from typing import List, Optional, Union, Tuple import torch from torch import Tensor from .optimizer import (Optimizer, ParamsT, _use_grad_for_differentiable, _get_value, _stack_if_compiling, _dispatch_sqrt, _default_to_fused_or_foreach, _get_scalar_dtype, _capturable_doc, _differentiable_doc, _foreach_doc, _fused_doc, _maximize_doc, _view_as_real) from torch.utils._foreach_utils import _get_fused_kernels_supported_devices __all__ = ['Adam', 'adam'] class Adam(Optimizer): def __init__(self, params: ParamsT, lr: Union[float, Tensor] = 1e-3, betas: Tuple[float, float] = (0.9, 0.999), eps: float = 1e-8, weight_decay: float = 0, amsgrad: bool = False, *, foreach: Optional[bool] = None, maximize: bool = False, capturable: bool = False, differentiable: bool = False, fused: Optional[bool] = None): if not 0.0 <= lr: raise ValueError(f"Invalid learning rate: {lr}") if isinstance(lr, Tensor) and foreach and not capturable: raise ValueError("lr as a Tensor is not supported for capturable=False and foreach=True") if not 0.0 <= eps: raise ValueError(f"Invalid epsilon value: {eps}") if not 0.0 <= betas[0] < 1.0: raise ValueError(f"Invalid beta parameter at index 0: {betas[0]}") if not 0.0 <= betas[1] < 1.0: raise ValueError(f"Invalid beta parameter at index 1: {betas[1]}") if not 0.0 <= weight_decay: raise ValueError(f"Invalid weight_decay value: {weight_decay}") defaults = dict(lr=lr, betas=betas, eps=eps, weight_decay=weight_decay, amsgrad=amsgrad, maximize=maximize, foreach=foreach, capturable=capturable, differentiable=differentiable, fused=fused) super().__init__(params, defaults) if fused: if differentiable: raise RuntimeError("`fused` does not support `differentiable`") self._step_supports_amp_scaling = True # TODO(crcrpar): [low prec params & their higher prec copy] # Support AMP with FP16/BF16 model params which would need # higher prec copy of params to do update math in higher prec to # alleviate the loss of information. fused_supported_devices = _get_fused_kernels_supported_devices() if not all( p.device.type in fused_supported_devices and torch.is_floating_point(p) for pg in self.param_groups for p in pg['params'] ): raise RuntimeError("`fused=True` requires all the params to be floating point Tensors of " f"supported devices: {fused_supported_devices}.") if foreach: raise RuntimeError("`fused` and `foreach` cannot be `True` together.") def __setstate__(self, state): super().__setstate__(state) for group in self.param_groups: group.setdefault('amsgrad', False) group.setdefault('maximize', False) group.setdefault('foreach', None) group.setdefault('capturable', False) group.setdefault('differentiable', False) fused = group.setdefault('fused', None) for p in group["params"]: p_state = self.state.get(p, []) if len(p_state) != 0 and not torch.is_tensor(p_state['step']): step_val = float(p_state["step"]) p_state["step"] = (torch.tensor(step_val, dtype=_get_scalar_dtype(is_fused=fused), device=p.device) if group['capturable'] or group['fused'] else torch.tensor(step_val, dtype=_get_scalar_dtype())) def _init_group( self, group, params_with_grad, grads, exp_avgs, exp_avg_sqs, max_exp_avg_sqs, state_steps ): has_complex = False for p in group['params']: if p.grad is not None: has_complex |= torch.is_complex(p) params_with_grad.append(p) if p.grad.is_sparse: raise RuntimeError('Adam does not support sparse gradients, please consider SparseAdam instead') grads.append(p.grad) state = self.state[p] # Lazy state initialization if len(state) == 0: # note(crcrpar): [special device hosting for step] # Deliberately host `step` on CPU if both capturable and fused are off. # This is because kernel launches are costly on CUDA and XLA. state['step'] = ( torch.zeros((), dtype=_get_scalar_dtype(is_fused=group['fused']), device=p.device) if group['capturable'] or group['fused'] else torch.tensor(0.0, dtype=_get_scalar_dtype()) ) # Exponential moving average of gradient values state['exp_avg'] = torch.zeros_like(p, memory_format=torch.preserve_format) # Exponential moving average of squared gradient values state['exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format) if group['amsgrad']: # Maintains max of all exp. moving avg. of sq. grad. values state['max_exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format) exp_avgs.append(state['exp_avg']) exp_avg_sqs.append(state['exp_avg_sq']) if group['amsgrad']: max_exp_avg_sqs.append(state['max_exp_avg_sq']) if group['differentiable'] and state['step'].requires_grad: raise RuntimeError('`requires_grad` is not supported for `step` in differentiable mode') # Foreach without capturable does not support a tensor lr if group['foreach'] and torch.is_tensor(group['lr']) and not group['capturable']: raise RuntimeError('lr as a Tensor is not supported for capturable=False and foreach=True') state_steps.append(state['step']) return has_complex @_use_grad_for_differentiable def step(self, closure=None): """Perform a single optimization step. Args: closure (Callable, optional): A closure that reevaluates the model and returns the loss. """ self._cuda_graph_capture_health_check() loss = None if closure is not None: with torch.enable_grad(): loss = closure() for group in self.param_groups: params_with_grad = [] grads = [] exp_avgs = [] exp_avg_sqs = [] max_exp_avg_sqs = [] state_steps = [] beta1, beta2 = group['betas'] has_complex = self._init_group( group, params_with_grad, grads, exp_avgs, exp_avg_sqs, max_exp_avg_sqs, state_steps) adam( params_with_grad, grads, exp_avgs, exp_avg_sqs, max_exp_avg_sqs, state_steps, amsgrad=group['amsgrad'], has_complex=has_complex, beta1=beta1, beta2=beta2, lr=group['lr'], weight_decay=group['weight_decay'], eps=group['eps'], maximize=group['maximize'], foreach=group['foreach'], capturable=group['capturable'], differentiable=group['differentiable'], fused=group['fused'], grad_scale=getattr(self, "grad_scale", None), found_inf=getattr(self, "found_inf", None), ) return loss Adam.__doc__ = r"""Implements Adam algorithm. .. math:: \begin{aligned} &\rule{110mm}{0.4pt} \\ &\textbf{input} : \gamma \text{ (lr)}, \beta_1, \beta_2 \text{ (betas)},\theta_0 \text{ (params)},f(\theta) \text{ (objective)} \\ &\hspace{13mm} \lambda \text{ (weight decay)}, \: \textit{amsgrad}, \:\textit{maximize} \\ &\textbf{initialize} : m_0 \leftarrow 0 \text{ ( first moment)}, v_0\leftarrow 0 \text{ (second moment)},\: \widehat{v_0}^{max}\leftarrow 0\\[-1.ex] &\rule{110mm}{0.4pt} \\ &\textbf{for} \: t=1 \: \textbf{to} \: \ldots \: \textbf{do} \\ &\hspace{5mm}\textbf{if} \: \textit{maximize}: \\ &\hspace{10mm}g_t \leftarrow -\nabla_{\theta} f_t (\theta_{t-1}) \\ &\hspace{5mm}\textbf{else} \\ &\hspace{10mm}g_t \leftarrow \nabla_{\theta} f_t (\theta_{t-1}) \\ &\hspace{5mm}\textbf{if} \: \lambda \neq 0 \\ &\hspace{10mm} g_t \leftarrow g_t + \lambda \theta_{t-1} \\ &\hspace{5mm}m_t \leftarrow \beta_1 m_{t-1} + (1 - \beta_1) g_t \\ &\hspace{5mm}v_t \leftarrow \beta_2 v_{t-1} + (1-\beta_2) g^2_t \\ &\hspace{5mm}\widehat{m_t} \leftarrow m_t/\big(1-\beta_1^t \big) \\ &\hspace{5mm}\widehat{v_t} \leftarrow v_t/\big(1-\beta_2^t \big) \\ &\hspace{5mm}\textbf{if} \: amsgrad \\ &\hspace{10mm}\widehat{v_t}^{max} \leftarrow \mathrm{max}(\widehat{v_t}^{max}, \widehat{v_t}) \\ &\hspace{10mm}\theta_t \leftarrow \theta_{t-1} - \gamma \widehat{m_t}/ \big(\sqrt{\widehat{v_t}^{max}} + \epsilon \big) \\ &\hspace{5mm}\textbf{else} \\ &\hspace{10mm}\theta_t \leftarrow \theta_{t-1} - \gamma \widehat{m_t}/ \big(\sqrt{\widehat{v_t}} + \epsilon \big) \\ &\rule{110mm}{0.4pt} \\[-1.ex] &\bf{return} \: \theta_t \\[-1.ex] &\rule{110mm}{0.4pt} \\[-1.ex] \end{aligned} For further details regarding the algorithm we refer to `Adam: A Method for Stochastic Optimization`_. """ + fr""" Args: params (iterable): iterable of parameters to optimize or dicts defining parameter groups lr (float, Tensor, optional): learning rate (default: 1e-3). A tensor LR is not yet supported for all our implementations. Please use a float LR if you are not also specifying fused=True or capturable=True. betas (Tuple[float, float], optional): coefficients used for computing running averages of gradient and its square (default: (0.9, 0.999)) eps (float, optional): term added to the denominator to improve numerical stability (default: 1e-8) weight_decay (float, optional): weight decay (L2 penalty) (default: 0) amsgrad (bool, optional): whether to use the AMSGrad variant of this algorithm from the paper `On the Convergence of Adam and Beyond`_ (default: False) {_foreach_doc} {_maximize_doc} {_capturable_doc} {_differentiable_doc} {_fused_doc} .. _Adam\: A Method for Stochastic Optimization: https://arxiv.org/abs/1412.6980 .. _On the Convergence of Adam and Beyond: https://openreview.net/forum?id=ryQu7f-RZ """ def adam(params: List[Tensor], grads: List[Tensor], exp_avgs: List[Tensor], exp_avg_sqs: List[Tensor], max_exp_avg_sqs: List[Tensor], state_steps: List[Tensor], # kwonly args with defaults are not supported by functions compiled with torchscript issue #70627 # setting this as kwarg for now as functional API is compiled by torch/distributed/optim foreach: Optional[bool] = None, capturable: bool = False, differentiable: bool = False, fused: Optional[bool] = None, grad_scale: Optional[Tensor] = None, found_inf: Optional[Tensor] = None, has_complex: bool = False, *, amsgrad: bool, beta1: float, beta2: float, lr: Union[float, Tensor], weight_decay: float, eps: float, maximize: bool): r"""Functional API that performs Adam algorithm computation. See :class:`~torch.optim.Adam` for details. """ # Respect when the user inputs False/True for foreach or fused. We only want to change # the default when neither have been user-specified. Note that we default to foreach # and pass False to use_fused. This is not a mistake--we want to give the fused impl # bake-in time before making it the default, even if it is typically faster. if fused is None and foreach is None: _, foreach = _default_to_fused_or_foreach(params, differentiable, use_fused=False) # Do not flip on foreach for the unsupported case where lr is a Tensor and capturable=False. if foreach and isinstance(lr, Tensor) and not capturable: foreach = False if fused is None: fused = False if foreach is None: foreach = False # this check is slow during compilation, so we skip it # if it's strictly needed we can add this check back in dynamo if not torch._utils.is_compiling() and not all(isinstance(t, torch.Tensor) for t in state_steps): raise RuntimeError("API has changed, `state_steps` argument must contain a list of singleton tensors") if foreach and torch.jit.is_scripting(): raise RuntimeError('torch.jit.script not supported with foreach optimizers') if fused and torch.jit.is_scripting(): raise RuntimeError("torch.jit.script not supported with fused optimizers") if fused and not torch.jit.is_scripting(): func = _fused_adam elif foreach and not torch.jit.is_scripting(): func = _multi_tensor_adam else: func = _single_tensor_adam