highperfocused/ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2025-08-25 09:41:25 +02:00
Code Issues Packages Projects Releases Wiki Activity

llama4

Browse Source
This commit is contained in:
Michael Yang
2025-04-03 15:18:29 -07:00
committed by Michael Yang
parent 54055a6dae
commit f0c66e6dea
13 changed files with 833 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
ml/backend.go
View File

@@ -133,6 +133,7 @@ type Tensor interface {
Mul(ctx Context, t2 Tensor) Tensor
Mulmat(ctx Context, t2 Tensor) Tensor
MulmatFullPrec(ctx Context, t2 Tensor) Tensor
MulmatID(ctx Context, t2, ids Tensor) Tensor
Softmax(ctx Context) Tensor
LayerNorm(ctx Context, weight, bias Tensor, eps float32) Tensor
@@ -150,6 +151,7 @@ type Tensor interface {
Tanh(ctx Context) Tensor
GELU(ctx Context) Tensor
SILU(ctx Context) Tensor
Sigmoid(ctx Context) Tensor
Reshape(ctx Context, shape ...int) Tensor
View(ctx Context, offset int, shape ...int) Tensor
@@ -168,6 +170,8 @@ type Tensor interface {
Rows(ctx Context, t2 Tensor) Tensor
Copy(ctx Context, t2 Tensor) Tensor
Duplicate(ctx Context) Tensor
TopK(ctx Context, k int) Tensor
}
// ScaledDotProductAttention implements a fused attention

39
ml/backend/ggml/ggml.go
View File

@@ -884,17 +884,32 @@ func (t *Tensor) MulmatFullPrec(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
}
}
func (t *Tensor) MulmatID(ctx ml.Context, t2, ids ml.Tensor) ml.Tensor {
return &Tensor{
b: t.b,
t: C.ggml_mul_mat_id(ctx.(*Context).ctx, t.t, t2.(*Tensor).t, ids.(*Tensor).t),
}
}
func (t *Tensor) LayerNorm(ctx ml.Context, w, b ml.Tensor, eps float32) ml.Tensor {
tt := (&Tensor{b: t.b, t: C.ggml_norm(ctx.(*Context).ctx, t.t, C.float(eps))}).Mul(ctx, w)
if b != nil {
tt = tt.Add(ctx, b)
tt := C.ggml_norm(ctx.(*Context).ctx, t.t, C.float(eps))
if w != nil {
tt = C.ggml_mul(ctx.(*Context).ctx, tt, w.(*Tensor).t)
if b != nil {
tt = C.ggml_add(ctx.(*Context).ctx, tt, b.(*Tensor).t)
}
}
return tt
return &Tensor{b: t.b, t: tt}
}
func (t *Tensor) RMSNorm(ctx ml.Context, w ml.Tensor, eps float32) ml.Tensor {
return (&Tensor{b: t.b, t: C.ggml_rms_norm(ctx.(*Context).ctx, t.t, C.float(eps))}).Mul(ctx, w)
tt := C.ggml_rms_norm(ctx.(*Context).ctx, t.t, C.float(eps))
if w != nil {
tt = C.ggml_mul(ctx.(*Context).ctx, tt, w.(*Tensor).t)
}
return &Tensor{b: t.b, t: tt}
}
func (t *Tensor) Pad(ctx ml.Context, shape ...int) ml.Tensor {
@@ -995,6 +1010,13 @@ func (t *Tensor) Tanh(ctx ml.Context) ml.Tensor {
}
}
func (t *Tensor) Sigmoid(ctx ml.Context) ml.Tensor {
return &Tensor{
b: t.b,
t: C.ggml_sigmoid_inplace(ctx.(*Context).ctx, t.t),
}
}
func (t *Tensor) Unpad(ctx ml.Context, shape ...int) ml.Tensor {
if len(shape) != 4 {
panic("expected 4 dimensions")
@@ -1158,3 +1180,10 @@ func (t *Tensor) Duplicate(ctx ml.Context) ml.Tensor {
t: C.ggml_dup(ctx.(*Context).ctx, t.t),
}
}
func (t *Tensor) TopK(ctx ml.Context, k int) ml.Tensor {
return &Tensor{
b: t.b,
t: C.ggml_top_k(ctx.(*Context).ctx, t.t, C.int(k)),
}
}