Cute Dsl Nvfp4 Dual GEMM Runner

class CuteDslNvfp4DualGemmRunner

Public Functions

cudaError_t prepare(

int32_t &maxActiveClusters,

bool quantizedOutput = false,

int32_t tactic = kDEFAULT_TACTIC

) noexcept

Initialize outside capture. Returns the persistent worker count for this device.

cudaError_t run(

Nvfp4DualGemmParams const &params,

cudaStream_t stream

) const noexcept

Launch only; a missing prepare() is an initialization error.

Public Static Functions

static bool canImplement(int32_t smVersion) noexcept

Public Static Attributes

static constexpr int32_t kDEFAULT_TACTIC = 0

static constexpr int32_t kFIXED_2X1_TACTIC = 1

static constexpr int32_t kMIXED_2X2_TACTIC = 2

static constexpr int32_t kMIXED_4X1_TACTIC = 3

static constexpr int32_t kNUM_QUANTIZED_TACTICS = kMIXED_4X1_TACTIC

struct Nvfp4DualGemmParams

Public Members

void const *activation

E2M1 bytes [M,K/2], low nibble first.

void const *packedWeight

E2M1 [2N,K/2], alternating 128-row gate/up tiles.

void const *activationScale

E4M3 SF-atom layout for [M,K/16].

void const *packedWeightScale

E4M3 SF-atom layout for interleaved [2N,K/16].

void const *alpha

Two device FP32 scalars: gate, up; applied before SiLU.

void *output

FP16 [M,N], or E2M1 [M,N/2] when outputScale is set.

int32_t numTokens

int32_t interSize

int32_t hiddenSize

int32_t maxActiveClusters

void *outputScale = {}

Optional E4M3 SF atoms for [M,N/16], padded to 128 rows.

float const *globalScale = {}

Device forward scale for quantized output.

int32_t tactic = {}

0/1: fixed 2x1, 2: mixed 2x2/2x1, 3: mixed 4x1/2x1 (quantized output).