This adds Python functions to get, set, and reset cuBLAS or cuBLASLt workspace sizes. Workspaces are updated lazily the next time they are accessed if the new workspace size is greater than the current size. @ngimel, is this what you were hoping for? I believe the current BLAS backend can already be queried with `torch.backends.cuda.preferred_blas_library`. Pull Request resolved: https://github.com/pytorch/pytorch/pull/177912 Approved by: https://github.com/ngimel, https://github.com/malfet, https://github.com/eqy
12 KiB
.. role:: hidden
:class: hidden-section
torch.backends
.. automodule:: torch.backends
torch.backends controls the behavior of various backends that PyTorch supports.
These backends include:
torch.backends.cputorch.backends.cudatorch.backends.cudnntorch.backends.cusparselttorch.backends.mhatorch.backends.mpstorch.backends.mkltorch.backends.mkldnntorch.backends.nnpacktorch.backends.openmptorch.backends.opt_einsumtorch.backends.python_nativetorch.backends.xeon
torch.backends.cpu
.. automodule:: torch.backends.cpu
.. autofunction:: torch.backends.cpu.get_cpu_capability
torch.backends.cuda
.. automodule:: torch.backends.cuda
.. autofunction:: torch.backends.cuda.is_built
.. currentmodule:: torch.backends.cuda.matmul
.. attribute:: allow_tf32
A :class:`bool` that controls whether TensorFloat-32 tensor cores may be used in matrix
multiplications on Ampere or newer GPUs. allow_tf32 is going to be deprecated. See :ref:`tf32_on_ampere`.
.. attribute:: allow_fp16_reduced_precision_reduction
A :class:`bool` that controls whether reduced precision reductions (e.g., with fp16 accumulation type) are allowed with fp16 GEMMs.
Assigning a tuple ``(allow_reduced_precision, allow_splitk)`` lets you also toggle whether
split-K heuristics may be used when dispatching to cuBLASLt. ``allow_splitk`` defaults to ``True``.
.. attribute:: allow_bf16_reduced_precision_reduction
A :class:`bool` that controls whether reduced precision reductions are allowed with bf16 GEMMs.
Assigning a tuple ``(allow_reduced_precision, allow_splitk)`` lets you also toggle whether
split-K heuristics may be used when dispatching to cuBLASLt. ``allow_splitk`` defaults to ``True``.
.. currentmodule:: torch.backends.cuda
.. attribute:: 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]`.
.. currentmodule:: torch.backends.cuda.cufft_plan_cache
.. attribute:: size
A readonly :class:`int` that shows the number of plans currently in a cuFFT plan cache.
.. attribute:: max_size
A :class:`int` that controls the capacity of a cuFFT plan cache.
.. method:: clear()
Clears a cuFFT plan cache.
.. autofunction:: torch.backends.cuda.preferred_blas_library
.. autofunction:: torch.backends.cuda.cublas_workspace_size
.. autofunction:: torch.backends.cuda.cublaslt_workspace_size
.. autofunction:: torch.backends.cuda.blas_workspace_size
.. autofunction:: torch.backends.cuda.preferred_rocm_fa_library
.. autofunction:: torch.backends.cuda.is_ck_sdpa_available
.. autofunction:: torch.backends.cuda.preferred_linalg_library
.. autoclass:: torch.backends.cuda.SDPAParams
.. autofunction:: torch.backends.cuda.flash_sdp_enabled
.. autofunction:: torch.backends.cuda.enable_mem_efficient_sdp
.. autofunction:: torch.backends.cuda.mem_efficient_sdp_enabled
.. autofunction:: torch.backends.cuda.enable_flash_sdp
.. autofunction:: torch.backends.cuda.math_sdp_enabled
.. autofunction:: torch.backends.cuda.enable_math_sdp
.. autofunction:: torch.backends.cuda.fp16_bf16_reduction_math_sdp_allowed
.. autofunction:: torch.backends.cuda.allow_fp16_bf16_reduction_math_sdp
.. autofunction:: torch.backends.cuda.cudnn_sdp_enabled
.. autofunction:: torch.backends.cuda.enable_cudnn_sdp
.. autofunction:: torch.backends.cuda.is_flash_attention_available
.. autofunction:: torch.backends.cuda.can_use_flash_attention
.. autofunction:: torch.backends.cuda.can_use_efficient_attention
.. autofunction:: torch.backends.cuda.can_use_cudnn_attention
.. autofunction:: torch.backends.cuda.sdp_kernel
torch.backends.cudnn
.. automodule:: torch.backends.cudnn
.. autofunction:: torch.backends.cudnn.version
.. autofunction:: torch.backends.cudnn.is_available
.. attribute:: enabled
A :class:`bool` that controls whether cuDNN is enabled.
.. attribute:: allow_tf32
A :class:`bool` that controls where TensorFloat-32 tensor cores may be used in cuDNN
convolutions on Ampere or newer GPUs. allow_tf32 is going to be deprecated. See :ref:`tf32_on_ampere`.
.. attribute:: deterministic
A :class:`bool` that, if True, causes cuDNN to only use deterministic convolution algorithms.
See also :func:`torch.are_deterministic_algorithms_enabled` and
:func:`torch.use_deterministic_algorithms`.
