Zero-Shot Image Classification
Transformers
Safetensors
siglip
vision

Remove the Gemma chat template from tokenizer_config.json

#23
by alexwengg - opened

SigLIP2's text tower reuses the Gemma tokenizer, and tokenizer_config.json still carries Gemma's chat_template. SigLIP2 has no chat format, so tools that honor a tokenizer's chat template wrap the text in <start_of_turn>user ... <end_of_turn> markers and drop EOS. For example, vLLM's /v1/embeddings with messages returns a different embedding than for the same plain text (11 tokens instead of 6 for "a photo of a cat"; vLLM-side fix in https://github.com/vllm-project/vllm/pull/66624).

This PR removes that one key. Tokenizer and processor output are unchanged (checked with Transformers); only apply_chat_template stops working, which has no meaning for this model.

The same template is in the other fixed-resolution SigLIP2 checkpoints, which could take the same one-line change: siglip2-base-patch16-256, -384, -512, siglip2-base-patch32-256, siglip2-large-patch16-256, -384, -512, siglip2-so400m-patch14-224, -384, siglip2-so400m-patch16-256, -384, -512, and siglip2-giant-opt-patch16-256, -384. The NaFlex checkpoints don't have it.

Ready to merge
This branch is ready to get merged automatically.

Sign up or log in to comment