feat: add GPUDirect support for intel xpu on Dynamo

[Spycsh]

Overview:

Add validated path of gdr for Intel XPU on Dynamo. Also fix related issues.

Details:

Validate running a vLLM 2P2D disaggregation example in Dynamo on Intel XPU, with GPUDirect KV transfer.
router-mode is set to be kv and smart routing is also validated to work correctly from my end.

Summary by CodeRabbit

• Bug Fixes

  • Improved compatibility with different vLLM version structures.

• New Features

  • Added example configuration for multi-process VLLM setup with KV routing, supporting separate prefill and decode worker management on distinct processing units.

_{✏️ Tip: You can customize this high-level summary in your review settings.}

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[github-actions]

👋 Hi Spycsh! Thank you for contributing to ai-dynamo/dynamo.

Just a reminder: The NVIDIA Test Github Validation CI runs an essential subset of the testing framework to quickly catch errors.Your PR reviewers may elect to test the changes comprehensively before approving your changes.

🚀

[coderabbitai]

Walkthrough

The PR introduces a fallback import mechanism for improved vLLM version compatibility and adds a new Bash script for orchestrating a distributed VLLM multi-worker setup with KV routing on XPU devices, using separate decode and prefill workers with GDR configuration.

Changes

Cohort / File(s)	Summary
Import Compatibility components/src/dynamo/vllm/multimodal_utils/chat_processor.py	Adds try/except fallback for AnyTokenizer import, enabling compatibility with different vLLM version structures by falling back to vllm.transformers_utils.tokenizer if vllm.tokenizers raises ImportError.
VLLM Launch Configuration examples/backends/vllm/launch/disagg_router_xpu_gdr.sh	New Bash script orchestrating multi-process VLLM setup with KV routing: starts Dynamo frontend in KV router mode, launches two decode workers and two prefill workers with separate side-channel ports, ZE_AFFINITY_MASK settings, and distinct KV events endpoints.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~12 minutes

Poem

🐰 With imports that fallback and workers aligned,
Our vLLM now routes with split-brain design!
KV caches dance across decode and prefill,
XPU acceleration bends to our will.
A rabbit's delight: when systems just work! 🚀

🚥 Pre-merge checks | ✅ 2 | ❌ 1

❌ Failed checks (1 inconclusive)

Check name	Status	Explanation	Resolution
Description check	❓ Inconclusive	The description addresses the Overview and Details sections but lacks specificity about which related issues are being fixed and does not follow the template's 'Related Issues' and 'Where should the reviewer start' sections.	Complete the 'Related Issues' section with specific GitHub issue numbers using action keywords (Closes/Fixes/Resolves), and add a 'Where should the reviewer start' section highlighting key files to review.

✅ Passed checks (2 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Title check	✅ Passed	The title clearly and concisely describes the main change: adding GPUDirect support for Intel XPU on Dynamo, which aligns with the changeset that introduces a fallback import path and adds a multi-worker vLLM configuration script.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

📝 Coding Plan

• Generate coding plan for human review comments

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

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

Tip

You can customize the tone of the review comments and chat replies.

Configure the tone_instructions setting to customize the tone of the review comments and chat replies. For example, you can set the tone to Act like a strict teacher, Act like a pirate and more.

[coderabbitai]

Actionable comments posted: 1

🤖 Fix all issues with AI agents

In `@examples/backends/vllm/launch/disagg_router_xpu_gdr.sh`:
- Around line 1-6: Remove the leading blank line so the shebang (#!/bin/bash) is
the very first byte of the file, and update the SPDX header year line (the
Copyright comment that currently contains "2025") to include 2026 (e.g.,
"2025-2026") so the copyright check passes; then run shellcheck quickly to
ensure the existing set -e and trap 'echo Cleaning up...; kill 0' EXIT are valid
and adjust quoting/exit handling if shellcheck flags issues.

[github-actions]

This PR is stale because it has been open 30 days with no activity. Remove stale label or comment or this will be closed in 5 days.

[Spycsh]

Hi @rmccorm4 , could you please help review this PR and merge this? This PR gives a simple example on 2P2D disaggregation on Intel XPU using GPUDirect. Dependencies have been resolved and upstreamed in vllm-project/vllm#35270. Thanks!

[Spycsh]

Kindly ping @PeaBrane @GuanLuo @tedzhouhk for review:)

[rmccorm4]

/ok to test 7997cf9

[rmccorm4]

Hi @Spycsh, thanks for this contribution. Can you start an xpu/ folder and move this script there for easier organization?

examples/backends/vllm/launch/xpu/disagg_router_xpu_gdr.sh

[Spycsh]

@rmccorm4 thanks, done.

[rmccorm4]

/ok to test e509793