.. attribute:: benchmark
A :class:`bool` that, if True, causes cuDNN to benchmark multiple convolution algorithms
and select the fastest.
.. attribute:: benchmark_limit
A :class:`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.
.. py:module:: torch.backends.cudnn.rnn
torch.backends.cusparselt
.. automodule:: torch.backends.cusparselt
.. autofunction:: torch.backends.cusparselt.version
.. autofunction:: torch.backends.cusparselt.is_available
torch.backends.mha
.. automodule:: torch.backends.mha
.. autofunction:: torch.backends.mha.get_fastpath_enabled
.. autofunction:: torch.backends.mha.set_fastpath_enabled
torch.backends.miopen
.. automodule:: torch.backends.miopen
.. attribute:: immediate
A :class:`bool` that, if True, causes MIOpen to use Immediate Mode
(https://rocm.docs.amd.com/projects/MIOpen/en/latest/how-to/find-and-immediate.html).
torch.backends.mps
.. automodule:: torch.backends.mps
.. autofunction:: torch.backends.mps.is_available
.. autofunction:: torch.backends.mps.is_built
torch.backends.mkl
.. automodule:: torch.backends.mkl
.. autofunction:: torch.backends.mkl.is_available
.. autoclass:: torch.backends.mkl.verbose
torch.backends.mkldnn
.. automodule:: torch.backends.mkldnn
.. autofunction:: torch.backends.mkldnn.is_available
.. autoclass:: torch.backends.mkldnn.verbose
torch.backends.nnpack
.. automodule:: torch.backends.nnpack
.. autofunction:: torch.backends.nnpack.is_available
.. autofunction:: torch.backends.nnpack.flags
.. autofunction:: torch.backends.nnpack.set_flags
torch.backends.openmp
.. automodule:: torch.backends.openmp
.. autofunction:: torch.backends.openmp.is_available
% Docs for other backends need to be added here. % Automodules are just here to ensure checks run but they don't actually % add anything to the rendered page for now.
.. py:module:: torch.backends.quantized
.. py:module:: torch.backends.xnnpack
.. py:module:: torch.backends.kleidiai
.. autofunction:: torch.backends.kleidiai.is_available
torch.backends.opt_einsum
.. automodule:: torch.backends.opt_einsum
.. autofunction:: torch.backends.opt_einsum.is_available
.. autofunction:: torch.backends.opt_einsum.get_opt_einsum
.. attribute:: 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.
.. attribute:: 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.python_native
.. automodule:: torch.backends.python_native
The torch.backends.python_native module provides user control over native operators implemented in python
via. DSLs (Domain Specific Languages) that are defined in torch._native. This allows users to selectively
enable or disable high-performance implementations from various DSLs like Triton and CuteDSL.
Module-level Functions
.. autofunction:: torch.backends.python_native.get_dsl_operations
.. autofunction:: torch.backends.python_native.disable_operations
.. autofunction:: torch.backends.python_native.enable_operations
.. autofunction:: torch.backends.python_native.disable_dispatch_keys
.. autofunction:: torch.backends.python_native.enable_dispatch_keys
.. autofunction:: torch.backends.python_native.operations_disabled
Module-level Properties
.. attribute:: available_dsls
A :class:`list` of :class:`str` containing the names of DSLs that are available at runtime.
This is a subset of :attr:`all_dsls` that have their runtime dependencies satisfied.
.. attribute:: all_dsls
A :class:`list` of :class:`str` containing the names of all registered DSLs, whether
available at runtime or not.
DSL Controllers
For each registered DSL (e.g., triton, cutedsl), auto-populated controller modules are available:
.. currentmodule:: torch.backends.python_native
DSL Properties
Each DSL controller (e.g., torch.backends.python_native.triton) provides the following properties:
| Property | Type | Description |
|---|---|---|
name |
str |
The name of the DSL |
available |
bool |
Whether the DSL's runtime dependencies are available |
enabled |
bool |
Controls whether all operations from this DSL are enabled. Setting to False disables all operations from the DSL, while True re-enables them |
version |
Version or None |
The version of the DSL runtime, if available. Returns None if the DSL is not available |
DSL Methods
Each DSL controller provides the following methods:
disable() Disable all operations from this DSL.
enable() Re-enable all operations from this DSL.
disabled() Context manager that temporarily disables all operations from this DSL. Operations are automatically re-enabled when exiting the context.
Example::
with torch.backends.python_native.triton.disabled():
# Triton operations are disabled here
result = model(input)
# Triton operations restored here
Usage Examples
.. code-block:: python
import torch.backends.python_native as pn
# Query available DSLs
print(pn.available_dsls) # ['triton', 'cutedsl']
# Disable all Triton operations
pn.triton.enabled = False
# Temporarily disable CuteDSL operations
with pn.cutedsl.disabled():
result = model(input) # CuteDSL ops disabled
# Disable specific operations across all DSLs
pn.disable_operations('scaled_mm', '_flash_attention_forward')
# Query operations for a specific DSL
triton_ops = pn.get_dsl_operations('triton')
torch.backends.xeon
.. automodule:: torch.backends.xeon
.. py:module:: torch.backends.xeon.run_cpu