Optim: APOLLO optimizer integration

docs/source/en/trainer.md

@@ -445,6 +445,139 @@ 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.
 
+
+### APOLLO
+
+ Approximated Gradient Scaling for Memory Efficient LLM Optimization (APOLLO) is a memory-efficient low-rank training strategy that allows full-parameter learning for both pre-training and fine-tuning, while maintaining AdamW-level performance with SGD-like memory efficiency.
+
+ * **Ultra-low rank efficiency** → Requires much lower rank than GaLore—even rank 1 (APOLLO-Mini) suffices.
+ * **No expensive SVD computations** → Unlike GaLore, APOLLO leverages random projection, avoiding training stalls.
+
+First make sure to install APOLLO from its official repository:
+
+```bash
+pip install apollo-torch
+```
+
+Then simply add one of `["apollo_adamw"]` in `optim` together with `optim_target_modules`, which can be a list of strings, regex or full path 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-apollo",
+    max_steps=100,
+    per_device_train_batch_size=2,
+    optim="apollo_adamw",
+    optim_target_modules=[r".*.attn.*", r".*.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()
+```
+
+To pass extra arguments supported by APOLLO, you should pass correctly `optim_args`, for example:
+
+```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=[r".*.attn.*", r".*.mlp.*"],
+    optim_args="proj=random,scale_type=tensor,rank=128,update_proj_gap=100,scale=1.0",
+
+)
+
+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()
+```
+
+Currently only Linear layers are considered to use the APOLLO optimizers, while the remaining modueles are still using AdamW.
+
+You can read more about the method in the [original repository](https://github.com/zhuhanqing/APOLLO) or the [paper](https://arxiv.org/abs/2412.05270).
+
+
+You can also perform layer-wise APOLLO by simply 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-apollo",
+    max_steps=100,
+    per_device_train_batch_size=2,
+    optim="apollo_adamw_layerwise",
+    optim_target_modules=[r".*.attn.*", r".*.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()
+```
+
+
 ### 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).

src/transformers/testing_utils.py

@@ -62,6 +62,7 @@
     GGUF_MIN_VERSION,
     is_accelerate_available,
     is_apex_available,
+    is_apollo_torch_available,
     is_aqlm_available,
     is_auto_awq_available,
     is_auto_gptq_available,
@@ -403,6 +404,14 @@ def require_galore_torch(test_case):
     return unittest.skipUnless(is_galore_torch_available(), "test requires GaLore")(test_case)
 
 
+def require_apollo_torch(test_case):
+    """
+    Decorator marking a test that requires GaLore. These tests are skipped when APOLLO isn't installed.
+    https://github.com/zhuhanqing/APOLLO
+    """
+    return unittest.skipUnless(is_apollo_torch_available(), "test requires APOLLO")(test_case)
+
+
 def require_lomo(test_case):
     """
     Decorator marking a test that requires LOMO. These tests are skipped when LOMO-optim isn't installed.

src/transformers/trainer.py

@@ -151,6 +151,7 @@
     find_labels,
     is_accelerate_available,
     is_apex_available,
+    is_apollo_torch_available,
     is_bitsandbytes_available,
     is_datasets_available,
     is_galore_torch_available,
@@ -1582,6 +1583,119 @@ def optimizer_hook(param):
 
