[Bug] CUDA error when running sample code

I am running the basic code to train a PPO agent on CartPole and it gives me an CUDA error:

RuntimeError: CUDA error: CUBLAS_STATUS_INTERNAL_ERROR when calling cublasCreate(handle)

in “/.local/lib/python3.6/site-packages/torch/nn/functional.py”, line 1753, in linear return torch._C._nn.linear(input, weight, bias)

Python version 3.6.9

Pytorch version 1.8.0

Code is this:

import gym

from stable_baselines3 import PPO

env = gym.make('CartPole-v1')

model = PPO('MlpPolicy', env, verbose=1)
model.learn(total_timesteps=10000)

obs = env.reset()
for i in range(1000):
    action, _states = model.predict(obs, deterministic=True)
    obs, reward, done, info = env.step(action)
    env.render()
    if done:
      obs = env.reset()

env.close()

[ac-93]

Same issue for me, seems to be torch 1.8 related (https://github.com/pytorch/pytorch/issues/53336)

Downgrading to torch 1.7.1 seems to work fine for now.

[youryz]

Actually no @Miffyli , the code araffin provides works. My own deep learning projects based on Pytorch also works (MLP training & inference). What I mean sample code is the code in my initial post, which is used to train a PPO agent, and I post it again here:

import gym

from stable_baselines3 import PPO

env = gym.make('CartPole-v1')

model = PPO('MlpPolicy', env, verbose=1)
model.learn(total_timesteps=10000)

obs = env.reset()
for i in range(1000):
    action, _states = model.predict(obs, deterministic=True)
    obs, reward, done, info = env.step(action)
    env.render()
    if done:
      obs = env.reset()

env.close()

I’ve checked my Pytorch/CUDA and there’s no problem with them. I will check again in case. But I don’t think there should be any problem with them since my past projects worked well.

the error that you have is definitely a pytorch/cuda error. I have tested SB3 with python 3.6 and pytorch 1.8 on my machine and on a google colab instance and did not get any error… i guess the code will also run if you use the cpu only.

I switched to cpu by passing ‘cpu’ to PPO, it trains two epochs, then it breaks again and gives me the same CUDA error.

I have two different machines and the code works on neither one.

Anyway I will check my CUDA installation.

Thanks