Skip to content

hack offload optimizer减少一次master weight的offload&reload #11111

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 2 commits into
base: develop
Choose a base branch
from
+9 −0
Open

hack offload optimizer减少一次master weight的offload&reload #11111

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
177c96a
hack offload optimizer减少一次master weight的offload&reload
Wennie396 Sep 23, 2025
74878bd
revert change
Wennie396 Sep 24, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 9 additions & 0 deletions paddlenlp/trainer/utils/offload_optimizer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,6 +11,7 @@
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os

import paddle
from paddle import _C_ops
Expand All @@ -37,6 +38,10 @@ def reload(tensor):
assert new_tensor is tensor, "to_device must be inplace operation"


def is_offload_opt_cache_master_weight():
return os.getenv("FLAGS_offload_opt_master_weight_cache", "0").lower() in ["true", "1"]


def hack_offload_optimizer():
# Step 1: mock _add_accumulator
origin_add_accumulator = getattr(Optimizer, "_add_accumulator")
Expand All @@ -60,6 +65,10 @@ def new_opt_op(*args):
ret = origin_op(*args)
is_offload_opt = getattr(args[0], "is_offload_opt", False)
for i, arg in enumerate(args):
# need_offload_arg = i >= 2 and isinstance(arg, paddle.Tensor) and is_offload_opt
# if is_offload_opt_cache_master_weight():
# need_offload_arg = need_offload_arg and i != 8
# if need_offload_arg: # do not offload parameter and gradient
if (
i >= 2 and isinstance(arg, paddle.Tensor) and is_offload_opt
): # do not offload parameter and gradient
Expand Down
Loading