Raise a clear error when a no-input model targets below iOS18

[john-rocky]

Summary

A model with no inputs requires Core ML specification version ≥ 9 (iOS18 / macOS15). Below that, the produced mlpackage fails to compile at predict time with:

validator error: Empty input is only valid in specification version >= 9. This model has version 6.

Per review feedback, rather than silently bumping the user's requested deployment target, surface the requirement at conversion time: when an mlprogram has no inputs and the requested target maps to a spec version below iOS18, raise a ValueError that names the requirement and points the user at minimum_deployment_target=ct.target.iOS18.

Test plan

• New test TestTorchExportConversionAPI::test_no_input_model_requires_ios18 covers both paths: the default (iOS15) target raises the actionable ValueError, and an explicit iOS18 target converts and predicts successfully.

Fixes #2578.

[TobyRoseman]

In this case I think we should fail, with a clear error message that telling the user they need to change their deployment target.

[TobyRoseman]

Let's also check the error case here, see my other comment.

[john-rocky]

Thanks — agreed, failing with a clear error is the right behavior. Reworked in 4f65a78:

Behavior change (converter.py)

• No-input + spec version < iOS18 now raises ValueError instead of bumping. Message points the user at the fix:

  Model has no inputs, which requires a deployment target of iOS18 / macOS15 or later (specification version 9); the requested target maps to specification version 6. Re-run ct.convert(...) with minimum_deployment_target=ct.target.iOS18 (or a later target).

• Dropped the now-unused _logger import.

Test change (test_torch_export_conversion_api.py, renamed test_no_input_model_requires_ios18)

• Asserts pytest.raises(ValueError, match=r"minimum_deployment_target=ct\.target\.iOS18") for the default target.
• Then re-runs with minimum_deployment_target=ct.target.iOS18 and verifies spec.specificationVersion >= ct.target.iOS18.value + verify_prediction(mlmodel) for the happy path.

Verified locally on the same toy model (forward(self): return torch.ones(5, 5)): default target raises the new error, iOS18 target converts through MIL cleanly.

[TobyRoseman]

CI: https://gitlab.com/coremltools1/coremltools/-/pipelines/2544917837

[john-rocky]

Reworked per your suggestion: instead of auto-bumping the spec version, conversion now raises a clear ValueError that points the user at minimum_deployment_target=ct.target.iOS18, and the test exercises both the error path and the iOS18 happy path. I've updated the title and description to match.

Verified locally that the default (iOS15) target now raises that error rather than producing a version-6 model that fails to compile with "Empty input is only valid in specification version >= 9".