For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).
Python class
ReasoningPipelineTokenizer
ReasoningPipelineTokenizer
class max.pipelines.modeling.types.ReasoningPipelineTokenizer(*args, **kwargs)
Bases: PipelineTokenizer[UnboundContextType, TokenizerEncoded, RequestType], Protocol[UnboundContextType, TokenizerEncoded, RequestType]
PipelineTokenizer that exposes its reasoning-delimiter token ids.
Implemented by architecture-specific tokenizers that drive a reasoning
parser (Gemma 4, Kimi K2.5, MiniMax M2). The tokenizer resolves the
delimiter ids once at construction and exposes them as instance
attributes so callers — for example
OverlapTextGenerationPipeline’s
thinking-mode temperature scaling — can read them directly without
re-encoding <think>/</think> or depending on the reasoning
parser registry.
reasoning_end_token_id
property reasoning_end_token_id: int
The token id that closes a reasoning span (e.g. <channel|>).
reasoning_start_token_id
property reasoning_start_token_id: int
The token id that opens a reasoning span (e.g. <|channel>).
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!