[BUG] MPS device crashes on float64 specs when moving environments with .to('mps')

[bsprenger]

Describe the bug

Creating a SyncDataCollector (or Collector) with env_device="mps" crashes with TypeError: Cannot convert a MPS Tensor to float64 dtype because gymnasium/MuJoCo environments produce float64 observation and action spaces, and MPS does not support float64 tensors.

To Reproduce

On a Mac with MPS:

import torch
from torchrl.envs import GymEnv
from torchrl.collectors import SyncDataCollector

env = GymEnv("HalfCheetah-v4")

collector = SyncDataCollector(
    env,
    policy=env.rand_action,
    frames_per_batch=10,
    total_frames=10,
    device="mps",
    env_device="mps",
)

Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/Users/bensprenger/repositories/torchRL/torchrl/collectors/_base.py", line 1050, in __call__
    return super().__call__(*args, **kwargs)
  File "/Users/bensprenger/repositories/torchRL/torchrl/collectors/_single.py", line 519, in __init__
    self._apply_env_device()
  File "/Users/bensprenger/repositories/torchRL/torchrl/collectors/_single.py", line 872, in _apply_env_device
    self.env: EnvBase = self.env.to(self.env_device)
  File "/Users/bensprenger/repositories/torchRL/torchrl/envs/common.py", line 76, in wrapper
    result = func(self, *args, **kwargs)
  File "/Users/bensprenger/repositories/torchRL/torchrl/envs/common.py", line 4001, in to
    self.__dict__["_output_spec"] = self.output_spec.to(device)
  File "/Users/bensprenger/repositories/torchRL/torchrl/data/tensor_specs.py", line 5962, in to
    kwargs[key] = value.to(dest)
  File "/Users/bensprenger/repositories/torchRL/torchrl/data/tensor_specs.py", line 5962, in to
    kwargs[key] = value.to(dest)
  File "/Users/bensprenger/repositories/torchRL/torchrl/data/tensor_specs.py", line 3165, in to
    return Unbounded(shape=self.shape, device=dest_device, dtype=dest_dtype)
  File "/Users/bensprenger/repositories/torchRL/torchrl/data/tensor_specs.py", line 3034, in __call__
    instance = super().__call__(*args, **kwargs)
  File "/Users/bensprenger/repositories/torchRL/torchrl/data/tensor_specs.py", line 3130, in __init__
    torch.full(
TypeError: Cannot convert a MPS Tensor to float64 dtype as the MPS framework doesn't support float64. Please use float32 instead.

Expected behavior

A clear and concise description of what you expected to happen.

Screenshots

System info

• Installed from source in a micromamba environment
• macOS (Apple Silicon M-series)
• Python version 3.10

>>> import torchrl, numpy, sys
>>> print(torchrl.__version__, numpy.__version__, sys.version, sys.platform)
0.10.0 2.2.6 3.10.18 | packaged by conda-forge | (main, Jun  4 2025, 14:46:00) [Clang 18.1.8 ] darwin

Checklist

• I have checked that there is no similar issue in the repo (required)
• I have read the documentation (required)
• I have provided a minimal working example to reproduce the bug (required)

[bsprenger]

I opened PR #3548 to fix this. LMK what you think 🙂