test(rocm): ReshapeAndCache->PagedAttention composition at real dims (issue #41)

[VikashLoomba]

Row

BACKEND-ROCM — test-hardening only (one additive cross-device case). Issue #41.

What changed

Adds a cross-device case for the KV-cache composition the in-tree suite doesn't cover: the existing paged-attention case hand-builds a contiguous cache, but the real model path writes KV via ReshapeAndCache and reads it back via PagedAttention. This case is that composition at real Qwen3.5-0.8B dims (Dh=256, Hq=8, Hkv=2, block_size 16), a shuffled block table, and a non-sequential slot mapping — the layout a stride/scatter bug would live in and the contiguous case cannot see.

Surfaced by the #41 Qwen3.5-0.8B divergence investigation: with this composition passing, every piece of the ROCm attention path validates in isolation, which is what localizes the residual divergence to bf16-softmax accumulation rather than a kernel defect (full causal chain in this #41 comment).

Evidence (4× gfx1100, ROCm 7.14, Release)

• New case: 7/7 vs the CPU oracle (runs on ROCm; the composition exercised end to end)
• Full cross-device suite green
• agent-preflight.sh --staged green; check-commit-trailers green

Speed claims

• This PR makes NO speed claim.

Honest gaps

• Test-only; no behavior change. The case validated the path as-is — it did not surface a defect (that was the point: it closes a coverage gap the divergence investigation needed ruled out).

Current-main integration cutoff (2026-08-18)

Rebased from 4a9dcc437cdcc01da36a9b7dd4d73a8d6df5bfb2 to 9afd404441ef45c44e60630c612775c5eb942425 on recorded current-main cutoff 65d6cdaed3e20e9bc70b4f9374fccafefefa7bd0. Exact two-commit replay. Fresh cutoff-head builds, focused gates, full repository preflight, and independent mutation review all pass.

FOLLOWING_AGENTS_PROTOCOL

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: pi:kimi-k3 [pi]
Assisted-by: codex:gpt-5.6-sol [codex]
Assisted-by: pi:gpt-5.6-sol [pi]

[localai-org-maint-bot]

The new composition test does not build a self-consistent logical sequence. ReshapeAndCache writes tokens to (i*7+3)%64, while PagedAttention reads logical positions 0..19 through block table {3,1,2,0} (physical slots 48..63 and 16..19). Most read positions were never written, so CPU/device agreement can compare zeros and miss the intended scatter/layout defect. Please derive each token slot from its logical position and the shuffled block table (for example block_table[i/BS]*BS + i%BS), then add a mutation/guard proving a wrong physical-block mapping makes the test fail. Also remove the unused padded qstride: the tensor is declared contiguous, so those extra uploaded elements do not exercise a fused-view stride.

[VikashLoomba]

Both findings accepted and reworked (commit 8edebbd, rebased onto current main):

1. Slot mapping derives from the block table now (slots[i] = block_table[i/BS]*BS + i%BS) — the engine's real mapping, so the reads actually hit the writes. You're right that the previous (i*7+3)%64 scatter was disjoint from the table and could compare zeros.
2. Anti-vacuity guard added: a corrupted block table (first two logical blocks swapped — both hold real tokens) must change the attention output, proving the composition is load-bearing. (First draft of the guard swapped two blocks outside the logical range and was itself vacuous — the guard caught my guard; the committed version swaps in-range blocks.)
3. The unused padded qstride is removed (the tensor is declared contiguous; the padding exercised nothing).

Gate: test_backend_cross_device 20/20 on gfx1100.

[VikashLoomba]

Current-main rebase and fresh review

Rebased from 10813f0c to 5df0b17c on current-main pin f22c6cc8 and force-updated with an explicit lease.

• Independent range-diff: exact two-commit replay; no base behavior lost.
• Fresh HIP build and full focused binary: 21/21 cases, 351/351 assertions.
• Commit trailers and git diff --check: pass.
• Repository preflight on the immutable head: All gates green.
• Fresh independent integration review: PASS; final tree clean.

The former review was against the superseded head. This immutable head is ready for fresh maintainer review.

[VikashLoomba]

@localai-org-maint-bot Re-review requested on immutable head 5df0b17c78332163dbaf8764ebc8632ad41aaa1b; the prior CHANGES_REQUESTED review targets the superseded pre-rework head. The slot-mapping and anti-vacuity repair, current-main rebase, focused gate, preflight, and fresh independent review all pass.

[VikashLoomba]

Final integration cutoff

Final head: 9afd404441ef45c44e60630c612775c5eb942425, exact replay on cutoff 65d6cdae.

• Fresh HIP gate: 21/21 cases, 351/351 assertions.
• Full repository preflight: All gates green.
• Independent cutoff review: PASS, exact range-diff and clean tree.

@localai-org-maint-bot Fresh review is requested on this immutable head; the existing CHANGES_REQUESTED verdict targets the superseded pre-rework history.

looking good