Edit LLM Inference Style (speedy_utils)

Use this skill to standardize or modify how speedy_utils builds prompts and consumes generation outputs. It focuses on chat templating, reasoning-style prefixes, and safe stopping rules for structured answer extraction.

When to Use This Skill

Use this skill when you need to:

• Insert or enforce a reasoning prefix (for example, <think>\n).
• Switch a flow to LLM.generate() instead of chat completion helpers.
• Apply a tokenizer chat template before generating.
• Stop generations on boxed answers (\boxed{}) or <|im_end|> tokens.
• Normalize outputs before evaluation (for example, GSM8K or math tasks).

Prerequisites

• A model-backed tokenizer available via transformers.AutoTokenizer.
• A speedy_utils LLM instance configured to point at the correct backend.

Core Capabilities