IR

Создано: 13 дек. 2022 | Последнее обновление: 3 дек. 2025

PyTorch 2.0 предлагает два набора IR для взаимодействия с бэкендами: Core Aten IR и Prims IR.

Core Aten IR

Операторы core aten — это основное подмножество операторов aten, из которых можно составлять другие операторы. Core aten IR является полностью функциональным, и в этом наборе операторов нет вариантов inplace или _out. В отличие от Prims IR, операторы core aten повторно используют существующие операторы aten из “native_functions.yaml” и не разлагают операторы на отдельные операции приведения типов и broadcasting. Этот набор операторов предназначен для использования в качестве функционального IR для взаимодействия с бэкендами.

Предупреждение

Этот набор операторов всё ещё активно разрабатывается; в будущем будут добавлены новые операторы.

Оператор	Схема
`aten._adaptive_avg_pool2d`	_adaptive_avg_pool2d(Tensor self, SymInt[2] output_size) -> Tensor
`aten._adaptive_avg_pool2d_backward`	_adaptive_avg_pool2d_backward(Tensor grad_output, Tensor self) -> Tensor
`aten._adaptive_avg_pool3d`	_adaptive_avg_pool3d(Tensor self, SymInt[3] output_size) -> Tensor
`aten._cdist_forward`	_cdist_forward(Tensor x1, Tensor x2, float p, int? compute_mode) -> Tensor
`aten._embedding_bag`	_embedding_bag(Tensor weight, Tensor indices, Tensor offsets, bool scale_grad_by_freq=False, int mode=0, bool sparse=False, Tensor? per_sample_weights=None, bool include_last_offset=False, int padding_idx=-1) -> (Tensor, Tensor, Tensor, Tensor)
`aten._fft_c2r`	_fft_c2r(Tensor self, int[] dim, int normalization, SymInt last_dim_size) -> Tensor
`aten._fft_r2c`	_fft_r2c(Tensor self, int[] dim, int normalization, bool onesided) -> Tensor
`aten._local_scalar_dense`	_local_scalar_dense(Tensor self) -> Scalar
`aten._log_softmax`	_log_softmax(Tensor self, int dim, bool half_to_float) -> Tensor
`aten._native_batch_norm_legit`	_native_batch_norm_legit(Tensor input, Tensor? weight, Tensor? bias, Tensor(a!) running_mean, Tensor(b!) running_var, bool training, float momentum, float eps) -> (Tensor, Tensor, Tensor)
`aten._native_batch_norm_legit.no_stats`	_native_batch_norm_legit.no_stats(Tensor input, Tensor? weight, Tensor? bias, bool training, float momentum, float eps) -> (Tensor, Tensor, Tensor)
`aten._native_batch_norm_legit_no_training`	_native_batch_norm_legit_no_training(Tensor input, Tensor? weight, Tensor? bias, Tensor running_mean, Tensor running_var, float momentum, float eps) -> (Tensor, Tensor, Tensor)
`aten._pdist_forward`	_pdist_forward(Tensor self, float p=2) -> Tensor
`aten._softmax`	_softmax(Tensor self, int dim, bool half_to_float) -> Tensor
`aten._to_copy`	_to_copy(Tensor self, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None, bool non_blocking=False, MemoryFormat? memory_format=None) -> Tensor
`aten.abs`	abs(Tensor self) -> Tensor
`aten.acos`	acos(Tensor self) -> Tensor
