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

deepseekocr

Browse Source
This commit is contained in:
Michael Yang
2025-10-31 19:15:32 -07:00
committed by Michael Yang
parent 8ed1adf3db
commit 92981ae3f2
14 changed files with 975 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
ml/backend.go
View File

@@ -173,6 +173,7 @@ type Tensor interface {
Cos(ctx Context) Tensor
Tanh(ctx Context) Tensor
GELU(ctx Context, up ...Tensor) Tensor
QuickGELU(ctx Context, up ...Tensor) Tensor
SILU(ctx Context, up ...Tensor) Tensor
RELU(ctx Context, up ...Tensor) Tensor
Sigmoid(ctx Context) Tensor
@@ -207,6 +208,8 @@ type Tensor interface {
Stddev(ctx Context) Tensor
Sqr(ctx Context) Tensor
Sqrt(ctx Context) Tensor
Interpolate(ctx Context, dims [4]int, samplingMode SamplingMode) Tensor
}
// ScaledDotProductAttention implements a fused attention
@@ -372,3 +375,10 @@ const (
DTypeI32
DTypeMXFP4
)
type SamplingMode int
const (
SamplingModeNearest SamplingMode = iota
SamplingModeBilinear
)

29
ml/backend/ggml/ggml.go
View File

@@ -314,7 +314,7 @@ func New(modelPath string, params ml.BackendParams) (ml.Backend, error) {
"altup_proj", "altup_unembd_proj",
"per_layer_token_embd", "per_layer_model_proj", "per_layer_proj_norm"):
createTensor(tensor{source: t}, output.bts, blocks)
case strings.HasPrefix(t.Name, "v.") || strings.HasPrefix(t.Name, "mm."):
case strings.HasPrefix(t.Name, "v.") || strings.HasPrefix(t.Name, "mm.") || strings.HasPrefix(t.Name, "s."):
// TODO: assign vision tensors to the gpu if possible
createTensor(tensor{source: t}, output.bts, blocks)
case contains(t.Name, "rope_freqs", "rope_factors_long", "rope_factors_short"):
@@ -1567,6 +1567,16 @@ func (t *Tensor) GELU(ctx ml.Context, t2 ...ml.Tensor) ml.Tensor {
}
}
func (t *Tensor) QuickGELU(ctx ml.Context, t2 ...ml.Tensor) ml.Tensor {
var tt *C.struct_ggml_tensor
if len(t2) > 0 {
tt = C.ggml_geglu_quick_split(ctx.(*Context).ctx, t.t, t2[0].(*Tensor).t)
} else {
tt = C.ggml_gelu_quick_inplace(ctx.(*Context).ctx, t.t)
}
return &Tensor{b: t.b, t: tt}
}
func (t *Tensor) SILU(ctx ml.Context, t2 ...ml.Tensor) ml.Tensor {
if len(t2) > 0 {
return &Tensor{
@@ -1724,6 +1734,23 @@ func (t *Tensor) Sqrt(ctx ml.Context) ml.Tensor {
}
}
func (t *Tensor) Interpolate(ctx ml.Context, dims [4]int, samplingMode ml.SamplingMode) ml.Tensor {
var mode C.uint32_t
switch samplingMode {
case ml.SamplingModeNearest:
mode = C.GGML_SCALE_MODE_NEAREST
case ml.SamplingModeBilinear:
mode = C.GGML_SCALE_MODE_BILINEAR
default:
panic("unsupported interpolate mode")
}
return &Tensor{
b: t.b,
t: C.ggml_interpolate(ctx.(*Context).ctx, t.t, C.int64_t(dims[0]), C.int64_t(dims[1]), C.int64_t(dims[2]), C.int64_t(dims[3]), mode),
}
}
// 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.