Fp16 Moe Assemble#

cudaError_t trt_edgellm::kernel::launchFp16MoeSourceBatch(
uint8_t const *const *firstSources,
uint8_t const *const *secondSources,
int32_t count,
void *output,
int32_t rows,
int32_t columns,
Fp16MoeSourceType sourceType,
cudaStream_t stream
)#

Cast and stack checkpoint matrices directly into an FP16 MoE buffer. When secondSources is non-null, each source pair is emitted in the Fp16MoePlugin’s fixed 64-row up/gate interleave.