Tile-Level Sparse V Skip Experiment
Date: 2026-03-27
Hardware: Apple M5 Max 128GB

Hypothesis: Skip entire V tile when max attention weight across all C positions < τ.
Avoids per-position branch overhead for fully-negligible tiles.

Implementation: Added tile_max_w scan before FOR_UNROLL cc loop.
If tile_max_w < 1e-6f, skip entire V accumulation for that tile.
Per-position sparse V still applies inside for mixed tiles.

PPL: 6.1756 (identical to per-position-only sparse V)

=== MoE (Qwen3.5-35B-A3B Q8_0) ===

| Test | Tile skip | Per-position only | Delta |
|------|-----------|-------------------|-------|
| Short (tg128) | 76.78 | 76.11 | +0.9% |
| pp8192+tg128 | 1456.50 | — | — |
| pp16384+tg128 | 1393.42 | — | — |
| pp32768+tg128 | 1120.08 | 1113.77 | +0.6% |

=== Dense (Qwen3.5-27B Q8_0) ===

| Test | Tile skip | Per-position only | Delta |
|------|-----------|-------------------|-------|
| Short (tg128) | 16.77 | 16.73 | +0.2% |
| pp8192+tg128 | 295.94 | 298.27 | -0.8% |
| pp16384+tg128 | 313.51 | 316.98 | -1.1% |

RESULT: Marginal on MoE (+0.6-0.9%), slightly negative on dense (-0.8-1.1%).
The tile max scan loop (8 reads from ss[]) costs more than it saves when most
tiles have at least one non-negligible position.

VERDICT: NOT WORTH KEEPING. Reverted.
