Skip to content

Get up and running with Kimi, GLM, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models.

7.7 KB
Raw Blame
package mlxrunner

import (
	"context"
	"errors"
	"math"
	"slices"
	"testing"

	"github.com/ollama/ollama/mlx"
	"github.com/ollama/ollama/mlx/mlxtest"
	"github.com/ollama/ollama/mlxrunner/batch"
	"github.com/ollama/ollama/mlxrunner/model"
	"github.com/ollama/ollama/mlxrunner/model/clef"
	"github.com/ollama/ollama/mlxrunner/model/qwen3_5"
	"github.com/ollama/ollama/mlxrunner/nn"
)

// A small real backbone exercises both recurrent state and attention history.
func prefillTestModel() *clef.Model {
	seed := 0
	weight := func(shape ...int) *mlx.Array {
		n := 1
		for _, d := range shape {
			n *= d
		}
		values := make([]float32, n)
		for i := range values {
			seed++
			values[i] = float32(math.Sin(float64(seed)*0.17)) * 0.1
		}
		return mlx.FromValues(values, shape...)
	}
	linear := func(out, in int) *nn.Linear { return nn.NewLinear(weight(out, in), nil) }
	norm := func(n int) *nn.RMSNorm { return nn.NewRMSNorm(mlx.FromValues(slices.Repeat([]float32{1}, n), n), 1e-6) }
	cfg := &qwen3_5.Config{
		HiddenSize: 8, NumAttentionHeads: 1, NumKeyValueHeads: 1, HeadDim: 8,
		RMSNormEps: 1e-6, RopeDim: 8, RopeTheta: 10000, Scale: 1 / float32(math.Sqrt(8)),
		LinearNumKeyHeads: 1, LinearNumValueHeads: 1, LinearKeyHeadDim: 32,
		LinearValueHeadDim: 32, LinearConvKernelDim: 4,
	}
	m := &qwen3_5.Model{Config: cfg, EmbedTokens: nn.NewEmbedding(weight(16, 8)), Norm: norm(8)}
	for _, recurrent := range []bool{true, false} {
		layer := &qwen3_5.Layer{
			InputNorm: norm(8), PostAttentionNorm: norm(8), IsLinear: recurrent,
			MLP: &qwen3_5.DenseMLP{GateProj: linear(16, 8), UpProj: linear(16, 8), DownProj: linear(8, 16)},
		}
		if recurrent {
			layer.Linear = &qwen3_5.GatedDeltaNet{
				InProjQKVZ: linear(128, 8), InProjBA: linear(2, 8), OutProj: linear(8, 32),
				Conv1D:     nn.NewConv1d(weight(96, 4, 1), nil, 1, 0, 1, 96),
				NormWeight: mlx.FromValues(slices.Repeat([]float32{1}, 32), 32), DtBias: weight(1), AExp: mlx.FromValues([]float32{1}, 1),
			}
		} else {
			layer.FullAttn = &qwen3_5.FullAttention{
				QProj: linear(16, 8), KProj: linear(8, 8), VProj: linear(8, 8), OProj: linear(8, 8),
				QNorm: norm(8), KNorm: norm(8),
			}
		}
		m.Layers = append(m.Layers, layer)
	}
	return &clef.Model{Model: m}
}

func TestPrefillPreservesHiddenStates(t *testing.T) {
	mlxtest.Run(t, func(t *mlxtest.T) {
		mlx.Scoped(func() {
			m := prefillTestModel()
			for _, n := range []int{1, 2048, 2049, 4097} {
				mlx.Scoped(func() {
					tokens := make([]int32, n)
					for i := range tokens {
						tokens[i] = int32(i % 16)
					}
					ids := mlx.FromValues(tokens, 1, n)
					want := mlx.ScopedEval(func() []*mlx.Array {
						h, _ := m.Model.Forward(&batch.Batch{InputIDs: ids, SeqOffsets: []int32{0}, SeqQueryLens: []int32{int32(n)}}, nil)
						return []*mlx.Array{h}
					})[0]
					r := &Runner{Model: m}
					defer func() { r.scoreCache.close() }()
					got, _, err := r.scoreHiddenRow(context.Background(), &model.PreparedRequest{Tokens: tokens}, nil)
					if err != nil {
						t.Fatal(err)
					}
					if !slices.Equal(got.Dims(), want.Dims()) {
						t.Fatalf("%d tokens: hidden shape %v, want %v", n, got.Dims(), want.Dims())
					}
					// Chunking changes the forward shapes and floating-point arithmetic.
					values := want.Floats()
					for i, v := range got.Floats() {
						w := values[i]
						if math.IsNaN(float64(v)) || math.Abs(float64(v-w)) > 1e-3*(1+math.Abs(float64(w))) {
							t.Fatalf("%d tokens: hidden[%d] = %g, want %g", n, i, v, w)
						}
					}
				})
			}
		})
	})
}

