[Qwen][Bugfix] Fixes sigmoid activation in torch impl of RMSNormGated. (#40245)

Signed-off-by: Tao He <linzhu.ht@alibaba-inc.com>
This commit is contained in:
Tao He
2026-04-20 04:28:19 +00:00
committed by GitHub
parent 67ed01c353
commit 8936118134
2 changed files with 13 additions and 2 deletions
+5 -2
View File
@@ -478,9 +478,12 @@ class RMSNormGated(CustomOp):
weight = self.weight.float()
z = z.float() if z is not None else None
assert self.activation in ["silu", "sigmoid", "swish"]
act_fn = F.sigmoid if self.activation == "sigmoid" else F.silu
# Apply gating before normalization if needed
if z is not None and not self.norm_before_gate:
x = x * F.silu(z)
x = x * act_fn(z)
# RMS Normalization
if self.group_size is None:
@@ -499,7 +502,7 @@ class RMSNormGated(CustomOp):
# Apply gating after normalization if needed
if z is not None and self.norm_before_gate:
out = out * F.silu(z)
out = out * act_fn(z)
return out.to(orig_dtype)
@@ -357,11 +357,19 @@ class GatedDeltaNetAttention(PluggableLayer, MambaBase):
set_weight_attrs(self.A_log, {"weight_loader": sharded_weight_loader(0)})
set_weight_attrs(self.dt_bias, {"weight_loader": sharded_weight_loader(0)})
output_gate_type = getattr(config, "output_gate_type", "silu")
if output_gate_type == "swish":
output_gate_type = "silu"
assert output_gate_type in ["silu", "swish", "sigmoid"], (
f"unsupported {output_gate_type=}"
)
self.norm = RMSNormGated(
self.head_v_dim,
eps=self.layer_norm_epsilon,
group_size=None,
norm_before_gate=True,
activation=output_gate_type,
device=current_platform.current_device(),
)