run recompute's real backward with amp disabled

Author: zhiqiu

python/paddle/distributed/fleet/utils/recompute.py

@@ -182,9 +182,10 @@ def backward(ctx, *args):
  "none of output has requires_grad=True, this recompute() is not necessary"
  )
 
- # actually backward 
- paddle.autograd.backward(forward_outputs_with_grad,
- backward_inputs_with_grad)
+ # actually backward
+ with paddle.amp.auto_cast(enable=False):
+ paddle.autograd.backward(forward_outputs_with_grad,
+ backward_inputs_with_grad)
 
  grads = list(inp._grad_ivar() for inp in detached_inputs
  if isinstance(inp, core.VarBase))