move tokenizers to separate package (#13825)

This commit is contained in:
Michael Yang
2026-02-05 17:44:11 -08:00
committed by GitHub
parent 8a4b77f9da
commit f1373193dc
79 changed files with 200 additions and 6720 deletions

View File

@@ -7,7 +7,7 @@ import (
"slices"
"github.com/ollama/ollama/llama"
"github.com/ollama/ollama/model"
"github.com/ollama/ollama/tokenizer"
)
// 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(tok tokenizer.Tokenizer, grammarStr string) (*GrammarSampler, error) {
vocabIds := make([]uint32, len(tok.Vocabulary().Values))
pieces := make([]string, len(tok.Vocabulary().Values))
for i := range tok.Vocabulary().Values {
pieces[i], _ = tok.Decode([]int32{int32(i)})
vocabIds[i] = uint32(i)
}
grammar := llama.NewGrammar(grammarStr, vocabIds, pieces, model.Vocabulary().EOS)
grammar := llama.NewGrammar(grammarStr, vocabIds, pieces, tok.Vocabulary().EOS)
if grammar == nil {
return nil, errors.New("sample: failed to initialize grammar")
}

View File

@@ -8,7 +8,7 @@ import (
"path/filepath"
"testing"
"github.com/ollama/ollama/model"
"github.com/ollama/ollama/tokenizer"
)
func TestWeighted(t *testing.T) {
@@ -60,10 +60,10 @@ func TestWeighted(t *testing.T) {
}
}
func modelHelper(t testing.TB) model.BytePairEncoding {
func modelHelper(t testing.TB) tokenizer.Tokenizer {
t.Helper()
f, err := os.Open(filepath.Join("..", "model", "testdata", "llama3.2", "encoder.json"))
f, err := os.Open(filepath.FromSlash("../tokenizer/testdata/llama3.2/encoder.json"))
if err != nil {
t.Fatal(err)
}
@@ -81,8 +81,8 @@ func modelHelper(t testing.TB) model.BytePairEncoding {
merges := make([]string, 0, 1)
// Only need vocab for Grammar Test
return model.NewBytePairEncoding(
&model.Vocabulary{
return tokenizer.NewBytePairEncoding(
&tokenizer.Vocabulary{
Values: tokens,
Types: make([]int32, len(vocab)),
Merges: merges,