Pytorch 2.0.1 pypi wheel does not install dependent cuda libraries #100974
Description
🐛 Describe the bug
With torch 2.0.1 the torch pypi wheel does not depend on cuda libraries anymore. Therefore when starting torch on a GPU enabled machine, it complains ValueError: libnvrtc.so.*[0-9].*[0-9] not found in the system path (stacktrace see at the end below).
When I show the dependency trees for torch=2.0.1 and torch=2.0.0 with poetry (installed on the same machine with same dependency file as before) it becomes clear that torch 2.0.1 is missing the nvidia dependencies:
└── torch 2.0.1
├── filelock *
├── jinja2 *
│ └── markupsafe >=2.0
├── networkx *
├── sympy *
│ └── mpmath >=0.19
└── typing-extensions *
└── torch 2.0.0
├── filelock *
├── jinja2 *
│ └── markupsafe >=2.0
├── networkx *
├── nvidia-cublas-cu11 11.10.3.66
│ ├── setuptools *
│ └── wheel *
├── nvidia-cuda-cupti-cu11 11.7.101
│ ├── setuptools * (circular dependency aborted here)
│ └── wheel * (circular dependency aborted here)
├── nvidia-cuda-nvrtc-cu11 11.7.99
│ ├── setuptools * (circular dependency aborted here)
│ └── wheel * (circular dependency aborted here)
├── nvidia-cuda-runtime-cu11 11.7.99
│ ├── setuptools * (circular dependency aborted here)
│ └── wheel * (circular dependency aborted here)
├── nvidia-cudnn-cu11 8.5.0.96
│ ├── setuptools * (circular dependency aborted here)
│ └── wheel * (circular dependency aborted here)
├── nvidia-cufft-cu11 10.9.0.58
├── nvidia-curand-cu11 10.2.10.91
│ ├── setuptools * (circular dependency aborted here)
│ └── wheel * (circular dependency aborted here)
├── nvidia-cusolver-cu11 11.4.0.1
│ ├── setuptools * (circular dependency aborted here)
│ └── wheel * (circular dependency aborted here)
├── nvidia-cusparse-cu11 11.7.4.91
│ ├── setuptools * (circular dependency aborted here)
│ └── wheel * (circular dependency aborted here)
├── nvidia-nccl-cu11 2.14.3
├── nvidia-nvtx-cu11 11.7.91
│ ├── setuptools * (circular dependency aborted here)
│ └── wheel * (circular dependency aborted here)
├── sympy *
│ └── mpmath >=0.19
├── triton 2.0.0
│ ├── cmake *
│ ├── filelock * (circular dependency aborted here)
│ ├── lit *
│ └── torch * (circular dependency aborted here)
└── typing-extensions *
Here the stacktrace of the error at runtime:
File "/home/ray/anaconda3/envs/myenv/lib/python3.10/site-packages/easyocr/recognition.py", line 2, in <module>
import torch
File "/home/ray/anaconda3/envs/myenv/lib/python3.10/site-packages/torch/__init__.py", line 228, in <module>
_load_global_deps()
File "/home/ray/anaconda3/envs/myenv/lib/python3.10/site-packages/torch/__init__.py", line 189, in _load_global_deps
_preload_cuda_deps(lib_folder, lib_name)
File "/home/ray/anaconda3/envs/myenv/lib/python3.10/site-packages/torch/__init__.py", line 154, in _preload_cuda_deps
raise ValueError(f"{lib_name} not found in the system path {sys.path}")
ValueError: libnvrtc.so.*[0-9].*[0-9] not found in the system path ['/home/ray', '/home/ray/anaconda3/lib/python3.10/site-packages/ray/dashboard', '/home/ray/anaconda3/envs/myenv/lib/python3.10/site-packages/ray/thirdparty_files', '/home/ray/anaconda3/lib/python3.10/site-packages/ray/_private/workers', '/home/ray/anaconda3/envs/myenv/lib/python3.10', '/home/ray/anaconda3/envs/myenv/lib/python3.10/lib-dynload', '/home/ray/anaconda3/envs/myenv/lib/python3.10/site-packages']
Versions
Version where the issue occurs is the pypi wheel of torch 2.0.1.
When trying to run python collect_env.py to collect the versions, two errors shows up:
"OSError: libcurand.so.10: cannot open shared object file: No such file or directory"
During handling of the above exception, another exception occurred:
"ValueError: libnvrtc.so.*[0-9].*[0-9] not found in the system path"