[Bug][Ops] 3 pre-existing GQA T.tma_copy tilelang-compat failures in tests/ops/ (follow-up to #1121)

[lcy-seso]

Description

Symptom / Motivation

Discovered during PR #1121. While running the full pytest tests/ops/ baseline to verify no regressions, three pre-existing GQA test failures were observed, all rooted in T.tma_copy tilelang-compat issues unrelated to the elementwise op alignment in #1121. They block any future PR from cleanly green-lighting the full tests/ops/ suite, forcing every contributor to scope tests to changed-file subsets.

Root Cause Analysis

The failures predate #1121 — they are a tilelang ↔ TileOPs compatibility gap on the T.tma_copy path used by GQA. They were excluded from #1121's scope because (a) the PR's diff does not touch GQA or T.tma_copy, and (b) fixing them requires a different domain (attention / TMA), not elementwise.

Related Files

To be confirmed by reproduction, but the failing nodes live under:

• tests/ops/test_gqa* (or whichever GQA test files invoke a kernel hitting T.tma_copy)
• The affected GQA op / kernel implementation under tileops/ops/ and tileops/kernels/

Goal

Restore a clean baseline for pytest tests/ops/ so future PRs can verify "no regressions in the full suite" without per-file scoping.

Plan

1. Reproduce: run pytest tests/ops/ -k gqa (or the full tests/ops/) on upstream/main and capture the three failing node IDs and tracebacks.
2. Identify whether the failures are an upstream tilelang change (TMA API shift) or a TileOPs misuse.
3. If tilelang side: pin / patch our adapter for T.tma_copy. If TileOPs side: fix the kernel call.
4. Re-run the full tests/ops/ suite; confirm 0 failures.

Constraints

• Must not regress PR [Feat][Ops] align where/clamp/masked_fill multi-input ops with PyTorch API #1121.
• Must not weaken existing GQA correctness/parity coverage (no xfail / skip as a substitute for fixing).

Acceptance Criteria

• AC-1: The three failing node IDs are listed in the PR description.
• AC-2: pytest tests/ops/ passes end-to-end with no failures and no new skips/xfails.
• AC-3: Root cause is explained in the PR body (tilelang API change vs TileOPs misuse).