BUG: Using a length-only coord causes it to be volatile

[JasonTam]

Describe the issue:

When a variable is defined with a coord, it's considered volatile. Then sample_posterior_predictive does not transfer the node over.
I believe the volatility condition that's met is

pymc/pymc/sampling/forward.py

Lines 208 to 212 in 9eaf92a

	or ( # SharedVariables, except RandomState/Generators
	isinstance(node, SharedVariable)
	and not isinstance(node, RandomStateSharedVariable | RandomGeneratorSharedVariable)
	and not shared_value_matches(node)
	)

The coord is intended to be immutable, but after #7047, coords always mutable

Somewhat related:

• Make sample_posterior_predictive API less surprising #7069
• sample_posterior_predictive does not consider changes in "ConstantData" and "Constant coords" #6876
• sample_posterior_predictive wrongly considers Deterministic variables as source of volatility #7183
  (lmk if this issue is considered a duplicate of anything)

Reproduceable code example:

import pymc as pm

y_obs = [0] * 3
with pm.Model() as m:
    m.add_coord("coord0", length=1)
    x = pm.Data("x", [0] * 3)
    b = pm.Normal("b", dims="coord0")  # Causes `b` to be volatile
    # b = pm.Normal("b", shape=1)      # Works as expected (`b` is transferred)
    y = pm.Normal("y", b[0] * x, observed=y_obs)
    idata = pm.sample()
    
    pm.sample_posterior_predictive(trace=idata, var_names='y')

Error message:

Sampling: [b, y]

PyMC version information:

5.15.1

Context for the issue:

Cannot use sample_posterior_predictive as intended

[welcome]

]
🎉 Welcome to PyMC! 🎉 We're really excited to have your input into the project! 💖

If you haven't done so already, please make sure you check out our Contributing Guidelines and Code of Conduct.

[ricardoV94]

The volatility logic is supposed to check if the values of mutable data actually changed (it's not enough to be mutable) so this is quite surprising

[JasonTam]

Ah I think I understand what the issue is. The condition check for constant_coords is:

pymc/pymc/sampling/forward.py

Lines 797 to 803 in 9eaf92a

	current_coord = model.coords.get(dim, None)
	if (
	current_coord is not None
	and len(coord) == len(current_coord)
	and np.all(coord == current_coord)
	):
	constant_coords.add(dim)

But if you add a coord without any values (just the length)

m.add_coord("coord0", length=1)

then the coord, will be None. We will only have dim_lengths information.

pymc/pymc/model/core.py

Lines 1030 to 1031 in 9eaf92a

	self._dim_lengths[name] = length
	self._coords[name] = values

I guess my question now: is this the intended?

[ricardoV94]

I see, we always always get tricked by coords without values (as in we tend to forget they exist all the time). In that case the volatility check should be whether the length changed or not?

What do you think? And do you want to open a PR to fix it?

[JasonTam]

For sure it threw me in for a loop. So I think it should either be considered constant if the length stays the same, or a log message should be fired.

I've opened a PR #7381 and we can continue any discussion there