`aten.acosh`	acosh(Tensor self) -> Tensor
`aten.adaptive_avg_pool1d`	adaptive_avg_pool1d(Tensor self, int[1] output_size) -> Tensor
`aten.add.Scalar`	add.Scalar(Tensor self, Scalar other, Scalar alpha=1) -> Tensor
`aten.add.Tensor`	add.Tensor(Tensor self, Tensor other, *, Scalar alpha=1) -> Tensor
`aten.addmm`	addmm(Tensor self, Tensor mat1, Tensor mat2, *, Scalar beta=1, Scalar alpha=1) -> Tensor
`aten.alias`	alias(Tensor(a) self) -> Tensor(a)
`aten.amax`	amax(Tensor self, int[1] dim=[], bool keepdim=False) -> Tensor
`aten.amin`	amin(Tensor self, int[1] dim=[], bool keepdim=False) -> Tensor
`aten.any`	any(Tensor self) -> Tensor
`aten.any.dim`	any.dim(Tensor self, int dim, bool keepdim=False) -> Tensor
`aten.any.dims`	any.dims(Tensor self, int[]? dim=None, bool keepdim=False) -> Tensor
`aten.arange.start_step`	arange.start_step(Scalar start, Scalar end, Scalar step=1, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
`aten.argmax`	argmax(Tensor self, int? dim=None, bool keepdim=False) -> Tensor
`aten.argmin`	argmin(Tensor self, int? dim=None, bool keepdim=False) -> Tensor
`aten.as_strided`	as_strided(Tensor(a) self, SymInt[] size, SymInt[] stride, SymInt? storage_offset=None) -> Tensor(a)
`aten.asin`	asin(Tensor self) -> Tensor
`aten.asinh`	asinh(Tensor self) -> Tensor
`aten.atan`	atan(Tensor self) -> Tensor
`aten.atan2`	atan2(Tensor self, Tensor other) -> Tensor
`aten.atan2.out`	atan2.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
`aten.atanh`	atanh(Tensor self) -> Tensor
`aten.avg_pool1d`	avg_pool1d(Tensor self, int[1] kernel_size, int[1] stride=[], int[1] padding=0, bool ceil_mode=False, bool count_include_pad=True) -> Tensor
`aten.avg_pool2d`	avg_pool2d(Tensor self, int[2] kernel_size, int[2] stride=[], int[2] padding=0, bool ceil_mode=False, bool count_include_pad=True, int? divisor_override=None) -> Tensor
`aten.avg_pool2d_backward`	avg_pool2d_backward(Tensor grad_output, Tensor self, int[2] kernel_size, int[2] stride, int[2] padding, bool ceil_mode, bool count_include_pad, int? divisor_override) -> Tensor
`aten.avg_pool3d`	avg_pool3d(Tensor self, int[3] kernel_size, int[3] stride=[], int[3] padding=0, bool ceil_mode=False, bool count_include_pad=True, int? divisor_override=None) -> Tensor
`aten.bitwise_and.Scalar`	bitwise_and.Scalar(Tensor self, Scalar other) -> Tensor
`aten.bitwise_and.Tensor`	bitwise_and.Tensor(Tensor self, Tensor other) -> Tensor
`aten.bitwise_not`	bitwise_not(Tensor self) -> Tensor
`aten.bitwise_or.Scalar`	bitwise_or.Scalar(Tensor self, Scalar other) -> Tensor
`aten.bitwise_or.Tensor`	bitwise_or.Tensor(Tensor self, Tensor other) -> Tensor
`aten.bitwise_xor.Scalar`	bitwise_xor.Scalar(Tensor self, Scalar other) -> Tensor
`aten.bitwise_xor.Tensor`	bitwise_xor.Tensor(Tensor self, Tensor other) -> Tensor
`aten.bmm`	bmm(Tensor self, Tensor mat2) -> Tensor
`aten.cat`	cat(Tensor[] tensors, int dim=0) -> Tensor

