GenerationOptions

Overview

Fm::GenerationOptions controls how the model generates responses, including temperature, sampling strategy, and maximum token count.

Constructor

Fm::GenerationOptions.new(
  temperature : Float64? = nil,
  sampling : Fm::Sampling? = nil,
  max_response_tokens : UInt32? = nil,
  seed : UInt64? = nil
)

All parameters are optional. When nil, the model uses its default values.

Parameter	Type	Default	Description
`temperature`	`Float64?`	`nil`	Sampling temperature (0.0 - 2.0)
`sampling`	`Sampling?`	`nil`	Sampling strategy
`max_response_tokens`	`UInt32?`	`nil`	Maximum tokens in the response
`seed`	`UInt64?`	`nil`	Random seed for reproducible output

Class Methods

`.default`

Fm::GenerationOptions.default : Fm::GenerationOptions

Returns an instance with all parameters set to nil, using the model's defaults.

Properties

Property	Type	Description
`temperature`	`Float64?`	Controls randomness. Lower values are more deterministic.
`sampling`	`Sampling?`	The sampling strategy to use.
`max_response_tokens`	`UInt32?`	Limits the length of the generated response.
`seed`	`UInt64?`	Random seed for reproducible output.

Fm::Sampling

Value	Description
`Random`	Random sampling with temperature (default)
`Greedy`	Always pick the most likely token

Fm::SamplingMode

Advanced sampling mode with top-k and top-p (nucleus sampling) support.

Factory Methods

`.greedy`

Fm::SamplingMode.greedy : Fm::SamplingMode

Deterministic sampling — always pick the most likely token.

`.random`

Fm::SamplingMode.random(
  top : Int32? = nil,
  probability_threshold : Float64? = nil,
  seed : UInt64? = nil
) : Fm::SamplingMode

Random sampling with optional constraints.

Parameter	Type	Description
`top`	`Int32?`	Top-k: only consider the `k` most likely tokens
`probability_threshold`	`Float64?`	Top-p (nucleus): cumulative probability threshold (0.0 - 1.0)
`seed`	`UInt64?`	Seed for reproducible output

Note: top and probability_threshold cannot both be specified.

Note: seed alone (without top or probability_threshold) is not supported by Apple's FoundationModels API and will be ignored. Use seed together with top or probability_threshold for reproducible sampling.

Properties

Property	Type	Description
`strategy`	`Sampling`	The base strategy (Random or Greedy)
`top`	`Int32?`	Top-k value
`probability_threshold`	`Float64?`	Top-p value
`seed`	`UInt64?`	Seed for reproducibility

Examples

Creative output

options = Fm::GenerationOptions.new(
  temperature: 1.5,
  sampling: Fm::Sampling::Random,
  max_response_tokens: 1000_u32
)

Deterministic output

options = Fm::GenerationOptions.new(
  temperature: 0.0,
  sampling: Fm::Sampling::Greedy
)

Top-k sampling

options = Fm::GenerationOptions.new(
  temperature: 0.8,
  sampling_mode: Fm::SamplingMode.random(top: 40)
)

Top-p (nucleus) sampling

options = Fm::GenerationOptions.new(
  temperature: 0.9,
  sampling_mode: Fm::SamplingMode.random(probability_threshold: 0.9)
)

Reproducible output

options = Fm::GenerationOptions.new(
  temperature: 0.0,
  sampling: Fm::Sampling::Greedy,
  seed: 42_u64
)

Concise output

options = Fm::GenerationOptions.new(
  max_response_tokens: 100_u32
)