hasSideEffects INTERNAL ASSERT FAILED when using .split method with JIT

[serhii-havrylov]

When using the code snippet bellow, I get this error: RuntimeError: kind_.is_prim() INTERNAL ASSERT FAILED at /pytorch/torch/csrc/jit/ir.cpp:904, please report a bug to PyTorch. Only prim ops are allowed to not have a registered operator but aten::cat doesn't have one either. We don't know if this op has side effects. (hasSideEffects at /pytorch/torch/csrc/jit/ir.cpp:904)

import torch
import numpy as np


@torch.jit.script
def get_loss(a, idxs):
    idxs = idxs.split(split_size=1, dim=1)
    b = a.index(idxs)
    return b.sum()


c = torch.zeros(3, 4, 5)
c.requires_grad = True
idxs = torch.tensor(np.array([[0, 1, 2],
                              [0, 0, 4],
                              [0, 2, 4],
                              [2, 3, 0]], dtype=np.long))

loss = get_loss(c, idxs)

If idxs.split(split_size=1, dim=1) replaced with list(idxs.split(split_size=1, dim=1)) the error disappears.
PyTorch version is 1.2.0

cc @ezyang @gchanan @suo

[suo]

@smessmer any progress on this? It repros for me

[eellison]

The repro no longer errors, and there is another issue similar issue that is outstanding so I'm closing this.

[WGY907]

I get this error ,RuntimeError: kind_.is_prim() INTERNAL ASSERT FAILED at /pytorch/torch/csrc/jit/ir.cpp:904, please report a bug to PyTorch. Only prim ops are allowed to not have a registered operator but aten::_convolution doesn't have one either. We don't know if this op has side effects. (hasSideEffects at /pytorch/torch/csrc/jit/ir.cpp:904)
How to deal with it
thanks a lot

[kjdorow7]

@WGY907, it looks like this problem goes away with newer PyTorch versions. I ran the above code with 1.5.1 and had no problems.

[n-gao]

@eellison This bug is not fixed! If you call get_loss a second time, it fails!

import torch
import numpy as np


@torch.jit.script
def get_loss(a, idxs):
    idxs = idxs.split(split_size=1, dim=1)
    b = a.index(idxs)
    return b.sum()


c = torch.zeros(3, 4, 5)
c.requires_grad = True
idxs = torch.tensor(np.array([[0, 1, 2],
                              [0, 0, 4],
                              [0, 2, 4],
                              [2, 3, 0]], dtype=np.long))

loss = get_loss(c, idxs)
loss = get_loss(c, idxs)
---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-9-f6dfaf23fa61> in <module>
----> 1 loss = get_loss(c, idxs)

RuntimeError: kind_.is_prim() INTERNAL ASSERT FAILED at "/pytorch/torch/csrc/jit/ir/ir.cpp":1098, please report a bug to PyTorch. Only prim ops are allowed to not have a registered operator but aten::cat doesn't have one either. We don't know if this op has side effects.

Reproduced in PyTroch 1.8

[MohamedAliRashad]

Is there any updates on this issue ?

[the5solae]

Any updates on this issue?