Skip to content

fix: prevent F.linear from saving dequantized weights in MatMul4Bit/MatMul8bitLt to save ~13GB VRAM and prevent OOM errors #1935

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
butterwecksolutions wants to merge 1 commit into
base: main
Choose a base branch
from
+6 −2
Open

fix: prevent F.linear from saving dequantized weights in MatMul4Bit/MatMul8bitLt to save ~13GB VRAM and prevent OOM errors #1935

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
8 changes: 6 additions & 2 deletions bitsandbytes/autograd/_functions.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -262,7 +262,8 @@ def forward(ctx, A, B, out=None, bias=None, state=MatmulLtState):
B = state.CB

CB = state.CB.data.to(A.dtype).mul_(state.SCB.unsqueeze(1).mul(1.0 / 127.0))
output = torch.nn.functional.linear(A, CB, bias)
with torch.no_grad():
output = torch.nn.functional.linear(A, CB, bias)
ctx.state = state
ctx.dtype_A = A.dtype
ctx.grad_shape = A.shape
Expand Down Expand Up @@ -317,7 +318,10 @@ def forward(ctx, A, B, out=None, bias=None, quant_state: Optional[F.QuantState]

# 1. Dequantize
# 2. MatmulnN
output = torch.nn.functional.linear(A, F.dequantize_4bit(B, quant_state).to(A.dtype).t(), bias)
# no_grad prevents F.linear from saving the dequantized weight
# (MatMul4Bit.backward re-dequantizes from NF4 anyway)
with torch.no_grad():
output = torch.nn.functional.linear(A, F.dequantize_4bit(B, quant_state).to(A.dtype).t(), bias)
if out is not None:
out.copy_(output)
output = out
Expand Down