search

FalkonML / falkon

Large-scale, multi-GPU capable, kernel solver
https://falkonml.github.io/falkon/
MIT License
181 stars 22 forks source link

installing falkon seems to be succesful but failed when calling import falkon #55

Open hep07 opened 1 year ago

hep07 commented 1 year ago

Hi

My system I can use keops on GPU without problem and I install cuda11.6 already. I have a 3090TI on Ubuntu 22.06.

When I install falkon as instructed using command "pip uninstall git+https://github.com/falkonml/falkon.git", everything is fine without warnings/erros. But when I test it in a notebook using import falkon, I got the following error:

OSError Traceback (most recent call last) /tmp/ipykernel_1096354/295832182.py in 6 plt.style.use('ggplot') 7 ----> 8 import falkon

~/anaconda3/envs/repo/lib/python3.8/site-packages/falkon/init.py in 8 "c_ext", [os.path.dirname(file)]) 9 if spec is not None: ---> 10 torch.ops.load_library(spec.origin) 11 else: 12 raise ImportError("Failed to find C-extension. Please recompile Falkon.")

~/anaconda3/envs/repo/lib/python3.8/site-packages/torch/_ops.py in load_library(self, path) 571 # static (global) initialization code in order to register custom 572 # operators with the JIT. --> 573 ctypes.CDLL(path) 574 self.loaded_libraries.add(path) 575

~/anaconda3/envs/repo/lib/python3.8/ctypes/init.py in init(self, name, mode, handle, use_errno, use_last_error, winmode) 371 372 if handle is None: --> 373 self._handle = _dlopen(self._name, mode) 374 else: 375 self._handle = handle

OSError: /home/mc/anaconda3/envs/repo/lib/python3.8/site-packages/falkon/c_ext.so: undefined symbol: _ZN2at4cuda28getCurrentCUDASolverDnHandleEv

It looks like something wrong with cuda?

CHDev93 commented 1 year ago

Am seeing the same error. Was fine with torch==1.10.2+cu113 and pykeops==2.1 though. This issue mentioned that torch==1.12 was known not to be compatible but that the issue appeared to be fixed in the next version of torch. Seems there's a different issue though

Giodiro commented 1 year ago

Hi @hep07, sorry again for the late answer. Have you managed to figure this out?

Giacomo

CHDev93 commented 1 year ago

Not the original author of this post but I was not able to find a workaround aside from just having two different virtual environments. One with a current version of torch and one with the older version against which I can build falkon