The latest PyTorch XPU wheel 2.7.0.dev20250117+xpu does not work on Windows

[pbchekin]

Steps:

# This installs 2.7.0.dev20250117+xpu
pip install torch --index-url https://download.pytorch.org/whl/nightly/xpu
python -c 'import torch;print(torch.__version__)'

Result:

OSError: [WinError 126] The specified module could not be found. Error loading "C:\.venv\lib\site-packages\torch\lib\shm.dll" or one of its dependencies.

The last known wheel that worked is 2.7.0.dev20250110+xpu:

# This installs 2.7.0.dev20250110+xpu
pip install torch --index-url https://download.pytorch.org/whl/nightly/xpu
python -c 'import torch;print(torch.__version__)'

Result:

2.7.0.dev20250110+xpu

cc @seemethere @malfet @osalpekar @atalman @peterjc123 @mszhanyi @skyline75489 @nbcsm @iremyux @Blackhex @gujinghui @EikanWang @fengyuan14 @guangyey

[chuanqi129]

Hi @pbchekin , I have tried to install the latest nightly whl, but I can't reproduce this failure both in conda env and python venv

(venv) (nightly) c:\Users\chuanqiw\Documents>pip install torch --index-url https://download.pytorch.org/whl/nightly/xpu
Looking in indexes: https://download.pytorch.org/whl/nightly/xpu
Collecting torch
  Obtaining dependency information for torch from https://download.pytorch.org/whl/nightly/xpu/torch-2.7.0.dev20250119%2Bxpu-cp310-cp310-win_amd64.whl.metadata
  Downloading https://download.pytorch.org/whl/nightly/xpu/torch-2.7.0.dev20250119%2Bxpu-cp310-cp310-win_amd64.whl.metadata (27 kB)
Collecting umf==0.9.1
  Downloading https://download.pytorch.org/whl/nightly/xpu/umf-0.9.1-py2.py3-none-win_amd64.whl (72 kB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 72.2/72.2 kB 794.5 kB/s eta 0:00:00
Collecting jinja2
  Downloading https://download.pytorch.org/whl/nightly/jinja2-3.1.4-py3-none-any.whl (133 kB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 133.3/133.3 kB 7.7 MB/s eta 0:00:00
Collecting intel-cmplr-lib-rt==2025.0.2
  Downloading https://download.pytorch.org/whl/nightly/xpu/intel_cmplr_lib_rt-2025.0.2-py2.py3-none-win_amd64.whl (17.3 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 17.3/17.3 MB 7.8 MB/s eta 0:00:00
Collecting intel-cmplr-lic-rt==2025.0.2
  Downloading https://download.pytorch.org/whl/nightly/xpu/intel_cmplr_lic_rt-2025.0.2-py2.py3-none-win_amd64.whl (49 kB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 49.5/49.5 kB 64.4 kB/s eta 0:00:00
Collecting fsspec
  Downloading https://download.pytorch.org/whl/nightly/fsspec-2024.10.0-py3-none-any.whl (179 kB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 179.6/179.6 kB 11.3 MB/s eta 0:00:00
Collecting networkx
  Downloading https://download.pytorch.org/whl/nightly/networkx-3.4.2-py3-none-any.whl (1.7 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1.7/1.7 MB 5.0 MB/s eta 0:00:00
Collecting intel-pti==0.10.0
  Downloading https://download.pytorch.org/whl/nightly/xpu/intel_pti-0.10.0-py2.py3-none-win_amd64.whl (395 kB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 395.9/395.9 kB 1.8 MB/s eta 0:00:00
Collecting intel-cmplr-lib-ur==2025.0.2
  Downloading https://download.pytorch.org/whl/nightly/xpu/intel_cmplr_lib_ur-2025.0.2-py2.py3-none-win_amd64.whl (1.2 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1.2/1.2 MB 2.3 MB/s eta 0:00:00
Collecting sympy==1.13.1
  Downloading https://download.pytorch.org/whl/nightly/sympy-1.13.1-py3-none-any.whl (6.2 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 6.2/6.2 MB 5.9 MB/s eta 0:00:00
Collecting typing-extensions>=4.10.0
  Downloading https://download.pytorch.org/whl/nightly/typing_extensions-4.12.2-py3-none-any.whl (37 kB)
Collecting tcmlib==1.2.0
  Downloading https://download.pytorch.org/whl/nightly/xpu/tcmlib-1.2.0-py2.py3-none-win_amd64.whl (338 kB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 338.4/338.4 kB 20.5 MB/s eta 0:00:00
Collecting intel-sycl-rt==2025.0.2
  Downloading https://download.pytorch.org/whl/nightly/xpu/intel_sycl_rt-2025.0.2-py2.py3-none-win_amd64.whl (9.5 MB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 9.5/9.5 MB 5.7 MB/s eta 0:00:00
Collecting filelock
  Downloading https://download.pytorch.org/whl/nightly/filelock-3.16.1-py3-none-any.whl (16 kB)
Collecting mpmath<1.4,>=1.1.0
  Downloading https://download.pytorch.org/whl/nightly/mpmath-1.3.0-py3-none-any.whl (536 kB)
     ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 536.2/536.2 kB 2.4 MB/s eta 0:00:00
Collecting MarkupSafe>=2.0
  Downloading https://download.pytorch.org/whl/nightly/MarkupSafe-2.1.5-cp310-cp310-win_amd64.whl (17 kB)
Downloading https://download.pytorch.org/whl/nightly/xpu/torch-2.7.0.dev20250119%2Bxpu-cp310-cp310-win_amd64.whl (1209.0 MB)
   ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1.2/1.2 GB 1.7 MB/s eta 0:00:00
Using cached https://download.pytorch.org/whl/nightly/xpu/torch-2.7.0.dev20250119%2Bxpu-cp310-cp310-win_amd64.whl (1209.0 MB)
Installing collected packages: tcmlib, mpmath, intel-pti, intel-cmplr-lic-rt, intel-cmplr-lib-rt, umf, typing-extensions, sympy, networkx, MarkupSafe, fsspec, filelock, jinja2, intel-cmplr-lib-ur, intel-sycl-rt, torch
Successfully installed MarkupSafe-2.1.5 filelock-3.16.1 fsspec-2024.10.0 intel-cmplr-lib-rt-2025.0.2 intel-cmplr-lib-ur-2025.0.2 intel-cmplr-lic-rt-2025.0.2 intel-pti-0.10.0 intel-sycl-rt-2025.0.2 jinja2-3.1.4 mpmath-1.3.0 networkx-3.4.2 sympy-1.13.1 tcmlib-1.2.0 torch-2.7.0.dev20250119+xpu typing-extensions-4.12.2 umf-0.9.1

