This document is relevant for: Inf1

PyTorch Neuron (torch-neuron) Supported operators#

Current operator lists may be generated with these commands inside python:

import torch.neuron
print(*torch.neuron.get_supported_operations(), sep='\n')

PyTorch Neuron release [2.5.0.0]#

Date: 11/23/2022

Added support for new operators:

  • aten::threshold

  • aten::roll

  • aten::instance_norm

  • aten::amin

  • aten::amax

  • aten::new_empty

  • aten::new_ones

  • aten::tril

  • aten::triu

  • aten::zero_

  • aten::all

  • aten::broadcast_tensors

  • aten::broadcast_to

  • aten::logical_and

  • aten::logical_not

  • aten::logical_or

  • aten::logical_xor

  • aten::_convolution_mode

Added limited support for new operators:

PyTorch Neuron release [2.2.0.0]#

Date: 03/25/2022

  • Added full support for aten::max_pool2d_with_indices - (Was previously supported only when indices were unused).

PyTorch Neuron release [2.1.7.0]#

Date: 01/20/2022

  • Added support for new operators:

    • aten::bucketize

    • aten::any

    • aten::remainder

    • aten::clip

    • aten::repeat_interleave

    • aten::tensor_split

    • aten::split_with_sizes

    • aten::isnan

    • aten::embedding_renorm_

    • aten::dot

    • aten::mv

    • aten::hardsigmoid

    • aten::hardswish

    • aten::trunc

    • aten::one_hot: Supported when num_classes is known at trace time. The dynamic version of this operation when num_classes = -1 is not supported.

    • aten::adaptive_max_pool1d

    • aten::adaptive_max_pool2d

PyTorch Neuron Release [2.0.536.0]#

  • The following are operators with limited support on Neuron. Unlike fully supported operators, these operators are not returned when using torch_neuron.get_supported_operations(). See each operator description for conditional support:

    • aten::max_pool2d_with_indices - Supported when indices outputs are not used by a downstream operation. This allows the operation to be compiled to Neuron when it is equivalent to an aten::max_pool2d.

    • aten::max_pool3d_with_indices - Supported when indices outputs are not used by a downstream operation. This allows the operation to be compiled to Neuron when it is equivalent to an aten::max_pool3d.

    • aten::where - Supported when used as a conditional selection (3-argument variant). Unsupported when used to generate a dynamic list of indices (1-argument variant). See torch.where().

PyTorch Neuron Release [2.0.318.0]#

  • Added support for new operators:

    • aten::empty_like

    • aten::log

    • aten::type_as

    • aten::movedim

    • aten::einsum

    • aten::argmax

    • aten::min

    • aten::argmin

    • aten::abs

    • aten::cos

    • aten::sin

    • aten::linear

    • aten::pixel_shuffle

    • aten::group_norm

    • aten::_weight_norm

PyTorch Neuron Release [1.5.21.0]#

No change

PyTorch Neuron Release [1.5.7.0]#

Added:

aten::erf
prim::DictConstruct

PyTorch Neuron Release [1.4.1.0]#

No change

PyTorch Neuron Release [1.3.5.0]#

Added:

aten::numel
aten::ones_like
aten::reciprocal
aten::topk

PyTorch Neuron Release [1.2.16.0]#

No change

PyTorch Neuron Release [1.2.15.0]#

No change

PyTorch Neuron Release [1.2.3.0]#

Added:

aten::silu
aten::zeros_like

PyTorch Neuron Release [1.1.7.0]#

Added:

aten::_shape_as_tensor
aten::chunk
aten::empty
aten::masked_fill

PyTorch Neuron Release [1.0.24045.0]#

Added:

aten::__and__
aten::bmm
aten::clone
aten::expand_as
aten::fill_
aten::floor_divide
aten::full
aten::hardtanh
aten::hardtanh_
aten::le
aten::leaky_relu
aten::lt
aten::mean
aten::ne
aten::softplus
aten::unbind
aten::upsample_bilinear2d

PyTorch Neuron Release [1.0.1720.00]#

Added:

aten::constant_pad_nd
aten::meshgrid

PyTorch Neuron Release [1.0.1532.0]#

Added:

aten::ones

PyTorch Neuron Release [1.0.1522.0]#

  • No change

PyTorch Neuron Release [1.0.1386.0]#

Added the following instructions. Please note, primitives are included in this list from this release.

aten::ceil
aten::clamp
aten::eq
aten::exp
aten::expand_as
aten::flip
aten::full_like
aten::ge
aten::gt
aten::log2
aten::log_softmax
aten::max
aten::neg
aten::relu
aten::rsqrt
aten::scalarImplicit
aten::sqrt
aten::squeeze
aten::stack
aten::sub
aten::sum
aten::true_divide
aten::upsample_nearest2d
prim::Constant
prim::GetAttr
prim::ImplicitTensorToNum
prim::ListConstruct
prim::ListUnpack
prim::NumToTensor
prim::TupleConstruct
prim::TupleUnpack

PyTorch Neuron Release [1.0.1168.0]#

Added aten::ScalarImplicit

PyTorch Neuron Release [1.0.1001.0]#

Added

aten::detach
aten::floor
aten::gelu
aten::pow
aten::sigmoid
aten::split

Removed ( Reasons given alongside )

aten::embedding (does not meet performance criteria)
aten::erf (error function does not meet accuracy criteria)
aten::tf_dtype_from_torch (internal support function, not an operator)

PyTorch Neuron Release [1.0.825.0]#

PyTorch Neuron Release [1.0.763.0]#

  • Please note. Starting with this release we will not publish primitives (prim::).

  • Previous release inaccurately listed these operators as aten ops, they are not.

aten::tf_broadcastable_slice
aten::tf_padding

The following new operators are added in this release.

aten::Int
aten::arange
aten::contiguous
aten::div
aten::embedding
aten::erf
aten::expand
aten::eye
aten::index_select
aten::layer_norm
aten::matmul
aten::mm
aten::permute
aten::reshape
aten::rsub
aten::select
aten::size
aten::slice
aten::softmax
aten::tf_dtype_from_torch
aten::to
aten::transpose
aten::unsqueeze
aten::view
aten::zeros

These operators were already supported previously (removing the two that were included by mistake)

aten::_convolution
aten::adaptive_avg_pool2d
aten::add
aten::add_
aten::addmm
aten::avg_pool2d
aten::batch_norm
aten::cat
aten::dimension_value
aten::dropout
aten::flatten
aten::max_pool2d
aten::mul
aten::relu_
aten::t
aten::tanh
aten::values
prim::Constant
prim::GetAttr
prim::ListConstruct
prim::ListUnpack
prim::TupleConstruct
prim::TupleUnpack

PyTorch Neuron Release [1.0.672.0]#

No change

PyTorch Neuron Release [1.0.552.0]#

aten::_convolution
aten::adaptive_avg_pool2d
aten::add
aten::add_
aten::addmm
aten::avg_pool2d
aten::batch_norm
aten::cat
aten::dimension_value
aten::dropout
aten::flatten
aten::max_pool2d
aten::mul
aten::relu_
aten::t
aten::tanh
aten::tf_broadcastable_slice
aten::tf_padding
aten::values
prim::Constant
prim::GetAttr
prim::ListConstruct
prim::ListUnpack
prim::TupleConstruct
prim::TupleUnpack

This document is relevant for: Inf1