--reset

# Use inplace to reduce memory consumption, most eltwise implementations do not
# use different implementations for inplace. Use smallest supported data type to
# minimize memory usage as well.

--inplace=true

--dt=u8
--dir=FWD_D
--attr-post-ops=,add:f32:per_tensor
--batch=shapes_large_buffer

--reset
--dt=bf16
--dir=BWD_D
--batch=shapes_large_buffer
