Fix #884: test for nvidia-smi instead of nvcc
#918
Conversation
It seems that checking for the program `nvidia-smi` is the more reliable way to determine if a system has a usable NVIDIA GPU and the necessary drivers installed for running CUDA applications. Checking for `nvcc` may only tell you if the system can compile CUDA code, not necessarily if it can run it.
| # an NVIDIA GPU. Checking for nvcc may only tell you if the system can compile | ||
| # CUDA code, not necessarily if it can run it.) | ||
|
|
||
| ifneq (,$(shell command -v nvidia-smi > /dev/null 2>&1)) |
There was a problem hiding this comment.
There might be two issues here.
-
You discard both the standard output and the error output. So
ifneq (,$(shell command -v nvidia-smi > /dev/null 2>&1))is always false. -
It seems checking for
nvidia-smican also be unreliable. One can installnvidia-utilsto getnvidia-smi(even on a machine that does not have an NVidia GPU) without installing cuda, etc. But we neednvccto compile the code.
There was a problem hiding this comment.
- You discard both the standard output and the error output. So
ifneq (,$(shell command -v nvidia-smi > /dev/null 2>&1))is always false.
Ack, you're right of course. That should have been only 2> /dev/null. And of course, I only bothered to test it on a system that didn't have nvidia-smi, so the resulting behavior was what I expected. Anyway, thank you very for catching that.
- It seems checking for
nvidia-smican also be unreliable. One can installnvidia-utilsto getnvidia-smi(even on a machine that does not have an NVidia GPU) without installing cuda, etc. But we neednvccto compile the code.
Darn it. I guess one would need to test for nvcc and nvidia-smi, then grep the output of running nvidia-smi for an indication of a gpu card. This is getting complicated :-(.
I'm doing to close this PR and issue #884 and try to reproduce whatever led me to think it was an issue in the first place.
|
This PR is flawed and does not even address the problem fully. Closing. |
2 participants
It seems that checking for the program
nvidia-smiis the more reliable way to determine if a system has a usable NVIDIA GPU and the necessary drivers installed for running CUDA applications. Checking fornvccmay only tell you if the system can compile CUDA code, not necessarily if it can run it.This PR also fixes a small portability issue, where the use of
command -vis preferrable to the use ofwhichfor testing whether a program is available.