Skip to content

fix(autograd/MatMul4Bit): save the packed weight via save_for_backward (#2034) #2035

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
Anai-Guo wants to merge 1 commit into
base: main
Choose a base branch
from
+11 −4
Open

fix(autograd/MatMul4Bit): save the packed weight via save_for_backward (#2034) #2035

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
15 changes: 11 additions & 4 deletions bitsandbytes/autograd/_functions.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -354,10 +354,16 @@ def forward(ctx, A, B, out=None, bias=None, quant_state: Optional[F.QuantState]
ctx.state = quant_state
ctx.dtype_A, ctx.dtype_B, ctx.dtype_bias = A.dtype, B.dtype, None if bias is None else bias.dtype

# Save the packed weight through save_for_backward rather than as a plain
# ctx attribute. torch.utils.checkpoint discards and recomputes *saved*
# tensors, so a caller that re-materialises the weight (weight offloading,
# layer streaming) into a recycled buffer under checkpointing gets a fresh
# copy in backward. A raw ctx attribute is invisible to that mechanism and
# would keep pointing at the original buffer -- silently wrong gradients.
if any(ctx.needs_input_grad[:2]):
ctx.tensors = (None, B)
ctx.save_for_backward(B)
else:
ctx.tensors = (None, None)
ctx.save_for_backward()

return output

Expand All @@ -368,7 +374,8 @@ def backward(ctx, grad_output):
return torch.zeros_like(ctx.A), torch.zeros_like(ctx.B), None, bias_grad, None

req_gradA, _, _, req_gradBias, _ = ctx.needs_input_grad
_, B = ctx.tensors
saved = ctx.saved_tensors
B = saved[0] if saved else None

grad_A, grad_B, grad_bias = None, None, None

Expand All @@ -379,7 +386,7 @@ def backward(ctx, grad_output):
# not supported by PyTorch. TODO: create work-around
# if req_gradB: grad_B = torch.matmul(grad_output.t(), A)
if req_gradA:
# B in ctx.tensors is already in canonical [(N*K+1)//2, 1] form (normalized in forward).
# B from saved_tensors is already in canonical [(N*K+1)//2, 1] form (normalized in forward).
# dequantize returns [N, K]; matmul(grad_output[M,N], [N,K]) = grad_A[M,K].
grad_A = torch.matmul(grad_output, F.dequantize_4bit(B, ctx.state).to(grad_output.dtype))

Expand Down