Sparse V on f16 (non-quantized) V Path Experiment
Date: 2026-03-27
Hardware: Apple M5 Max 128GB
Model: Qwen3.5-35B-A3B Q8_0 (MoE)

Hypothesis: Apply sparse V skip to the native f16 V path (no dequant).
Saves FMA work for negligible positions even without dequant overhead.

Implementation: Added `if (float(sst[cc*NE]) < 1e-6f) continue;` before
the inner FMA loop in the is_same<vd4_t, v4_t> branch.

PPL: 6.1756 (unchanged)

=== f16 V cache (MoE) ===
| Test | With sparse V | Without | Delta |
|------|-------------|---------|-------|
| Short (tg128) | 85.60 | 85.90 | -0.3% |
| pp32768+tg128 | 1393.67 | 1375.35 | +1.3% |

RESULT: Flat at short context, marginal +1.3% at long context blended.
f16 V path is too cheap (memory load + FMA only) for the branch to pay off.

VERDICT: NOT WORTH KEEPING. Reverted.