`aten.ceil`	ceil(Tensor self) -> Tensor
`aten.clamp`	clamp(Tensor self, Scalar? min=None, Scalar? max=None) -> Tensor
`aten.clamp.Tensor`	clamp.Tensor(Tensor self, Tensor? min=None, Tensor? max=None) -> Tensor
`aten.clone`	clone(Tensor self, *, MemoryFormat? memory_format=None) -> Tensor
`aten.col2im`	col2im(Tensor self, SymInt[2] output_size, int[2] kernel_size, int[2] dilation, int[2] padding, int[2] stride) -> Tensor
`aten.constant_pad_nd`	constant_pad_nd(Tensor self, SymInt[] pad, Scalar value=0) -> Tensor
`aten.convolution`	convolution(Tensor input, Tensor weight, Tensor? bias, SymInt[] stride, SymInt[] padding, SymInt[] dilation, bool transposed, SymInt[] output_padding, SymInt groups) -> Tensor
`aten.convolution_backward`	convolution_backward(Tensor grad_output, Tensor input, Tensor weight, SymInt[]? bias_sizes, SymInt[] stride, SymInt[] padding, SymInt[] dilation, bool transposed, SymInt[] output_padding, SymInt groups, bool[3] output_mask) -> (Tensor, Tensor, Tensor)
`aten.copy`	copy(Tensor self, Tensor src, bool non_blocking=False) -> Tensor
`aten.cos`	cos(Tensor self) -> Tensor
`aten.cosh`	cosh(Tensor self) -> Tensor
`aten.cumsum`	cumsum(Tensor self, int dim, *, ScalarType? dtype=None) -> Tensor
`aten.diagonal`	diagonal(Tensor(a) self, int offset=0, int dim1=0, int dim2=1) -> Tensor(a)
`aten.div.Scalar`	div.Scalar(Tensor self, Scalar other) -> Tensor
`aten.div.Scalar_mode`	div.Scalar_mode(Tensor self, Scalar other, *, str? rounding_mode) -> Tensor
`aten.div.Tensor`	div.Tensor(Tensor self, Tensor other) -> Tensor
`aten.div.Tensor_mode`	div.Tensor_mode(Tensor self, Tensor other, *, str? rounding_mode) -> Tensor
`aten.elu`	elu(Tensor self, Scalar alpha=1, Scalar scale=1, Scalar input_scale=1) -> Tensor
`aten.embedding`	embedding(Tensor weight, Tensor indices, SymInt padding_idx=-1, bool scale_grad_by_freq=False, bool sparse=False) -> Tensor
`aten.embedding_dense_backward`	embedding_dense_backward(Tensor grad_output, Tensor indices, SymInt num_weights, SymInt padding_idx, bool scale_grad_by_freq) -> Tensor
`aten.empty.memory_format`	empty.memory_format(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None, MemoryFormat? memory_format=None) -> Tensor
`aten.empty_strided`	empty_strided(SymInt[] size, SymInt[] stride, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
`aten.eq.Scalar`	eq.Scalar(Tensor self, Scalar other) -> Tensor
`aten.eq.Tensor`	eq.Tensor(Tensor self, Tensor other) -> Tensor
`aten.erf`	erf(Tensor self) -> Tensor
`aten.exp`	exp(Tensor self) -> Tensor
`aten.expand`	expand(Tensor(a) self, SymInt[] size, *, bool implicit=False) -> Tensor(a)
`aten.expm1`	expm1(Tensor self) -> Tensor
`aten.fill.Scalar`	fill.Scalar(Tensor self, Scalar value) -> Tensor
`aten.flip`	flip(Tensor self, int[] dims) -> Tensor
`aten.floor`	floor(Tensor self) -> Tensor
`aten.fmod.Scalar`	fmod.Scalar(Tensor self, Scalar other) -> Tensor
`aten.fmod.Tensor`	fmod.Tensor(Tensor self, Tensor other) -> Tensor
`aten.full`	full(SymInt[] size, Scalar fill_value, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
`aten.full_like`	full_like(Tensor self, Scalar fill_value, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None, MemoryFormat? memory_format=None) -> Tensor
`aten.gather`	gather(Tensor self, int dim, Tensor index, *, bool sparse_grad=False) -> Tensor
`aten.ge.Scalar`	ge.Scalar(Tensor self, Scalar other) -> Tensor
`aten.ge.Tensor`	ge.Tensor(Tensor self, Tensor other) -> Tensor
`aten.gelu`	gelu(Tensor self, *, str approximate=’none’) -> Tensor
`aten.grid_sampler_2d`	grid_sampler_2d(Tensor input, Tensor grid, int interpolation_mode, int padding_mode, bool align_corners) -> Tensor
`aten.gt.Scalar`	gt.Scalar(Tensor self, Scalar other) -> Tensor
`aten.gt.Tensor`	gt.Tensor(Tensor self, Tensor other) -> Tensor
`aten.hardtanh`	hardtanh(Tensor self, Scalar min_val=-1, Scalar max_val=1) -> Tensor
`aten.index.Tensor`	index.Tensor(Tensor self, Tensor?[] indices) -> Tensor
`aten.index_put`	index_put(Tensor self, Tensor?[] indices, Tensor values, bool accumulate=False) -> Tensor
`aten.index_select`	index_select(Tensor self, int dim, Tensor index) -> Tensor
`aten.isinf`	isinf(Tensor self) -> Tensor
`aten.isnan`	isnan(Tensor self) -> Tensor
`aten.le.Scalar`	le.Scalar(Tensor self, Scalar other) -> Tensor
`aten.le.Tensor`	le.Tensor(Tensor self, Tensor other) -> Tensor
`aten.leaky_relu`	leaky_relu(Tensor self, Scalar negative_slope=0.01) -> Tensor
`aten.log`	log(Tensor self) -> Tensor
`aten.log10`	log10(Tensor self) -> Tensor
`aten.log1p`	log1p(Tensor self) -> Tensor
`aten.log2`	log2(Tensor self) -> Tensor

`aten.logical_and`	logical_and(Tensor self, Tensor other) -> Tensor
`aten.logical_not`	logical_not(Tensor self) -> Tensor
`aten.logical_or`	logical_or(Tensor self, Tensor other) -> Tensor
`aten.logical_xor`	logical_xor(Tensor self, Tensor other) -> Tensor
`aten.lt.Scalar`	lt.Scalar(Tensor self, Scalar other) -> Tensor
`aten.lt.Tensor`	lt.Tensor(Tensor self, Tensor other) -> Tensor
`aten.masked_scatter`	masked_scatter(Tensor self, Tensor mask, Tensor source) -> Tensor
`aten.max.dim`	max.dim(Tensor self, int dim, bool keepdim=False) -> (Tensor values, Tensor indices)
`aten.max_pool2d_with_indices`	max_pool2d_with_indices(Tensor self, int[2] kernel_size, int[2] stride=[], int[2] padding=0, int[2] dilation=1, bool ceil_mode=False) -> (Tensor, Tensor)
`aten.max_pool2d_with_indices_backward`	max_pool2d_with_indices_backward(Tensor grad_output, Tensor self, int[2] kernel_size, int[2] stride, int[2] padding, int[2] dilation, bool ceil_mode, Tensor indices) -> Tensor
`aten.max_pool3d_with_indices`	max_pool3d_with_indices(Tensor self, int[3] kernel_size, int[3] stride=[], int[3] padding=0, int[3] dilation=1, bool ceil_mode=False) -> (Tensor, Tensor)
`aten.maximum`	maximum(Tensor self, Tensor other) -> Tensor
`aten.mean`	mean(Tensor self, *, ScalarType? dtype=None) -> Tensor
`aten.mean.dim`	mean.dim(Tensor self, int[1]? dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor
`aten.min.dim`	min.dim(Tensor self, int dim, bool keepdim=False) -> (Tensor values, Tensor indices)
`aten.minimum`	minimum(Tensor self, Tensor other) -> Tensor
`aten.mm`	mm(Tensor self, Tensor mat2) -> Tensor
`aten.mul.Scalar`	mul.Scalar(Tensor self, Scalar other) -> Tensor
`aten.mul.Tensor`	mul.Tensor(Tensor self, Tensor other) -> Tensor
`aten.native_dropout`	native_dropout(Tensor input, float p, bool? train) -> (Tensor, Tensor)
`aten.native_group_norm`	native_group_norm(Tensor input, Tensor? weight, Tensor? bias, SymInt N, SymInt C, SymInt HxW, int group, float eps) -> (Tensor, Tensor, Tensor)
`aten.native_group_norm_backward`	native_group_norm_backward(Tensor grad_out, Tensor input, Tensor mean, Tensor rstd, Tensor? weight, SymInt N, SymInt C, SymInt HxW, int group, bool[3] output_mask) -> (Tensor, Tensor, Tensor)
`aten.native_layer_norm`	native_layer_norm(Tensor input, SymInt[] normalized_shape, Tensor? weight, Tensor? bias, float eps) -> (Tensor, Tensor, Tensor)
