llamacpp

type AttentionType int32

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

type BatchEmbeddings struct {
	Embeddings [][]float32 // One embedding vector per sequence/token
	Dimension  int         // Embedding dimension
}

type ChatMessage struct {
	Role    string // "system", "user", or "assistant"
	Content string // The message content
}

type CompletionOptions struct {
	// MaxTokens is the maximum number of tokens to generate (0 = default 128)
	MaxTokens int

	// StopWords are strings that stop generation when encountered
	StopWords []string

	// OnToken callback is called for each generated token
	// Return false to stop generation early
	OnToken func(token string) bool

	// Sampler parameters
	SamplerParams SamplerParams

	// EnablePrefixCaching reuses KV cache for matching prompt prefix
	EnablePrefixCaching bool

	// AbortContext cancels generation when done
	AbortContext context.Context
}

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

type ContextInfo struct {
	NCtx     uint32 // Context size
	NBatch   uint32 // Batch size
	NUBatch  uint32 // Micro-batch size
	NSeqMax  uint32 // Max sequences
	NThreads int32  // Thread count
}

type ContextParams struct {
	NCtx          uint32        // Context size (0 = from model)
	NBatch        uint32        // Logical maximum batch size
	NUBatch       uint32        // Physical maximum batch size
	NSeqMax       uint32        // Max number of sequences
	NThreads      int32         // Threads for generation
	NThreadsBatch int32         // Threads for batch processing
	RopeFreqBase  float32       // RoPE base frequency (0 = from model)
	RopeFreqScale float32       // RoPE frequency scaling (0 = from model)
	TypeK         GGMLType      // KV cache K type (-1 = default F16)
	TypeV         GGMLType      // KV cache V type (-1 = default F16)
	AttentionType AttentionType // Attention type for embeddings (-1 = auto)
	FlashAttn     FlashAttnType // Flash attention mode (-1 = auto, 0 = disabled, 1 = enabled)
	Embeddings    bool          // Extract embeddings
	OffloadKQV    bool          // Offload KQV ops to GPU
	KVUnified     bool          // Use unified KV cache (required for encoder/BERT models)
	NoPerf        bool          // Disable performance timings
}

type DetokenizeOptions struct {
	RemoveSpecial  bool // Remove BOS/EOS tokens from output
	UnparseSpecial bool // Render special tokens as text
}

type EmbeddingOptions struct {
	// Normalize whether to L2-normalize embeddings (recommended for cosine similarity)
	Normalize bool
	// AddBOS whether to add BOS token at the start
	AddBOS bool
	// AddEOS whether to add EOS token at the end
	AddEOS bool
}

type FlashAttnType int32

type GGMLType int32

type GPUInfo struct {
	DeviceID         int32
	DeviceName       string
	FreeMemoryBytes  int64 // -1 if unknown
	TotalMemoryBytes int64 // -1 if unknown
}

type GrammarSampler struct {
	Sampler
}

type LazyGrammarOptions struct {
	// GrammarStr is the GBNF grammar string
	GrammarStr string

	// GrammarRoot is the start symbol name (default: "root")
	GrammarRoot string

	// TriggerPatterns are regex patterns that activate the grammar
	// Pattern matches from start of generation, grammar gets content from first match group
	TriggerPatterns []string

	// TriggerTokens are token IDs that activate the grammar
	// Grammar gets content starting from the trigger token (included)
	TriggerTokens []Token
}

type LazyGrammarSampler struct {
	Sampler
}

type LogCallback func(level LogLevel, message string)

type LogLevel int

type MirostatV2Params struct {
	Seed uint32  // Random seed
	Tau  float32 // Target cross-entropy (higher = more random)
	Eta  float32 // Learning rate for mu updates
}

type MirostatV2Sampler struct {
	Sampler
}

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

type ModelInfo struct {
	NLayer    int32  // Total number of layers
	NHead     int32  // Number of attention heads
	NHeadKV   int32  // Number of KV heads (for GQA/MQA)
	NEmbd     int32  // Embedding dimension
	NCtxTrain int32  // Training context length
	NParams   uint64 // Total parameter count
	ModelSize uint64 // Model size in bytes
}

type ModelParams struct {
	NGPULayers int32 // Number of layers to offload to GPU (-1 = all)
	MainGPU    int32 // Main GPU index
	UseMmap    bool  // Use memory mapping for model loading
	UseMlock   bool  // Lock model in memory
}

type PerfData struct {
	StartMs     float64 // Absolute start time (ms)
	LoadMs      float64 // Model loading time (ms)
	PromptMs    float64 // Prompt processing time (ms)
	GenerateMs  float64 // Token generation time (ms)
	PromptCount int32   // Prompt tokens processed
	TokenCount  int32   // Tokens generated
}

type PoolingType int32

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

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

type SamplerParams struct {
	// Seed for random sampling (0 = random seed)
	Seed uint32

	// Temperature controls randomness (1.0 = normal, <1.0 = more deterministic, >1.0 = more random)
	// Set to 0 for greedy sampling
	Temperature float32

	// TopK limits sampling to top K tokens (0 = disabled)
	TopK int32

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

	// MinP filters tokens with probability < P * max_prob (0.0 = disabled)
	MinP float32

	// RepeatPenalty penalizes repeated tokens (1.0 = disabled)
	RepeatPenalty float32

	// RepeatLastN is the number of tokens to consider for repetition penalty
	RepeatLastN int32

	// FrequencyPenalty reduces probability of frequent tokens (0.0 = disabled)
	FrequencyPenalty float32

	// PresencePenalty reduces probability of tokens that appeared at all (0.0 = disabled)
	PresencePenalty float32
}

type Token = int32

type TokenizeOptions struct {
	AddSpecial   bool // Add BOS/EOS tokens
	ParseSpecial bool // Parse special tokens in text (like <|endoftext|>)
}

type XTCParams struct {
	Probability float32 // Probability threshold (0.0-1.0)
	Threshold   float32 // Logit threshold
	MinKeep     int     // Minimum tokens to keep
	Seed        uint32  // Random seed
}

type XTCSampler struct {
	Sampler
}