sigilant-runner · Phi-3.5-mini-instruct · A10G · vllm · 4 configs

Config | TPS | TPS p95 | TTFT | TTFT p95 | ITL | PPL | TPS% | TTFT% | PPL% | Score
FP16_BASELINE · ctx:8192 · kv:k16v16 · default  <- best | 53.8 | n/a | 1503.0 | n/a | 18.74 | 2.88 | 100.0 | 100.0 | 100.0 | 100
FP16_BASELINE · ctx:16384 · kv:k8v8 · default | 52.9 | n/a | 1506.1 | n/a | 18.97 | 2.88 | 98.3 | 99.8 | 100.0 | 99
INT8_W8A8 · ctx:8192 · kv:k16v16 · default | 52.9 | n/a | 1513.0 | n/a | 18.96 | 2.88 | 98.3 | 99.3 | 100.0 | 99
INT8_W8A8 · ctx:16384 · kv:k8v8 · default | 52.9 | n/a | 1558.1 | n/a | 18.97 | 2.88 | 98.3 | 96.5 | 100.0 | 99

Best config: FP16_BASELINE · ctx:8192 · kv:k16v16 · default
Auto baseline compare: score Δ=0.00 TPS Δ=0.00 TTFT Δ=0.0ms PPL Δ=0.00
Confidence: target=medium gap_before=1.00% var_before=n/a% replay=False(disabled) gap_after=1.00%

Depth profile (not cross-depth comparable):
  - bucket winners: best_at_8k=FP16_BASELINE · ctx:8192 · kv:k16v16 · default | best_at_14k=FP16_BASELINE · ctx:16384 · kv:k8v8 · default | best_at_28k=n/a
  - 8k: winner=FP16_BASELINE · ctx:8192 · kv:k16v16 · default error=none
  - 14k: winner=FP16_BASELINE · ctx:16384 · kv:k8v8 · default error=none
  - 28k: winner=n/a error=none

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