`aten.native_layer_norm_backward`	native_layer_norm_backward(Tensor grad_out, Tensor input, SymInt[] normalized_shape, Tensor mean, Tensor rstd, Tensor? weight, Tensor? bias, bool[3] output_mask) -> (Tensor, Tensor, Tensor)
`aten.ne.Scalar`	ne.Scalar(Tensor self, Scalar other) -> Tensor
`aten.ne.Tensor`	ne.Tensor(Tensor self, Tensor other) -> Tensor
`aten.neg`	neg(Tensor self) -> Tensor
`aten.nonzero`	nonzero(Tensor self) -> Tensor
`aten.permute`	permute(Tensor(a) self, int[] dims) -> Tensor(a)
`aten.pow.Scalar`	pow.Scalar(Scalar self, Tensor exponent) -> Tensor
`aten.pow.Tensor_Scalar`	pow.Tensor_Scalar(Tensor self, Scalar exponent) -> Tensor
`aten.pow.Tensor_Tensor`	pow.Tensor_Tensor(Tensor self, Tensor exponent) -> Tensor
`aten.prod`	prod(Tensor self, *, ScalarType? dtype=None) -> Tensor
`aten.prod.dim_int`	prod.dim_int(Tensor self, int dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor
`aten.rand`	rand(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
`aten.randn`	randn(SymInt[] size, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
`aten.randperm`	randperm(SymInt n, *, ScalarType? dtype=long, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
`aten.reciprocal`	reciprocal(Tensor self) -> Tensor
`aten.reflection_pad1d`	reflection_pad1d(Tensor self, SymInt[2] padding) -> Tensor
`aten.reflection_pad2d`	reflection_pad2d(Tensor self, SymInt[4] padding) -> Tensor
`aten.reflection_pad3d`	reflection_pad3d(Tensor self, SymInt[6] padding) -> Tensor
`aten.relu`	relu(Tensor self) -> Tensor
`aten.remainder.Scalar`	remainder.Scalar(Tensor self, Scalar other) -> Tensor
`aten.remainder.Tensor`	remainder.Tensor(Tensor self, Tensor other) -> Tensor
`aten.repeat`	repeat(Tensor self, SymInt[] repeats) -> Tensor
`aten.replication_pad2d`	replication_pad2d(Tensor self, SymInt[4] padding) -> Tensor
`aten.replication_pad3d`	replication_pad3d(Tensor self, SymInt[6] padding) -> Tensor
`aten.resize_`	resize_(Tensor(a!) self, SymInt[] size, *, MemoryFormat? memory_format=None) -> Tensor(a!)
`aten.round`	round(Tensor self) -> Tensor
`aten.rsqrt`	rsqrt(Tensor self) -> Tensor
`aten.scalar_tensor`	scalar_tensor(Scalar s, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None) -> Tensor
`aten.scatter.src`	scatter.src(Tensor self, int dim, Tensor index, Tensor src) -> Tensor
`aten.scatter.value`	scatter.value(Tensor self, int dim, Tensor index, Scalar value) -> Tensor

