Why adaptive_avg_pool1d became adaptive_avg_pool2d in make_fx?

torch.compile FX (Functional Transformations)
1

I need to use the make_fx interface to output aten ops and convert them into corresponding torch diaclects. But I found that make_fx will replace some ops. How does this happen and how should I solve it?

import torch
import os
from torch.fx.experimental.proxy_tensor import make_fx
print("os:pid:", os.getpid())
class Aten_Adaptive_avg_pool1d(torch.nn.Module):
    def __init__(self, output_size):
        super().__init__()
        self.output_size=output_size
    def forward(self, self_):
        return torch.ops.aten.adaptive_avg_pool1d(self_, output_size=self.output_size)

if __name__ == "__main__":
    model = Aten_Adaptive_avg_pool1d(3)
    model.train(False)
    model = model.eval()
    #"adaptive_avg_pool1d": {"inputs": 1, "outputs": 1, "attrs": {"output_size": [3]}, "input_shapes": ["1,32,128"]},
    input = torch.randn(1,32,128)
    traced_model = make_fx(model)(input)
    print(traced_model.graph)
    pass

and It print out:

graph():
    %arg0_1 : [num_users=1] = placeholder[target=arg0_1]
    %unsqueeze : [num_users=1] = call_function[target=torch.ops.aten.unsqueeze.default](args = (%arg0_1, -2), kwargs = {})
    %_adaptive_avg_pool2d : [num_users=1] = call_function[target=torch.ops.aten._adaptive_avg_pool2d.default](args = (%unsqueeze, [1, 3]), kwargs = {})
    %squeeze : [num_users=1] = call_function[target=torch.ops.aten.squeeze.dim](args = (%_adaptive_avg_pool2d, -2), kwargs = {})
    return squeeze

I need some help, thanks~

2

I find adaptive_avg_pool1d expand it to adaptive_avg_pool2d.in cpp library .
How does make_fx trace back to function calls in the cpp library?