import functools import numbers import operator import sys from enum import Enum from functools import partial, reduce from itertools import chain, product from typing import Any, Callable, cast, Iterable, List, Optional, Tuple, Union import torch import torch._prims as prims import torch._prims_common as utils import torch.nn.functional as F from torch import sym_float, sym_int, Tensor from torch._decomp import register_decomposition from torch._higher_order_ops.out_dtype import out_dtype from torch._prims_common import IntLike, NumberType, TensorLike, TensorSequenceType from torch._prims_common.wrappers import ( _maybe_convert_to_dtype, _maybe_resize_out, _safe_copy_out, out_wrapper, ) from torch.utils import _pytree as pytree from torch.utils._pytree import tree_map DispatchKey = torch._C.DispatchKey # type: ignore[attr-defined] # None of these functions are publicly accessible; get at them # from torch._decomps __all__: List[str] = [] aten = torch._ops.ops.aten class Reduction(Enum): NONE = 0 MEAN = 1 SUM = 2 # This wraps a decomposition and performs various type promotion logic within it, depending on the strategy provided # We're currently re-using ELEMENTWISE_TYPE_PROMOTION_KIND, although some of the usages are on non-elementwise ops # Will need to validate the non-elementwise uses def type_casts( f: Callable, type_promotion: utils.ELEMENTWISE_TYPE_PROMOTION_KIND, compute_dtype_only: bool = False, ): @functools.wraps(f) def inner(*args, **kwargs): flat_args = [ x for x in pytree.arg_tree_leaves(*args, **kwargs) if isinstance(x, Tensor) ] computation_dtype, result_dtype = utils.elementwise_dtypes( *flat_args, type_promotion_kind=type_promotion ) # TODO: pretty sure this is not quite right def increase_prec(x): if isinstance(x, Tensor): return x.to(computation_dtype) else: return x def decrease_prec(x): if isinstance(x, Tensor): return x.to(result_dtype) else: return x r = f(*tree_map(increase_prec, args), **tree_map(increase_prec, kwargs)) if compute_dtype_only: return r else: return tree_map(decrease_prec, r) return inner compute_only_pw_cast_for_opmath = partial( type_casts, type_promotion=utils.ELEMENTWISE_TYPE_PROMOTION_KIND.DEFAULT, compute_dtype_only=True, ) pw_cast_for_opmath = partial( type_casts, type_promotion=utils.ELEMENTWISE_TYPE_PROMOTION_KIND.DEFAULT ) pw_cast_for_int_to_real = partial( type_casts, type_promotion=utils.ELEMENTWISE_TYPE_PROMOTION_KIND.INT_TO_FLOAT ) # This expands x until x.dim() == dim. Might be useful as an operator def _unsqueeze_to_dim(x: Tensor, dim: int) -> Tensor: for _ in range(dim - x.dim()): x = x.unsqueeze(-1) return x @register_decomposition(aten.tanh_backward) @out_wrapper("grad_input") @pw_cast_for_opmath def tanh_backward(out_grad: Tensor, y: Tensor): return out_grad * (1 - y * y).conj_physical() @register_decomposition(aten.sigmoid_backward) @out_wrapper("grad_input") @pw_cast_for_opmath def sigmoid_backward(out_grad: Tensor, y: Tensor): return out_grad * (y * (1 - y)).conj_physical() @register_decomposition(aten.softplus_backward) @out_wrapper("grad_input") @pw_cast_for_opmath def softplus_backward(out_grad: Tensor, x: Tensor, beta: float, threshold: float): z = (x * beta).exp() return torch.where((x * beta) > threshold, out_grad, out_grad * z / (z + 1.0)) @register_decomposition(aten.elu_backward) @out_wrapper("grad_input") @pw_cast_for_opmath def elu_backward( grad_output: Tensor, alpha: float, scale: float, input_scale: float, is_result: bool, self_or_result: Tensor, ): negcoef = alpha * scale poscoef = scale negiptcoef = input_scale if is_result: return torch.where( self_or_result <= 0, grad_output * negiptcoef * (self_or_result + negcoef), grad_output * poscoef, ) else: return torch.where( self_or_result <= 0, grad_output * negiptcoef * negcoef * torch.exp(self_or_result * negiptcoef), grad_output * poscoef, ) @register_decomposition([aten.fill.Scalar]) def fill_scalar(self, value): return torch.full_like(self, value) @register_decomposition([aten.fill.Tensor]) def fill_tensor(self, value: Tensor): torch._check( value.dim() == 0, lambda: f"fill only supports 0-dimension value tensor but got tensor with {value.dim()} dimensions", ) return aten.copy(self, value) @register_decomposition(aten.hardsigmoid) @out_wrapper() @pw_cast_for_opmath def hardsigmoid(self: Tensor) -> Tensor: return torch.clamp(torch.clamp(self + 3, min=0), max=6) / 6 @register_decomposition(aten.hardsigmoid_backward) @out_wrapper("grad_input") @pw_cast_for_opmath def hardsigmoid_backward(grad_output: Tensor, self: Tensor): return torch.where( (self > -3.0) & (self < 3.0), grad_output * (1.0 / 6.0), 0.0, ) @register_decomposition(aten.hardtanh_backward) @out_wrapper("grad_input") def hardtanh_backward( grad_output: Tensor, self: Tensor, min_val: float, max_val: float ): return torch.where((self <= min_val) | (self >= max_val), 0.0, grad_output) @register_decomposition(aten.hardswish) @out_wrapper() @pw_cast_for_opmath def hardswish(self: Tensor) -> Tensor: return self * torch.clamp(torch.clamp(self + 3, min=0), max=6) / 6 @register_decomposition(aten.hardswish_backward) @out_wrapper() @pw_cast_for_opmath def hardswish_backward(grad_output: Tensor, self: Tensor) -> Tensor: return torch.where( self < -3, 0.0, torch.where(self <= 3, grad_output * ((self / 3) + 0.5), grad_output), ) @register_decomposition(aten.threshold_backward) @out_wrapper("grad_input") def threshold_backward(grad_output: Tensor, self: Tensor, threshold: float): return torch.where(self <= threshold, 0, grad_output) @register_decomposition(aten.leaky_relu_backward) @out_wrapper("grad_input") @pw_cast_for_opmath def leaky_relu_backward( grad_output: Tensor, self: Tensor, negative_slope: float, self_is_result: bool ): return torch.where(self > 0, grad_output, grad_output * negative_slope) @register_decomposition(aten.gelu_backward) @out_wrapper("grad_input") @pw_cast_for_opmath def gelu_backward(grad: Tensor, self: Tensor, approximate: str = "none"): M_SQRT2 = 1.41421356237309504880 M_SQRT1_2 = 0.70710678118654752440 M_2_SQRTPI = 1.12837916709551257390 if approximate == "tanh": kBeta = M_SQRT2 * M_2_SQRTPI * 0.5 kKappa = 0.044715 x_sq = self * self x_cube = x_sq * self inner = kBeta * (self + kKappa * x_cube) tanh_inner = torch.tanh(inner) left = 0.5 * self right = 1 + tanh_inner left_derivative = 0.5 * right tanh_derivative = 1 - tanh_inner * tanh_inner inner_derivative = kBeta * (1 + 3 * kKappa * x_sq) right_derivative = left * tanh_derivative * inner_derivative return grad * (left_derivative + right_derivative) else: kAlpha = M_SQRT1_2 kBeta = M_2_SQRTPI * M_SQRT1_2 * 0.5 cdf = 0.5 * (1 + torch.erf(self * kAlpha)) pdf = kBeta * torch.exp(self * self * -0.5) return grad * (cdf + self * pdf) @register_decomposition(aten.mish_backward) @pw_cast_for_opmath def mish_backward(grad_output: Tensor, input: Tensor): input_tanh_softplus = torch.tanh(F.softplus(input)) input_sigmoid = torch.sigmoid(input) out = input * input_sigmoid * (1 - input_tanh_softplus * input_tanh_softplus) return grad_output * (input_tanh_softplus + out) @register_decomposition(aten.silu) @out_wrapper() @pw_cast_for_opmath def silu(self: Tensor) -> Tensor: return self * torch.sigmoid(self) @register_decomposition(aten.silu_backward) @out_wrapper("grad_input") @pw_cast_for_opmath def silu_backward(grad_output: Tensor, self: Tensor) -> Tensor: sigmoid = 1 / (1 + torch.exp(-self)) return grad_output * sigmoid * (1 + self * (1 - sigmoid)) @register_decomposition(aten._prelu_kernel) def _prelu_kernel(self: Tensor, weight: Tensor) -> Tensor: return torch.where(self > 0, self, weight * self) @register_decomposition(aten._prelu_kernel_backward) def _prelu_kernel_backward( grad_output: Tensor, self: Tensor, weight: Tensor, ) -> Tuple[Tensor, Tensor]: input_grad = torch.where(self > 0, grad_output, weight * grad_output) weight_grad = torch.where(self > 0, 0.0, self * grad_output) return (input_grad, weight_grad) @register_decomposition(aten.rrelu_with_noise) @aten.rrelu_with_noise.default.py_impl(DispatchKey.AutogradCUDA) @out_wrapper() @pw_cast_for_opmath def rrelu_with_noise( self: Tensor, noise: Tensor, lower: float = 0.125, upper: float = 0.3333333333333333, training: bool = False, generator: Optional[torch.Generator] = None, ) -> Tensor: assert generator is None if training: not_positive = self <= 0 r = aten.uniform(self, lower, upper) output = torch.where(not_positive, self * r, self) noise.copy_(torch.where(not_positive, r, 1)) return output else: negative_slope = (lower + upper) / 2 return aten.leaky_relu(self, negative_slope) @register_decomposition(aten.rrelu_with_noise_) @aten.rrelu_with_noise_.default.py_impl(DispatchKey.AutogradCUDA) @pw_cast_for_opmath def rrelu_with_noise_( self: Tensor, noise: Tensor, lower: float, upper: float, training: bool = False, generator: Optional[torch.Generator] = None, ) -> Tensor: return self.copy_(rrelu_with_noise(self, noise, lower, upper, training, generator)) @register_decomposition(aten.rrelu_with_noise_backward) @out_wrapper() @pw_cast_for_opmath def rrelu_with_noise_backward( grad_output: Tensor, self: Tensor, noise: Tensor, lower: float, upper: float, training: bool, self_is_result: bool, ) -> Tensor: if training and upper - lower > 1e-6: return grad_output.mul(noise) else: negative_slope = (lower + upper) / 2 return aten.leaky_relu_backward( grad_output, self, negative_slope, self_is_result ) @register_decomposition(aten.log_sigmoid_backward) @out_wrapper("grad_input") @pw_cast_for_opmath def log_sigmoid_backward(grad_output: Tensor, self: Tensor, buffer: Tensor) -> Tensor: in_negative = self < 0 max_deriv = torch.where(in_negative, 1, 0) sign = torch.where(in_negative, 1, -1) z = torch.exp(-torch.abs(self)) return grad_output * (max_deriv - sign * (z / (1 + z))) # CPU has a special formula that uses buffer, but disabled for convenience sake # return (max_deriv - sign * (buffer / (1 + buffer))) * grad_output def apply_loss_reduction(loss: Tensor, reduction: int): if reduction == Reduction.MEAN.value: return torch.mean(loss) elif reduction == Reduction.SUM.value: return torch.sum(loss) else: return loss def to_real_dtype(dtype: torch.dtype): if dtype == torch.complex32: return torch.float16 elif dtype == torch.complex64: return torch.float32 elif dtype == torch.complex128: return torch.float64 # TODO: None of these loss castings are quite correct, see # https://github.com/pytorch/pytorch/issues/76870. Also, the ATen kernels # perform the pointwise portion in opmath, but don't maintain it between the # pointwise portion and the reduction @register_decomposition(aten.mse_loss) @out_wrapper() @pw_cast_for_opmath def mse_loss( self: Tensor, target: Tensor, reduction: int = Reduction.MEAN.value ) -> Tensor: loss = (self - target) ** 2 return apply_loss_reduction(loss, reduction) @register_decomposition(aten.mse_loss_backward) @out_wrapper("grad_input") @pw_cast_for_opmath def mse_loss_backward( grad_output: Tensor, input: Tensor, target: Tensor, reduction: int ): norm = 2.0 / input.numel() if reduction == Reduction.MEAN.value else 2.0 return norm * (input - target) * grad_output @register_decomposition(aten.smooth_l1_loss) @out_wrapper() @pw_cast_for_opmath def smooth_l1_loss( self: Tensor, target: Tensor, reduction: int = Reduction.MEAN.value, beta: float = 1.0, ): loss = (self - target).abs() loss = torch.where(loss < beta, 0.5 * loss**2 / beta, loss - 0.5 * beta) return apply_loss_reduction(loss, reduction) @register_decomposition(aten.smooth_l1_loss_backward.default) @pw_cast_for_opmath def smooth_l1_loss_backward( grad_output: Tensor, self: Tensor, target: Tensor, reduction: int, beta: float ): norm = 1.0 / self.numel() if reduction == Reduction.MEAN.value else 1.0 x = self - target abs_x = torch.abs(x) norm_grad = norm * grad_output return torch.where( abs_x < beta, norm_grad * x / beta, norm_grad * torch.sign(x), ) @register_decomposition(aten.smooth_l1_loss_backward.grad_input) @pw_cast_for_opmath def smooth_l1_loss_backward_out( grad_output: Tensor, self: Tensor, target: Tensor, reduction: int, beta: float, grad_input: Tensor, ): result = smooth_l1_loss_backward(grad_output, self, target, reduction, beta) _maybe_resize_out(grad_input, result.shape) return _safe_copy_out(copy_from=result, copy_to=grad_input, exact_dtype=True) @register_decomposition(aten.huber_loss_backward.default) @pw_cast_for_opmath def huber_loss_backward( grad_output: Tensor, self: Tensor, target: Tensor, reduction: int, delta: float ): norm = 1.0 / self.numel() if reduction == Reduction.MEAN.value else 1.0 x = self - target return torch.where( x < -delta, -norm * grad_output * delta, torch.where(x > delta, norm * grad_output * delta, norm * x * grad_output), ) # We cannot use @out_wrapper() here, because the output tensor is not named 'out', it's 'grad_input' @register_decomposition(aten.huber_loss_backward.out) @pw_cast_for_opmath def huber_loss_backward_out( grad_output: Tensor, self: Tensor, target: Tensor, reduction: int, delta: float, grad_input: Tensor, ): result = huber_loss_backward(grad_output, self, target, reduction, delta) _maybe_resize_out(grad_input, result.shape) return _safe_copy_out(copy_from=result, copy_to=grad_input, exact_dtype=True) def _nll_loss_backward( grad_output: Tensor, self: Tensor, target: Tensor, weight: Optional[Tensor], reduction: int, ignore_index: int, total_weight: Tensor, ) -> Tensor: channel_dim = 0 if self.dim() < 2 else 1 if reduction == Reduction.MEAN.value: grad_output = grad_output / total_weight target = target.unsqueeze(channel_dim) safe_target = torch.where(target != ignore_index, target, 0) grad_input = torch.zeros_like(self) grad_input = torch.scatter(grad_input, channel_dim, safe_target, -1.0) if grad_input.dim() > grad_output.dim() > 0: grad_output = grad_output.unsqueeze(channel_dim) if weight is not None: new_shape = [1 for _ in range(self.dim())] new_shape[channel_dim] = weight.shape[0] weight = weight.reshape(new_shape) grad_output = grad_output * weight grad_output = torch.where(target != ignore_index, grad_output, 0) return grad_input * grad_output @register_decomposition(aten.glu_backward) @out_wrapper("grad_input") @pw_cast_for_opmath def glu_backward(grad_output: Tensor, self: Tensor, dim: int) -> Tensor: assert self.dim() > 0, "glu does not support 0-dimensional tensors" wrap_dim = utils.canonicalize_dim(self.dim(), dim) nIn = self.size(wrap_dim) assert ( nIn % 2 == 0 ), f"Halving dimension must be even, but dimension {wrap_dim} is size {nIn}" inputSize = nIn // 2 firstHalf = self.narrow(wrap_dim, 0, inputSize) secondHalf = self.narrow(wrap_dim, inputSize, inputSize) gradInputFirstHalf = torch.sigmoid(secondHalf) gradInputSecondHalf = ( (1.0 - gradInputFirstHalf) * gradInputFirstHalf * firstHalf * grad_output ) gradInputFirstHalf = gradInputFirstHalf * grad_output return torch.cat([gradInputFirstHalf, gradInputSecondHalf], dim=wrap_dim) @register_decomposition(aten.nll_loss_backward) @out_wrapper("grad_input") def nll_loss_backward( grad_output: Tensor, self: Tensor, target: Tensor, weight: Optional[Tensor], reduction: int, ignore_index: int, total_weight: Tensor, ) -> Tensor: assert 0 <= self.dim() <= 2, "input tensor should be 1D or 2D" assert ( target.dim() <= 1 ), "0D or 1D target tensor expected, multi-target not supported" no_batch_dim = self.dim() == 1 and target.dim() == 0 assert no_batch_dim or ( self.shape[0] == target.shape[0] ), f"size mismatch (got input: {self.shape}, target: {target.shape})" assert total_weight.numel() == 1, ( "expected total_weight to be a single element tensor, got: ", f"{total_weight.shape} ({total_weight.numel()} elements)", ) assert ( weight is None or weight.numel() == self.shape[-1] ), "weight tensor should be defined either for all or no classes" if reduction == Reduction.NONE.value and self.dim() == 2: assert grad_output.dim() == 1 and grad_output.shape[0] == self.shape[0], ( f"Expected a tensor of dimension 1 and tensor.size[0] == {self.shape[0]} but " f"got: dimension {grad_output.dim()} and tensor.size[0] == {grad_output.shape[0]}" ) else: assert ( grad_output.dim() <= 1 and grad_output.numel() == 1 ), f"Expected a single element grad_output tensor, but got: {grad_output.shape}" return _nll_loss_backward( grad_output, self, target, weight, reduction, ignore_index, total_weight ) @register_decomposition(aten.nll_loss2d_backward) @out_wrapper("grad_input") def nll_loss2d_backward( grad_output: Tensor, self: Tensor, target: Tensor, weight: Optional[Tensor], reduction: int, ignore_index: int, total_weight: Tensor, ) -> Tensor: assert ( self.dim() == 4 ), f"only batches of spatial inputs supported (4D tensors), but got input of dimension: {self.dim()}" assert ( target.dim() == 3 ), f"only batches of spatial targets supported (3D tensors) but got targets of dimension: {target.dim()}" assert ( self.shape[0] == target.shape[0] and self.shape[2] == target.shape[1] and self.shape[3] == target.shape[2] ), f"size mismatch (got input: {self.shape}, target: {target.shape}" assert total_weight.numel() == 1, ( "expected total_weight to be a single element tensor, " f"got: {total_weight.shape} ( {total_weight.numel()}, elements)" ) return _nll_loss_backward( grad_output, self, target, weight, reduction, ignore_index, total_weight ) @register_decomposition(aten.binary_cross_entropy) @out_wrapper() @pw_cast_for_opmath def binary_cross_entropy( self: Tensor, target: Tensor, weight: Optional[Tensor] = None, reduction: int = Reduction.MEAN.value, ) -> Tensor: # We cannot currently model this without introducing data-dependent control flow # TORCH_CHECK( # (input_val >= 0) && (input_val <= 1), # "all elements of input should be between 0 and 1" # ) loss = (target - 1) * torch.maximum( torch.log1p(-self), self.new_full((), -100) ) - target * torch.maximum(torch.log(self), self.new_full((), -100)) if weight is not None: loss = loss * weight return apply_loss_reduction(loss, reduction) @register_decomposition(aten.binary_cross_entropy_backward) @out_wrapper("grad_input") @pw_cast_for_opmath def binary_cross_entropy_backward( grad_output: Tensor, self: Tensor, target: Tensor, weight: Optional[Tensor] = None, reduction: int = Reduction.MEAN.value, ) -> Tensor: EPSILON = 1e-12 result = grad_output * (self - target) / torch.clamp(self * (1 - self), min=EPSILON) if weight is not None: result = result * weight if reduction == Reduction.MEAN.value: result = result / self.numel() return result @register_decomposition(aten.soft_margin_loss) @out_wrapper() @pw_cast_for_opmath def soft_margin_loss( input: Tensor, target: Tensor, reduction: int = Reduction.MEAN.value, ) -> Tensor: loss = torch.log1p(torch.exp(-input * target)) return apply_loss_reduction(loss, reduction) @register_decomposition(aten.soft_margin_loss_backward) @out_wrapper("grad_input") @pw_cast_for_opmath def soft_margin_loss_backward( grad_output: Tensor, self: Tensor, target: Tensor, reduction: int = Reduction.MEAN.value, ) -> Tensor: grad_input = target * grad_output * (torch.sigmoid(target * self) - 1) if reduction == Reduction.MEAN.value: grad_input = grad_input / self.numel() return grad_input @register_decomposition(aten.dist) @out_wrapper() def dist(input: Tensor, other: Tensor, p: float = 2): return aten.norm(input - other, p=p) @register_decomposition(aten._euclidean_dist) @out_wrapper() def _euclidean_dist(x1: Tensor, x2: Tensor) -> Tensor: x1_norm = x1.pow(2).sum(-1, True) x1_pad = torch.ones_like(x1_norm, memory_format=torch.contiguous_format) x2_norm = x2.pow(2).sum(-1, True) x2_pad = torch.ones_like(x2_norm, memory_format=torch.contiguous_format) x1_ = torch.cat([x1.mul(-2), x1_norm, x1_pad], -1) x2_ = torch.cat([x2, x2_pad, x2_norm], -1) result = x1_.matmul(x2_.mT) return result.clamp_min(0).sqrt() @register_decomposition(aten.slice_backward) @out_wrapper() def slice_backward( grad_output: Tensor, input_sizes: List[int], dim: int, start: int, end: int, step: int, ): grad_input = grad_output.new_zeros(input_sizes) return torch.slice_scatter(grad_input, grad_output, dim, start, end, step) @register_decomposition(aten.slice.Tensor) def slice_forward( # Tensor(a) self, int dim=0, SymInt? start=None, SymInt? end=None, SymInt step=1 self: Tensor, dim: int = 0, start: Optional[int] = None, end: Optional[int] = None, step: int = 1, ): ndim = self.dim() if ndim == 0: raise RuntimeError("slice() cannot be applied to a 0-dim tensor.") dim = utils.canonicalize_dim(self.dim(), dim) sizes = list(self.size()) strides = list(self.stride()) if step <= 0: raise RuntimeError("slice step must be positive") start_val = start if start is not None else 0 end_val = end if end is not None else sys.maxsize # 2^63 – 1 if start_val < 0: start_val += sizes[dim] if end_val < 0: end_val += sizes[dim] if start_val < 0: start_val = 0 elif start_val > sizes[dim]: start_val = sizes[dim] if end_val < start_val: end_val = start_val elif end_val > sizes[dim]: end_val = sizes[dim] storage_offset = self.storage_offset() + start_val * strides[dim] len = end_val - start_val sizes[dim] = (len + step - 1) // step strides[dim] *= step if self.is_quantized: raise NotImplementedError( "Slice decomposition for quantized tensors aren't implemented" ) else: return self.as_strided(sizes, strides, storage_offset) @register_decomposition(aten.select_backward) @out_wrapper() def select_backward(grad_output: Tensor, input_sizes: List[int], dim: int, index: int): grad_input = grad_output.new_zeros(input_sizes) return torch.select_scatter(grad_input, grad_output, dim, index) @register_decomposition(aten.diagonal_backward) @out_wrapper() def diagonal_backward( grad_output: Tensor, input_sizes: List[int], offset: int, dim1: int, dim2: int ): grad_input = grad_output.new_zeros(input_sizes) return torch.diagonal_scatter(grad_input, grad_output, offset, dim1, dim2) def _cast_grad_to_input_dtype( grad_output: Tensor, grad_input: Tensor, input_dtype: torch.dtype ): if grad_output.dtype != input_dtype: grad_input = grad_input.to(input_dtype) return grad_input @register_decomposition(aten._softmax_backward_data) @out_wrapper("grad_input") @compute_only_pw_cast_for_opmath def _softmax_backward_data( grad_output: Tensor, output: Tensor, dim: int, input_dtype: torch.dtype ): new_grad_output = grad_output * output grad_input = new_grad_output - output * torch.sum( new_grad_output, dim=dim, keepdim=True ) # CPU kernel doesn't respect input_dtype, but following check doesn't work for meta tensor # if grad_output.device == torch.device("cpu"): # return grad_input.contiguous() return _cast_grad_to_input_dtype(grad_output, grad_input, input_dtype).contiguous() @register_decomposition(aten._log_softmax_backward_data) @out_wrapper() @compute_only_pw_cast_for_opmath def _log_softmax_backward_data( grad_output: Tensor, output: Tensor, dim: int, input_dtype: torch.dtype ): grad_input = grad_output - torch.exp(output) * torch.sum( grad_output, dim=dim, keepdim=True ) return _cast_grad_to_input_dtype(grad_output, grad_input, input_dtype) def _im2col_col2im_indices_along_dim( input_d, kernel_d, dilation_d, padding_d, stride_d, device ): """Utility function to implement im2col and col2im""" blocks_d = input_d + padding_d * 2 - dilation_d * (kernel_d - 1) arange_kw = partial(torch.arange, dtype=torch.int64, device=device) # Stride kernel over input and find starting indices along dim d blocks_d_indices = arange_kw(0, blocks_d, stride_d).unsqueeze(0) # Apply dilation on kernel and find its indices along dim d kernel_grid = arange_kw(0, kernel_d * dilation_d, dilation_d).unsqueeze(-1) # Broadcast and add kernel starting positions (indices) with # kernel_grid along dim d, to get block indices along dim d return blocks_d_indices + kernel_grid @register_decomposition(aten.im2col) @out_wrapper() def im2col( input: Tensor, kernel_size: List[int], dilation: List[int], padding: List[int], stride: List[int], ) -> Tensor: torch._check(len(kernel_size) == 2, lambda: "im2col(): only 2D kernel supported") torch._check(len(dilation) == 2, lambda: "im2col(): only 2D dilation supported") torch._check(len(padding) == 2, lambda: "im2col(): only 2D padding supported") torch._check(len(stride) == 2, lambda: "im2col(): only 2D stride supported") def check_positive(param, param_name, strict=True): cond = all(p > 0 for p in param) if strict else all(p >= 0 for p in param) torch._check( cond, lambda: "{param_name} should be greater {'than' zero, but got {param}" ) check_positive(kernel_size, "kernel_size") check_positive(dilation, "dilation") check_positive(dilation, "padding", strict=False) check_positive(stride, "stride") shape = input.shape ndim = len(shape) torch._check( ndim in (3, 4) and all(d != 0 for d in shape[-3:]), lambda: "Expected 3D or 4D (batch mode) tensor for input with possible 0 batch size " f"and non-zero dimensions, but got: {tuple(shape)}", ) output_size = tuple( 1 + (out + 2 * pad - dil * (ker - 1) - 1) // st for out, pad, dil, ker, st in zip( shape[-2:], padding, dilation, kernel_size, stride ) ) torch._check( all(c > 0 for c in output_size), lambda: f"Given an input with spacial size {tuple(shape[-2:])}, " f"kernel_size={kernel_size}, dilation={dilation}, " f"padding={padding}, stride={stride}, " "the calculated shape of the array of sliding blocks " f"is {output_size}, but its components must be at least one.", ) batched_input = ndim == 4 if not batched_input: input = input.unsqueeze(0) batch_dim, channel_dim, input_h, input_w = input.shape stride_h, stride_w = stride padding_h, padding_w = padding dilation_h, dilation_w = dilation kernel_h, kernel_w = kernel_size blocks_row_indices = _im2col_col2im_indices_along_dim( input_h, kernel_h, dilation_h, padding_h, stride_h, input.device ) blocks_col_indices = _im2col_col2im_indices_along_dim( input_w, kernel_w, dilation_w, padding_w, stride_w, input.device ) # Note that F.pad takes (padding_left, padding_right, padding_top, padding_bottom) # ugh padded_input = F.pad(input, (padding_w, padding_w, padding_h, padding_h)) blocks_row_indices = blocks_row_indices.unsqueeze(-1).unsqueeze(-1) output = padded_input[:, :, blocks_row_indices, blocks_col_indices] output = output.permute(0, 1, 2, 4, 3, 5) num_blocks_row = blocks_row_indices.size(1) num_blocks_col = blocks_col_indices.size(1) output = output.reshape( batch_dim, channel_dim * kernel_h * kernel_w, num_blocks_row * num_blocks_col ) if not batched_input: output = output.squeeze(0) return output @register_decomposition(aten.col2im) @out_wrapper() @pw_cast_for_opmath def col2im( input: Tensor, output_size: List[int], kernel_size: List[int], dilation: List[int], padding: List[int], stride: List[int], ) -> Tensor: torch._check(len(output_size) == 2, lambda: "only 2D output_size supported") torch._check(len(kernel_size) == 2, lambda: "only 2D kernel supported") torch._check(len(dilation) == 2, lambda: "only 2D dilation supported") torch._check(len(padding) == 2, lambda: "only 2D padding supported") torch._check(len(stride) == 2, lambda: "only 2D stride supported") def check_positive(param, param_name, strict=True): cond = all(p > 0 for p in param) if strict else all(p >= 0 for p in param) torch._check( cond, lambda: "{param_name} should be greater than zero, but got {param}" ) check_positive(kernel_size, "kernel_size") check_positive(dilation, "dilation") check_positive(padding, "padding", strict=False) check_positive(stride, "stride") check_positive(output_size, "output_size") shape = input.shape ndim = len(shape) torch._check( ndim in (2, 3) and all(d != 0 for d in shape[-2:]), lambda: "Expected 2D or 3D (batch mode) tensor for input with possible 0 batch size " f"and non-zero dimensions, but got: {tuple(shape)}", ) prod_kernel_size = kernel_size[0] * kernel_size[1] torch._check( shape[-2] % prod_kernel_size == 0, lambda: "Expected size of input's first non-batch dimension to be divisible by the " f"product of kernel_size, but got input.shape[-2] = {shape[-2]} and " f"kernel_size={kernel_size}", ) col = [ 1 + (out + 2 * pad - dil * (ker - 1) - 1) // st for out, pad, dil, ker, st in zip( output_size, padding, dilation, kernel_size, stride ) ] L = col[0] * col[1] torch._check( shape[-1] == L, lambda: f"Given output_size={output_size}, kernel_size={kernel_size}, " f"dilation={dilation}, padding={padding}, stride={stride}, " f"expected input.size(-1) to be {L} but got {shape[-1]}.", ) torch._check( L > 0, lambda: f"Given output_size={output_size}, kernel_size={kernel_size}, " f"dilation={dilation}, padding={padding}, stride={stride}, " f"expected input.size(-1) to be {L} but got {shape[-1]}.", ) batched_input = ndim == 3 if not batched_input: input = input.unsqueeze(0) shape = input.shape out_h, out_w = output_size stride_h, stride_w = stride padding_h, padding_w = padding dilation_h, dilation_w = dilation kernel_h, kernel_w = kernel_size # col2im is defined as the backwards of im2col, so we differentiate its decomposition by hand input = input.reshape([shape[0], shape[1] // prod_kernel_size] + kernel_size + col) input = input.permute(0, 1, 2, 4, 3, 5) indices_row = _im2col_col2im_indices_along_dim( out_h, kernel_h, dilation_h, padding_h, stride_h, input.device ) indices_row = _unsqueeze_to_dim(indices_row, 4) indices_col = _im2col_col2im_indices_along_dim( out_w, kernel_w, dilation_w, padding_w, stride_w, input.device ) output_padded_size = [o + 2 * p for o, p in zip(output_size, padding)] output = input.new_zeros( [shape[0], shape[1] // prod(kernel_size)] + output_padded_size ) idx = (None, None, indices_row, indices_col) output = aten._unsafe_index_put(output, idx, input, accumulate=True) output = F.pad(output, (-padding_w, -padding_w, -padding_h, -padding_h)) if not batched_input: output = output.squeeze(0) return output @register_decomposition(aten.native_dropout_backward) @out_wrapper() def native_dropout_backward(grad_output: Tensor, mask: Tensor, scale: float): # According to the CUDA kernel implementation we should have this test; # but it seems to fail tests! # torch._check(mask.dtype == torch.bool, lambda: f"Mask should be Bool Scalar Type {mask.dtype}") # Mimicking CUDA kernel's behavior for output stride: output follow input's memory format # This different from TensorIterator's behavior r = (grad_output * (mask.type_as(grad_output) * scale)).clone( memory_format=utils.suggest_memory_format(grad_output) ) return r @register_decomposition(aten.unfold_backward) @out_wrapper() def unfold_backward( grad: Tensor, input_size: List[int], dimension: int, size: int, step: int ) -> Tensor: if len(input_size) == 0: return torch.squeeze_copy(grad, 0) dim = utils.canonicalize_dim(len(input_size), dimension) idx = torch.arange(input_size[dim], device=grad.device, dtype=torch.int32) idx = idx.unfold(0, size, step).flatten() grad = grad.movedim(-1, dim + 1).flatten(dim, dim + 1) # nb. At the moment this generates two kernels in triton # It could potentially be fused into one call to scatter_reduce, # in the case step <= size provided scatter_reduce generates 1 kernel grad_input = grad.new_zeros(input_size) index = (None,) * dim + (idx,) return aten._unsafe_index_put(grad_input, index, grad, accumulate=True).contiguous() @register_decomposition(aten.logit_backward.default) @pw_cast_for_opmath def logit_backward( grad_output: Tensor, self: Tensor, eps: Optional[float] = None ) -> Tensor: if eps is not None: lo = eps hi = 1.0 - lo return torch.where( torch.logical_and(self >= lo, self <= hi), grad_output / (self * (1.0 - self)), 0.0, ) else: return torch.where( torch.logical_and(self >= 0.0, self <= 1.0), grad_output / (self * (1.0 - self)), self.new_full((), float("nan")), ) @register_decomposition(aten.dropout) @aten.dropout.default.py_impl(DispatchKey.CompositeImplicitAutograd) @aten.dropout.default.py_impl(DispatchKey.Autograd) def dropout(input: Tensor, p: float, train: Optional[bool]): if train and p != 0: return aten.native_dropout(input, p, train)[0] else: return input.clone() @register_decomposition(aten.native_dropout) @out_wrapper("out0", "out1") def native_dropout(input: Tensor, p: float, train: Optional[bool]): if train and p != 0: if p == 1: return (torch.zeros_like(input), torch.zeros_like(input, dtype=torch.bool)) if not input.dtype.is_floating_point: raise RuntimeError( "result type Float can't be cast to the desired output type Long" ) bool_mask = torch.rand_like(input) > p res = bool_mask * input * float(1.0 / (1.0 - p)) return (res, bool_mask) else: return (input, torch.ones_like(input, dtype=torch.bool)) @register_decomposition(aten._softmax) @out_wrapper() def _softmax(x: Tensor, dim: int, half_to_float: bool): # eager softmax returns a contiguous tensor. Ensure that decomp also returns # a contiguous tensor. x = x.contiguous() if half_to_float: assert x.dtype == torch.half computation_dtype, result_dtype = utils.elementwise_dtypes( x, type_promotion_kind=utils.ELEMENTWISE_TYPE_PROMOTION_KIND.DEFAULT ) x = x.to(computation_dtype) if x.numel() == 0: unnormalized = torch.exp(x) else: x_max = torch.amax(x, dim, keepdim=True) unnormalized = torch.exp(x - x_max) result = unnormalized / torch.sum(unnormalized, dim, keepdim=True) if not half_to_float: result = result.to(result_dtype) return result @register_decomposition(aten._log_softmax) @out_wrapper() def _log_softmax(x: Tensor, dim: int, half_to_float: bool): # eager log_softmax returns a contiguous tensor. Ensure that decomp also # returns a contiguous tensor. x = x.contiguous() if half_to_float: assert x.dtype == torch.half computation_dtype, result_dtype = utils.elementwise_dtypes( x, type_promotion_kind=utils.ELEMENTWISE_TYPE_PROMOTION_KIND.DEFAULT ) x = x.to(computation_dtype) if x.numel() == 0: shifted = x else: x_max = torch.amax(x, dim, keepdim=True) shifted = x - x_max shifted_logsumexp = torch.log(torch.sum(torch.exp(shifted), dim, keepdim=True)) result = shifted - shifted_logsumexp if not half_to_float: result = result.to(result_dtype) return result @register_decomposition(aten.embedding) @out_wrapper() def embedding( weight: Tensor, indices: Tensor, padding_idx: int = -1, scale_grad_by_freq: bool = False, sparse: bool = False, ) -> Tensor: assert weight.dim() == 2, "'weight' must be 2-D" # Nb. scale_grad_by_freq is not used in the forward if indices.ndim <= 1: # We need this one as weight[indices] calls item() in these cases out = weight.index_select(0, indices) if indices.ndim == 0: out = out.squeeze(0) return out else: return weight[indices] @register_decomposition(aten.embedding_dense_backward) @out_wrapper() def embedding_dense_backward( grad_output: Tensor, indices: Tensor, num_weights: int, padding_idx: int, scale_grad_by_freq: bool, ): computation_dtype, result_dtype = utils.elementwise_dtypes( grad_output, type_promotion_kind=utils.ELEMENTWISE_TYPE_PROMOTION_KIND.DEFAULT ) grad_output = grad_output.to(computation_dtype) indices = _maybe_convert_to_dtype(indices, torch.long) # type: ignore[assignment] if scale_grad_by_freq: counts = indices.new_zeros((num_weights,)) ones = torch.ones_like(indices) counts = aten._unsafe_index_put(counts, [indices], ones, accumulate=True) grad_weights_scale = counts[indices] grad_output = grad_output / grad_weights_scale.unsqueeze(-1) mask = _unsqueeze_to_dim(indices == padding_idx, grad_output.ndim) grad = grad_output.masked_fill(mask, 0) grad_weight = grad_output.new_zeros( (num_weights,) + grad_output.shape[indices.ndim :] ) return aten._unsafe_index_put(grad_weight, [indices], grad, accumulate=True).to( result_dtype ) def prod(x: List[int]): r = 1 for i in x: r *= i return r def _pad_chunk( tensors: List[Tensor], dim: int, num_chunks: int, ) -> List[Tensor]: padded_tensors = [] for tensor in tensors: tensor_size = tensor.size() pad_along_dim = (tensor_size[dim] + num_chunks - 1) // num_chunks * num_chunks if pad_along_dim != tensor_size[dim]: # Use aten.constant_pad_nd instead of copy_ for functionalization pad = [0] * 2 * (tensor.ndim - dim - 1) + [ 0, pad_along_dim - tensor_size[dim], ] tensor = aten.constant_pad_nd(tensor, pad, 0) view_size = tensor_size[:dim] + torch.Size([num_chunks, -1]) padded_tensors.append(tensor.view(view_size)) return padded_tensors def have_same_ndims(tensors: List[Tensor]): ndim = tensors[0].ndim for tensor in tensors: if tensor.ndim != ndim: return False return True def leading_dimension_matches(tensors: List[Tensor], dim: int): leading_dim_sizes = tensors[0].size()[:dim] for tensor in tensors: torch._check( tensor.size()[:dim] == leading_dim_sizes, lambda: "_chunk_cat expects same sizes of 0,...,dim-1 dimensions for all tensors", ) def _preprocess_chunk_cat_inputs( tensors: List[Tensor], dim: int, num_chunks: int, ): torch._check(num_chunks >= 1, lambda: "_chunk_cat expects positive num_chunks") torch._check( len(tensors) > 0, lambda: "_chunk_cat expects a non-empty input tensor list" ) expected_dtype = tensors[0].dtype expected_device = tensors[0].device for tensor in tensors: torch._check(tensor.numel() > 0, lambda: "_chunk_cat expects non-empty tensor") torch._check( tensor.dtype == expected_dtype, lambda: "_chunk_cat expects all input tensors with the same dtype", ) torch._check( tensor.device == expected_device, lambda: "_chunk_cat expects all inputs tensors on the same device", ) if have_same_ndims(tensors): dim = utils.canonicalize_dim(tensors[0].dim(), dim) else: torch._check( dim >= 0, lambda: "_chunk_cat expects non-negative dim when input tensors have different ndims", ) for tensor in tensors: torch._check( dim < tensor.ndim, lambda: "_chunk_cat expects dim < ndim for all input tensors", ) leading_dimension_matches(tensors, dim) return dim @register_decomposition([aten._chunk_cat.default, aten._chunk_cat.out]) def _chunk_cat( tensors: List[Tensor], dim: int, num_chunks: int, out: Optional[Tensor] = None, ) -> Tensor: dim = _preprocess_chunk_cat_inputs(tensors, dim, num_chunks) padded_tensors = _pad_chunk(tensors, dim, num_chunks) if out is None: return torch.cat(padded_tensors, dim + 1) else: torch.cat(padded_tensors, dim + 1, out=out) return out @register_decomposition(aten.split_with_sizes) def split_with_sizes( self: Tensor, split_sizes: List[int], dim: int = 0 ) -> List[Tensor]: # NB: Perform the check_is_size tests first so that the # sum test does not try to do a replacement for i in range(len(split_sizes)): torch._check_is_size( split_sizes[i], lambda: "split_with_sizes expects split_sizes have only non-negative entries", ) torch._check_with( ValueError, sum(split_sizes) == self.shape[dim], lambda: f"Split sizes add up to {sum(split_sizes)} but got the tensor's size of {self.shape[dim]}", ) num_splits = len(split_sizes) splits = [] start_idx = 0 # Avoid importing sympy at a module level from torch.fx.experimental.symbolic_shapes import expect_true for i in range(num_splits): length = split_sizes[i] # We know this is true thanks to the sum, but this assertion helps # out our internal reasoning expect_true(start_idx + length <= self.shape[dim]) splits.append(self.narrow(dim, start_idx, length)) start_idx += length return splits # out_wrapper currently does not allow optional outputs @register_decomposition( [aten.split_with_sizes_copy.default, aten.split_with_sizes_copy.out] ) def split_with_sizes_copy( self: Tensor, split_sizes: List[int], dim: int = 0, out: Optional[List[Tensor]] = None, ) -> Optional[List[Tensor]]: splits = split_with_sizes(self, split_sizes, dim=dim) if out is None: return [s.clone(memory_format=torch.contiguous_format) for s in splits] else: for output, split in zip(out, splits): _maybe_resize_out(output, split.shape) _safe_copy_out(copy_from=split, copy_to=output, exact_dtype=True) return None @register_decomposition(aten.unsafe_split.Tensor) def unsafe_split(input: Tensor, split_size: int, dim: int = 0) -> Tuple[Tensor, ...]: return aten.split.Tensor(input, split_size, dim) @register_decomposition(aten.unsafe_split_with_sizes.default) def unsafe_split_with_sizes( input: Tensor, split_sizes: List[int], dim: int = 0 ) -> Tuple[Tensor, ...]: return aten.split_with_sizes.default(input, split_sizes, dim) @register_decomposition(aten.split.Tensor) def split(self: Tensor, split_size: int, dim: int = 0) -> Tuple[Tensor, ...]: input_sizes = self.shape dim_size = input_sizes[dim] if split_size == 0: assert dim_size == 0 return (self,) chunks = (dim_size + split_size - 1) // split_size # Avoid importing sympy at a module level from torch.fx.experimental.symbolic_shapes import guard_int chunks = guard_int(chunks) split_sizes = [split_size for i in range(chunks)] split_sizes[-1] = split_size - (split_size * chunks - dim_size) return torch.split(self, split_sizes, dim) @aten.tensor_split.tensor_indices_or_sections.py_impl( DispatchKey.CompositeImplicitAutograd ) def tensor_split_tensor_indices_or_sections_py_impl( self: Tensor, tensor_indices_or_sections: Tensor, dim: int = 0, ) -> Tuple[Tensor, ...]: assert tensor_indices_or_sections.device.type == "cpu" assert tensor_indices_or_sections.dtype == torch.int64 split_dim = tensor_indices_or_sections.dim() torch._check( split_dim == 1 or split_dim == 0, lambda: "tensor_split expected tensor_indices_or_sections to be a zero-dimensional " f"or one-dimensional tensor, but got a tensor with {split_dim} dims", ) if split_dim == 0: sections = tensor_indices_or_sections.item() assert isinstance(sections, IntLike) return self.tensor_split(sections, dim) else: indices = [i.item() for i in tensor_indices_or_sections] return self.tensor_split(indices, dim) # TODO: this doesn't appear to have enough precision in bfloat16 @register_decomposition(aten.addmm) @out_wrapper() @pw_cast_for_opmath def addmm(self: Tensor, mat1: Tensor, mat2: Tensor, beta: int = 1, alpha: int = 1): if not self.is_floating_point() and not self.is_complex(): beta = int(beta) alpha = int(alpha) out = alpha * torch.mm(mat1, mat2) if beta == 0: return out # The output of aten.addmm is contiguous, we need to match this behavior in the decomposition. # The original implementation 'beta * self + out' would return a strided tensor if `self` is strided. # We thus use `out`, the output of torch.mm, which is always contiguous, as the first argument for addition. # This is relying on TensorIterator's behavior that it takes higher precedence on the stride of first input. # Alternative, we can write `(beta * self + out).contiguous()`, but it introduces another copy in some cases. # This implementation is not ideal, and we should revisit this when we have a better solution. return out + beta * self @register_decomposition(aten._addmm_activation) @out_wrapper() @pw_cast_for_opmath def _addmm_activation( self: Tensor, mat1: Tensor, mat2: Tensor, beta: int = 1, alpha: int = 1, use_gelu: bool = False, ): out = addmm(self, mat1, mat2, beta, alpha) if use_gelu: if self.is_cuda: return aten.gelu(out, approximate="tanh") else: return aten.gelu(out) return aten.relu(out) @register_decomposition(aten.addmv) @out_wrapper() @pw_cast_for_opmath def addmv(self: Tensor, mat1: Tensor, vec: Tensor, beta: int = 1, alpha: int = 1): if not self.is_floating_point() and not self.is_complex(): beta = int(beta) alpha = int(alpha) out = alpha * torch.mv(mat1, vec) if beta == 0: return out return out + beta * self @register_decomposition(aten.native_group_norm_backward.default) @pw_cast_for_opmath def native_group_norm_backward( grad_output: Tensor, input: Tensor, mean: Tensor, rstd: Tensor, gamma: Optional[Tensor], N: int, C: int, HxW: int, group: int, output_mask: List[bool], ) -> Tuple[Optional[Tensor], Optional[Tensor], Optional[Tensor]]: utils.check_same_device( grad_output, input, mean, rstd, allow_cpu_scalar_tensors=False ) utils.check_same_shape(input, grad_output, allow_cpu_scalar_tensors=False) utils.check_same_shape(mean, rstd, allow_cpu_scalar_tensors=False) torch._check( input.numel() == N * C * HxW, lambda: f"Expect input to have { N * C * HxW} elements", ) torch._check( mean.shape == (N, group), lambda: f"Expect mean to have shape ({N}, {group}, but got {mean.shape}", ) torch._check( gamma is None or gamma.numel() == C, lambda: f"Expect gamma to have {C} elements but got {gamma.numel() if gamma is not None else -1}", ) cpg, _rem = divmod(C, group) torch._check( _rem == 0, lambda: f"Expect number of channels {C} to be evenly-divisible by number of groups {group}", ) # Compute Internal gradients ds = torch.mul(grad_output, input).view(N, C, HxW).sum(dim=[2]) db = grad_output.view(N, C, HxW).sum(dim=[2]) d_input: Optional[Tensor] = None d_gamma: Optional[Tensor] = None d_bias: Optional[Tensor] = None if output_mask[0]: s = 1.0 / (HxW * cpg) if gamma is not None: ds_val = torch.mul(ds, gamma.unsqueeze(0)).reshape(N, group, cpg).sum(2) db_val = torch.mul(db, gamma.unsqueeze(0)).reshape(N, group, cpg).sum(2) c1 = torch.mul( rstd.unsqueeze(-1), gamma.reshape(1, group, cpg), ) else: ds_val = ds.reshape(N, group, cpg).sum(2) db_val = db.reshape(N, group, cpg).sum(2) c1 = torch.mul( rstd.unsqueeze(-1), torch.ones((1, group, cpg), device=rstd.device), ) c2 = (db_val * mean - ds_val) * rstd * rstd * rstd * s c3 = -c2 * mean - db_val * rstd * s c1 = c1.unsqueeze(-1) c2 = _unsqueeze_to_dim(c2, 4) c3 = _unsqueeze_to_dim(c3, 4) d_input = ( torch.mul(grad_output.reshape(N, group, cpg, HxW), c1) + torch.mul(input.reshape(N, group, cpg, HxW), c2) + c3 ) d_input = d_input.reshape(input.shape).to(input.dtype) if output_mask[1]: d_gamma = ( ( (ds.view(N, group, cpg) - db.view(N, group, cpg) * mean.unsqueeze(-1)) * rstd.unsqueeze(-1) ) .sum(dim=[0]) .reshape(C) ) if output_mask[2]: d_bias = db.sum(dim=[0]) return (d_input, d_gamma, d_bias) # out_wrapper currently does not allow optional outputs @register_decomposition(aten.native_group_norm_backward.out) def native_group_norm_backward_out( grad_output: Tensor, input: Tensor, mean: Tensor, rstd: Tensor, gamma: Optional[Tensor], N: int, C: int, HxW: int, group: int, output_mask: List[bool], *, out0: torch.Tensor, out1: torch.Tensor, out2: torch.Tensor, ) -> Tuple[Optional[Tensor], Optional[Tensor], Optional[Tensor]]: result = native_group_norm_backward( grad_output, input, mean, rstd, gamma, N, C, HxW, group, output_mask ) grad_input = (out0, out1, out2) for i, r in enumerate(result): if r is not None: _maybe_resize_out(grad_input[i], r.shape) _safe_copy_out(copy_from=r, copy_to=grad_input[i], exact_dtype=True) return grad_input def _maybe_cast(x: Optional[Tensor], dtype) -> Optional[Tensor]: if x is not None: return x.to(dtype) return x # TODO: Take a closer look at the type promotion semantics @register_decomposition(aten.native_layer_norm_backward.default) def native_layer_norm_backward( grad_out: Tensor, input: Tensor, normalized_shape: List[int], mean: Tensor, rstd: Tensor, weight: Optional[Tensor], bias: Optional[Tensor], output_mask: List[bool], ) -> Tuple[Optional[Tensor], Optional[Tensor], Optional[Tensor]]: input_shape = input.shape input_ndim = input.dim() computation_dtype = utils.get_computation_dtype(input.dtype) grad_out_cast, input_cast, weight_cast, bias_cast = ( x.to(computation_dtype).contiguous() if x is not None else x for x in (grad_out, input, weight, bias) ) assert grad_out_cast is not None axis = input_ndim - len(normalized_shape) inner_dims = input_shape[axis:] outer_dims = input_shape[:axis] inner_dim_indices: List[int] = [] outer_dim_indices: List[int] = [] for i in range(input_ndim): if i >= axis: inner_dim_indices.append(i) else: outer_dim_indices.append(i) N = prod(inner_dims) # type: ignore[arg-type] M = prod(outer_dims) # type: ignore[arg-type] if M <= 0 or N <= 0: return ( input.new_zeros(input_shape) if output_mask[0] else None, input.new_zeros(input_shape[axis:]) if output_mask[1] else None, input.new_zeros(input_shape[axis:]) if output_mask[2] else None, ) mean = _unsqueeze_to_dim(mean, input_cast.dim()) # type: ignore[union-attr] rstd = _unsqueeze_to_dim(rstd, input_cast.dim()) # type: ignore[union-attr] x_hat = (input_cast - mean) * rstd if weight_cast is not None: grad_x_hat = grad_out_cast * weight_cast else: grad_x_hat = grad_out_cast a = grad_x_hat * N b = torch.sum(grad_x_hat, inner_dim_indices, True) c1 = torch.mul(grad_x_hat, x_hat) c2 = torch.sum(c1, inner_dim_indices, True) c3 = torch.mul(x_hat, c2) inner = a - b - c3 d_input: Optional[Tensor] = None d_weight: Optional[Tensor] = None d_bias: Optional[Tensor] = None if output_mask[0]: d_input = (rstd / N) * inner if output_mask[1] and weight_cast is not None: if len(outer_dim_indices) > 0: d_weight = torch.sum(grad_out_cast * x_hat, outer_dim_indices, False) else: d_weight = grad_out_cast * x_hat if output_mask[2] and bias_cast is not None: if len(outer_dim_indices) > 0: d_bias = torch.sum(grad_out_cast, outer_dim_indices, False) else: d_bias = grad_out_cast.clone() return ( _maybe_cast(d_input, input.dtype), _maybe_cast(d_weight, input.dtype), _maybe_cast(d_bias, input.dtype), ) # out_wrapper currently does not allow optional outputs @register_decomposition(aten.native_layer_norm_backward.out) def native_layer_norm_backward_out( grad_out: Tensor, input: Tensor, normalized_shape: List[int], mean: Tensor, rstd: Tensor, weight: Optional[Tensor], bias: Optional[Tensor], output_mask: List[bool], *, out0: torch.Tensor, out1: torch.Tensor, out2: torch.Tensor, ) -> Tuple[Optional[Tensor], Optional[Tensor], Optional[Tensor]]: result = native_layer_norm_backward( grad_out, input, normalized_shape, mean, rstd, weight, bias, output_mask ) grad_input = (out0, out1, out2) for i, r in enumerate(result): if r is not None: _maybe_resize_out(grad_input[i], r.shape) _safe_copy_out(copy_from=r, copy_to=grad_input[i], exact_dtype=True) return grad_input def native_batch_norm_helper( input: Tensor, weight: Optional[Tensor], bias: Optional[Tensor], running_mean: Optional[Tensor], running_var: Optional[Tensor], training: bool, momentum: float, eps: float, functional: bool, ) -> Tuple[Tensor, Tensor, Tensor, Optional[Tensor], Optional[Tensor]]: reduction_dims = [0] + list(range(2, input.dim())) computation_dtype = utils.get_computation_dtype(input.dtype) new_running_mean = running_mean new_running_var = running_var if training: computation_dtype = utils.get_computation_dtype(input.dtype) input_acc = input.to(dtype=computation_dtype) biased_var, mean = torch.var_mean( input_acc, dim=reduction_dims, correction=0, keepdim=True ) rstd = torch.rsqrt(biased_var + eps) output = (input - mean) * rstd save_mean = torch.squeeze(mean, reduction_dims) save_rstd = torch.squeeze(rstd, reduction_dims) if running_mean is not None: new_running_mean = momentum * save_mean + (1 - momentum) * running_mean if not functional: running_mean.copy_(new_running_mean) if running_var is not None: n = input.numel() / input.shape[1] # This doesn't strictly match eager's numerics, which accumulates var sum and then directly applies the correction # But... that would require re-implementing var here, for negligible numerics gain on a tensor whose # numerics probably don't matter. squeezed_var = torch.squeeze(biased_var, reduction_dims) unbiased_var = squeezed_var * (n / (n - 1)) new_running_var = momentum * unbiased_var + (1 - momentum) * running_var if not functional: running_var.copy_(new_running_var) else: assert running_mean is not None and running_var is not None running_mean = running_mean.to(dtype=computation_dtype, copy=True) new_running_mean = running_mean running_var = running_var.to(dtype=computation_dtype, copy=True) new_running_var = running_var mean = running_mean invstd = 1 / (torch.sqrt(running_var + eps)) # Very annoying inconsistency where CPU and CUDA give different shapes if input.device.type != "cpu": save_mean = running_mean save_rstd = invstd else: save_mean = input.new_zeros((0,)) save_rstd = input.new_zeros((0,)) mean = _unsqueeze_to_dim(mean, input.dim() - 1) invstd = _unsqueeze_to_dim(invstd, input.dim() - 1) output = (input - mean) * invstd if weight is not None: weight = weight.flatten() weight = _unsqueeze_to_dim(weight, input.dim() - 1) output = output * weight if bias is not None: bias = bias.flatten() bias = _unsqueeze_to_dim(bias, input.dim() - 1) output = output + bias if input.device.type == "cpu": save_mean = save_mean.to(dtype=input.dtype) save_rstd = save_rstd.to(dtype=input.dtype) return ( output.to(dtype=input.dtype), save_mean, save_rstd, new_running_mean, new_running_var, ) @register_decomposition(aten.native_batch_norm) @out_wrapper("out", "save_mean", "save_invstd") def native_batch_norm( input: Tensor, weight: Optional[Tensor], bias: Optional[Tensor], running_mean: Optional[Tensor], running_var: Optional[Tensor], training: bool, momentum: float, eps: float, ) -> Tuple[Tensor, Tensor, Tensor]: output, save_mean, save_rstd, _, _ = native_batch_norm_helper( input, weight, bias, running_mean, running_var, training, momentum, eps, False ) return output, save_mean, save_rstd # TODO: this decomposition is NOT here to stay. We would much prefer replacing native_batch_norm # with our new correctly schema'd _native_batch_norm_legit and its variants, but # we cannot do that immediately in the C++ because it would be forwards incompatible # with some mobile use cases. # # Since this change is most impactful for aot autograd/functionalization, we simply # register this decomposition on the Autograd key for the python dispatcher (which is # currently only used by aot autograd/functionalization and no one else, really). # In two weeks or so, we should remove this decomposition and phase out the current native_batch_norm # to be _native_batch_norm_legit and have the right schema (stating that there are input mutations). @aten.native_batch_norm.default.py_impl(DispatchKey.Autograd) @aten.native_batch_norm.default.py_impl(DispatchKey.CompositeImplicitAutograd) def native_batch_norm_decomposition( input: Tensor, weight: Optional[Tensor], bias: Optional[Tensor], running_mean: Optional[Tensor], running_var: Optional[Tensor], training: bool, momentum: float, eps: float, ) -> Tuple[Tensor, Tensor, Tensor]: if running_mean is None and running_var is None: return aten._native_batch_norm_legit( input, weight, bias, training, momentum, eps ) if running_mean is None: raise RuntimeError( "running_mean is None, but running_var is provided. " "They should both be None or both be provided." ) if running_var is None: raise RuntimeError( "running_var is None, but running_mean is provided. " "They should both be None or both be provided." ) if training: # HACK: batch norm consolidation should clean this up so this op doesn't take in a training arg. return aten._native_batch_norm_legit( input, weight, bias, running_mean, running_var, training, momentum, eps ) else: return aten._native_batch_norm_legit_no_training( input, weight, bias, running_mean, running_var, momentum, eps ) @aten.unsafe_chunk.default.py_impl(DispatchKey.CompositeImplicitAutograd) def unsafe_chunk_py_impl(tensor, chunks, dim=0) -> List[Tensor]: dim_size = tensor.size(dim) split_size = (dim_size + chunks - 1) // chunks if split_size == 0 and dim_size == 0: split_sizes = [split_size for _ in chunks] split_sizes[chunks - 1] = split_size - (split_size * chunks - dim_size) return torch.ops.aten.unsafe_split_with_sizes.default(tensor, split_sizes, dim) return torch.ops.aten.unsafe_split.Tensor(tensor, split_size, dim) @register_decomposition(aten._native_batch_norm_legit_no_training.default) def _native_batch_norm_legit_no_training( input: Tensor, weight: Optional[Tensor], bias: Optional[Tensor], running_mean: Tensor, running_var: Tensor, momentum: float, eps: float, ) -> Tuple[Tensor, Tensor, Tensor]: return aten._native_batch_norm_legit.default( input, weight, bias, running_mean, running_var, False, # training momentum, eps, ) @register_decomposition(aten._native_batch_norm_legit.default) def _native_batch_norm_legit( input: Tensor, weight: Optional[Tensor], bias: Optional[Tensor], running_mean: Tensor, running_var: Tensor, training: bool, momentum: float, eps: float, ) -> Tuple[Tensor, Tensor, Tensor]: output, save_mean, save_rstd, _, _ = native_batch_norm_helper( input, weight, bias, running_mean, running_var, training, momentum, eps, False ) return output, save_mean, save_rstd @register_decomposition(aten._native_batch_norm_legit.no_stats) def _native_batch_norm_legit_no_stats( input: Tensor, weight: Optional[Tensor], bias: Optional[Tensor], training: bool, momentum: float, eps: float, ) -> Tuple[Tensor, Tensor, Tensor]: output, save_mean, save_rstd, _, _ = native_batch_norm_helper( input, weight, bias, None, None, training, momentum, eps, False ) return output, save_mean, save_rstd @register_decomposition(aten._native_batch_norm_legit_functional.default) def _native_batch_norm_legit_functional( input: Tensor, weight: Optional[Tensor], bias: Optional[Tensor], running_mean: Tensor, running_var: Tensor, training: bool, momentum: float, eps: float, ) -> Tuple[Tensor, Tensor, Tensor, Tensor, Tensor]: ( output, save_mean, save_rstd, new_running_mean, new_running_var, ) = native_batch_norm_helper( input, weight, bias, running_mean, running_var, training, momentum, eps, True ) assert new_running_mean is not None, "new_running_mean should not be None" assert new_running_var is not None, "new_running_var should not be None" return output, save_mean, save_rstd, new_running_mean, new_running_var @register_decomposition(aten._fused_dropout) @out_wrapper("out0", "out1") @pw_cast_for_opmath def _fused_dropout_decomposition(input, p, generator=None): assert generator is None mask = (torch.rand_like(input) < p).to(dtype=torch.uint8) res = mask.type_as(input) * input * (1.0 / p) return (res, mask) def device_hint(tensor): if isinstance(tensor, torch._subclasses.FakeTensor): return tensor.fake_device else: return None @register_decomposition(aten._to_copy) @out_wrapper() def _to_copy( x: Tensor, *, dtype: Optional[torch.dtype] = None, layout=None, device: Optional[torch.device] = None, pin_memory: bool = False, non_blocking: bool = False, memory_format: Optional[torch.memory_format] = None, ): assert not layout or layout == torch.strided, "TODO" assert not pin_memory, "TODO" if device is None and dtype is None and memory_format is None: return x.clone() dtype_converted = False common_device = device_hint(x) if device is not None and device != x.device: # avoid conversions on cpu if dtype is not None and device.type == "cpu": x = torch._prims.convert_element_type(x, dtype) dtype_converted = True x = torch._prims.device_put(x, device) if dtype is not None and not dtype_converted: x = torch._prims.convert_element_type(x, dtype) dtype_converted = True if memory_format is not None: # no ref/prim for memory format return torch.clone(x, memory_format=memory_format) return x # Questionable decompositions # This is only valid if we're running the graph without autograd, such as if the backward pass has been traced. # Note that this decomposition causes issues with in-place ops @register_decomposition([aten.detach, aten.lift, aten.lift_fresh]) @out_wrapper() def nop_decomposition(x): return aten.alias(x) # Also register to the Autograd dispatch key, so this decomp can run above autograd. # native_batch_norm needs to decompose into other ops before autograd. @aten.cudnn_batch_norm.default.py_impl(DispatchKey.Autograd) @register_decomposition(aten.cudnn_batch_norm) @out_wrapper("out0", "out1", "out2", "out3") def cudnn_batch_norm( input: Tensor, weight: Tensor, bias: Optional[Tensor], running_mean: Optional[Tensor], running_var: Optional[Tensor], training: bool, exponential_average_factor: float, epsilon: float, ): a, b, c = aten.native_batch_norm( input, weight, bias, running_mean, running_var, training, exponential_average_factor, epsilon, ) # Cudnn return running mean and variance when training is True if training: return (a, b, c, input.new_zeros((0,), dtype=torch.uint8)) return ( a, weight.new_zeros((0,)), weight.new_zeros((0,)), input.new_zeros((0,), dtype=torch.uint8), ) def _broadcast_batch_norm_backward(x, broadcast_mask): for axis, mask in enumerate(broadcast_mask): if mask == 1 and not (axis < x.ndim and x.shape[axis] == broadcast_mask[axis]): x = x.unsqueeze(axis) return x @register_decomposition(aten.native_batch_norm_backward.default) def native_batch_norm_backward( grad_out: Tensor, input: Tensor, weight: Optional[Tensor], running_mean: Optional[Tensor], running_var: Optional[Tensor], save_mean: Optional[Tensor], save_invstd: Optional[Tensor], train: bool, eps: float, output_mask: List[bool], ) -> Tuple[Tensor, Optional[Tensor], Optional[Tensor]]: input_dtype = input.dtype if weight is not None: weight_dtype = weight.dtype else: weight_dtype = input_dtype computation_dtype = utils.get_computation_dtype(input.dtype) ( grad_out_cast, input_cast, weight_cast, running_mean_cast, running_var_cast, save_mean_cast, save_invstd_cast, ) = ( x.to(computation_dtype) if x is not None else x for x in ( grad_out, input, weight, running_mean, running_var, save_mean, save_invstd, ) ) input_shape = input.shape input_rank = input.dim() assert input_rank >= 2, "rank of the input must be at least 2" axis = 1 num_features = prod(list(input_shape)) / input_shape[axis] mean = save_mean_cast invstd = save_invstd_cast if train: assert save_mean_cast is not None and save_invstd_cast is not None else: assert running_mean_cast is not None and running_var_cast is not None mean = running_mean_cast invstd = torch.rsqrt(running_var_cast + eps) broadcast_mask: List[int] = [1] * input_rank broadcast_mask[axis] = input_shape[axis] reduction_axes: List[int] = [] for i in range(input_rank): if i != axis: reduction_axes.append(i) mean = _broadcast_batch_norm_backward(mean, broadcast_mask) # type: ignore[arg-type] norm = 1.0 / num_features grad_output_sum = torch.sum(grad_out_cast, reduction_axes) # type: ignore[arg-type] dot_p = torch.sum(grad_out_cast * (input_cast - mean), reduction_axes) # type: ignore[operator] grad_mean = _broadcast_batch_norm_backward(grad_output_sum * norm, broadcast_mask) proj_scale = _broadcast_batch_norm_backward(torch.mul(dot_p * norm, invstd * invstd), broadcast_mask) # type: ignore[operator] if weight_cast is None: grad_scale = _broadcast_batch_norm_backward(invstd, broadcast_mask) * 1.0 # type: ignore[arg-type] else: grad_scale = _broadcast_batch_norm_backward( invstd * weight_cast, broadcast_mask ) if train: proj = (input_cast - mean) * proj_scale # type: ignore[operator] grad_input = ((grad_out_cast - proj) - grad_mean) * grad_scale else: grad_input = grad_out_cast * grad_scale if output_mask[1]: grad_weight = dot_p * invstd else: grad_weight = None # "None" doesn't work with vjp, should use zeros for vjp if output_mask[2]: grad_bias = grad_output_sum else: grad_bias = None # "None" doesn't work with vjp, should use zeros for vjp return ( grad_input.to(input_dtype), _maybe_cast(grad_weight, weight_dtype), _maybe_cast(grad_bias, weight_dtype), ) # out_wrapper currently does not allow optional outputs @register_decomposition(aten.native_batch_norm_backward.out) def native_batch_norm_backward_out( grad_out: Tensor, input: Tensor, weight: Optional[Tensor], running_mean: Optional[Tensor], running_var: Optional[Tensor], save_mean: Optional[Tensor], save_invstd: Optional[Tensor], train: bool, eps: float, output_mask: List[bool], *, out0: torch.Tensor, out1: torch.Tensor, out2: torch.Tensor, ) -> Tuple[Tensor, Optional[Tensor], Optional[Tensor]]: result = native_batch_norm_backward( grad_out, input, weight, running_mean, running_var, save_mean, save_invstd, train, eps, output_mask, ) grad_input = (out0, out1, out2) for i, r in enumerate(result): if r is not None: _maybe_resize_out(grad_input[i], r.shape) _safe_copy_out(copy_from=r, copy_to=grad_input[i], exact_dtype=True) return grad_input @register_decomposition(aten.cudnn_batch_norm_backward) @out_wrapper("out0", "out1", "out2") def cudnn_batch_norm_backward( input: Tensor, grad_output: Tensor, weight: Tensor, running_mean: Optional[Tensor], running_var: Optional[Tensor], save_mean: Optional[Tensor], save_var: Optional[Tensor], epsilon: float, reserveSpace: Tensor, ): return aten.native_batch_norm_backward( grad_output, input, weight, running_mean, running_var, save_mean, save_var, True, epsilon, [True, True, True], ) @register_decomposition(aten._adaptive_avg_pool2d) @out_wrapper() @pw_cast_for_opmath def adaptive_avg_pool2d(input: Tensor, output_size: Tuple[int, int]): # Preconditions device = input.device shape = input.shape ndim = len(shape) torch._check( ndim in (3, 4), lambda: f"adaptive_avg_pool2d(): Expected 3D or 4D tensor, but got {ndim}", ) for d in input.shape[-2:]: torch._check( d != 0, lambda: "adaptive_avg_pool2d(): Expected input to have non-zero size for " f"non-batch dimensions, but input has shape {tuple(shape)}.", ) # Optimisation (we should also do this in the kernel implementation) if shape[-2] % output_size[-2] == 0 and shape[-1] % output_size[-1] == 0: stride = tuple(i // o for i, o in zip(shape[-2:], output_size)) kernel = tuple( i - (o - 1) * s for i, o, s in zip(shape[-2:], output_size, stride) ) return torch.nn.functional.avg_pool2d(input, kernel, stride) def start_index(a, b, c): return torch.div(a * c, b, rounding_mode="trunc") def end_index(a, b, c): return torch.div((a + 1) * c + b - 1, b, rounding_mode="trunc") def compute_idx(in_size, out_size): orange = torch.arange(out_size, device=device, dtype=torch.int64) i0 = start_index(orange, out_size, in_size) # Let length = end_index - start_index, i.e. the length of the pooling kernels # length.max() can be computed analytically as follows: maxlength = in_size // out_size + 1 in_size_mod = in_size % out_size # adaptive = True iff there are kernels with different lengths adaptive = not (in_size_mod == 0 or out_size % in_size_mod == 0) if adaptive: maxlength += 1 elif in_size_mod == 0: maxlength -= 1 range_max = torch.arange(maxlength, device=device, dtype=torch.int64) idx = i0.unsqueeze(-1) + range_max if adaptive: # Need to clamp to avoid accessing out-of-bounds memory # TODO make minimum accept scalars maxval = torch.scalar_tensor( in_size - 1, dtype=idx.dtype, device=idx.device ) idx = torch.minimum(idx, maxval) # Compute the length i1 = end_index(orange, out_size, in_size) length = i1 - i0 else: length = maxlength return idx, length, range_max, adaptive # length is not None if it's constant, otherwise we'll need to compute it idxh, length_h, range_max_h, adaptive_h = compute_idx(shape[-2], output_size[-2]) idxw, length_w, range_max_w, adaptive_w = compute_idx(shape[-1], output_size[-1]) vals = input[..., _unsqueeze_to_dim(idxh, 4), idxw] # Shortcut for the simpler case if not adaptive_h and not adaptive_w: return torch.mean(vals, dim=(-3, -1)) def maybe_mask(vals, length, range_max, adaptive, dim): if isinstance(length, IntLike): return vals, length else: # zero-out the things we didn't really want to select assert dim < 0 # hack mask = range_max >= length.unsqueeze(-1) if dim == -2: mask = _unsqueeze_to_dim(mask, 4) vals = torch.masked_fill(vals, mask, 0.0) # Compute the length of each window length = _unsqueeze_to_dim(length, -dim) return vals, length vals, length_h = maybe_mask( vals, length_h, range_max_h, adaptive=adaptive_h, dim=-2 ) vals, length_w = maybe_mask( vals, length_w, range_max_w, adaptive=adaptive_w, dim=-1 ) # We unroll the sum as we assume that the kernels are going to be small ret = None for i, j in product(range(vals.shape[-3]), range(vals.shape[-1])): if ret is None: ret = vals[..., i, :, j] else: ret = ret + vals[..., i, :, j] return ret / (length_h * length_w) @register_decomposition(aten.index_add_) def index_add_( x: TensorLike, dim: int, index: TensorLike, tensor: TensorLike, *, alpha: NumberType = 1, ): return _index_add(x, dim, index, tensor, inplace=True, alpha=alpha) @register_decomposition(aten.index_add) @out_wrapper() def index_add( x: TensorLike, dim: int, index: TensorLike, tensor: TensorLike, *, alpha: NumberType = 1, ): return _index_add(x, dim, index, tensor, inplace=False, alpha=alpha) def _index_add( x: TensorLike, dim: int, index: TensorLike, tensor: TensorLike, *, inplace: bool, alpha: NumberType = 1, ): dim = utils.canonicalize_dims(x.ndim, dim) torch._check( index.ndim <= 1, lambda: f"Index should have dimension 1 or 0 (got {index.ndim})", ) index_size = index.size(0) if index.ndim == 1 else 1 tensor_size = tensor.size(dim) if tensor.ndim > 0 else 1 torch._check( tensor_size == index_size, lambda: f"Number of indices ({index_size}) should be equal to tensor.size(dim) ({tensor_size}), for {dim=}", ) if alpha != 1: python_type = utils.dtype_to_type(x.dtype) torch._check( python_type == bool or utils.is_weakly_lesser_type(type(alpha), python_type), lambda: f"alpha argument of type {type(alpha)} cannot be safely cast to type {python_type}!", ) tensor = tensor * alpha # Treat scalars as elements of \R^1 zero_dim = x.ndim == 0 x1 = x.unsqueeze(0) if zero_dim else x idx = (None,) * dim + (index,) index_put = aten.index_put_ if inplace else aten.index_put out = index_put(x1, idx, tensor, accumulate=True) if inplace: return x else: return out.squeeze(0) if zero_dim else out.contiguous() @register_decomposition(aten.pad_sequence.default) @aten.pad_sequence.default.py_impl(DispatchKey.CompositeImplicitAutograd) def pad_sequence(sequences, batch_first=False, padding_value=0.0): torch._check(len(sequences) > 0, lambda: "received an empty list of sequences") sequences_size = len(sequences) max_size = sequences[0].size() trailing_dims = max_size[1:] max_len = max(x.size(0) for x in sequences) if batch_first: out_dims = (sequences_size, max_len) else: out_dims = (max_len, sequences_size) out_dims = out_dims + trailing_dims out = sequences[0].new_full(out_dims, padding_value) dim_paddings = (0, 0) * len(trailing_dims) for i in range(sequences_size): currseq = sequences[i] row = aten.constant_pad_nd( currseq, dim_paddings + (0, max_len - currseq.size(0)), padding_value ) if batch_first: out = aten.select_scatter(out, row, dim=0, index=i) else: out = aten.select_scatter(out, row, dim=1, index=i) return out @register_decomposition(aten.index_copy_) def index_copy_(x: TensorLike, dim: int, index: TensorLike, tensor: TensorLike): return _index_copy(x, dim, index, tensor, inplace=True) @register_decomposition(aten.index_copy) @out_wrapper() def index_copy(x: TensorLike, dim: int, index: TensorLike, tensor: TensorLike): return _index_copy(x, dim, index, tensor, inplace=False) def _index_copy( x: TensorLike, dim: int, index: TensorLike, tensor: TensorLike, *, inplace: bool ): dim = utils.canonicalize_dims(x.ndim, dim) torch._check( index.ndim <= 1, lambda: f"Index should have dimension 1 or 0 (got {index.ndim})", ) # Treat scalars as elements of \R^1 zero_dim = x.ndim == 0 x1 = x.unsqueeze(0) if zero_dim else x index = index.unsqueeze(0) if index.ndim == 0 else index idx = (None,) * dim + (index,) index_put = aten.index_put_ if inplace else aten.index_put out = index_put(x1, idx, tensor) if inplace: return x else: return out.squeeze(0) if zero_dim else out.contiguous() # nb: Should use acc_t, not op_math @register_decomposition(aten.log_sigmoid_forward) @out_wrapper("output", "buffer") @pw_cast_for_opmath def log_sigmoid_forward(self: Tensor) -> Tuple[Tensor, Tensor]: min = torch.minimum(self.new_zeros(()), self) z = torch.exp(-torch.abs(self)) if self.is_cuda: buffer = self.new_zeros((0,)) else: buffer = z return min - torch.log1p(z), buffer @register_decomposition(aten.uniform) @out_wrapper() def uniform( x: Tensor, low: Union[bool, int, float] = 0.0, high: Union[bool, int, float] = 1.0, generator: Optional[torch.Generator] = None, ): return prims._uniform_helper( x.shape, low=sym_float(low), high=sym_float(high), dtype=x.dtype, device=x.device, generator=generator, ) @register_decomposition(aten.uniform_) def uniform_(self, low=0, high=1, generator=None): return self.copy_(uniform(self, low, high, generator)) # aten/src/ATen/native/UpSample.cpp compute_output_size def upsample_compute_output_size(input_size, output_size, scale_factors): spatial_dimensions = len(input_size) - 2 if output_size is not None: torch._check( scale_factors is None, lambda: "Must specify exactly one of output_size and scale_factors", ) torch._check(len(output_size) == spatial_dimensions, lambda: "") return output_size if scale_factors is not None: # NB: this isn't necessary lol torch._check( output_size is None, lambda: "Must specify exactly one of output_size and scale_factors", ) torch._check(len(scale_factors) == spatial_dimensions, lambda: "") output_size = [] for i, s in enumerate(scale_factors): if int(s) == s: output_size.append(input_size[i + 2] * int(s)) else: output_size.append(sym_int(input_size[i + 2] * s)) return output_size torch._check( False, lambda: "Must specify exactly one of output_size and scale_factors" ) def get_scale_value(scales, idx): if scales is None: return None return scales[idx] @register_decomposition(aten.upsample_nearest1d.vec) @aten.upsample_nearest1d.vec.py_impl(DispatchKey.CompositeImplicitAutograd) @aten.upsample_nearest1d.vec.py_impl(DispatchKey.Autograd) def upsample_nearest1d_vec(input, output_size, scale_factors): osize = upsample_compute_output_size(input.size(), output_size, scale_factors) scale = get_scale_value(scale_factors, 0) return aten.upsample_nearest1d.default(input, osize, scale) @register_decomposition(aten._upsample_nearest_exact1d.vec) @aten._upsample_nearest_exact1d.vec.py_impl(DispatchKey.CompositeImplicitAutograd) @aten._upsample_nearest_exact1d.vec.py_impl(DispatchKey.Autograd) def _upsample_nearest_exact1d_vec(input, output_size, scale_factors): osize = upsample_compute_output_size(input.size(), output_size, scale_factors) scale = get_scale_value(scale_factors, 0) return aten._upsample_nearest_exact1d.default(input, osize, scale) @register_decomposition(aten.upsample_nearest2d.vec) @aten.upsample_nearest2d.vec.py_impl(DispatchKey.CompositeImplicitAutograd) @aten.upsample_nearest2d.vec.py_impl(DispatchKey.Autograd) def upsample_nearest2d_vec(input, output_size, scale_factors): osize = upsample_compute_output_size(input.size(), output_size, scale_factors) scale_h = get_scale_value(scale_factors, 0) scale_w = get_scale_value(scale_factors, 1) return aten.upsample_nearest2d.default(input, osize, scale_h, scale_w) @register_decomposition(aten._upsample_nearest_exact2d.vec) @aten._upsample_nearest_exact2d.vec.py_impl(DispatchKey.CompositeImplicitAutograd) @aten._upsample_nearest_exact2d.vec.py_impl(DispatchKey.Autograd) def _upsample_nearest_exact2d_vec(input, output_size, scale_factors): osize = upsample_compute_output_size(input.size(), output_size, scale_factors) scale_h = get_scale_value(scale_factors, 0) scale_w = get_scale_value(scale_factors, 1) return aten._upsample_nearest_exact2d.default(input, osize, scale_h, scale_w) @register_decomposition(aten.upsample_nearest3d.vec) @aten.upsample_nearest3d.vec.py_impl(DispatchKey.CompositeImplicitAutograd) @aten.upsample_nearest3d.vec.py_impl(DispatchKey.Autograd) def upsample_nearest3d_vec(input, output_size, scale_factors): osize = upsample_compute_output_size(input.size(), output_size, scale_factors) scale_d = get_scale_value(scale_factors, 0) scale_h = get_scale_value(scale_factors, 1) scale_w = get_scale_value(scale_factors, 2) return aten.upsample_nearest3d.default(input, osize, scale_d, scale_h, scale_w) @register_decomposition(aten._upsample_nearest_exact3d.vec) @aten._upsample_nearest_exact3d.vec.py_impl(DispatchKey.CompositeImplicitAutograd) @aten._upsample_nearest_exact3d.vec.py_impl(DispatchKey.Autograd) def _upsample_nearest_exact3d_vec(input, output_size, scale_factors): osize = upsample_compute_output_size(input.size(), output_size, scale_factors) scale_d = get_scale_value(scale_factors, 0) scale_h = get_scale_value(scale_factors, 1) scale_w = get_scale_value(scale_factors, 2) return aten._upsample_nearest_exact3d.default( input, osize, scale_d, scale_h, scale_w ) def _compute_upsample_nearest_indices(input, output_size, scales, exact=False): # For each dim in output_size, compute the set of input indices used # to produce the upsampled output. indices = [] num_spatial_dims = len(output_size) offset = 0.5 if exact else 0.0 for d in range(num_spatial_dims): # Math matches aten/src/ATen/native/cpu/UpSampleKernel.cpp # # Indices are computed as following: # scale = isize / osize # Case: exact=False # input_index = floor(output_index * scale) # Same as OpenCV INTER_NEAREST # # Case: exact=False # index_f32 = (output_index + 0.5) * scale - 0.5 # input_index = round(index_f32) # Same as Pillow and Scikit-Image/Scipy ndi.zoom osize = output_size[d] isize = input.shape[-num_spatial_dims + d] scale = isize / (isize * scales[d]) if scales[d] is not None else isize / osize output_indices = torch.arange(osize, dtype=torch.float32, device=input.device) input_indices = ((output_indices + offset) * scale).to(torch.int64) for _ in range(num_spatial_dims - 1 - d): input_indices = input_indices.unsqueeze(-1) indices.append(input_indices) return tuple(indices) @register_decomposition(aten.upsample_nearest1d.default) @aten.upsample_nearest1d.default.py_impl(DispatchKey.Autograd) @pw_cast_for_opmath def upsample_nearest1d( input: Tensor, output_size: List[int], scales: Optional[float] = None, ) -> Tensor: (l_indices,) = _compute_upsample_nearest_indices(input, output_size, (scales,)) return aten._unsafe_index(input, (None, None, l_indices)) @register_decomposition(aten._upsample_nearest_exact1d.default) @aten._upsample_nearest_exact1d.default.py_impl(DispatchKey.Autograd) @pw_cast_for_opmath def _upsample_nearest_exact1d( input: Tensor, output_size: List[int], scales: Optional[float] = None, ) -> Tensor: (l_indices,) = _compute_upsample_nearest_indices( input, output_size, (scales,), exact=True ) return aten._unsafe_index(input, (None, None, l_indices)) def _upsample_nearest2d_common(input, h_indices, w_indices): result = aten._unsafe_index(input, (None, None, h_indices, w_indices)) # convert output to correct memory format, if necessary memory_format = utils.suggest_memory_format(input) # following "heuristic: only use channels_last path when it's faster than the contiguous path" _, n_channels, _, _ = input.shape if input.device.type == "cuda" and n_channels < 4: memory_format = torch.contiguous_format result = result.contiguous(memory_format=memory_format) return result @register_decomposition(aten.upsample_nearest2d.default) @aten.upsample_nearest2d.default.py_impl(DispatchKey.Autograd) @pw_cast_for_opmath def upsample_nearest2d( input: Tensor, output_size: List[int], scales_h: Optional[float] = None, scales_w: Optional[float] = None, ) -> Tensor: h_indices, w_indices = _compute_upsample_nearest_indices( input, output_size, (scales_h, scales_w) ) return _upsample_nearest2d_common(input, h_indices, w_indices) @register_decomposition(aten._upsample_nearest_exact2d.default) @aten._upsample_nearest_exact2d.default.py_impl(DispatchKey.Autograd) @pw_cast_for_opmath def _upsample_nearest_exact2d( input: Tensor, output_size: List[int], scales_h: Optional[float] = None, scales_w: Optional[float] = None, ) -> Tensor: h_indices, w_indices = _compute_upsample_nearest_indices( input, output_size, (scales_h, scales_w), exact=True ) return _upsample_nearest2d_common(input, h_indices, w_indices) @register_decomposition(aten.upsample_nearest3d.default) @aten.upsample_nearest3d.default.py_impl(DispatchKey.Autograd) @pw_cast_for_opmath def upsample_nearest3d( input: Tensor, output_size: List[int], scales_d: Optional[float] = None, scales_h: Optional[float] = None, scales_w: Optional[float] = None, ) -> Tensor: d_indices, h_indices, w_indices = _compute_upsample_nearest_indices( input, output_size, (scales_d, scales_h, scales_w) ) result = aten._unsafe_index(input, (None, None, d_indices, h_indices, w_indices)) return result @register_decomposition(aten._upsample_nearest_exact3d.default) @aten._upsample_nearest_exact3d.default.py_impl(DispatchKey.Autograd) @pw_cast_for_opmath def _upsample_nearest_exact3d( input: Tensor, output_size: List[int], scales_d: Optional[float] = None, scales_h: Optional[float] = None, scales_w: Optional[float] = None, ) -> Tensor: d_indices, h_indices, w_indices = _compute_upsample_nearest_indices( input, output_size, (scales_d, scales_h, scales_w), exact=True ) result = aten._unsafe_index(input, (None, None, d_indices, h_indices, w_indices)) return result def gather_params(params, has_biases, has_projections): if has_biases and has_projections: group_size = 5 elif has_biases: group_size = 4 elif has_projections: group_size = 3 else: group_size = 2 assert len(params) % group_size == 0, len(params) return [ tuple(params[i : i + group_size]) for i in range(0, len(params), group_size) ] def params_hiddens(params, hiddens, i, bidirectional): if bidirectional: cur_params, cur_hidden = params[2 * i], hiddens[2 * i] bidir_params, bidir_hidden = params[2 * i + 1], hiddens[2 * i + 1] else: cur_params, cur_hidden = params[i], hiddens[i] bidir_params, bidir_hidden = None, None return cur_params, cur_hidden, bidir_params, bidir_hidden def update_hidden_for_packed(cur_hidden, last_batch_size, batch_size, hiddens): assert last_batch_size > batch_size hiddens.append(cur_hidden.narrow(0, batch_size, last_batch_size - batch_size)) return cur_hidden.narrow(0, 0, batch_size) def update_hidden_for_packed_reverse( cur_hidden, last_batch_size, batch_size, inp_hidden ): if last_batch_size == batch_size: return cur_hidden assert last_batch_size < batch_size return torch.concat( ( cur_hidden, inp_hidden.narrow(0, last_batch_size, batch_size - last_batch_size), ) ) def one_layer_rnn_data( inp, hidden, params, has_biases, hidden_fn, batch_sizes, reverse=False ): ih_weight = params[0] hh_weight = params[1] ih_bias = params[2] if has_biases else None hh_bias = params[3] if has_biases else None step_output = [] hiddens: List[torch.Tensor] = [] last_batch_size = batch_sizes[-1] if reverse else batch_sizes[0] cur_hidden = hidden.narrow(0, 0, last_batch_size) split_inp = torch.split(inp, list(batch_sizes)) if reverse: split_inp = split_inp[::-1] for inp in split_inp: i = inp.shape[0] if last_batch_size == i: pass # don't update cur_hidden # this will only happen when reverse=False, since batch sizes are sorted largest -> smallest elif reverse: cur_hidden = update_hidden_for_packed_reverse( cur_hidden, last_batch_size, i, hidden ) else: cur_hidden = update_hidden_for_packed( cur_hidden, last_batch_size, i, hiddens ) cur_hidden = hidden_fn(inp, cur_hidden, ih_weight, ih_bias, hh_weight, hh_bias) last_batch_size = i step_output.append(cur_hidden) if reverse: step_output.reverse() else: hiddens.append(cur_hidden) hiddens.reverse() out = torch.cat(step_output, 0) hidden_out = torch.cat(hiddens, 0) if not reverse else cur_hidden return out, hidden_out def rnn_cell(nonlinearity): def inner(i, cur_hidden, ih_weight, ih_bias, hh_weight, hh_bias): return nonlinearity(F.linear(cur_hidden, hh_weight, hh_bias) + i) return inner def rnn_cell_data(nonlinearity): def inner(i, cur_hidden, ih_weight, ih_bias, hh_weight, hh_bias): i = F.linear(i, ih_weight, ih_bias) return nonlinearity(F.linear(cur_hidden, hh_weight, hh_bias) + i) return inner def one_layer_rnn(inp, hidden, params, has_biases, hidden_fn, reverse=False): ih_weight = params[0] hh_weight = params[1] ih_bias = params[2] if has_biases else None hh_bias = params[3] if has_biases else None precomputed_input = F.linear(inp, ih_weight, ih_bias) precomputed_input = precomputed_input.flip(0) if reverse else precomputed_input cur_hidden = hidden.unsqueeze(0) step_output = [] for i in precomputed_input: cur_hidden = hidden_fn(i, cur_hidden, ih_weight, ih_bias, hh_weight, hh_bias) step_output.append(cur_hidden) if reverse: step_output.reverse() out = torch.cat(step_output, 0) return out, cur_hidden.squeeze(0) def mkldnn_one_layer_lstm(inp, hidden, params, has_biases, reverse=False): w0 = params[0] w1 = params[1] if has_biases: w2 = params[2] w3 = params[3] else: w2 = torch.zeros(w0.size()) w3 = torch.zeros(w1.size()) hx = hidden[0].unsqueeze(0) cx = hidden[1].unsqueeze(0) batch_sizes: List[int] = [] mode = 2 # third_party/ideep/include/ideep/abstract_types.hpp: ideep::rnn_kind::LSTM = 2 hidden_size = hx.size(2) num_layers = 1 # _rnn_helper already handles bidirectional and batch_first so we hard-code them to False here bidirectional = False batch_first = False train = False # If batch_first, inp has been permuted in _rnn_helper. Convert to contiguous here. # Same as aten/src/ATen/native/mkldnn/RNN.cpp: mkldnn_rnn: input = input.contiguous(); inp = inp.contiguous() hx = hx.contiguous() cx = cx.contiguous() outputs = torch.ops.aten.mkldnn_rnn_layer.default( inp, w0, w1, w2, w3, hx, cx, reverse, batch_sizes, mode, hidden_size, num_layers, has_biases, bidirectional, batch_first, train, ) y, hy, cy = outputs[0], outputs[1], outputs[2] return y, (hy.squeeze(0), cy.squeeze(0)) def _rnn_helper( input, hidden, params, has_biases, num_layers, dropout, train, bidirectional, batch_first, layer_fn, ): input = input.transpose(0, 1) if batch_first else input final_hiddens = [] for i in range(num_layers): cur_params, cur_hidden, bidir_params, bidir_hidden = params_hiddens( params, hidden, i, bidirectional ) dropout = dropout if (train and num_layers < i - 1) else 0.0 fwd_inp, fwd_hidden = layer_fn(input, cur_hidden, cur_params, has_biases) final_hiddens.append(fwd_hidden) if bidirectional: bwd_inp, bwd_hidden = layer_fn( input, bidir_hidden, bidir_params, has_biases, reverse=True ) final_hiddens.append(bwd_hidden) if bidirectional: input = torch.cat([fwd_inp, bwd_inp], fwd_inp.dim() - 1) # type: ignore[possibly-undefined] else: input = fwd_inp if dropout != 0 and train and i < num_layers - 1: input = torch.dropout(input, dropout, train=True) input = input.transpose(0, 1) if batch_first else input return input, final_hiddens @register_decomposition(aten.rnn_tanh.input) @aten.rnn_tanh.input.py_impl(DispatchKey.CompositeImplicitAutograd) @aten.rnn_tanh.input.py_impl(DispatchKey.Autograd) def rnn_tanh_input( input, hx, params, has_biases, num_layers, dropout, train, bidirectional, batch_first, ): hidden = hx.unbind(0) params = gather_params(params, has_biases, False) out, final_hiddens = _rnn_helper( input, hidden, params, has_biases, num_layers, dropout, train, bidirectional, batch_first, partial(one_layer_rnn, hidden_fn=rnn_cell(torch.tanh)), ) return out, torch.stack(final_hiddens, 0) @register_decomposition(aten.rnn_relu.input) @aten.rnn_relu.input.py_impl(DispatchKey.CompositeImplicitAutograd) @aten.rnn_relu.input.py_impl(DispatchKey.Autograd) def rnn_relu_input( input, hx, params, has_biases, num_layers, dropout, train, bidirectional, batch_first, ): hidden = hx.unbind(0) params = gather_params(params, has_biases, False) out, final_hiddens = _rnn_helper( input, hidden, params, has_biases, num_layers, dropout, train, bidirectional, batch_first, partial(one_layer_rnn, hidden_fn=rnn_cell(torch.relu)), ) return out, torch.stack(final_hiddens, 0) @register_decomposition(aten.rnn_relu.data) @aten.rnn_relu.data.py_impl(DispatchKey.CompositeImplicitAutograd) @aten.rnn_relu.data.py_impl(DispatchKey.Autograd) def rnn_relu_data( data, batch_sizes, hx, params, has_biases, num_layers, dropout, train, bidirectional, ): hidden = hx.unbind(0) params = gather_params(params, has_biases, False) out, final_hiddens = _rnn_helper( data, hidden, params, has_biases, num_layers, dropout, train, bidirectional, False, partial( one_layer_rnn_data, batch_sizes=batch_sizes, hidden_fn=rnn_cell_data(torch.relu), ), ) return out, torch.stack(final_hiddens, 0) @register_decomposition(aten.rnn_tanh.data) @aten.rnn_tanh.data.py_impl(DispatchKey.CompositeImplicitAutograd) @aten.rnn_tanh.data.py_impl(DispatchKey.Autograd) def rnn_tanh_data( data, batch_sizes, hx, params, has_biases, num_layers, dropout, train, bidirectional, ): hidden = hx.unbind(0) params = gather_params(params, has_biases, False) out, final_hiddens = _rnn_helper( data, hidden, params, has_biases, num_layers, dropout, train, bidirectional, False, partial( one_layer_rnn_data, batch_sizes=batch_sizes, hidden_fn=rnn_cell_data(torch.tanh), ), ) return out, torch.stack(final_hiddens, 0) def lstm_cell(inp, hx, cx, hh_weight, hh_bias, hr_weight, chunk_dim): gates = F.linear(hx, hh_weight, hh_bias) + inp chunked_gates = gates.chunk(4, chunk_dim) in_gate = chunked_gates[0].sigmoid() forget_gate = chunked_gates[1].sigmoid() cell_gate = chunked_gates[2].tanh() out_gate = chunked_gates[3].sigmoid() cy = forget_gate * cx + (in_gate * cell_gate) hy = out_gate * cy.tanh() hy = hy if hr_weight is None else F.linear(hy, hr_weight, None) return hy, cy def one_layer_lstm(inp, hidden, params, has_biases, reverse=False): ih_weight = params[0] hh_weight = params[1] ih_bias = params[2] if has_biases else None hh_bias = params[3] if has_biases else None hr_weight = ( params[4] if len(params) == 5 else params[2] if len(params) == 3 else None ) hx = hidden[0].unsqueeze(0) cx = hidden[1].unsqueeze(0) precomputed_input = F.linear(inp, ih_weight, ih_bias) precomputed_input = precomputed_input.flip(0) if reverse else precomputed_input step_output = [] for inp in precomputed_input: hx, cx = lstm_cell(inp, hx, cx, hh_weight, hh_bias, hr_weight, chunk_dim=2) step_output.append(hx) if reverse: step_output.reverse() out = torch.cat(step_output, 0) return out, (hx.squeeze(1), cx.squeeze(1)) def one_layer_lstm_data(inp, hidden, params, has_biases, batch_sizes, reverse=False): ih_weight = params[0] hh_weight = params[1] ih_bias = params[2] if has_biases else None hh_bias = params[3] if has_biases else None hr_weight = ( params[4] if len(params) == 5 else params[2] if len(params) == 3 else None ) step_output = [] hiddens = [] last_batch_size = batch_sizes[-1] if reverse else batch_sizes[0] split_inp = torch.split(inp, list(batch_sizes)) if reverse: split_inp = split_inp[::-1] orig_hx = hidden[0] orig_cx = hidden[1] hx, cx = orig_hx.narrow(0, 0, last_batch_size), orig_cx.narrow( 0, 0, last_batch_size ) for inp in split_inp: i = inp.shape[0] inp = F.linear(inp, ih_weight, ih_bias) # this will only happen when reverse=False, since batch sizes are sorted largest -> smallest if i < last_batch_size: hiddens.append( ( hx.narrow(0, i, last_batch_size - i), cx.narrow(0, i, last_batch_size - i), ) ) hx, cx = hx.narrow(0, 0, i), cx.narrow(0, 0, i) # this will only happen when reverse=True if i > last_batch_size: hx = torch.concat( (hx, orig_hx.narrow(0, last_batch_size, i - last_batch_size)), 0 ) cx = torch.concat( (cx, orig_cx.narrow(0, last_batch_size, i - last_batch_size)), 0 ) hx, cx = lstm_cell(inp, hx, cx, hh_weight, hh_bias, hr_weight, chunk_dim=1) last_batch_size = i step_output.append(hx) if reverse: step_output.reverse() hidden_out = (hx, cx) else: hiddens.append((hx, cx)) hiddens.reverse() hidden0, hidden1 = zip(*hiddens) hidden_out = torch.cat(hidden0, 0), torch.cat(hidden1, 0) out = torch.cat(step_output, 0) return out, hidden_out def select_one_layer_lstm_function(input, hx, params): r"""Check whether we could use decompose lstm with mkldnn_rnn_layer. All the below conditions need to be met: * ``torch._C._get_mkldnn_enabled()`` returns ``True``. * All the input args are on CPU. * The dtypes of args are either torch.float or torch.bfloat16. * Inference. * ``has_projections`` returns ``False``. Args: * input: the input sequence to LSTM * hx: a tuple of the input hidden state and cell state ``(h_0, c_0)`` to LSTM * params: the weight and bias tensors of LSTM """ def use_mkldnn(input, hx, params): if not torch._C._get_mkldnn_enabled(): return False tensors = [input] + list(hx) + list(chain.from_iterable(params)) devices = {t.device for t in tensors} if len(devices) != 1: return False device = devices.pop() if device != torch.device("cpu"): return False # With autocast, possible to have mixed dtype here dtypes = {t.dtype for t in tensors} for dtype in dtypes: if dtype not in [torch.float, torch.bfloat16]: return False if input.requires_grad: return False has_projections = hx[0].size(2) != hx[1].size(2) if has_projections: return False return True # mkldnn_one_layer_lstm does not depend on seq_len while one_layer_lstm # will expand over the seq_len dim if use_mkldnn(input, hx, params): return mkldnn_one_layer_lstm else: return one_layer_lstm @register_decomposition(aten.lstm.input) @aten.lstm.input.py_impl(DispatchKey.CompositeImplicitAutograd) @aten.lstm.input.py_impl(DispatchKey.Autograd) def lstm_impl( input, hx, params, has_biases, num_layers, dropout, train, bidirectional, batch_first, ): assert len(hx) == 2, "lstm expects two hidden states" params = gather_params(params, has_biases, hx[0].size(2) != hx[1].size(2)) hidden = list(zip(hx[0], hx[1])) layer_fn = select_one_layer_lstm_function(input, hx, params) out, final_hiddens = _rnn_helper( input, hidden, params, has_biases, num_layers, dropout, train, bidirectional, batch_first, layer_fn, ) final_hiddens = list(zip(*final_hiddens)) return out, torch.stack(final_hiddens[0], 0), torch.stack(final_hiddens[1], 0) @register_decomposition(aten.lstm.data) @aten.lstm.data.py_impl(DispatchKey.CompositeImplicitAutograd) @aten.lstm.data.py_impl(DispatchKey.Autograd) def lstm_data_impl( data, batch_sizes, hx, params, has_biases, num_layers, dropout, train, bidirectional, ): assert len(hx) == 2, "lstm expects two hidden states" params = gather_params(params, has_biases, hx[0].size(2) != hx[1].size(2)) hidden = list(zip(hx[0], hx[1])) out, final_hiddens = _rnn_helper( data, hidden, params, has_biases, num_layers, dropout, train, bidirectional, False, partial(one_layer_lstm_data, batch_sizes=batch_sizes), ) final_hiddens = list(zip(*final_hiddens)) return out, torch.stack(final_hiddens[0], 0), torch.stack(final_hiddens[1], 0) def gru_cell(inp, cur_hidden, ih_weight, ih_bias, hh_weight, hh_bias): chunked_igates = inp.chunk(3, 1) chunked_hgates = F.linear(cur_hidden, hh_weight, hh_bias).chunk(3, 2) reset_gate = (chunked_hgates[0] + chunked_igates[0]).sigmoid() input_gate = (chunked_hgates[1] + chunked_igates[1]).sigmoid() new_gate = (chunked_igates[2] + (chunked_hgates[2] * reset_gate)).tanh() return (cur_hidden - new_gate) * input_gate + new_gate def gru_cell_data(inp, cur_hidden, ih_weight, ih_bias, hh_weight, hh_bias): chunked_igates = F.linear(inp, ih_weight, ih_bias).chunk(3, 1) chunked_hgates = F.linear(cur_hidden, hh_weight, hh_bias).chunk(3, 1) reset_gate = (chunked_hgates[0] + chunked_igates[0]).sigmoid() input_gate = (chunked_hgates[1] + chunked_igates[1]).sigmoid() new_gate = (chunked_igates[2] + (chunked_hgates[2] * reset_gate)).tanh() return (cur_hidden - new_gate) * input_gate + new_gate @register_decomposition(aten.gru.data) @aten.gru.data.py_impl(DispatchKey.CompositeImplicitAutograd) @aten.gru.data.py_impl(DispatchKey.Autograd) def gru_impl_data( data, batch_sizes, hx, params, has_biases, num_layers, dropout, train, bidirectional, ): params = gather_params(params, has_biases, False) out, final_hiddens = _rnn_helper( data, hx.unbind(0), params, has_biases, num_layers, dropout, train, bidirectional, False, partial(one_layer_rnn_data, batch_sizes=batch_sizes, hidden_fn=gru_cell_data), ) return out, torch.stack(final_hiddens, 0) @register_decomposition(aten.gru.input) @aten.gru.input.py_impl(DispatchKey.CompositeImplicitAutograd) @aten.gru.input.py_impl(DispatchKey.Autograd) def gru_impl( input, hx, params, has_biases, num_layers, dropout, train, bidirectional, batch_first, ): params = gather_params(params, has_biases, False) out, final_hiddens = _rnn_helper( input, hx.unbind(0), params, has_biases, num_layers, dropout, train, bidirectional, batch_first, partial(one_layer_rnn, hidden_fn=gru_cell), ) return out, torch.stack(final_hiddens, 0) @register_decomposition(aten._upsample_bilinear2d_aa.vec) @aten._upsample_bilinear2d_aa.vec.py_impl(DispatchKey.CompositeImplicitAutograd) @aten._upsample_bilinear2d_aa.vec.py_impl(DispatchKey.Autograd) def upsample_bilinear2d_aa_vec(input, output_size, align_corners, scale_factors): osize = upsample_compute_output_size(input.size(), output_size, scale_factors) scale_h = get_scale_value(scale_factors, 0) scale_w = get_scale_value(scale_factors, 1) return torch.ops.aten._upsample_bilinear2d_aa( input, osize, align_corners, scale_h, scale_w ) @register_decomposition(aten._upsample_bicubic2d_aa.vec) @aten._upsample_bicubic2d_aa.vec.py_impl(DispatchKey.CompositeImplicitAutograd) @aten._upsample_bicubic2d_aa.vec.py_impl(DispatchKey.Autograd) def upsample_bicubic2d_aa_vec(input, output_size, align_corners, scale_factors): osize = upsample_compute_output_size(input.size(), output_size, scale_factors) scale_h = get_scale_value(scale_factors, 0) scale_w = get_scale_value(scale_factors, 1) return torch.ops.aten._upsample_bicubic2d_aa( input, osize, align_corners, scale_h, scale_w ) @register_decomposition(aten.upsample_bilinear2d.vec) @register_decomposition(aten.upsample_trilinear3d.vec) @aten.upsample_linear1d.vec.py_impl(DispatchKey.CompositeImplicitAutograd) @aten.upsample_linear1d.vec.py_impl(DispatchKey.Autograd) @aten.upsample_bilinear2d.vec.py_impl(DispatchKey.CompositeImplicitAutograd) @aten.upsample_bilinear2d.vec.py_impl(DispatchKey.Autograd) @aten.upsample_trilinear3d.vec.py_impl(DispatchKey.CompositeImplicitAutograd) @aten.upsample_trilinear3d.vec.py_impl(DispatchKey.Autograd) def _upsample_linear_vec(input, output_size, align_corners, scale_factors): osize = upsample_compute_output_size(input.size(), output_size, scale_factors) scales = scale_factors if scale_factors else [None] * len(osize) return _upsample_linear(input, osize, align_corners, scales) @register_decomposition([aten.upsample_linear1d.default, aten.upsample_linear1d.out]) @out_wrapper() def upsample_linear1d( input: Tensor, output_size: List[int], align_corners: bool, scales_w: Optional[float] = None, ) -> Tensor: return _upsample_linear(input, output_size, align_corners, [scales_w]) @register_decomposition( [aten.upsample_bilinear2d.default, aten.upsample_bilinear2d.out] ) @aten.upsample_bilinear2d.default.py_impl(DispatchKey.Autograd) @out_wrapper() def upsample_bilinear2d( input: Tensor, output_size: List[int], align_corners: bool, scales_h: Optional[float] = None, scales_w: Optional[float] = None, ) -> Tensor: return _upsample_linear(input, output_size, align_corners, [scales_h, scales_w]) @register_decomposition( [aten.upsample_trilinear3d.default, aten.upsample_trilinear3d.out] ) @out_wrapper() def upsample_trilinear3d( input: Tensor, output_size: List[int], align_corners: bool, scales_d: Optional[float] = None, scales_h: Optional[float] = None, scales_w: Optional[float] = None, ) -> Tensor: return _upsample_linear( input, output_size, align_corners, [scales_d, scales_h, scales_w] ) def _compute_scale(in_size, out_size, align_corners, scale=None): if align_corners: return (in_size - 1.0) / (out_size - 1.0) if out_size > 1 else 0 else: return 1.0 / scale if scale is not None and scale > 0 else in_size / out_size def _compute_source_index(scale, dst_index, align_corners): if align_corners: return scale * dst_index else: return scale * (dst_index + 0.5) - 0.5 @pw_cast_for_opmath def _upsample_linear( input: Tensor, output_size: List[int], align_corners: bool, scales: List[Optional[float]], ) -> Tensor: # get dimensions of original image n_batch, n_channels = input.shape[:2] inp_sizes = input.shape[2:] n_dims = len(inp_sizes) _, dtype = utils.elementwise_dtypes( input, type_promotion_kind=utils.ELEMENTWISE_TYPE_PROMOTION_KIND.INT_TO_FLOAT, ) def get_values(inp_size, out_size, scales, nsqueeze): # First Calculate scaling factor scale_factor = _compute_scale(inp_size, out_size, align_corners, scales) # We have to create arange with int64 dtype and use .to in order to avoid # additional kernels creation in inductor and get a perf slowdown i = torch.arange(out_size, device=input.device).to(dtype=dtype) x_f32 = _compute_source_index(scale_factor, i, align_corners).clamp(min=0.0) x_f32 = x_f32.reshape(x_f32.shape[0], *[1] * (nsqueeze)) x = x_f32.to(torch.int64) xp1 = (x + 1).clamp(max=inp_size - 1) return x_f32, x, xp1 values = [ get_values(inp_size, out_size, scales, n_dims - 1 - i) for i, (inp_size, out_size, scales) in enumerate( zip(inp_sizes, output_size, scales) ) ] xs_f32, xs, xp1s = list(zip(*values)) vs = [] for a in product(*[[0, 1]] * n_dims): idx = [None, None] + [xs[k] if a[k] == 0 else xp1s[k] for k in range(n_dims)] v = aten._unsafe_index(input, idx) v = _maybe_convert_to_dtype(v, dtype) vs.append(v) for i in reversed(range(n_dims)): xscale = (xs_f32[i] - xs[i]).clamp(0.0, 1.0).to(dtype) vs = [ # x1 * (1 - alpha) + x2 * alpha == x1 + (x2 - x1) * alpha v1 + torch.mul(v2 - v1, xscale) for v1, v2 in zip(vs[::2], vs[1::2]) ] assert len(vs) == 1 result = vs[0] # convert output to correct memory format, if necessary memory_format = utils.suggest_memory_format(input) # following "heuristic: only use channels_last path when it's faster than the contiguous path" if input.device.type == "cuda" and n_channels < 16: memory_format = torch.contiguous_format assert isinstance(result, torch.Tensor) result = result.contiguous(memory_format=memory_format) if not input.is_floating_point(): result = result.round() return result # We should be applying decompositions after all transformations @register_decomposition(aten.is_same_size.default) def is_same_size(a: Tensor, b: Tensor) -> bool: return a.shape == b.shape @register_decomposition([aten._reshape_alias, aten._unsafe_view]) @out_wrapper() def _reshape_alias(x, shape, *args): return aten.view(x, shape) @register_decomposition([aten._unsafe_index]) def _index(x, indices): return aten.index(x, indices) def _nll_loss_forward( self: Tensor, target: Tensor, weight: Optional[Tensor], reduction: int, ignore_index: int, ) -> Tuple[Tensor, Tensor]: # self can be [N, C] or [C] # target can be [N] or [] n_dims = self.dim() channel_dim = 1 if n_dims < 2: channel_dim = 0 if weight is not None: if n_dims > 1: shape = [ 1, ] * n_dims shape[channel_dim] = weight.shape[0] w = weight.view(shape) else: w = weight self = self * w safe_target = torch.where(target != ignore_index, target, 0) safe_target_ = safe_target.unsqueeze(channel_dim) # target can be [N, 1] or [1] result = -torch.gather(self, channel_dim, safe_target_).squeeze(channel_dim) result = torch.where(target != ignore_index, result, 0) if reduction == Reduction.NONE.value and n_dims > 1: total_weight = self.new_full((), 0.0) return result, total_weight if weight is not None: w = w.expand(self.shape) wsum = torch.gather(w, channel_dim, safe_target_).squeeze(channel_dim) wsum = torch.where(target != ignore_index, wsum, 0) total_weight = wsum.sum() else: total_weight = (target != ignore_index).sum().to(self) if reduction == Reduction.SUM.value: result = result.sum() elif reduction == Reduction.MEAN.value: result = result.sum() / total_weight return result, total_weight @register_decomposition(aten.nll_loss_forward) @out_wrapper("output", "total_weight") def nll_loss_forward( self: Tensor, target: Tensor, weight: Optional[Tensor], reduction: int, ignore_index: int, ) -> Tuple[Tensor, Tensor]: assert self.dim() > 0 and self.dim() <= 2, "input tensor should be 1D or 2D" assert ( target.dim() <= 1 ), "0D or 1D target tensor expected, multi-target