func(params, grads, exp_avgs, exp_avg_sqs, max_exp_avg_sqs, state_steps, amsgrad=amsgrad, has_complex=has_complex, beta1=beta1, beta2=beta2, lr=lr, weight_decay=weight_decay, eps=eps, maximize=maximize, capturable=capturable, differentiable=differentiable, grad_scale=grad_scale, found_inf=found_inf) def _single_tensor_adam(params: List[Tensor], grads: List[Tensor], exp_avgs: List[Tensor], exp_avg_sqs: List[Tensor], max_exp_avg_sqs: List[Tensor], state_steps: List[Tensor], grad_scale: Optional[Tensor], found_inf: Optional[Tensor], *, amsgrad: bool, has_complex: bool, beta1: float, beta2: float, lr: Union[float, Tensor], weight_decay: float, eps: float, maximize: bool, capturable: bool, differentiable: bool): assert grad_scale is None and found_inf is None if torch.jit.is_scripting(): # this assert is due to JIT being dumb and not realizing that the ops below # have overloads to handle both float and Tensor lrs, so we just assert it's # a float since most people using JIT are using floats assert isinstance(lr, float) for i, param in enumerate(params): grad = grads[i] if not maximize else -grads[i] exp_avg = exp_avgs[i] exp_avg_sq = exp_avg_sqs[i] step_t = state_steps[i] # If compiling, the compiler will handle cudagraph checks, see note [torch.compile x capturable] if not torch._utils.is_compiling() and capturable: assert ( (param.is_cuda and step_t.is_cuda) or (param.is_xla and step_t.is_xla) ), "If capturable=True, params and state_steps must be CUDA or XLA tensors." # update step step_t += 1 if weight_decay != 0: grad = grad.add(param, alpha=weight_decay) if torch.is_complex(param): grad = torch.view_as_real(grad) exp_avg = torch.view_as_real(exp_avg) exp_avg_sq = torch.view_as_real(exp_avg_sq) if amsgrad: max_exp_avg_sqs[i] = torch.view_as_real(max_exp_avg_sqs[i]) param = torch.view_as_real(param) # Decay the first and second moment running average coefficient exp_avg.lerp_(grad, 1 - beta1) exp_avg_sq.mul_(beta2).addcmul_(grad, grad.conj(), value=1 - beta2) if capturable or differentiable: step = step_t bias_correction1 = 1 - beta1 ** step bias_correction2 = 1 - beta2 ** step step_size = lr / bias_correction1 step_size_neg = step_size.neg() bias_correction2_sqrt = bias_correction2.sqrt() if amsgrad: # Maintains the maximum of all 2nd moment running avg. till now if differentiable: max_exp_avg_sq = max_exp_avg_sqs[i].clone() else: max_exp_avg_sq = max_exp_avg_sqs[i] max_exp_avg_sqs[i].copy_(torch.maximum(max_exp_avg_sq, exp_avg_sq)) # Uses the max. for normalizing running avg. of gradient # Folds in (admittedly ugly) 1-elem step_size math here to avoid extra param-set-sized read+write # (can't fold it into addcdiv_ below because addcdiv_ requires value is a Number, not a Tensor) denom = (max_exp_avg_sqs[i].sqrt() / (bias_correction2_sqrt * step_size_neg)).add_(eps / step_size_neg) else: denom = (exp_avg_sq.sqrt() / (bias_correction2_sqrt * step_size_neg)).add_(eps / step_size_neg) param.addcdiv_(exp_avg, denom) else: step = _get_value(step_t) bias_correction1 = 1 - beta1 ** step bias_correction2 = 1 - beta2 ** step step_size = lr / bias_correction1 bias_correction2_sqrt = _dispatch_sqrt(bias_correction2) if amsgrad: # Maintains the maximum of all 2nd moment running avg. till now torch.maximum(max_exp_avg_sqs[i], exp_avg_sq, out=max_exp_avg_sqs[i]) # Use the max. for normalizing running avg. of gradient denom = (max_exp_avg_sqs[i].sqrt() / bias_correction2_sqrt).add_(eps) else: denom = (exp_avg_sq.sqrt() / bias_correction2_sqrt).add_(eps) param.addcdiv_(exp_avg, denom, value=-step_size) # Lastly, switch back to complex view if amsgrad and torch.is_complex(params[i]): max_exp_avg_sqs[i] = torch.view_as_complex(max_exp_avg_sqs[i]) def _multi_tensor_adam(params: List[Tensor], grads: List[Tensor], exp_avgs: List[Tensor], exp_avg_sqs: List[Tensor], max_exp_avg_sqs: List[Tensor], state_steps: List[Tensor], grad_scale: Optional[Tensor], found_inf: Optional[Tensor], *, amsgrad: bool, has_complex: bool, beta1: float, beta2: float, lr: Union[float, Tensor], weight_decay: float, eps: float, maximize: bool, capturable: bool, differentiable: bool): if len(params) == 0: return if isinstance(lr, Tensor) and not capturable: raise RuntimeError("lr as a Tensor is not supported for capturable=False and foreach=True") # If compiling, the compiler will handle cudagraph checks, see note [torch.compile x capturable] if not torch._utils.is_compiling() and capturable: assert all(p.is_cuda and step.is_cuda for p, step in zip(params, state_steps)), \ "If capturable=True, params and state_steps must be CUDA tensors." assert grad_scale is None and found_inf is None assert not differentiable, "_foreach ops don't support autograd" grouped_tensors = Optimizer._group_tensors_by_device_and_dtype( [params, grads, exp_avgs, exp_avg_sqs, max_exp_avg_sqs, state_steps]) for (( device_params, device_grads, device_exp_avgs, device_exp_avg_sqs, device_max_exp_avg_sqs, device_state_steps, ), _) in grouped_tensors.values(): # Handle complex parameters if has_complex: if amsgrad: _view_as_real(device_params, device_grads, device_exp_avgs, device_exp_avg_sqs, device_max_exp_avg_sqs) else: _view_as_real(device_params, device_grads, device_exp_avgs, device_exp_avg_sqs) if maximize: device_grads = torch._foreach_neg(device_grads) # Update steps # If steps are on CPU, foreach will fall back to the slow path, which is a for-loop calling t.add(1) over # and over. 1 will then be wrapped into a Tensor over and over again, which is slower than if we just # wrapped it once now. The alpha is required to assure we go to the right overload. if device_state_steps[0].is_cpu: torch._foreach_add_(device_state_steps, torch.tensor(1.0, device='cpu'), alpha=1.0) else: torch._foreach_add_(device_state_steps, 1) if weight_decay != 0: # Re-use the intermediate memory (device_grads) already allocated for maximize if maximize: torch._foreach_add_(device_grads, device_params, alpha=weight_decay) else: device_grads = torch._foreach_add(device_grads, device_params, alpha=weight_decay) # Decay the first and second moment running average coefficient torch._foreach_lerp_(device_exp_avgs, device_grads, 1 - beta1) torch._foreach_mul_(device_exp_avg_sqs, beta2) torch._foreach_addcmul_(device_exp_avg_sqs, device_grads, device_grads, 1 - beta2) # Delete the local intermediate since it won't be used anymore to save on peak memory del device_grads if capturable: bias_correction1 = torch._foreach_pow(beta1, device_state_steps) bias_correction2 = torch._foreach_pow(beta2, device_state_steps) # foreach_sub doesn't allow a scalar as the first arg torch._foreach_sub_(bias_correction1, 1) torch._foreach_sub_(bias_correction2, 1) # we do not negate bias_correction1 as it'll need to be negated later anyway torch._foreach_neg_(bias_correction2) # foreach_div doesn't allow a scalar as the first arg torch._foreach_div_(bias_correction1, lr) torch._foreach_reciprocal_(bias_correction1) torch._foreach_sqrt_(bias_correction2) # Re-assign for clarity as we maintain minimal intermediates: we'll have # step_size = - lr / (1 - beta1 ^ t) where t = num_steps # bias_correction2_sqrt = sqrt(1 - beta2 ^ t) step_size = bias_correction1 bias_correction2_sqrt = bias_correction2 if amsgrad: # Maintains the maximum of all 2nd moment running avg. till now torch._foreach_maximum_(device_max_exp_avg_sqs, device_exp_avg_sqs) # type: