10 # list[list[T]] -> list[T]
20 (options) -> (inputs, params, forward, backward_setup, backward)
29 backward: Given `output = backward_setup(*forward(*inputs))`, performs
46 def simple_backward_setup(output, seed=None):  argument
47     assert isinstance(output, torch.Tensor)
50     grad_output = torch.randn_like(output)
51     return output, grad_output
54 def simple_backward(output, grad_output, **kwargs):  argument
55     return output.backward(grad_output, **kwargs)
69 def lstm_creator(script=True, **kwargs):  argument
75         forward=lstm_factory(lstm_cell, script),
81 def lnlstm_creator(script=True, decompose_layernorm=False, **kwargs):  argument
82     assert script is True
110 def dropoutlstm_creator(script=True, **kwargs):  argument
111     assert script is True
138 def lstm_premul_creator(script=True, **kwargs):  argument
144         forward=lstm_factory_premul(premul_lstm_cell, script),
150 def lstm_premul_bias_creator(script=True, **kwargs):  argument
156         forward=lstm_factory_premul_bias(premul_lstm_cell_no_bias, script),
162 def lstm_simple_creator(script=True, **kwargs):  argument
168         forward=lstm_factory_simple(flat_lstm_cell, script),
174 def lstm_multilayer_creator(script=True, **kwargs):  argument
180         forward=lstm_factory_multilayer(lstm_cell, script),
251         # XXX: It's more efficient to store the output in its padded form,
253         # Un-padding the output also makes the backward pass 2x slower...
266 def varlen_lstm_factory(cell, script):  argument
274     ) -> Tuple[List[Tensor], Tuple[List[Tensor], List[Tensor]]]:
278         # List of: (output, hx, cx)
284             output = []
292                 output += [hy]
293             outputs += [torch.stack(output)]
299     if script:
300         cell = torch.jit.script(cell)
301         dynamic_rnn = torch.jit.script(dynamic_rnn)
306 def varlen_lstm_creator(script=False, **kwargs):  argument
312         forward=varlen_lstm_factory(lstm_cell, script),
357 # output: packed_weights with format
370     # XXX: script fns have problems indexing multidim lists, so we try to
406 def lstm_factory(cell, script):  argument
414     ) -> Tuple[Tensor, Tuple[Tensor, Tensor]]:
424     if script:
425         cell = torch.jit.script(cell)
426         dynamic_rnn = torch.jit.script(dynamic_rnn)
432 def lstm_factory_premul(premul_cell, script):  argument
440     ) -> Tuple[Tensor, Tuple[Tensor, Tensor]]:
450     if script:
451         premul_cell = torch.jit.script(premul_cell)
452         dynamic_rnn = torch.jit.script(dynamic_rnn)
458 def lstm_factory_premul_bias(premul_cell, script):  argument
466     ) -> Tuple[Tensor, Tuple[Tensor, Tensor]]:
470 …# add bias for all timesteps instead of going step-by-step, results in a single reduction kernel i…
474         inputs = torch.mm(input.view(-1, inpSize[2]), wih.t()) + bih
475         inputs = inputs.view(inpSize[0], inpSize[1], -1).unbind(0)
482     if script:
483         premul_cell = torch.jit.script(premul_cell)
484         dynamic_rnn = torch.jit.script(dynamic_rnn)
491 def lstm_factory_simple(cell, script):  argument
500     if script:
501         cell = torch.jit.script(cell)
502         dynamic_rnn = torch.jit.script(dynamic_rnn)
507 def lstm_factory_multilayer(cell, script):  argument
510     ) -> Tuple[Tensor, Tuple[Tensor, Tensor]]:
529     if script:
530         cell = torch.jit.script(cell)
531         dynamic_rnn = torch.jit.script(dynamic_rnn)