DeepSpeed

1b2bdbd3 - perf: cache static A_log/dt_bias fp32 casts — 48 launches off the decode step (Loop 6)

perf: cache static A_log/dt_bias fp32 casts — 48 launches off the decode step (Loop 6) The rank-sliced fp32 casts of A_log/dt_bias re-ran every decode step inside the graph; they are static across the rollout so cache once. Signed-off-by: Guokai Ma <guokai.ma@intel.com>