// Cancel at an actual forward boundary without a timer or a production hook.
type cancellingEmbedding struct {
	nn.EmbeddingLayer
	cancel      context.CancelFunc
	cancelAfter int
	calls       int
}

func (e *cancellingEmbedding) Forward(ids *mlx.Array) *mlx.Array {
	e.calls++
	if e.calls == e.cancelAfter {
		e.cancel()
	}
	return e.EmbeddingLayer.Forward(ids)
}

func TestPrefillCancellation(t *testing.T) {
	mlxtest.Run(t, func(t *mlxtest.T) {
		mlx.Scoped(func() {
			m := prefillTestModel()
			embedding := &cancellingEmbedding{EmbeddingLayer: m.EmbedTokens}
			m.EmbedTokens = embedding
			prepared := &model.PreparedRequest{Tokens: make([]int32, 4097)}
			r := &Runner{Model: m}
			defer func() { r.scoreCache.close() }()
			want, _, err := r.scoreHiddenRow(context.Background(), prepared, nil)
			if err != nil {
				t.Fatal(err)
			}
			mlx.Eval(want)
			for _, after := range []int{0, 1, 2, 3, 1, 2} {
				r.scoreCache.close()
				r.scoreCache, r.scoreHidden = nil, nil
				ctx, cancel := context.WithCancel(context.Background())
				embedding.calls, embedding.cancelAfter, embedding.cancel = 0, after, cancel
				if after == 0 {
					cancel()
				}
				got, _, err := r.scoreHiddenRow(ctx, prepared, nil)
				cancel()
				if !errors.Is(err, context.Canceled) || got != nil {
					t.Fatalf("cancel after %d chunks: hidden=%v, err=%v", after, got, err)
				}
				if embedding.calls != after {
					t.Fatalf("cancel after %d chunks ran %d forwards", after, embedding.calls)
				}
				embedding.cancelAfter = 0
				mlx.Scoped(func() {
					got, _, err := r.scoreHiddenRow(context.Background(), prepared, nil)
					if err != nil {
						t.Fatalf("request after cancellation: %v", err)
					}
					if !slices.Equal(got.Floats(), want.Floats()) {
						t.Fatal("cancelled request changed the next request's hidden states")
					}
				})
			}
		})
	})
}

// Reuse must preserve every earlier hidden state read by decision heads, not
// just the final token used by ordinary vocabulary scoring.
func TestScoreHiddenPrefixReuse(t *testing.T) {
	mlxtest.Run(t, func(t *mlxtest.T) {
		m := prefillTestModel()
		r := &Runner{Model: m}
		defer func() { r.scoreCache.close() }()
		original := slices.Repeat([]int32{1, 2, 3, 4}, 520)
		branch := slices.Clone(original)
		branch[1024] = 5
		rows := [][]int32{original, branch, original[:1024], branch, original}
		for _, tokens := range rows {
			cold := &Runner{Model: m}
			want, _, err := cold.scoreHiddenRow(context.Background(), &model.PreparedRequest{Tokens: tokens}, nil)
			if err != nil {
				t.Fatal(err)
			}
			wantValues := want.Floats()
			cold.scoreCache.close()
			for repeat := range 2 {
				got, cached, err := r.scoreHiddenRow(context.Background(), &model.PreparedRequest{Tokens: tokens}, nil)
				if err != nil {
					t.Fatal(err)
				}
				if repeat == 1 && cached != len(tokens) {
					t.Fatalf("exact hit reused %d/%d tokens", cached, len(tokens))
				}
				for i, v := range got.Floats() {
					if math.IsNaN(float64(v)) || math.Abs(float64(v-wantValues[i])) > 1e-3*(1+math.Abs(float64(wantValues[i]))) {
						t.Fatalf("cached hidden[%d]=%g, cold=%g", i, v, wantValues[i])
					}
				}
			}
		}
	})
}

func TestScoreHiddenSharedStateSnapshot(t *testing.T) {
	mlxtest.Run(t, func(t *mlxtest.T) {
		m := prefillTestModel()
		first := slices.Repeat([]int32{1, 2, 3, 4}, 520)
		second := slices.Clone(first)
		const shared = 1024
		second[shared] = 5

		cold := &Runner{Model: m}
		want, _, err := cold.scoreHiddenRow(context.Background(), &model.PreparedRequest{Tokens: second}, nil)
		if err != nil {
			t.Fatal(err)
		}
		wantValues := want.Floats()
		cold.scoreCache.close()

		r := &Runner{Model: m}
		defer func() { r.scoreCache.close() }()
		if _, _, err := r.scoreHiddenRow(context.Background(), &model.PreparedRequest{Tokens: first, ScorePrefix: shared}, nil); err != nil {
			t.Fatal(err)
		}
		got, cached, err := r.scoreHiddenRow(context.Background(), &model.PreparedRequest{Tokens: second, ScorePrefix: shared}, nil)
		if err != nil {
			t.Fatal(err)
		}
		if cached != shared {
			t.Fatalf("shared state reused %d/%d tokens", cached, shared)
		}
		for i, v := range got.Floats() {
			if math.IsNaN(float64(v)) || math.Abs(float64(v-wantValues[i])) > 1e-3*(1+math.Abs(float64(wantValues[i]))) {
				t.Fatalf("cached hidden[%d]=%g, cold=%g", i, v, wantValues[i])
			}
		}
	})
}