[Bug] Incorrect traced graph for in-place ops in current master #29008
Description
🐛 Bug
In PyTorch master build (275adb1), tracing the below snippet produces empty traced graph.
import torch
class IndexPutModel(torch.nn.Module):
def forward(self, x, update):
x[..., torch.tensor([2, 1, 3]), 2:4] += update
return x
x = torch.randn(3, 4, 5, 6, 7)
update = torch.randn(3, 1, 1, 3, 2)
trace_graph, torch_out, inputs_states = torch.jit.get_trace_graph(IndexPutModel(), (x, update), _force_outplace=True, _return_inputs_states=True)
print(trace_graph)traced graph:
graph(%0 : Float(3, 4, 5, 6, 7),
%1 : Float(3, 1, 1, 3, 2)):
return (%0)
The traced graph was correct in PyTorch release 1.3.
graph(%0 : Float(3, 4, 5, 6, 7),
%1 : Float(3, 1, 1, 3, 2)):
%2 : Float(3, 4, 5, 6, 7) = aten::clone(%0) # /home/work/anaconda3/envs/torch/lib/python3.7/site-packages/torch/jit/__init__.py:338:0
%3 : Float(3, 1, 1, 3, 2) = aten::clone(%1) # /home/work/anaconda3/envs/torch/lib/python3.7/site-packages/torch/jit/__init__.py:338:0
%4 : Long(3) = prim::Constant[value= 2 1 3 [ Variable[CPULongType]{3} ]](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%5 : Device = prim::Constant[value="cpu"](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%6 : int = prim::Constant[value=4](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%7 : bool = prim::Constant[value=0](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%8 : bool = prim::Constant[value=0](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%9 : Long(3) = aten::to(%4, %5, %6, %7, %8), scope: IndexPutModel # repro_trace_inplace.py:5:0
%10 : Long(3) = aten::detach(%9), scope: IndexPutModel # repro_trace_inplace.py:5:0
%11 : int = prim::Constant[value=4](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%12 : int = prim::Constant[value=2](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%13 : int = prim::Constant[value=4](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%14 : int = prim::Constant[value=1](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%15 : Float(3, 4, 5, 6, 2) = aten::slice(%0, %11, %12, %13, %14), scope: IndexPutModel # repro_trace_inplace.py:5:0
%16 : int = prim::Constant[value=4](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%17 : int = prim::Constant[value=0](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%18 : Device = prim::Constant[value="cpu"](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%19 : bool = prim::Constant[value=0](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%20 : bool = prim::Constant[value=0](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%21 : bool = prim::Constant[value=0](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%22 : Long(3) = aten::to(%10, %16, %17, %18, %19, %20, %21), scope: IndexPutModel # repro_trace_inplace.py:5:0
%23 : None = prim::Constant(), scope: IndexPutModel
%24 : None = prim::Constant(), scope: IndexPutModel
%25 : None = prim::Constant(), scope: IndexPutModel
%26 : Tensor?[] = prim::ListConstruct(%23, %24, %25, %22), scope: IndexPutModel
%27 : Float(3, 4, 5, 3, 2) = aten::index(%15, %26), scope: IndexPutModel # repro_trace_inplace.py:5:0
%28 : int = prim::Constant[value=1](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%29 : Float(3, 4, 5, 3, 2) = aten::add(%27, %1, %28), scope: IndexPutModel # repro_trace_inplace.py:5:0
%30 : int = prim::Constant[value=4](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%31 : int = prim::Constant[value=2](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%32 : int = prim::Constant[value=4](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%33 : int = prim::Constant[value=1](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%34 : Float(3, 4, 5, 6, 2) = aten::slice(%0, %30, %31, %32, %33), scope: IndexPutModel # repro_trace_inplace.py:5:0
%35 : int = prim::Constant[value=4](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%36 : int = prim::Constant[value=0](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%37 : Device = prim::Constant[value="cpu"](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%38 : bool = prim::Constant[value=0](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%39 : bool = prim::Constant[value=0](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%40 : bool = prim::Constant[value=0](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%41 : Long(3) = aten::to(%22, %35, %36, %37, %38, %39, %40), scope: IndexPutModel # repro_trace_inplace.py:5:0
%42 : None = prim::Constant(), scope: IndexPutModel
%43 : None = prim::Constant(), scope: IndexPutModel
%44 : None = prim::Constant(), scope: IndexPutModel
%45 : Tensor?[] = prim::ListConstruct(%42, %43, %44, %41), scope: IndexPutModel
%46 : bool = prim::Constant[value=0](), scope: IndexPutModel # repro_trace_inplace.py:5:0
%47 : Float(3, 4, 5, 6, 2) = aten::index_put(%34, %45, %29, %46), scope: IndexPutModel # repro_trace_inplace.py:5:0
return (%0)