Long-Context Sparse V PPL Validation — Raw Results
Date: 2026-03-27
Hardware: Apple M5 Max 128GB

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

c=512 sparse_v=ON:  PPL = 6.1756 +/- 0.33034
c=512 sparse_v=OFF: PPL = 6.1756 +/- 0.33034  (delta: 0.0000)

c=8192 sparse_v=ON:  PPL = 5.5700 +/- 0.07234
c=8192 sparse_v=OFF: PPL = 5.5700 +/- 0.07234  (delta: 0.0000)

c=16384 sparse_v=ON:  PPL = 5.1122 +/- 0.04536
c=16384 sparse_v=OFF: PPL = 5.1122 +/- 0.04536  (delta: 0.0000)

c=32768 sparse_v=ON:  PPL = 6.1293 +/- 0.08156  (2 chunks)
c=32768 sparse_v=OFF: PPL = 6.1293 +/- 0.08156  (2 chunks, delta: 0.0000)

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

c=8192 sparse_v=ON:  PPL = 7.0152 +/- 0.10647
c=8192 sparse_v=OFF: PPL = 7.0152 +/- 0.10647  (delta: 0.0000)

=== SUMMARY ===
All deltas: 0.0000 (bit-for-bit identical)
Tested contexts: 512, 8K, 16K, 32K
Models: MoE 35B-A3B, Dense 27B
KV cache: turbo3 (3.5-bit TurboQuant)
Sparse V threshold: 1e-6

Commands:
  ON:  llama-perplexity -m $MODEL -f wiki.test.raw -c $CTX -ctk turbo3 -ctv turbo3 -fa on --chunks 8 -ngl 99
  OFF: TURBO_SPARSE_V=0 llama-perplexity -m $MODEL -f wiki.test.raw -c $CTX -ctk turbo3 -ctv turbo3 -fa on --chunks 8 -ngl 99
  (32K used --chunks 2)
