sample: do all sorting in topK

README.md

@@ -54,10 +54,6 @@ Here are some example models that can be downloaded:
 
 | Model              | Parameters | Size  | Download                         |
 | ------------------ | ---------- | ----- | -------------------------------- |
-| Gemma 3            | 1B         | 815MB | `ollama run gemma3:1b`           |
-| Gemma 3            | 4B         | 3.3GB | `ollama run gemma3`              |
-| Gemma 3            | 12B        | 8.1GB | `ollama run gemma3:12b`          |
-| Gemma 3            | 27B        | 17GB  | `ollama run gemma3:27b`          |
 | QwQ                | 32B        | 20GB  | `ollama run qwq`                 |
 | DeepSeek-R1        | 7B         | 4.7GB | `ollama run deepseek-r1`         |
 | DeepSeek-R1        | 671B       | 404GB | `ollama run deepseek-r1:671b`    |
@@ -70,6 +66,9 @@ Here are some example models that can be downloaded:
 | Llama 3.1          | 405B       | 231GB | `ollama run llama3.1:405b`       |
 | Phi 4              | 14B        | 9.1GB | `ollama run phi4`                |
 | Phi 4 Mini         | 3.8B       | 2.5GB | `ollama run phi4-mini`           |
+| Gemma 2            | 2B         | 1.6GB | `ollama run gemma2:2b`           |
+| Gemma 2            | 9B         | 5.5GB | `ollama run gemma2`              |
+| Gemma 2            | 27B        | 16GB  | `ollama run gemma2:27b`          |
 | Mistral            | 7B         | 4.1GB | `ollama run mistral`             |
 | Moondream 2        | 1.4B       | 829MB | `ollama run moondream`           |
 | Neural Chat        | 7B         | 4.1GB | `ollama run neural-chat`         |

cmd/interactive.go

@@ -195,10 +195,6 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error {
 			opts.Messages = []api.Message{}
 			fmt.Printf("Loading model '%s'\n", opts.Model)
 			if err := loadOrUnloadModel(cmd, &opts); err != nil {
-				if strings.Contains(err.Error(), "not found") {
-					fmt.Printf("error: %v\n", err)
-					continue
-				}
 				return err
 			}
 			continue

model/models/gemma3/model_text.go

@@ -15,6 +15,7 @@ type TextOptions struct {
 	attnKeyLen, attnValLen           int
 	eps, ropeScale                   float32
 	ropeLocalBase, ropeGlobalBase    float32
+	finalLogitSoftcap                float32
 	largeModelScaling                bool
 }
 
@@ -56,15 +57,16 @@ func newTextModel(c ml.Config) *TextModel {
 		),
 		Layers: make([]TextLayer, numBlocks),
 		TextOptions: &TextOptions{
-			hiddenSize:     int(c.Uint("embedding_length")),
-			numHeads:       int(c.Uint("attention.head_count")),
-			numKVHeads:     int(c.Uint("attention.head_count_kv")),
-			attnKeyLen:     int(c.Uint("attention.key_length", 256)),
-			attnValLen:     int(c.Uint("attention.value_length", 256)),
-			eps:            c.Float("attention.layer_norm_rms_epsilon", 1e-06),
-			ropeLocalBase:  c.Float("rope.local.freq_base", 10000.0),
-			ropeGlobalBase: c.Float("rope.global.freq_base", 1000000.0),
-			ropeScale:      c.Float("rope.freq_scale", 1.0),
+			hiddenSize:        int(c.Uint("embedding_length")),
+			numHeads:          int(c.Uint("attention.head_count")),
+			numKVHeads:        int(c.Uint("attention.head_count_kv")),
+			attnKeyLen:        int(c.Uint("attention.key_length", 256)),
+			attnValLen:        int(c.Uint("attention.value_length", 256)),
+			eps:               c.Float("attention.layer_norm_rms_epsilon", 1e-06),
+			ropeLocalBase:     c.Float("rope.local.freq_base", 10000.0),
+			ropeGlobalBase:    c.Float("rope.global.freq_base", 1000000.0),
+			ropeScale:         c.Float("rope.freq_scale", 1.0),
+			finalLogitSoftcap: c.Float("final_logit_softcapping", 30.0),
 		},
 	}
 
