move tokenizers to separate package

llm/server.go

@@ -34,6 +34,7 @@ import (
 	"github.com/ollama/ollama/logutil"
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/tokenizers"
 )
 
 type filteredEnv []string
@@ -115,7 +116,7 @@ type llamaServer struct {
 type ollamaServer struct {
 	llmServer
 
-	textProcessor model.TextProcessor // textProcessor handles text encoding/decoding
+	tokenizer tokenizers.Tokenizer // textProcessor handles text encoding/decoding
 }
 
 // LoadModel will load a model from disk. The model must be in the GGML format.
@@ -141,11 +142,11 @@ func LoadModel(model string, maxArraySize int) (*ggml.GGML, error) {
 // NewLlamaServer will run a server for the given GPUs
 func NewLlamaServer(systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, modelPath string, f *ggml.GGML, adapters, projectors []string, opts api.Options, numParallel int) (LlamaServer, error) {
 	var llamaModel *llama.Model
-	var textProcessor model.TextProcessor
+	var tokenizer tokenizers.Tokenizer
 	var err error
 	if envconfig.NewEngine() || f.KV().OllamaEngineRequired() {
 		if len(projectors) == 0 {
-			textProcessor, err = model.NewTextProcessor(modelPath)
+			tokenizer, err = model.NewTextProcessor(modelPath)
 		} else {
 			err = errors.New("split vision models aren't supported")
 		}
@@ -154,7 +155,7 @@ func NewLlamaServer(systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, modelPath st
 			slog.Debug("model not yet supported by Ollama engine, switching to compatibility mode", "model", modelPath, "error", err)
 		}
 	}
-	if textProcessor == nil {
+	if tokenizer == nil {
 		llamaModel, err = llama.LoadModelFromFile(modelPath, llama.ModelParams{VocabOnly: true})
 		if err != nil {
 			return nil, err
@@ -210,7 +211,7 @@ func NewLlamaServer(systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, modelPath st
 
 	kvct := strings.ToLower(envconfig.KvCacheType())
 
-	if textProcessor == nil {
+	if tokenizer == nil {
 		flashAttention := ml.FlashAttentionAuto
 		if faUserSet {
 			if fa {
@@ -260,7 +261,7 @@ func NewLlamaServer(systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, modelPath st
 	gpuLibs := ml.LibraryPaths(gpus)
 	status := NewStatusWriter(os.Stderr)
 	cmd, port, err := StartRunner(
-		textProcessor != nil,
+		tokenizer != nil,
 		modelPath,
 		gpuLibs,
 		status,
@@ -309,8 +310,8 @@ func NewLlamaServer(systemInfo ml.SystemInfo, gpus []ml.DeviceInfo, modelPath st
 		}
 	}()
 
-	if textProcessor != nil {
-		return &ollamaServer{llmServer: s, textProcessor: textProcessor}, nil
+	if tokenizer != nil {
+		return &ollamaServer{llmServer: s, tokenizer: tokenizer}, nil
 	} else {
 		return &llamaServer{llmServer: s, ggml: f}, nil
 	}
@@ -1772,7 +1773,7 @@ func (s *llamaServer) Tokenize(ctx context.Context, content string) ([]int, erro
 }
 
 func (s *ollamaServer) Tokenize(ctx context.Context, content string) ([]int, error) {
-	tokens, err := s.textProcessor.Encode(content, false)
+	tokens, err := s.tokenizer.Encode(content, false)
 	if err != nil {
 		return nil, err
 	}
@@ -1807,7 +1808,7 @@ func (s *ollamaServer) Detokenize(ctx context.Context, tokens []int) (string, er
 		toks[i] = int32(t)
 	}
 
-	content, err := s.textProcessor.Decode(toks)
+	content, err := s.tokenizer.Decode(toks)
 	if err != nil {
 		return "", err
 	}

model/model.go

@@ -23,6 +23,7 @@ import (
 	_ "github.com/ollama/ollama/ml/backend"
 	"github.com/ollama/ollama/ml/nn/pooling"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizers"
 )
 
 var (
@@ -119,7 +120,7 @@ func New(modelPath string, params ml.BackendParams) (Model, error) {
 	return m, nil
 }
 
-func NewTextProcessor(s string) (TextProcessor, error) {
+func NewTextProcessor(s string) (tokenizers.Tokenizer, error) {
 	r, err := os.Open(s)
 	if err != nil {
 		return nil, err
@@ -136,7 +137,7 @@ func NewTextProcessor(s string) (TextProcessor, error) {
 		return nil, err
 	}
 
