ValueError: Jacobian shape ... constraint limitations got me. Will they for others?

[rsamf]

I recently ran into this exception when training an RL policy:

...

  File "/opt/conda/lib/python3.11/site-packages/genesis/engine/solvers/rigid/constraint_solver_decomp.py", line 42, in __init__
    self.constraint_state = array_class.get_constraint_state(self, self._solver)
                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/opt/conda/lib/python3.11/site-packages/genesis/utils/array_class.py", line 169, in get_constraint_state
    raise ValueError(
ValueError: Jacobian shape (np.int64(7577), 35, 16384) is too large for int32. Consider reducing the number of constraints or the number of degrees of freedom.

I won't pretend to know what's going on under the hood, but I'm hoping someone may know here. I am trying to saturate an H100 GPU. The model itself is pretty small allowing me to create a large number of environments, but I'm limited by some Jacobian that is bounded by 32 bits. Would a simple bump up to int64 fix this or is that unwise? If this seems like an issue, I can create one for it, but I wasn't sure what to call this.

[duburcqa]

Would a simple bump up to int64 fix this or is that unwise?

This is not supported by GsTaichi for now. You have no other solution than doing what this error message suggest: reduce the number of constraints. To do so, the easiest way is to reduce the maximum number of collision pairs:

            rigid_options=gs.options.RigidOptions(
                max_collision_pairs=50,
            ),

[rsamf]

Ooo I'll play around with parameter. I wasn't aware that it was related to collisions. Thanks for the insight. That's helpful.

[duburcqa]

We should probably clarify this point in the documentation.