"""contains the Cerebras Adafactor implementation"""
# coding=utf-8
#
# This code is adapted from
# https://github.com/huggingface/transformers/blob/master/src/transformers/optimization.py
#
# Copyright 2016-2023 Cerebras Systems
# SPDX-License-Identifier: Apache-2.0
#
# Copyright 2018 The Google AI Language Team Authors and The HuggingFace Inc. team.
# Modifications Copyright 2023 Cerebras.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import torch
from torch.optim import Optimizer
import cerebras.pytorch as cstorch
from .optimizer import Optimizer
[docs]class Adafactor(Optimizer):
"""
Adafactor optimizer implemented to conform to execution within the
constraints of the Cerebras WSE.
"""
[docs] def __init__(
self,
params,
lr,
eps=(1e-30, 1e-3),
clip_threshold=1.0,
decay_rate=-0.8,
beta1=None,
weight_decay=0.0,
scale_parameter=True,
relative_step=False,
warmup_init=False,
):
if lr is not None and relative_step:
raise ValueError(
"Cannot combine manual `lr` and `relative_step=True` options"
)
if warmup_init and not relative_step:
raise ValueError("`warmup_init=True` is not supported yet")
if clip_threshold != 1.0:
raise ValueError(
f"Only `clip_threshold=1.0` is supported now. "
f"It was set to {clip_threshold}."
)
if beta1 is not None:
raise ValueError(
f"Only `beta1=None` is supported now. It was set to {beta1}."
)
if relative_step:
raise ValueError("`relative_step=True` is not supported yet")
defaults = dict(
lr=lr,
eps=eps,
clip_threshold=clip_threshold,
decay_rate=decay_rate,
beta1=beta1,
weight_decay=weight_decay,
scale_parameter=scale_parameter,
relative_step=relative_step,
warmup_init=warmup_init,
)
super().__init__(params, defaults, enable_global_step=True)
[docs] def preinitialize(self):
"""
Allocates tensors for the optimizer state to allow direct compilation
of the model before the first step.
"""
for group in self.param_groups:
for p in group["params"]:
state = self.state[p]
grad_shape = p.shape
factored = len(grad_shape) >= 2
use_first_moment = group["beta1"] is not None
if use_first_moment:
state["exp_avg"] = cstorch.zeros_like(p)
if factored:
state["exp_avg_sq_row"] = cstorch.zeros(grad_shape[:-1])
state["exp_avg_sq_col"] = cstorch.zeros(
grad_shape[:-2] + grad_shape[-1:]
)
else:
state["exp_avg_sq"] = cstorch.zeros_like(p)
@staticmethod
def _get_lr(param_group, rms):
rel_step_sz = param_group["lr"]
param_scale = 1.0
if param_group["scale_parameter"]:
eps = param_group["eps"][1]
if not isinstance(eps, torch.Tensor):
eps = torch.tensor(eps)
param_scale = torch.maximum(rms, eps)
return param_scale * rel_step_sz
@staticmethod
def _rms(tensor):
return tensor.square().mean().sqrt()
@staticmethod
def _approx_sq_grad(exp_avg_sq_row, exp_avg_sq_col):
r_factor = (
(exp_avg_sq_row / exp_avg_sq_row.mean(dim=-1, keepdim=True))
.rsqrt()
.unsqueeze(-1)
)
c_factor = exp_avg_sq_col.rsqrt().unsqueeze(-2)
return torch.mul(r_factor, c_factor)
@torch.no_grad()
def step(self, closure=None):
"""
Performs a single optimization step.
Arguments:
closure (:obj:`Callable`, `optional`): A closure that reevaluates
the model and returns the loss.
"""
loss = None
if closure is not None:
loss = closure()
for group in self.param_groups:
for p in group["params"]:
if p.grad is None:
continue
grad = p.grad
if grad.is_sparse:
raise RuntimeError(
"Adafactor does not support sparse gradients."
)
state = self.state[p]
factored = "exp_avg_sq_row" in state
use_first_moment = "exp_avg" in state
global_step_fp32 = self.increment_global_step(p)
lr = self._get_lr(group, self._rms(p))
decay_rate = group["decay_rate"]
if not isinstance(decay_rate, torch.Tensor):
decay_rate = torch.tensor(decay_rate)
beta2t = 1.0 - torch.pow(
global_step_fp32, decay_rate.to(p.device)
)
update = (grad**2) + group["eps"][0]
if factored:
exp_avg_sq_row = state["exp_avg_sq_row"]
exp_avg_sq_col = state["exp_avg_sq_col"]
exp_avg_sq_row.mul_(beta2t).add_(
update.mean(dim=-1).mul(1.0 - beta2t)
)
exp_avg_sq_col.mul_(beta2t).add_(
update.mean(dim=-2).mul(1.0 - beta2t)
)
# Approximation of exponential moving average of square of gradient
update = self._approx_sq_grad(
exp_avg_sq_row, exp_avg_sq_col
)
update.mul_(grad)
else:
exp_avg_sq = state["exp_avg_sq"]
exp_avg_sq.mul_(beta2t).add_(update.mul(1.0 - beta2t))
update = exp_avg_sq.rsqrt().mul_(grad)
update.div_(
torch.maximum(
self._rms(update) / group["clip_threshold"],
torch.tensor(1.0, dtype=torch.float32, device=p.device),
)
)
update.mul_(lr)
if use_first_moment:
exp_avg = state["exp_avg"]
exp_avg.mul_(group["beta1"]).add_(
update.mul(1 - group["beta1"])
)
update = exp_avg
if group["weight_decay"] > 0.0:
p.sub_(p.mul(group["weight_decay"] * lr))
p.sub_(update)
return loss