Sparse V quality validation on q8_0 (non-TurboQuant) KV cache
Hardware: Apple M5 Max 128GB
Model: Qwen3.5-35B-A3B Q8_0
Date: 2026-03-27

=== PPL (8-chunk wikitext-2) ===
q8_0 + sparse V:    6.1109 +/- 0.32553
q8_0 (no sparse V): 6.1109 +/- 0.32553
Delta: 0.0000 (identical)

=== NIAH Single Needle (seed 42, depths 0%/50%/100%, contexts 4K/8K/16K) ===

q8_0 + sparse V (7/9):
| Depth  | 4K   | 8K   | 16K  |
|--------|------|------|------|
| 0   %  | PASS | PASS | PASS |
| 50  %  | PASS | PASS | PASS |
| 100 %  | PASS | FAIL | FAIL |

q8_0 without sparse V (7/9):
| Depth  | 4K   | 8K   | 16K  |
|--------|------|------|------|
| 0   %  | PASS | PASS | PASS |
| 50  %  | PASS | PASS | PASS |
| 100 %  | PASS | FAIL | FAIL |

Delta: identical (same 2 failures at 100% depth for 8K and 16K)

=== SUMMARY ===
Sparse V has zero quality or retrieval impact on q8_0.
PPL identical. NIAH identical (same pass/fail pattern).
Combined with +5% decode speedup: pure win.

Raw NIAH JSON: niah_q8_0_sparse_v/ and niah_q8_0_no_sparse_v/
