sigilant-runner · Phi-3.5-mini-instruct-quantized.w8a8 · A10G · vllm · 3 configs

Config | TPS | TPS p95 | TTFT | TTFT p95 | ITL | PPL | TPS% | TTFT% | PPL% | Score
FP16_BASELINE · ctx:16384 · kv:k16v16 · long  <- best | 36.0 | n/a | 1771.5 | n/a | 28.78 | 3.02 | 100.0 | 100.0 | 100.0 | 100
FP16_BASELINE · ctx:16384 · kv:k8v8 · default | 36.0 | n/a | 2353.9 | n/a | 28.80 | 3.06 | 100.0 | 75.3 | 98.7 | 95
FP16_BASELINE · ctx:32768 · kv:k8v8 · long | 32.3 | n/a | 3066.4 | n/a | 32.07 | 3.08 | 89.7 | 57.8 | 98.1 | 87

Best config: FP16_BASELINE · ctx:16384 · kv:k16v16 · long
Auto baseline compare: score Δ=0.00 TPS Δ=0.00 TTFT Δ=0.0ms PPL Δ=0.00
Confidence: target=medium gap_before=5.00% var_before=n/a% replay=False(disabled) gap_after=5.00%

Depth profile (not cross-depth comparable):
  - bucket winners: best_at_8k=FP16_BASELINE · ctx:16384 · kv:k16v16 · long | best_at_14k=FP16_BASELINE · ctx:16384 · kv:k16v16 · long | best_at_28k=FP16_BASELINE · ctx:32768 · kv:k8v8 · long
  - 8k: winner=FP16_BASELINE · ctx:16384 · kv:k16v16 · long error=none
  - 14k: winner=FP16_BASELINE · ctx:16384 · kv:k16v16 · long error=none
  - 28k: winner=FP16_BASELINE · ctx:32768 · kv:k8v8 · long error=none

PPL is a quality proxy, not production validation.
Full production safety and long-context certification require Sigilant Optimizer.
