mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-21 22:33:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

use split activations when possible (#12293)

Browse Source 
* use ggml_*_split activations when possible

* forward qkv
This commit is contained in:
Michael Yang
2025-09-16 09:51:19 -07:00
committed by GitHub
parent c253433d68
commit ad95d5b30b
16 changed files with 59 additions and 50 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
ml/backend.go
View File

@@ -430,12 +430,13 @@ type Tensor interface {
Sin(ctx Context) Tensor
Cos(ctx Context) Tensor
Tanh(ctx Context) Tensor
GELU(ctx Context) Tensor
QuickGELU(ctx Context) Tensor
SILU(ctx Context) Tensor
RELU(ctx Context) Tensor
GELU(ctx Context, up ...Tensor) Tensor
SILU(ctx Context, up ...Tensor) Tensor
RELU(ctx Context, up ...Tensor) Tensor
Sigmoid(ctx Context) Tensor
SwiGLU(ctx Context, up Tensor, alpha, limit float32) 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
Reshape(ctx Context, shape ...int) Tensor
View(ctx Context, offset int, shape ...int) Tensor

31
ml/backend/ggml/ggml.go
View File

@@ -1431,35 +1431,46 @@ func (t *Tensor) IM2Col(ctx ml.Context, t2 ml.Tensor, s0, s1, p0, p1, d0, d1 int
}
}
func (t *Tensor) GELU(ctx ml.Context) ml.Tensor {
func (t *Tensor) GELU(ctx ml.Context, t2 ...ml.Tensor) ml.Tensor {
if len(t2) > 0 {
return &Tensor{
b: t.b,
t: C.ggml_geglu_split(ctx.(*Context).ctx, t.t, t2[0].(*Tensor).t),
}
}
return &Tensor{
b: t.b,
t: C.ggml_gelu_inplace(ctx.(*Context).ctx, t.t),
}
}
func (t *Tensor) QuickGELU(ctx ml.Context) ml.Tensor {
return &Tensor{
b: t.b,
t: C.ggml_gelu_quick_inplace(ctx.(*Context).ctx, t.t),
func (t *Tensor) SILU(ctx ml.Context, t2 ...ml.Tensor) ml.Tensor {
if len(t2) > 0 {
return &Tensor{
b: t.b,
t: C.ggml_swiglu_split(ctx.(*Context).ctx, t.t, t2[0].(*Tensor).t),
}
}
}
func (t *Tensor) SILU(ctx ml.Context) ml.Tensor {
return &Tensor{
b: t.b,
t: C.ggml_silu_inplace(ctx.(*Context).ctx, t.t),
}
}
func (t *Tensor) RELU(ctx ml.Context) ml.Tensor {
func (t *Tensor) RELU(ctx ml.Context, t2 ...ml.Tensor) ml.Tensor {
if len(t2) > 0 {
return &Tensor{
b: t.b,
t: C.ggml_reglu_split(ctx.(*Context).ctx, t.t, t2[0].(*Tensor).t),
}
}
return &Tensor{
b: t.b,
t: C.ggml_relu_inplace(ctx.(*Context).ctx, t.t),
}
}
func (t *Tensor) SwiGLU(ctx ml.Context, up ml.Tensor, alpha, limit float32) ml.Tensor {
func (t *Tensor) SILUAlphaLimit(ctx ml.Context, up ml.Tensor, alpha, limit float32) ml.Tensor {
return &Tensor{
b: t.b,
t: C.ggml_swiglu_oai(ctx.(*Context).ctx, t.t, up.(*Tensor).t, C.float(alpha), C.float(limit)),

2
ml/nn/attention.go
View File

@@ -26,6 +26,7 @@ func Attention(ctx ml.Context, query, key, value ml.Tensor, scale float64, cache
}
func AttentionWithSinks(ctx ml.Context, query, key, value, sinks ml.Tensor, scale float64, cache kvcache.Cache) ml.Tensor {
ctx.Forward(query)
if key != nil && value != nil {
if query.Dim(0) != key.Dim(0) {
panic(fmt.Errorf("d_k in attention operation does not match between query(%v) and key(%v)", query.Dim(0), key.Dim(0)))
@@ -39,6 +40,7 @@ func AttentionWithSinks(ctx ml.Context, query, key, value, sinks ml.Tensor, scal
panic(fmt.Errorf("seq_len_k in attention operation does not match between key(%v) and value(%v)", key.Dim(2), value.Dim(2)))
}
ctx.Forward(key, value)
if cache != nil {
cache.Put(ctx, key, value)
}