Nvfp4 A16 Blackwell Moe Jit Compiler

struct Nvfp4A16BlackwellMoeJitKey

Public Functions

inline auto asTuple() const noexcept

inline bool operator==(

Nvfp4A16BlackwellMoeJitKey const &other

) const noexcept

Public Members

int32_t sm = {nvfp4_a16_blackwell::kTargetSm}

uint32_t layout = {kNVFP4_A16_BLACKWELL_MOE_LAYOUT_ABI}

int32_t numExperts = {}

int32_t topK = {}

int32_t hiddenSize = {}

int32_t interSize = {}

int32_t interSizePadded = {}

int32_t fc1SplitK = {1}

int32_t fc2SplitK = {1}

int32_t fc2PrefetchSlots = {0}

Nvfp4A16BlackwellMoeDataType dataType = {Nvfp4A16BlackwellMoeDataType::kHALF}

uint32_t sourceAbi = {kNVFP4_A16_BLACKWELL_MOE_SOURCE_ABI}

struct Nvfp4A16BlackwellMoeJitDigest

Public Functions

inline bool operator==(

Nvfp4A16BlackwellMoeJitDigest const &other

) const noexcept

Public Members

uint64_t lo = {}

uint64_t hi = {}

struct Nvfp4A16BlackwellMoeJitKernel

Public Members

Nvfp4A16BlackwellMoeJitKey key

Nvfp4A16BlackwellMoeJitDigest digest

std::vector<uint8_t> cubin

int32_t trt_edgellm::getNvfp4A16BlackwellMoeFc1SharedBytes(

Nvfp4A16BlackwellMoeJitKey const &key

) noexcept

Static shared memory the baked FC1 / FC2 kernels declare (must stay within the 48 KB static limit; the key is rejected otherwise).

int32_t trt_edgellm::getNvfp4A16BlackwellMoeFc2SharedBytes(

Nvfp4A16BlackwellMoeJitKey const &key

) noexcept

char const *trt_edgellm::describeNvfp4A16BlackwellMoeJitKeyProblem(

Nvfp4A16BlackwellMoeJitKey const &key

) noexcept

Why a key cannot be compiled, or nullptr when it can.

bool trt_edgellm::canCompileNvfp4A16BlackwellMoeJitKernel(

Nvfp4A16BlackwellMoeJitKey const &key

) noexcept

Nvfp4A16BlackwellMoeJitDigest trt_edgellm::computeNvfp4A16BlackwellMoeJitDigest(

Nvfp4A16BlackwellMoeJitKey const &key,

void const *const cubinData,

size_t const cubinSize

)

Nvfp4A16BlackwellMoeJitKernel trt_edgellm::compileNvfp4A16BlackwellMoeJitKernel(

Nvfp4A16BlackwellMoeJitKey const &key

)

std::vector<uint8_t> trt_edgellm::serializeNvfp4A16BlackwellMoeJitKernel(

Nvfp4A16BlackwellMoeJitKernel const &kernel

)

Nvfp4A16BlackwellMoeJitKernel trt_edgellm::deserializeNvfp4A16BlackwellMoeJitKernel(

void const *const data,

size_t const size

)