admin管理员组

文章数量:1278793

I'm trying to train a GPT-2–style language model on a system at my institute that uses a Tesla K20m GPU (compute capability 3.5).

I initially attempted to build PyTorch from source to enable proper CUDA kernel support for my GPU by setting the environment variable TORCH_CUDA_ARCH_LIST="3.5". However, am running into several issues while building from source. I would therefore prefer to use a pre-built version of PyTorch (along with the appropriate CUDA toolkit and Python version) that explicitly supports compute capability 3.5.

My Environment Details:

  • GPU: Tesla K20m (compute capability 3.5)
  • System-wide CUDA: CUDA 11.2 is installed
  • Python: I have tried using Python 3.7 and 3.8 in different conda environments
  • Current Attempts:
    • I have also attempted to use different combinations of pre-built binaries and python versions, but they do not include the necessary CUDA kernels for sm_35, resulting in runtime errors such as: 
      RuntimeError: CUDA error: no kernel image is available for execution on the device

My Question:
Is there a recommended pre-built configuration (i.e. specific versions of PyTorch, CUDA toolkit, and Python) known to support Tesla K20m (compute capability 3.5) out-of-the-box? Preferably, I need a solution that works with a system-wide CUDA installation (even if it means using an older CUDA version like 10.2) and that avoids the need for building from source with an updated GCC. Any pointers to known configurations, conda channels, or documentation would be greatly appreciated.

Thank you in advance for your help!

Edit: I got my institute to install GCC 9.5.0, but the remaining building process itself is very error prone. I am trying to navigate through it.

本文标签：

版权声明：本文标题：python - Looking for Pre-built PyTorch Configuration Supporting Tesla K20m (sm_35) on an Institute System - Stack Overflow 内容由网友自发贡献，该文观点仅代表作者本人， 转载请联系作者并注明出处：http://www.betaflare.com/web/1741289102a2370444.html， 本站仅提供信息存储空间服务，不拥有所有权，不承担相关法律责任。如发现本站有涉嫌抄袭侵权/违法违规的内容，一经查实，本站将立刻删除。