-	tp, ok := m.(TextProcessor)
+	tp, ok := m.(tokenizers.Tokenizer)
 	if !ok {
 		return nil, ErrUnsupportedTokenizer
 	}

model/models/bert/embed.go

@@ -10,11 +10,12 @@ import (
 	"github.com/ollama/ollama/ml/nn/pooling"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizers"
 )
 
 type Model struct {
 	model.Base
-	model.TextProcessor
+	tokenizers.Tokenizer
 
 	TokenEmbedding     *nn.Embedding `gguf:"token_embd"`
 	TypeEmbedding      *nn.Embedding `gguf:"token_types"`
@@ -129,7 +130,7 @@ func (o Options) headDim() int {
 }
 
 func New(c fs.Config) (model.Model, error) {
-	vocab := &model.Vocabulary{
+	vocab := &tokenizers.Vocabulary{
 		Values: c.Strings("tokenizer.ggml.tokens"),
 		Scores: c.Floats("tokenizer.ggml.scores"),
 		Types:  c.Ints("tokenizer.ggml.token_type"),
@@ -153,17 +154,17 @@ func New(c fs.Config) (model.Model, error) {
 		},
 	}
 
-	var processor model.TextProcessor
+	var tokenizer tokenizers.Tokenizer
 	switch c.String("tokenizer.ggml.model", "bert") {
 	case "bert":
-		processor = model.NewWordPiece(vocab, true)
+		tokenizer = tokenizers.NewWordPiece(vocab, true)
 	default:
 		return nil, model.ErrUnsupportedTokenizer
 	}
 
 	return &Model{
-		TextProcessor: processor,
-		Layers:        make([]EncoderLayer, c.Uint("block_count")),
+		Tokenizer: tokenizer,
+		Layers:    make([]EncoderLayer, c.Uint("block_count")),
 		Options: Options{
 			hiddenSize:  int(c.Uint("embedding_length")),
 			numHeads:    int(c.Uint("attention.head_count")),

model/models/deepseek2/model.go

@@ -13,6 +13,7 @@ import (
 	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizers"
 )
 
 type Options struct {
@@ -222,7 +223,7 @@ func (t *Layer) Forward(ctx ml.Context, hiddenStates, positions, outputs ml.Tens
 
 type Model struct {
 	model.Base
-	model.BytePairEncoding
+	tokenizers.Tokenizer
 
 	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
 	Layers         []Layer       `gguf:"blk"`
@@ -277,8 +278,8 @@ func New(c fs.Config) (model.Model, error) {
 	}
 
 	m := Model{
-		BytePairEncoding: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		Tokenizer: tokenizers.NewBytePairEncoding(
+			&tokenizers.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

model/models/deepseekocr/model.go

@@ -10,11 +10,12 @@ import (
 	"github.com/ollama/ollama/ml/nn"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizers"
 )
 
 type Model struct {
 	model.Base
-	model.TextProcessor
+	tokenizers.Tokenizer
 
 	Sam    *samModel    `gguf:"s"`
 	Vision *visionModel `gguf:"v"`
@@ -134,8 +135,8 @@ func init() {
 		}
 
 		m := Model{
-			TextProcessor: model.NewBytePairEncoding(
-				&model.Vocabulary{
+			Tokenizer: tokenizers.NewBytePairEncoding(
+				&tokenizers.Vocabulary{
 					Values: c.Strings("tokenizer.ggml.tokens"),
 					Types:  c.Ints("tokenizer.ggml.token_type"),
 					Merges: c.Strings("tokenizer.ggml.merges"),

model/models/gemma2/model.go

@@ -10,6 +10,7 @@ import (
 	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizers"
 )
 
 type Options struct {
@@ -27,7 +28,7 @@ func (o Options) applyRotaryPositionEmbeddings(ctx ml.Context, states, positions
 
 type Model struct {
 	model.Base
-	model.SentencePiece
+	tokenizers.Tokenizer
 
 	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
 	Layers         []Layer       `gguf:"blk"`
@@ -43,8 +44,8 @@ const (
 
 func New(c fs.Config) (model.Model, error) {
 	m := Model{
-		SentencePiece: model.NewSentencePiece(
-			&model.Vocabulary{
+		Tokenizer: tokenizers.NewSentencePiece(
+			&tokenizers.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Scores: c.Floats("tokenizer.ggml.scores"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),

model/models/gemma3/embed.go

@@ -7,11 +7,12 @@ import (
 	"github.com/ollama/ollama/ml/nn/pooling"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizers"
 )
 
 type embedModel struct {
 	model.Base
-	model.SentencePiece
+	tokenizers.Tokenizer
 
