158158
159159
160160 < div class ="version ">
161- < a href ='http://pytorch.org/docs/versions.html '> 1.7.0a0+8a83851 ▼</ a >
161+ < a href ='http://pytorch.org/docs/versions.html '> 1.7.0a0+be34aa1 ▼</ a >
162162 </ div >
163163
164164
@@ -642,10 +642,67 @@ <h1>Source code for torch</h1><div class="highlight"><pre>
642642 < span class ="n "> _C</ span > < span class ="o "> .</ span > < span class ="n "> _set_default_dtype</ span > < span class ="p "> (</ span > < span class ="n "> d</ span > < span class ="p "> )</ span >
643643
644644< span class ="k "> def</ span > < span class ="nf "> set_deterministic</ span > < span class ="p "> (</ span > < span class ="n "> d</ span > < span class ="p "> ):</ span >
645- < span class ="sa "> r</ span > < span class ="sd "> """Sets a global flag to force all operations to use a deterministic</ span >
646- < span class ="sd "> implementation if available. If an operation that does not have a</ span >
647- < span class ="sd "> deterministic implementation is called while this setting is True, the</ span >
648- < span class ="sd "> operation will throw a RuntimeError.</ span >
645+ < span class ="sa "> r</ span > < span class ="sd "> """ Sets whether native PyTorch operations must use deterministic</ span >
646+ < span class ="sd "> algorithms. When True, operations without deterministic algorithms</ span >
647+ < span class ="sd "> will throw a :class:RuntimeError when called.</ span >
648+
649+ < span class ="sd "> .. warning::</ span >
650+ < span class ="sd "> This feature is a beta feature, so it does not affect every</ span >
651+ < span class ="sd "> nondeterministic operation yet. The following operations are</ span >
652+ < span class ="sd "> affected by this flag.</ span >
653+
654+ < span class ="sd "> The following normally-nondeterministic operations will act</ span >
655+ < span class ="sd "> deterministically when `d=True`:</ span >
656+
657+ < span class ="sd "> * :class:`torch.nn.Conv1d` when called on CUDA tensor</ span >
658+ < span class ="sd "> * :class:`torch.nn.Conv2d` when called on CUDA tensor</ span >
659+ < span class ="sd "> * :class:`torch.nn.Conv3d` when called on CUDA tensor</ span >
660+ < span class ="sd "> * :class:`torch.nn.ConvTranspose1d` when called on CUDA tensor</ span >
661+ < span class ="sd "> * :class:`torch.nn.ConvTranspose2d` when called on CUDA tensor</ span >
662+ < span class ="sd "> * :class:`torch.nn.ConvTranspose3d` when called on CUDA tensor</ span >
663+ < span class ="sd "> * :func:`torch.bmm` when called on sparse-dense CUDA tensors</ span >
664+
665+ < span class ="sd "> The following normally-nondeterministic operations will throw a</ span >
666+ < span class ="sd "> :class:`RuntimeError` when `d=True`:</ span >
667+
668+ < span class ="sd "> * :class:`torch.nn.AvgPool3d` when called on a CUDA tensor that requires grad</ span >
669+ < span class ="sd "> * :class:`torch.nn.AdaptiveAvgPool2d` when called on a CUDA tensor that requires grad</ span >
670+ < span class ="sd "> * :class:`torch.nn.AdaptiveAvgPool3d` when called on a CUDA tensor that requires grad</ span >
671+ < span class ="sd "> * :class:`torch.nn.MaxPool3d` when called on a CUDA tensor that requires grad</ span >
672+ < span class ="sd "> * :class:`torch.nn.AdaptiveMaxPool2d` when called on a CUDA tensor that requires grad</ span >
673+ < span class ="sd "> * :class:`torch.nn.FractionalMaxPool2d` when called on a CUDA tensor that requires grad</ span >
674+ < span class ="sd "> * :class:`torch.nn.FractionalMaxPool3d` when called on a CUDA tensor that requires grad</ span >
675+ < span class ="sd "> * :func:`torch.nn.functional.interpolate` when called on a CUDA tensor that requires grad</ span >
676+ < span class ="sd "> and one of the following modes is used:</ span >
677+ < span class ="sd "> - `linear`</ span >
678+ < span class ="sd "> - `bilinear`</ span >
679+ < span class ="sd "> - `bicubic`</ span >
680+ < span class ="sd "> - `trilinear`</ span >
681+ < span class ="sd "> * :class:`torch.nn.ReflectionPad1d` when called on a CUDA tensor that requires grad</ span >
682+ < span class ="sd "> * :class:`torch.nn.ReflectionPad2d` when called on a CUDA tensor that requires grad</ span >
