jeans/ollama
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

sample: update tests and add test logits

Browse Source
This commit is contained in:
ParthSareen 2025-03-12 00:46:41 -04:00
parent 310b235626
commit 7fa6ea0da7
3 changed files with 122 additions and 23 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
sample/testdata/logits.bin vendored Normal file
View File

File diff suppressed because one or more lines are too long

4
sample/transforms.go
View File

@ -151,8 +151,8 @@ func partialSortLogits(ts []token, n int) []token {
if ts[i].value >= pivotValue {
ts[storeIndex], ts[i] = ts[i], ts[storeIndex]
storeIndex++
}
}
}
// Move pivot to its final position
ts[right], ts[storeIndex] = ts[storeIndex], ts[right]
@ -180,7 +180,7 @@ func partialSortLogits(ts []token, n int) []token {
})
return ts[:n]
}
}
// sortLogits uses partialSortLogits to efficiently sort tokens
// It sorts approximately sqrt(len(tokens)) elements which balances

140
sample/transforms_test.go
View File

@ -1,39 +1,44 @@
package sample
import (
"encoding/binary"
"errors"
"math"
"math/rand/v2"
"os"
"path/filepath"
"runtime"
"testing"
)
// Helper to convert float64 slice to logit slice
func toTokens(values []float64) []token {
// Helper to convert float32 slice to logit slice
func toTokens(values []float32) []token {
tokens := make([]token, len(values))
for i, v := range values {
tokens[i] = token{
id: int32(i),
value: float32(v),
value: v,
}
}
return tokens
}
// Helper to compare logit slices
func compareLogits(t *testing.T, name string, want []float64, got []token) {
func compareLogits(t *testing.T, name string, want []float32, got []token) {
t.Helper()
if len(want) != len(got) {
t.Errorf("%s: length mismatch: want %d, got %d", name, len(want), len(got))
return
}
for i := range want {
if math.Abs(float64(got[i].value)-want[i]) > 1e-6 {
if math.Abs(float64(got[i].value-want[i])) > 1e-6 {
t.Errorf("%s: index %d: want %f, got %f", name, i, want[i], got[i].value)
}
}
}
func TestTemperatureAndSoftmax(t *testing.T) {
input := []float64{1, 4, -2, 0}
input := []float32{1, 4, -2, 0}
got := temperature(toTokens(input), 0.5)
// Check probabilities sum to 1
@ -41,7 +46,7 @@ func TestTemperatureAndSoftmax(t *testing.T) {
for _, token := range got {
sum += token.value
}
if math.Abs(float64(sum)-1.0) > 1e-6 {
if math.Abs(float64(sum-1.0)) > 1e-6 {
t.Errorf("probabilities don't sum to 1: got %f", sum)
}
@ -51,30 +56,31 @@ func TestTemperatureAndSoftmax(t *testing.T) {
for _, token := range got {
sum += token.value
}
if math.Abs(float64(sum)-1.0) > 1e-6 {
if math.Abs(float64(sum-1.0)) > 1e-6 {
t.Errorf("probabilities don't sum to 1: got %f", sum)
}
}
func TestTopK(t *testing.T) {
input := []float64{-3, -2, -1, 0, 1, 2, 4}
input := []float32{0.026986899, 0.043722924, 0.036774673, 0.27755088, 0.0046718004, 0.08582123, 0.20409796, 0.00412893, 0.15720603, 0.045046154, 0.0030491839, 0.01681367}
// Test k=3
got := topK(toTokens(input), 3)
if len(got) != 3 {
t.Errorf("topK(3): wrong length: want 3, got %d", len(got))
got := topK(toTokens(input), 5)
if len(got) != 5 {
t.Errorf("topK(5): wrong length: want 5, got %d", len(got))
}
// Should keep highest 3 values: 4, 2, 1
want := []float64{4, 2, 1}
// Should keep highest 3 values in descending order
want := []float32{0.27755088, 0.20409796, 0.15720603, 0.08582123, 0.045046154}
compareLogits(t, "topK(3)", want, got)
// Test k > len
got = topK(toTokens(input), 10)
compareLogits(t, "topK(10)", input, got)
got = topK(toTokens(input), 20)
if len(got) != len(input) {
t.Errorf("topK(20): wrong length: want %d, got %d", len(input), len(got))