 	*TextModel
 	poolingType pooling.Type
@@ -31,8 +32,8 @@ func (m *embedModel) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, erro
 
 func newEmbedModel(c fs.Config) (model.Model, error) {
 	m := &embedModel{
-		SentencePiece: model.NewSentencePiece(
-			&model.Vocabulary{
+		Tokenizer: tokenizers.NewSentencePiece(
+			&tokenizers.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Scores: c.Floats("tokenizer.ggml.scores"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),

model/models/gemma3/model.go

@@ -12,11 +12,12 @@ import (
 	"github.com/ollama/ollama/ml/nn"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizers"
 )
 
 type Model struct {
 	model.Base
-	model.TextProcessor
+	tokenizers.Tokenizer
 
 	*VisionModel `gguf:"v"`
 	*TextModel
@@ -54,7 +55,7 @@ func (p *MultiModalProjector) Forward(ctx ml.Context, visionOutputs ml.Tensor, i
 }
 
 func New(c fs.Config) (model.Model, error) {
-	vocabulary := model.Vocabulary{
+	vocabulary := tokenizers.Vocabulary{
 		Values: c.Strings("tokenizer.ggml.tokens"),
 		Scores: c.Floats("tokenizer.ggml.scores"),
 		Types:  c.Ints("tokenizer.ggml.token_type"),
@@ -70,19 +71,19 @@ func New(c fs.Config) (model.Model, error) {
 		),
 	}
 
-	var processor model.TextProcessor
+	var tokenizer tokenizers.Tokenizer
 	switch c.String("tokenizer.ggml.model") {
 	case "gpt2":
-		processor = model.NewBytePairEncoding(&vocabulary)
+		tokenizer = tokenizers.NewBytePairEncoding(&vocabulary)
 	default:
 		// Previous uploads of Gemma 3 on Ollama did not have token 106
 		// (i.e. "<end_of_turn>") so we need to add in case it's not already present
 		vocabulary.EOS = append(vocabulary.EOS, int32(c.Uint("tokenizer.ggml.eot_token_id", 106)))
-		processor = model.NewSentencePiece(&vocabulary)
+		tokenizer = tokenizers.NewSentencePiece(&vocabulary)
 	}
 
 	m := Model{
-		TextProcessor:  processor,
+		Tokenizer:      tokenizer,
 		ImageProcessor: newImageProcessor(c),
 		VisionModel:    newVisionModel(c),
 		TextModel:      newTextModel(c),

model/models/gemma3n/model.go

@@ -6,11 +6,12 @@ import (
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizers"
 )
 
 type Model struct {
 	model.Base
-	model.SentencePiece
+	tokenizers.Tokenizer
 
 	*TextModel
 }
@@ -23,8 +24,8 @@ func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
 func New(c fs.Config) (model.Model, error) {
 	m := Model{
 		TextModel: newTextModel(c),
-		SentencePiece: model.NewSentencePiece(
-			&model.Vocabulary{
+		Tokenizer: tokenizers.NewSentencePiece(
+			&tokenizers.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Scores: c.Floats("tokenizer.ggml.scores"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),

model/models/gptoss/model.go

@@ -12,11 +12,12 @@ import (
 	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizers"
 )
 
 type Transformer struct {
 	model.Base
-	model.BytePairEncoding
+	tokenizers.Tokenizer
 
 	TokenEmbedding    *nn.Embedding      `gguf:"token_embd"`
 	TransformerBlocks []TransformerBlock `gguf:"blk"`
@@ -196,8 +197,8 @@ func (mlp *MLPBlock) Forward(ctx ml.Context, hiddenStates ml.Tensor, opts *Optio
 func New(c fs.Config) (model.Model, error) {
 	m := Transformer{
 		TransformerBlocks: make([]TransformerBlock, c.Uint("block_count")),
-		BytePairEncoding: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		Tokenizer: tokenizers.NewBytePairEncoding(
+			&tokenizers.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

model/models/llama/model.go

@@ -11,6 +11,7 @@ import (
 	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizers"
 )
 
 type Options struct {
@@ -25,7 +26,7 @@ func (o Options) applyRotaryPositionEmbeddings(ctx ml.Context, states, positions
 
 type Model struct {
 	model.Base
-	model.TextProcessor
+	tokenizers.Tokenizer
 
 	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
 	Layers         []Layer       `gguf:"blk"`
@@ -41,8 +42,8 @@ func New(c fs.Config) (model.Model, error) {
 		return nil, model.ErrUnsupportedModel
 	}
 