`aten.scatter_add`	scatter_add(Tensor self, int dim, Tensor index, Tensor src) -> Tensor
`aten.scatter_reduce.two`	scatter_reduce.two(Tensor self, int dim, Tensor index, Tensor src, str reduce, *, bool include_self=True) -> Tensor
`aten.select.int`	select.int(Tensor(a) self, int dim, SymInt index) -> Tensor(a)
`aten.select_scatter`	select_scatter(Tensor self, Tensor src, int dim, SymInt index) -> Tensor
`aten.sigmoid`	sigmoid(Tensor self) -> Tensor
`aten.sign`	sign(Tensor self) -> Tensor
`aten.sin`	sin(Tensor self) -> Tensor
`aten.sinh`	sinh(Tensor self) -> Tensor
`aten.slice.Tensor`	slice.Tensor(Tensor(a) self, int dim=0, SymInt? start=None, SymInt? end=None, SymInt step=1) -> Tensor(a)
`aten.slice_scatter`	slice_scatter(Tensor self, Tensor src, int dim=0, SymInt? start=None, SymInt? end=None, SymInt step=1) -> Tensor
`aten.sort`	sort(Tensor self, int dim=-1, bool descending=False) -> (Tensor values, Tensor indices)
`aten.split_with_sizes`	split_with_sizes(Tensor(a -> *) self, SymInt[] split_sizes, int dim=0) -> Tensor(a)[]
`aten.sqrt`	sqrt(Tensor self) -> Tensor
`aten.squeeze.dim`	squeeze.dim(Tensor(a) self, int dim) -> Tensor(a)
`aten.squeeze.dims`	squeeze.dims(Tensor(a) self, int[] dim) -> Tensor(a)
`aten.sub.Scalar`	sub.Scalar(Tensor self, Scalar other, Scalar alpha=1) -> Tensor
`aten.sub.Tensor`	sub.Tensor(Tensor self, Tensor other, *, Scalar alpha=1) -> Tensor
`aten.sum.dim_IntList`	sum.dim_IntList(Tensor self, int[1]? dim, bool keepdim=False, *, ScalarType? dtype=None) -> Tensor
`aten.sym_is_contiguous`	sym_is_contiguous(Tensor self, MemoryFormat memory_format=contiguous_format) -> SymBool
`aten.sym_numel`	sym_numel(Tensor self) -> SymInt
`aten.sym_size.int`	sym_size.int(Tensor self, int dim) -> SymInt
`aten.sym_storage_offset`	sym_storage_offset(Tensor self) -> SymInt
`aten.sym_stride.int`	sym_stride.int(Tensor self, int dim) -> SymInt
`aten.tan`	tan(Tensor self) -> Tensor
`aten.tanh`	tanh(Tensor self) -> Tensor
`aten.topk`	topk(Tensor self, SymInt k, int dim=-1, bool largest=True, bool sorted=True) -> (Tensor values, Tensor indices)
`aten.trunc`	trunc(Tensor self) -> Tensor
`aten.unsqueeze`	unsqueeze(Tensor(a) self, int dim) -> Tensor(a)
`aten.upsample_bilinear2d.vec`	upsample_bilinear2d.vec(Tensor input, SymInt[]? output_size, bool align_corners, float[]? scale_factors) -> Tensor
`aten.upsample_nearest2d.vec`	upsample_nearest2d.vec(Tensor input, SymInt[]? output_size, float[]? scale_factors) -> Tensor
`aten.var.correction`	var.correction(Tensor self, int[1]? dim=None, *, Scalar? correction=None, bool keepdim=False) -> Tensor
`aten.var.dim`	var.dim(Tensor self, int[1]? dim, bool unbiased=True, bool keepdim=False) -> Tensor
`aten.view`	view(Tensor(a) self, SymInt[] size) -> Tensor(a)
`aten.where.self`	where.self(Tensor condition, Tensor self, Tensor other) -> Tensor

