torch.backends

torch.backends controls the behavior of various backends that PyTorch supports.

These backends include:

• torch.backends.cpu

• torch.backends.cuda

• torch.backends.cudnn

• torch.backends.mha

• torch.backends.mps

• torch.backends.mkl

• torch.backends.mkldnn

• torch.backends.nnpack

• torch.backends.openmp

• torch.backends.opt_einsum

• torch.backends.xeon

torch.backends.cpu

torch.backends.cpu.get_cpu_capability()

Return cpu capability as a string value.

Possible values: - “DEFAULT” - “VSX” - “Z VECTOR” - “NO AVX” - “AVX2” - “AVX512”

Return type

str

torch.backends.cuda

torch.backends.cuda.is_built()

Return whether PyTorch is built with CUDA support.

Note that this doesn’t necessarily mean CUDA is available; just that if this PyTorch binary were run on a machine with working CUDA drivers and devices, we would be able to use it.

torch.backends.cuda.matmul.allow_tf32

A bool that controls whether TensorFloat-32 tensor cores may be used in matrix multiplications on Ampere or newer GPUs. See TensorFloat-32 (TF32) on Ampere (and later) devices.

torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction

A bool that controls whether reduced precision reductions (e.g., with fp16 accumulation type) are allowed with fp16 GEMMs.

torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction

A bool that controls whether reduced precision reductions are allowed with bf16 GEMMs.

torch.backends.cuda.cufft_plan_cache

cufft_plan_cache contains the cuFFT plan caches for each CUDA device. Query a specific device i’s cache via torch.backends.cuda.cufft_plan_cache[i].

torch.backends.cuda.cufft_plan_cache.size

A readonly int that shows the number of plans currently in a cuFFT plan cache.

torch.backends.cuda.cufft_plan_cache.max_size

A int that controls the capacity of a cuFFT plan cache.

torch.backends.cuda.cufft_plan_cache.clear()

Clears a cuFFT plan cache.

torch.backends.cuda.preferred_linalg_library(backend=None)

Override the heuristic PyTorch uses to choose between cuSOLVER and MAGMA for CUDA linear algebra operations.

Warning

This flag is experimental and subject to change.

When PyTorch runs a CUDA linear algebra operation it often uses the cuSOLVER or MAGMA libraries, and if both are available it decides which to use with a heuristic. This flag (a str) allows overriding those heuristics.

• If “cusolver” is set then cuSOLVER will be used wherever possible.

• If “magma” is set then MAGMA will be used wherever possible.

• If “default” (the default) is set then heuristics will be used to pick between cuSOLVER and MAGMA if both are available.

• When no input is given, this function returns the currently preferred library.

• User may use the environment variable TORCH_LINALG_PREFER_CUSOLVER=1 to set the preferred library to cuSOLVER globally. This flag only sets the initial value of the preferred library and the preferred library may still be overridden by this function call later in your script.

Note: When a library is preferred other libraries may still be used if the preferred library doesn’t implement the operation(s) called. This flag may achieve better performance if PyTorch’s heuristic library selection is incorrect for your application’s inputs.

Currently supported linalg operators:

• torch.linalg.inv()

• torch.linalg.inv_ex()

• torch.linalg.cholesky()

• torch.linalg.cholesky_ex()

• torch.cholesky_solve()

• torch.cholesky_inverse()

• torch.linalg.lu_factor()

• torch.linalg.lu()

• torch.linalg.lu_solve()

• torch.linalg.qr()

• torch.linalg.eigh()

• torch.linalg.eighvals()

• torch.linalg.svd()

• torch.linalg.svdvals()

Return type

_LinalgBackend

class torch.backends.cuda.SDPAParams

torch.backends.cuda.flash_sdp_enabled()

Warning

This flag is beta and subject to change.

Returns whether flash scaled dot product attention is enabled or not.

torch.backends.cuda.enable_mem_efficient_sdp(enabled)

Warning

This flag is beta and subject to change.