-	var processor model.TextProcessor
-	vocabulary := model.Vocabulary{
+	var processor tokenizers.Tokenizer
+	vocabulary := tokenizers.Vocabulary{
 		Values: c.Strings("tokenizer.ggml.tokens"),
 		Scores: c.Floats("tokenizer.ggml.scores"),
 		Types:  c.Ints("tokenizer.ggml.token_type"),
@@ -80,16 +81,16 @@ func New(c fs.Config) (model.Model, error) {
 				"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
 			}
 		}
-		processor = model.NewBytePairEncoding(&vocabulary, pretokenizers...)
+		processor = tokenizers.NewBytePairEncoding(&vocabulary, pretokenizers...)
 	case "llama":
-		processor = model.NewSentencePiece(&vocabulary)
+		processor = tokenizers.NewSentencePiece(&vocabulary)
 	default:
 		return nil, model.ErrUnsupportedTokenizer
 	}
 
 	m := Model{
-		TextProcessor: processor,
-		Layers:        make([]Layer, c.Uint("block_count")),
+		Tokenizer: processor,
+		Layers:    make([]Layer, c.Uint("block_count")),
 		Options: Options{
 			hiddenSize: int(c.Uint("embedding_length")),
 			numHeads:   int(c.Uint("attention.head_count")),

model/models/llama4/model.go

@@ -11,11 +11,12 @@ import (
 	"github.com/ollama/ollama/ml/nn"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizers"
 )
 
 type Model struct {
 	model.Base
-	model.BytePairEncoding
+	tokenizers.Tokenizer
 	ImageProcessor
 
 	*VisionModel `gguf:"v"`
@@ -33,8 +34,8 @@ func (p *Projector) Forward(ctx ml.Context, visionOutputs ml.Tensor) ml.Tensor {
 
 func New(c fs.Config) (model.Model, error) {
 	m := Model{
-		BytePairEncoding: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		Tokenizer: tokenizers.NewBytePairEncoding(
+			&tokenizers.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

model/models/mistral3/model.go

@@ -11,11 +11,12 @@ import (
 	"github.com/ollama/ollama/ml/nn"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizers"
 )
 
 type Model struct {
 	model.Base
-	model.BytePairEncoding
+	tokenizers.Tokenizer
 
 	*TextModel
 	*VisionModel         `gguf:"v"`
@@ -28,12 +29,12 @@ type Model struct {
 var _ model.MultimodalProcessor = (*Model)(nil)
 
 // Implement TextProcessor interface
-var _ model.TextProcessor = (*Model)(nil)
+var _ tokenizers.Tokenizer = (*Model)(nil)
 
 func New(c fs.Config) (model.Model, error) {
 	m := &Model{
-		BytePairEncoding: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		Tokenizer: tokenizers.NewBytePairEncoding(
+			&tokenizers.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

model/models/mllama/model.go

@@ -11,11 +11,12 @@ import (
 	"github.com/ollama/ollama/ml/nn"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizers"
 )
 
 type Model struct {
 	model.Base
-	model.BytePairEncoding
+	tokenizers.Tokenizer
 
 	*VisionModel `gguf:"v"`
 	*TextModel
@@ -32,8 +33,8 @@ const (
 
 func New(c fs.Config) (model.Model, error) {
 	m := Model{
-		BytePairEncoding: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		Tokenizer: tokenizers.NewBytePairEncoding(
+			&tokenizers.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

model/models/nomicbert/model.go

@@ -11,11 +11,12 @@ import (
 	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizers"
 )
 
 type Model struct {
 	model.Base
-	model.TextProcessor
+	tokenizers.Tokenizer
 
 	TokenEmbedding     *nn.Embedding `gguf:"token_embd"`
 	TypeEmbedding      *nn.Embedding `gguf:"token_types"`
@@ -178,8 +179,8 @@ func New(c fs.Config) (model.Model, error) {
 	numHeads := int(c.Uint("attention.head_count"))
 	headDim := hiddenSize / numHeads
 
