mirror of
https://github.com/ollama/ollama.git
synced 2026-11-08 02:16:57 -06:00
Daniel HiltgenandGitHub
195f4cdca1
* mlx: match publisher tokenizer semantics Honor pretokenizer stage order, split behavior, Unicode boundaries, added-token normalization, and ranked BPE merges. Handle empty added tokens and empty Metaspace input consistently. Add shared Go/Python reference cases using published tokenizers, pulling missing models directly and failing on errors, plus focused regressions for configuration precedence, byte fallback, and parallel encoding. * address comments
50 lines
1.3 KiB
Go
50 lines
1.3 KiB
Go
package tokenizer_test
|
|
|
|
import (
|
|
"strings"
|
|
"testing"
|
|
|
|
"github.com/ollama/ollama/mlxrunner/tokenizer"
|
|
)
|
|
|
|
func BenchmarkTokenizerReference(b *testing.B) {
|
|
for _, model := range []struct{ name, special string }{
|
|
{"qwen3.5:0.8b-mxfp8", "<|endoftext|>"},
|
|
{"nemotron-3.5-lightning:30b-a3b-nvfp4", "<unk>"},
|
|
{"muse-glimmer:30b-nvfp4", "<|begin_of_text|>"},
|
|
} {
|
|
b.Run(model.name, func(b *testing.B) {
|
|
tok, err := tokenizer.LoadFromBytes(loadTokenizerReference(b, model.name))
|
|
if err != nil {
|
|
b.Fatal(err)
|
|
}
|
|
if _, ok := tok.GetSpecialToken(model.special); !ok {
|
|
b.Fatalf("missing special token %q", model.special)
|
|
}
|
|
for _, input := range []struct {
|
|
name string
|
|
text string
|
|
}{
|
|
{"short", "The quick brown fox jumps over the lazy dog. "},
|
|
{"270KB", strings.Repeat("The quick brown fox jumps over the lazy dog. ", 6000)},
|
|
{"special100", strings.Repeat("x"+model.special, 100)},
|
|
{"special800", strings.Repeat("x"+model.special, 800)},
|
|
} {
|
|
b.Run(input.name, func(b *testing.B) {
|
|
b.ReportAllocs()
|
|
b.SetBytes(int64(len(input.text)))
|
|
b.ResetTimer()
|
|
var ids []int32
|
|
for range b.N {
|
|
ids = tok.Encode(input.text, false)
|
|
}
|
|
b.StopTimer()
|
|
if len(ids) == 0 {
|
|
b.Fatal("encoding produced no tokens")
|
|
}
|
|
})
|
|
}
|
|
})
|
|
}
|
|
}
|