Data Structures
class	Context

struct	GGMAConfig

struct	KVCache

struct	ModelConfig

class	SentencePieceTokenizer

class	Tokenizer

class	TokenizerFactory

Enumerations
enum class	KVCacheDataType { FLOAT32 , UINT8 }

Functions
template<typename T >
void	load_config_field (const Json::Value &root, const std::string &field_name, T &target, bool is_optional=false)

template<>
void	load_config_field< bool > (const Json::Value &root, const std::string &field_name, bool &target, bool is_optional)

template<>
void	load_config_field< std::optional< int > > (const Json::Value &root, const std::string &field_name, std::optional< int > &target, bool is_optional)

bool	validate_model_config (const ModelConfig &config)

std::string	to_string (const ModelConfig &config)

nnfw_session *	create_and_prepare_session (const std::string &model_path)

uint64_t	num_elems (const nnfw_tensorinfo *tensor_info)

uint64_t	bufsize_for (const nnfw_tensorinfo *ti)

template void	Context::decode_impl< false, std::vector< uint8_t > > (ggma_token token_id, std::vector< uint8_t > &output)

template void	Context::decode_impl< true, std::vector< float > > (ggma_token token_id, std::vector< float > &output)

const char *	to_string (KVCacheDataType type)

KVCacheDataType	from_string (const std::string &type_str)

bool	is_supported_type (KVCacheDataType type)

Enumeration Type Documentation

◆ KVCacheDataType

enum class ggma::KVCacheDataType

strong

Enumerator
FLOAT32
UINT8

Definition at line 36 of file KVCache.h.

{
  FLOAT32,
  UINT8
};

Function Documentation

◆ bufsize_for()

uint64_t ggma::bufsize_for ( const nnfw_tensorinfo * ti )

Definition at line 63 of file Context.cc.

{
  static int elmsize[] = {
    sizeof(float),   /* NNFW_TYPE_TENSOR_FLOAT32 */
    sizeof(int),     /* NNFW_TYPE_TENSOR_INT32 */
    sizeof(uint8_t), /* NNFW_TYPE_TENSOR_QUANT8_ASYMM */
    sizeof(bool),    /* NNFW_TYPE_TENSOR_BOOL = 3 */
    sizeof(uint8_t), /* NNFW_TYPE_TENSOR_UINT8 = 4 */
    sizeof(int64_t), /* NNFW_TYPE_TENSOR_INT64 = 5 */
    sizeof(int8_t),  /* NNFW_TYPE_TENSOR_QUANT8_ASYMM_SIGNED = 6 */
    sizeof(int16_t), /* NNFW_TYPE_TENSOR_QUANT16_SYMM_SIGNED = 7 */
  };
  return elmsize[ti->dtype] * num_elems(ti);
}

References nnfw_tensorinfo::dtype, and num_elems().

Referenced by ggma::Context::prefill(), and ggma::Context::unemb().

◆ Context::decode_impl< false, std::vector< uint8_t > >()

template void ggma::Context::decode_impl< false, std::vector< uint8_t > >	(	ggma_token	token_id,
		std::vector< uint8_t > &	output
	)

◆ Context::decode_impl< true, std::vector< float > >()

template void ggma::Context::decode_impl< true, std::vector< float > >	(	ggma_token	token_id,
		std::vector< float > &	output
	)

◆ create_and_prepare_session()

nnfw_session * ggma::create_and_prepare_session ( const std::string & model_path )

Definition at line 45 of file Context.cc.

{
  nnfw_session *session = nullptr;
  NNFW_ENSURE_STATUS(nnfw_create_session(&session));
  NNFW_ENSURE_STATUS(nnfw_load_model_from_file(session, model_path.c_str()));
  NNFW_ENSURE_STATUS(nnfw_prepare(session));
  return session;
}

References nnfw_create_session(), NNFW_ENSURE_STATUS, nnfw_load_model_from_file(), and nnfw_prepare().

Referenced by ggma::Context::prefill(), and ggma::Context::unemb().

◆ from_string()

KVCacheDataType ggma::from_string ( const std::string & type_str )

Definition at line 42 of file KVCache.cc.

{
  if (type_str == "FLOAT32" || type_str == "float32")
  {
    return KVCacheDataType::FLOAT32;
  }
  else if (type_str == "UINT8" || type_str == "uint8")
  {
    return KVCacheDataType::UINT8;
  }
  throw std::runtime_error("Unsupported KV cache data type: " + type_str);
}

References FLOAT32, and UINT8.

◆ is_supported_type()

bool ggma::is_supported_type ( KVCacheDataType type )

Definition at line 56 of file KVCache.cc.

{
  switch (type)
  {
    case KVCacheDataType::FLOAT32:
    case KVCacheDataType::UINT8:
      return true;
    default:
      return false;
  }
}

References FLOAT32, type, and UINT8.

◆ load_config_field()

template<typename T >

void ggma::load_config_field	(	const Json::Value &	root,
		const std::string &	field_name,
		T &	target,
		bool	is_optional = `false`
	)

Definition at line 28 of file Config.cc.

{
  if (root.isMember(field_name))
    target = root[field_name].asInt();
  else if (!is_optional)
    throw std::runtime_error(field_name + " not found in config.json");
}

Referenced by ggma::ModelConfig::load_from_json().

◆ load_config_field< bool >()

template<>

void ggma::load_config_field< bool >	(	const Json::Value &	root,
		const std::string &	field_name,
		bool &	target,
		bool	is_optional
	)

Definition at line 39 of file Config.cc.

{
  if (root.isMember(field_name))
    target = root[field_name].asBool();
  else if (!is_optional)
    throw std::runtime_error(field_name + " not found in config.json");
}

◆ load_config_field< std::optional< int > >()

template<>

void ggma::load_config_field< std::optional< int > >	(	const Json::Value &	root,
		const std::string &	field_name,
		std::optional< int > &	target,
		bool	is_optional
	)

Definition at line 50 of file Config.cc.

{
  if (root.isMember(field_name))
    target = root[field_name].asInt();
}

◆ num_elems()

uint64_t ggma::num_elems ( const nnfw_tensorinfo * tensor_info )

Definition at line 55 of file Context.cc.

{
  uint64_t n = 1;
  for (int32_t i = 0; i < tensor_info->rank; ++i)
    n *= tensor_info->dims[i];
  return n;
}

References nnfw_tensorinfo::dims, and nnfw_tensorinfo::rank.

Referenced by bufsize_for(), and ggma::Context::unemb().

◆ to_string() [1/2]

std::string ggma::to_string ( const ModelConfig & config )

Definition at line 144 of file Config.cc.

144{ return config.to_string(); }

◆ to_string() [2/2]

const char * ggma::to_string ( KVCacheDataType type )

Definition at line 28 of file KVCache.cc.

{
  switch (type)
  {
    case KVCacheDataType::FLOAT32:
      return "FLOAT32";
    case KVCacheDataType::UINT8:
      return "UINT8";
    default:
      return "UNKNOWN";
  }
}