Add vLLM production deployment notes

[chris-colinsky]

Summary

Second of five docs-polish items. Extends docs/model-providers/vllm.md with cross-cutting gotchas surfaced in real production work.

Tool-calling section additions:

• --tool-call-parser family table verified against vLLM's docs: Llama 3.x (llama3_json), Llama 4 (llama4_pythonic), Mistral (mistral), Hermes / Qwen 2.5 (hermes), Qwen3 / Qwen3-Coder (qwen3_xml), DeepSeek V3 (deepseek_v3), GPT-OSS (openai)
• Explicit "not supported here" callouts for Anthropic / Gemini (proprietary cloud) and mainstream Gemma 2 / 3 / CodeGemma (no parser ships; only the specialized FunctionGemma 270M has one). Heads off three predictable user confusions.
• Qwen3-VL caveat: vLLM's docs don't currently document a parser for Qwen3-VL specifically; check release notes before assuming the Qwen3 row carries over.

New "Production deployment" H2:

• VLLM_HTTP_TIMEOUT_KEEP_ALIVE: stock 5s uvicorn keep-alive lapses pooled OA-side httpx connections, surfaces as ProviderUnavailable; widen to roughly 300s. Same rule applies behind a reverse proxy.
• systemd unit skeleton: structural pattern, no model-specific paths; uses EnvironmentFile so units ship across hosts.
• Throughput knobs (--max-model-len, --max-num-seqs, --gpu-memory-utilization) framed OA-side: when fan-out concurrency exceeds vLLM's cap, expect ProviderRateLimit.retry_after; wrap the LLM-calling node in RetryMiddleware.

Docs-only; no code or test changes.

Test plan

• uv run pytest -q (1080 pass, no regressions)
• ruff check + ruff format --check clean
• No em dashes in new content (matches the polish convention from PR Add three patterns from graduated agent shapes #111)
• Local mkdocs serve and confirm the new sections render correctly under Model Providers → Self-hosted vLLM

Out of scope

• Items 3-5 of the docs-polish sweep (deep docs sweep, README pointer-block thinning, version on homepage). Separate PRs.
• Sweeping the pre-existing 16 em dashes elsewhere in vllm.md. Falls under item 3's broader docs sweep.

[Copilot]

Pull request overview

Docs-only PR extending docs/model-providers/vllm.md with a tool-call-parser family table (plus Gemma/Gemini/Qwen3-VL caveats) and a new "Production deployment" section covering keep-alive timeouts, a systemd unit skeleton, and throughput knobs.

Changes:

• Replace single-model parser example with a model-family table and explicit caveats for proprietary/Gemma/Qwen3-VL cases.
• Add a "Production deployment" H2 covering VLLM_HTTP_TIMEOUT_KEEP_ALIVE, a systemd unit skeleton, and --max-model-len/--max-num-seqs/--gpu-memory-utilization knobs.
• CHANGELOG entry summarizing the additions.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
docs/model-providers/vllm.md	Adds parser-family table, model-family caveats, and a production deployment section.
CHANGELOG.md	Adds Unreleased Added entry describing the new docs sections.

---

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