[REFACTOR][MANIFEST] drop generative-op carrier-scalar; allow empty signature.inputs

[lcy-seso]

Description

Symptom / Motivation

The current manifest schema enforces len(signature.inputs) >= 1 via test_every_signature_has_inputs_and_outputs. Two reduction-family ops (AlibiFwdOp, SinusoidalFwdOp) synthesize their output entirely from construction-time scalar parameters — they take zero tensor inputs. To satisfy the schema, both manifest entries declare a fake device_carrier 0-dim scalar tensor as their sole input; the carrier exists only as a workaround. This same carrier is also threaded through _register_generative_custom_op and the validator carries a is_generative carve-out that detects the pattern via ref_api == "none" + 0 positional args and skips L1/C4 alignment checks. The hack documents itself in the manifest YAML preamble; the design doc previously enshrined it as well (removed in #1439).

The real defect is one schema invariant being too strict, not a missing "generative op kind". Every op has inputs — they are not always tensors. ALiBi and sinusoidal have integer + dtype params as their real inputs.

Root Cause Analysis

• scripts/validate_manifest.py:test_every_signature_has_inputs_and_outputs requires len(signature.inputs) >= 1. Generative ops physically violate this.
• scripts/validate_manifest.py threads is_generative through check_l1 and L4 (lines around 542, 836-862, 3398-3408) to skip the inputs↔forward positional alignment. The flag is derived from a ref_api == "none" + 0 positional heuristic.
• tileops/ops/elementwise/_base.py:_register_generative_custom_op registers the op with device_carrier: Tensor as its first arg so torch.compile / register_fake has a tensor to trace.
• tileops/ops/elementwise/alibi.py and tileops/ops/elementwise/sinusoidal.py allocate self._device_carrier = torch.empty((), dtype=dtype, device="cuda") in __init__ and pass it to the custom op call in forward().
• tileops/manifest/elementwise_generative.yaml declares device_carrier as the sole input on both entries; output dtype uses same_as(device_carrier) to channel dtype through the carrier.

The cleanup vector is: relax the schema invariant, drop the carrier and its plumbing, simplify the validator.

Related Files

• scripts/validate_manifest.py
• tileops/ops/elementwise/_base.py
• tileops/ops/elementwise/alibi.py
• tileops/ops/elementwise/sinusoidal.py
• tileops/manifest/elementwise_generative.yaml
• tests/test_validate_manifest.py
• tests/ops/test_special_elementwise.py

Goal

Replace the carrier-scalar hack with the minimum-correct implementation that allows signature.inputs: {} for ops whose output is fully derived from construction-time params. After this change, no "generative-op" carve-out concept exists anywhere — the schema is permissive enough that the alignment checks become natural no-ops on empty input lists.

Plan

1. Relax scripts/validate_manifest.py:test_every_signature_has_inputs_and_outputs from len(signature.inputs) >= 1 to len(signature.outputs) >= 1 AND (len(signature.inputs) >= 1 OR len(signature.params) >= 1). Add a test that an op with inputs: {} plus non-empty params is accepted.
2. Remove the is_generative parameter from check_l1 and L4 in scripts/validate_manifest.py. Remove the ref_api == "none" + 0 positional heuristic detection (around lines 836-862 and 3398-3408). The inputs↔forward positional alignment loop becomes a natural no-op when inputs == [].
3. Remove all "Generative-op carve-out" comments from scripts/validate_manifest.py.
4. Update tileops/manifest/elementwise_generative.yaml: change inputs: { device_carrier: ... } to inputs: {} on both entries. Promote dtype from a fake same_as(device_carrier) channel to an explicit params: dtype: {type: torch.dtype} entry. Update output dtype expression accordingly (e.g., same_as(dtype) or direct reference). Delete the file's preamble comment block describing the carve-out.
5. Replace tileops/ops/elementwise/_base.py:_register_generative_custom_op with a tensor-input-less impl. Two acceptable approaches: (a) drop the custom_op registration entirely and have forward() call the kernel directly (eager-only; the ops cannot be torch.compile-graph-captured); or (b) register via torch.library.impl directly without going through dispatcher tracing. Pick (a) for this PR unless graph capture is required by a downstream consumer.
6. Delete self._device_carrier allocation in tileops/ops/elementwise/alibi.py:__init__ and tileops/ops/elementwise/sinusoidal.py:__init__. Simplify forward() to call the kernel directly without passing a carrier.
7. Delete validator tests that assert the carve-out fires (in tests/test_validate_manifest.py); add one test confirming inputs: {} is accepted by the relaxed invariant.
8. Run pytest tests/ops/test_special_elementwise.py to confirm AlibiFwdOp / SinusoidalFwdOp still produce correct output without the carrier path.
9. Run python scripts/validate_manifest.py and pytest tests/test_validate_manifest.py — both pass.

Constraints

• Joint manifest + validator + ops change. The schema relaxation and the impl changes that depend on it land in the same PR — splitting them would leave the manifest mid-state for one round.
• MUST NOT touch other manifest entries that currently use ref_api: "none" with real tensor inputs (the 11 attention.yaml entries and 3 elementwise_fused_gated.yaml entries). They are not affected.
• MUST NOT introduce a new "generative op kind" concept anywhere. The fix is schema relaxation, not categorization.
• MUST keep AlibiFwdOp / SinusoidalFwdOp working end-to-end. tests/ops/test_special_elementwise.py must stay green.
• Dropping torch.compile graph capture for ALiBi / sinusoidal is acceptable in this PR. If a downstream consumer requires it, file a separate follow-up; do not solve it here with another carrier-style placeholder.

Acceptance Criteria

• Modified files pass unit tests (pytest tests/test_validate_manifest.py tests/ops/test_special_elementwise.py, all green on CUDA).
• python scripts/validate_manifest.py exits 0.
• grep -rn 'device_carrier\|_register_generative_custom_op\|is_generative\|Generative-op carve-out\|generative-op carve-out' tileops/ scripts/ tests/ returns no matches (every trace of the hack is gone).
• grep -rn 'signature.inputs' scripts/validate_manifest.py tests/test_validate_manifest.py shows no >= 1 invariant; the relaxed outputs >= 1 AND (inputs >= 1 OR params >= 1) is in place.
• tileops/manifest/elementwise_generative.yaml declares inputs: {} on AlibiFwdOp and SinusoidalFwdOp; dtype is a param; the file preamble has no carve-out narrative.
• A new test test_signature_inputs_may_be_empty_when_params_present (or equivalent) is added to tests/test_validate_manifest.py and passes.