add SOAP optimizer

_quarto.yml

@@ -243,7 +243,6 @@ website:
             - docs/unsloth.qmd
             - docs/torchao.qmd
             - docs/custom_integrations.qmd
-            - docs/sequence_parallelism.qmd
 
         - section: "Troubleshooting"
           contents:

docs/config.qmd

@@ -658,9 +658,6 @@ ddp_broadcast_buffers:
 # subsequences, or set to 4 to split into four equal-sized subsequences.
 # See https://axolotl-ai-cloud.github.io/axolotl/docs/sequence_parallelism.html for more details.
 sequence_parallel_degree:
-# Optional; strides across the key dimension. Larger values use more memory but should make training faster.
-# Must evenly divide the number of KV heads in your model.
-heads_k_stride: 1
 
 # Path to torch distx for optim 'adamw_anyprecision'
 torchdistx_path:

docs/multi-gpu.qmd

@@ -18,7 +18,6 @@ Axolotl supports several methods for multi-GPU training:
 
 - DeepSpeed (recommended)
 - FSDP (Fully Sharded Data Parallel)
-- Sequence parallelism
 - FSDP + QLoRA
 
 ## DeepSpeed {#sec-deepspeed}
@@ -67,28 +66,6 @@ fsdp_config:
   fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
 ```
 
-## Sequence parallelism {#sec-sequence-parallelism}
-
-We support sequence parallelism (SP) via the
-[ring-flash-attention](https://github.com/zhuzilin/ring-flash-attention) project. This
-allows one to split up sequences across GPUs, which is useful in the event that a
-single sequence causes OOM errors during model training.
-
-First, install `ring-flash-attn`, recommended via `pip install axolotl[ring-flash-attn]`,
-or from source with `pip install .[ring-flash-attn]`.
-
-Your Axolotl YAML config should contain the following lines:
-
-```{.yaml}
-sequence_parallel_degree: 4  # Split each sequence into 4 parts, one per GPU
-flash_attention: true  # Required with sequence parallelism
-
-# Optional; strides across the key dimension. Larger values use more memory but will make training faster.
-heads_k_stride: 1
-```
-
-See our [dedicated guide](sequence_parallelism.qmd) for more details.
-
 ### FSDP + QLoRA {#sec-fsdp-qlora}
 
 For combining FSDP with QLoRA, see our [dedicated guide](fsdp_qlora.qmd).

docs/sequence_parallelism.qmd

@@ -25,8 +25,6 @@ To enable sequence parallelism, add the following to your configuration file:
 ```yaml
 # Set to a divisor (> 1) of the number of GPUs available
 sequence_parallel_degree: 4  # Split sequences across 4 GPUs
-# Optional; strides across the key dimension. Larger values use more memory but should make training faster.
-heads_k_stride: 1
 ```
 
 The `sequence_parallel_degree` should be a divisor of the total number of GPUs. For example:
@@ -60,16 +58,11 @@ To use sequence parallelism, you need:
 ## Example
 
 ```yaml
+# Example config with sequence parallelism
 base_model: meta-llama/Llama-3-8B-Instruct
 sequence_len: 8192
-
+sequence_parallel_degree: 2  # Split each sequence into 4 parts
-...
-
-sequence_parallel_degree: 4  # Split each sequence into 4 parts, one per GPU
 flash_attention: true  # Required with sequence parallelism
-# Optional; strides across the key dimension. Larger values use more memory but should make training faster.
-heads_k_stride: 1
-
 ...
