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

model: support for mistral-small in the ollama runner

Browse Source 
Mistral is a popular research lab making open source models. This updates
the forward pass of llama architecture models to support both llama models
and mistral models by accounting for additional metadata present in mistral
models, and finding the correct dimensions for the output projection.
This commit is contained in:
Bruce MacDonald
2025-03-14 16:56:32 -07:00
committed by Michael Yang
parent 1861fbdeb5
commit 6bd0a983cd
27 changed files with 1116 additions and 350 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
ml/backend.go
View File

@@ -118,6 +118,7 @@ type Tensor interface {
Bytes() []byte
Floats() []float32
Neg(ctx Context) Tensor
Add(ctx Context, t2 Tensor) Tensor
Mul(ctx Context, t2 Tensor) Tensor
Mulmat(ctx Context, t2 Tensor) Tensor
@@ -132,7 +133,10 @@ type Tensor interface {
Conv2D(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor
RoPE(ctx Context, positionIDs, ropeFactors Tensor, dim, ropeType uint32, base, scale float32) Tensor
IM2Col(ctx Context, weight Tensor, s0, s1, p0, p1, d0, d1 int) Tensor
Sin(ctx Context) Tensor
Cos(ctx Context) Tensor
Tanh(ctx Context) Tensor
GELU(ctx Context) Tensor
SILU(ctx Context) Tensor
@@ -147,9 +151,13 @@ type Tensor interface {
Unpad(ctx Context, shape ...int) Tensor
Stack(ctx Context, dim int, s ...Tensor) Tensor
// Repeat repeats the tensor n times along dimension dim
Repeat(ctx Context, dim, n int) Tensor
Concat(ctx Context, t2 Tensor, dim int) Tensor
Rows(ctx Context, t2 Tensor) Tensor
Copy(ctx Context, t2 Tensor) Tensor
Duplicate(ctx Context) Tensor
}
// ScaledDotProductAttention implements a fused attention
@@ -214,7 +222,7 @@ func Dump(ctx Context, t Tensor, opts ...DumpOptions) string {
return strconv.FormatFloat(float64(f), 'f', opts[0].Precision, 32)
})
case DTypeF16, DTypeQ80, DTypeQ40:
f32 := ctx.Empty(DTypeF32, t.Shape()...)
f32 := ctx.Input().Empty(DTypeF32, t.Shape()...)
f32 = t.Copy(ctx, f32)
return dump[[]float32](ctx, f32, opts[0].Items, func(f float32) string {
return strconv.FormatFloat(float64(f), 'f', opts[0].Precision, 32)

