docs(examples): add Laguna-S-2.1 on Nebius H100 recipe

[ytsarev]

Description of your changes

Adds a Modelplane example for Poolside's Laguna models: Laguna-S-2.1 served FP8
as a single Standalone engine on one Nebius 8x H100 node (gpu-h100-sxm,
eu-north). Two engine options: vLLM (model-deployment.yaml) and SGLang
(model-deployment-sglang.yaml), which has native Laguna support and returns the
reasoning block in a separate reasoning_content field. Mirrors the kimi-k2 and
glm-4.5-air recipes: InferenceClass, InferenceCluster, ModelCache, both
ModelDeployments, ModelService, plus the recipe page.

Validated end to end on Nebius (TP=8): serving, tool calling, and — on a build
with the Anthropic endpoint picker (#360) — the Anthropic Messages API. Evidence
below.

Reasoning and enable_thinking

Behavior is identical across both engines (vLLM, SGLang) and both endpoints
(/v1/chat/completions and /v1/messages) — always send enable_thinking
explicitly:

enable_thinking	content	reasoning
omitted	a leading </think> leaks into content	—
false	clean	none
true	clean	extracted — OpenAI: reasoning (vLLM) / reasoning_content (SGLang); Anthropic: a native thinking content block

Tool calling is unaffected in all cases. The </think> leak is a Laguna
chat-template artifact — the template opens a think block on the assistant turn,
and the poolside_v1 reasoning parser is passthrough unless thinking is set — so
it is not an engine bug and switching engines does not change it. The fix is
to send enable_thinking explicitly (or set a server-side default on the engine).

Notes:

• [Bug]: Poolside Laguna-S-2.1 poolside_v1 reasoning parser falls back to IdentityReasoningParser and fails auto initialization vllm-project/vllm#49379 (not reproduced here): reports the vLLM poolside_v1
  reasoning parser failing to auto-initialize on v0.25.1 and falling back to a
  passthrough parser. This deployment extracted reasoning fine with per-request
  enable_thinking=true; flagged since it targets the pinned version.
• Anthropic Messages API (/v1/messages): rejected on v0.2.0 (the endpoint
  picker rejects the request shape); works on builds that include feat(docs): Serve the Anthropic Messages API (Claude Code) end to end #360 (EPP v0.9.0
  Anthropic parser) — not yet in a tagged release. Confirmed live on a feat(docs): Serve the Anthropic Messages API (Claude Code) end to end #360 build
  (HTTP 200), see below.

E2E Tested

vLLM — Nebius 8xH100 (gpu-h100-sxm, eu-north), Modelplane v0.2.0, vLLM 0.25.1, TP=8.

# 1. Serving - /v1/chat/completions
curl "$ADDR/v1/chat/completions" -H 'Content-Type: application/json' \
  -d '{"model":"laguna","messages":[{"role":"user","content":"Write a Python function that reverses a string. Output only the code."}]}'
-> content: "...</think>def reverse_string(s):\n    return s[::-1]"
   finish_reason: "stop"   system_fingerprint: "vllm-0.25.1-tp8"

# 2. Tool calling - tools + tool_choice:auto
curl "$ADDR/v1/chat/completions" -H 'Content-Type: application/json' \
  -d '{"model":"laguna","messages":[{"role":"user","content":"What is the weather in Paris? Use the tool."}],
       "tools":[{"type":"function","function":{"name":"get_weather","parameters":{"type":"object","properties":{"city":{"type":"string"}},"required":["city"]}}}],"tool_choice":"auto"}'
-> tool_calls: [{"function":{"name":"get_weather","arguments":"{\"city\": \"Paris\"}"}}]
   finish_reason: "tool_calls"

# 3. Full tool-use loop (feed the tool result back as role:tool)
-> content: "The current weather in Paris is 18C with light rain."   finish_reason: "stop"

# 4. Reasoning - /v1/chat/completions with chat_template_kwargs.enable_thinking=true
-> content: "To find the product of 17 and 23 ... 161 + 230 = 391"  (clean, no </think>)
   reasoning: "Okay, so I need to find the product of 17 and 23. Hmm, let me think..."

SGLang — same node, SGLang v0.5.12.post1-cu129, on a build with the Anthropic endpoint picker (#360), TP=8.

