[question] how dynamo or aot-autograd optimize slice operation?

compiler
1

For the following case, the fx graph obtained by the inductor entry no longer contains the slice op. Does anyone know where this part is optimized?

e.g. slice ut case

import torch

class NaiveModel(torch.nn.Module):
    def __init__(self):
        super().__init__()
    def forward(self, arg0_1):
        slice_1 = torch.ops.aten.slice.Tensor(arg0_1, 0, 0, 9223372036854775807)
        return slice_1

model = NaiveModel()
opt_mod = torch.compile(model, backend="inductor")

x0 = torch.rand(256, 256)

golden = model(x0)
outs = opt_mod(x0)

stage1. pdb in inductor_compile_fx_inner

image
image1486×886 267 KB

stage2. print(gm.code)
image

2

It seems torch/_inductor/fx_passes/joint_graph.py::joint_graph_passes deal with the optimize. :face_with_monocle: