# round 33 track b: decode qsa attention and top-k

## decision: no ship

none of six candidates meets the predeclared acceptance rule: bitwise parity plus a reproducible >=3% attention improvement without material shape regressions. production files were never edited. `qsa_score` was unchanged; this experiment evaluated the proposed attention and top-k changes only.

`/tmp/sol_r33_qsa/sol_r33_qsa.patch` is intentionally empty. it passes `git -C vendor/strata apply --check --allow-empty /tmp/sol_r33_qsa/sol_r33_qsa.patch`; ordinary `apply --check` requires a nonempty patch and is not applicable to a no-change outcome. do not apply the separately labelled rejected patches.

## setup and contract

- local apple m5, as reported by qsa_parity. live server and :8080 untouched; no model benchmarks.
- baseline vendor head: `296b59d2f73c8f04a805df29c2fa97655adace78`, plus current working-tree kernel snapshots captured before edits.
- immutable baseline: `/tmp/sol_r33_qsa/reference/`.
- qsa.metal sha256: `d2d3f311030eae507ab83148afe267b7746a15ac2231aa50aa0a22cc4403db4c`
- qsa2.metal sha256: `1e1937f95e2224ce09944c8cb549af82edc02a0ab38a1cceeb203f4f7b58d59b`
- existing `vendor/strata/build-metal/qsa_parity` executable; no evaluator edits or rebuild needed.
- kernels selected through `STRATA_METAL_KERNELS=/tmp/sol_r33_qsa/strata/metal/kernels`; attention selected with `QSA_ATTN=1` or `2`.
- each benchmark invocation ran serially, normally <0.4 s; broad bitwise invocations each ~1.3 s. no invocation exceeded 10 s.
- timing method: existing paired benchmark, alternating order, median of nine rounds, 16 repeated dispatches; random bitwise mode uses 32 repeated dispatches.
- source snapshots and logs for every explored candidate retained under `/tmp/sol_r33_qsa/`. scratch kernels restored to baseline after experiments.

## candidates

all candidates include the proposed final top-k pass cache (`k_local[4]`, fallback for `per > 4`):

1. `allscale`: precomputed row index, half4 key/value scales preloaded before existing barrier; v1 value pointer/scale offset hoisted.
2. `vsonly`: as above but preload only value scales.
3. `rowonly`: precompute row index, no scale staging.
4. `byteoffset`: store row byte offset rather than row index, no scale staging.
5. `ksonly`: precomputed row index plus key scale staging only.
6. `alias` (parent-requested follow-up): value scales alias dead `qs` storage after the post-q.k barrier, populated before the existing post-softmax barrier. no extra threadgroup storage or barriers. passes paired pos0=3000 ntok=1/2/4/8 on both versions, but attn1 still regresses (n=4: 162.943 / 165.719 us; n=8: 320.544 / 326.474 us), and attn2 regresses at n=8 (379.219 / 381.766 us). this rules out added threadgroup storage as the sole explanation of this regression; no profiler evidence establishes a precise cause.

## correctness

- `allscale`: all 24 paired cases passed full-buffer memcmp (13 buffers), covering both attention versions, pos0=37/2048/3000, ntok=1/2/4/8.
- `allscale`: both random-input bitwise suites passed: **0/40 differences for each version, 80/80 total**. these cover pos0=0,1,63,64,512,2047,2048,2050,8192,30000 and ntok=1/2/4/8; both partial attention and merged output are compared by memcmp.
- other candidates: paired pos0=3000 at ntok=1/4/8 passed for both versions. no equivalence claim beyond these tested shapes.
- final-cache `per > 4` path was not independently exercised for topk; no topk change ships.

## measured results

representative confirmation at pos0=3000; microseconds, reference / candidate:

| candidate | kernel | ntok=1 | ntok=4 | ntok=8 |
|---|---|---:|---:|---:|
| allscale | attn1 | 46.659 / 47.404 | 163.211 / 166.000 | 321.250 / 326.747 |
| allscale | attn2 | 50.935 / 50.180 | 193.630 / 191.810 | 380.570 / 380.727 |
| byteoffset | attn1 | inconclusive clock transition | 162.992 / 162.130 | 322.190 / 320.750 |
| byteoffset | attn2 | 50.802 / 50.852 | 193.391 / 192.997 | 379.760 / 378.534 |

allscale consistently makes attn1 ~1.6–1.8% slower; attn2 gains ~1–2% at small windows but disappear at ntok=8. value-only staging behaves similarly. key-only staging does not establish an improvement. row-only and byte-offset changes yield roughly 0–0.6% gains on stable longer-window measurements, below the planned selection threshold.

top-k differences are small and inconsistent, with occasional much larger apparent clock/thermal transitions affecting both reference and unchanged stages. for example allscale confirmation attn1 topk at n=1/4/8 was 8.250/8.156, 13.677/13.687, 8.289/8.086 us; initial same-shape results did not establish a repeatable win. do not infer a top-k speedup from isolated favorable samples. the baseline same-versus-same smoke was effectively equal (attn1 162.992 / 162.904 us at n=4).

## reproduction and evidence

```
STRATA_METAL_KERNELS=/tmp/sol_r33_qsa/strata/metal/kernels QSA_ATTN=1 \
  vendor/strata/build-metal/qsa_parity --paired /tmp/sol_r33_qsa/reference 3000 --tokens 4
STRATA_METAL_KERNELS=/tmp/sol_r33_qsa/strata/metal/kernels QSA_ATTN=2 \
  vendor/strata/build-metal/qsa_parity --bitwise /tmp/sol_r33_qsa/reference/qsa2.metal
```

copy the two kernels from the named candidate snapshot (`allscale/`, `byteoffset/`, etc.) into scratch `strata/metal/kernels/` first; the final scratch tree deliberately equals baseline. logs: `allscale_av*_p*_n*.log`, `bitwise_av*.log`, `confirm_*.log`, and candidate-specific logs. rejected diffs: `*.rejected.patch`. experiment contract: `plan.md`.

this is a valid negative performance result, not a failed implementation: the suggested transformations preserve tested outputs, but do not justify a production change on this m5.