github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-03-09 07:16:38 -05:00
Code Issues Packages Projects Releases Wiki Activity

qwen3next: avoid inplace sigmoid for shared gate (#14077)

Browse Source
This commit is contained in:
Jeffrey Morgan
2026-02-04 15:50:02 -08:00
committed by GitHub
parent c323161f24
commit d25535c3f3
3 changed files with 9 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
ml/backend.go
View File

@@ -175,6 +175,7 @@ type Tensor interface {
SILU(ctx Context, up ...Tensor) Tensor
RELU(ctx Context, up ...Tensor) Tensor
Sigmoid(ctx Context) Tensor
SigmoidOut(ctx Context) Tensor
// AlphaLimitSILU is a variant of SILU that clamps the input to the range [-limit, limit]
SILUAlphaLimit(ctx Context, up Tensor, alpha, limit float32) Tensor

7
ml/backend/ggml/ggml.go
View File

@@ -1468,6 +1468,13 @@ func (t *Tensor) Sigmoid(ctx ml.Context) ml.Tensor {
}
}
func (t *Tensor) SigmoidOut(ctx ml.Context) ml.Tensor {
return &Tensor{
b: t.b,
t: C.ggml_sigmoid(ctx.(*Context).ctx, t.t),
}
}
func (t *Tensor) View(ctx ml.Context, offset int, shape ...int) ml.Tensor {
switch len(shape) {
case 1:

2
model/models/qwen3next/model.go
View File

@@ -135,7 +135,7 @@ func (mlp *sparse) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *Options
// Apply shared expert gating
if mlp.SharedGateInp != nil {
sharedGateVal := mlp.SharedGateInp.Forward(ctx, hiddenStates2D)
sharedGateVal = sharedGateVal.Sigmoid(ctx)
sharedGateVal = sharedGateVal.SigmoidOut(ctx)
// Broadcast gate to match dimensions
sharedGateVal = sharedGateVal.Repeat(ctx, 0, sharedOut.Dim(0))
sharedOut = sharedOut.Mul(ctx, sharedGateVal)