makarna/pkg/engine/scheduler_test.go

package engine

import (
	"context"
	"testing"

	"makarna/pkg/backend/cpu"
	"makarna/pkg/backend/cpu/nn"
	"makarna/pkg/graph"
	"makarna/pkg/kvcache"
	"makarna/pkg/model"
	"makarna/pkg/tensor"
)

func TestSchedulerOffsetsPositionsAndContext(t *testing.T) {
	cfg := &model.Config{
		VocabSize:  5,
		NumLayers:  1,
		NumKVHeads: 1,
		HeadDim:    2,
	}
	mock := &mockModel{cfg: cfg}
	engine := &Engine{model: mock}

	plan := graph.BuildPlan(graph.RequestSpec{
		ID:           "req-1",
		MaxContext:   8,
		BlockSize:    4,
		NumLayers:    1,
		UseAttention: true,
		LayerDevices: []tensor.DevicePlacement{{Type: tensor.CUDA, GPU: 0}},
	})
	pool, err := kvcache.NewBlockPool(kvcache.BlockPoolConfig{
		NumLayers:  1,
		NumKVHeads: 1,
		HeadDim:    2,
		BlockSize:  4,
		NumBlocks:  8,
		Device:     tensor.CPU,
		GPU:        0,
	})
	if err != nil {
		t.Fatalf("NewBlockPool: %v", err)
	}
	cache := kvcache.NewPagedKVCache(pool, kvcache.PagedCacheConfig{
		NumLayers:  1,
		NumKVHeads: 1,
		HeadDim:    2,
		BlockSize:  4,
		MaxSeqLen:  8,
		Device:     tensor.CPU,
		GPU:        0,
	}, "sched")
	defer cache.Free()

	sched := engine.NewScheduler(plan, cache)

	if _, err := sched.Prefill([]int{1, 2, 3}); err != nil {
		t.Fatalf("prefill failed: %v", err)
	}
	if _, err := sched.Decode([]int{4, 5}); err != nil {
		t.Fatalf("decode failed: %v", err)
	}

	if cache.SeqLen() != 5 {
		t.Fatalf("expected seqLen 5, got %d", cache.SeqLen())
	}
	if remaining := sched.RemainingContext(); remaining != 3 {
		t.Fatalf("expected remaining context 3, got %d", remaining)
	}

	if len(mock.positions) != 2 {
		t.Fatalf("expected 2 forward calls, got %d", len(mock.positions))
	}
	expectSlice(t, []int{0, 1, 2}, mock.positions[0])
	expectSlice(t, []int{3, 4}, mock.positions[1])

	if plan.Layers[0].Device.Type != tensor.CUDA || plan.Layers[0].Device.GPU != 0 {
		t.Fatalf("expected layer device cuda:0, got %+v", plan.Layers[0].Device)
	}
}

type mockModel struct {
	cfg       *model.Config
	positions [][]int
}

func (m *mockModel) Forward(ctx context.Context, input tensor.Tensor, positions tensor.Tensor, kv model.KVCache) (tensor.Tensor, error) {
	seq := input.Shape()[0]
	pos := nn.ParsePositions(positions, seq)
	m.positions = append(m.positions, pos)
	return cpu.NewTensor(tensor.Shape{seq, m.cfg.VocabSize}, nil), nil
}

func (m *mockModel) Config() *model.Config { return m.cfg }
func (m *mockModel) Close() error          { return nil }
func (m *mockModel) SetTensor(string, tensor.Tensor) error {
	return nil
}

func expectSlice(t *testing.T, want, got []int) {
	if len(want) != len(got) {
		t.Fatalf("length mismatch want=%v got=%v", want, got)
	}
	for i := range want {
		if want[i] != got[i] {
			t.Fatalf("slice mismatch at %d: want=%d got=%d", i, want[i], got[i])
		}
	}
}