From e3c7d4dc4c9d960fce86af64bb03490900457ac4 Mon Sep 17 00:00:00 2001
From: Vineeth Sai <vineethsai4444@gmail.com>
Date: Tue, 23 Jun 2026 15:40:50 -0700
Subject: [PATCH] Match IGNORED_TOKENIZER_NAMES case-insensitively (#6620)

---
 tests/test_ignored_tokenizer_casing.py | 60 ++++++++++++++++++++++++++
 unsloth/tokenizer_utils.py             |  2 +-
 2 files changed, 61 insertions(+), 1 deletion(-)
 create mode 100644 tests/test_ignored_tokenizer_casing.py

diff --git a/tests/test_ignored_tokenizer_casing.py b/tests/test_ignored_tokenizer_casing.py
new file mode 100644
index 0000000000..aeeea981fc
--- /dev/null
+++ b/tests/test_ignored_tokenizer_casing.py
@@ -0,0 +1,60 @@
+# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Regression test: the IGNORED_TOKENIZER_NAMES guard must match case-insensitively."""
+
+import types
+from unittest.mock import patch
+
+import unsloth.tokenizer_utils as tu
+
+
+class _Tok:
+    def __init__(self, kind):
+        self.kind = kind
+
+
+def _fake_auto_tokenizer():
+    # _load_correct_tokenizer loads a slow tokenizer (from_slow = True) and a fast one;
+    # return distinguishable stand-ins so we can see which one is returned.
+    def from_pretrained(name, **kwargs):
+        return _Tok("slow" if kwargs.get("from_slow") else "fast")
+
+    return types.SimpleNamespace(from_pretrained = from_pretrained)
+
+
+def test_ignored_tokenizer_name_matched_case_insensitively():
+    # IGNORED_TOKENIZER_NAMES is stored fully lowercased, but users pass canonical
+    # mixed-case ids. The names in this list (e.g. the Qwen2.5-Coder tokenizers) must be
+    # returned untouched; before the fix the mixed-case name never matched, so it fell
+    # through into the slow/fast reconciliation + conversion path.
+    name = "unsloth/Qwen2.5-Coder-7B-Instruct"  # name.lower() is in IGNORED_TOKENIZER_NAMES
+    assert name.lower() in tu.IGNORED_TOKENIZER_NAMES  # guard the test's own premise
+
+    with (
+        patch.object(tu, "AutoTokenizer", _fake_auto_tokenizer()),
+        patch.object(tu, "assert_same_tokenization", lambda a, b: False),
+        patch.object(tu, "convert_to_fast_tokenizer", lambda slow, **k: _Tok("converted")),
+    ):
+        result = tu._load_correct_tokenizer(name, fix_tokenizer = True)
+
+    assert result.kind == "fast", (
+        "an ignored tokenizer (matched case-insensitively) must be returned unchanged, "
+        f"but it fell through to conversion (got {result.kind!r})"
+    )
+
+
+if __name__ == "__main__":
+    test_ignored_tokenizer_name_matched_case_insensitively()
+    print("ok")
diff --git a/unsloth/tokenizer_utils.py b/unsloth/tokenizer_utils.py
index 3cec002012..eb18b1966c 100644
--- a/unsloth/tokenizer_utils.py
+++ b/unsloth/tokenizer_utils.py
@@ -601,7 +601,7 @@ def _load_correct_tokenizer(
         cache_dir = cache_dir,
     )
 
-    if not fix_tokenizer or tokenizer_name in IGNORED_TOKENIZER_NAMES:
+    if not fix_tokenizer or tokenizer_name.lower() in IGNORED_TOKENIZER_NAMES:
         return fast_tokenizer
     # Ignore Mistral ones - they're a bit weird to handle!
     elif "mistral" in tokenizer_name.lower():