242242 < div class ="pytorch-left-menu-search ">
243243
244244 < div class ="version ">
245- < a href ='https://pytorch.org/docs/versions.html '> main (2.3.0a0+git45c4a0c ) ▼</ a >
245+ < a href ='https://pytorch.org/docs/versions.html '> main (2.3.0a0+git34f7dc9 ) ▼</ a >
246246 </ div >
247247
248248
@@ -485,7 +485,7 @@ <h1>Source code for torch.__future__</h1><div class="highlight"><pre>
485485
486486< div class ="viewcode-block " id ="set_overwrite_module_params_on_conversion "> < a class ="viewcode-back " href ="../../future_mod.html#torch.__future__.set_overwrite_module_params_on_conversion "> [docs]</ a > < span class ="k "> def</ span > < span class ="nf "> set_overwrite_module_params_on_conversion</ span > < span class ="p "> (</ span > < span class ="n "> value</ span > < span class ="p "> :</ span > < span class ="nb "> bool</ span > < span class ="p "> )</ span > < span class ="o "> -></ span > < span class ="kc "> None</ span > < span class ="p "> :</ span >
487487< span class ="w "> </ span > < span class ="sd "> """</ span >
488- < span class ="sd "> Sets whether to assign new tensors to the parameters instead of changing the</ span >
488+ < span class ="sd "> Sets whether to assign new tensors to the parameters instead of changing the</ span >
489489< span class ="sd "> existing parameters in-place when converting an ``nn.Module``.</ span >
490490
491491< span class ="sd "> When enabled, the following methods will assign new parameters to the module:</ span >
@@ -495,6 +495,9 @@ <h1>Source code for torch.__future__</h1><div class="highlight"><pre>
495495< span class ="sd "> (for converting a module to a different dtype)</ span >
496496< span class ="sd "> #. ``module.to()``</ span >
497497
498+ < span class ="sd "> Args:</ span >
499+ < span class ="sd "> value (bool): Whether to assign new tensors or not.</ span >
500+
498501< span class ="sd "> """</ span >
499502 < span class ="k "> global</ span > < span class ="n "> _overwrite_module_params_on_conversion</ span >
500503 < span class ="n "> _overwrite_module_params_on_conversion</ span > < span class ="o "> =</ span > < span class ="n "> value</ span > </ div >
@@ -503,28 +506,41 @@ <h1>Source code for torch.__future__</h1><div class="highlight"><pre>
503506< div class ="viewcode-block " id ="get_overwrite_module_params_on_conversion "> < a class ="viewcode-back " href ="../../future_mod.html#torch.__future__.get_overwrite_module_params_on_conversion "> [docs]</ a > < span class ="k "> def</ span > < span class ="nf "> get_overwrite_module_params_on_conversion</ span > < span class ="p "> ()</ span > < span class ="o "> -></ span > < span class ="nb "> bool</ span > < span class ="p "> :</ span >
504507< span class ="w "> </ span > < span class ="sd "> """</ span >
505508< span class ="sd "> Returns whether to assign new tensors to the parameters instead of changing the</ span >
506- < span class ="sd "> existing parameters in-place when converting an ``nn.Module`. Defaults to ``False``.</ span >
509+ < span class ="sd "> existing parameters in-place when converting an ``nn.Module`` . Defaults to ``False``.</ span >
507510
508- < span class ="sd "> See :func:`~torch.nn.utils .set_overwrite_module_params_on_conversion` for more information.</ span >
511+ < span class ="sd "> See :func:`~torch.__future__ .set_overwrite_module_params_on_conversion` for more information.</ span >
509512< span class ="sd "> """</ span >
510513 < span class ="k "> return</ span > < span class ="n "> _overwrite_module_params_on_conversion</ span > </ div >
511514
512515
513516< div class ="viewcode-block " id ="set_swap_module_params_on_conversion "> < a class ="viewcode-back " href ="../../future_mod.html#torch.__future__.set_swap_module_params_on_conversion "> [docs]</ a > < span class ="k "> def</ span > < span class ="nf "> set_swap_module_params_on_conversion</ span > < span class ="p "> (</ span > < span class ="n "> value</ span > < span class ="p "> :</ span > < span class ="nb "> bool</ span > < span class ="p "> )</ span > < span class ="o "> -></ span > < span class ="kc "> None</ span > < span class ="p "> :</ span >
