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

Config | TPS | TPS p95 | TTFT | TTFT p95 | ITL | PPL | TPS% | TTFT% | PPL% | Score
GPTQ4_MARLIN · ctx:16384 · kv:k16v16 · long  <- best | 179.5 | n/a | 137.2 | n/a | 5.59 | 3.09 | 100.0 | 100.0 | 100.0 | 100
GPTQ4_MARLIN · ctx:16384 · kv:k8v8 · default | 179.0 | n/a | 914.3 | n/a | 5.61 | 3.09 | 99.7 | 15.0 | 100.0 | 83
GPTQ4_MARLIN · ctx:8192 · kv:k16v16 · default | 179.4 | n/a | 920.6 | n/a | 5.60 | 3.09 | 99.9 | 14.9 | 100.0 | 83
GPTQ4_MARLIN · ctx:32768 · kv:k8v8 · long | 35.0 | n/a | 1490.5 | n/a | 28.65 | 3.09 | 19.5 | 9.2 | 100.0 | 50

Best config: GPTQ4_MARLIN · 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=17.00% var_before=n/a% replay=False(disabled) gap_after=17.00%

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