683+ < span class ="sd "> * :class:`torch.nn.ReplicationPad1d` when called on a CUDA tensor that requires grad</ span >
684+ < span class ="sd "> * :class:`torch.nn.ReplicationPad2d` when called on a CUDA tensor that requires grad</ span >
685+ < span class ="sd "> * :class:`torch.nn.ReplicationPad3d` when called on a CUDA tensor that requires grad</ span >
686+ < span class ="sd "> * :class:`torch.nn.NLLLoss` when called on a CUDA tensor that requires grad</ span >
687+ < span class ="sd "> * :class:`torch.nn.CTCLoss` when called on a CUDA tensor that requires grad</ span >
688+ < span class ="sd "> * :class:`torch.nn.EmbeddingBag` when called on a CUDA tensor that requires grad</ span >
689+ < span class ="sd "> * :func:`torch.scatter_add_` when called on a CUDA tensor</ span >
690+ < span class ="sd "> * :func:`torch.index_add_` when called on a CUDA tensor</ span >
691+ < span class ="sd "> * :func:`torch.index_select` when called on a CUDA tensor that requires grad</ span >
692+ < span class ="sd "> * :func:`torch.repeat_interleave` when called on a CUDA tensor that requires grad</ span >
693+ < span class ="sd "> * :func:`torch.histc` when called on a CUDA tensor</ span >
694+ < span class ="sd "> * :func:`torch.bincount` when called on a CUDA tensor</ span >
695+
696+ < span class ="sd "> A handful of CUDA operations are nondeterministic if the CUDA version is</ span >
697+ < span class ="sd "> 10.2 or greater, unless the environment variable `CUBLAS_WORKSPACE_CONFIG=:4096:8`</ span >
698+ < span class ="sd "> or `CUBLAS_WORKSPACE_CONFIG=:16:8` is set. See the CUDA documentation for more</ span >
699+ < span class ="sd "> details: `<https://docs.nvidia.com/cuda/cublas/index.html#cublasApi_reproducibility>`_</ span >
700+ < span class ="sd "> If one of these environment variable configurations is not set, a :class:`RuntimeError`</ span >
701+ < span class ="sd "> will be raised from these operations when called with CUDA tensors:</ span >
702+
703+ < span class ="sd "> * :func:`torch.mm`</ span >
704+ < span class ="sd "> * :func:`torch.mv`</ span >
705+ < span class ="sd "> * :func:`torch.bmm`</ span >
649706
650707< span class ="sd "> Note that deterministic operations tend to have worse performance than</ span >
651708< span class ="sd "> non-deterministic operations.</ span >
@@ -656,11 +713,11 @@ <h1>Source code for torch</h1><div class="highlight"><pre>
656713< span class ="sd "> """</ span >
657714 < span class ="n "> _C</ span > < span class ="o "> .</ span > < span class ="n "> _set_deterministic</ span > < span class ="p "> (</ span > < span class ="n "> d</ span > < span class ="p "> )</ span >
658715
659- < span class ="k "> def</ span > < span class ="nf "> is_deterministic</ span > < span class ="p "> ():</ span >
660- < span class ="sa "> r</ span > < span class ="sd "> """Returns True if the global deterministic flag is turned on and </ span >
661- < span class ="sd "> operations are being forced to use a deterministic implementation .</ span >
716+ < div class =" viewcode-block " id =" is_deterministic " > < a class =" viewcode-back " href =" ../generated/torch.is_deterministic.html#torch.is_deterministic " > [docs] </ a > < span class ="k "> def</ span > < span class ="nf "> is_deterministic</ span > < span class ="p "> ():</ span >
717+ < span class ="sa "> r</ span > < span class ="sd "> """Returns True if the global deterministic flag is turned on. Refer to </ span >
718+ < span class ="sd "> :func:`torch.set_deterministic` documentation for more details .</ span >
662719< span class ="sd "> """</ span >
663- < span class ="k "> return</ span > < span class ="n "> _C</ span > < span class ="o "> .</ span > < span class ="n "> _get_deterministic</ span > < span class ="p "> ()</ span >
720+ < span class ="k "> return</ span > < span class ="n "> _C</ span > < span class ="o "> .</ span > < span class ="n "> _get_deterministic</ span > < span class ="p "> ()</ span > </ div >
664721
665722< span class ="c1 "> ################################################################################</ span >
666723< span class ="c1 "> # Define Storage and Tensor classes</ span >
0 commit comments