You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Me too, finally got it working on Linux, but no GPU yet.
This is really a bit of a black art, others say downgrading with:
python -m pip install tensorflow-gpu==2.3.1 may work, but for me I just get:
RuntimeError: CUDA runtime implicit initialization on GPU:0 failed. Status: device kernel image is invalid
I am running a 40 series GPU so perhaps it's not compatible. I'm still working out how to get all the python / cuda / system versions and packages to gain a blessing from the universe. May it be blessed today.
GPU works on other ML apps so it must be something simple, will report back if I get a fix.
EDIT
It turns out using 2.2 instead of 2.3 above nearly gets it to work, like it detects the GPU and initialises the models to 60% before hanging.
EDIT2
Nope on a second attempt it did actually work. But it took ages, longer than in CPU.
Now waiting on the starting attempt and wondering if it will be as slow to train as it will to load. It's way faster on windows so far, I can see the GPU is idle while all this is going on.
Activity
marshalleq commentedon Jul 17, 2023
Me too, finally got it working on Linux, but no GPU yet.
This is really a bit of a black art, others say downgrading with:
python -m pip install tensorflow-gpu==2.3.1 may work, but for me I just get:
RuntimeError: CUDA runtime implicit initialization on GPU:0 failed. Status: device kernel image is invalid
I am running a 40 series GPU so perhaps it's not compatible. I'm still working out how to get all the python / cuda / system versions and packages to gain a blessing from the universe. May it be blessed today.
GPU works on other ML apps so it must be something simple, will report back if I get a fix.
EDIT
It turns out using 2.2 instead of 2.3 above nearly gets it to work, like it detects the GPU and initialises the models to 60% before hanging.
EDIT2
Nope on a second attempt it did actually work. But it took ages, longer than in CPU.
Now waiting on the starting attempt and wondering if it will be as slow to train as it will to load. It's way faster on windows so far, I can see the GPU is idle while all this is going on.
tailangjun commentedon Sep 5, 2023
my tensorflow-gpu version is 2.0.4, it can't work with GPU too.