Skip to content

Fix SDPA dropout during eval #160

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

Fix SDPA dropout during eval #160

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
8d175bd
fix stochastic in inference
massimilianoviola May 6, 2026
ac6ccbe
also fix vjepa2
massimilianoviola May 6, 2026
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 app/vjepa_2_1/models/utils/modules.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -285,7 +285,7 @@ def forward(
if self.use_sdpa:
with torch.backends.cuda.sdp_kernel():
x = F.scaled_dot_product_attention(
q, k, v, dropout_p=self.proj_drop_prob, is_causal=self.is_causal
q, k, v, dropout_p=self.proj_drop_prob if self.training else 0.0, is_causal=self.is_causal
)
attn = None
else:
Expand Down Expand Up @@ -338,7 +338,7 @@ def forward(self, x):
if self.use_sdpa:
with torch.backends.cuda.sdp_kernel():
x = F.scaled_dot_product_attention(
q, k, v, dropout_p=self.proj_drop_prob, is_causal=self.is_causal
q, k, v, dropout_p=self.proj_drop_prob if self.training else 0.0, is_causal=self.is_causal
)
attn = None
else:
Expand Down
9 changes: 6 additions & 3 deletions src/models/utils/modules.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -248,7 +248,8 @@ def merge_(tx, ta):
if attn_mask is not None or self.use_sdpa:
with torch.backends.cuda.sdp_kernel():
x = F.scaled_dot_product_attention(
q, k, v, dropout_p=self.proj_drop_prob, is_causal=self.is_causal, attn_mask=attn_mask
q, k, v, dropout_p=self.proj_drop_prob if self.training else 0.0,
is_causal=self.is_causal, attn_mask=attn_mask
)
attn = None
else:
Expand Down Expand Up @@ -372,7 +373,8 @@ def forward(self, x, mask=None, attn_mask=None, T=None, H_patches=None, W_patche
if attn_mask is not None or self.use_sdpa:
with torch.backends.cuda.sdp_kernel():
x = F.scaled_dot_product_attention(
q, k, v, dropout_p=self.proj_drop_prob, is_causal=self.is_causal, attn_mask=attn_mask
q, k, v, dropout_p=self.proj_drop_prob if self.training else 0.0,
is_causal=self.is_causal, attn_mask=attn_mask
)
attn = None
else:
Expand Down Expand Up @@ -419,7 +421,8 @@ def forward(self, x, mask=None, attn_mask=None):
if attn_mask is not None or self.use_sdpa:
with torch.backends.cuda.sdp_kernel():
x = F.scaled_dot_product_attention(
q, k, v, dropout_p=self.proj_drop_prob, is_causal=self.is_causal, attn_mask=attn_mask
q, k, v, dropout_p=self.proj_drop_prob if self.training else 0.0,
is_causal=self.is_causal, attn_mask=attn_mask
)
attn = None
else:
Expand Down