lamb.py

# Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
# Lamb optimizer directly copied from https://github.com/facebookresearch/online-dt
from __future__ import annotations

import math

import torch
from torch.optim import Optimizer


class Lamb(Optimizer):
    """Implements a pure pytorch variant of FuseLAMB (NvLamb variant) optimizer from apex.optimizers.FusedLAMB
    reference: https://github.com/NVIDIA/DeepLearningExamples/blob/master/PyTorch/LanguageModeling/Transformer-XL/pytorch/lamb.py
    LAMB was proposed in `Large Batch Optimization for Deep Learning: Training BERT in 76 minutes`_.
    Arguments:
        params (iterable): iterable of parameters to optimize or dicts defining parameter groups.
        lr (:obj:`float`, optional): learning rate. (default: 1e-3)
        betas (Tuple[float, float], optional): coefficients used for computing
            running averages of gradient and its norm. (default: (0.9, 0.999))
        eps (:obj:`float`, optional): term added to the denominator to improve
            numerical stability. (default: 1e-8)
        weight_decay (:obj:`float`, optional): weight decay (L2 penalty) (default: 0)
        grad_averaging (bool, optional): whether apply (1-beta2) to grad when
            calculating running averages of gradient. (default: True)
        max_grad_norm (:obj:`float`, optional): value used to clip global grad norm (default: 1.0)
        trust_clip (bool): enable LAMBC trust ratio clipping (default: False)
        always_adapt (boolean, optional): Apply adaptive learning rate to 0.0
            weight decay parameter (default: False)
    .. _Large Batch Optimization for Deep Learning - Training BERT in 76 minutes:
        https://arxiv.org/abs/1904.00962
    .. _On the Convergence of Adam and Beyond:
        https://openreview.net/forum?id=ryQu7f-RZ
    """

    def __init__(
        self,
        params,
        lr=1e-3,
        bias_correction=True,
        betas=(0.9, 0.999),
        eps=1e-6,
        weight_decay=0.01,
        grad_averaging=True,
        max_grad_norm=1.0,
        trust_clip=False,
        always_adapt=False,
    ):
        defaults = {
            "lr": lr,
            "bias_correction": bias_correction,
            "betas": betas,
            "eps": eps,
            "weight_decay": weight_decay,
            "grad_averaging": grad_averaging,
            "max_grad_norm": max_grad_norm,
            "trust_clip": trust_clip,
            "always_adapt": always_adapt,
        }
        super().__init__(params, defaults)

    @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.
        """
        loss = None
        if closure is not None:
            with torch.enable_grad():
                loss = closure()

        device = self.param_groups[0]["params"][0].device
        one_tensor = torch.tensor(
            1.0, device=device
        )  # because torch.where doesn't handle scalars correctly
        global_grad_norm = torch.zeros(1, device=device)
        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(
                        "Lamb does not support sparse gradients, consider SparseAdam instad."
                    )
                global_grad_norm.add_(grad.pow(2).sum())

        global_grad_norm = torch.sqrt(global_grad_norm)
        # FIXME it'd be nice to remove explicit tensor conversion of scalars when torch.where promotes
        # scalar types properly https://github.com/pytorch/pytorch/issues/9190
        max_grad_norm = torch.tensor(self.defaults["max_grad_norm"], device=device)
        clip_global_grad_norm = torch.where(
            global_grad_norm > max_grad_norm,
            global_grad_norm / max_grad_norm,
            one_tensor,
        )

        for group in self.param_groups:
            bias_correction = 1 if group["bias_correction"] else 0
            beta1, beta2 = group["betas"]
            grad_averaging = 1 if group["grad_averaging"] else 0
            beta3 = 1 - beta1 if grad_averaging else 1.0

            # assume same step across group now to simplify things
            # per parameter step can be easily support by making it tensor, or pass list into kernel
            if "step" in group:
                group["step"] += 1
            else:
                group["step"] = 1

            if bias_correction:
                bias_correction1 = 1 - beta1 ** group["step"]
                bias_correction2 = 1 - beta2 ** group["step"]
            else:
                bias_correction1, bias_correction2 = 1.0, 1.0

            for p in group["params"]:
                if p.grad is None:
                    continue
                grad = p.grad.div_(clip_global_grad_norm)
                state = self.state[p]

                # State initialization
                if len(state) == 0:
                    # Exponential moving average of gradient valuesa
                    state["exp_avg"] = torch.zeros_like(p)
                    # Exponential moving average of squared gradient values
                    state["exp_avg_sq"] = torch.zeros_like(p)

                exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"]

                # Decay the first and second moment running average coefficient
                exp_avg.mul_(beta1).add_(grad, alpha=beta3)  # m_t
                exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)  # v_t

                denom = (exp_avg_sq.sqrt() / math.sqrt(bias_correction2)).add_(
                    group["eps"]
                )
                update = (exp_avg / bias_correction1).div_(denom)

                weight_decay = group["weight_decay"]
                if weight_decay != 0:
                    update.add_(p, alpha=weight_decay)

                if weight_decay != 0 or group["always_adapt"]:
                    # Layer-wise LR adaptation. By default, skip adaptation on parameters that are
                    # excluded from weight decay, unless always_adapt == True, then always enabled.
                    w_norm = p.norm(2.0)
                    g_norm = update.norm(2.0)
                    # FIXME nested where required since logical and/or not working in PT XLA
                    trust_ratio = torch.where(
                        w_norm > 0,
                        torch.where(g_norm > 0, w_norm / g_norm, one_tensor),
                        one_tensor,
                    )
                    if group["trust_clip"]:
                        # LAMBC trust clipping, upper bound fixed at one
                        trust_ratio = torch.minimum(trust_ratio, one_tensor)
                    update.mul_(trust_ratio)

                p.add_(update, alpha=-group["lr"])

        return loss