514517< span class ="w "> </ span > < span class ="sd "> """</ span >
515518< span class ="sd "> Sets whether to use :func:`~torch.utils.swap_tensors` instead of setting ``.data`` to</ span >
516- < span class ="sd "> change the existing parameters in-place when converting an ``nn.Module``.</ span >
519+ < span class ="sd "> change the existing parameters in-place when converting an ``nn.Module`` and instead</ span >
520+ < span class ="sd "> of ``param.copy_(state_dict[key])`` when loading a state dict into an ``nn.Module``.</ span >
517521
518522< span class ="sd "> .. note::</ span >
519523< span class ="sd "> If :func:`~torch.__future__.get_overwrite_module_params_on_conversion` returns ``True``,</ span >
520- < span class ="sd "> no swapping will occur.</ span >
524+ < span class ="sd "> for methods other than :meth:`~nn.Module.load_state_dict` no swapping will occur.</ span >
521525
522526< span class ="sd "> When enabled, the following methods will swap the existing parameters in-place:</ span >
523527
524528< span class ="sd "> #. ``module.{device}()`` (e.g. ``module.cuda()``) for moving a module between devices</ span >
525529< span class ="sd "> #. ``module.{dtype}()`` (e.g. ``module.float()``) for converting a module to a different dtype</ span >
526530< span class ="sd "> (for converting a module to a different dtype)</ span >
527531< span class ="sd "> #. ``module.to()``</ span >
532+ < span class ="sd "> #. ``module.load_state_dict(state_dict)``</ span >
533+
534+ < span class ="sd "> The semantics for :meth:`~nn.Module.load_state_dict` when this is set are as follows:</ span >
535+
536+ < span class ="sd "> #. For each parameter/buffer, its corresponding``state_dict['key']`` is transformed via</ span >
537+ < span class ="sd "> :meth:`~torch.Tensor.module_load` (i.e. ``res = param.module_load(state_dict['key'])``)</ span >
538+ < span class ="sd "> #. If necessary, ``res`` will be wrapped in an :class:`~nn.Parameter`</ span >
539+ < span class ="sd "> #. The parameter/buffer in the module will be swapped via :func:`~torch.utils.swap_tensors`</ span >
540+ < span class ="sd "> with ``res``</ span >
541+
542+ < span class ="sd "> Args:</ span >
543+ < span class ="sd "> value (bool): Whether to use :func:`~torch.utils.swap_tensors` or not.</ span >
528544
529545< span class ="sd "> """</ span >
530546 < span class ="k "> global</ span > < span class ="n "> _swap_module_params_on_conversion</ span >
@@ -534,9 +550,9 @@ <h1>Source code for torch.__future__</h1><div class="highlight"><pre>
534550< div class ="viewcode-block " id ="get_swap_module_params_on_conversion "> < a class ="viewcode-back " href ="../../future_mod.html#torch.__future__.get_swap_module_params_on_conversion "> [docs]</ a > < span class ="k "> def</ span > < span class ="nf "> get_swap_module_params_on_conversion</ span > < span class ="p "> ()</ span > < span class ="o "> -></ span > < span class ="nb "> bool</ span > < span class ="p "> :</ span >
535551< span class ="w "> </ span > < span class ="sd "> """</ span >
536552< span class ="sd "> Returns whether to use :func:`~torch.utils.swap_tensors` instead of setting .data to</ span >
537- < span class ="sd "> change the existing parameters in-place when converting an nn.Module. Defaults to ``False``.</ span >
553+ < span class ="sd "> change the existing parameters in-place when converting an `` nn.Module`` . Defaults to ``False``.</ span >
538554
539- < span class ="sd "> See :func:`~torch.nn.utils .set_swap_module_params_on_conversion` for more information.</ span >
555+ < span class ="sd "> See :func:`~torch.__future__ .set_swap_module_params_on_conversion` for more information.</ span >
540556< span class ="sd "> """</ span >
541557 < span class ="k "> return</ span > < span class ="n "> _swap_module_params_on_conversion</ span > </ div >
542558</ pre > </ div >
0 commit comments