Skip to content

Add the missing b_hn bias to nnx.GRUCell #5595

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
vineethsaivs wants to merge 3 commits into
base: main
Choose a base branch
from
+13 −0
Open

Add the missing b_hn bias to nnx.GRUCell #5595

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
776f9dd
Add the missing b_hn bias to nnx.GRUCell
vineethsaivs Sep 21, 2026
50c4db8
Drop the GRUCell equivalence test
vineethsaivs Sep 28, 2026
998b0bf
Broadcast and promote the GRUCell hn bias like Linear
vineethsaivs Oct 6, 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
13 changes: 13 additions & 0 deletions flax/nnx/nn/recurrent.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -693,6 +693,15 @@ def __init__(
kernel_metadata=recurrent_kernel_metadata,
)

# `b_hn` from the docstring. Only the reset-gated term carries a hidden
# bias, which the fused `dense_h` above cannot express: a bias there would
# reach r and z as well. `flax.linen.GRUCell` builds it as a separate `hn`
# bias for the same reason.
self.hn_bias = nnx.Param(
bias_init(rngs.params(), (hidden_features,), self.param_dtype),
**bias_metadata,
)

if carry_init:
warnings.warn(
"carry_init is provided in __init__. "
Expand Down Expand Up @@ -730,6 +739,10 @@ def __call__(self, carry: Array, inputs: Array) -> tuple[Array, Array]: # type:
z = self.gate_fn(xi_z + hh_z)

# Compute n with an additional linear transformation on h
hh_n, hn_bias = self.promote_dtype(
(hh_n, self.hn_bias[...]), dtype=self.dtype
)
hh_n += jnp.reshape(hn_bias, (1,) * (hh_n.ndim - 1) + (-1,))
n = self.activation_fn(xi_n + r * hh_n)

# Update hidden state
Expand Down
Loading