Nvfp4 Moe Types#

struct MoELayout#

CPU-side layout result. Populated by the test-only CPU reference builder.

Public Members

rt::Tensor tileIdxToGroupIdx#
rt::Tensor tileIdxToMnLimit#
rt::Tensor permutedIdxToExpandedIdx#
rt::Tensor numNonExitingTiles#
int32_t numTiles = {}#
int32_t mPadded = {}#
std::vector<int32_t> tileIdxToGroupIdxHost#
std::vector<int32_t> tileIdxToMnLimitHost#
struct MoELayoutBuffers#

Pre-allocated GPU buffers for the layout builder kernel. maxTiles = tileIdxToGroupIdx.getShape()[0], maxMPadded = permutedIdxToExpandedIdx.getShape()[0].

Public Members

rt::Tensor tileIdxToGroupIdx#
rt::Tensor tileIdxToMnLimit#
rt::Tensor permutedIdxToExpandedIdx#
rt::Tensor numNonExitingTiles#