ignore[assignment] # Set intermediate to the max. for normalizing running avg. of gradient when amsgrad exp_avg_sq_sqrt = torch._foreach_sqrt(device_max_exp_avg_sqs) else: exp_avg_sq_sqrt = torch._foreach_sqrt(device_exp_avg_sqs) torch._foreach_div_(exp_avg_sq_sqrt, bias_correction2_sqrt) torch._foreach_add_(exp_avg_sq_sqrt, eps) torch._foreach_div_(exp_avg_sq_sqrt, step_size) # at this point, exp_avg_sq_sqrt = - (1 - beta^t) * [sqrt(exp_avg_sq / (1 - beta2^t)) + eps] / lr torch._foreach_addcdiv_(device_params, device_exp_avgs, exp_avg_sq_sqrt) else: bias_correction1 = [1 - beta1 ** _get_value(step) for step in device_state_steps] bias_correction2 = [1 - beta2 ** _get_value(step) for step in device_state_steps] step_size = _stack_if_compiling([(lr / bc) * -1 for bc in bias_correction1]) bias_correction2_sqrt = [_dispatch_sqrt(bc) for bc in bias_correction2] if amsgrad: # Maintains the maximum of all 2nd moment running avg. till now torch._foreach_maximum_(device_max_exp_avg_sqs, device_exp_avg_sqs) # Use the max. for normalizing running avg. of gradient exp_avg_sq_sqrt = torch._foreach_sqrt(device_max_exp_avg_sqs) else: exp_avg_sq_sqrt = torch._foreach_sqrt(device_exp_avg_sqs) torch._foreach_div_(exp_avg_sq_sqrt, bias_correction2_sqrt) torch._foreach_add_(exp_avg_sq_sqrt, eps) torch._foreach_addcdiv_(device_params, device_exp_avgs, exp_avg_sq_sqrt, step_size) def _fused_adam( params: List[Tensor], grads: List[Tensor], exp_avgs: List[Tensor], exp_avg_sqs: List[Tensor], max_exp_avg_sqs: List[Tensor], state_steps: List[Tensor], grad_scale: Optional[Tensor], found_inf: Optional[Tensor], *, amsgrad: bool, has_complex: bool, # Needed for consistency. beta1: float, beta2: float, lr: Union[float, Tensor], weight_decay: float, eps: float, maximize: bool, capturable: bool, # Needed for consistency. differentiable: bool, ) -> None: if not params: return if differentiable: raise RuntimeError("Adam with fused=True does not support differentiable=True") grad_scale_dict = {grad_scale.device: grad_scale} if grad_scale is not None else None found_inf_dict = {found_inf.device: found_inf} if found_inf is not None else None # We only shuffle around the lr when it is a Tensor and on CUDA, otherwise, we prefer # treating it as a scalar. lr_dict = {lr.device: lr} if isinstance(lr, Tensor) and str(lr.device) != "cpu" else None grouped_tensors = Optimizer._group_tensors_by_device_and_dtype( [params, grads, exp_avgs, exp_avg_sqs, max_exp_avg_sqs, state_steps]) for (device, _), ((device_params, device_grads, device_exp_avgs, device_exp_avg_sqs, device_max_exp_avg_sqs, device_state_steps,), _) in grouped_tensors.items(): device_grad_scale, device_found_inf = None, None if grad_scale is not None: if device not in grad_scale_dict: grad_scale_dict[device] = grad_scale.to(device, non_blocking=True) device_grad_scale = grad_scale_dict[device] if found_inf is not None: if found_inf not in found_inf_dict: found_inf_dict[device] = found_inf.to(device, non_blocking=True) device_found_inf = found_inf_dict[device] if lr_dict is not None and device not in lr_dict: lr_dict[device] = lr.to(device=device, non_blocking=True) lr = lr_dict[device] torch._foreach_add_(device_state_steps, 1) torch._fused_adam_( device_params, device_grads, device_exp_avgs, device_exp_avg_sqs, device_max_exp_avg_sqs, device_state_steps, amsgrad=amsgrad, lr=lr, beta1=beta1, beta2=beta2, weight_decay=weight_decay, eps=eps, maximize=maximize, grad_scale=device_grad_scale, found_inf=device_found_inf, ) if device_found_inf is not None: torch._foreach_sub_(device_state_steps, [device_found_inf] * len(device_state_steps))