Enables or disables memory efficient scaled dot product attention.

torch.backends.cuda.mem_efficient_sdp_enabled()

Warning

This flag is beta and subject to change.

Returns whether memory efficient scaled dot product attention is enabled or not.

torch.backends.cuda.enable_flash_sdp(enabled)

Warning

This flag is beta and subject to change.

Enables or disables flash scaled dot product attention.

torch.backends.cuda.math_sdp_enabled()

Warning

This flag is beta and subject to change.

Returns whether math scaled dot product attention is enabled or not.

torch.backends.cuda.enable_math_sdp(enabled)

Warning

This flag is beta and subject to change.

Enables or disables math scaled dot product attention.

torch.backends.cuda.cudnn_sdp_enabled()

Warning

This flag is beta and subject to change.

Returns whether cuDNN scaled dot product attention is enabled or not.

torch.backends.cuda.enable_cudnn_sdp(enabled)

Warning

This flag is beta and subject to change.

Enables or disables cuDNN scaled dot product attention.

torch.backends.cuda.can_use_flash_attention(params, debug=False)

Check if FlashAttention can be utilized in scaled_dot_product_attention.

Parameters

• params (_SDPAParams) – An instance of SDPAParams containing the tensors for query, key, value, an optional attention mask, dropout rate, and a flag indicating if the attention is causal.

• debug (bool) – Whether to logging.warn debug information as to why FlashAttention could not be run. Defaults to False.

Returns

True if FlashAttention can be used with the given parameters; otherwise, False.

Return type

bool

Note

This function is dependent on a CUDA-enabled build of PyTorch. It will return False in non-CUDA environments.

torch.backends.cuda.can_use_efficient_attention(params, debug=False)

Check if efficient_attention can be utilized in scaled_dot_product_attention.

Parameters

• params (_SDPAParams) – An instance of SDPAParams containing the tensors for query, key, value, an optional attention mask, dropout rate, and a flag indicating if the attention is causal.

• debug (bool) – Whether to logging.warn with information as to why efficient_attention could not be run. Defaults to False.

Returns

True if efficient_attention can be used with the given parameters; otherwise, False.

Return type

bool

Note

This function is dependent on a CUDA-enabled build of PyTorch. It will return False in non-CUDA environments.

torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=True, enable_mem_efficient=True, enable_cudnn=True)

Warning

This flag is beta and subject to change.

This context manager can be used to temporarily enable or disable any of the three backends for scaled dot product attention. Upon exiting the context manager, the previous state of the flags will be restored.

torch.backends.cudnn

torch.backends.cudnn.version()

Return the version of cuDNN.

torch.backends.cudnn.is_available()

Return a bool indicating if CUDNN is currently available.

torch.backends.cudnn.enabled

A bool that controls whether cuDNN is enabled.

torch.backends.cudnn.allow_tf32

A bool that controls where TensorFloat-32 tensor cores may be used in cuDNN convolutions on Ampere or newer GPUs. See TensorFloat-32 (TF32) on Ampere (and later) devices.

torch.backends.cudnn.deterministic

A bool that, if True, causes cuDNN to only use deterministic convolution algorithms. See also torch.are_deterministic_algorithms_enabled() and torch.use_deterministic_algorithms().

torch.backends.cudnn.benchmark

A bool that, if True, causes cuDNN to benchmark multiple convolution algorithms and select the fastest.

torch.backends.cudnn.benchmark_limit

A int that specifies the maximum number of cuDNN convolution algorithms to try when torch.backends.cudnn.benchmark is True. Set benchmark_limit to zero to try every available algorithm. Note that this setting only affects convolutions dispatched via the cuDNN v8 API.

torch.backends.mha

torch.backends.mha.get_fastpath_enabled()

Returns whether fast path for TransformerEncoder and MultiHeadAttention is enabled, or True if jit is scripting.

..note:

The fastpath might not be run even if get_fastpath_enabled returns True unless all conditions on inputs are met.

