Skip to content

Update vLLM-SR RouterArena submission - #147

Merged
yl231 merged 1 commit into
RouteWorks:mainfrom
Xunzhuo:vllm/vllm-sr-routerarena-refresh
Jun 18, 2026
Merged

Update vLLM-SR RouterArena submission#147
yl231 merged 1 commit into
RouteWorks:mainfrom
Xunzhuo:vllm/vllm-sr-routerarena-refresh

Conversation

@Xunzhuo

@Xunzhuo Xunzhuo commented Jun 17, 2026

Copy link
Copy Markdown
Contributor

Summary

  • Refresh the vLLM-SR RouterArena config and prediction artifacts.
  • Keep deterministic signal/projection routing over a compact model pool.
  • Include full and robustness prediction files, with generated-result entries for the full split.

Validation

  • PYTHONPATH=. python3 router_inference/check_config_prediction_files.py --check-generated-result vllm-sr full
  • PYTHONPATH=. python3 router_inference/check_config_prediction_files.py vllm-sr robustness

Copilot AI review requested due to automatic review settings June 17, 2026 04:05
@Xunzhuo

Xunzhuo commented Jun 17, 2026

Copy link
Copy Markdown
Contributor Author

/evaluate

Copilot AI left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Updates the vLLM semantic router artifacts by re-pointing a subset of stored “prediction” model IDs in the robustness predictions file and refreshing the router config description.

Changes:

  • Swapped the prediction field for multiple benchmark entries in vllm-sr-robustness.json to different model identifiers (e.g., Gemini / DeepSeek / Grok).
  • Updated the human-readable description string in vllm-sr.json.

Reviewed changes

Copilot reviewed 2 out of 3 changed files in this pull request and generated 2 comments.

File Description
router_inference/predictions/vllm-sr-robustness.json Updates many per-sample prediction model identifiers used for robustness runs.
router_inference/config/vllm-sr.json Adjusts the semantic-router config description text.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

"global index": "MMLUPro_computer science_9086",
"prompt": "Please review the subsequent multiple-choice queries and select the answer most likely to be accurate from the provided alternatives.\n\nContext: None\n\nQuestion: To which Internet Security Protocol does Encapsulating Security Payload (ESP) pertain?\n\nOptions: \nA. Secure File Transfer Protocol\nB. Secure IP Protocol\nC. Secure Socket Layer Protocol\nD. Internet Key Exchange Protocol\nE. Transport Layer Security Protocol\nF. Secure Shell Protocol\nG. Secure Http Protocol\nH. Point-to-Point Tunneling Protocol\nI. Hypertext Transfer Protocol Secure\nJ. Network Security Protocol\n\nIndicate the correct letter choice in \\boxed{X}, where X represents the correct letter. Keep any explanation or feedback concise, limited to three sentences.",
"prediction": "grok-4-1-fast-reasoning",
"prediction": "deepseek/deepseek-v4-flash",
Comment thread router_inference/config/vllm-sr.json Outdated
Comment on lines +13 to +16
"deepseek-reasoner"
],
"default_model": "google/gemini-3.1-flash-lite",
"description": "vLLM Semantic Router submission using deterministic semantic, structural, domain-aware, and cost-aware routing over a compact model pool."
"description": "vLLM Semantic Router using deterministic signal and projection routing."
@github-actions

Copy link
Copy Markdown

Router Evaluation Results

Router: vllm-sr
Dataset Split: full

RouterArena Metrics

Metric Value
RouterArena Score 0.7481
Accuracy 76.61%
Total Cost $2.526625
Avg Cost per Query $0.000301
Avg Cost per 1K Queries $0.3008
Number of Queries 8400
Abnormal Entries 0
Robustness Score 0.7143

Optimality Metrics

Metric Value
Opt.Sel (Optimal Selection) 0.1659
Opt.Cost (Cost Efficiency) 0.1809
Opt.Acc (Accuracy vs Optimal) 0.9060

Evaluation completed by RouterArena automated workflow

@Xunzhuo
Xunzhuo force-pushed the vllm/vllm-sr-routerarena-refresh branch from 4d728ff to 4cec1c2 Compare June 18, 2026 03:28
@Xunzhuo

Xunzhuo commented Jun 18, 2026

Copy link
Copy Markdown
Contributor Author

/evaluate

@github-actions

Copy link
Copy Markdown

Router Evaluation Results

Router: vllm-sr
Dataset Split: full

RouterArena Metrics

Metric Value
RouterArena Score 0.7530
Accuracy 77.18%
Total Cost $2.537787
Avg Cost per Query $0.000302
Avg Cost per 1K Queries $0.3021
Number of Queries 8400
Abnormal Entries 0
Robustness Score 0.6762

Optimality Metrics

Metric Value
Opt.Sel (Optimal Selection) 0.1681
Opt.Cost (Cost Efficiency) 0.2510
Opt.Acc (Accuracy vs Optimal) 0.8937

Evaluation completed by RouterArena automated workflow

@yl231
yl231 merged commit e95ba67 into RouteWorks:main Jun 18, 2026
6 checks passed
@Xunzhuo

Xunzhuo commented Jun 20, 2026

Copy link
Copy Markdown
Contributor Author

yl231 added a commit that referenced this pull request Jun 23, 2026
Transcribes the official /evaluate metrics for two merged submissions and
re-ranks the README leaderboard accordingly.

- vLLM-SR (#147): 72.15/73.19/$0.23 -> 75.30/77.18/$0.30, opt 16.81/25.10/89.37,
  robustness 67.62. Reclaims #1 (its resubmission fixed the empty token_usage
  rows from #135, so the previously-zeroed correct answers count again).
- Hybrid Router (#150): new entry at #7 — 72.08/71.38/$0.04, opt 89.87/94.19/92.81,
  robustness 96.67 (@mikemao27).

Also adds Hybrid Router to leaderboard_manifest.yaml so the website sync
(build_site_data.py) includes it; verified the regenerated leaderboard.json and
flip_labels reflect both routers.

Co-authored-by: Louie Lu <yl231@datalab2.cs.rice.edu>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
@yl231

yl231 commented Aug 5, 2026

Copy link
Copy Markdown
Contributor

Heads-up on a leaderboard correction affecting vLLM-SR (#172, closes #166).

xAI retired grok-4-1-fast-reasoning on 2026-05-15 and silently redirects it to grok-4.3 (stronger, ~5–6× pricier). vLLM-SR routed 257 calls to that slug via OpenRouter, confirmed served as grok-4.3 (model_used=x-ai/grok-4.3) but billed at the old price. The evaluator now bills the actually-served model.

Impact: Arena 75.30 → 74.86, cost $0.30 → $0.42/1K — rank #2#3. Nothing is required on your end. If you'd like to re-optimize routing now that grok-4.3's real price is visible, you're welcome to resubmit. Thanks, @Xunzhuo!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants