Minimax H3 Engine

class MiniMaxH3Engine

One MiniMax-H3 engine, its runtime config, and the context that runs it.

The context memory and the weights the engine binds from the checkpoint are claimed here, once.

Public Functions

MiniMaxH3Engine(

std::string const &artifactDir,

std::string const &component,

cudaStream_t stream

)

Load <artifactDir>/<component>.engine, its config.json, and its checkpoint weights.

inline nlohmann::json const &config() const noexcept

void bind(char const *name, rt::Tensor &tensor)

Bind a tensor address, and its shape when it is an input.

void bind(char const *name, void *data, rt::Coords const &inputShape)

void enqueue(cudaStream_t stream)