cannot compile the depthwise conv when it is included in other models

[wcshds]

cannot compile the depthwise conv when it is included in other models. But it can be compiled normally when used alone.

Here is the minimal reproducible example:

from torch import nn
import torch
import intel_npu_acceleration_library


model = nn.Sequential(
    nn.Conv2d(
        in_channels=32,
        out_channels=32,
        groups=32,
        kernel_size=3,
        padding=1
    )
)
model = intel_npu_acceleration_library.compile(model)
x = torch.rand([1, 32, 48, 200])
output = model(x)
print(output.shape)

[alessandropalla]

Yes it is not supported yet.

[alessandropalla]

Implemented in #61