 ```
 

src/axolotl/core/trainer_builder.py

@@ -69,6 +69,7 @@ from axolotl.utils.callbacks import (
     LossWatchDogCallback,
     SaveAxolotlConfigtoWandBCallback,
     SaveBetterTransformerModelCallback,
+    SaveModelCallback,
     bench_eval_callback_factory,
     causal_lm_bench_eval_callback_factory,
     log_prediction_callback_factory,
@@ -248,6 +249,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
 
         if self.cfg.gc_steps:
             callbacks.append(GCCallback(gc_steps=self.cfg.gc_steps))
+        callbacks.append(SaveModelCallback())
 
         return callbacks
 
@@ -661,6 +663,11 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
 
                 optimizer_cls = MuonOptimizerFactory
                 optimizer_kwargs.update(adam_kwargs)
+            elif self.cfg.optimizer == "soap":
+                from axolotl.utils.optimizers.soap import SOAP
+
+                optimizer_cls = SOAP
+                optimizer_kwargs.update(adam_kwargs)
             elif self.cfg.optimizer == "optimi_adamw":
                 from optimi import AdamW
 
@@ -935,6 +942,7 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
 
     def get_callbacks(self):
         callbacks = super().get_callbacks()
+        callbacks.append(SaveModelCallback())
 
         return callbacks
 

src/axolotl/evaluate.py

@@ -15,7 +15,6 @@ from axolotl.logging_config import configure_logging
 from axolotl.train import TrainDatasetMeta
 from axolotl.utils import set_pytorch_cuda_alloc_conf
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.distributed import cleanup_distributed
 from axolotl.utils.models import load_model, load_processor, load_tokenizer
 from axolotl.utils.trainer import setup_trainer
 
@@ -160,6 +159,4 @@ def evaluate(*, cfg: DictDefault, dataset_meta: TrainDatasetMeta) -> Dict[str, f
     del model
     del tokenizer
 
-    cleanup_distributed()
-
     return all_metrics

src/axolotl/monkeypatch/attention/ring_attn.py

@@ -38,19 +38,13 @@ def set_ring_attn_group(ring_attn_group: dist.ProcessGroup | None):
     RING_ATTN_GROUP = ring_attn_group
 
 
-def register_ring_attn(sequence_parallel_degree: int, heads_k_stride: int | None):
+def register_ring_attn(sequence_parallel_degree: int):
     """
     Create ring attention group and substitute flash attn with ring flash attn.
 
     Args:
         sequence_parallel_degree: Sequence parallelism factor.
-        heads_k_stride: Sequence parallelism K head stride size. Passed
-            through to `ring_flash_attn.substitute_hf_flash_attn`.
     """
-    if get_ring_attn_group() is not None:
-        LOG.info("Ring attention already registered, exiting early...")
-        return
-
     LOG.info(
         "Enabling ring attention sequence parallelism: "
         f"each sequence will be processed across {sequence_parallel_degree} GPUs"
@@ -90,11 +84,6 @@ def register_ring_attn(sequence_parallel_degree: int, heads_k_stride: int | None
     if rank == 0:
         LOG.info(f"Sequence parallel group assignments: {group_assignments}")
 
-    if heads_k_stride is None:
-        heads_k_stride = 1
-
     from ring_flash_attn import substitute_hf_flash_attn
 
-    substitute_hf_flash_attn(
+    substitute_hf_flash_attn(get_ring_attn_group(), sequence_parallel_degree)
-        process_group=get_ring_attn_group(), heads_k_stride=heads_k_stride
-    )

src/axolotl/train.py

@@ -27,7 +27,6 @@ from axolotl.contribs.lgpl import (  # pylint: disable = no-name-in-module
 from axolotl.core.trainer_builder import HFCausalTrainerBuilder, HFRLTrainerBuilder
 from axolotl.logging_config import configure_logging
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.distributed import cleanup_distributed
 from axolotl.utils.freeze import freeze_layers_except
 from axolotl.utils.models import load_model, load_processor, load_tokenizer
 from axolotl.utils.trainer import setup_trainer
@@ -158,8 +157,6 @@ def setup_signal_handler(
                 _model.save_pretrained(
                     cfg.output_dir, safe_serialization=safe_serialization
                 )
-
-            cleanup_distributed()
             sys.exit(0)
 
         _model_weakref = weakref.ref(model)
@@ -481,7 +478,7 @@ def train(
     Returns:
         Tuple of (model, tokenizer) after training
     """
-    # Setup model, tokenizer, (causal or RLHF) trainer, etc.
+    # Setup model, tokenizer, (causal or RLHF) trainer etc.
