Checkpoints cannot use CUDA

[Theo-Fan]

I trained a model using GPU, and everything worked fine during training. However, when I try to load the saved checkpoint later (on a CPU-only machine), I get the following error:

025-07-24 22:22:56,431	ERROR actor_manager.py:873 -- Ray error (System error: Attempting to deserialize object on a CUDA device but torch.cuda.is_available() is False. 
If you are running on a CPU-only machine, please use torch.load with map_location=torch.device('cpu') to map your storages to the CPU.

Steps to reproduce

1. Train a model on a GPU machine
2. Save the model or checkpoint
3. Move the checkpoint to a CPU-only machine
4. Try to load it (using ·torch.load()·)

Test:

from ray import train, tune
from ray.rllib.algorithms.ppo import PPOConfig

config = (
    PPOConfig()
    .environment("Pendulum-v1")
    
    .training(
        lr=tune.grid_search([0.001, 0.0001]),
    )
    .env_runners(
        num_env_runners=2,
        batch_mode="complete_episodes"
    )
    .learners(
        num_learners=1,
        num_gpus_per_learner=1, # gpu config
    )
    
)


tuner = tune.Tuner(
    config.algo_class,
    param_space=config,
    run_config=train.RunConfig(
        stop={
        	"training_iteration": 5,
        },
        checkpoint_config=tune.CheckpointConfig(
            checkpoint_at_end=True,  # Problem
        ),
    ),
)

results = tuner.fit()

Environment:

• RLlib version: 2.47.1 (Ray 2.47.1)
• CUDA available during training: Yes
• CUDA available during restore: No
• OS: Ubuntu 22.04
• Python version: 3.10.18

[simonsays1980]

Originally fixed in #43906 and reintroduced