(venv) (nightly) c:\Users\chuanqiw\Documents>python -c "import torch; print(torch.__version__)"
c:\Users\chuanqiw\Documents\venv\lib\site-packages\torch\_subclasses\functional_tensor.py:275: UserWarning: Failed to initialize NumPy: No module named 'numpy' (Triggered internally at C:\actions-runner\_work\pytorch\pytorch\pytorch\torch\csrc\utils\tensor_numpy.cpp:81.)
  cpu = _conversion_method_template(device=torch.device("cpu"))
2.7.0.dev20250119+xpu

Could you please double check it? And provide more information of the environment

[pbchekin]

I think you are testing in the environment with oneAPI (or Deep Learning Essentials) installed and initialized. Try in a clean environment.

Version 2.7.0.dev20250110+xpu

# Does not show missing runtime dependencies
Dependencies.exe -chain C:\Users\vagrant\.venv\lib\site-packages\torch\lib\shm.dll -depth 2 | grep NOT_FOUND

Versions 2.7.0.dev20250117+xpu, 2.7.0.dev20250119+xpu

# Shows missing runtime dependencies
Dependencies.exe -chain C:\Users\vagrant\.venv\lib\site-packages\torch\lib\shm.dll -depth 2 | grep NOT_FOUND
|  |  ├ sycl8.dll (NOT_FOUND) :
|  |  ├ pti_view-0-10.dll (NOT_FOUND) :

In the newest builds torch_cpu.dll depends on sycl8.dll and pti_view-0-10.dll:

├ shm.dll (ROOT) : C:\Users\vagrant\.venv\lib\site-packages\torch\lib\shm.dll
|  ├ torch_cpu.dll (ApplicationDirectory) : C:\Users\vagrant\.venv\lib\site-packages\torch\lib\torch_cpu.dll
|  |  ├ sycl8.dll (NOT_FOUND) :
|  |  ├ pti_view-0-10.dll (NOT_FOUND) :
...

[chuanqi129]

Hi @pbchekin, I didn't install oneapi on my test machine. Could you share your pip list? Those dlls should be installed as torch whl dependencies directly

[pbchekin]

Could you share your pip list?