56
ml/backend/ggml/ggml.go
View File

@@ -711,6 +711,13 @@ func (t *Tensor) DType() ml.DType {
}
}
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,
@@ -718,6 +725,27 @@ func (t *Tensor) Add(ctx ml.Context, t2 ml.Tensor) ml.Tensor {
}
}
func (t *Tensor) Repeat(ctx ml.Context, dim, n int) ml.Tensor {
if dim < 0 || dim >= C.GGML_MAX_DIMS {
panic("invalid dimension")
}
shape := make([]C.int64_t, C.GGML_MAX_DIMS)
for i := range C.GGML_MAX_DIMS {
if i == dim {
shape[i] = C.int64_t(t.Dim(i) * n)
} else {
shape[i] = C.int64_t(t.Dim(i))
}
}
tmpl := C.ggml_new_tensor(ctx.(*Context).ctx, t.t._type, C.int(len(shape)), unsafe.SliceData(shape))
return &Tensor{
b: t.b,
t: C.ggml_repeat(ctx.(*Context).ctx, t.t, tmpl),
}
}
func (t *Tensor) Stack(ctx ml.Context, dim int, s ...ml.Tensor) ml.Tensor {
if len(s) > 0 {
return t.Concat(ctx, s[0].Stack(ctx, dim, s[1:]...), dim)
@@ -854,6 +882,20 @@ func (t *Tensor) Softmax(ctx ml.Context) ml.Tensor {
}
}
func (t *Tensor) Sin(ctx ml.Context) ml.Tensor {
return &Tensor{
b: t.b,
t: C.ggml_sin(ctx.(*Context).ctx, t.t),
}
}
func (t *Tensor) Cos(ctx ml.Context) ml.Tensor {
return &Tensor{
b: t.b,
t: C.ggml_cos(ctx.(*Context).ctx, t.t),
}
}
func (t *Tensor) Tanh(ctx ml.Context) ml.Tensor {
return &Tensor{
b: t.b,
@@ -942,6 +984,13 @@ func (t *Tensor) RoPE(ctx ml.Context, positionIDs, ropeFactors ml.Tensor, ropeDi
}
}
func (t *Tensor) IM2Col(ctx ml.Context, t2 ml.Tensor, s0, s1, p0, p1, d0, d1 int) ml.Tensor {
return &Tensor{
b: t.b,
t: C.ggml_im2col(ctx.(*Context).ctx, t.t, t2.(*Tensor).t, C.int(s0), C.int(s1), C.int(p0), C.int(p1), C.int(d0), C.int(d1), true, C.GGML_TYPE_F32),
}
}
func (t *Tensor) GELU(ctx ml.Context) ml.Tensor {
return &Tensor{
b: t.b,
@@ -1010,3 +1059,10 @@ func (t *Tensor) ScaledDotProductAttention(ctx ml.Context, key, value, mask ml.T
return kqv.Permute(ctx, 0, 2, 1, 3).Contiguous(ctx)
}
}
func (t *Tensor) Duplicate(ctx ml.Context) ml.Tensor {
return &Tensor{
b: t.b,
t: C.ggml_dup(ctx.(*Context).ctx, t.t),
}
}

7
ml/backend/ggml/ggml/src/ggml-metal/ggml-metal-embed.metal vendored
View File

@@ -3083,6 +3083,13 @@ kernel void kernel_cos(
dst[tpig] = cos(src0[tpig]);
}
kernel void kernel_neg(
device const float * src0,
device float * dst,
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = -src0[tpig];
}
kernel void kernel_sum_rows(
device const float * src0,
device float * dst,

15
ml/backend/ggml/ggml/src/ggml-metal/ggml-metal.m vendored
View File

@@ -423,6 +423,7 @@ enum ggml_metal_kernel_type {
GGML_METAL_KERNEL_TYPE_SQRT,
GGML_METAL_KERNEL_TYPE_SIN,
GGML_METAL_KERNEL_TYPE_COS,
GGML_METAL_KERNEL_TYPE_NEG,
GGML_METAL_KERNEL_TYPE_SUM_ROWS,
GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,
GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,
@@ -1039,6 +1040,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQRT, sqrt, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIN, sin, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_COS, cos, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NEG, neg, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS, sum_rows, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGMAX, argmax, true);
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32, pool_2d_avg_f32, true);
@@ -1202,6 +1204,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
case GGML_UNARY_OP_GELU_QUICK:
case GGML_UNARY_OP_SILU:
case GGML_UNARY_OP_ELU:
case GGML_UNARY_OP_NEG:
return ggml_is_contiguous(op->src[0]);
default:
return false;
@@ -1873,6 +1876,18 @@ static void ggml_metal_encode_node(
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
case GGML_UNARY_OP_NEG:
{
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_NEG].pipeline;
[encoder setComputePipelineState:pipeline];
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
[encoder setBuffer:id_dst offset:offs_dst atIndex:1];
const int64_t n = ggml_nelements(dst);
[encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
} break;
default:
{
GGML_LOG_WARN("%s: node %3d, op = %8s not implemented\n", __func__, idx, ggml_op_name(dst->op));

7
ml/backend/ggml/ggml/src/ggml-metal/ggml-metal.metal vendored
View File

@@ -945,6 +945,13 @@ kernel void kernel_cos(
dst[tpig] = cos(src0[tpig]);
}
kernel void kernel_neg(
device const float * src0,
device float * dst,
uint tpig[[thread_position_in_grid]]) {
dst[tpig] = -src0[tpig];
}
kernel void kernel_sum_rows(
device const float * src0,
device float * dst,