TPU initialization failed: open(/dev/accel2): Operation not permitted

[JaySDarji]

❓ Questions and Help

On a fresh Google Cloud TPU VM (created today), I cannot get torch-xla to recognize the TPU.

I tried to delete and recreate the tpu vm many times and even in all TPU VMs these same errors came for that:

scd@t1v-n-1219602b-w-0:~$ PJRT_DEVICE=TPU python3
Python 3.10.12 (main, Aug 15 2025, 14:32:43) [GCC 11.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch_xla.core.xla_model as xm
>>> print(xm.get_xla_supported_devices(\"TPU\"))
  File "<stdin>", line 1
    print(xm.get_xla_supported_devices(\"TPU\"))
                                        ^
SyntaxError: unexpected character after line continuation character
>>> print(xm.get_xla_supported_devices("TPU\"))
  File "<stdin>", line 1
    print(xm.get_xla_supported_devices("TPU\"))
                                       ^
SyntaxError: unterminated string literal (detected at line 1)
>>> print(xm.get_xla_supported_devices())
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/scd/.local/lib/python3.10/site-packages/torch_xla/core/xla_model.py", line 79, in get_xla_supported_devices
    devices = torch_xla._XLAC._xla_get_devices()
RuntimeError: Bad StatusOr access: UNKNOWN: TPU initialization failed: open(/dev/accel2): Operation not permitted: Operation not permitted; Couldn't open device: /dev/accel2; [/dev/accel2] 
>>> import torch
>>> import torch_xla.core.xla_model as xm
>>> 
>>> dev = xm.xla_device()
<stdin>:1: DeprecationWarning: Use torch_xla.device instead
F0828 01:09:54.071199  532138 runtime.cpp:21] Check failed: !g_computation_client_initialized InitializeComputationClient() can only be called once.
*** Check fail

Steps tried:

• Checked /dev/accel* → exists but owned by root:root:
  Then i tried (but failed):

• Tried adding my user to tpu group, but group does not exist.

• Attempted to install libtpu-nightly, but package not found:

• TPU training script fails with RuntimeError: tensorflow/compiler/xla/pjrt/pjrt_client.cc: Could not initialize PJRT client: Not found: Unable to access /dev/accel0.

Environment

• TPU VM: v4-8 (freshly created Aug 21, 2025)
• OS: Ubuntu 22.04 (default TPU VM image)
• torch &torch-xla: by using commands from this reference (https://cloud.google.com/tpu/docs/run-calculation-pytorch)

What I’ve tried

• Updating apt sources
• Installing libtpu-nightly (fails)
• Checking device permissions
• Following official setup docs

Request

Please help me with torch xla on my tpu device on google cloud

[ysiraichi]

Thank you for filing this issue.
Could you the PyTorch/XLA version you are using by printing torch_xla.__version__?
Also, could you try the nightly release?