Hello

Running an inferencing script on x86 with the following  stack works ok
# conda list | grep torch
intel-extension-for-pytorch 2.1.100                  pypi_0    pypi
torch                     2.3.0.dev20231219+cpu          pypi_0    pypi
torchaudio                2.2.0.dev20231219+cpu          pypi_0    pypi
torchvision               0.18.0.dev20231219+cpu          pypi_0    pypi


However on P10 with the following conda environment

   conda list | grep torch
_pytorch_select           1.0                       cpu_2    rocketce
pytorch-base              2.0.1           cpu_py310_pb4.21.12_1    rocketce
pytorch-cpu               2.0.1                   py310_1    rocketce
torchdata                 0.6.0                   py310_2    rocketce
torchtext-base            0.15.2              cpu_py310_1    rocketce
torchtext-cpu             0.15.2                  py310_1    rocketce
torchvision-base          0.15.2              cpu_py310_1    rocketce
torchvision-cpu           0.15.2                  py310_1    rocketce

and  running the same  inferencing script, getting the following error

    torch.torch._inductor.cudagraph_mark_step_begin()
AttributeError: module 'torch._inductor' has no attribute 'cudagraph_mark_step_begin'

------------------------------
Mel Bakhshi
Systems Engineer
Rocket Forum Shared Account
------------------------------

Hi Mel Bakhshi,

Thanks for reaching out to us.

We will try to reproduce and resolve this issue in our environment.

Could you please provide us following details from both x86 and p10 env?

conda -V

"conda config --show" (double - before show)

conda list

If possible, could you please also share the code snippet where this issue occurs?

------------------------------
Suyog Jadhav
Rocket Internal - All Brands
------------------------------

Original Message Original Message:
Sent: 01-15-2024 13:14
From: Mel Bakhshi
Subject: AttributeError on IBM Power 10 system

Hello

Running an inferencing script on x86 with the following  stack works ok
# conda list | grep torch
intel-extension-for-pytorch 2.1.100                  pypi_0    pypi
torch                     2.3.0.dev20231219+cpu          pypi_0    pypi
torchaudio                2.2.0.dev20231219+cpu          pypi_0    pypi
torchvision               0.18.0.dev20231219+cpu          pypi_0    pypi


However on P10 with the following conda environment

   conda list | grep torch_pytorch_select           1.0                       cpu_2    rocketcepytorch-base              2.0.1           cpu_py310_pb4.21.12_1    rocketcepytorch-cpu               2.0.1                   py310_1    rocketcetorchdata                 0.6.0                   py310_2    rocketcetorchtext-base            0.15.2              cpu_py310_1    rocketcetorchtext-cpu             0.15.2                  py310_1    rocketcetorchvision-base          0.15.2              cpu_py310_1    rocketcetorchvision-cpu           0.15.2                  py310_1    rocketce

and  running the same  inferencing script, getting the following error

    torch.torch._inductor.cudagraph_mark_step_begin()
AttributeError: module 'torch._inductor' has no attribute 'cudagraph_mark_step_begin'

------------------------------
Mel Bakhshi
Systems Engineer
Rocket Forum Shared Account
------------------------------