Prims IR

Prims IR — это набор примитивных операторов, которые можно использовать для составления других операторов. Prims IR — это набор операторов более низкого уровня, чем core aten IR; он дополнительно раскладывает операции на явные операции приведения типов и широковещательного распространения: prims.convert_element_type и prims.broadcast_in_dim. Этот набор операторов предназначен для взаимодействия с серверными компонентами компилятора.

Предупреждение

Этот набор операторов всё ещё активно разрабатывается; в будущем будут добавлены новые операции.

Оператор	Схема
`prims.abs`	(Tensor self) -> Tensor
`prims.acos`	(Tensor self) -> Tensor
`prims.acosh`	(Tensor self) -> Tensor
`prims.asin`	(Tensor self) -> Tensor
`prims.asinh`	(Tensor self) -> Tensor
`prims.atan`	(Tensor self) -> Tensor
`prims.atanh`	(Tensor self) -> Tensor
`prims.cos`	(Tensor self) -> Tensor
`prims.cosh`	(Tensor self) -> Tensor
`prims.bessel_i0`	(Tensor self) -> Tensor
`prims.bessel_i0e`	(Tensor self) -> Tensor
`prims.bessel_i1`	(Tensor self) -> Tensor
`prims.bessel_i1e`	(Tensor self) -> Tensor
`prims.bessel_j0`	(Tensor self) -> Tensor
`prims.bessel_j1`	(Tensor self) -> Tensor
`prims.bitwise_not`	(Tensor self) -> Tensor
`prims.cbrt`	(Tensor self) -> Tensor
`prims.ceil`	(Tensor self) -> Tensor
`prims.conj_physical`	(Tensor self) -> Tensor
`prims.digamma`	(Tensor self) -> Tensor
`prims.erf`	(Tensor self) -> Tensor
`prims.erf_inv`	(Tensor self) -> Tensor
`prims.erfc`	(Tensor self) -> Tensor
`prims.erfcx`	(Tensor self) -> Tensor
`prims.exp`	(Tensor self) -> Tensor
`prims.expm1`	(Tensor self) -> Tensor
`prims.exp2`	(Tensor self) -> Tensor
`prims.fill`	(Tensor self, Scalar value) -> Tensor
`prims.floor`	(Tensor self) -> Tensor
`prims.imag`	(Tensor(a) self) -> Tensor(a)
`prims.isfinite`	(Tensor self) -> Tensor
`prims.lgamma`	(Tensor self) -> Tensor
`prims.log`	(Tensor self) -> Tensor
`prims.log1p`	(Tensor self) -> Tensor
`prims.log2`	(Tensor self) -> Tensor
`prims.log10`	(Tensor self) -> Tensor
`prims.ndtri`	(Tensor self) -> Tensor
`prims.neg`	(Tensor self) -> Tensor
`prims.real`	(Tensor(a) self) -> Tensor(a)
`prims.reciprocal`	(Tensor self) -> Tensor
`prims.round`	(Tensor self) -> Tensor
`prims.sign`	(Tensor self) -> Tensor
`prims.signbit`	(Tensor self) -> Tensor
`prims.sin`	(Tensor self) -> Tensor
`prims.sinh`	(Tensor self) -> Tensor
`prims.spherical_bessel_j0`	(Tensor self) -> Tensor
`prims.sqrt`	(Tensor self) -> Tensor
`prims.tan`	(Tensor self) -> Tensor
`prims.tanh`	(Tensor self) -> Tensor
`prims.trunc`	(Tensor self) -> Tensor
`prims.add`	(Tensor self, Tensor other) -> Tensor
`prims.atan2`	(Tensor self, Tensor other) -> Tensor
`prims.bitwise_and`	(Tensor self, Tensor other) -> Tensor
`prims.bitwise_or`	(Tensor self, Tensor other) -> Tensor
`prims.bitwise_xor`	(Tensor self, Tensor other) -> Tensor
`prims.div`	(Tensor self, Tensor other) -> Tensor
`prims.eq`	(Tensor self, Tensor other) -> Tensor
`prims.fmax`	(Tensor self, Tensor other) -> Tensor
`prims.fmin`	(Tensor self, Tensor other) -> Tensor
`prims.fmod`	(Tensor self, Tensor other) -> Tensor
`prims.frexp`	(Tensor self) -> (Tensor mantissa, Tensor exponent)
`prims.gcd`	(Tensor self, Tensor other) -> Tensor
`prims.ge`	(Tensor self, Tensor other) -> Tensor
`prims.gt`	(Tensor self, Tensor other) -> Tensor
`prims.hypot`	(Tensor self, Tensor other) -> Tensor
`prims.igamma`	(Tensor self, Tensor other) -> Tensor
`prims.igammac`	(Tensor self, Tensor other) -> Tensor
`prims.le`	(Tensor self, Tensor other) -> Tensor
`prims.lt`	(Tensor self, Tensor other) -> Tensor
`prims.maximum`	(Tensor self, Tensor other) -> Tensor
`prims.minimum`	(Tensor self, Tensor other) -> Tensor
`prims.mul`	(Tensor self, Tensor other) -> Tensor
`prims.ne`	(Tensor self, Tensor other) -> Tensor
`prims.nextafter`	(Tensor self, Tensor other) -> Tensor
`prims.pow`	(Tensor self, Tensor other) -> Tensor

