stop tokens

**Stop tokens** is the **special token IDs that instruct the decoder to terminate generation immediately when emitted** - they provide low-level termination control at token granularity. **What Is Stop tokens?** - **Definition**: Model-recognized token markers treated as hard completion boundaries. - **Typical Examples**: EOS markers and custom control tokens reserved by tokenizer vocabulary. - **Execution Behavior**: When generated, decoding loop exits without adding further tokens. - **Scope**: Used internally by runtimes and exposed through API configuration in some systems. **Why Stop tokens Matters** - **Termination Precision**: Enables deterministic ending behavior independent of text matching. - **Format Integrity**: Helps close structured outputs cleanly at expected boundaries. - **Runtime Simplicity**: Token-based checks are fast and reliable compared with string scans. - **Safety**: Supports strict cutoffs for guarded completion flows. - **Interoperability**: Aligns behavior across serving backends using shared token IDs. **How It Is Used in Practice** - **Vocabulary Mapping**: Verify stop-token IDs against tokenizer version and model checkpoint. - **Priority Rules**: Define interactions between stop tokens and stop sequences. - **Regression Tests**: Validate no premature stops under multilingual and code-generation prompts. Stop tokens is **a foundational primitive for deterministic decode termination** - correct token mapping is essential to avoid truncation or runaway output.