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

Jump to bottom

Fix row parallel lora layers parameters initialization bug #9427

Open
wants to merge 2 commits into
base: develop
Choose a base branch
from
Open

Fix row parallel lora layers parameters initialization bug #9427

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
1e2ff58
Fix row parallel lora layers parameters initialization bug
will-jl944 Nov 14, 2024
4892fff
Fix ColumnParallel LoRA layers
will-jl944 Nov 15, 2024
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
78 changes: 44 additions & 34 deletions paddlenlp/peft/lora/lora_layers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,7 @@
# limitations under the License.

import math
from contextlib import nullcontext
from typing import Optional

import paddle
Expand All @@ -22,6 +23,7 @@
from paddle.distributed.fleet.meta_parallel import (
ColumnParallelLinear,
RowParallelLinear,
get_rng_state_tracker,
)

from ...transformers import linear_utils
Expand Down Expand Up @@ -50,6 +52,10 @@
from .lora_quick_layers import quick_lora


def rng_ctx(is_mp: bool, in_dynamic_mode: bool):
return get_rng_state_tracker().rng_state() if (is_mp and in_dynamic_mode) else nullcontext()

Check warning on line 56 in paddlenlp/peft/lora/lora_layers.py

View check run for this annotation

Codecov / codecov/patch

paddlenlp/peft/lora/lora_layers.py#L56 

Added line #L56 was not covered by tests


class LoRALinear(nn.Linear):
# LoRA implemented in a dense layer
def __init__(
Expand Down Expand Up @@ -198,14 +204,15 @@
self.name = self._name

# Actual trainable parameters
self.lora_A = self.create_parameter(
shape=[self.input_size_per_partition, r],
dtype=self._dtype,
is_bias=False,
attr=paddle.ParamAttr(
initializer=nn.initializer.KaimingUniform(negative_slope=math.sqrt(5), nonlinearity="leaky_relu")
),
)
with rng_ctx(self.is_mp, paddle.in_dynamic_mode()):
self.lora_A = self.create_parameter(

Check warning on line 208 in paddlenlp/peft/lora/lora_layers.py

View check run for this annotation

Codecov / codecov/patch

paddlenlp/peft/lora/lora_layers.py#L207-L208 

Added lines #L207 - L208 were not covered by tests
shape=[self.input_size_per_partition, r],
dtype=self._dtype,
is_bias=False,
attr=paddle.ParamAttr(
initializer=nn.initializer.KaimingUniform(negative_slope=math.sqrt(5), nonlinearity="leaky_relu")
),
)
self.lora_B = self.create_parameter(
shape=[r, self.out_features],
dtype=self._dtype,
Expand Down Expand Up @@ -345,14 +352,15 @@
self.name = self._name

# Actual trainable parameters
self.lora_A = self.create_parameter(
shape=[self.input_size_per_partition, r],
dtype=self._dtype,
is_bias=False,
attr=paddle.ParamAttr(
initializer=nn.initializer.KaimingUniform(negative_slope=math.sqrt(5), nonlinearity="leaky_relu")
),
)
with rng_ctx(self.is_mp, paddle.in_dynamic_mode()):
self.lora_A = self.create_parameter(

Check warning on line 356 in paddlenlp/peft/lora/lora_layers.py

View check run for this annotation

Codecov / codecov/patch

paddlenlp/peft/lora/lora_layers.py#L355-L356 

Added lines #L355 - L356 were not covered by tests
shape=[self.input_size_per_partition, r],
dtype=self._dtype,
is_bias=False,
attr=paddle.ParamAttr(
initializer=nn.initializer.KaimingUniform(negative_slope=math.sqrt(5), nonlinearity="leaky_relu")
),
)
self.lora_B = self.create_parameter(
shape=[r, self.out_features],
dtype=self._dtype,
Expand Down Expand Up @@ -468,15 +476,16 @@
attr=lora_A_weight_attr,
)
self.lora_A.is_distributed = False
self.lora_B = self.create_parameter(
shape=[r, self.output_size_per_partition],
dtype=self._dtype,
is_bias=False,
attr=paddle.ParamAttr(
initializer=paddle.nn.initializer.Constant(value=0.0),
learning_rate=lora_plus_scale,
),
)
with rng_ctx(self.is_mp, paddle.in_dynamic_mode()):
self.lora_B = self.create_parameter(

Check warning on line 480 in paddlenlp/peft/lora/lora_layers.py

View check run for this annotation

Codecov / codecov/patch

paddlenlp/peft/lora/lora_layers.py#L479-L480 

Added lines #L479 - L480 were not covered by tests
shape=[r, self.output_size_per_partition],
dtype=self._dtype,
is_bias=False,
attr=paddle.ParamAttr(
initializer=paddle.nn.initializer.Constant(value=0.0),
learning_rate=lora_plus_scale,
),
)

self.lora_B.is_distributed = True
self.lora_B.split_axis = 1
Expand Down Expand Up @@ -599,15 +608,16 @@
self.lora_A.is_distributed = False
mark_as_sequence_parallel_parameter(self.lora_A)

self.lora_B = self.create_parameter(
shape=[r, self.output_size_per_partition],
dtype=self._dtype,
is_bias=False,
attr=paddle.ParamAttr(
initializer=paddle.nn.initializer.Constant(value=0.0),
learning_rate=lora_plus_scale,
),
)
with rng_ctx(self.is_mp, paddle.in_dynamic_mode()):
self.lora_B = self.create_parameter(

Check warning on line 612 in paddlenlp/peft/lora/lora_layers.py

View check run for this annotation

Codecov / codecov/patch

paddlenlp/peft/lora/lora_layers.py#L611-L612 

Added lines #L611 - L612 were not covered by tests
shape=[r, self.output_size_per_partition],
dtype=self._dtype,
is_bias=False,
attr=paddle.ParamAttr(
initializer=paddle.nn.initializer.Constant(value=0.0),
learning_rate=lora_plus_scale,
),
)

self.lora_B.is_distributed = True
self.lora_B.split_axis = 1
Expand Down
Loading