     (
         trainer,
         model,
@@ -490,26 +487,34 @@ def train(
         processor,
     ) = setup_model_and_trainer(cfg, dataset_meta)
 
-    # Handle untrained tokens if configured
+    # Determine if we need to resume from a checkpoint
+    resume_from_checkpoint = determine_resume_checkpoint(cfg)
+
+    # Configuration for saving
     safe_serialization = cfg.save_safetensors is True
+
+    # Handle untrained tokens if configured
     train_dataset = dataset_meta.train_dataset
     handle_untrained_tokens_fix(
         cfg, model, tokenizer, train_dataset, safe_serialization
     )
 
-    # Additional setup
+    # Save initial configs
     save_initial_configs(cfg, tokenizer, model, peft_config, processor)
+
+    # Set up signal handler for graceful termination
     setup_signal_handler(cfg, model, safe_serialization)
+
+    # Set up badges and config info for model card
     setup_model_card(cfg)
 
     # Execute the training
-    resume_from_checkpoint = determine_resume_checkpoint(cfg)
     execute_training(cfg, trainer, resume_from_checkpoint)
 
-    # Save the trained model and cleanup
+    # Save the trained model
     save_trained_model(cfg, trainer, model, safe_serialization)
+
+    # Create model card
     create_model_card(cfg, trainer)
-    if not cfg.use_ray:
-        cleanup_distributed()
 
     return model, tokenizer, trainer

src/axolotl/utils/callbacks/__init__.py

@@ -816,6 +816,27 @@ class SaveAxolotlConfigtoWandBCallback(TrainerCallback):
         return control
 
 
+class SaveModelCallback(TrainerCallback):
+    """Callback to save model on train end"""
+
+    def on_step_end(  # pylint: disable=unused-argument
+        self,
+        args: TrainingArguments,
+        state: TrainerState,
+        control: TrainerControl,
+        **kwargs,
+    ):
+        # Save
+        if state.global_step >= state.max_steps:
+            control.should_save = True
+
+    def on_train_end(  # pylint: disable=unused-argument
+        self, args, state, control, **kwargs
+    ):
+        control.should_save = True
+        return control
+
+
 class GCCallback(TrainerCallback):
     """Callback to garbage collect torch cache"""
 

src/axolotl/utils/distributed.py

@@ -71,8 +71,8 @@ def barrier():
 
 def is_main_process():
     """
-    Check if the current process is the main process. If not in distributed mode,
+    Check if the current process is the main process.
-    always return `True`.
+    If not in distributed mode, always return True.
     """
     if not is_distributed():
         return True
@@ -87,18 +87,6 @@ def get_world_size():
     return int(os.getenv("WORLD_SIZE", "1"))
 
 
-def cleanup_distributed():
-    """
-    Destroy process group if torch distributed is initialized. Called in training early
-    termination or when training successfully completes.
-    """
-    # Ensure that all operations are completed before destroying the process group
-    torch.cuda.synchronize()
-    # Destroy the process group
-    if torch.distributed.is_initialized():
-        torch.distributed.destroy_process_group()
-
-
 @contextmanager
 def zero_only():
     """

src/axolotl/utils/models.py

@@ -609,10 +609,7 @@ class ModelLoader:
             # Initialize ring attn for sequence parallelism. This must be done after
             # model init but before the first forward pass, since it modifies flash
             # attn to use ring comm for SP training across multiple GPUs.
-            register_ring_attn(
+            register_ring_attn(self.cfg.sequence_parallel_degree)
-                sequence_parallel_degree=self.cfg.sequence_parallel_degree,
-                heads_k_stride=self.cfg.heads_k_stride,
-            )
 
     def patch_attention(self) -> None:
         if hasattr(self.model_config, "model_type"):

src/axolotl/utils/optimizers/soap/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2024 Nikhil Vyas
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

src/axolotl/utils/optimizers/soap/__init__.py

@@ -0,0 +1,495 @@
+# pylint: skip-file
+# Copied from https://github.com/nikhilvyas/SOAP
+from itertools import chain
+
+import torch
+import torch.optim as optim
+
+# Parts of the code are modifications of Pytorch's AdamW optimizer
+# Parts of the code are modifications of code from https://github.com/jiaweizzhao/GaLore/blob/master/galore_torch/galore_projector.py
+
+
+class SOAP(optim.Optimizer):
+    """
+    Implements SOAP algorithm (https://arxiv.org/abs/2409.11321).
