Model Parameters

no_llm provides two parameter classes: ModelParameters for runtime use and ConfigurableModelParameters for model configuration.

Parameters vs Configurable Parameters

• ModelParameters: Used at runtime for parameter passing, with simple value validation
• ConfigurableModelParameters: Used in model configurations, with full validation rules, variants, and capability checks

Supported Parameters

Sampling Parameters

Parameter	Type	Description	Default
temperature	float	Controls randomness in generation	1.0
top_p	float	Nucleus sampling threshold	1.0
top_k	int | None	Top-k sampling threshold	None

Penalty Parameters

Parameter	Type	Description	Default
frequency_penalty	float | None	Penalty for token frequency	0.0
presence_penalty	float | None	Penalty for token presence	0.0
logit_bias	dict[str, float] | None	Token biasing dictionary	None

Output Parameters

Parameter	Type	Description	Default
max_tokens	int | None	Maximum tokens to generate	None
stop	list[str] | None	Stop sequences	None
logprobs	int | None	Number of logprobs to return	None
top_logprobs	int | None	Most likely tokens to return	None
seed	int | None	Random seed for reproducibility	None

Request Parameters

Parameter	Type	Description	Default
timeout	float | None	Request timeout in seconds	None

Reasoning Parameters

Parameter	Type	Description	Default	Required Capability
include_reasoning	bool | None	Include reasoning steps	False	REASONING
reasoning_effort	"low"|"medium"|"high" | None	Reasoning level	None	REASONING

Usage Example

# Runtime parameter passing
from no_llm.config.parameters import ModelParameters

params = ModelParameters(
    temperature=0.7,
    max_tokens=100,
    include_reasoning=True
)

# Model configuration
from no_llm.config.parameters import ConfigurableModelParameters

config_params = ConfigurableModelParameters(
    temperature=ParameterValue(
        variant=ParameterVariant.VARIABLE,
        value=0.7,
        validation_rule=RangeValidation(min_value=0.0, max_value=2.0)
    )
)

NOT_GIVEN vs None

In ModelParameters, unset parameters use the special value 'NOT_GIVEN' instead of None. This allows distinguishing between explicitly setting a parameter to None and not setting it at all.

Parameter Conversion Flow

ConfigurableModelParameters can be converted to ModelParameters through validation:

# Model configuration with validation rules
config_params = ConfigurableModelParameters(
    temperature=ParameterValue(
        variant=ParameterVariant.VARIABLE,
        value=0.7,
        validation_rule=RangeValidation(min_value=0.0, max_value=2.0)
    )
)

# Get validated runtime parameters
validated_params = config_params.validate_parameters(
    capabilities={ModelCapability.STREAMING},
    temperature=0.8  # Will be validated against rules
)

# Create ModelParameters from validated values
model_params = ModelParameters(**validated_params)

Extensibility

The parameter system can be extended through inheritance to support additional parameters Check out the Parameter Customization section for more information.