not supported" no_batch_dim = self.dim() == 1 and target.dim() == 0 assert no_batch_dim or ( self.shape[0] == target.shape[0] ), f"size mismatch (got input: {self.shape}, target: {target.shape})" n_classes = self.shape[-1] assert weight is None or ( weight.dim() == 1 and weight.numel() == n_classes ), f"weight tensor should be defined either for all {n_classes} classes or no classes but got weight tensor of shape: {weight.shape}" # noqa: B950 return _nll_loss_forward(self, target, weight, reduction, ignore_index) @register_decomposition(aten.nll_loss2d_forward) @out_wrapper("output", "total_weight") def nll_loss2d_forward( self: Tensor, target: Tensor, weight: Optional[Tensor], reduction: int, ignore_index: int, ) -> Tuple[Tensor, Tensor]: return _nll_loss_forward(self, target, weight, reduction, ignore_index) # These are adapted from aten/src/ATen/native/UpSample.h, wich is based on # https://en.wikipedia.org/wiki/Bicubic_interpolation#Bicubic_convolution_algorithm def _upsample_cubic_convolution1(x: Tensor, A: float) -> Tensor: return ((A + 2) * x - (A + 3)) * x * x + 1 def _upsample_cubic_convolution2(x: Tensor, A: float) -> Tensor: return ((A * x - 5 * A) * x + 8 * A) * x - 4 * A def _upsample_get_cubic_coefficients(t: Tensor) -> TensorSequenceType: A = -0.75 return ( _upsample_cubic_convolution2(t + 1.0, A), _upsample_cubic_convolution1(t, A), _upsample_cubic_convolution1(1.0 - t, A), _upsample_cubic_convolution2(2.0 - t, A), ) def _upsample_cubic_interp1d(coeffs: TensorSequenceType, ts: Tensor) -> Tensor: coeffs2 = _upsample_get_cubic_coefficients(ts) return _sum_tensors(c1 * c2 for (c1, c2) in zip(coeffs, coeffs2)) # Need this instead of just sum() to keep mypy happy def _sum_tensors(ts: Iterable[Tensor]) -> Tensor: return reduce(torch.add, ts) def _linspace_from_neg_one( num_steps: int, align_corners: bool, dtype: torch.dtype, device: torch.device ): if num_steps <= 1: return torch.tensor(0, device=device, dtype=dtype) a = ((num_steps - 1) / num_steps) if not align_corners else 1 return torch.linspace(-a, a, steps=num_steps, device=device, dtype=dtype) def _make_base_grid_4d(theta: Tensor, h: int, w: int, align_corners: bool): dtype = theta.dtype device = theta.device # Using padding and summation generates a single kernel vs using torch.stack where 3 kernels generated # corresponding to each individual tensor: grid_x, grid_y, grid_one grid_x = _linspace_from_neg_one(w, align_corners, dtype, device).view(1, w, 1) grid_y = _linspace_from_neg_one(h, align_corners, dtype, device).view(h, 1, 1) grid_one = torch.ones((1, 1, 1), dtype=dtype, device=device) # this is just a temporary hack and we should use torch.stack here once #104480 is merged grid_x = torch.nn.functional.pad(grid_x, pad=(0, 2), mode="constant", value=0) grid_y = torch.nn.functional.pad(grid_y, pad=(1, 1), mode="constant", value=0) grid_one = torch.nn.functional.pad(grid_one, pad=(2, 0), mode="constant", value=0) return grid_x + grid_y + grid_one def _make_base_grid_5d(theta: Tensor, d: int, h: int, w: int, align_corners: bool): dtype = theta.dtype device = theta.device grid_x = _linspace_from_neg_one(w, align_corners, dtype, device).view(1, 1, w, 1) grid_y = _linspace_from_neg_one(h, align_corners, dtype, device).view(1, h, 1, 1) grid_z = _linspace_from_neg_one(d, align_corners, dtype, device).view(d, 1, 1, 1) grid_one = torch.ones((1, 1, 1, 1), dtype=dtype, device=device) # this is just a temporary hack and we should use torch.stack here once #104480 is merged grid_x = torch.nn.functional.pad(grid_x, pad=(0, 3), mode="constant", value=0) grid_y = torch.nn.functional.pad(grid_y, pad=(1, 2), mode="constant", value=0) grid_z = torch.nn.functional.pad(grid_z, pad=(2, 1), mode="constant", value=0) grid_one = torch.nn.functional.pad(grid_one, pad=(3, 0), mode="constant", value=0) return grid_x + grid_y + grid_z + grid_one def _affine_grid_generator_4d(theta: Tensor, size: List[int], align_corners: bool): n, _, h, w = size base_grid = _make_base_grid_4d(theta, h, w, align_corners=align_corners) # base_grid shape is (h, w, 3) and theta shape is (n, 2, 3) # We do manually a matrix multiplication which is faster than mm() # (h * w, 3, 1) * (n, 1, 3, 2) -> (n, h * w, 2) grid = (base_grid.view(-1, 3, 1) * theta.mT.unsqueeze(1)).sum(-2) return grid.view(n, h, w, 2) def _affine_grid_generator_5d(theta: Tensor, size: List[int], align_corners: bool): n, _, d, h, w = size base_grid = _make_base_grid_5d(theta, d, h, w, align_corners=align_corners) # base_grid shape is (d, h, w, 4) and theta shape is (n, 3, 4) # We do manually a matrix multiplication which is faster than mm() # (d * h * w, 4, 1) * (n, 1, 4, 3) -> (n, h * w, 3) grid = (base_grid.view(-1, 4, 1) * theta.mT.unsqueeze(1)).sum(-2) return grid.view(n, d, h, w, 3) @register_decomposition(aten.affine_grid_generator) @out_wrapper() @pw_cast_for_opmath def affine_grid_generator(theta: Tensor, size: List[int], align_corners: bool): torch._check( len(size) in (4, 5), lambda: "affine_grid_generator needs 4d (spatial) or 5d (volumetric) inputs.", ) if len(size) == 4: return _affine_grid_generator_4d(theta, size, align_corners=align_corners) else: return _affine_grid_generator_5d(theta, size, align_corners=align_corners) def _grid_sampler_2d( a: Tensor, grid: Tensor, interpolation_mode: int = 0, padding_mode: int = 0, align_corners: bool = False, _expand_grid: bool = True, ) -> Tensor: # This method is a copy of grid_sampler_2d implementation and introduced with additional arg _expand_grid to # optionally expand the input grid for performance reasons. # Experimenting locally it was found that compiled CUDA code is accelerated by ~5x # and CPU code by ~2x on bicubic mode, if we expand the grid from (N, H, W, 2) into (N, C, H, W, 2) # However, this leads to a slowdown around ~0.8x on CPU bilinear mode, channels first. # Thus we apply this hack to not expand the grid for this case. torch._check( interpolation_mode in (0, 1, 2), lambda: f"Invalid interpolation mode {interpolation_mode}", ) torch._check( padding_mode in (0, 1, 2), lambda: f"Invalid padding mode {padding_mode}" ) def unnormalize(coords: Tensor, size: int) -> Tensor: # Rescale coordinates from [-1, 1] to: # [0, size - 1] if align_corners is True # [-.5, size -.5] if align_corners is False mul = (size * 0.5 - 0.5) if align_corners else (size * 0.5) ofs = size * 0.5 - 0.5 return coords * mul + ofs # Reflects coordinates until they fall between low and high (inclusive). # The bounds are passed as twice their value so that half-integer values # can be represented as ints. def reflect_coordinates(coords: Tensor, twice_low: int, twice_high: int) -> Tensor: if twice_low == twice_high: return torch.zeros_like(coords) coords_min = twice_low / 2 coords_span = (twice_high - twice_low) / 2 coords2 = (coords - coords_min).abs() extra = torch.fmod(coords2, coords_span) flips = (coords2 / coords_span).floor().to(dtype=torch.int8) return torch.where( flips & 1 == 0, extra + coords_min, coords_span + coords_min - extra ) def compute_coordinates(coords: Tensor, size: int) -> Tensor: if padding_mode == 0: # Zero return coords elif padding_mode == 1: # Borders return torch.clamp(coords, 0, size - 1) else: # padding_mode == 2, Reflection if align_corners: coords_reflected = reflect_coordinates(coords, 0, 2 * (size - 1)) else: coords_reflected = reflect_coordinates(coords, -1, 2 * size - 1) return torch.clamp(coords_reflected, 0, size - 1) def compute_source_index(coords: Tensor, size: int) -> Tensor: coords_un = unnormalize(coords, size) return compute_coordinates(coords_un, size) N, C, iH, iW = a.shape _, oH, oW, two = grid.shape assert two == 2 if _expand_grid: # Let's expand grid to [N, C, oH, oW, 2] # This allows to generate a single triton cuda kernel instead of two kernels. # Two kernels are due source indices, weights have shape (N, 1, oH, oW), xnumel=N*oH*oW # and output has shape (N, C, oH, oW), xnumel=N*C*oH*oW # Expanding grid to (N, C, oH, oW, two) unifies xnumel to N*C*oH*oW grid = grid.view(N, 1, oH, oW, two).expand(N, C, oH, oW, 2) def in_bounds_cond(xs: Tensor, ys: Tensor) -> Tensor: return torch.logical_and( 0 <= xs, torch.logical_and(xs < iW, torch.logical_and(0 <= ys, ys < iH)) ) N_idx = torch.arange(N, device=a.device).view(N, 1, 1, 1) C_idx = torch.arange(C, device=a.device).view(1, C, 1, 1) def clip(xs: Tensor, ys: Tensor, ws: Tensor) -> TensorSequenceType: cond = in_bounds_cond(xs, ys) # To clip to inside valid coordinates, we map the coordinates # to (x, y) = (0, 0) and also set the weight to 0 # We also change the shape of the tensor to the appropriate one for # broadcasting with N_idx, C_idx for the purposes of advanced indexing c = C if _expand_grid else 1 return tuple( torch.where(cond, t, 0).view(N, c, oH, oW) for t in (xs.to(dtype=torch.int64), ys.to(dtype=torch.int64), ws) ) def get_summand(ix: Tensor, iy: Tensor, w) -> Tensor: # Perform clipping, index into input tensor and multiply by weight idx_x, idx_y, w_ = clip(ix, iy, w) return a[N_idx, C_idx, idx_y, idx_x] * w_ x = grid[..., 0] y = grid[..., 1] if interpolation_mode == 0: # Bilinear ix = compute_source_index(x, iW) iy = compute_source_index(y, iH) ix_nw, iy_nw = ix.floor(), iy.floor() ix_ne, iy_ne = ix_nw + 1, iy_nw ix_sw, iy_sw = ix_nw, iy_nw + 1 ix_se, iy_se = ix_ne, iy_sw w_nw = (ix_se - ix) * (iy_se - iy) w_ne = (ix - ix_sw) * (iy_sw - iy) w_sw = (ix_ne - ix) * (iy - iy_ne) w_se = (ix - ix_nw) * (iy - iy_nw) return _sum_tensors( get_summand(ix, iy, w) for (ix, iy, w) in ( (ix_nw, iy_nw, w_nw), (ix_ne, iy_ne, w_ne), (ix_sw, iy_sw, w_sw), (ix_se, iy_se, w_se), ) ) elif interpolation_mode == 1: # Nearest ix = compute_source_index(x, iW) iy = compute_source_index(y, iH) ix_nearest = ix.round() iy_nearest = iy.round() return get_summand(ix_nearest, iy_nearest, 1) else: # interpolation_mode == 2, Bicubic ix = unnormalize(x, iW) iy = unnormalize(y, iH) ix_nw = ix.floor() iy_nw = iy.floor() tx = ix - ix_nw ty = iy - iy_nw if not _expand_grid: tx = tx.unsqueeze(1) ty = ty.unsqueeze(1) def get_value_bounded(ix: Tensor, iy: Tensor) -> Tensor: x = compute_coordinates(ix, iW) y = compute_coordinates(iy, iH) return get_summand(x, y, 1) def get_coeff(ofs: int) -> Tensor: iy_ofs = iy_nw + (ofs - 1) cs = ( get_value_bounded(ix_nw - 1, iy_ofs), get_value_bounded(ix_nw, iy_ofs), get_value_bounded(ix_nw + 1, iy_ofs), get_value_bounded(ix_nw + 2, iy_ofs), ) return _upsample_cubic_interp1d(cs, tx) coeffs = tuple(get_coeff(ofs) for ofs in range(4)) return _upsample_cubic_interp1d(coeffs, ty) @register_decomposition(aten.grid_sampler_2d) @out_wrapper() @pw_cast_for_opmath def grid_sampler_2d( a: Tensor, grid: Tensor, interpolation_mode: int = 0, padding_mode: int = 0, align_corners: bool = False, ) -> Tensor: return _grid_sampler_2d( a, grid=grid, interpolation_mode=interpolation_mode, padding_mode=padding_mode, align_corners=align_corners, ) @register_decomposition(aten.mv) @out_wrapper() @pw_cast_for_opmath def mv(self, vec): torch._check( self.dim() == 2 and vec.dim() == 1, lambda: f"matrix @ vector expected, got {self.dim()}, {vec.dim()}", ) torch._check( self.size(1) == vec.size(0), lambda: f"size mismatch, got input ({self.size(0)}x{self.size(1)}), vec ({vec.size(0)})", ) return (self * vec).sum(dim=1) @register_decomposition(aten.binary_cross_entropy_with_logits) @out_wrapper() def binary_cross_entropy_with_logits( self, target, weight=None, pos_weight=None, reduction=Reduction.MEAN.value ): if pos_weight is not None: log_weight = (pos_weight - 1) * target + 1 loss = (1 - target) * self - (log_weight * F.logsigmoid(self)) else: loss = (1 - target) * self - F.logsigmoid(self) if weight is not None: loss = loss * weight return apply_loss_reduction(loss, reduction) def should_fold(tensor1: torch.Tensor, tensor2: torch.Tensor, is_out: bool) -> bool: # For comments of the logic of this function see eager in /native/LinearAlgebra.cpp t1, t2 = (tensor1, tensor2) if tensor1.ndim >= tensor2.ndim else (tensor2, tensor1) from torch.fx.experimental.symbolic_shapes import guard_size_oblivious if not (t1.ndim >= 3 and t2.ndim <= 2): return False if t2.requires_grad and not is_out: return True if tensor1.ndim == 2: return False if guard_size_oblivious(t1.numel() == 0): return True t1_shape = t1.shape t1_stride = t1.stride() return all( st1 == st2 * s2 for (st1, st2, s2) in zip(t1_stride[:-2], t1_stride[1:-1], t1_shape[1:-1]) ) @aten.matmul.default.py_impl(DispatchKey.CompositeImplicitAutograd) @out_wrapper(pass_is_out=True) def matmul(tensor1, tensor2, *, is_out=False): dim_tensor1 = tensor1.dim() dim_tensor2 = tensor2.dim() assert dim_tensor1 != 0 and dim_tensor2 != 0 if dim_tensor1 == 1 and dim_tensor2 == 1: return torch.dot(tensor1, tensor2) elif dim_tensor1 == 2 and dim_tensor2 == 1: return torch.mv(tensor1, tensor2) elif dim_tensor1 == 1 and dim_tensor2 == 2: return torch.squeeze(torch.mm(torch.unsqueeze(tensor1, 0), tensor2), 0) elif dim_tensor1 == 2 and dim_tensor2 == 2: return torch.mm(tensor1, tensor2) elif should_fold(tensor1, tensor2, is_out): # dim_tensor1 >=3 && (dim_tensor2 == 1 || dim_tensor2 == 2) || # dim_tensor2 >=3 && (dim_tensor1 == 1 || dim_tensor1 == 2) # and some condition on the strides is fulfilled # optimization: use mm instead of bmm by folding the batch of the larger tensor # into its leading matrix dimension transpose = dim_tensor2 > dim_tensor1 t1 = tensor2.mT if transpose else tensor1 t2 = ( tensor2 if not transpose else (tensor1.t() if dim_tensor1 == 2 else tensor1) ) # Invariant: t1.dim() >= 3 && (t2.dim() == 1 || t2.dim() == 2) # and t1 and t2 are matmul-compatible # Why not t1.view(-1, sizes_1[-1])? # If the last dim is 0, then view(-1, 0) won't work because the -1 becomes ambiguous. # This can happen in e.g. [3, 5, 0] @ [0, 0]. sizes_1 = t1.shape output_shape = list(sizes_1[:-1]) folded_dim1 = reduce(operator.mul, output_shape) # Readjust output_shape if we are multiplying by a matrix t2_is_matrix = t2.dim() == 2 if t2_is_matrix: output_shape.append(t2.shape[1]) # This will almost always be a view. # It may not be a view if t2->requires_grad(). See should_fold in aten/ for an explanation t1_folded = t1.reshape(folded_dim1, sizes_1[-1]) if t2_is_matrix: # This copies if we perform a 2D @ 3D and the first tensor requires_grad # See should_fold native/LinearAlgebra.cpp for why. output = t1_folded.mm(t2).view(output_shape) return output.mT.contiguous() if transpose else output else: return t1_folded.mv(t2).view(output_shape) elif dim_tensor1 >= 1 and dim_tensor2 >= 1: # We are multiplying b1 x n x m1 by x2 x m2 x p (where b1 can be a list); # we track m1 vs m2 separately even though they must match for nicer error messages n = tensor1.size(-2) if dim_tensor1 > 1 else 1 m1 = tensor1.size(-1) batch_tensor1 = tensor1.shape[:-2] m2 = tensor2.size(-2) if dim_tensor2 > 1 else tensor2.size(-1) p = tensor2.size(-1) if dim_tensor2 > 1 else 1 batch_tensor2: List[int] = [] # TODO: handling of slice for i in range(dim_tensor2 - 2): batch_tensor2.append(tensor2.size(i)) # Same optimization for the gradients as that in should_fold # If we're going to broadcast, we force it to go through the should_fold branch if ( dim_tensor1 == 3 and dim_tensor2 == 3 and batch_tensor1[0] != batch_tensor2[0] ): if batch_tensor1[0] == 1 and tensor1.requires_grad: return matmul(tensor1.squeeze(0), tensor2) if batch_tensor2[0] == 1 and tensor2.requires_grad: return matmul(tensor1, tensor2.squeeze(0)) # expand the batch portion (i.e. cut off matrix dimensions and expand rest) expand_batch_portion = list( torch.broadcast_shapes(batch_tensor1, batch_tensor2) ) tensor1_expand_size = expand_batch_portion + [n, m1] expand_batch_product = prod(expand_batch_portion) # HACK: We need reshape with symint support tensor1_expanded = tensor1.expand(tensor1_expand_size).reshape( expand_batch_product, n, m1 ) vector_rhs = dim_tensor2 == 1 if vector_rhs: tensor2_expand_size = expand_batch_portion + [m2] tensor2_expanded = ( tensor2.expand(tensor2_expand_size) .reshape(expand_batch_product, m2) .unsqueeze(2) ) else: tensor2_expand_size = expand_batch_portion + [m2, p] tensor2_expanded = tensor2.expand(tensor2_expand_size).reshape( expand_batch_product, m2, p ) output_shape = expand_batch_portion if dim_tensor1 > 1: output_shape.append(n) if dim_tensor2 > 1: output_shape.append(p) if vector_rhs: return tensor1_expanded.bmm(tensor2_expanded).squeeze(-1).view(output_shape) else: return tensor1_expanded.bmm(tensor2_expanded).view(output_shape) else: torch._check(False, lambda: "both arguments to matmul need to be at least 1D") @register_decomposition(aten.upsample_bicubic2d.default) @pw_cast_for_opmath def upsample_bicubic2d_default( a: Tensor, output_size: Tuple[int, int], align_corners: bool, scale_h: Optional[float] = None, scale_w: Optional[float] = None, ) -> Tensor: N, C, iH, iW = a.shape oH, oW = output_size def compute_scale(in_size, out_size, align_corners, scale=None): if align_corners: return (in_size - 1) / (out_size - 1) if out_size > 1 else 0 else: return 1 / scale if scale is not None and scale > 0 else in_size / out_size def compute_source_index(scale, dst_index, align_corners): if align_corners: return scale * dst_index