@@ -243,5 +245,10 @@ func (m *TextModel) Forward(ctx ml.Context, inputs, positions, outputs ml.Tensor
 	}
 
 	hiddenState = m.OutputNorm.Forward(ctx, hiddenState, m.eps)
-	return m.Output.Forward(ctx, hiddenState)
+	hiddenState = m.Output.Forward(ctx, hiddenState)
+
+	// final logit softcap
+	hiddenState = hiddenState.Scale(ctx, 1.0/float64(m.TextOptions.finalLogitSoftcap))
+	hiddenState = hiddenState.Tanh(ctx)
+	return hiddenState.Scale(ctx, float64(m.TextOptions.finalLogitSoftcap))
 }

readline/readline.go

@@ -116,9 +116,19 @@ func (i *Instance) Readline() (string, error) {
 
 			switch r {
 			case KeyUp:
-				i.historyPrev(buf, &currentLineBuf)
+				if i.History.Pos > 0 {
+					if i.History.Pos == i.History.Size() {
+						currentLineBuf = []rune(buf.String())
+					}
+					buf.Replace([]rune(i.History.Prev()))
+				}
 			case KeyDown:
-				i.historyNext(buf, &currentLineBuf)
+				if i.History.Pos < i.History.Size() {
+					buf.Replace([]rune(i.History.Next()))
+					if i.History.Pos == i.History.Size() {
+						buf.Replace(currentLineBuf)
+					}
+				}
 			case KeyLeft:
 				buf.MoveLeft()
 			case KeyRight:
@@ -175,10 +185,6 @@ func (i *Instance) Readline() (string, error) {
 			esc = true
 		case CharInterrupt:
 			return "", ErrInterrupt
-		case CharPrev:
-			i.historyPrev(buf, &currentLineBuf)
-		case CharNext:
-			i.historyNext(buf, &currentLineBuf)
 		case CharLineStart:
 			buf.MoveToStart()
 		case CharLineEnd:
@@ -240,24 +246,6 @@ func (i *Instance) HistoryDisable() {
 	i.History.Enabled = false
 }
 
-func (i *Instance) historyPrev(buf *Buffer, currentLineBuf *[]rune) {
-	if i.History.Pos > 0 {
-		if i.History.Pos == i.History.Size() {
-			*currentLineBuf = []rune(buf.String())
-		}
-		buf.Replace([]rune(i.History.Prev()))
-	}
-}
-
-func (i *Instance) historyNext(buf *Buffer, currentLineBuf *[]rune) {
-	if i.History.Pos < i.History.Size() {
-		buf.Replace([]rune(i.History.Next()))
-		if i.History.Pos == i.History.Size() {
-			buf.Replace(*currentLineBuf)
-		}
-	}
-}
-
 func NewTerminal() (*Terminal, error) {
 	fd := os.Stdin.Fd()
 	termios, err := SetRawMode(fd)

sample/samplers.go

@@ -1,10 +1,11 @@
 package sample
 
 import (
+	"errors"
 	"math"
-	"math/rand"
+	"math/rand/v2"
+	"slices"
 	"sync"
-	"time"
 
 	"github.com/ollama/ollama/llama"
 )
@@ -86,53 +87,53 @@ func (s *Sampler) sample(tokens []token) (token, error) {
 	// topK also sorts the tokens in descending order of logits
 	tokens = topK(tokens, s.topK)
 
+	// token logit values are updated to probabilities
+	tokens = temperature(tokens, s.temperature)
+
 	tokens = topP(tokens, s.topP)
 	tokens = minP(tokens, s.minP)
 
-	// token logit values are updated to probabilities
-	temperature(tokens, s.temperature)
-	softmax(tokens)
-	return tokens[dist(tokens, s.rng.Int63())], nil
+	// TODO: this should fall back to greedy sampling
+	// or topP, topK values etc should be such that
+	// there are always tokens to sample from
+	if len(tokens) == 0 {
+		return token{}, errors.New("no tokens to sample from")
+	}
 
-	// // TODO: this should fall back to greedy sampling
-	// // or topP, topK values etc should be such that
-	// // there are always tokens to sample from
-	// if len(tokens) == 0 {
-	// 	return token{}, errors.New("no tokens to sample from")
-	// }
+	var r float32
+	if s.rng != nil {
+		r = s.rng.Float32()
+	} else {
+		r = rand.Float32()
+	}
 
-	// var r float32
-	// if s.rng != nil {
-	// 	r = s.rng.Float32()
-	// } else {
-	// 	r = rand.Float32()
-	// }
+	// Calculate cumulative sum of probabilities
+	var sum float32
+	for i := range tokens {
+		sum += tokens[i].value
+		tokens[i].value = sum
+	}
+	r *= tokens[len(tokens)-1].value
 
