makarna/pkg/quant/quantize_test.go

package quant

import (
	"math/rand"
	"testing"
	"time"
)

func BenchmarkQuantizeQ8K(b *testing.B) {
	// 1M floats (typical small tensor)
	data := make([]float32, 1024*1024)
	r := rand.New(rand.NewSource(42))
	for i := range data {
		data[i] = r.Float32()*2 - 1 // [-1, 1]
	}

	b.ResetTimer()
	for i := 0; i < b.N; i++ {
		_ = QuantizeQ8K(data)
	}
}

func BenchmarkQuantizeQ6K(b *testing.B) {
	data := make([]float32, 1024*1024)
	r := rand.New(rand.NewSource(42))
	for i := range data {
		data[i] = r.Float32()*2 - 1
	}

	b.ResetTimer()
	for i := 0; i < b.N; i++ {
		_ = QuantizeQ6K(data)
	}
}

func BenchmarkQuantizeQ4K(b *testing.B) {
	data := make([]float32, 1024*1024)
	r := rand.New(rand.NewSource(42))
	for i := range data {
		data[i] = r.Float32()*2 - 1
	}

	b.ResetTimer()
	for i := 0; i < b.N; i++ {
		_ = QuantizeQ4K(data)
	}
}

func TestQuantizeQ8K_Basic(t *testing.T) {
	// Simple test: 256 elements
	data := make([]float32, 256)
	for i := range data {
		data[i] = float32(i-128) / 128.0 // [-1, ~1]
	}

	start := time.Now()
	result := QuantizeQ8K(data)
	elapsed := time.Since(start)

	// Expect 292 bytes (1 block)
	if len(result) != 292 {
		t.Errorf("Expected 292 bytes, got %d", len(result))
	}

	t.Logf("Q8K: 256 floats -> %d bytes in %v", len(result), elapsed)
}

func TestQuantizeQ6K_Basic(t *testing.T) {
	data := make([]float32, 256)
	for i := range data {
		data[i] = float32(i-128) / 128.0
	}

	start := time.Now()
	result := QuantizeQ6K(data)
	elapsed := time.Since(start)

	// Expect 210 bytes (1 block)
	if len(result) != 210 {
		t.Errorf("Expected 210 bytes, got %d", len(result))
	}

	t.Logf("Q6K: 256 floats -> %d bytes in %v", len(result), elapsed)
}

func TestQuantizeQ4K_Basic(t *testing.T) {
	data := make([]float32, 256)
	for i := range data {
		data[i] = float32(i-128) / 128.0
	}

	start := time.Now()
	result := QuantizeQ4K(data)
	elapsed := time.Since(start)

	// Expect 144 bytes (1 block)
	if len(result) != 144 {
		t.Errorf("Expected 144 bytes, got %d", len(result))
	}

	t.Logf("Q4K: 256 floats -> %d bytes in %v", len(result), elapsed)
}

func TestLargeQuantization(t *testing.T) {
	// Test with 4M elements (typical large weight matrix)
	size := 4 * 1024 * 1024
	data := make([]float32, size)
	r := rand.New(rand.NewSource(42))
	for i := range data {
		data[i] = r.Float32()*2 - 1
	}

	t.Run("Q8K_4M", func(t *testing.T) {
		start := time.Now()
		result := QuantizeQ8K(data)
		elapsed := time.Since(start)
		mbps := float64(size*4) / elapsed.Seconds() / (1024 * 1024)
		t.Logf("Q8K: %d floats (%.1f MB) -> %d bytes in %v (%.1f MB/s)", 
			size, float64(size*4)/(1024*1024), len(result), elapsed, mbps)
	})

	t.Run("Q6K_4M", func(t *testing.T) {
		start := time.Now()
		result := QuantizeQ6K(data)
		elapsed := time.Since(start)
		mbps := float64(size*4) / elapsed.Seconds() / (1024 * 1024)
		t.Logf("Q6K: %d floats (%.1f MB) -> %d bytes in %v (%.1f MB/s)", 
			size, float64(size*4)/(1024*1024), len(result), elapsed, mbps)
	})

	t.Run("Q4K_4M", func(t *testing.T) {
		start := time.Now()
		result := QuantizeQ4K(data)
		elapsed := time.Since(start)
		mbps := float64(size*4) / elapsed.Seconds() / (1024 * 1024)
		t.Logf("Q4K: %d floats (%.1f MB) -> %d bytes in %v (%.1f MB/s)", 
			size, float64(size*4)/(1024*1024), len(result), elapsed, mbps)
	})
}