Building a Sequence with GPU tensor parameters fails

[sgrava]

When building a Parametric sequence from a torch.Tensor that lives on GPU, I get the following error:

Traceback (most recent call last):
  File "/home/stefano/Workspace/pulser-gym/gpu_parametric_sequence.py", line 32, in <module>
    built_seq = seq.build(omega=omega, area=area)
  File "/home/stefano/Workspace/pulser-gym/.hatch/pulser-gym/lib/python3.10/site-packages/pulser/sequence/sequence.py", line 1688, in build
    args_ = [
  File "/home/stefano/Workspace/pulser-gym/.hatch/pulser-gym/lib/python3.10/site-packages/pulser/sequence/sequence.py", line 1689, in <listcomp>
    arg.build() if isinstance(arg, Parametrized) else arg
  File "/home/stefano/Workspace/pulser-gym/.hatch/pulser-gym/lib/python3.10/site-packages/pulser/parametrized/paramobj.py", line 198, in build
    self._instance = obj(*args_, **kwargs_)
  File "/home/stefano/Workspace/pulser-gym/.hatch/pulser-gym/lib/python3.10/site-packages/pulser/pulse.py", line 107, in __init__
    if np.any(amplitude.samples.as_array(detach=True) < 0):
  File "/home/stefano/Workspace/pulser-gym/.hatch/pulser-gym/lib/python3.10/site-packages/pulser/waveforms.py", line 121, in samples
    return self._samples.copy()
  File "/usr/lib/python3.10/functools.py", line 981, in __get__
    val = self.func(instance)
  File "/home/stefano/Workspace/pulser-gym/.hatch/pulser-gym/lib/python3.10/site-packages/pulser/waveforms.py", line 533, in _samples
    return self._value * np.ones(self.duration)
  File "/home/stefano/Workspace/pulser-gym/.hatch/pulser-gym/lib/python3.10/site-packages/pulser/math/abstract_array.py", line 216, in __mul__
    return AbstractArray(operator.mul(*self._binary_operands(other)))
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

Since some computation is done with that tensor to build the sequence, tensors need to be on the same device.
Two solutions:

1. Pulser handle moving the parameter tensor to CPU before building the sequence
2. Every backend ensures to move all parameters to CPU before building a sequence
   Any idea if one solution is better?

To reproduce

from pulser import Sequence, Pulse, Register
from pulser.devices import MockDevice
import torch

# create sequence and declare channels
reg = Register.square(2, spacing=8, prefix="q")
seq = Sequence(reg, MockDevice)
seq.declare_channel("rydberg_global", "rydberg_global")

# declare sequence variables
omega_param = seq.declare_variable("omega")

# create and add parametric pulse
pulse_const = Pulse.ConstantPulse(1000, omega_param, 0.0, 0.0)
seq.add(pulse_const, "rydberg_global")

# CPU works fine!
omega = torch.tensor(1.0, device="cpu")
built_seq = seq.build(omega=omega)

# GPU error
omega = torch.tensor(1.0, device="cuda")
built_seq = seq.build(omega=omega)

[sgrava]

I opened this issue just to keep in mind this problem.
Solution 2 seems easier for the moment, so it will be explored first in the backend emulators.

[HGSilveri]

Thanks @sgrava !