Trying to run basic examples, but I think I have some issues with my config #177
Comments
|
The 0.1.6 version is only compatible with PyTorch 1.4, this probably explains your compilation errors. |
|
Thank you for your quick answer ! :) |
|
Ok, I used the master tip and it works, I just had to add /lib in my path |
|
I'm trying to use CUDA, used the correct libtorch from website, but I encounter this issue, any idea on how to pass that ? prints False Thanks in advance |
|
Right, the cuda detection and linking has become a bit messier since the 1.5 release, see [1] for some context. The main culprit is that cargo does not let you pass arbitrary flags so it is difficult to pass the |
|
With your update, I now have a new error : Maybe I have a missing dll or something ? I can successfully use tensorflow on my system as well as pyTorch python with cuda 10.2 and dll 7.5 though . |
|
I'll be quite available this week end if you need me to test anything |
|
That's weird, the println!("cargo:rustc-link-lib=c10_cuda");It doesn't seem likely to help but worth trying when you get a chance. |
|
I just tried this : and got this result : |
|
Thanks, that's sadly the same error as previously, I can see the symbol being exported by torch_cuda.lib using winedump but the linker doesn't seem to pick this up - I'm afraid I don't know enough about the windows linker to be helpful here. |
|
Ok, I'll try to work on the CPU for now, thanks for trying, if I find something I'll let you know. Thank you again and have a great day ! |
|
I was able to get access to a win+nvidia box and was getting the same error as you did when using cuda. I just pushed some changes that should hopefully fix the issue (at least they worked on this box), let me know if this still doesn't work for you. |
|
Wow thank you a lot ! It indeed prints :) I just have to manage to specify GPU for your MNIST example, I replaced I have this error now : On this last line : I think I miss some device specifications somewhere. |
|
Haha that's was a stupid mistake, I did put train_images into train_labels ... So everything seems to work greatly ! Now I have no excuses to translate my whole python RL system to Rust. I'll keep you posted ! Thank you a lot ! :) |
|
Glad that it ended up working, and let me know if there is anything else that could be helpful to you. |
|
Yes, that's why in the first place I was interested in your repository :) I just encountered another issue though, when I run with --release I encounter the same issue as before : Any idea ? |
|
Indeed it seems that symbols get stripped more aggressively in release mode on windows than they do on linux. use torch_sys::dummy_cuda_dependency();
fn main() {
unsafe{ dummy_cuda_dependency(); }
// your normal code there
}If this ends up working well for you, we can try to make this nicer, at least removing the |
|
Great! That indeed did the trick, do you want me to close the issue then or do you prefer making another test before ? Thank you again. |
|
Cool, we can probably keep this issue open until we have a less hacky way of triggering the dependency. |
|
That trick works indeed, thanks, but one needs to add the dependency to your torch-sys as well. Here is my basic code to test it, and print main.rs -> extern crate tch;
use tch::Device;
fn main() {
unsafe{ torch_sys::dummy_cuda_dependency(); }
let device = Device::cuda_if_available();
println!("using {:?}", device);
} |
|
I think I just got hit by this but I'm on Ubuntu. I get this error : |
|
Not sure to understand what your setup is: if you didn't install libtorch manually nor have set the |
|
Apologies, it seems like I've not been very specific and given you context. |
|
Closing this for now as not much has been done recently. The PR[1] allowing cargo to pass arbitrary linker flags is making slow but steady progress so hopefully the hacky bits can be removed at some point. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Hello there :)
I wanted to give this crate a go and was attracted by the apparent simplijcity of usage.
I downloaded libtorch 1.5 from pytorch website and filled the corresponding Environment Variables
I tried to start a new cargo project and just put :
in my cargo.toml file.
I used the first main in example :
When I try to run it I encounter som compilation issue :
Do you have any idea of what I should do or what is missing on my system ?
I'm on Windows 10
using CLion with rust plugin
Thank you in advance.
Nicolas
The text was updated successfully, but these errors were encountered: