localmodels

type AssetPuller interface {
	PullModel(ctx context.Context, runtimeCfg *factoryconfig.LoadedFactoryConfig, modelName string) (apisurface.ModelPullResult, error)
	EnsureModelAvailable(ctx context.Context, runtimeCfg *factoryconfig.LoadedFactoryConfig, worker *interfaces.WorkerConfig) error
	ResolveModelCache(ctx context.Context, runtimeCfg *factoryconfig.LoadedFactoryConfig, worker *interfaces.WorkerConfig) (CacheLayout, error)
}

type CacheLayout struct {
	ModelName string
	CachePath string
	Revision  string
	Files     []string
}

type CatalogEntry struct {
	Summary factoryapi.ModelSummary
	Detail  factoryapi.ModelDetail
}

type Handle interface {
	Invoke(context.Context, InvocationRequest) (interfaces.InferenceResponse, error)
}

type Hooks struct {
	MarkResourceWaitStarted  func(context.Context, time.Time)
	MarkResourceWaitFinished func(context.Context, time.Time, bool)
	MarkLoadRequested        func(context.Context, time.Time)
	MarkLoadFinished         func(context.Context, time.Time)
	MarkLoadReused           func(context.Context)
}

type InvocationRequest struct {
	Resource interfaces.ResourceConfig
	Worker   *interfaces.WorkerConfig
	Request  interfaces.RunnerExecutionRequest
}

type LoadRequest struct {
	Resource  interfaces.ResourceConfig
	Worker    *interfaces.WorkerConfig
	ModelName string
	CachePath string
	Revision  string
	Files     []string
}

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

type OmniVoiceInvocationPayload struct {
	Operation  string                                     `json:"operation"`
	ModelName  string                                     `json:"modelName"`
	Revision   string                                     `json:"revision,omitempty"`
	OutputFile string                                     `json:"outputFile"`
	Text       string                                     `json:"text"`
	Bindings   []interfaces.ResolvedModelOperationBinding `json:"bindings,omitempty"`
}

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

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

type Runtime interface {
	Supports(resource interfaces.ResourceConfig, worker *interfaces.WorkerConfig) bool
	Load(context.Context, LoadRequest) (Handle, error)
}