+
+    Parameters:
+        params (`Iterable[nn.parameter.Parameter]`):
+            Iterable of parameters to optimize or dictionaries defining parameter groups.
+        lr (`float`, *optional*, defaults to 0.003):
+            The learning rate to use.
+        betas (`Tuple[float,float]`, *optional*, defaults to `(0.95, 0.95)`):
+            Adam's betas parameters (b1, b2).
+        shampoo_beta (`float`, *optional*, defaults to -1):
+            If >= 0, use this beta for the preconditioner (L and R in paper, state["GG"] below) moving average instead of betas[1].
+        eps (`float`, *optional*, defaults to 1e-08):
+            Adam's epsilon for numerical stability.
+        weight_decay (`float`, *optional*, defaults to 0.01): weight decay coefficient.
+        precondition_frequency (`int`, *optional*, defaults to 10):
+            How often to update the preconditioner.
+        max_precond_dim (`int`, *optional*, defaults to 10000):
+            Maximum dimension of the preconditioner.
+            Set to 10000, so that we exclude most common vocab sizes while including layers.
+        merge_dims (`bool`, *optional*, defaults to `False`):
+            Whether or not to merge dimensions of the preconditioner.
+        precondition_1d (`bool`, *optional*, defaults to `False`):
+            Whether or not to precondition 1D gradients.
+        normalize_grads (`bool`, *optional*, defaults to `False`):
+            Whether or not to normalize gradients per layer.
+            Helps at large precondition_frequency (~100 in our experiments),
+            but hurts performance at small precondition_frequency (~10 in our experiments).
+        data_format (`str`, *optional*, defaults to `channels_first`):
+            Data format of the input for convolutional layers.
+            Should be "channels_last" for data_format of NHWC and "channels_first" for NCHW.
+        correct_bias (`bool`, *optional*, defaults to `True`):
+            Whether or not to use bias correction in Adam.
+    """
+
+    def __init__(
+        self,
+        params,
+        lr: float = 3e-3,
+        betas=(0.95, 0.95),
+        shampoo_beta: float = -1,
+        eps: float = 1e-8,
+        weight_decay: float = 0.01,
+        precondition_frequency: int = 10,
+        max_precond_dim: int = 10000,  #
+        merge_dims: bool = False,  # Merge dimensions till the product of the dimensions is less than or equal to max_precond_dim.
+        precondition_1d: bool = False,
+        normalize_grads: bool = False,
+        data_format: str = "channels_first",
+        correct_bias: bool = True,
+    ):
+        defaults = {
+            "lr": lr,
+            "betas": betas,
+            "shampoo_beta": shampoo_beta,
+            "eps": eps,
+            "weight_decay": weight_decay,
+            "precondition_frequency": precondition_frequency,
+            "max_precond_dim": max_precond_dim,
+            "merge_dims": merge_dims,
+            "precondition_1d": precondition_1d,
+            "normalize_grads": normalize_grads,
+            "correct_bias": correct_bias,
+        }
+        super().__init__(params, defaults)
+        self._data_format = data_format
+
+    def merge_dims(self, grad, max_precond_dim):
+        """
+        Merges dimensions of the gradient tensor till the product of the dimensions is less than or equal to max_precond_dim.
+        """
+        assert self._data_format in ["channels_first", "channels_last"]
+        if self._data_format == "channels_last" and grad.dim() == 4:
+            grad = grad.permute(0, 3, 1, 2)
+        shape = grad.shape
+        new_shape = []
+
+        curr_shape = 1
+        for sh in shape:
+            temp_shape = curr_shape * sh
+            if temp_shape > max_precond_dim:
+                if curr_shape > 1:
+                    new_shape.append(curr_shape)
+                    curr_shape = sh
+                else:
+                    new_shape.append(sh)
+                    curr_shape = 1
+            else:
+                curr_shape = temp_shape
+
+        if curr_shape > 1 or len(new_shape) == 0:
+            new_shape.append(curr_shape)
+
+        new_grad = grad.reshape(new_shape)
+        return new_grad
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """
+        Performs a single optimization step.
