]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commitdiff
vocab : adopt leading TemplateProcessing special token as BOS (#24428)
authoro7si <redacted>
Thu, 11 Jun 2026 07:37:23 +0000 (15:37 +0800)
committerGitHub <redacted>
Thu, 11 Jun 2026 07:37:23 +0000 (10:37 +0300)
gguf-py/gguf/vocab.py

index f8d3b3e7406a21f8d2e78128d798844abccf8d10..d93b94f2d792147276be21db004dbd8d4edef82c 100644 (file)
@@ -256,6 +256,11 @@ class SpecialVocab:
                             if special_first := tmpl_single[0].get('SpecialToken', {}).get('id'):
                                 if not tokenizer_config:
                                     special_bos = special_first
+                                elif special_first not in (special_bos, special_cls):
+                                    if not special_bos:
+                                        tokenizer_config['bos_token'] = special_bos = special_first
+                                    if not special_cls:
+                                        tokenizer_config['cls_token'] = special_cls = special_first
                                 self.add_special_token['bos'] = True if special_first in (special_bos, special_cls) else False
                                 if special_first not in (special_bos, special_cls):
                                     logger.warning(f'Unknown leading special token {special_first!r} in TemplateProcessing<single>')