DeepSpeed

dac60fad - fix: custom-AR latency path also serves no-grad decode of training engines

fix: custom-AR latency path also serves no-grad decode of training engines is_inference_mode stays false under rollout engines that configure a train batch; the decode forward is still grad-free and mathematically identical through the one-shot kernel. Signed-off-by: Guokai Ma <guokai.ma@intel.com>