Skip to content
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

Fix master weight bug for multi_tensor optimizer(momentum, adam) #38991

Merged
merged 2 commits into from
Jan 20, 2022
Merged

Fix master weight bug for multi_tensor optimizer(momentum, adam) #38991

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
98194d5
fix mp
zhangbo9674 Jan 17, 2022
4159708
support merged_momentum for mp
zhangbo9674 Jan 18, 2022
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: 4 additions & 5 deletions python/paddle/optimizer/adam.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -551,8 +551,7 @@ def _append_optimize_multi_tensor_op(self, target_block,
multi_tensor_list = ['FP32_LODTensor', 'FP16_LODTensor']
for key in multi_tensor_list:
if len(self._param_dict[key]) > 0:
if key == 'FP32_LODTensor':
self._multi_precision = False
find_master = self._multi_precision and key == 'FP16_LODTensor'

_beta1 = self._beta1 if not isinstance(
self._beta1, Variable) else self._beta1.numpy().item(0)
Expand All @@ -571,7 +570,7 @@ def _append_optimize_multi_tensor_op(self, target_block,
self._beta2_pow_acc_dict[key],
self._master_weight_dict[key], 'epsilon', self._epsilon,
'beta1', _beta1, 'beta2', _beta2, 'multi_precision',
self._multi_precision)
find_master)
else:
inputs = {
"Param": self._param_dict[key],
Expand All @@ -594,11 +593,11 @@ def _append_optimize_multi_tensor_op(self, target_block,
"beta1": _beta1,
"beta2": _beta2
}
if self._multi_precision:
if find_master:
inputs["MasterParam"] = self._master_weight_dict[key]
outputs["MasterParamOut"] = self._master_weight_dict[
key]
attrs["multi_precision"] = self._multi_precision
attrs["multi_precision"] = find_master
target_block.append_op(
type="merged_adam",
inputs=inputs,
Expand Down
9 changes: 4 additions & 5 deletions python/paddle/optimizer/momentum.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -464,8 +464,7 @@ def _append_optimize_multi_tensor_op(self, target_block,
multi_tensor_list = ['FP32_LODTensor', 'FP16_LODTensor']
for key in multi_tensor_list:
if len(self._param_dict[key]) > 0:
if key == 'FP32_LODTensor':
self._multi_precision = False
find_master = self._multi_precision and key == 'FP16_LODTensor'

if framework.in_dygraph_mode():
_, _, _ = _C_ops.merged_momentum(
Expand All @@ -478,7 +477,7 @@ def _append_optimize_multi_tensor_op(self, target_block,
self._regularization_method_dict[key],
'regularization_coeff',
self._regularization_coeff_dict[key], 'multi_precision',
self._multi_precision)
find_master)
else:
inputs = {
"Param": self._param_dict[key],
Expand All @@ -498,11 +497,11 @@ def _append_optimize_multi_tensor_op(self, target_block,
"regularization_coeff":
self._regularization_coeff_dict[key],
}
if self._multi_precision:
if find_master:
inputs["MasterParam"] = self._master_weight_dict[key]
outputs["MasterParamOut"] = self._master_weight_dict[
key]
attrs["multi_precision"] = self._multi_precision
attrs["multi_precision"] = find_master
target_block.append_op(
type="merged_momentum",
inputs=inputs,
Expand Down