Safety FMHA Launch#

struct Artifact#

Public Members

unsigned char const *cubin#
size_t bytes#
char const *symbol#
unsigned int sharedMemory#
int headSize#
bool visual#
int tileM#
int threads#
struct PreparedBindings#

One immutable binding set per plugin instance, populated during Initialization warmup.

Public Members

CUtensorMap maps[4] = {}#
std::array<void const*, 4> pointers = {}#
std::array<int32_t, 7> dimensions = {}#
bool ready = {}#
struct ContiguousArguments#

Public Members

void const *validity#
void const *q#
void const *k#
void const *v#
void *output#
int32_t const *cuQ#
int32_t const *cuK#
int32_t queryLength#
int32_t qHeads#
int32_t kvHeads#
int32_t capacity#
float scale#
struct VisualArguments#

Public Members

void const *validity#
void const *q#
void const *k#
void const *v#
void *output#
int32_t const *cuSeqLens#
int32_t totalTokens#
int32_t qHeads#
int32_t kvHeads#
int32_t sequences#
int32_t maxSeqLength#
float scale#