# 1. Anthropic Messages API - /v1/messages
curl "$ADDR/v1/messages" -H 'Content-Type: application/json' -H 'anthropic-version: 2023-06-01' \
  -d '{"model":"laguna","max_tokens":512,"messages":[{"role":"user","content":"Hello!"}]}'
-> HTTP 200  {"type":"message","role":"assistant","content":[{"type":"text","text":"..."}],"stop_reason":"end_turn"}

# 2. Reasoning - /v1/chat/completions with chat_template_kwargs.enable_thinking=true
curl "$ADDR/v1/chat/completions" -H 'Content-Type: application/json' \
  -d '{"model":"laguna","max_tokens":1024,"chat_template_kwargs":{"enable_thinking":true},"messages":[{"role":"user","content":"What is 17*23? Think step by step."}]}'
-> content: "17 * 23 = 391"   reasoning_content: "Okay, let me try to figure out..." (940 reasoning tokens)

# 3. Tool calling - tools + tool_choice:auto
-> tool_calls: [{"function":{"name":"get_weather","arguments":"{\"city\": \"Paris\"}"}}]
   finish_reason: "tool_calls"

I have:

• Read and followed Modelplane's contribution process.
• Run nix flake check (or ./nix.sh flake check) and made sure it passes.
• Added or updated tests covering any composition function changes. Docs and manifests only; no composition function changes.
• Signed off every commit with git commit -s.

[Copilot]

Pull request overview

Adds a new end-to-end documentation example for serving Poolside’s Laguna-S-2.1 FP8 on Nebius (single 8×H100 node) using Modelplane, including the full set of manifests and a corresponding recipe page.

Changes:

• Extend Vale vocabulary to accept “NVLink” and “Laguna*” terms used in the new docs.
• Add a complete “laguna” example manifest set (InferenceClass, InferenceCluster, ModelCache, ModelDeployment, ModelService).
• Add a new docs recipe page that embeds the new manifests in the Examples section.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
docs/utils/vale/styles/config/vocabularies/Modelplane/accept.txt	Adds new accepted terms used by the Laguna docs (NVLink, Laguna*).
docs/manifests/examples/laguna/model-service.yaml	New ModelService manifest for exposing the Laguna deployment via one OpenAI-compatible endpoint.
docs/manifests/examples/laguna/model-deployment.yaml	New ModelDeployment manifest configuring vLLM FP8 serving (TP=8) on a single 8×H100 node.
docs/manifests/examples/laguna/model-cache.yaml	New ModelCache manifest to stage Laguna weights from Hugging Face onto shared RWX storage.
docs/manifests/examples/laguna/inference-cluster-nebius.yaml	New Nebius-backed InferenceCluster manifest for a single 8×H100 node pool.
docs/manifests/examples/laguna/inference-class-nebius.yaml	New Nebius InferenceClass describing the 8×H100 SXM node shape and DRA device properties.
docs/content/examples/laguna.md	New recipe page wiring the manifests into the docs site.

---

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

[Copilot]

Pull request overview

Copilot reviewed 8 out of 8 changed files in this pull request and generated no new comments.

Comments suppressed due to low confidence (2)

docs/manifests/examples/laguna/model-service.yaml:5

• The sample curl command omits chat_template_kwargs.enable_thinking. Per the PR’s own note, omitting it can cause a leading </think> to leak into content, so the copy/paste example will produce degraded output. Update the example to explicitly set enable_thinking:false (or document the difference).

# then call it, e.g.:
#   curl "$ADDR/v1/chat/completions" -H 'Content-Type: application/json' \
#     -d '{"model":"laguna","messages":[{"role":"user","content":"hello"}]}'

docs/content/examples/laguna.md:18

• The recipe page doesn’t mention the enable_thinking gotcha described in the PR (omitting it can leak a leading </think> into content). Adding a short note here will prevent users from thinking the deployment is broken when they do a first test call.

This recipe was run end to end on Nebius (`eu-north`): serving and tool calling
validated on a single 8x H100 node. `poolside/Laguna-S-2.1-FP8` is a public
repository, so no Hugging Face token or Secret is needed. Apply the platform
side first, then the ML side.