[serve][llm][sglang] Chat tokenize ignores user's add_generation_prompt setting

[eureka928]

Context

In SGLangServer.tokenize(), when handling a TokenizeChatRequest, the method delegates to _render_chat_prompt() which hardcodes add_generation_prompt=True in the apply_chat_template call.

The vLLM TokenizeChatRequest exposes add_generation_prompt as a user-configurable field, but reusing _render_chat_prompt silently ignores this setting. For tokenization (unlike chat completions), users often need to control whether the generation prompt is included in the token count.

Relevant code

• tokenize() at python/ray/llm/examples/sglang/modules/sglang_engine.py (lines 392-394)
• _render_chat_prompt() at the same file (lines 140-145)

Suggested fix

Pass add_generation_prompt from the request to _render_chat_prompt (or call apply_chat_template directly in tokenize() with the request's field value) instead of hardcoding True.

References

• Found by Cursor Bugbot review on PR [serve][llm] Add tokenize/detokenize to SGLang example engine #61446
• Reviewer @eicherseiji requested issue creation: [serve][llm] Add tokenize/detokenize to SGLang example engine #61446 (comment)