# vLLM NVFP4 with FULL MiaAI recipe (MTP-3 + turboquant_4bit_nc KV + PR#40914 patch)
# RadixArk/Qwen3.8-27B-NVFP4, sm_120 RTX 5090, 2026-08-20
# Recipe source: github.com/MiaAI-Lab/Qwen3.8-27B-NVFP4-RTX-5090 (== our tools/qwen38-nvfp4-vllm)

# --- Single-session recipe as-shipped (max-num-seqs 1, 256K ctx) ---
N=1  agg=147.0  per-req=147.0  p50=3.77s   (554 tok/3.77s)
N=2  agg=147.2  per-req=98.1   p50=7.52s   (serializes - single slot)
N=4  agg=147.1  per-req=58.8   p50=11.29s

# --- Concurrent variant (max-num-seqs 8, 32K ctx, MTP-3 + TQ KV kept) ---
N=1  agg=145.0  per-req=145.0  p50=4.53s
N=2  agg=275.4  per-req=137.8  p50=3.69s
N=4  agg=613.3  per-req=153.4  p50=3.07s
N=8  agg=1080.3 per-req=135.2  p50=3.8s    (4104 tok/3.8s)  <-- sweet spot (=max-num-seqs)
N=12 agg=867.5  per-req=103.0  p95=6.9s    (oversubscribed, 2 waves)
N=16 agg=1053.5 per-req=87.6   p95=7.38s   (oversubscribed, 2 waves)

# Coherence under 8-way load: "the quick brown fox" -> CLEAN (no garble/repetition; #40914 patch holds)
# Unit survived every load level: no crash. The recipe's "MTP+concurrency crashes this KV path"
# note did NOT reproduce on sm_120 Blackwell with the #40914 patch at max-num-seqs 8 / 32K ctx.