[XPU] Kineto profiler fails on XPU with `PTI_ERROR_NOT_IMPLEMENTED`

[frost-intel]

🐛 Describe the bug

Since the 2025.1 oneAPI upgrade (#151899), any attempt to use the Pytorch Profiler fails with RuntimeError: Fail to enable Kineto Profiler on XPU due to error code: 2. The detailed error message is: PTI_ERROR_NOT_IMPLEMENTED.

To reproduce:

import torch
from torch.profiler import profile, ProfilerActivity

with profile(activities=[ProfilerActivity.XPU,], with_stack=True) as p:
    x = torch.ones(10).xpu()

Versions

PyTorch version: 2.8.0.dev20250515+xpu

cc @gujinghui @EikanWang @fengyuan14 @guangyey

[frost-intel]

@zejun-chen @chuanqi129

[zejun-chen]

Hi, @frost-intel

I run the reproducer you provide and print the profiling results as below. The profiler works fine and i am using the same oneAPI toolkit.

Which driver are you using for now? and Which OS are you running on?

[frost-intel]

I've reproduced this on two different machines:

Ubuntu 22.04, Driver Version 24.52.32224
SUSE Linux Enterprise Server 15, Driver Version 24.35.30872

Both are using clean conda environments with a pip-install of xpu/nightly

[zejun-chen]

Hi, @frost-intel

Could you provide the whole log when running the reproducer?

[frost-intel]

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
Cell In[1], line 3
      1 import torch
      2 from torch.profiler import profile
----> 3 with profile() as p:
      4     x = torch.ones(10).xpu()

File ~/miniforge3/envs/ptpy_nightly/lib/python3.10/site-packages/torch/profiler/profiler.py:810, in profile.__enter__(self)
    809 def __enter__(self):
--> 810     self.start()
    811     return self

File ~/miniforge3/envs/ptpy_nightly/lib/python3.10/site-packages/torch/profiler/profiler.py:820, in profile.start(self)
    819 def start(self):
--> 820     self._transit_action(ProfilerAction.NONE, self.current_action)
    821     if self.record_steps:
    822         self.step_rec_fn = prof.record_function(
    823             "ProfilerStep#" + str(self.step_num)
    824         )

File ~/miniforge3/envs/ptpy_nightly/lib/python3.10/site-packages/torch/profiler/profiler.py:876, in profile._transit_action(self, prev_action, current_action)
    874 if action_list:
    875     for action in action_list:
--> 876         action()

File ~/miniforge3/envs/ptpy_nightly/lib/python3.10/site-packages/torch/profiler/profiler.py:198, in _KinetoProfile.prepare_trace(self)
    184 if (self.profiler is None) or (not self.acc_events):
    185     self.profiler = prof.profile(
    186         use_cpu=(ProfilerActivity.CPU in self.activities),
    187         use_device=self.use_device,
   (...)
    196         custom_trace_id_callback=self.custom_trace_id_callback,
    197     )
--> 198 self.profiler._prepare_trace()

File ~/miniforge3/envs/ptpy_nightly/lib/python3.10/site-packages/torch/autograd/profiler.py:362, in profile._prepare_trace(self)
    360 self.entered = True
    361 t0 = perf_counter_ns()
--> 362 _prepare_profiler(self.config(create_trace_id=True), self.kineto_activities)
    363 t1 = perf_counter_ns()
    364 self._stats.profiler_prepare_call_duration_us = int((t1 - t0) / 1000)

RuntimeError: Fail to enable Kineto Profiler on XPU due to error code: 2. The detailed error message is: PTI_ERROR_NOT_IMPLEMENTED

[zejun-chen]

OK. Thank you. Let me verify and dive the issue.

[mschilling0]

Hi. I was able to reproduce this. The issue is our runtime package (and therefore our pip package) is missing the library where we implement PTI. Our development package has this library, but it is not bundled with pip.

So, libtorch_cpu.so links libpti_view.so.0 and that will try to dynamically load libpti.so. The package is missing libpti.so. We will figure out the best way to fix this, however, I will try to find a workaround in the mean time.

[mschilling0]

So, in the mean time, @frost-intel (or whoever is encountering this issue),

On Ubuntu, if you have the machine configured with oneAPI apt repository

After PyTorch is downloaded and installed via pip.

source venv/bin/activate
sudo apt install intel-pti-dev
source /opt/intel/oneapi/pti/latest/env/vars.sh
python reprod.py

/conda/venv/lib/python3.12/site-packages/torch/_subclasses/functional_tensor.py:279: UserWarning: Failed to initialize NumPy: No module named 'numpy' (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:81.)
  cpu = _conversion_method_template(device=torch.device("cpu"))
------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------
                                      Name    Self CPU %      Self CPU   CPU total %     CPU total  CPU time avg      Self XPU    Self XPU %     XPU total  XPU time avg    # of Calls
------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------
                          urUSMDeviceAlloc        22.79%     173.591us        22.79%     173.591us     173.591us       0.000us         0.00%       0.000us       0.000us             1
                        urEnqueueUSMMemcpy        77.21%     587.969us        77.21%     587.969us     587.969us       0.000us         0.00%       0.000us       0.000us             1
    Memcpy M2D (MEMORY(Unknown) -> DEVICE)         0.00%       0.000us         0.00%       0.000us       0.000us       6.800us       100.00%       6.800us       6.800us             1
------------------------------------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------  ------------
Self CPU time total: 761.560us
Self XPU time total: 6.800us

I imagine SLES is similar with dnf or yum.