def test_galore_adamw_8bit(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="galore_adamw_8bit",
+                galore_target_modules=["attn", "mlp"],
+            )
+            trainer = Trainer(tiny_llama, args, train_dataset=train_dataset)
+
+            # Check this works
+            _ = trainer.train()
+
+    @require_galore_torch
+    def test_galore_adafactor(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="galore_adafactor",
+                galore_target_modules=["attn", "mlp"],
+            )
+            trainer = Trainer(tiny_llama, args, train_dataset=train_dataset)
+
+            # Check this works
+            _ = trainer.train()
+
     @require_torch_multi_accelerator
     def test_data_is_not_parallelized_when_model_is_parallel(self):
         model = RegressionModel()