`prims.remainder`	(Tensor self, Tensor other) -> Tensor
`prims.rsqrt`	(Tensor self) -> Tensor
`prims.shift_left`	(Tensor self, Tensor other) -> Tensor
`prims.shift_right_arithmetic`	(Tensor self, Tensor other) -> Tensor
`prims.sub`	(Tensor self, Tensor other) -> Tensor
`prims.zeta`	(Tensor self, Tensor other) -> Tensor
`prims.as_strided`	(Tensor(a!) a, SymInt[] size, SymInt[] stride, SymInt storage_offset) -> Tensor(a!)
`prims.broadcast_in_dim`	(Tensor(a) a, SymInt[] shape, int[] broadcast_dimensions) -> Tensor(a)
`prims.collapse_view`	(Tensor(a) a, int start, int end) -> Tensor(a)
`prims.conj`	(Tensor(a) a) -> Tensor(a)
`prims.split_dim`	(Tensor(a) a, int dim, SymInt outer_length) -> Tensor(a)
`prims.squeeze`	(Tensor(a) a, int[] dimensions) -> Tensor(a)
`prims.transpose`	(Tensor(a) a, int[] permutation) -> Tensor(a)
`prims.view_of`	(Tensor(a) a) -> Tensor(a)
`prims.view_of_dtype`	(Tensor(a) a, ScalarType dtype) -> Tensor(a)
`prims.as_strided_scatter`	(Tensor self, Tensor src, SymInt[] size, SymInt[] stride, SymInt storage_offset) -> Tensor
`prims.collapse`	(Tensor a, int start, int end) -> Tensor
`prims.cat`	(Tensor[] tensors, int dim) -> Tensor
`prims.reshape`	(Tensor a, SymInt[] shape) -> Tensor
`prims.rev`	(Tensor a, int[] dims) -> Tensor
`prims.where`	(Tensor pred, Tensor a, Tensor b) -> Tensor
`prims.clone`	(Tensor self, *, MemoryFormat? memory_format=None) -> Tensor
`prims.convert_element_type`	(Tensor a, ScalarType dtype) -> Tensor
`prims.device_put`	(Tensor a, Device device, bool non_blocking=False) -> Tensor
`prims.item`	(Tensor a) -> Scalar
`prims.maximum_value`	(ScalarType dtype) -> Scalar
`prims.minimum_value`	(ScalarType dtype) -> Scalar
`prims.copy_strided`	(Tensor a, SymInt[] stride) -> Tensor
`prims.copy_to`	(Tensor(a!) a, Tensor b) -> Tensor(a!)
`prims.resize`	(Tensor(a!) a, SymInt[] shape) -> Tensor(a!)
`prims.amax`	(Tensor inp, int[]? dims, *, ScalarType? output_dtype=None) -> Tensor
`prims.amin`	(Tensor inp, int[]? dims, *, ScalarType? output_dtype=None) -> Tensor
`prims.prod`	(Tensor inp, int[]? dims, *, ScalarType? output_dtype=None) -> Tensor
`prims.sum`	(Tensor inp, int[]? dims, *, ScalarType? output_dtype=None) -> Tensor
`prims.xor_sum`	(Tensor inp, int[]? dims, *, ScalarType? output_dtype=None) -> Tensor
`prims.var`	(Tensor inp, int[]? dims, float? correction=1, *, ScalarType? output_dtype=None) -> Tensor
`prims.empty_strided`	(SymInt[] shape, SymInt[] strides, *, ScalarType dtype, Device device, bool requires_grad) -> Tensor
`prims.empty_permuted`	(SymInt[] shape, int[] physical_layout, *, ScalarType dtype, Device device, bool requires_grad) -> Tensor
`prims.scalar_tensor`	(Scalar s, *, ScalarType? dtype=None, Device? device=None) -> Tensor
`prims.iota`	(SymInt length, *, SymInt start, SymInt step, ScalarType dtype, Device device, bool requires_grad) -> Tensor
`prims.svd`	(Tensor A, *, bool full_matrices) -> (Tensor U, Tensor S, Tensor Vh)
`prims.normal`	(SymInt[] shape, *, Scalar mean, Scalar std, ScalarType dtype, Device device, bool requires_grad, Generator? generator=None) -> Tensor
`prims.uniform`	(SymInt[] shape, *, Scalar low, Scalar high, ScalarType dtype, Device device, SymInt[] stride, Generator? generator=None) -> Tensor
`prims.fft_r2c`	(Tensor self, *, int[] dim, bool onesided) -> Tensor
`prims.fft_c2c`	(Tensor self, *, int[] dim, bool forward) -> Tensor
`prims.fft_c2r`	(Tensor self, *, int[] dim, SymInt last_dim_size) -> Tensor
`prims._make_token`	() -> Tensor
`prims._sink_tokens`	(Tensor[] tokens) -> ()

© 2026, PyTorch Contributors
PyTorch has a BSD-style license, as found in the LICENSE file.
https://docs.pytorch.org/docs/2.14/user_guide/torch_compiler/torch.compiler_ir.html