Package            Version
------------------ ---------------------
certifi            2024.12.14
cmake              3.31.4
colorama           0.4.6
cython             3.0.11
exceptiongroup     1.2.2
execnet            2.1.1
filelock           3.16.1
fsspec             2024.10.0
iniconfig          2.0.0
intel-cmplr-lib-rt 2025.0.2
intel-cmplr-lib-ur 2025.0.2
intel-cmplr-lic-rt 2025.0.2
intel-pti          0.10.0
intel-sycl-rt      2025.0.2
jinja2             3.1.4
markupsafe         2.1.5
mpmath             1.3.0
networkx           3.4.2
numpy              1.26.4
packaging          24.2
pluggy             1.5.0
pybind11           2.13.6
pytest             8.3.4
pytest-select      0.1.2
pytest-xdist       3.6.1
scipy              1.13.1
setuptools         75.8.0
sympy              1.13.1
tcmlib             1.2.0
tomli              2.2.1
torch              2.7.0.dev20250110+xpu
triton             3.2.0+git1f735815
typing-extensions  4.12.2
umf                0.9.1
wheel              0.45.1

[pbchekin]

Let me clarify what is the problem:

• The old build (2.7.0.dev20250110+xpu) does not have import time dependency on sycl8.dll and pti_view-0-10.dll.
• The newer builds (2.7.0.dev20250117+xpu, 2.7.0.dev20250119+xpu) do have import time dependency on sycl8.dll and pti_view-0-10.dll.

Those dlls should be installed as torch whl dependencies directly

These DLLs are installed. In my case, they are installed to

C:\Users\vagrant\.venv\Library\bin\sycl8.dll
C:\Users\vagrant\.venv\bin\pti_view-0-10.dll

My PATH is

C:\Users\vagrant\.venv/Scripts;C:\Program Files\OpenSSH;C:\WINDOWS\system32;C:\WINDOWS;C:\WINDOWS\System32\Wbem;C:\WINDOWS\System32\WindowsPowerShell\v1.0\;C:\WINDOWS\System32\OpenSSH\;C:\Program Files\PowerShell\7\;C:\Program Files\rsync;C:\Program Files\OpenSSH;C:\ProgramData\chocolatey\bin;C:\Program Files\Git\cmd;C:\Program Files\Git\mingw64\bin;C:\Program Files\Git\usr\bin;C:\Program Files\dotnet\;C:\Users\vagrant\.local\bin;C:\Users\vagrant\AppData\Local\Microsoft\WindowsApps;

If .venv\Library\bin and .venv\bin are not in the PATH then OS will not find these DLLs.

[xuhancn]

Hi @pbchekin

We enabled kineto for Windows XPU during 0110 and 0117 by PR: #144316

The sycl and pti_view modules are engaged by kineto. Could you please try to force-reinstall to fix it.

pip install torch --index-url https://download.pytorch.org/whl/nightly/xpu  --upgrade --force-reinstall

[etaf]

@pbchekin I was able to reproduce your problem when I created a conda env by conda create -n reproduce (acutall not clean).

conda create -n reproduce
conda activate reproduce
#actuall the way we created this new conda env is not clean.
pip install torch --index-url https://download.pytorch.org/whl/nightly/xpu
python -c "import torch;print(torch.__version__)"
(reproduce) C:\Users\sdp>python -c "import torch;print(torch.__version__)"
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "C:\Users\sdp\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.10_qbz5n2kfra8p0\LocalCache\local-packages\Python310\site-packages\torch\__init__.py", line 274, in <module>
    _load_dll_libraries()
  File "C:\Users\sdp\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.10_qbz5n2kfra8p0\LocalCache\local-packages\Python310\site-packages\torch\__init__.py", line 270, in _load_dll_libraries
    raise err
OSError: [WinError 126] The specified module could not be found. Error loading "C:\Users\sdp\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.10_qbz5n2kfra8p0\LocalCache\local-packages\Python310\site-packages\torch\lib\shm.dll" or one of its dependencies.

But if I create a real clean conda env like conda create -n torch-py3.9 python=3.9, there is no prblem here, all dependency can be found:

conda create -n torch-py3.9 python=3.9
conda activate torch-py3.9 
pip install torch --index-url https://download.pytorch.org/whl/nightly/xpu
(torch-py3.9) C:\Users\sdp>python -c "import torch;print(torch.__version__)"
2.7.0.dev20250119+xpu

(torch-py3.9) C:\Users\sdp>dlldiag deps "C:\Users\sdp\miniforge3\envs\torch-py3.9\lib\site-packages\torch\lib\torch_cpu.dll"
DLL Diagnostic Tools version 0.0.18
Copyright (c) 2019-2021 Adam Rehn

Parsing module header and identifying direct dependencies... done.

