feat: upgrade vLLM to 0.16.0

[vivekkalyan]

Summary

Upgrades vllm from 0.15.1 to 0.16.0

Testing Scope

• Models evaluated:
  • OpenPipe/Qwen3-14B-Instruct
  • Qwen/Qwen3-30B-A3B-Instruct-2507
• Modes evaluated:
  • inference-only
  • strict replay
  • ART-E
• Experiment discipline:
  • H200 single-GPU
  • one fresh VM/cluster per run

Inference-only (single GPU, c=8)

• 14B:
  • throughput: 621.88 -> 620.69 tok/s (-0.19%)
  • latency avg: 1.2766s -> 1.2775s (+0.07%)
• 30B-A3B:
  • 0.16.0: 620.38 tok/s, 1.1545s latency avg
  • 0.15.1: 618.06 tok/s, 1.1585s latency avg
  • +0.38% throughput, -0.34% latency for 0.16.0

ART-E

Task-quality metrics stayed stable across both models; 30B-A3B ART-E showed slight latency/throughput improvement on 0.16.0.

14B:

• latency_mean: 0.384353 -> 0.384383 (+0.0076%)
• latency_p95: 0.877676 -> 0.877771 (+0.0108%)
• completion_tokens_per_sec: 335.3655 -> 335.0151 (-0.10%)

30B-A3B:

• latency_mean: 0.306657 -> 0.300918 (-1.87%)
• latency_p95: 0.740726 -> 0.733300 (-1.00%)
• completion_tokens_per_sec: 428.8369 -> 442.5670 (+3.20%)

Compatibility notes

• Upstream protocol paths changed around reasoning_content; thinking-model flows should be exercised for ART paths that still emit it.
• Tinker renderer paths are Tinker API-specific and not the primary local vLLM path.

[vivekkalyan]

closed by #610