+
+        Arguments:
+            closure (`Callable`, *optional*): A closure that reevaluates the model and returns the loss.
+        """
+        if closure is None:
+            loss = None
+        else:
+            loss = closure()
+
+        for group in self.param_groups:
+            for p in group["params"]:
+                if p.grad is None:
+                    continue
+                grad = p.grad
+
+                state = self.state[p]
+
+                if "step" not in state:
+                    state["step"] = 0
+
+                    # State initialization
+                if "exp_avg" not in state:
+                    # Exponential moving average of gradient values
+                    state["exp_avg"] = torch.zeros_like(grad)
+                    # Exponential moving average of squared gradient values
+                    state["exp_avg_sq"] = torch.zeros_like(grad)
+
+                if "Q" not in state:
+                    self.init_preconditioner(
+                        grad,
+                        state,
+                        precondition_frequency=group["precondition_frequency"],
+                        precondition_1d=group["precondition_1d"],
+                        shampoo_beta=(
+                            group["shampoo_beta"]
+                            if group["shampoo_beta"] >= 0
+                            else group["betas"][1]
+                        ),
+                        max_precond_dim=group["max_precond_dim"],
+                        merge_dims=group["merge_dims"],
+                    )
+                    self.update_preconditioner(
+                        grad,
+                        state,
+                        max_precond_dim=group["max_precond_dim"],
+                        merge_dims=group["merge_dims"],
+                        precondition_1d=group["precondition_1d"],
+                    )
+                    continue  # first step is skipped so that we never use the current gradients in the projection.
+
+                # Projecting gradients to the eigenbases of Shampoo's preconditioner
+                # i.e. projecting to the eigenbases of matrices in state["GG"]
+                grad_projected = self.project(
+                    grad,
+                    state,
+                    merge_dims=group["merge_dims"],
+                    max_precond_dim=group["max_precond_dim"],
+                )
+
+                exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"]
+                beta1, beta2 = group["betas"]
+
+                state["step"] += 1
+
+                # Decay the first and second moment running average coefficient
+                # In-place operations to update the averages at the same time
+                exp_avg.mul_(beta1).add_(grad_projected, alpha=(1.0 - beta1))
+                exp_avg_sq.mul_(beta2).add_(
+                    grad_projected.square(), alpha=(1.0 - beta2)
+                )
+
+                denom = exp_avg_sq.sqrt().add_(group["eps"])
+
+                # Projecting the exponential moving average of gradients to the eigenbases of Shampoo's preconditioner
+                # i.e. projecting to the eigenbases of matrices in state["GG"]
+                # exp_avg_projected = self.project(
+                #     exp_avg,
+                #     state,
+                #     merge_dims=group["merge_dims"],
+                #     max_precond_dim=group["max_precond_dim"],
+                # )
+                exp_avg_projected = exp_avg
+
+                step_size = group["lr"]
+                if group["correct_bias"]:
+                    bias_correction1 = 1.0 - beta1 ** (state["step"])
+                    bias_correction2 = 1.0 - beta2 ** (state["step"])
+                    step_size = step_size * (bias_correction2**0.5) / bias_correction1
+
+                # Projecting back the preconditioned (by Adam) exponential moving average of gradients
+                # to the original space
+                norm_grad = self.project_back(
+                    exp_avg_projected / denom,
+                    state,
+                    merge_dims=group["merge_dims"],
+                    max_precond_dim=group["max_precond_dim"],
+                )
+
+                if group["normalize_grads"]:
+                    norm_grad = norm_grad / (1e-30 + torch.mean(norm_grad**2) ** 0.5)
+
+                p.add_(norm_grad, alpha=-step_size)
+
+                # From AdamW code: Just adding the square of the weights to the loss function is *not*
+                # the correct way of using L2 regularization/weight decay with Adam,
+                # since that will interact with the m and v parameters in strange ways.
