feat(vllm): upgrade to 0.16.0 with single-GPU validation

[vivekkalyan]

Summary

Upgrades vllm from 0.15.1 to 0.16.0 (via uv) and refreshes uv.lock.

Scope

• Models evaluated:
  • OpenPipe/Qwen3-14B-Instruct
  • Qwen/Qwen3-30B-A3B-Instruct-2507
• Modes evaluated:
  • inference-only
  • strict replay
  • ART-E
• Experiment discipline:
  • uv run sky ...
  • H200 single-GPU
  • one fresh VM/cluster per run
  • non-FP8 focus

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.
• enable_dbo=true is not used in the recommended single-GPU config here (requires DeepEP all2all backend/kernels in this setup).

Benchmark highlights

Inference-only (single GPU, c=8)

• 14B is effectively flat:
  • throughput: 621.88 -> 620.69 tok/s (-0.19%)
  • latency avg: 1.2766s -> 1.2775s (+0.07%)
• 30B default regresses, but tuned config removes it:
  • default: 660.95 -> 549.96 tok/s (-16.79%)
  • tuned best (max_num_batched_tokens):
    • 0.16.0: 620.38 tok/s, 1.1545s latency avg
    • 0.15.1: 618.06 tok/s, 1.1585s latency avg
    • best-vs-best: +0.38% throughput, -0.34% latency for 0.16.0

Replay follow-up (single GPU, c=8)

For 30B replay, defaults were best for both versions.

• 0.15.1 default: 358.66 tok/s, 0.14798s latency mean
• 0.16.0 default: 364.74 tok/s, 0.14284s latency mean
• best-vs-best: +1.70% throughput, -3.47% latency mean for 0.16.0

Forcing max_num_batched_tokens (8192/16384) reduced replay performance on both versions.

ART-E

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

Recommendation

• Proceed with vllm==0.16.0 for single-GPU non-FP8 with mode-specific config:
  • inference-only (30B): set max_num_batched_tokens explicitly (8192 or 16384)
  • replay (30B): keep default server settings (do not force mnbt)
• Do not assume one server config is optimal across modes.

[vivekkalyan]

Superseded by #584 after branch rename to feat/vllm-0.16.0.