-	processor := model.NewWordPiece(
-		&model.Vocabulary{
+	tokenizer := tokenizers.NewWordPiece(
+		&tokenizers.Vocabulary{
 			Values: c.Strings("tokenizer.ggml.tokens"),
 			Scores: c.Floats("tokenizer.ggml.scores"),
 			Types:  c.Ints("tokenizer.ggml.token_type"),
@@ -219,8 +220,8 @@ func New(c fs.Config) (model.Model, error) {
 	}
 
 	return &Model{
-		TextProcessor: processor,
-		Layers:        layers,
+		Tokenizer: tokenizer,
+		Layers:    layers,
 		Options: Options{
 			hiddenSize:      hiddenSize,
 			numHeads:        numHeads,

model/models/olmo3/model.go

@@ -11,6 +11,7 @@ import (
 	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizers"
 )
 
 const (
@@ -33,7 +34,7 @@ type Options struct {
 
 type Model struct {
 	model.Base
-	model.TextProcessor
+	tokenizers.Tokenizer
 
 	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
 	Layers         []Layer       `gguf:"blk"`
@@ -44,7 +45,7 @@ type Model struct {
 }
 
 func New(c fs.Config) (model.Model, error) {
-	vocabulary := model.Vocabulary{
+	vocabulary := tokenizers.Vocabulary{
 		Values: c.Strings("tokenizer.ggml.tokens"),
 		Scores: c.Floats("tokenizer.ggml.scores"),
 		Types:  c.Ints("tokenizer.ggml.token_type"),
@@ -58,14 +59,14 @@ func New(c fs.Config) (model.Model, error) {
 		),
 	}
 
-	processor := model.NewBytePairEncoding(
+	tokenizer := tokenizers.NewBytePairEncoding(
 		&vocabulary,
 		"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
 	)
 
 	m := Model{
-		TextProcessor: processor,
-		Layers:        make([]Layer, c.Uint("block_count")),
+		Tokenizer: tokenizer,
+		Layers:    make([]Layer, c.Uint("block_count")),
 		Options: Options{
 			hiddenSize:            int(c.Uint("embedding_length")),
 			numHeads:              int(c.Uint("attention.head_count")),

model/models/qwen2/model.go

@@ -13,6 +13,7 @@ import (
 	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizers"
 )
 
 type Options struct {
@@ -92,7 +93,7 @@ func (d DecoderLayer) Forward(ctx ml.Context, hiddenStates, positions, outputs m
 
 type Model struct {
 	model.Base
-	model.BytePairEncoding
+	tokenizers.Tokenizer
 
 	TokenEmbedding *nn.Embedding  `gguf:"token_embd"`
 	Layers         []DecoderLayer `gguf:"blk"`
@@ -139,8 +140,8 @@ func New(c fs.Config) (model.Model, error) {
 	}
 	m := Model{
 		Layers: make([]DecoderLayer, c.Uint("block_count")),
-		BytePairEncoding: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		Tokenizer: tokenizers.NewBytePairEncoding(
+			&tokenizers.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

model/models/qwen25vl/model.go

@@ -10,11 +10,12 @@ import (
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizers"
 )
 
 type Model struct {
 	model.Base
-	model.BytePairEncoding
+	tokenizers.Tokenizer
 
 	*TextModel
 	*VisionModel `gguf:"v"`
@@ -27,8 +28,8 @@ var _ model.MultimodalProcessor = (*Model)(nil)
 
 func New(c fs.Config) (model.Model, error) {
 	m := &Model{
-		BytePairEncoding: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		Tokenizer: tokenizers.NewBytePairEncoding(
+			&tokenizers.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

model/models/qwen3/embed.go

@@ -7,11 +7,12 @@ import (
 	"github.com/ollama/ollama/ml/nn/pooling"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizers"
 )
 
 type embedModel struct {
 	model.Base
-	model.BytePairEncoding
+	tokenizers.Tokenizer
 
 	*Model
 	poolingType pooling.Type
@@ -34,8 +35,8 @@ func newEmbed(c fs.Config) (model.Model, error) {
 		layers[i].MLP = &dense{}
 	}
 	m := embedModel{
-		BytePairEncoding: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		Tokenizer: tokenizers.NewBytePairEncoding(
+			&tokenizers.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

model/models/qwen3/model.go

@@ -12,6 +12,7 @@ import (
 	"github.com/ollama/ollama/ml/nn/rope"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizers"
 )
 
 type Options struct {
@@ -159,7 +160,7 @@ func (d *Layer) Forward(ctx ml.Context, hiddenStates, positions, outputs ml.Tens
 
 type Model struct {
 	model.Base
-	model.BytePairEncoding
+	tokenizers.BytePairEncoding
 
