[XPU] update dp rank w/o env-var isolation

[zhenwei-intel]

Purpose

The old vLLM worker handles DP on XPU by calling set_device_control_env_var to set ZE_AFFINITY_MASK, so each DP worker only sees a device subset. In practice, this approach can introduce accuracy issues when DP+EP.

Example (ZE_AFFINITY_MASK='2,3'):

• ZE_AFFINITY_MASK='2' for DP0, using torch.device("xpu:0")
• ZE_AFFINITY_MASK='3' for DP1, also using torch.device("xpu:0")

This PR switches XPU worker device assignment to follow the same approach as the GPU worker
where the device is assigned using:
DP_LOCAL_RANK * TP_PP_WORLD_SIZE + TP_LOCAL_RANK

Example (ZE_AFFINITY_MASK='2,3'):

• DP0, using torch.device("xpu:0")
• DP1, using torch.device("xpu:1")

Test Plan

ZE_AFFINITY_MASK=4,5 vllm serve Qwen/Qwen3-30B-A3B --enforce-eager --tensor-parallel-size 1 --data-parallel-size 2 --quantization fp8 -ep --port 8222 --block-size 64 --gpu-memory-utilization 0.8 --max-model-len 8192

curl http://localhost:8222/v1/completions -H "Content-Type: application/json" -d '{"model": "Qwen/Qwen3-30B-A3B", "prompt": "Red H
at is the best company in the world to work for because it works on open source software, which means that all the contributions are delivered to the community. As a result, when w
orking on projects like vLLM we are able to meet many amazing people from various organizations like AMD, Google, NVIDIA, We hold these truths to be self-evident, that all men are
created equal, that they are endowed by their Creator with certain unalienable Rights, that among these are Life, Liberty and the pursuit of Happiness.--That to secure these rights
, Governments are instituted among Men, deriving their just powers from the consent of the governed, ", "max_tokens": 20}'

output:

{"id":"cmpl-bfd8e7a7ecbe4688","object":"text_completion","created":1776220625,"model":"Qwen/Qwen3-30B-A3B","choices":[{"index":0,"text":"1776, Declaration of Independence, United S
tates of America. It's a great place to","logprobs":null,"finish_reason":"length","stop_reason":null,"token_ids":null,"prompt_logprobs":null,"prompt_token_ids":null}],"service_tier
":null,"system_fingerprint":null,"usage":{"prompt_tokens":131,"total_tokens":151,"completion_tokens":20,"prompt_tokens_details":null},"kv_transfer_params":null}⏎

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.

[gemini-code-assist]

Code Review

This pull request introduces XPU support for device control and worker initialization in the V1 engine. It updates vllm/v1/engine/utils.py to bypass device environment isolation for XPU platforms and modifies vllm/v1/worker/xpu_worker.py to calculate the local rank offset in data-parallel scenarios. Feedback was provided regarding a potential bug where modifying self.local_rank might incorrectly affect the LOCAL_RANK environment variable, along with suggestions to remove redundant calls and improve assertion clarity.

[gemini-code-assist]

There are a few issues in this block:

1. Potential Bug with LOCAL_RANK: Modifying self.local_rank to the absolute device index will cause the environment variable LOCAL_RANK to be set incorrectly at line 105 (os.environ["LOCAL_RANK"] = str(self.local_rank)). In a data-parallel setup, LOCAL_RANK is typically expected by distributed backends (like oneCCL) to be the rank relative to the local collective group (e.g., 0 to TP*PP-1), not the absolute device index on the node. Since environment variable isolation is being disabled for XPU, you should ensure that LOCAL_RANK remains the relative rank while only using the offset value for device selection and init_distributed_environment (to match CUDA behavior).
2. Redundancy: torch.accelerator.device_count() is called twice redundantly.
3. Assertion Clarity: The assertion message for the rank check is missing the actual limit, which makes debugging harder.

Suggested change

	self.local_rank += dp_local_rank * tp_pp_world_size
	assert self.local_rank < torch.accelerator.device_count(), (
	f"DP adjusted local rank {self.local_rank} is out of bounds. "
	)
	visible_device_count = torch.accelerator.device_count()
	assert parallel_config.local_world_size <= visible_device_count, (
	f"local_world_size ({parallel_config.local_world_size}) must "
	f"be less than or equal to the number of visible devices "
	f"({visible_device_count})."
	)
	# DP_LOCAL_RANK * TP_PP_WORLD_SIZE + TP_LOCAL_RANK
	self.local_rank += dp_local_rank * tp_pp_world_size
	visible_device_count = torch.accelerator.device_count()
	assert self.local_rank < visible_device_count, (
	f"DP adjusted local rank {self.local_rank} is out of bounds "
	f"(max {visible_device_count})."
	)
	assert parallel_config.local_world_size <= visible_device_count, (
	f"local_world_size ({parallel_config.local_world_size}) must "
	f"be less than or equal to the number of visible devices "
	f"({visible_device_count})."
	)