Skip to content

[BugFix] Fix qk_norm optimization #6080

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

Merged
zhoutianzi666 merged 1 commit into from
Jan 19, 2026
+5 −4
Merged

[BugFix] Fix qk_norm optimization #6080

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
7 changes: 4 additions & 3 deletions fastdeploy/model_executor/layers/normalization.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,6 +20,7 @@
import paddle
from paddle import nn

from fastdeploy.model_executor.forward_meta import ForwardMeta
from fastdeploy.platforms import current_platform

if current_platform.is_gcu():
Expand All @@ -28,7 +29,6 @@
from paddle.incubate.nn.functional import fused_layer_norm, fused_rms_norm

from fastdeploy.config import FDConfig
from fastdeploy.model_executor.forward_meta import ForwardMeta
from fastdeploy.model_executor.ops.triton_ops import _TRITON_AVAILABLE, qk_rmsnorm_fused

from .utils import get_tensor
Expand Down Expand Up @@ -307,7 +307,7 @@ def __init__(
prefix=f"{prefix}.k_norm",
begin_norm_axis=begin_norm_axis,
)
self.qk_norm_fused = _TRITON_AVAILABLE
self.qk_norm_fused = current_platform.is_cuda() and _TRITON_AVAILABLE

def load_state_dict(self, state_dict):
self.q_norm.load_state_dict(state_dict)
Expand All @@ -316,8 +316,9 @@ def load_state_dict(self, state_dict):
def forward(
self,
qkv_out,
forward_meta,
) -> paddle.Tensor:
if self.qk_norm_fused:
if self.qk_norm_fused and forward_meta.step_use_cudagraph:
qkv_out = qk_rmsnorm_fused(
qkv_out,
self.q_norm.weight,
Expand Down
2 changes: 1 addition & 1 deletion fastdeploy/model_executor/models/qwen3.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -103,7 +103,7 @@ def forward(
):
""" """
qkv_out = self.qkv_proj(hidden_states)
qkv_out = self.qk_norm(qkv_out)
qkv_out = self.qk_norm(qkv_out, forward_meta)
atten_out = self.attn(
qkv=qkv_out,
forward_meta=forward_meta,
Expand Down
Loading