}
}
func TestTopP(t *testing.T) {
input := []float64{-3, -2, -1, 0, 1, 2, 4}
input := []float32{-3, -2, -1, 0, 1, 2, 4}
tokens := toTokens(input)
// First apply temperature and softmax to get probabilities
@ -92,7 +98,7 @@ func TestTopP(t *testing.T) {
}
func TestMinP(t *testing.T) {
input := []float64{-3, -2, -1, 0, 1, 2, 4, 3}
input := []float32{-3, -2, -1, 0, 1, 2, 4, 3}
tokens := toTokens(input)
// First apply temperature and softmax
@ -108,7 +114,7 @@ func TestMinP(t *testing.T) {
}
func TestSortLogits(t *testing.T) {
input := []float64{3, 1, 4, 2, -1, 0, -2}
input := []float32{0.026986899, 0.043722924, 0.036774673, 0.27755088, 0.0046718004, 0.08582123, 0.20409796, 0.00412893, 0.15720603, 0.045046154, 0.0030491839, 0.01681367}
tokens := toTokens(input)
sortLogits(tokens)
@ -120,10 +126,102 @@ func TestSortLogits(t *testing.T) {
}
}
want := []float64{4, 3, 2, 1, 0, -1, -2}
want := []float32{0.27755088, 0.20409796, 0.15720603, 0.08582123, 0.045046154, 0.043722924, 0.036774673, 0.026986899, 0.01681367, 0.0046718004, 0.00412893, 0.0030491839}
compareLogits(t, "sortLogits", want, tokens)
}
// TestSortLogitsWithRealData tests sorting behavior using real model logit distributions
func TestSortLogitsWithRealData(t *testing.T) {
// This will be populated from testdata/logits.bin
// Format: 32-bit float array in binary format
logits, err := loadTestLogits(t)
if err != nil {
t.Skipf("Skipping real logit test: %v", err)
return
}
tokens := toTokens(logits)
sortLogits(tokens)
// Calculate n for verification
n := int(math.Sqrt(float64(len(tokens)))) + 1
if n > 1000 {
n = 1000
} else if n < 100 {
n = 100
}
t.Logf("Testing with %d tokens, partial sorting top %d", len(tokens), n)
// Only verify the top n elements are sorted (which is what we guarantee)
// This is much faster than checking the entire array
topN := tokens[:n]
for i := 1; i < len(topN); i++ {
if topN[i].value > topN[i-1].value {
t.Fatalf("top %d tokens not properly sorted at index %d: %.15f > %.15f",
n, i, topN[i].value, topN[i-1].value)
}
}
// Verify we didn't lose any high value tokens by checking that
// all tokens after position n are <= the nth token
// Do this in chunks to avoid timeouts on large arrays
nthValue := tokens[n-1].value
const chunkSize = 1000
for start := n; start < len(tokens); start += chunkSize {
end := min(start+chunkSize, len(tokens))
for i := start; i < end; i++ {
if tokens[i].value > nthValue {
t.Fatalf("found higher value token after position %d: tokens[%d].value = %.15f > %.15f",
n, i, tokens[i].value, nthValue)
}
}
}
}
// loadTestLogits loads logit test data from testdata/logits.bin
func loadTestLogits(t *testing.T) ([]float32, error) {
t.Helper()
_, currFile, _, ok := runtime.Caller(0)
if !ok {
return nil, errors.New("could not determine test file path")
}
testDataPath := filepath.Join(filepath.Dir(currFile), "testdata", "logits.bin")
file, err := os.Open(testDataPath)
if err != nil {
return nil, err
}
defer file.Close()
stat, err := file.Stat()
if err != nil {
return nil, err
}
numFloats := stat.Size() / 4 // each float32 is 4 bytes
if numFloats*4 != stat.Size() {
return nil, errors.New("logits.bin has invalid size: not a multiple of 4 bytes")
}
logits := make([]float32, numFloats)
for i := range logits {
var val uint32
if err := binary.Read(file, binary.LittleEndian, &val); err != nil {
return nil, err
}
logits[i] = math.Float32frombits(val)
}
if len(logits) == 0 {
return nil, errors.New("logits.bin is empty")
}
return logits, nil
}
func BenchmarkTransforms(b *testing.B) {
// Generate random logits
tokens := make([]token, 1<<16)