Skip to content

[Fix] Forward the configured epsilon in the bit-linear norm #1303

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

Merged
yzhangcs merged 5 commits into from
Oct 11, 2026
Merged

[Fix] Forward the configured epsilon in the bit-linear norm #1303

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
532cae0
[Fix] Forward the configured epsilon in the bit-linear norm
aniruddhaadak80 Sep 28, 2026
f6a909f
[Fix] Update BitLinear epsilon forwarding for current modules
yzhangcs Oct 11, 2026
dfe4777
[Test] Keep focused BitLinear epsilon forwarding checks
yzhangcs Oct 11, 2026
c7a35a0
[Test] Remove added BitLinear epsilon tests
yzhangcs Oct 11, 2026
d0d9bbd
[Refactor] Use keyword arguments in BitLinear helpers
yzhangcs Oct 11, 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
10 changes: 6 additions & 4 deletions fla/modules/fused_bitlinear.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,7 +24,7 @@
def activation_quant(x: torch.Tensor) -> torch.Tensor:
"""Quantize activations to 8 bits per token and return dequantized values."""
scale = 127.0 / x.abs().max(dim=-1, keepdim=True).values.clamp_(min=1e-5)
y = (x * scale).round().clamp_(-128, 127) / scale
y = (x * scale).round().clamp_(min=-128, max=127) / scale
return y


Expand Down Expand Up @@ -94,6 +94,7 @@ def bit_linear(
norm_bias=norm_bias,
linear_weight=weight,
linear_bias=bias,
eps=eps,
is_rms_norm=True,
)

Expand Down Expand Up @@ -126,11 +127,11 @@ def __repr__(self) -> str:

def forward(self, x: torch.Tensor) -> torch.Tensor:
w = self.weight
x_norm = self.norm(x)
x_norm = self.norm(x=x)

# straight-through gradients bypass quantization
x_quant = x_norm + (activation_quant(x_norm) - x_norm).detach()
w_quant = w + (weight_quant(w) - w).detach()
x_quant = x_norm + (activation_quant(x=x_norm) - x_norm).detach()
w_quant = w + (weight_quant(w=w) - w).detach()
y = F.linear(input=x_quant, weight=w_quant)

return y
Expand All @@ -149,5 +150,6 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:
norm_bias=self.norm.bias,
linear_weight=self.weight,
linear_bias=self.bias,
eps=self.norm.eps,
is_rms_norm=True,
)
Loading