frozenleaves/accelerate
1
0
Fork 0
mirror of https://github.com/huggingface/accelerate.git synced 2025-10-20 10:03:46 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix (skip) cache flush when original device is cpu and offloaded to disk meta (#3796)

Browse Source
This commit is contained in:
Qubitium-ModelCloud
2025-10-08 17:48:04 +08:00
committed by GitHub
parent 14383311c2
commit 057edec226
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/accelerate/utils/modeling.py
View File

@ -404,7 +404,7 @@ def set_module_tensor_to_device(
module.weight = module.weight.cuda(device_index)
# clean pre and post forward hook
if clear_cache and device != "cpu":
if clear_cache and device not in ("cpu", "meta"):
clear_device_cache()
# When handling tied weights, we update tied_params_map to keep track of the tied weights that have already been allocated on the device in