Dion optimizer support

_quarto.yml

@@ -284,6 +284,7 @@ website:
             - docs/sequence_parallelism.qmd
             - docs/gradient_checkpointing.qmd
             - docs/nd_parallelism.qmd
+            - docs/optimizers.qmd
 
         - section: "Troubleshooting"
           contents:

docs/nd_parallelism.qmd

@@ -1,5 +1,5 @@
 ---
-title: "N-D Parallelism"
+title: "N-D Parallelism (Beta)"
 ---
 
 Axolotl enables training models at scale by composing different parallelism techniques. This is essential when:

docs/optimizers.qmd

@@ -0,0 +1,18 @@
+---
+title: Optimizers
+description: Configuring optimizers
+---
+
+### Dion Optimizer
+
+Microsoft's Dion (DIstributed OrthoNormalization) optimizer is a scalable and communication-efficient
+orthonormalizing optimizer that uses low-rank approximations to reduce gradient communication.
+
+Usage:
+
+```yaml
+optimizer: dion
+dion_lr: 0.01
+dion_momentum: 0.95
+lr: 0.00001  # learning rate for embeddings and parameters that fallback to AdamW
+```

requirements.txt

@@ -66,6 +66,6 @@ torchao==0.12.0
 schedulefree==1.4.1
 
 axolotl-contribs-lgpl==0.0.6
-axolotl-contribs-mit==0.0.3
+axolotl-contribs-mit==0.0.4
 
 mistral-common==1.8.3

src/axolotl/core/builders/base.py

@@ -267,6 +267,17 @@ def _configure_custom_optimizer(
 
                 optimizer_cls = MuonOptimizerFactory
                 optimizer_kwargs.update(adam_kwargs)
+            elif self.cfg.optimizer == "dion":
+                from axolotl.contribs.mit.dion import (  # pylint: disable=no-name-in-module
+                    DionOptimizerFactory,
+                )
+
+                optimizer_cls = DionOptimizerFactory
+                optimizer_kwargs["dion_lr"] = training_args_kwargs["dion_learning_rate"]
+                optimizer_kwargs["dion_mu"] = training_args_kwargs["dion_momentum"]
+                optimizer_kwargs.update(adam_kwargs)
+                partial_state = PartialState()
+                optimizer_kwargs["device_mesh"] = partial_state.device_mesh
             elif self.cfg.optimizer == "optimi_adamw":
                 from optimi import AdamW
 
@@ -516,10 +527,20 @@ def _set_base_training_args(
             "include_tokens_per_second",
             "weight_decay",
             "seed",
+            "dion_momentum",
+            "dion_rank_fraction",
+            "dion_rank_multiple_of",
         ]:
             if hasattr(self.cfg, arg) and getattr(self.cfg, arg) is not None:
                 training_args_kwargs[arg] = getattr(self.cfg, arg)
 
+        arg_map = {
+            "dion_learning_rate": "dion_lr",
+        }
+        for kwarg, cfg_arg in arg_map.items():
+            if hasattr(self.cfg, cfg_arg) and getattr(self.cfg, cfg_arg) is not None:
+                training_args_kwargs[kwarg] = getattr(self.cfg, cfg_arg)
+
         training_args_kwargs["per_device_train_batch_size"] = self.cfg.micro_batch_size
         training_args_kwargs["average_tokens_across_devices"] = False
 

src/axolotl/core/training_args_base.py

@@ -243,3 +243,18 @@ class AxolotlTrainingMixins:
     )
 
     # end of multi-modal section
+
+    dion_learning_rate: float | None = field(
+        default=None,
+        metadata={"help": "The learning rate for Dion"},
+    )
+    dion_momentum: float | None = field(
+        default=None,
+        metadata={"help": "The momentum for Dion"},
+    )
+    dion_rank_fraction: float | None = field(
+        default=None,
+    )
+    dion_rank_multiple_of: int | None = field(
+        default=None,
+    )

src/axolotl/integrations/base.py

@@ -26,9 +26,11 @@
 from typing import TYPE_CHECKING, Callable, OrderedDict, Union
 
 from peft import PeftModel
+from torch import nn
 from torch.optim import Optimizer
 from torch.optim.lr_scheduler import LRScheduler
 from transformers import PreTrainedModel, Trainer
+from transformers.trainer_pt_utils import get_parameter_names
 