Parsed module details:
Module:          C:\Users\sdp\miniforge3\envs\torch-py3.9\lib\site-packages\torch\lib\torch_cpu.dll
Type:            Dynamic-Link Library
Architecture:    x64

Attempting to load the module's direct dependencies:

api-ms-win-crt-convert-l1-1-0.dll        Loaded successfully
api-ms-win-crt-environment-l1-1-0.dll    Loaded successfully
api-ms-win-crt-filesystem-l1-1-0.dll     Loaded successfully
api-ms-win-crt-heap-l1-1-0.dll           Loaded successfully
api-ms-win-crt-locale-l1-1-0.dll         Loaded successfully
api-ms-win-crt-math-l1-1-0.dll           Loaded successfully
api-ms-win-crt-process-l1-1-0.dll        Loaded successfully
api-ms-win-crt-runtime-l1-1-0.dll        Loaded successfully
api-ms-win-crt-stdio-l1-1-0.dll          Loaded successfully
api-ms-win-crt-string-l1-1-0.dll         Loaded successfully
api-ms-win-crt-time-l1-1-0.dll           Loaded successfully
api-ms-win-crt-utility-l1-1-0.dll        Loaded successfully
c10.dll                                  Loaded successfully
fbgemm.dll                               Loaded successfully
KERNEL32.dll                             Loaded successfully
libiomp5md.dll                           Loaded successfully
MSVCP140.dll                             Loaded successfully
pti_view-0-10.dll                        Loaded successfully
sycl8.dll                                Loaded successfully
uv.dll                                   Loaded successfully
VCRUNTIME140.dll                         Loaded successfully
VCRUNTIME140_1.dll                       Loaded successfully
WS2_32.dll                               Loaded successfully

So it seems like a env issue.
You can check pip list to see if it's clean if you created a new env.

[pbchekin]

You can check pip list to see if it's clean if you created a new env.

Yes, the error is reproducible in a new clean venv.

[pbchekin]

Steps to reproduce:

On a clean Windows 11 machine install Python. I've used chocolatey to install the latest Python 3.9:

> choco install python39
> python --version

Python 3.9.13

Create a new venv and activate it:

> python -m venv venv
> call venv\Scripts\activate.bat

Double check nothing is installed:

> pip list

Package    Version
---------- -------
pip        22.0.4
setuptools 58.1.0

Install PyTorch XPU:

> pip install torch --index-url https://download.pytorch.org/whl/nightly/xpu
> pip list

Package            Version
------------------ ---------------------
filelock           3.16.1
fsspec             2024.10.0
intel-cmplr-lib-rt 2025.0.2
intel-cmplr-lib-ur 2025.0.2
intel-cmplr-lic-rt 2025.0.2
intel-pti          0.10.0
intel-sycl-rt      2025.0.2
Jinja2             3.1.4
MarkupSafe         2.1.5
mpmath             1.3.0
networkx           3.2.1
pip                22.0.4
setuptools         58.1.0
sympy              1.13.1
tcmlib             1.2.0
torch              2.7.0.dev20250120+xpu
typing_extensions  4.12.2
umf                0.9.1 

Reproduce the error:

> python -c "import torch; print(torch.__version__)"

Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "C:\Users\vagrant\venv\lib\site-packages\torch\__init__.py", line 274, in <module>
    _load_dll_libraries()
  File "C:\Users\vagrant\venv\lib\site-packages\torch\__init__.py", line 270, in _load_dll_libraries
    raise err
OSError: [WinError 126] The specified module could not be found. Error loading "C:\Users\vagrant\venv\lib\site-packages\torch\lib\shm.dll" or one of its dependencies.

[yinghu5]

Hi @pbchekin ,

thank you very much for reporting the issues. we found that if we use venv to create test environment

python -m venv test
we can reproduce the problem.
and if we use conda
conda create -n test python=3.10 xx
then everything looks fine.

the root cause as you discovered, is that the pti dlls were in C:\Users\vagrant.venv\bin , which is not in path and also not be linkable in venv environment, while conda environment will enable the link path.

so as a workaroud, would you please copy the two dlls under C:\Users\vagrant.venv\bin to C:\Users\vagrant.venv\Library\bin
and see if it works?

and we will fix the issue and update if any result.

thank you!