CI: run vLLM disagg from upstream main

[JiaoliangYu]

Summary

• replace the itej89/vllm CI fork dependency with vllm-project/vllm@main
• temporarily fetch and cherry-pick vLLM PR [ROCm][DistInf] Enable vLLM DI CI with buildkite/slurm vllm-project/vllm#47030 until it merges
• reuse the single-node vLLM CI pattern to build the current AITER wheel in rocm/vllm-dev:nightly
• stage that wheel with the upstream disaggregated-inference scripts and load it from a writable overlay in every compute-node container
• derive the execution matrix directly from vLLM PR #47030's pipeline-disagg.yaml; AITER only filters for:
  • DeepSeek-V3 and DeepSeek-R1-MXFP4
  • 1P1D TP8 and 2P2D TP8
  • vllm-router
• run the selected upstream commands serially so the fixed Spur nodes are not shared by concurrent cases

AITER does not duplicate the upstream model flags, topology parameters, MoRIIO settings, router settings, or time limits. It only selects the desired upstream cases, overlays the current AITER wheel, and maps the upstream 2-node/4-node request onto the AITER Spur runners.

PR execution is explicitly gated by the ci:vllm-di or ci:all label.

Validation

• Actionlint 1.7.7
• git diff --check
• verified vllm-project/vllm@main + refs/pull/47030/head cherry-picks cleanly when the PR parent is fetched
• verified the AITER overlay hook leaves the upstream launcher shell syntax valid
• the live selector job successfully applied #47030 and derived the four expected upstream cases: https://github.com/ROCm/aiter/actions/runs/31148637525/job/92773486262
• full live run: https://github.com/ROCm/aiter/actions/runs/31148637525

Follow-up

Remove the temporary PR fetch/cherry-pick after vLLM PR #47030 merges.

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
ci:sglang	SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom	ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full	ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm	vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all	All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 4619 --add-label <label>