#2221 sm_75: Q2_0 matvec with one warp per four rows (bitwise)

Beschreibung

On Turing the Q2_0 matvec of the decode windows is far from the card's bandwidth: the QSA layers' `attn_q` (2560 → 12288, Q2_0 in the canonical pack) takes 80 µs at three columns, ~110 GB/s on an RTX 2080 Ti. The exact layout gives each row four warps and half a 64-value block per thread, then a cross-warp reduction; at the engine's 2,560 inputs a row has 40 blocks, so 48 of the 128 threads have no chunk and the reduction is most of the kernel.

`q2_0_warp_rows_kernel` gives a warp four rows. Lane `l` runs the chunks of the exact kernel's threads `l`, `32 + l`, `64 + l` and `96 + l`, in their loop order, adds the four sums the way warp 0 adds the other warps' partials, then the same `warp_sum`. Every output is bit for bit `native_q2_0_mmvq_kernel`'s (one column) and `launch_multi`'s exact layout (two to four columns); the upstream (non-exact) layout keeps its kernel.

Default on sm_75 only (the only card I can measure); `STRATA_Q2_WARP=0` keeps the old kernels, `=1` takes this one on any CUDA card.

## Setup

RTX 2080 Ti 22 GB (sm_75), i7-6700, CUDA 13.0, SM 1650 MHz, 240 W. Flash-Next, canonical Q2_0 pack, `--max-context 524288`, `--serve`.

## Kernel (2560 × 12288, cold weights: six copies in rotation, 300 calls)

| columns | exact layout | warp rows | |
|---:|---:|---:|---:|
| 1 | 36.6 µs | 24.8 µs | 1.48× |
| 2 | 59.8 µs | 31.4 µs | 1.90× |
| 3 | 68.3 µs | 31.3 µs | 2.18× |
| 4 | 88.4 µs | 36.6 µs | 2.42× |

The outputs are compared with `memcmp` against the engine's own launch for every column count: identical. One and two rows per warp, and eight warps per block, were measured too; four rows and four warps was best.

## Engine

Together with the QSA attention change of #2220 (both bitwise, measured as one step), 512 new tokens, two runs each: Spanish chat 86.0 → 87.8 t/s, code 103.1 → 106.1, English at 32K 75.3 → 78.4. The same kernel also serves the shared expert's Q2_0 gate/up on its side stream.

Part of the Turing set tracked in #2197.

🤖 Generated with [Claude Code](https://claude.com/claude-code)