else: return scale * (dst_index + 0.5) - 0.5 height_scale = compute_scale(iH, oH, align_corners, scale_h) width_scale = compute_scale(iW, oW, align_corners, scale_w) N_idx = torch.arange(N, device=a.device).view(N, 1, 1, 1) C_idx = torch.arange(C, device=a.device).view(1, C, 1, 1) out_y = torch.arange(oH, device=a.device).view((1, 1, oH, 1)) out_x = torch.arange(oW, device=a.device).view((1, 1, 1, oW)) real_x = compute_source_index(width_scale, out_x, align_corners) in_x = real_x.floor() t_x = real_x - in_x ix = in_x.to(dtype=torch.int64) real_y = compute_source_index(height_scale, out_y, align_corners) in_y = real_y.floor() t_y = real_y - in_y iy = in_y.to(dtype=torch.int64) iys_ofs = (iy - 1, iy, iy + 1, iy + 2) ixs_ofs = (ix - 1, ix, ix + 1, ix + 2) def load_bounded(ys, xs): y_idx = torch.clamp(ys, 0, iH - 1) x_idx = torch.clamp(xs, 0, iW - 1) return aten._unsafe_index(a, [N_idx, C_idx, y_idx, x_idx]) def get_x_interp(y): coeffs_x = tuple(load_bounded(y, x_ofs) for x_ofs in ixs_ofs) return _upsample_cubic_interp1d(coeffs_x, t_x) coeffs_y = tuple(get_x_interp(y_ofs) for y_ofs in iys_ofs) result = _upsample_cubic_interp1d(coeffs_y, t_y) # convert output to correct memory format, if necessary memory_format = utils.suggest_memory_format(a) result = result.contiguous(memory_format=memory_format) return result @register_decomposition(aten.upsample_bicubic2d.vec) @aten.upsample_bicubic2d.vec.py_impl(DispatchKey.CompositeImplicitAutograd) @aten.upsample_bicubic2d.vec.py_impl(DispatchKey.Autograd) @out_wrapper() @pw_cast_for_opmath def upsample_bicubic2d_vec( a: Tensor, output_size: Optional[Tuple[int, int]], align_corners: bool, scale_factors: Optional[Tuple[float, float]] = None, ) -> Tensor: torch._check( bool(output_size) + bool(scale_factors) == 1, lambda: "Must specify exactly one of output_size and scale_factors.", ) if output_size is None: assert scale_factors is not None output_size = cast( Tuple[int, int], tuple( sym_int(sym_float(w) * scale) for w, scale in zip(a.shape[2:], scale_factors) ), ) scale_h, scale_w = scale_factors if scale_factors else (None, None) return upsample_bicubic2d_default(a, output_size, align_corners, scale_h, scale_w) @register_decomposition(aten.reflection_pad1d) @register_decomposition(aten.reflection_pad2d) @register_decomposition(aten.reflection_pad3d) @pw_cast_for_opmath @out_wrapper() def _reflection_pad(a: Tensor, padding: Tuple[int, ...]) -> Tensor: def idx(left, middle, right): dim_idx = torch.arange(-left, middle + right, device=a.device) return middle - 1 - (middle - 1 - dim_idx.abs()).abs() return _reflection_or_replication_pad( a, padding, idx, ) @register_decomposition(aten.replication_pad1d) @register_decomposition(aten.replication_pad2d) @register_decomposition(aten.replication_pad3d) @pw_cast_for_opmath @out_wrapper() def _replication_pad(a: Tensor, padding: Tuple[int, ...]) -> Tensor: def idx(left, middle, right): dim_idx = torch.arange(-left, middle + right, device=a.device) return torch.clamp(dim_idx, 0, middle - 1) return _reflection_or_replication_pad( a, padding, idx, ) def _reflection_or_replication_pad( a: Tensor, padding: Tuple[int, ...], idx_fn: Callable[[int, int, int], Tensor], ) -> Tensor: dim = len(padding) // 2 torch._check( a.dim() in (dim + 1, dim + 2), lambda: f"reflection_pad{dim}d requires {dim + 1}D or {dim + 2}D input", ) inp_shape = a.shape[-dim:] nc_dim = a.dim() - dim padding_left = [padding[2 * (dim - 1 - i)] for i in range(dim)] padding_right = [padding[2 * (dim - 1 - i) + 1] for i in range(dim)] result = a for i in range(dim): idx: List[Any] = [None] * result.dim() idx[i + nc_dim] = idx_fn(padding_left[i], inp_shape[i], padding_right[i]) result = aten._unsafe_index(result, idx) # convert output to correct memory format, if necessary memory_format = utils.suggest_memory_format(result) result = result.contiguous(memory_format=memory_format) return result @register_decomposition(aten.aminmax) @out_wrapper("min", "max") def aminmax(self, *, dim=None, keepdim=False): amin = torch.amin(self, dim=dim, keepdim=keepdim) amax = torch.amax(self, dim=dim, keepdim=keepdim) return amin, amax @register_decomposition(aten.nansum) @out_wrapper() def nansum(self, dim=None, keepdim=False, *, dtype=None): return aten.sum(torch.where(torch.isnan(self), 0, self), dim, keepdim, dtype=dtype) @register_decomposition([aten.arange.default, aten.arange.out]) @out_wrapper() def arange_default( end: NumberType, *, dtype: Optional[torch.dtype] = None, layout: torch.layout = torch.strided, device: Optional[torch.device] = None, pin_memory: bool = False, ): return aten.arange.start_step( 0, end, 1, dtype=dtype, layout=layout, device=device, pin_memory=pin_memory ) @register_decomposition([aten.arange.start]) def arange_start( start: NumberType, end: NumberType, *, dtype: Optional[torch.dtype] = None, layout: torch.layout = torch.strided, device: Optional[torch.device] = None, pin_memory: bool = False, ): return aten.arange.start_step( start, end, 1, dtype=dtype, layout=layout, device=device, pin_memory=pin_memory ) @register_decomposition(out_dtype) def out_dtype_decomp(*args, **kwargs): from torch._higher_order_ops.out_dtype import out_dtype_dense return out_dtype_dense(*args, **kwargs) @register_decomposition(aten.multi_margin_loss) @aten.multi_margin_loss.default.py_impl(DispatchKey.Autograd) @out_wrapper() def multi_margin_loss( input: Tensor, target: Tensor, p: NumberType = 1, margin: NumberType = 1, weight: Optional[Tensor] = None, reduction: int = Reduction.MEAN.value, ) -> Tensor: input = torch.atleast_2d(input) target = torch.atleast_1d(target) nframe = input.shape[0] dim = input.shape[1] torch._check(p == 1 or p == 2, lambda: "only p == 1 and p == 2 supported") torch._check( input.ndim == 2 and dim != 0, lambda: f"Expected non-empty vector or matrix with optional 0-dim batch size, but got: {input.shape}", ) torch._check( target.ndim == 1 and target.numel() == nframe, lambda: f"inconsistent target size, expected {nframe} but got {target.shape}", ) if weight is not None: weight = torch.atleast_1d(weight) torch._check( weight.ndim == 1 and weight.numel() == dim, # type: ignore[union-attr] lambda: f"inconsistent weight size, expected {dim} but got {weight.shape}", # type: ignore[union-attr] ) target = target.unsqueeze(1) u = torch.gather(input, dim=1, index=target) z = margin - u + input z = z.clamp_min(0) z = z if p == 1 else z * z if weight is not None: z = z * weight[target] idx = torch.arange(dim, device=input.device) z = torch.where(idx != target, z, 0) if reduction == Reduction.MEAN.value: return z.mean() elif reduction == Reduction.SUM.value: return z.sum() / z.shape[1] else: return z.mean(dim=1) @register_decomposition(aten.multilabel_margin_loss_forward) @aten.multilabel_margin_loss_forward.default.py_impl(DispatchKey.Autograd) @out_wrapper("output", "is_target") def multilabel_margin_loss_forward( input: Tensor, target: Tensor, reduction: int, ) -> Tuple[Tensor, Tensor]: orig_input_shape = input.shape orig_target_shape = target.shape input = torch.atleast_2d(input) target = torch.atleast_2d(target) dim = input.shape[1] torch._check( len(orig_input_shape) <= 2 and dim != 0, lambda: f"Expected non-empty vector or matrix with optional 0-dim batch size, but got: {orig_input_shape}", ) torch._check( len(orig_target_shape) <= 2 and orig_target_shape == orig_input_shape, lambda: f"inconsistent target size: {orig_target_shape} for input of size: {orig_input_shape}", ) # ignores labels after the first -1, detects when -1 is not present idx = torch.arange(dim, device=target.device) is_end = target == -1 end_idx = torch.amin(torch.where(is_end, idx, dim), dim=-1, keepdim=True) # target indices target_mask = idx < end_idx # masks target to be able to use gather, which doesn't allow -1 tidx0 = torch.where(target_mask, target, 0) u = torch.gather(input, dim=-1, index=tidx0) # is_target tidx1 = torch.where(target_mask, target, -1) is_target = torch.any(idx == tidx1.unsqueeze(dim=-1), dim=1) # loss z = 1.0 - u.T.unsqueeze(dim=-1) + input z = z.clamp_min(0) z = z / dim # masks loss z = torch.where(is_target, 0, z) # reduction if reduction == Reduction.MEAN.value: z = z.sum(dim=(0, -1)).mean() elif reduction == Reduction.SUM.value: z = z.sum() else: z = z.sum(dim=(0, -1)) # result is_target = is_target.to(input.dtype).reshape(orig_target_shape) return z, is_target # scaled_dot_product_attention used to be decomposed in pre-autograd, given that # it calls _scaled_dot_product_attention_math and # _scaled_dot_product_attention_math only has a CompositeImplicitAutograd # kernel. As a result it's decomposed into ops with finer granularity. # However recent PRs (#103826 #105131 #115913) added new logic in # scaled_dot_product_attention and now it calls # _scaled_dot_product_flash_attention_for_cpu in export path. This results # in _scaled_dot_product_flash_attention_for_cpu showing up in export result. # This decomposition ensures scaled_dot_product_attention is still decomposed # the same way as before, i.e., going through # _scaled_dot_product_attention_math. Notice that this decomp rule should be # excluded by inductor. @register_decomposition(aten._scaled_dot_product_flash_attention_for_cpu.default) def scaled_dot_product_flash_attention_for_cpu( query: Tensor, key: Tensor, value: Tensor, dropout_p: float = 0.0, is_causal: bool = False, *, attn_mask: Optional[Tensor] = None, scale: Optional[float] = None, ) -> Tuple[Tensor, Tensor]: dtype = query.dtype torch._check( torch.is_floating_point(query), lambda: f"query must be FP32, FP64, BF16, FP16 but got {query.dtype}", ) torch._check( query.dim() == 4 and key.dim() == 4 and value.dim() == 4, lambda: f"q, k, v must be a 4 dimensional tensor, got {query.dim()}, {key.dim()}, {value.dim()}", ) torch._check( dropout_p == 0.0, lambda: f"dropout probability must be zero, got {dropout_p}" ) torch._check( query.shape[3] == value.shape[3] and key.shape[3] == value.shape[3], lambda: "q, k, v should have the same head size", ) output, attn = aten._scaled_dot_product_attention_math.default( query, key, value, attn_mask=attn_mask, dropout_p=dropout_p, is_causal=is_causal, dropout_mask=None, scale=scale, ) # Why this change? # In pre-dispatch export scaled_dot_product_attention is executed via # * flash_attention. # flash_attention allocates output tensor as (N, L, H, E) # it then transposes that to get (N, H, L, E) which is supposed to be the return # tensor dim for scaled_dot_product_attention # assume x: [N, H, L, E] is the output sdpa # In MHA code, this output is then permuted via (2, 0, 1, 3) to get # (L, N, H, E) dim tensor # x = x.permute(2, 0, 1, 3).contiguous() and the viewed via # x = x.view(L * N, H * E) # During pre autograd dispatch call to contiguous is not traced because # flash_attention output after the x.permute is already contiguous # on which the view is valid # However, during 2nd stage export, post-dispatch, we run _match variant # instead of flash* to get the decomposition. _match variant returns # x: [N, H, L, E] applying x.permute(2, 0, 1, 3) returns # x: [L, N, H, E] and without converting this to contiguous tensor # subsequent view is not valid and the export fails # solution is to maintain the return tensor view from the decomp to be # exactly same as *flash* variant. # flash variants output is contiguous as [N, L, H, E] # _match variant out is contiguous as [N, H, L, E] # out = out.transpose(1, 2).contiguous gets output as contiguous # in [N, L, H, E]. # Subsrequent transpose(1, 2) then returns a view on which # aforementioned code snippet, as showm below, is valid # x = x.permute(2, 0, 1, 3).contiguous() and the viewed via # x = x.view(L * N, H * E) # Really the invariant you want to maintain is: # pre-dispatch op-output and its decomposed representation must # return tensor with same view and dims output = output.transpose(1, 2).contiguous(memory_format=torch.contiguous_format) return (output.transpose(1, 2), attn) def register_inplace(aten_op, outplace_op): @register_decomposition(aten_op) def inplace_op(*args, **kwargs): out = outplace_op(*args, **kwargs) return args[0].copy_(out) return inplace_op @register_decomposition([aten.baddbmm]) @out_wrapper() @pw_cast_for_opmath def baddbmm(self, batch1, batch2, beta=1, alpha=1): if not self.is_floating_point() and not self.is_complex(): beta = int(beta) alpha = int(alpha) result = torch.bmm(batch1, batch2) if not isinstance(alpha, numbers.Number) or alpha != 1: result = result * alpha if beta == 0: return result if not isinstance(beta, numbers.Number) or beta != 1: self = self * beta return self + result @register_decomposition(aten.floor_divide) @out_wrapper() def floor_divide(self, other): return torch.div(self, other, rounding_mode="floor") @register_decomposition(aten.sym_numel) def sym_numel(t): return functools.reduce(operator.mul, t.shape, 1) @register_decomposition([aten.sum.default, aten.sum.out]) def sum_default( self: Tensor, *, dtype: Optional[torch.dtype] = None, out: Optional[Tensor] = None, ) -> Tensor: if out is None: return aten.sum.dim_IntList(self, [], dtype=dtype) else: return aten.sum.IntList_out(self, [], dtype=dtype, out=out) @register_decomposition([aten.squeeze.default, aten.squeeze.dim]) def squeeze_default(self: Tensor, dim: Optional[int] = None): if dim is None: return aten.squeeze.dims(self, list(range(self.dim()))) else: return aten.squeeze.dims(self, [dim]) @register_decomposition(torch.ops.aten._weight_norm_interface) def _weight_norm_interface(x, y, dim=0): # https://github.com/pytorch/pytorch/blob/852f8526c52190125446adc9a6ecbcc28fb66182/aten/src/ATen/native/WeightNorm.cpp#L58 keep_dim = tuple(i for i in range(len(x.shape)) if i != dim) norm = x.norm(2, keep_dim, keepdim=True) return x * (y / norm), norm @register_decomposition(aten.isin) @out_wrapper() def isin(elements, test_elements, *, assume_unique=False, invert=False): # handle when either elements or test_elements are Scalars (they can't both be) if not isinstance(elements, torch.Tensor): elements = torch.tensor(elements, device=test_elements.device) if not isinstance(test_elements, torch.Tensor): test_elements = torch.tensor(test_elements, device=elements.device) if test_elements.numel() < 10.0 * pow(elements.numel(), 0.145): return isin_default(elements, test_elements, invert=invert) else: return isin_sorting( elements, test_elements, assume_unique=assume_unique, invert=invert ) def isin_default(elements, test_elements, *, invert=False): if elements.numel() == 0: return torch.empty_like(elements, dtype=torch.bool) x = elements.view(*elements.shape, *((1,) * test_elements.ndim)) if not invert: cmp = x == test_elements else: cmp = x != test_elements dim = tuple(range(-1, -test_elements.ndim - 1, -1)) return cmp.any(dim=dim) def isin_sorting(elements, test_elements, *, assume_unique=False, invert=False): elements_flat = elements.flatten() test_elements_flat = test_elements.flatten() if assume_unique: # This is the same as the aten implementation. For # assume_unique=False, we cannot use unique() here, so we use a # version with searchsorted instead. all_elements = torch.cat([elements_flat, test_elements_flat]) sorted_elements, sorted_order = torch.sort(all_elements, stable=True) duplicate_mask = sorted_elements[1:] == sorted_elements[:-1] duplicate_mask = torch.constant_pad_nd(duplicate_mask, [0, 1], False) if invert: duplicate_mask = duplicate_mask.logical_not() mask = torch.empty_like(duplicate_mask) mask = mask.index_copy(0, sorted_order, duplicate_mask) return mask[0 : elements.numel()] else: sorted_test_elements, _ = torch.sort(test_elements_flat) idx = torch.searchsorted(sorted_test_elements, elements_flat) test_idx = torch.where(idx < sorted_test_elements.numel(), idx, 0) cmp = sorted_test_elements[test_idx] == elements_flat cmp = cmp.logical_not() if invert else cmp return cmp.reshape(elements.shape) @register_decomposition(aten.take) @out_wrapper() def take(self, index): flattened = self.reshape(-1) return flattened[index] register_inplace(aten.addbmm_, aten.addbmm) register_inplace(aten.addmm_, aten.addmm) register_inplace(aten.addmv_, aten.addmv) register_inplace(aten.baddbmm_, aten.baddbmm) register_inplace(aten.fill_, aten.fill) register_inplace(aten.gelu_, aten.gelu) register_inplace(aten.hardswish_, aten.hardswish) register_inplace(aten.hardtanh_, aten.hardtanh) register_inplace(aten.hardsigmoid_, aten.hardsigmoid) register_inplace(aten.__iand__, aten.__and__) register_inplace(aten.__ilshift__, aten.__lshift__) register_inplace(aten.index_put_, aten.index_put) register_inplace(aten.index_reduce_, aten.index_reduce) register_inplace(aten.__ior__, aten.__or__) register_inplace(aten.__irshift__, aten.__rshift__) register_inplace(aten.__ixor__, aten.__xor__) register_inplace(aten.leaky_relu_, aten.leaky_relu) register_inplace(aten.logit_, aten.logit) register_inplace(aten.relu_, aten.relu) register_inplace(aten.renorm_, aten.renorm) register_inplace(aten.round_, aten.round) register_inplace(aten.scatter_, aten.scatter) register_inplace(aten.scatter_add_, aten.scatter_add) register_inplace(aten.scatter_reduce_, aten.scatter_reduce) register_inplace(aten.silu_, aten.silu)