Fix IndexError in cast_inputs for ops with no formal inputs

[bodapatisaikrishna]

Motivation

autocast.cast_inputs() is the shared implementation behind both eager-mode
execution (dynamic_cast_inputs) and static script-translation
(static_cast_inputs). When it sees more positional arguments than the op's
formal input parameters, it's supposed to raise a clear error:

raise ValueError(
    f"Number of actual parameters {len(args)} "
    f"exceeds number of formal parameters {len(expected_inputs)}."
)

But the branch that decides whether to raise that error indexes into
expected_inputs[-1] unconditionally:

if i < len(expected_inputs):
    expected = expected_inputs[i]
elif expected_inputs[-1].variadic:   # <-- IndexError when expected_inputs == []
    ...
else:
    raise ValueError(...)

For an op with zero formal input parameters (only attributes — e.g.
RandomNormal, RandomUniform), expected_inputs is [], so
expected_inputs[-1] raises an unhandled IndexError: list index out of range instead of the intended ValueError.

This is reachable end-to-end through values.Op.__call__, which doesn't
pre-validate argument counts before dispatching to
evaluator.eval_op -> autocast.dynamic_cast_inputs -> cast_inputs. This
path is used for any op accessed dynamically (e.g. custom-domain ops via
Opset.__getattr__, or an Op constructed directly), as opposed to the
code-generated opset functions (e.g. opset18.RandomNormal), which have
fixed Python signatures that Python itself rejects extra arguments for.

Repro:

from onnxscript._internal import values
from onnxscript import opset18

values.Op(opset18, "RandomNormal")(1.0, 2.0, 3.0)
# IndexError: list index out of range

Fix

Guard the variadic check with expected_inputs and ... so an empty input
list falls through to the existing else branch and raises the intended,
descriptive ValueError instead of crashing with an unrelated IndexError.
One-line change; no other behavior affected — the non-empty variadic and
non-variadic paths are untouched.

Testing

Added onnxscript/_internal/autocast_test.py (no test file existed for this
module) covering:

• the empty-formal-inputs overflow case (was IndexError, now ValueError)
• the empty-formal-inputs no-args case (still returns (), no regression)
• the pre-existing non-variadic overflow case (still raises ValueError)
• the homogeneous-variadic case (guards against regressing the condition
  just touched)

$ python -m pytest onnxscript/_internal/autocast_test.py -q
4 passed
$ python -m pytest onnxscript/_internal/ -q
188 passed, 1 skipped, 1 xfailed, 3 xpassed

Confirmed the new failing-case test fails with the original IndexError
when the fix is reverted (git stash), and passes with the fix applied.

ruff check / ruff format --check clean on both changed files.

Found by reading autocast.py and testing cast_inputs against its own
error-message contract, not from a reported issue.

[bodapatisaikrishna]

@microsoft-github-policy-service agree