Return type

bool

torch.backends.mha.set_fastpath_enabled(value)

Sets whether fast path is enabled

torch.backends.mps

torch.backends.mps.is_available()

Return a bool indicating if MPS is currently available.

Return type

bool

torch.backends.mps.is_built()

Return whether PyTorch is built with MPS support.

Note that this doesn’t necessarily mean MPS is available; just that if this PyTorch binary were run a machine with working MPS drivers and devices, we would be able to use it.

Return type

bool

torch.backends.mkl

torch.backends.mkl.is_available()

Return whether PyTorch is built with MKL support.

class torch.backends.mkl.verbose(enable)

On-demand oneMKL verbosing functionality.

To make it easier to debug performance issues, oneMKL can dump verbose messages containing execution information like duration while executing the kernel. The verbosing functionality can be invoked via an environment variable named MKL_VERBOSE. However, this methodology dumps messages in all steps. Those are a large amount of verbose messages. Moreover, for investigating the performance issues, generally taking verbose messages for one single iteration is enough. This on-demand verbosing functionality makes it possible to control scope for verbose message dumping. In the following example, verbose messages will be dumped out for the second inference only.

import torch
model(data)
with torch.backends.mkl.verbose(torch.backends.mkl.VERBOSE_ON):
    model(data)

Parameters

level – Verbose level - VERBOSE_OFF: Disable verbosing - VERBOSE_ON: Enable verbosing

torch.backends.mkldnn

torch.backends.mkldnn.is_available()

Return whether PyTorch is built with MKL-DNN support.

class torch.backends.mkldnn.verbose(level)

On-demand oneDNN (former MKL-DNN) verbosing functionality.

To make it easier to debug performance issues, oneDNN can dump verbose messages containing information like kernel size, input data size and execution duration while executing the kernel. The verbosing functionality can be invoked via an environment variable named DNNL_VERBOSE. However, this methodology dumps messages in all steps. Those are a large amount of verbose messages. Moreover, for investigating the performance issues, generally taking verbose messages for one single iteration is enough. This on-demand verbosing functionality makes it possible to control scope for verbose message dumping. In the following example, verbose messages will be dumped out for the second inference only.

import torch
model(data)
with torch.backends.mkldnn.verbose(torch.backends.mkldnn.VERBOSE_ON):
    model(data)

Parameters

level – Verbose level - VERBOSE_OFF: Disable verbosing - VERBOSE_ON: Enable verbosing - VERBOSE_ON_CREATION: Enable verbosing, including oneDNN kernel creation

torch.backends.nnpack

torch.backends.nnpack.is_available()

Return whether PyTorch is built with NNPACK support.

torch.backends.nnpack.flags(enabled=False)

Context manager for setting if nnpack is enabled globally

torch.backends.nnpack.set_flags(_enabled)

Set if nnpack is enabled globally

torch.backends.openmp

torch.backends.openmp.is_available()

Return whether PyTorch is built with OpenMP support.

torch.backends.opt_einsum

torch.backends.opt_einsum.is_available()

Return a bool indicating if opt_einsum is currently available.

Return type

bool

torch.backends.opt_einsum.get_opt_einsum()

Return the opt_einsum package if opt_einsum is currently available, else None.

Return type

Any

torch.backends.opt_einsum.enabled

A :class:bool that controls whether opt_einsum is enabled (True by default). If so, torch.einsum will use opt_einsum (https://optimized-einsum.readthedocs.io/en/stable/path_finding.html) if available to calculate an optimal path of contraction for faster performance.

If opt_einsum is not available, torch.einsum will fall back to the default contraction path of left to right.

torch.backends.opt_einsum.strategy

A :class:str that specifies which strategies to try when torch.backends.opt_einsum.enabled is True. By default, torch.einsum will try the “auto” strategy, but the “greedy” and “optimal” strategies are also supported. Note that the “optimal” strategy is factorial on the number of inputs as it tries all possible paths. See more details in opt_einsum’s docs (https://optimized-einsum.readthedocs.io/en/stable/path_finding.html).

torch.backends.xeon