[Doc] Update DeepSeek V3.1/R1 2P1D doc

[dragondream-chen]

What this PR does / why we need it?

The PR updates the documentation for DeepSeek-V3.1 and DeepSeek-R1 in the scenario of prefill-decode disaggregation.

Updated some PD separation-related setting parameters and optimal configurations. This script has been verified.

• vLLM version: release/v0.13.0
• vLLM main: vllm-project/vllm@bc0a5a0

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

[gemini-code-assist]

Code Review

This pull request updates the documentation for DeepSeek V3.1 and R1 models, primarily focusing on the 2P1D deployment configurations. The changes involve updating command-line parameters, replacing an embedded script with a link for better maintainability, and removing hardcoded version numbers. My review identified a critical issue in the data parallelism configuration for prefill nodes in DeepSeek-V3.1.md. The provided commands are inconsistent and would result in a non-functional deployment. I have included a detailed comment with a suggested fix for this issue. The other documentation changes appear correct and improve clarity.

[gemini-code-assist]

The data parallelism configuration for the prefill nodes (p0, p1) appears to be incorrect and will likely lead to errors.

Here are the inconsistencies:

1. dp-size vs. number of workers: dp-size is set to 2, but with dp-size-local=2 on two nodes (p0 and p1), you are attempting to launch 4 workers in total. The total number of workers should match dp-size.
2. Rank Conflict: dp-rank-start is 0 for both nodes. This will cause data parallel rank conflicts as both nodes will try to launch workers with ranks 0 and 1.
3. DP Master Address: dp-address is different for p0 and p1. For a single data parallel group, all workers should point to the same master address.
4. Connector Config: The kv-transfer-config in the run_dp_template.sh for prefill nodes specifies "dp_size": 2, which is inconsistent with launching 4 workers.

To fix this for a 2-node, 4-worker prefill setup, you should:

• Use dp-size=4.
• Assign unique ranks (e.g., dp-rank-start=0 for p0, dp-rank-start=2 for p1).
• Use a single master dp-address.
• Update the dp_size to 4 in kv-transfer-config in the run_dp_template.sh scripts.

I've provided a code suggestion to correct the launch commands, assuming 141.xx.xx.1 is the master. You will also need to update the dp_size to 4 in the prefill section of kv_connector_extra_config within the run_dp_template.sh scripts for both prefill nodes.

Suggested change

	python launch_online_dp.py --dp-size 2 --tp-size 8 --dp-size-local 2 --dp-rank-start 0 --dp-address 141.xx.xx.1 --dp-rpc-port 12321 --vllm-start-port 7100
	# p1
	python launch_dp_program.py --dp-size 2 --tp-size 8 --dp-size-local 2 --dp-rank-start 0 --dp-address 141.xx.xx.2 --dp-rpc-port 12321 --vllm-start-port 7100
	python launch_online_dp.py --dp-size 2 --tp-size 8 --dp-size-local 2 --dp-rank-start 0 --dp-address 141.xx.xx.2 --dp-rpc-port 12321 --vllm-start-port 7100
	python launch_online_dp.py --dp-size 4 --tp-size 8 --dp-size-local 2 --dp-rank-start 0 --dp-address 141.xx.xx.1 --dp-rpc-port 12321 --vllm-start-port 7100
	# p1
	python launch_online_dp.py --dp-size 4 --tp-size 8 --dp-size-local 2 --dp-rank-start 2 --dp-address 141.xx.xx.1 --dp-rpc-port 12321 --vllm-start-port 7100