[Doc] Add documents for multi-node distributed serving with MP backend

[Isotr0py]

Purpose

• Add missing documentation for [V1] Support MP Executor for multi node distributed inference #23691
• Add basic usage introduction to launch multi-node distributed serving with MP backend

Test Plan

Test Result

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[mergify]

Documentation preview: https://vllm--30509.org.readthedocs.build/en/30509/

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Fix multi-node init address for multiprocessing backend

Multi-node serving with the multiprocessing executor still cannot work because the distributed process group is initialized with get_distributed_init_method(get_loopback_ip(), get_open_port()), forcing every node to bind to 127.0.0.1 and a local port instead of the configured master_addr. When following the new multi-node docs (nnodes>1, differing node_ranks), each node forms its own local group and torch.distributed.init_process_group never connects across nodes, so startup will hang/fail. The init method needs to use the shared master address/port for multi-node runs.

Useful? React with 👍 / 👎.

[gemini-code-assist]

Code Review

This pull request adds documentation for using the multiprocessing backend in a multi-node distributed serving setup. My review focuses on the clarity of the new documentation and the correctness of the related code changes. I've found some areas for improvement in the documentation examples to enhance clarity and consistency. More importantly, I've identified a critical issue in the MultiprocExecutor that appears to prevent the new multi-node functionality from working as intended. Please see my detailed comments for suggestions.

[gemini-code-assist]

The distributed_init_method is hardcoded to use get_loopback_ip(), which is only suitable for single-node deployments. For multi-node deployments to work as described in the new documentation, this needs to use the master_addr and master_port from the parallel configuration.

Without this change, the multi-node feature with the multiprocessing backend will fail to initialize the distributed process group correctly across nodes.

Suggested change

	# use the loopback address get_loopback_ip() for communication.
	distributed_init_method = get_distributed_init_method(
	get_loopback_ip(), get_open_port()
	)
	if self.parallel_config.nnodes > 1:
	distributed_init_method = get_distributed_init_method(
	self.parallel_config.master_addr, self.parallel_config.master_port)
	else:
	# use the loopback address get_loopback_ip() for communication.
	distributed_init_method = get_distributed_init_method(
	get_loopback_ip(), get_open_port())

[gemini-code-assist]

For consistency with other examples in this document and for better clarity, I suggest using the long-form arguments and a placeholder for the IP address instead of a hardcoded one.

Suggested change

	vllm serve /path/to/the/model/in/the/container \
	-tp=8 -pp=2 --nnodes 2 --node-rank 0 \
	--master-addr 172.16.98.223
	vllm serve /path/to/the/model/in/the/container \
	--tensor-parallel-size 8 --pipeline-parallel-size 2 --nnodes 2 --node-rank 0 \
	--master-addr <HEAD_NODE_IP>

[gemini-code-assist]

For consistency with other examples in this document and for better clarity, I suggest using the long-form arguments and a placeholder for the IP address instead of a hardcoded one.

Suggested change

	vllm serve /path/to/the/model/in/the/container \
	-tp=8 -pp=2 --nnodes 2 --node-rank 1 \
	--master-addr 172.16.98.223 --headless
	vllm serve /path/to/the/model/in/the/container \
	--tensor-parallel-size 8 --pipeline-parallel-size 2 --nnodes 2 --node-rank 1 \
	--master-addr <HEAD_NODE_IP> --headless

[njhill]

Thank you @Isotr0py! cc @luccafong