[None][doc] Add trtllm-bench LWS launch guidance

[zcxGGmu]

Summary

• Add multi-node trtllm-bench guidance for Kubernetes LeaderWorkerSet-style deployments.
• Document the trtllm-llmapi-launch rank-per-GPU launch contract, using the existing Slurm MGMN script as the canonical reference.
• Include a two-node / four-GPU-per-node command sketch and call out gpus_per_node via --config because it is not exposed as a throughput CLI flag.

Test Plan

• Static doc smoke check for the new TOC entry, section, reference script link, and required launch tokens
• git diff --check
• Not run: full Sphinx build (python -m sphinx --version fails locally with No module named sphinx)

Refs #15104

Dev Engineer Review

• Added documentation for multi-node trtllm-bench execution in Kubernetes LeaderWorkerSet deployments.
• Documented the trtllm-llmapi-launch rank-per-GPU launch contract.
• Added dataset and configuration preparation steps.
• Documented MPI/LWS execution, parallelism settings, GPU allocation, shared paths, checkpoint handling, and launcher limitations.
• Added a two-node, four-GPU-per-node command example.
• Clarified that gpus_per_node must be set through --config.
• Added an SPDX header and a Table of Contents entry.
• No code, public API, configuration, or test-list changes were made.
• Validation included static documentation checks and git diff --check.
• A full Sphinx build was not run because Sphinx was not installed locally.

QA Engineer Review

No test changes.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: d13a1a6c-d234-4968-9616-56205b5f3ad6

📥 Commits

Reviewing files that changed from the base of the PR and between cfd1c45 and 0eb74eb.

📒 Files selected for processing (1)

• docs/source/developer-guide/perf-benchmarking.md

🚧 Files skipped from review as they are similar to previous changes (1)

• docs/source/developer-guide/perf-benchmarking.md

---

Walkthrough

The performance benchmarking guide adds instructions for multi-node PyTorch throughput benchmarks using trtllm-llmapi-launch, MPI, and Kubernetes LeaderWorkerSet.

Changes

Multi-node benchmarking

Layer / File(s)	Summary
Multi-node launcher instructions docs/source/developer-guide/perf-benchmarking.md	Adds the license header, table of contents entry, and guidance for multi-node benchmark configuration, GPU and rank allocation, parallelism sizing, shared paths, checkpoints, MPI, and LWS. Includes an eight-rank example and an MPI connectivity fallback.

Estimated code review effort: 1 (Trivial) | ~5 minutes

Suggested reviewers: nv-guomingz, kaiyux

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title follows the required format and clearly states that it adds documentation for trtllm-bench LWS launch guidance.
Description check	✅ Passed	The description explains the changes and test plan, but it omits the template's PR Checklist section.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

Actionable comments posted: 3

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@docs/source/developer-guide/perf-benchmarking.md`:
- Around line 311-313: Update the gpus_per_node guidance in the benchmarking
documentation to describe it as a per-node override: instruct users to pass it
only when the runtime-inferred value from local MPI ranks and visible GPUs
differs from the actual GPUs available per Pod, not when Pod GPUs are fewer than
the total world size.
- Line 255: Add an NVIDIA copyright header with the current year at the
beginning of the Markdown file, before the existing developer-guide content and
heading.
- Around line 303-310: Update the parallel-world invariant in the benchmarking
guide so cp_size=1 requires --tp * --pp to equal the LLM world size, while other
configurations use tp_size * pp_size * cp_size; correct the --tp 8 --pp 2
example to use -np 16 when cp_size=1. Add the NVIDIA copyright header with the
current year at the document’s header.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: 2701101a-1c36-4e1b-a50b-d0bf4ba6a5a4

📥 Commits

Reviewing files that changed from the base of the PR and between 048ae4a and cfd1c45.

📒 Files selected for processing (1)

• docs/source/developer-guide/perf-benchmarking.md