-	// // Calculate cumulative sum of probabilities
-	// var sum float32
-	// for i := range tokens {
-	// 	sum += tokens[i].value
-	// 	tokens[i].value = sum
-	// }
-	// r *= tokens[len(tokens)-1].value
+	idx, _ := slices.BinarySearchFunc(tokens, r, func(token token, target float32) int {
+		if token.value < target {
+			return -1
+		}
+		return 1
+	})
 
-	// idx, _ := slices.BinarySearchFunc(tokens, r, func(token token, target float32) int {
-	// 	if token.value < target {
-	// 		return -1
-	// 	}
-	// 	return 1
-	// })
-
-	// return tokens[idx], nil
+	return tokens[idx], nil
 }
 
 // TODO(parthsareen): update sampler interface to use json unmarshal https://github.com/ollama/ollama/issues/9278
 func NewSampler(temperature float32, topK int, topP float32, minP float32, seed int, grammar *Grammar) Sampler {
 	var rng *rand.Rand
 	if seed != -1 {
-		rng = rand.New(rand.NewSource(int64(seed)))
-	} else {
-		rng = rand.New(rand.NewSource(time.Now().UnixNano()))
+		// PCG requires two parameters: sequence and stream
+		// Use original seed for sequence
+		sequence := uint64(seed)
+		// Use golden ratio hash to generate statistically independent seeds
+		rng = rand.New(rand.NewPCG(sequence, sequence^0x9E3779B9))
 	}
 	if temperature < 0.0 {
 		temperature = 0.0

sample/transforms.go

@@ -3,7 +3,6 @@ package sample
 import (
 	"container/heap"
 	"math"
-	"math/rand"
 	"slices"
 )
 
@@ -26,6 +25,32 @@ func (h *tokenHeap) Pop() any {
 	return x
 }
 
+// temperature applies scaling and softmax to the logits
+func temperature(ts []token, temp float32) []token {
+	// Find max logit for numerical stability
+	maxLogit := float32(math.Inf(-1))
+	for _, t := range ts {
+		if t.value > maxLogit {
+			maxLogit = t.value
+		}
+	}
+
+	// Apply temperature and compute exp(x - max)
+	temp = max(temp, 1e-7)
+	var sum float32
+	for i, v := range ts {
+		ts[i].value = float32(math.Exp(float64((v.value - maxLogit) / temp)))
+		sum += ts[i].value
+	}
+
+	// Normalize
+	for i := range ts {
+		ts[i].value /= sum
+	}
+
+	return ts
+}
+
 // topK limits the number of tokens considered to the k highest logits
 func topK(ts []token, k int) []token {
 	if k >= len(ts) || k <= 0 {
@@ -109,59 +134,3 @@ func minP(ts []token, p float32) []token {
 	ts = validTokens
 	return ts
 }
-
-func temperature(ts []token, temp float32) {
-	for i := range ts {
-		ts[i].value /= temp
-	}
-}
-
-func softmax(ts []token) {
-	if len(ts) == 0 {
-		return
-	}
-
-	// Find max logit for numerical stability
-	maxLogit := ts[0].value
-	for _, t := range ts {
-		if t.value > maxLogit {
-			maxLogit = t.value
-		}
-	}
-
-	// Compute exp(logit - maxLogit) and sum them
-	var sumExp float32
-	for i, t := range ts {
-		expVal := float32(math.Exp(float64(t.value - maxLogit)))
-		ts[i].value = expVal
-		sumExp += expVal
-	}
-
-	// Normalize probabilities
-	for i := range ts {
-		ts[i].value /= sumExp
-	}
-}
-
-// applyDist selects a token based on probabilities and seed
-func dist(ts []token, seed int64) int {
-	rng := rand.New(rand.NewSource(seed))
-
-	cdf := make([]float32, len(ts))
-	var cumSum float32
-	for i, t := range ts {
-		cumSum += t.value
-		cdf[i] = cumSum
-	}
-
-	r := rng.Float32() * cumSum
-
-	// Select token based on CDF
-	for i, probSum := range cdf {
-		if r < probSum {
-			return i
-		}
-	}
-
-	return len(ts) - 1
-}

sample/transforms_test.go

@@ -1,13 +1,8 @@
 package sample
 
 import (
-	"encoding/binary"
-	"errors"
 	"math"
 	"math/rand/v2"
-	"os"
-	"path/filepath"
-	"runtime"
 	"testing"
 )
 
@@ -148,98 +143,6 @@ func TestSortLogits(t *testing.T) {
 	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)