docs: add inference benchmarks for Kubernetes and standalone

[danehans]

Company

N/A

Contribution summary

• Add inference benchmark pages for Kubernetes and standalone modes in latest and main.
• Publish the Qwen3-32B/H100 campaign from Add top-level README and publish H100 benchmark reports benchmarks#3.

Checklist

Logo file and listing requirements: not applicable.

• Built the staged changes in an isolated directory with Hugo.
• Verified summary and per-rate measurements against the published reports.
• Checked benchmark pages, overview links, chart assets, and section ordering in both modes and versions.

Preview

Reviewed the pages locally.

[cloudflare-workers-and-pages]

Deploying agentproxy with    Cloudflare Pages

Latest commit:	dd91fc9
Status:	⚡️  Build in progress...

View logs

[danehans]

@artberger PTAL

[danehans]

@kristin-kronstain-brown PTAL

[kristin-kronstain-brown]

Thank you for this, @danehans. The numbers check out: I compared every table value against metrics.csv at 522fc04 and they all match, and the reuse structure across both modes and versions is clean.

Blocking

environment.md line 3: "It uses no EPP or scoring" reads as if the campaign itself uses no EPP, which contradicts the next paragraph. Suggested:

This {{< gloss "Benchmark campaign" >}}campaign{{< /gloss >}} compares EPP-based agentgateway routing with a plain Kubernetes Service across eight vLLM model servers. The Kubernetes Service baseline uses no [EPP](https://llm-d.ai/docs/architecture/core/router#llm-d-endpoint-picker-epp) or [scoring](https://llm-d.ai/docs/architecture/core/router/epp/configuration#plugins) to route inference requests to a model server endpoint.

Questions only you can answer

1. The summary compares at 60 requests/s, where the Service baseline has been saturated at about 6.7 requests/s since 20 QPS. At 10 QPS, the Service has slightly higher output throughput (5,113 vs 4,739). Can you add one sentence that says where the results diverge, so that the TTFT reduction reads as the effect of saturation?
2. failure_rate is 0.0 for every stage in both CSVs. Can the latency caveat state that no requests failed, instead of only asking readers to check?
3. The chart legends say "k8s service (RR)". How does the baseline Service balance load? The prose should match the legend.

Smaller items

• Pin the benchmark links to 522fc04 instead of main, as the README describes.
• Move benchmarking.md into inference/, so that the generated cards list it. You can then remove disableCards, the manual card grid, and the ## Benchmarking section from the four _index.md files. Weight 47 also ties with realtime.md and virtual-models.md.
• Replace {{< callout >}} with GitHub alerts (> [!NOTE], > [!WARNING]), which this repo uses elsewhere.
• Nit: remove "intelligently" in inference.md, and the n/a sentence, because no n/a values appear in the tables.

[danehans]

@danehans thanks for the review. I resoled your #1050 (comment), PTAL.

[kristin-kronstain-brown]

Just one question, but not a deal breaker.

[kristin-kronstain-brown]

Is it confusing to say "Both agentgateway modes" when the topic is centered around one? Could that part of the sentence just be removed?

[danehans]

This is a shared page used by both agentgateway deployment modes. I prefer this being a shared page to avoid future drift.

[kristin-kronstain-brown]

Same question here.

[danehans]

Same response as #1050 (comment).

[danehans]

@kristin-kronstain-brown thanks for the review. I responded to your latest comments. What else is needed to merge the PR?

[kristin-kronstain-brown]

Nothing. Merging.