makarna/pkg/engine/placement_test.go

package engine

import (
	"testing"

	"makarna/pkg/backend/cuda"
	"makarna/pkg/loader"
	"makarna/pkg/tensor"
)

func TestPlanLayerDevicesFallsBackToCPUWhenNoCUDA(t *testing.T) {
	cudaDeviceCountFn = func() (int, error) { return 0, nil }
	defer func() { cudaDeviceCountFn = cuda.DeviceCount }()

	cfg := &loader.ModelConfig{
		Params: map[string]any{
			"num_hidden_layers":       float64(4),
			"num_attention_heads":     float64(8),
			"num_key_value_heads":     float64(8),
			"head_dim":                float64(128),
			"max_position_embeddings": float64(2048),
		},
	}
	md := &loader.ModelData{}
	placements := PlanLayerDevices(md, cfg, 0.9, nil)
	if len(placements) != 4 {
		t.Fatalf("expected 4 placements, got %d", len(placements))
	}
	for i, p := range placements {
		if p.Type != tensor.CPU {
			t.Fatalf("layer %d expected CPU, got %v", i, p.Type)
		}
	}
}

func TestPlanLayerDevicesFitsInBudget(t *testing.T) {
	cudaDeviceCountFn = func() (int, error) { return 1, nil }
	cudaMemoryInfoDeviceFn = func(gpu int) (uint64, uint64, error) {
		return 8 * 1024 * 1024 * 1024, 8 * 1024 * 1024 * 1024, nil
	}
	defer func() { cudaDeviceCountFn = cuda.DeviceCount }()
	defer func() { cudaMemoryInfoDeviceFn = cuda.MemoryInfoDevice }()

	cfg := &loader.ModelConfig{
		Params: map[string]any{
			"num_hidden_layers":       float64(2),
			"num_attention_heads":     float64(8),
			"num_key_value_heads":     float64(8),
			"head_dim":                float64(128),
			"max_position_embeddings": float64(128),
		},
	}
	md := &loader.ModelData{
		Metadata: loader.Metadata{
			Tensors: map[string]loader.TensorEntry{
				"w": {Shape: []uint64{1024, 1024}, DType: loader.F32},
			},
			ModelConfig: *cfg,
		},
	}
	placements := PlanLayerDevices(md, cfg, 0.5, nil)
	for i, p := range placements {
		if p.Type != tensor.CUDA {
			t.Fatalf("layer %d expected CUDA, got %v", i, p.Type)
		}
	}
}