mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-21 14:26:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

chore: update models to use slice/chunk/chunksections (#12934)

Browse Source 
* use slice/chunks

* bert

* llama4

* gemma3n

* gptoss

* mistral3

* qwen3vl

* qwen25vl

* deepseek2

* remove unused ops
This commit is contained in:
Michael Yang
2025-11-13 15:20:12 -08:00
committed by GitHub
parent c114987523
commit 333203d871
13 changed files with 59 additions and 135 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
ml/backend.go
View File

@@ -146,7 +146,6 @@ type Tensor interface {
FromFloats([]float32)
FromInts([]int32)
Neg(ctx Context) Tensor
Add(ctx Context, t2 Tensor) Tensor
Sub(ctx Context, t2 Tensor) Tensor
Mul(ctx Context, t2 Tensor) Tensor
@@ -185,7 +184,6 @@ type Tensor interface {
View(ctx Context, offset int, shape ...int) Tensor
Permute(ctx Context, shape ...int) Tensor
Contiguous(ctx Context, shape ...int) Tensor
Set(ctx Context, t2 Tensor, offset int, strides ...int) Tensor
Pad(ctx Context, shape ...int) Tensor
@@ -209,7 +207,6 @@ type Tensor interface {
Stddev(ctx Context) Tensor
Sqr(ctx Context) Tensor
Sqrt(ctx Context) Tensor
Clamp(ctx Context, min, max float32) Tensor
}
// ScaledDotProductAttention implements a fused attention

28
ml/backend/ggml/ggml.go
View File

@@ -1137,13 +1137,6 @@ func (t *Tensor) Cast(ctx ml.Context, dtype ml.DType) ml.Tensor {
}
}
func (t *Tensor) Neg(ctx ml.Context) ml.Tensor {
return &Tensor{
b: t.b,
t: C.ggml_neg(ctx.(*Context).ctx, t.t),
}
}
func (t *Tensor) Add(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
return &Tensor{
b: t.b,
@@ -1632,20 +1625,6 @@ func (t *Tensor) AvgPool2D(ctx ml.Context, k, s int, p float32) ml.Tensor {
}
}
func (t *Tensor) Set(ctx ml.Context, t2 ml.Tensor, offset int, strides ...int) ml.Tensor {
var tt *C.struct_ggml_tensor
switch len(strides) {
case 0:
tt = C.ggml_set_1d(ctx.(*Context).ctx, t.t, t2.(*Tensor).t, C.size_t(offset))
case 1:
tt = C.ggml_set_2d(ctx.(*Context).ctx, t.t, t2.(*Tensor).t, C.size_t(offset), C.size_t(strides[0]))
default:
panic("unsupported number of dimensions")
}
return &Tensor{b: t.b, t: tt}
}
func (t *Tensor) ScaledDotProductAttention(ctx ml.Context, key, value, mask, sinks ml.Tensor, scale float64) ml.Tensor {
var kqMask *C.struct_ggml_tensor
if mask != nil {
@@ -1732,13 +1711,6 @@ func (t *Tensor) Sqrt(ctx ml.Context) ml.Tensor {
}
}
func (t *Tensor) Clamp(ctx ml.Context, min, max float32) ml.Tensor {
return &Tensor{
b: t.b,
t: C.ggml_clamp(ctx.(*Context).ctx, t.t, C.float(min), C.float(max)),
}
}
// Slice returns a view of the tensor sliced along dim from low to high in step steps.
// Slice panics if the dimension is invalid or the slice parameters are out of range.
// If dim=0 and step>1, the tensor is a copy rather than a view to ensure proper shape.

5
ml/nn/pooling/pooling.go
View File

@@ -32,10 +32,9 @@ func (t Type) Forward(ctx ml.Context, hiddenStates ml.Tensor) ml.Tensor {
hiddenStates = hiddenStates.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx).Mean(ctx)
return hiddenStates.Permute(ctx, 1, 0, 2, 3).Contiguous(ctx)
case TypeCLS:
return hiddenStates.View(ctx, 0, hiddenStates.Dim(0))
return hiddenStates.Slice(ctx, 1, 0, 1, 1)
case TypeLast:
hiddenStates = hiddenStates.View(ctx, (hiddenStates.Dim(1)-1)*hiddenStates.Stride(1), hiddenStates.Dim(0))
return hiddenStates
return hiddenStates.Slice(ctx, 1, hiddenStates.Dim(1)-1, hiddenStates.Dim(1), 1)
default:
panic("unknown pooling type")
}