Skip to content

Get up and running with Kimi, GLM, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models.

use ollama engine for bert models (#13501)

Michael Yang903b1fc97f37
diff --git a/fs/ggml/ggml.go b/fs/ggml/ggml.go
index 56614a321..44a48511c 100644
--- a/fs/ggml/ggml.go
+++ b/fs/ggml/ggml.go
@@ -241,19 +241,20 @@ func (kv KV) Bools(key string, defaultValue ...[]bool) []bool {
func (kv KV) OllamaEngineRequired() bool {
return slices.Contains([]string{
+ "bert",
+ "deepseek2",
+ "deepseekocr",
"gemma3",
"gemma3n",
"gptoss", "gpt-oss",
"llama4",
"mistral3",
"mllama",
+ "nomic-bert",
+ "olmo3",
"qwen25vl",
"qwen3", "qwen3moe",
"qwen3vl", "qwen3vlmoe",
- "deepseekocr",
- "deepseek2",
- "nomic-bert",
- "olmo3",
}, kv.Architecture())
}
@@ -839,6 +840,7 @@ func (f GGML) SupportsFlashAttention() bool {
// FlashAttention checks if the model should enable flash attention
func (f GGML) FlashAttention() bool {
return slices.Contains([]string{
+ "bert",
"gemma3",
"gptoss", "gpt-oss",
"mistral3",
diff --git a/model/models/bert/embed.go b/model/models/bert/embed.go
index 5e7ca5e92..705c63138 100644
--- a/model/models/bert/embed.go
+++ b/model/models/bert/embed.go
@@ -129,35 +129,36 @@ func (o Options) headDim() int {
}
func New(c fs.Config) (model.Model, error) {
+ vocab := &model.Vocabulary{
+ Values: c.Strings("tokenizer.ggml.tokens"),
+ Scores: c.Floats("tokenizer.ggml.scores"),
+ Types: c.Ints("tokenizer.ggml.token_type"),
+ AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
+ BOS: []int32{
+ int32(cmp.Or(
+ c.Uint("tokenizer.ggml.cls_token_id"),
+ c.Uint("tokenizer.ggml.bos_token_id"),
+ )),
+ },
+ AddEOS: c.Bool("tokenizer.ggml.add_eos_token", true),
+ EOS: []int32{
+ int32(cmp.Or(
+ c.Uint("tokenizer.ggml.separator_token_id"),
+ //nolint:misspell
+ // NOTE: "seperator_token_id" is a typo in model metadata but we need to
+ // support it for compatibility.
+ c.Uint("tokenizer.ggml.seperator_token_id"),
+ c.Uint("tokenizer.ggml.eos_token_id"),
+ )),
+ },
+ }
+
var processor model.TextProcessor
switch c.String("tokenizer.ggml.model", "bert") {
case "bert":
- processor = model.NewWordPiece(
- &model.Vocabulary{
- Values: c.Strings("tokenizer.ggml.tokens"),
- Scores: c.Floats("tokenizer.ggml.scores"),
- Types: c.Ints("tokenizer.ggml.token_type"),
- AddBOS: c.Bool("tokenizer.ggml.add_bos_token", true),
- BOS: []int32{
- int32(cmp.Or(
- c.Uint("tokenizer.ggml.cls_token_id"),
- c.Uint("tokenizer.ggml.bos_token_id"),
- )),
- },
- AddEOS: c.Bool("tokenizer.ggml.add_eos_token", true),
- EOS: []int32{
- int32(cmp.Or(
- c.Uint("tokenizer.ggml.separator_token_id"),
- //nolint:misspell
- // NOTE: "seperator_token_id" is a typo in model metadata but we need to
- // support it for compatibility.
- c.Uint("tokenizer.ggml.seperator_token_id"),
- c.Uint("tokenizer.ggml.eos_token_id"),
- )),
- },
- },
- true,
- )
+ processor = model.NewWordPiece(vocab, true)
+ case "gpt2":
+ processor = model.NewBytePairEncoding(vocab)
default:
return nil, model.ErrUnsupportedTokenizer
}