DeepSpeed

2c3c3a04 - perf(tp): vocab-parallel lm_head decode — half weight traffic per rank (Loop 8)

perf(tp): vocab-parallel lm_head decode — half weight traffic per rank (Loop 8) Each rank GEMMs its contiguous half of the replicated lm_head weight (saves ~630MB/step/rank of HBM traffic at bs=32) and the rank-ordered all-gather reassembles full logits — argmax and sampling paths see identical semantics downstream. Decode path only; prefill unchanged. Signed-off-by: Guokai Ma <guokai.ma@intel.com>