# round 21 track b — qsa2

## experiment contract

owner: qsa2 attention and merge kernels; standalone parity evaluator. scratch-only implementation; no server or full-model runs.

hypothesis: fixed-trip unrolling and unconditional full eight-cell v loads reduce attention cost; shared merge weights and the existing prepared-plane writer reduce redundant work. keep chunk layout, sequential fma/reduction order, selection, interfaces and defaults unchanged. reject on any part/xp bit mismatch. compare baseline/candidate on identical seeded inputs in paired rounds; >3% consistent speedup is useful, smaller differences are noise. 40 position/window cases include partial chunks and the 2051-cell budget boundary. short runs only, no port 8080. measure six workloads, two confirmation runs; stop on correctness failure or after two setup repairs.

baseline: snapshot of current dirty vendor tree, saved in /tmp/sol_r21_qsa2; patch relative to those exact starting files, not vendor HEAD.

## implementation and review

- unroll query staging, head/dimension dot loops, accumulator init and stores; hoist query lane pointer.
- split v accumulation into unconditional unrolled eight-cell blocks and scalar tail, preserving sequential fma order.
- merge caches maxima/weights in `mk[128]`/`ek[128]`, computes denominator once, then uses `qsa_write_prepared` from preceding qsa.metal. all barriers remain uniform. threadgroup storage is 1028 bytes versus the old 1024, not a claimed memory reduction; prepared-plane stage itself becomes barrier-free.
- `QSA_ATTN=2 qsa_parity --bitwise REF-qsa2.metal` selects matching 65-chunk/four-group dispatch and allocation; absent that environment variable preserves original v1 mode. baseline/candidate share the unchanged qsa.metal dependency. separate part/xp failure diagnostics; alternate ab/ba timing order.
- reviewed full diff with view_diff, then test setup and merge context with view_code. bytes from qsa_topk2 onward are unchanged. no tracked vendor source was edited.

## build and reproduction

vendor base HEAD: `296b59d2f73c8f04a805df29c2fa97655adace78`, with pre-existing dirty changes. baseline file sha256:

- qsa2.metal: `f8587c2fa8016cc695e101a8c38cb52b535766491c57748e6e6feea9b385b7e1`
- qsa_parity.cpp: `ea7a66b94683790e582afed0f47752bc2d2167973e7b9e84e0939cde11b75c3d`

scratch: `/tmp/sol_r21_qsa2/strata`; originals: `/tmp/sol_r21_qsa2/qsa2.ref.metal` and `qsa.ref.metal`. scratch git index holds starting files; working files hold candidate. no code modifications during any measured run.

```
cmake -S metal -B build-metal -G Ninja -DCMAKE_BUILD_TYPE=Release   -DMETAL_CPP_DIR=/tmp/sol_r21_qsa2/strata/third_party/metal-cpp   -DSTRATA_GGML_DIR=/Users/dawn/proj/swift1.5-qwen3.8-next-flash-ps-scr-qat-persona/vendor/llama.cpp
cmake --build build-metal --target qsa_parity strata_window_cli -j 4
QSA_ATTN=2 /tmp/sol_r21_qsa2/strata/build-metal/qsa_parity --bitwise /tmp/sol_r21_qsa2/qsa2.ref.metal
/tmp/sol_r21_qsa2/strata/build-metal/qsa_parity --bitwise /tmp/sol_r21_qsa2/qsa.ref.metal
```

both build targets passed; cli was built only, never invoked. logs: `/tmp/sol_r21_qsa2/build.log`, `rebuild.log`. device: apple m5. seeded random generator 41, 32768 kv cells, ntok 1/2/4/8; pos0 0,1,63,64,512,2047,2048,2050,8192,30000. compiler runtime default fast_math=false. live server was not touched; external server contention remains a timing caveat.

## measured results

v2 runs 1 and 2: 40/40 cases have exact `memcmp == 0` on **both full part and full xp buffers**, including zeroed inactive regions. final formatting-only evaluator rebuild reran v2: 40/40 again. original v1 mode also passed 40/40. each invocation exited 0 in 2.00s, 1.39s, 1.61s, and 1.21s respectively, all under 5s.

paired kernel times below are run 2 medians, microseconds, 9 alternating paired rounds x 32 dispatches after 50ms warm-up. run 1 independently showed the same improvements.

| pos0 | ntok | attn baseline / candidate | attn reduction | merge baseline / candidate | merge reduction |
|---:|---:|---:|---:|---:|---:|
| 512 | 1 | 37.14 / 22.07 | 40.6% | 6.61 / 6.06 | 8.3% |
| 512 | 4 | 82.67 / 50.49 | 38.9% | 10.83 / 9.25 | 14.6% |
| 2048 | 1 | 82.11 / 50.65 | 38.3% | 14.27 / 12.40 | 13.1% |
| 2048 | 4 | 309.16 / 192.27 | 37.8% | 55.30 / 50.73 | 8.3% |
| 8192 | 1 | 82.21 / 51.04 | 37.9% | 14.30 / 12.45 | 12.9% |
| 8192 | 4 | 309.24 / 195.33 | 36.8% | 55.28 / 50.70 | 8.3% |

logs: `/tmp/sol_r21_qsa2/bitwise-v2-run1.log`, `bitwise-v2-run2.log`, `bitwise-v2-final.log`, `bitwise-v1.log`. final run had a startup outlier at 512/1 attention (68.18/34.13us); remaining workloads agree closely with earlier runs. do not interpret that outlier as additional speedup. no formal confidence intervals were collected.

## decision

accept as a validated local kernel improvement: attention time drops 36.8–40.6%, merge 8.3–14.6% in the confirmation table, with exact outputs on all requested cases. no end-to-end speedup claimed, and v2 is still slower than the existing v1 attention path in these standalone workloads; no default change recommended.

patch: `/tmp/sol_r21_qsa2/sol_r21_qsa2.patch`. `git -C vendor/strata apply --check` passed, exit 0. patch changes only the two owned files and only qsa_attn2/qsa_merge2 in the kernel file (original lines 1–120; added lines shift subsequent numbering).