 	TokenEmbedding *nn.Embedding `gguf:"token_embd"`
 	OutputNorm     *nn.RMSNorm   `gguf:"output_norm"`
@@ -218,8 +219,8 @@ func New(c fs.Config) (model.Model, error) {
 	}
 
 	m := Model{
-		BytePairEncoding: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		BytePairEncoding: tokenizers.NewBytePairEncoding(
+			&tokenizers.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

model/models/qwen3vl/model.go

@@ -10,11 +10,12 @@ import (
 	"github.com/ollama/ollama/ml"
 	"github.com/ollama/ollama/model"
 	"github.com/ollama/ollama/model/input"
+	"github.com/ollama/ollama/tokenizers"
 )
 
 type Model struct {
 	model.Base
-	model.TextProcessor
+	tokenizers.Tokenizer
 
 	*TextModel
 	*VisionModel `gguf:"v"`
@@ -172,8 +173,8 @@ func (m *Model) Forward(ctx ml.Context, batch input.Batch) (ml.Tensor, error) {
 
 func New(c fs.Config) (model.Model, error) {
 	m := Model{
-		TextProcessor: model.NewBytePairEncoding(
-			&model.Vocabulary{
+		Tokenizer: tokenizers.NewBytePairEncoding(
+			&tokenizers.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),
 				Merges: c.Strings("tokenizer.ggml.merges"),

runner/ollamarunner/runner.go

@@ -37,6 +37,7 @@ import (
 	"github.com/ollama/ollama/model/input"
 	"github.com/ollama/ollama/runner/common"
 	"github.com/ollama/ollama/sample"
+	"github.com/ollama/ollama/tokenizers"
 
 	_ "github.com/ollama/ollama/model/models"
 )
@@ -210,9 +211,9 @@ func (s *Server) NewSequence(prompt string, images []llm.ImageData, params NewSe
 }
 
 // calculateLogprobs converts raw logits to log probabilities and finds top K tokens
-func calculateLogprobs(logits []float32, selectedToken int32, topK int, textProcessor model.TextProcessor) []llm.Logprob {
+func calculateLogprobs(logits []float32, selectedToken int32, topK int, tokenizer tokenizers.Tokenizer) []llm.Logprob {
 	decoder := func(tokenID int) string {
-		text, _ := textProcessor.Decode([]int32{int32(tokenID)})
+		text, _ := tokenizer.Decode([]int32{int32(tokenID)})
 		return text
 	}
 	return common.CalculateLogprobs(logits, int(selectedToken), topK, decoder)
@@ -242,7 +243,7 @@ func (s *Server) inputs(prompt string, images []llm.ImageData) ([]*input.Input,
 
 	for i, part := range parts {
 		// text - tokenize
-		tokens, err := s.model.(model.TextProcessor).Encode(part, i == 0)
+		tokens, err := s.model.(tokenizers.Tokenizer).Encode(part, i == 0)
 		if err != nil {
 			return nil, nil, nil, err
 		}
@@ -766,7 +767,7 @@ func (s *Server) computeBatch(activeBatch batchState) {
 		nextBatchTokens[i].Token = token
 
 		// if it's an end of sequence token, break
-		if s.model.(model.TextProcessor).Is(token, model.SpecialEOS) {
+		if s.model.(tokenizers.Tokenizer).Is(token, tokenizers.SpecialEOS) {
 			// TODO (jmorganca): we should send this back
 			// as it's important for the /api/generate context
 			// seq.responses <- piece
@@ -775,14 +776,14 @@ func (s *Server) computeBatch(activeBatch batchState) {
 			continue
 		}
 
-		piece, err := s.model.(model.TextProcessor).Decode([]int32{token})
+		piece, err := s.model.(tokenizers.Tokenizer).Decode([]int32{token})
 		if err != nil {
 			panic("failed to decode token")
 		}
 
 		// Calculate logprobs if requested (after EOS check to avoid logprobs for EOS tokens)
 		if seq.logprobs {
-			logprobs := calculateLogprobs(logits, token, seq.topLogprobs, s.model.(model.TextProcessor))
+			logprobs := calculateLogprobs(logits, token, seq.topLogprobs, s.model.(tokenizers.Tokenizer))
 			seq.pendingLogprobs = append(seq.pendingLogprobs, logprobs...)
 		}
 
@@ -873,7 +874,7 @@ func (s *Server) completion(w http.ResponseWriter, r *http.Request) {
 	var grammar *sample.GrammarSampler
 	var err error
 	if req.Grammar != "" {
-		grammar, err = sample.NewGrammarSampler(s.model.(model.TextProcessor), req.Grammar)
+		grammar, err = sample.NewGrammarSampler(s.model.(tokenizers.Tokenizer), req.Grammar)
 		if err != nil {
 			http.Error(w, "failed to load model vocabulary required for format", http.StatusInternalServerError)
 			return

sample/samplers.go

@@ -7,7 +7,7 @@ import (
 	"slices"
 
