Int4 Moe Scale Repack#
- cudaError_t trt_edgellm::kernel::launchInt4MoeScaleRepackSourceBatch(
- uint16_t const *const *firstSources,
- uint16_t const *const *secondSources,
- uint16_t *marlinOutput,
- int32_t count,
- int32_t G,
- int32_t projectionN,
- cudaStream_t stream
Repack non-contiguous expert scales directly from CUDA-mapped checkpoint storage.
secondSourcesis null for a single projection.