+                #
+                # Instead we want to decay the weights in a manner that doesn't interact
+                # with the m/v parameters. This is equivalent to adding the square
+                # of the weights to the loss with plain (non-momentum) SGD.
+                # Add weight decay at the end (fixed version)
+                if group["weight_decay"] > 0.0:
+                    p.add_(p, alpha=(-group["lr"] * group["weight_decay"]))
+
+                # Update is done after the gradient step to avoid using current gradients in the projection.
+                self.update_preconditioner(
+                    grad,
+                    state,
+                    max_precond_dim=group["max_precond_dim"],
+                    merge_dims=group["merge_dims"],
+                    precondition_1d=group["precondition_1d"],
+                )
+
+        return loss
+
+    def init_preconditioner(
+        self,
+        grad,
+        state,
+        precondition_frequency=10,
+        shampoo_beta=0.95,
+        max_precond_dim=10000,
+        precondition_1d=False,
+        merge_dims=False,
+    ):
+        """
+        Initializes the preconditioner matrices (L and R in the paper).
+        """
+        state["GG"] = (
+            []
+        )  # Will hold all the preconditioner matrices (L and R in the paper).
+        if grad.dim() == 1:
+            if not precondition_1d or grad.shape[0] > max_precond_dim:
+                state["GG"].append([])
+            else:
+                state["GG"].append(
+                    torch.zeros(grad.shape[0], grad.shape[0], device=grad.device)
+                )
+        else:
+            if merge_dims:
+                grad = self.merge_dims(grad, max_precond_dim)
+
+            for sh in grad.shape:
+                if sh > max_precond_dim:
+                    state["GG"].append([])
+                else:
+                    state["GG"].append(torch.zeros(sh, sh, device=grad.device))
+
+        state["Q"] = None  # Will hold all the eigenbases of the preconditioner.
+        state["precondition_frequency"] = precondition_frequency
+        state["shampoo_beta"] = shampoo_beta
+
+    def project(self, grad, state, merge_dims=False, max_precond_dim=10000):
+        """
+        Projects the gradient to the eigenbases of the preconditioner.
+        """
+        original_shape = grad.shape
+        if merge_dims:
+            if grad.dim() == 4 and self._data_format == "channels_last":
+                permuted_shape = grad.permute(0, 3, 1, 2).shape
+            grad = self.merge_dims(grad, max_precond_dim)
+
+        for mat in state["Q"]:
+            if len(mat) > 0:
+                grad = torch.tensordot(
+                    grad,
+                    mat,
+                    dims=[[0], [0]],
+                )
+            else:
+                permute_order = list(range(1, len(grad.shape))) + [0]
+                grad = grad.permute(permute_order)
+
+        if merge_dims:
+            if self._data_format == "channels_last" and len(original_shape) == 4:
+                grad = grad.reshape(permuted_shape).permute(0, 2, 3, 1)
+            else:
+                grad = grad.reshape(original_shape)
+        return grad
+
+    def update_preconditioner(
+        self,
+        grad,
+        state,
+        max_precond_dim=10000,
+        merge_dims=False,
+        precondition_1d=False,
+    ):
+        """
+        Updates the preconditioner matrices and the eigenbases (L, R, Q_L, Q_R in the paper).
+        """
+        if state["Q"] is not None:
+            state["exp_avg"] = self.project_back(
+                state["exp_avg"],
+                state,
+                merge_dims=merge_dims,
+                max_precond_dim=max_precond_dim,
+            )
+        if grad.dim() == 1:
+            if precondition_1d and grad.shape[0] <= max_precond_dim:
+                state["GG"][0].lerp_(
+                    grad.unsqueeze(1) @ grad.unsqueeze(0), 1 - state["shampoo_beta"]
+                )
+        else:
+            if merge_dims:
+                new_grad = self.merge_dims(grad, max_precond_dim)
+                for idx, sh in enumerate(new_grad.shape):
+                    if sh <= max_precond_dim:
+                        outer_product = torch.tensordot(
+                            new_grad,
+                            new_grad,
+                            dims=[
+                                [
+                                    *chain(
+                                        range(idx), range(idx + 1, len(new_grad.shape))
+                                    )
+                                ]
+                            ]
+                            * 2,
+                        )
+                        state["GG"][idx].lerp_(outer_product, 1 - state["shampoo_beta"])
+            else:
+                for idx, sh in enumerate(grad.shape):
+                    if sh <= max_precond_dim:
+                        outer_product = torch.tensordot(
+                            grad,
+                            grad,
+                            # Contracts across all dimensions except for k.