 	"github.com/ollama/ollama/llama"
-	"github.com/ollama/ollama/model"
+	"github.com/ollama/ollama/tokenizers"
 )
 
 // token represents information about a single token during sampling
@@ -168,15 +168,15 @@ type GrammarSampler struct {
 	grammar *llama.Grammar
 }
 
-func NewGrammarSampler(model model.TextProcessor, grammarStr string) (*GrammarSampler, error) {
-	vocabIds := make([]uint32, len(model.Vocabulary().Values))
-	pieces := make([]string, len(model.Vocabulary().Values))
-	for i := range model.Vocabulary().Values {
-		pieces[i], _ = model.Decode([]int32{int32(i)})
+func NewGrammarSampler(tokenizer tokenizers.Tokenizer, grammarStr string) (*GrammarSampler, error) {
+	vocabIds := make([]uint32, len(tokenizer.Vocabulary().Values))
+	pieces := make([]string, len(tokenizer.Vocabulary().Values))
+	for i := range tokenizer.Vocabulary().Values {
+		pieces[i], _ = tokenizer.Decode([]int32{int32(i)})
 		vocabIds[i] = uint32(i)
 	}
 
-	grammar := llama.NewGrammar(grammarStr, vocabIds, pieces, model.Vocabulary().EOS)
+	grammar := llama.NewGrammar(grammarStr, vocabIds, pieces, tokenizer.Vocabulary().EOS)
 	if grammar == nil {
 		return nil, errors.New("sample: failed to initialize grammar")
 	}

tokenizers/bytepairencoding.go

@@ -1,8 +1,10 @@
-package model
+package tokenizers
 
 import (
 	"cmp"
+	"fmt"
 	"iter"
+	"log/slog"
 	"slices"
 	"strings"
 
@@ -16,7 +18,7 @@ type BytePairEncoding struct {
 	regexps []*regexp2.Regexp
 }
 
-var _ TextProcessor = (*BytePairEncoding)(nil)
+var _ Tokenizer = (*BytePairEncoding)(nil)
 
 func NewBytePairEncoding(vocab *Vocabulary, pretokenizers ...string) BytePairEncoding {
 	if len(pretokenizers) == 0 {
@@ -243,6 +245,14 @@ func (bpe BytePairEncoding) Encode(s string, addSpecial bool) ([]int32, error) {
 	return ids, nil
 }
 
+type lazyIdsString struct {
+	ids []int32
+}
+
+func (l lazyIdsString) LogValue() slog.Value {
+	return slog.AnyValue(fmt.Sprint(l.ids))
+}
+
 func (bpe BytePairEncoding) Decode(ids []int32) (string, error) {
 	var sb strings.Builder
 	for _, id := range ids {
@@ -267,6 +277,6 @@ func (bpe BytePairEncoding) Decode(ids []int32) (string, error) {
 		}
 	}
 
-	logutil.Trace("decoded", "string", sb.String(), "from", ids)
+	logutil.Trace("decoded", "string", sb.String(), "from", lazyIdsString{ids: ids})
 	return sb.String(), nil
 }

tokenizers/bytepairencoding_test.go

@@ -1,4 +1,4 @@
-package model
+package tokenizers
 
 import (
 	"bufio"
@@ -17,7 +17,7 @@ import (
 func llama(t testing.TB) BytePairEncoding {
 	t.Helper()
 
-	f, err := os.Open(filepath.Join("testdata", "llama3.2", "encoder.json"))
+	f, err := os.Open(filepath.Join("..", "..", "model", "testdata", "llama3.2", "encoder.json"))
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -43,7 +43,7 @@ func llama(t testing.TB) BytePairEncoding {
 		}
 	}
 
-	f, err = os.Open(filepath.Join("testdata", "llama3.2", "vocab.bpe"))
+	f, err = os.Open(filepath.Join("..", "..", "model", "testdata", "llama3.2", "vocab.bpe"))
 	if err != nil {
 		t.Fatal(err)
 	}

tokenizers/sentencepiece.go

@@ -1,4 +1,4 @@
-package model
+package tokenizers
 
 import (
 	"container/heap"
@@ -17,7 +17,7 @@ type SentencePiece struct {
 	vocab       *Vocabulary
 }
 
