From 4845b90d9b5f1ef643355a4886d745c1fdd6b4fb Mon Sep 17 00:00:00 2001 From: yurekami Date: Thu, 25 Dec 2025 03:24:05 +0900 Subject: [PATCH] fix: preserve original tokenizer class name for HuggingFace compatibility MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Address review feedback: correctly set CachedTokenizer.__name__ to the original tokenizer's class name instead of 'CachedTokenizer'. This ensures transformers processors can validate the tokenizer type by checking __name__, which previously failed validation. Fixes #31080 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 Signed-off-by: yurekami --- vllm/tokenizers/hf.py | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/vllm/tokenizers/hf.py b/vllm/tokenizers/hf.py index a7b565dca5d8f..2a95f9c82fbd3 100644 --- a/vllm/tokenizers/hf.py +++ b/vllm/tokenizers/hf.py @@ -58,7 +58,9 @@ def get_cached_tokenizer(tokenizer: HfTokenizer) -> HfTokenizer: def __reduce__(self): return get_cached_tokenizer, (tokenizer,) - CachedTokenizer.__name__ = f"Cached{tokenizer.__class__.__name__}" + # Preserve original class name for HuggingFace compatibility. + # Some processors validate tokenizer type by checking __name__. + CachedTokenizer.__name__ = tokenizer.__class__.__name__ cached_tokenizer.__class__ = CachedTokenizer return cached_tokenizer