DeepSpeed

deafce7b - perf(tp): reduce-scatter + all-gather in inference RowParallel

perf(tp): reduce-scatter + all-gather in inference RowParallel Replace the single all-reduce in RowParallel's inference path with reduce-scatter followed by all-gather. Same result, but each collective moves N/TP bytes per rank instead of N — smaller messages for the latency-bound decode step and the architectural hook for overlapping the all-gather with chunked sequence-parallel compute (residual/norm on 1/TP of the data) in a follow-up. Signed-off-by: Guokai Ma <guokai.ma@intel.com>