Feature request

A built-in way to cap how many tokens a reasoning model spends inside its <think> … </think> block. Today, we can only control the total response length via max_new_tokens. No parameter limits the internal reasoning segment when enable_thinking=True.

Motivation

• Reasoning models (e.g., Qwen3 series) often produce very long thought blocks, which can blow past latency budgets before the final answer starts.
• Users need a simple, model-agnostic control to bound that “thinking” cost without disabling reasoning entirely.
• The Qwen docs (https://qwen.readthedocs.io/en/latest/getting_started/quickstart.html#thinking-budget) already describe a brute-force approach (two-step generation) to implement “thinking budgets”.

Your contribution

I want to submit a PR that:

• Extends GenerationConfig with:
  max_thinking_tokens: integer budget for reasoning tokens.
  begin_thinking_token_id / end_thinking_token_id: marker IDs so generation knows where the thinking span begins/ends.
• Add a MaxThinkingTokensLogitsProcessor that watches the active <think> block. Once the budget is reached, it forces end_thinking_token_id, ensuring the model exits reasoning and continues with the final response.
• Document the new parameter in reasoning-model guides (EXAONE, CWM, etc.) and show how to wire the thinking-token IDs until configs do it automatically.
• Provide unit coverage so _get_logits_processor injects the new processor whenever the config is fully specified.