Skip to content

Fix training crash in addmm_act by falling back to unfused ops when grad is enabled #613

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
Rudr4Khunt1 wants to merge 1 commit into
base: main
Choose a base branch
from
+8 −1
Open

Fix training crash in addmm_act by falling back to unfused ops when grad is enabled #613

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
9 changes: 8 additions & 1 deletion sam3/perflib/fused.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,14 @@

def addmm_act(activation, linear, mat1):
if torch.is_grad_enabled():
raise ValueError("Expected grad to be disabled.")
# Training path: the fused kernel below is inference-only (detached
# weights, bf16 cast), so fall back to standard autograd-friendly ops.
y = linear(mat1)
if activation in [torch.nn.functional.relu, torch.nn.ReLU]:
return torch.nn.functional.relu(y)
if activation in [torch.nn.functional.gelu, torch.nn.GELU]:
return torch.nn.functional.gelu(y)
raise ValueError(f"Unexpected activation {activation}")
self = linear.bias.detach()
mat2 = linear.weight.detach()
self = self.to(torch.bfloat16)
Expand Down