CI: add top-priority models to vLLM benchmark

[gyohuangxin]

Refs #2826

Summary

• replace the generic vLLM benchmark matrix with explicit top-priority lanes for openai/gpt-oss-120b, deepseek-ai/DeepSeek-R1-0528, and moonshotai/Kimi-K2.5
• keep the benchmark execution flow close to the existing workflow while allowing model-specific TP, extra args, and Kimi-only env overrides
• align main-branch vLLM CI coverage with the current top-priority benchmark list from issue Increase vLLM downstream test coverage #2826

Test plan

• parse .github/workflows/vllm_benchmark.yaml with python3 and yaml.safe_load
• verify the isolated workflow diff only changes vllm_benchmark.yaml
• run the GitHub Actions workflow end-to-end

[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
ci:atom	ATOM benchmark (DeepSeek-R1 + GPT-OSS)
ci:vllm	vLLM benchmark
ci:all	All of the above

Add labels via the sidebar or gh pr edit 2952 --add-label <label>

[Copilot]

Pull request overview

Updates the vLLM downstream benchmark CI workflow to run a small, explicit set of “top-priority” model lanes with per-model configuration (TP sizing, extra CLI args, and optional env overrides) instead of the previous cartesian matrix.

Changes:

• Replaces the model/kv-cache cartesian matrix with an explicit matrix.include list for three prioritized models.
• Adds per-lane parameters (tp, extra_args, extra_env_args) and wires them into the benchmark invocation.
• Keeps kv-cache dtype handling while letting each lane opt into fp8 KV cache and/or additional flags.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.