Skip to content

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

temporary work around for converting spm

Patrick Devine631fecc6d9fe
diff --git a/convert/tokenizer_spm.go b/convert/tokenizer_spm.go
index d8a012c08..340c3d581 100644
--- a/convert/tokenizer_spm.go
+++ b/convert/tokenizer_spm.go
@@ -47,6 +47,12 @@ func parseSentencePiece(fsys fs.FS) (*Vocabulary, error) {
v.Types = append(v.Types, int32(t))
default:
tt := int32(sentencepiece.ModelProto_SentencePiece_NORMAL)
+
+ // temporary fix to handle gemma3 broken configs
+ if slices.Contains([]string{"<end_of_turn>", "<start_of_turn>"}, piece.GetPiece()) {
+ tt = int32(sentencepiece.ModelProto_SentencePiece_CONTROL)
+ }
+
for _, t := range ast {
if t.Content == piece.GetPiece() {
tt = int32(sentencepiece.ModelProto_SentencePiece_CONTROL)