mirror of
https://github.com/pytorch/pytorch.git
synced 2025-10-20 21:14:14 +08:00
1022443168
Test Plan: revert-hammer
Differential Revision:
D30279364 (b004307252)
Original commit changeset: c1ed77dfe43a
fbshipit-source-id: eab50857675c51e0088391af06ec0ecb14e2347e
20 lines
813 B
Python
20 lines
813 B
Python
"""
|
|
This global flag controls whether to assign new tensors to the parameters
|
|
instead of changing the existing parameters in-place when converting an `nn.Module`
|
|
using the following methods:
|
|
1. `module.cuda()` / `.cpu()` (for moving `module` between devices)
|
|
2. `module.float()` / `.double()` / `.half()` (for converting `module` to a different dtype)
|
|
3. `module.to()` / `.type()` (for changing `module`'s device or dtype)
|
|
4. `module._apply(fn)` (for generic functions applied to `module`)
|
|
|
|
Default: False
|
|
"""
|
|
_overwrite_module_params_on_conversion = False
|
|
|
|
def set_overwrite_module_params_on_conversion(value):
|
|
global _overwrite_module_params_on_conversion
|
|
_overwrite_module_params_on_conversion = value
|
|
|
|
def get_overwrite_module_params_on_conversion():
|
|
return _overwrite_module_params_on_conversion
|