From 0a784858283fc0539313c42e675af04eb937f676 Mon Sep 17 00:00:00 2001
From: huni <chenweize1202@gmail.com>
Date: Mon, 24 Aug 2026 18:07:08 +0800
Subject: [PATCH] add fix_tokenizer.py

---
 fix_tokenizer.py | 71 ++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 71 insertions(+)
 create mode 100644 fix_tokenizer.py

diff --git a/fix_tokenizer.py b/fix_tokenizer.py
new file mode 100644
index 0000000..8e012a3
--- /dev/null
+++ b/fix_tokenizer.py
@@ -0,0 +1,71 @@
+import os
+import shutil
+import json
+from detect_tokenizer import detect
+
+MODEL_DIR = os.environ.get("MODEL_DIR", "/model")
+OUT_DIR = os.environ.get("FIX_TOKENIZER_DIR", "/tmp/fixed_tokenizer")
+
+os.makedirs(OUT_DIR, exist_ok=True)
+
+def copy_if_exists(name):
+    src = os.path.join(MODEL_DIR, name)
+    if os.path.exists(src):
+        shutil.copy(src, OUT_DIR)
+
+# 复制所有可能相关文件
+for f in [
+    "tokenizer.json",
+    "tokenizer_config.json",
+    "special_tokens_map.json",
+    "vocab.json",
+    "merges.txt",
+    "tokenizer.model",
+]:
+    copy_if_exists(f)
+
+typ, orig_cls = detect(MODEL_DIR)
+
+cfg_path = os.path.join(OUT_DIR, "tokenizer_config.json")
+
+if os.path.exists(cfg_path):
+    with open(cfg_path) as f:
+        cfg = json.load(f)
+else:
+    cfg = {}
+
+# ===== 自动修复策略 =====
+if typ == "fast":
+    cfg["tokenizer_class"] = "PreTrainedTokenizerFast"
+    cfg["from_slow"] = False
+    cfg.pop("backend", None)
+
+elif typ == "sentencepiece":
+    cfg["tokenizer_class"] = "LlamaTokenizer"
+
+elif typ == "bpe":
+    cfg["tokenizer_class"] = "GPT2TokenizerFast"
+
+else:
+    cfg["tokenizer_class"] = "PreTrainedTokenizerFast"
+
+# 特殊 case 修复
+bad_classes = [
+    "TokenizersBackend",
+    "TiktokenTokenizer",
+]
+
+if orig_cls in bad_classes:
+    print(f"[fix] override bad tokenizer_class: {orig_cls} → {cfg['tokenizer_class']}")
+
+# 修复 extra_special_tokens: list → dict 格式
+if "extra_special_tokens" in cfg and isinstance(cfg["extra_special_tokens"], list):
+    orig_list = cfg["extra_special_tokens"]
+    cfg["extra_special_tokens"] = {token: token for token in orig_list}
+    print(f"[fix] converted extra_special_tokens from list ({len(orig_list)} items) to dict format")
+
+# 写回
+with open(cfg_path, "w") as f:
+    json.dump(cfg, f)
+
+print(f"[fix_tokenizer] done → {OUT_DIR}")