SAEnhancerConfig#

class tensorrt_llm.llmapi.SAEnhancerConfig( *, threshold: Annotated[int, Gt(gt=0)] = 4, enable_global_pool: bool = False, )[source]#

Bases: StrictBaseModel

Configuration for the Suffix Automaton (SA) draft enhancer.

Use this to combine SA pattern-matching drafting with another speculative decoding method (Eagle3, MTP, PARD). When provided as sa_config on a decoding config, SA drafting is enabled and may override neural draft tokens when the suffix match length meets the threshold.

For standalone SA speculative decoding (no neural drafter), use SADecodingConfig instead.

field enable_global_pool: bool = False#: When True, each request searches all active SA states for the longest match, not just its own. Improves acceptance rates when requests share common patterns.

field threshold: Annotated[int, Gt(gt=0)] = 4#

Minimum suffix match length required for the SA output to override neural draft tokens.

Constraints:

gt = 0

__init__(**data: Any) → None#

Create a new model by parsing and validating input data from keyword arguments.

Raises [ValidationError][pydantic_core.ValidationError] if the input data cannot be validated to form a valid model.

self is explicitly positional-only to allow self as a field name.