             if args.optim == OptimizerNames.GALORE_ADAFACTOR:
                 optimizer_kwargs.update({"scale_parameter": False, "relative_step": False})
+        elif args.optim in [
+            OptimizerNames.APOLLO_ADAMW,
+            OptimizerNames.APOLLO_ADAMW_LAYERWISE,
+        ]:
+            if not is_apollo_torch_available():
+                raise ImportError(
+                    "You need to install `apollo_torch` in order to use APOLLO optimizers"
+                    " install it with `pip install git+https://github.com/zhuhanqing/APOLLO`"
+                )
+            from apollo_torch import APOLLOAdamW
+
+            is_layerwise = args.optim.lower().endswith("layerwise")
+            if is_layerwise and args.parallel_mode == ParallelMode.DISTRIBUTED:
+                raise NotImplementedError("Layer-wise APOLLO does not support DDP at this time")
+
+            optimizer_mapping = {
+                OptimizerNames.APOLLO_ADAMW: APOLLOAdamW,
+                OptimizerNames.APOLLO_ADAMW_LAYERWISE: APOLLOAdamW,
+            }
+
+            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 APOLLO 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 APOLLO optimizer.")
+
+            all_linear = (
+                isinstance(args.optim_target_modules, str)
+                and args.optim_target_modules.replace("_", "-") == "all-linear"
+            )
+
+            apollo_params = []
+            apollo_params_names = []
+            for module_name, module in model.named_modules():
+                target_module_exists, is_regex = check_target_module_exists(
+                    args.optim_target_modules, module_name, return_is_regex=True
+                )
+
+                if not isinstance(module, nn.Linear):
+                    # Warn in case we match but it's not a linear layer
+                    if target_module_exists and not is_regex:
+                        logger.warning(
+                            f"{module_name} has been matched but ignored as APOLLO only supports linear layers. Please double check your `optim_target_modules`!"
+                        )
+
+                    continue
+
+                if not target_module_exists and not all_linear:
+                    continue
+
+                apollo_params.append(module.weight)
+                apollo_params_names.append(module_name + ".weight")
+
+            if len(apollo_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_apollo_params = [p for n, p in model.named_parameters() if n not in apollo_params_names]
+            apollo_optim_kwargs = {
+                "rank": int(optim_args.pop("rank", 128)),
+                "proj": optim_args.pop("proj", "random"),
+                "scale_type": optim_args.pop("scale_type", "channel"),
+                "update_proj_gap": int(optim_args.pop("update_proj_gap", 200)),
+                "scale": float(optim_args.pop("scale", 1.0)),
+                "proj_type": optim_args.pop("proj_type", "std"),
+            }
+
+            # The default args are from the official repository: https://github.com/zhuhanqing/APOLLO
+            param_groups = [
+                {"params": non_apollo_params},
+                {"params": apollo_params, **apollo_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 APOLLO optimizer do not support gradient accumulation !")
+
+                optimizer_dict = {}
+                for param in non_apollo_params:
+                    param_groups = [{"params": [param]}]
+                    optimizer_dict[param] = optimizer_cls(param_groups, **optimizer_kwargs)
+                for param in apollo_params:
+                    param_groups = [{"params": [param], **apollo_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():
+                    if param.requires_grad:
+                        param.register_post_accumulate_grad_hook(optimizer_hook)
+
+                optimizer_cls = LayerWiseDummyOptimizer
+                optimizer_kwargs.update({"optimizer_dict": optimizer_dict})
+
+            optimizer_kwargs.update({"params": param_groups})
         elif args.optim in [OptimizerNames.LOMO, OptimizerNames.ADALOMO]:
             if not is_lomo_available():
                 raise ImportError(

src/transformers/training_args.py

@@ -184,6 +184,8 @@ class OptimizerNames(ExplicitEnum):
     GROKADAMW = "grokadamw"
     SCHEDULE_FREE_ADAMW = "schedule_free_adamw"
     SCHEDULE_FREE_SGD = "schedule_free_sgd"
+    APOLLO_ADAMW = "apollo_adamw"
+    APOLLO_ADAMW_LAYERWISE = "apollo_adamw_layerwise"
 
 
 # Sometimes users will pass in a `str` repr of a dict in the CLI

src/transformers/utils/__init__.py

@@ -118,6 +118,7 @@
     get_torch_version,
     is_accelerate_available,
     is_apex_available,
+    is_apollo_torch_available,
     is_aqlm_available,
     is_auto_awq_available,
     is_auto_gptq_available,

src/transformers/utils/import_utils.py

@@ -98,6 +98,7 @@ def _is_package_available(pkg_name: str, return_version: bool = False) -> Union[
 
 _accelerate_available, _accelerate_version = _is_package_available("accelerate", return_version=True)
 _apex_available = _is_package_available("apex")
+_apollo_torch_available = _is_package_available("apollo_torch")
 _aqlm_available = _is_package_available("aqlm")
 _vptq_available, _vptq_version = _is_package_available("vptq", return_version=True)
 _av_available = importlib.util.find_spec("av") is not None
@@ -402,6 +403,10 @@ def is_galore_torch_available():
     return _galore_torch_available
 
 
+def is_apollo_torch_available():
+    return _apollo_torch_available
+
+
 def is_lomo_available():
     return _lomo_available