Skip to content

fix sdpa attn mask with 1 channel #6402

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
merged 3 commits into from
Nov 12, 2025
Merged

fix sdpa attn mask with 1 channel #6402

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
cc98196
fix sdpa attn mask with 1 channel
nihui Nov 12, 2025
249a1c5
test++
nihui Nov 12, 2025
bdbf4ba
unpool
nihui Nov 12, 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
4 changes: 2 additions & 2 deletions src/layer/sdpa.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -94,7 +94,7 @@ int SDPA::forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_bl

if (attn_mask)
{
const Mat& maskm = attn_mask_blob.dims == 3 ? attn_mask_blob.channel(q) : attn_mask_blob;
const Mat& maskm = attn_mask_blob.c > 1 ? attn_mask_blob.channel(q) : attn_mask_blob;

for (int i = 0; i < src_seqlen; i++)
{
Expand Down Expand Up @@ -317,7 +317,7 @@ int SDPA::forward_int8(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& t

if (attn_mask)
{
const Mat& maskm = attn_mask_blob.dims == 3 ? attn_mask_blob.channel(q) : attn_mask_blob;
const Mat& maskm = attn_mask_blob.c > 1 ? attn_mask_blob.channel(q) : attn_mask_blob;

for (int i = 0; i < src_seqlen; i++)
{
Expand Down
15 changes: 10 additions & 5 deletions tools/pnnx/tests/ncnn/test_F_scaled_dot_product_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,18 +10,22 @@ class Model(nn.Module):
def __init__(self):
super(Model, self).__init__()

def forward(self, q, k, v, m, k2, v2):
def forward(self, q, k, v, m, k2, v2, m2):
x = F.scaled_dot_product_attention(q, k, v)
y = F.scaled_dot_product_attention(q, k, v, attn_mask=m)

if version.parse(torch.__version__) >= version.parse('2.5'):
z = F.scaled_dot_product_attention(q, k2, v2, enable_gqa=True)
z2 = F.scaled_dot_product_attention(q, k2, v2, attn_mask=m2, enable_gqa=True)
else:
k2_stack = k2.repeat_interleave(q.size(-3)//k2.size(-3), -3)
v2_stack = v2.repeat_interleave(q.size(-3)//v2.size(-3), -3)
z = F.scaled_dot_product_attention(q, k2_stack, v2_stack)
k2_stack = k2.clone().repeat_interleave(q.size(-3)//k2.size(-3), -3)
v2_stack = v2.clone().repeat_interleave(q.size(-3)//v2.size(-3), -3)
z2 = F.scaled_dot_product_attention(q, k2_stack, v2_stack, attn_mask=m2)

return x, y, z
return x, y, z, z2

def test():
if version.parse(torch.__version__) < version.parse('2.0'):
Expand All @@ -37,16 +41,17 @@ def test():
m = torch.rand(1, 8, 128, 48)
k2 = torch.rand(1, 2, 48, 64)
v2 = torch.rand(1, 2, 48, 77)
m2 = torch.rand(1, 1, 128, 48)

a = net(q, k, v, m, k2, v2)
a = net(q, k, v, m, k2, v2, m2)

# export torchscript
mod = torch.jit.trace(net, (q, k, v, m, k2, v2))
mod = torch.jit.trace(net, (q, k, v, m, k2, v2, m2))
mod.save("test_F_scaled_dot_product_attention.pt")

# torchscript to pnnx
import os
os.system("../../src/pnnx test_F_scaled_dot_product_attention.pt inputshape=[1,8,128,64],[1,8,48,64],[1,8,48,77],[1,8,128,48],[1,2,48,64],[1,2,48,77]")
os.system("../../src/pnnx test_F_scaled_dot_product_attention.pt inputshape=[1,8,128,64],[1,8,48,64],[1,8,48,77],[1,8,128,48],[1,2,48,64],[1,2,48,77],[1,1,128,48]")

# ncnn inference
import test_F_scaled_dot_product_attention_ncnn
Expand Down
13 changes: 8 additions & 5 deletions tools/pnnx/tests/test_F_scaled_dot_product_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10,18 +10,20 @@ class Model(nn.Module):
def __init__(self):
super(Model, self).__init__()

def forward(self, q, k, v, m, k2, v2):
def forward(self, q, k, v, m, k2, v2, m2):
x = F.scaled_dot_product_attention(q, k, v)
y = F.scaled_dot_product_attention(q, k, v, attn_mask=m)

if version.parse(torch.__version__) >= version.parse('2.5'):
z = F.scaled_dot_product_attention(q, k2, v2, enable_gqa=True)
z2 = F.scaled_dot_product_attention(q, k2, v2, attn_mask=m2, enable_gqa=True)
else:
k2_stack = k2.repeat_interleave(q.size(-3)//k2.size(-3), -3)
v2_stack = v2.repeat_interleave(q.size(-3)//v2.size(-3), -3)
z = F.scaled_dot_product_attention(q, k2_stack, v2_stack)
z2 = F.scaled_dot_product_attention(q, k2_stack, v2_stack, attn_mask=m2)

return x, y, z
return x, y, z, z2

def test():
if version.parse(torch.__version__) < version.parse('2.0'):
Expand All @@ -37,16 +39,17 @@ def test():
m = torch.rand(3, 8, 128, 48)
k2 = torch.rand(3, 2, 48, 64)
v2 = torch.rand(3, 2, 48, 77)
m2 = torch.rand(3, 1, 128, 48)

a = net(q, k, v, m, k2, v2)
a = net(q, k, v, m, k2, v2, m2)

# export torchscript
mod = torch.jit.trace(net, (q, k, v, m, k2, v2))
mod = torch.jit.trace(net, (q, k, v, m, k2, v2, m2))
mod.save("test_F_scaled_dot_product_attention.pt")

# torchscript to pnnx
import os
os.system("../src/pnnx test_F_scaled_dot_product_attention.pt inputshape=[3,8,128,64],[3,8,48,64],[3,8,48,77],[3,8,128,48],[3,2,48,64],[3,2,48,77]")
os.system("../src/pnnx test_F_scaled_dot_product_attention.pt inputshape=[3,8,128,64],[3,8,48,64],[3,8,48,77],[3,8,128,48],[3,2,48,64],[3,2,48,77],[3,1,128,48]")

# pnnx inference
import test_F_scaled_dot_product_attention_pnnx
Expand Down
Loading