mirror of
https://github.com/ollama/ollama.git
synced 2026-11-22 00:01:45 +00:00
* mlxrunner: add Inkling 2 Mini * mlx: fuse Inkling 2 Mini residual mixing on Metal Fuse RMS normalization and query scoring in one kernel. Apply the softmax weights in a second kernel to avoid storing per-query products. * create: expose Inkling 2 Mini context length in model metadata Copy model_max_length to max_position_embeddings during import so ollama show reports the checkpoint's context length. Reject configs that specify conflicting limits. * mlxrunner: reduce Metal dispatches in Inkling 2 Mini decode Reuse the prefill mask shader to compute relative position bias in one dispatch for single-query attention. On M3 Ultra with Metal 3, compute the loop controller's gain and forget projections in one matrix multiply.
48 lines
1.5 KiB
Go
48 lines
1.5 KiB
Go
package mlx
|
|
|
|
import (
|
|
"testing"
|
|
|
|
"github.com/ollama/ollama/mlx/mlxthread/mlxthreadtest"
|
|
)
|
|
|
|
func TestMetalTensorOpsFallback(t *testing.T) {
|
|
withMLXThread(t, func(t *mlxthreadtest.T) {
|
|
if hasMetalTensorOps() {
|
|
t.Skip("requires a backend without Metal tensor ops")
|
|
}
|
|
// These shapes qualify on a Metal 4 driver. An older driver must
|
|
// return to the native model path before constructing the shader.
|
|
q := Zeros(DTypeFloat32, 1, 1, 16, 128)
|
|
k := Zeros(DTypeBFloat16, 1, 1, 16, 128)
|
|
table := Zeros(DTypeBFloat16, 1, 1, 16, 16)
|
|
sink := Zeros(DTypeBFloat16, 1, 1, 16, 1)
|
|
if RelativePositionAttentionBF16Probabilities(q, k, k, table, sink, 0, 16, 0.125) != nil {
|
|
t.Fatal("selected MPP attention")
|
|
}
|
|
x := Zeros(DTypeBFloat16, 16, 1, 1, 32)
|
|
w := Zeros(DTypeUint32, 1, 64, 4)
|
|
s := Zeros(DTypeUint8, 1, 64, 2)
|
|
ids := Zeros(DTypeInt32, 16, 1)
|
|
if PrefersCompactNVFP4QMM() || GatherNVFP4QMM(x, w, s, ids) != nil {
|
|
t.Fatal("selected MPP grouped matmul")
|
|
}
|
|
})
|
|
}
|
|
|
|
func TestM5MatrixDevice(t *testing.T) {
|
|
for _, tc := range []struct {
|
|
name, architecture string
|
|
want bool
|
|
}{
|
|
{"Apple M5 Max", "applegpu_g17s", true},
|
|
{"Apple M5 Ultra", "applegpu_g17d", true},
|
|
{"Apple M5 Max", "applegpu_g17d", false},
|
|
{"Apple M5 Pro", "applegpu_g17s", true},
|
|
{"Apple M3 Ultra", "applegpu_g15d", false},
|
|
} {
|
|
if got := isM5MatrixDevice(tc.name, tc.architecture); got != tc.want {
|
|
t.Errorf("device %s/%s returns %t, want %t", tc.name, tc.architecture, got, tc.want)
|
|
}
|
|
}
|
|
}
|