-var _ TextProcessor = (*SentencePiece)(nil)
+var _ Tokenizer = (*SentencePiece)(nil)
 
 func (spm SentencePiece) Vocabulary() *Vocabulary {
 	return spm.vocab

tokenizers/sentencepiece_test.go

@@ -1,4 +1,4 @@
-package model
+package tokenizers
 
 import (
 	"log/slog"
@@ -15,7 +15,7 @@ import (
 func loadSentencePieceVocab(t *testing.T) SentencePiece {
 	t.Helper()
 
-	bts, err := os.ReadFile(filepath.Join("testdata", "gemma2", "tokenizer.model"))
+	bts, err := os.ReadFile(filepath.Join("..", "..", "model", "testdata", "gemma2", "tokenizer.model"))
 	if err != nil {
 		t.Fatal(err)
 	}

tokenizers/tokenizers.go

@@ -1,4 +1,4 @@
-package model
+package tokenizers
 
 const (
 	TOKEN_TYPE_NORMAL = iota + 1
@@ -9,7 +9,7 @@ const (
 	TOKEN_TYPE_BYTE
 )
 
-type TextProcessor interface {
+type Tokenizer interface {
 	Encode(s string, addSpecial bool) ([]int32, error)
 	Decode([]int32) (string, error)
 	Is(int32, Special) bool

tokenizers/vocabulary.go

@@ -1,4 +1,4 @@
-package model
+package tokenizers
 
 import (
 	"log/slog"

tokenizers/vocabulary_test.go

@@ -1,4 +1,4 @@
-package model
+package tokenizers
 
 import (
 	"testing"

tokenizers/wordpiece.go

@@ -1,4 +1,4 @@
-package model
+package tokenizers
 
 import (
 	"fmt"
@@ -32,7 +32,7 @@ var wordPieceReplacer = strings.NewReplacer(
 	" 're", "'re",
 )
 
-// Decode implements TextProcessor.
+// Decode implements Tokenizer.
 func (wpm WordPiece) Decode(ids []int32) (string, error) {
 	var sb strings.Builder
 	for i, id := range ids {
@@ -96,7 +96,7 @@ func (wpm WordPiece) words(s string) iter.Seq[string] {
 	}
 }
 
-// Encode implements TextProcessor.
+// Encode implements Tokenizer.
 func (wpm WordPiece) Encode(s string, addSpecial bool) ([]int32, error) {
 	var ids []int32
 
@@ -151,17 +151,17 @@ func (wpm WordPiece) Encode(s string, addSpecial bool) ([]int32, error) {
 	return ids, nil
 }
 
-// Is implements TextProcessor.
+// Is implements Tokenizer.
 func (wpm WordPiece) Is(id int32, special Special) bool {
 	return wpm.vocab.Is(id, special)
 }
 
-// Vocabulary implements TextProcessor.
+// Vocabulary implements Tokenizer.
 func (wpm WordPiece) Vocabulary() *Vocabulary {
 	return wpm.vocab
 }
 
-var _ TextProcessor = (*WordPiece)(nil)
+var _ Tokenizer = (*WordPiece)(nil)
 
 func NewWordPiece(vocab *Vocabulary, lowercase bool) WordPiece {
 	return WordPiece{

tokenizers/wordpiece_test.go

@@ -1,4 +1,4 @@
-package model
+package tokenizers
 
 import (
 	"slices"

x/model/models/gemma3/model.go

@@ -9,6 +9,7 @@ import (
 	"slices"
 
 	"github.com/ollama/ollama/fs"
+	"github.com/ollama/ollama/tokenizers"
 	"github.com/ollama/ollama/x/kvcache"
 	"github.com/ollama/ollama/x/ml"
 	"github.com/ollama/ollama/x/ml/nn"
@@ -18,7 +19,7 @@ import (
 
 type Model struct {
 	model.Base
-	model.SentencePiece
+	tokenizers.Tokenizer
 
 	*VisionModel `gguf:"vision_tower.vision_model"`
 	*TextModel   `gguf:"language_model.model"`
@@ -58,8 +59,8 @@ func (p *MultiModalProjector) Forward(ctx ml.Context, visionOutputs ml.Tensor, i
 func New(c fs.Config) (model.Model, error) {
 	// slog.Info("XXX Config", "c", c)
 	m := Model{
-		SentencePiece: model.NewSentencePiece(
-			&model.Vocabulary{
+		Tokenizer: tokenizers.NewSentencePiece(
+			&tokenizers.Vocabulary{
 				Values: c.Strings("tokenizer.ggml.tokens"),
 				Scores: c.Floats("tokenizer.ggml.scores"),
 				Types:  c.Ints("tokenizer.ggml.token_type"),