[Docs] Add GB300 NVFP4 multi-node GLM-5.2 recipe

[Ankur-singh]

Motivation

The GLM-5.2 Deployment selector has no GB300/NVFP4/Low-Latency multi-node cell, so it disables Multi-Nodes. This adds a compact recipe for one TP8 worker spanning two four-GPU GB300 nodes.

Modifications

• Enable GB300 → NVFP4 → Low-Latency → Multi-Nodes with 12 serving flags and two NCCL fabric environment variables. The renderer supplies node count, node rank, and head-node address. Keep NVFP4, EAGLE 5-1-6, concurrency 8, and memory fraction 0.8 explicit.
• Put the four HiCache flags and their 540 GB-per-node host-memory requirement in a collapsed Optional HiCache section. Omit redundant EP1, loading/observability options, and benchmark-specific graph, prefill, context, and environment overrides from the main command.
• Document fabric and node setup. Keep the recipe Not Verified, with no borrowed benchmark results; FP8 multi-node remains unavailable.

Source and qualification

Derived from InferenceX #3659's TP8 C4 recipe at e79239ba, tested in run 36997573694, attempt 2.

That sweep used Dynamo and lmsysorg/sglang:nightly-dev-20260914-4358a161. This simplified standalone command uses the cookbook's existing GB300 image selection (lmsysorg/sglang:latest) and does not reproduce the full benchmark configuration. The standalone adaptation and current image have not been tested on GPUs.

Related #42726 covers broader Dynamo/disaggregated recipes. This patch addresses the existing selector and does not claim to complete all of #3659's recipe requirements.

Accuracy Tests

• Cookbook config check, Mintlify build validation, internal links/anchors/redirects, and whitespace check against the PR base — passed.
• Playwright: Python/Docker output, Single Node → Multi-Nodes, TP8/MTP settings, Playground inheritance without default HiCache, and optional HiCache accordion — passed. No browser console errors.

GPU startup and model evaluation were not run. The recipe remains unverified.

Speed Tests and Profiling

Not run; no benchmark numbers are added.

Checklist

• Validate documentation, config, links, and rendered controls.
• Keep the change declarative; shared rendering engines are unchanged.
• Validate standalone serving on GPUs before marking it verified.

Prepared with Codex (GPT-6); the exact runtime model version was not exposed.

---

CI States

Latest PR Test (Base): ✅ Run #37545675299
Latest PR Test (Extra): ❌ Run #37545675082
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.

[Ankur-singh]

@faradawn Can you please help review this?