Build Layout

void trt_edgellm::kernel::buildLayoutGpu(

MoELayoutBuffers &buf,

int32_t const *tokenSelectedExperts,

int32_t numTokens,

int32_t topK,

int32_t localNumExperts,

int32_t tileSize,

cudaStream_t stream

)

GPU-side layout builder via single-CTA kernel (~3-5 us). All device pointers in buffers must be pre-allocated by the caller. tokenSelectedExperts must contain LOCAL expert indices in [0, L).