generate

type GenerateConfig struct {
	// MaxTokens is the maximum number of tokens to generate.
	MaxTokens int

	// MinTokens is the minimum number of tokens before stopping.
	MinTokens int

	// StopStrings are strings that trigger stopping.
	StopStrings []string

	// StopTokens are token IDs that trigger stopping.
	StopTokens []int32

	// Stream enables streaming generation.
	Stream bool

	// EchoPrompt includes the prompt in output.
	EchoPrompt bool

	// Sampling is the sampling configuration.
	Sampling SamplingConfig
}

type GenerateResult struct {
	Token   string // Decoded token text
	TokenID int32  // Token ID
	Done    bool   // Is generation complete
	Reason  string // Stop reason: "eos", "max_tokens", "stop_string", "stop_token"
	Error   error  // Error if any
}

type GeneratorOption func(*generatorOptions)

type KVCache interface {
	Clear()
}

type LLMModel interface {
	// Forward runs a forward pass and returns logits.
	// Input shape: [batch, seq_len]
	// Output shape: [batch, seq_len, vocab_size]
	Forward(input *tensor.RawTensor, cache KVCache, startPos int) *tensor.RawTensor

	// VocabSize returns the vocabulary size.
	VocabSize() int
}

type Sampler struct {
	// contains filtered or unexported fields
}

type SamplingConfig struct {
	// Temperature controls randomness. 0 = greedy, 1 = normal, >1 = more random.
	Temperature float32

	// TopK limits sampling to top K tokens. 0 = disabled.
	TopK int

	// TopP (nucleus sampling) limits to tokens with cumulative prob < P. 1.0 = disabled.
	TopP float32

	// MinP filters tokens with prob < max_prob * MinP. 0 = disabled.
	MinP float32

	// Repetition control
	RepeatPenalty    float32 // Penalty for repeated tokens. 1.0 = no penalty.
	FrequencyPenalty float32 // Penalty based on frequency. 0 = disabled.
	PresencePenalty  float32 // Penalty for presence. 0 = disabled.
	RepeatWindow     int     // Number of tokens to consider. 0 = all.

	// Seed for reproducibility. -1 = random.
	Seed int64
}

type SpeculativeConfig struct {
	// DraftModel is the small fast model for speculation.
	DraftModel LLMModel

	// TargetModel is the large accurate model for verification.
	TargetModel LLMModel

	// NumSpeculate is the number of tokens to speculate (default: 5).
	NumSpeculate int

	// Sampling is the sampling configuration.
	Sampling SamplingConfig
}

type SpeculativeGenerator struct {
	// contains filtered or unexported fields
}

type TextGenerator struct {
	// contains filtered or unexported fields
}