+                            dims=[[*chain(range(idx), range(idx + 1, len(grad.shape)))]]
+                            * 2,
+                        )
+                        state["GG"][idx].lerp_(outer_product, 1 - state["shampoo_beta"])
+
+        if state["Q"] is None:
+            state["Q"] = self.get_orthogonal_matrix(state["GG"])
+        if state["step"] > 0 and state["step"] % state["precondition_frequency"] == 0:
+            state["Q"] = self.get_orthogonal_matrix_QR(
+                state, max_precond_dim, merge_dims
+            )
+            # state["Q"] = self.get_fast_QR(state, max_precond_dim, merge_dims)
+
+        if state["step"] > 0:
+            state["exp_avg"] = self.project(
+                state["exp_avg"],
+                state,
+                merge_dims=merge_dims,
+                max_precond_dim=max_precond_dim,
+            )
+
+    def project_back(self, grad, state, merge_dims=False, max_precond_dim=10000):
+        """
+        Projects the gradient back to the original space.
+        """
+        original_shape = grad.shape
+        if merge_dims:
+            if self._data_format == "channels_last" and grad.dim() == 4:
+                permuted_shape = grad.permute(0, 3, 1, 2).shape
+            grad = self.merge_dims(grad, max_precond_dim)
+        for mat in state["Q"]:
+            if len(mat) > 0:
+                grad = torch.tensordot(
+                    grad,
+                    mat,
+                    dims=[[0], [1]],
+                )
+            else:
+                permute_order = list(range(1, len(grad.shape))) + [0]
+                grad = grad.permute(permute_order)
+
+        if merge_dims:
+            if self._data_format == "channels_last" and len(original_shape) == 4:
+                grad = grad.reshape(permuted_shape).permute(0, 2, 3, 1)
+            else:
+                grad = grad.reshape(original_shape)
+        return grad
+
+    def get_orthogonal_matrix(self, mat):
+        """
+        Computes the eigenbases of the preconditioner using torch.linalg.eigh decomposition.
+        """
+        matrix = []
+        for m in mat:
+            if len(m) == 0:
+                matrix.append([])
+                continue
+            if m.data.dtype != torch.float:
+                float_data = False
+                original_type = m.data.dtype
+                original_device = m.data.device
+                matrix.append(m.data.float())
+            else:
+                float_data = True
+                matrix.append(m.data)
+
+        final = []
+        for m in matrix:
+            if len(m) == 0:
+                final.append([])
+                continue
+            try:
+                _, Q = torch.linalg.eigh(
+                    m + 1e-30 * torch.eye(m.shape[0], device=m.device)
+                )
+            except:  # pylint: disable=bare-except # noqa: E722
+                _, Q = torch.linalg.eigh(
+                    m.to(torch.float64) + 1e-30 * torch.eye(m.shape[0], device=m.device)
+                )
+                Q = Q.to(m.dtype)
+            Q = torch.flip(Q, [1])
+
+            if not float_data:
+                Q = Q.to(original_device).type(original_type)
+            final.append(Q)
+        return final
+
+    def get_orthogonal_matrix_QR(self, state, max_precond_dim=10000, merge_dims=False):
+        """
+        Computes the eigenbases of the preconditioner using one round of power iteration
+        followed by torch.linalg.qr decomposition.
