Skip to content

Cache ctx.saved_tensors in Mamba3 backward to fix FSDP activation checkpointing #909

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
anasiri wants to merge 1 commit into
base: main
Choose a base branch
from
+3 −2
Open

Cache ctx.saved_tensors in Mamba3 backward to fix FSDP activation checkpointing #909

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
5 changes: 3 additions & 2 deletions mamba_ssm/ops/triton/mamba3/mamba3_siso_combined.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -165,7 +165,8 @@ def backward(
except Exception:
pass

if len(ctx.saved_tensors) == 0:
_saved = ctx.saved_tensors
if len(_saved) == 0:
raise RuntimeError(
"Backward called but forward ran without gradient tracking. "
"Ensure inputs require grad or run under torch.enable_grad()."
Expand All @@ -176,7 +177,7 @@ def backward(
(Q, K, V, ADT, DT, Trap, Q_bias, K_bias, Angles, Angles_Cumsum,
D_save, Z_save, Input_SSM_State_save, Input_K_State_save, Input_V_State_save,
Out, Out_v, SSM_States, DA_CS, DA_CS_SUM, Q_rot, K_scaled, QK_dot, Scale, Gamma,
Final_SSM_State_save, cu_seqlens_save) = ctx.saved_tensors
Final_SSM_State_save, cu_seqlens_save) = _saved

D = D_save if ctx.has_D else None
Z = Z_save if ctx.has_Z else None
Expand Down