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.logging import get_logger
@@ -641,3 +643,24 @@ def __call__(
         self, opt_model, training_args, **optimizer_kwargs
     ) -> Optimizer | None:
         pass
+
+    # duplicated from transformers
+    def get_decay_parameter_names(self, model) -> list[str]:
+        """
+        Get all parameter names that weight decay will be applied to.
+
+        This function filters out parameters in two ways:
+        1. By layer type (instances of layers specified in ALL_LAYERNORM_LAYERS)
+        2. By parameter name patterns (containing 'bias', or variation of 'norm')
+        """
+        forbidden_name_patterns = [
+            r"bias",
+            r"layernorm",
+            r"rmsnorm",
+            r"(?:^|\.)norm(?:$|\.)",
+            r"_norm(?:$|\.)",
+        ]
+        decay_parameters = get_parameter_names(
+            model, [nn.LayerNorm], forbidden_name_patterns
+        )
+        return decay_parameters

src/axolotl/utils/schemas/enums.py

@@ -79,6 +79,7 @@ class CustomSupportedOptimizers(str, Enum):
     adopt_adamw = "adopt_adamw"
     came_pytorch = "came_pytorch"
     muon = "muon"
+    dion = "dion"
 
 
 class RingAttnFunc(str, Enum):

src/axolotl/utils/schemas/training.py

@@ -138,6 +138,26 @@ class HyperparametersConfig(BaseModel):
     adam_beta3: float | None = Field(
         default=None, json_schema_extra={"description": "only used for CAME Optimizer"}
     )
+
+    dion_lr: float | None = Field(
+        default=None, json_schema_extra={"description": "Dion Optimizer learning rate"}
+    )
+    dion_momentum: float | None = Field(
+        default=None, json_schema_extra={"description": "Dion Optimizer momentum"}
+    )
+    dion_rank_fraction: float | None = Field(
+        default=1.0,
+        json_schema_extra={
+            "description": "Dion Optimizer: r/d fraction for low-rank approximation. Used to compute the low-rank dimension."
+        },
+    )
+    dion_rank_multiple_of: int | None = Field(
+        default=1,
+        json_schema_extra={
+            "description": "Dion Optimizer: Round up the low-rank dimension to a multiple of this number. This may be useful to ensure even sharding."
+        },
+    )
+
     max_grad_norm: float | None = Field(
         default=None, json_schema_extra={"description": "Gradient clipping max norm"}
     )

tests/e2e/test_optimizers.py

@@ -13,6 +13,7 @@
     check_model_output_exists,
     require_torch_2_5_1,
     require_torch_2_6_0,
+    require_torch_2_7_0,
     with_temp_dir,
 )
 
@@ -160,6 +161,49 @@ def test_muon(self, temp_dir):
         check_model_output_exists(temp_dir, cfg)
         assert "Muon" in trainer.optimizer.optimizer.__class__.__name__
 
+    @with_temp_dir
+    @require_torch_2_7_0
+    def test_dion(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "model_type": "AutoModelForCausalLM",
+                "tokenizer_type": "AutoTokenizer",
+                "sequence_len": 1024,
+                "val_set_size": 0.0,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                },
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 5,
+                "micro_batch_size": 8,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "dion",
+                "dion_lr": 0.01,
+                "dion_momentum": 0.95,
+                "lr_scheduler": "cosine",
+                "weight_decay": 0.01,
+                "save_first_step": False,
+            }
+        )
+
+        cfg = validate_config(cfg)
+        normalize_config(cfg)
+        dataset_meta = load_datasets(cfg=cfg)
+
+        _, _, trainer = train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)
+        assert "Dion" in trainer.optimizer.optimizer.__class__.__name__
+
     @with_temp_dir
     def test_fft_schedule_free_adamw(self, temp_dir):
         # pylint: disable=duplicate-code