+        """
+        precond_list = state["GG"]
+        orth_list = state["Q"]
+
+        matrix = []
+        orth_matrix = []
+        for m, o in zip(precond_list, orth_list):
+            if len(m) == 0:
+                matrix.append([])
+                orth_matrix.append([])
+                continue
+            if m.data.dtype != torch.float:
+                float_data = False
+                original_type = m.data.dtype
+                original_device = m.data.device
+                matrix.append(m.data.float())
+                orth_matrix.append(o.data.float())
+            else:
+                float_data = True
+                matrix.append(m.data.float())
+                orth_matrix.append(o.data.float())
+
+        orig_shape = state["exp_avg_sq"].shape
+        if self._data_format == "channels_last" and len(orig_shape) == 4:
+            permuted_shape = state["exp_avg_sq"].permute(0, 3, 1, 2).shape
+        if merge_dims:
+            exp_avg_sq = self.merge_dims(state["exp_avg_sq"], max_precond_dim)
+        else:
+            exp_avg_sq = state["exp_avg_sq"]
+
+        final = []
+        for ind, (m, o) in enumerate(zip(matrix, orth_matrix)):
+            if len(m) == 0:
+                final.append([])
+                continue
+            est_eig = torch.diag(o.T @ m @ o)
+            sort_idx = torch.argsort(est_eig, descending=True)
+            exp_avg_sq = exp_avg_sq.index_select(ind, sort_idx)
+            o = o[:, sort_idx]
+            power_iter = m @ o
+            Q, _ = torch.linalg.qr(power_iter)
+
+            if not float_data:
+                Q = Q.to(original_device).type(original_type)
+            final.append(Q)
+
+        if merge_dims:
+            if self._data_format == "channels_last" and len(orig_shape) == 4:
+                exp_avg_sq = exp_avg_sq.reshape(permuted_shape).permute(0, 2, 3, 1)
+            else:
+                exp_avg_sq = exp_avg_sq.reshape(orig_shape)
+
+        state["exp_avg_sq"] = exp_avg_sq
+        return final

src/axolotl/utils/schemas/config.py

@@ -248,7 +248,6 @@ class AxolotlInputConfig(
     val_set_size: float | None = Field(default=0.0)
 
     sequence_parallel_degree: int | None = None
-    heads_k_stride: int | None = None
 
     special_tokens: SpecialTokensConfig | None = None
     tokens: list[str] | None = None
@@ -1109,7 +1108,7 @@ class AxolotlInputConfig(
 
     @field_validator("sequence_parallel_degree", mode="before")
     @classmethod
-    def check_sequence_parallel_degree(cls, value, info):
+    def check_sequence_parallel_config(cls, value, info):
         if not value:
             value = 1
 

src/axolotl/utils/schemas/enums.py

@@ -52,3 +52,4 @@ class CustomSupportedOptimizers(str, Enum):
     ao_adamw_fp8 = "ao_adamw_fp8"  # pylint: disable=invalid-name
     adopt_adamw = "adopt_adamw"  # pylint: disable=invalid-name
     muon = "muon"  # pylint: disable=invalid-name
+    soap = "soap"  # pylint: disable=invalid-name

tests/e2e/patched/test_sp.py

@@ -110,7 +110,7 @@ class TestRingAttention:
         mock_new_group.return_value = mock_group
 
         # Call register_ring_attn with size 4
-        register_ring_attn(sequence_parallel_degree=4, heads_k_stride=1)
+        register_ring_attn(sequence_parallel_degree=4)
 
         # Verify the number of calls without examining the arguments
         assert mock_new_group.call_count == 2

tests/e2e/test_optimizers.py

@@ -201,3 +201,46 @@ class TestCustomOptimizers(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+
+    @with_temp_dir
+    def test_soap(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM-135M",
+                "sequence_len": 1024,
+                "load_in_8bit": True,
+                "adapter": "lora",
+                "lora_r": 8,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "val_set_size": 0.1,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                },
+                "datasets": [
+                    {
+                        "path": "vicgalle/alpaca-gpt4",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "micro_batch_size": 8,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "soap",
+                "adam_beta1": 0.9,
+                "adam_beta2": 0.95,
+                "lr_scheduler": "cosine",
+            }
+        )
+
+        cfg = validate_config(cfg)
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)