[Compatibility Issue] `np.einsum_path` contraction tuple format change in numpy >= 2.4

[xyf5432]

Affected File

src/qibo/quantum_info/quantum_networks.py:1023-1048

Current Code

_, contracrtion_list = np.einsum_path(
    subscripts, *tensors, optimize=False, einsum_call=True
)

inds, idx_rm, einsum_str, _, _ = contracrtion_list[0]  # 5-element unpack
input_str, results_index = einsum_str.split("->")
inputs = input_str.split(",")

for ind in idx_rm:   # line 1033 — idx_rm actively used
    ...

Root Cause

In numpy 2.4, the contraction tuples returned by einsum_path were reduced from 5 elements to 3 elements (einsumfunc.py:874):

numpy version	Contraction tuple	Count
≤ 2.3.3	(contract_inds, idx_removed, einsum_str, remaining, do_blas)	5
≥ 2.4.0	(contract_inds, einsum_str, remaining)	3

The idx_removed and do_blas fields were removed. Line 1027 unpacks 5 elements → crash:

ValueError: not enough values to unpack (expected 5, got 3)

Impact

• Severity: High — runtime crash on numpy >= 2.4.0
• idx_rm (i.e., idx_removed) is used in business logic at line 1033 and cannot simply be discarded

Solution

Inspect the tuple length rather than parsing the numpy version string. This is robust across numpy versions, forks, and backports.

_, contraction_list = np.einsum_path(
    subscripts, *tensors, optimize=False, einsum_call=True
)

contraction = contraction_list[0]

if len(contraction) == 5:
    # numpy <= 2.3.x: (contract_inds, idx_removed, einsum_str, remaining, do_blas)
    inds, idx_rm, einsum_str, _, _ = contraction
elif len(contraction) == 3:
    # numpy >= 2.4.x: (contract_inds, einsum_str, remaining)
    inds, einsum_str, _remaining = contraction
    # idx_rm: indices contracted away in this step
    # = indices appearing in the LHS of einsum_str but not in the RHS
    input_str, results_index = einsum_str.split("->")
    idx_rm = set(input_str.replace(",", "")) - set(results_index)
else:
    raise ValueError(
        f"Unexpected einsum_path contraction tuple length: "
        f"{len(contraction)} (expected 3 or 5)"
    )

input_str, results_index = einsum_str.split("->")
inputs = input_str.split(",")

# ... rest of the code unchanged ...

References

• NumPy einsumfunc.py — contraction tuple construction
• pymc-labs/CausalPy#723 — same root cause
• pytensor fix for the same issue

[alecandido]

@renatomello seems a good point

[xyf5432]

A fix is now up: PR #1876 implements the tuple-length check and reconstructs idx_removed from einsum_str. Verified on both tuple formats (numpy 2.3.3: 13/13 tests pass; numpy 2.5.2: 13/13 with the fix, 6 failures without). qibo declares numpy = "^2.0.0", so numpy 2.4+ is a fully allowed version and this is a real regression for current installs.