torch.future — PyTorch 2.7 documentation (original) (raw)
torch.__future__.set_overwrite_module_params_on_conversion(value)[source][source]¶
Sets whether to assign new tensors to the parameters instead of changing the existing parameters in-place when converting an nn.Module
.
When enabled, the following methods will assign new parameters to the module:
module.{device}()
(e.g.nn.Module.cuda()
) for moving a module between devicesmodule.{dtype}()
(e.g.nn.Module.float()
) for converting a module to a different dtypenn.Module.to()
nn.Module.to_empty()
Parameters
value (bool) – Whether to assign new tensors or not.
torch.__future__.get_overwrite_module_params_on_conversion()[source][source]¶
Returns whether to assign new tensors to the parameters instead of changing the existing parameters in-place when converting an torch.nn.Module. Defaults to False
.
See set_overwrite_module_params_on_conversion() for more information.
Return type
torch.__future__.set_swap_module_params_on_conversion(value)[source][source]¶
Sets whether to use swap_tensors() instead of setting .data
to change the existing parameters in-place when converting an nn.Module
and instead of param.copy_(state_dict[key])
when loading a state dict into an nn.Module
.
Note
This function takes precedence over get_overwrite_module_params_on_conversion()
When enabled, the following methods will swap the existing parameters in-place:
module.{device}()
(e.g.nn.Module.cuda()
) for moving a module between devicesmodule.{dtype}()
(e.g.nn.Module.float()
) for converting a module to a different dtypenn.Module.to()
nn.Module.to_empty()
nn.Module.load_state_dict()
The semantics for load_state_dict()
when this is set are as follows:
- For each parameter/buffer, its corresponding
state_dict['key']
is transformed viamodule_load() (i.e.res = param.module_load(state_dict['key'])
) - If necessary,
res
will be wrapped in anParameter
- The parameter/buffer in the module will be swapped via swap_tensors()with
res
Parameters
value (bool) – Whether to use swap_tensors() or not.
torch.__future__.get_swap_module_params_on_conversion()[source][source]¶
Returns whether to use swap_tensors() instead of setting .data to change the existing parameters in-place when converting an nn.Module
. Defaults to False
.
See set_swap_module_params_on_conversion() for more information.
Return type