ci(workflows): use gpu runner label

[njzjz-bot]

Problem

• deepmodeling has a new dynamic runner using the gpu label.
• CUDA workflows still targeted the old nvidia runner label.

Change

• Update CUDA workflow runs-on labels from nvidia to gpu.

Notes

• Verified with git diff --check and confirmed no runs-on: nvidia remains under .github/workflows.

Authored by OpenClaw (model: custom-chat-jinzhezeng-group/gpt-5.5)

[njzjz-bot]

Current dynamic GPU runner setup

This PR switches GPU jobs from runs-on: nvidia to runs-on: gpu because deepmodeling has deployed the new dynamic runner setup. In this setup, the gpu label is the stable GitHub Actions entry point, and the backend is already based on Actions Runner Controller (ARC).

The current scheme can be summarized as:

ARC + Alibaba Cloud ACK / ECI for dynamically provisioned GPU runners

This is the recommended cloud-native container-based runner architecture when CI jobs can run inside Docker containers.

• Core component: GitHub's officially supported Actions Runner Controller (ARC).
• How it works:
  1. ARC is deployed in an Alibaba Cloud Kubernetes environment, such as ACK.
  2. ARC watches the GitHub Actions job queue through the runner scale set API.
  3. When a CI job using the gpu label enters the queued state, ARC dynamically schedules a fresh runner Pod in the cluster.
  4. After the job finishes, the Pod is automatically released.
• Elastic and cost-efficient setup: with Alibaba Cloud ECI virtual nodes, runner Pods can be backed by serverless container instances instead of pre-reserved ECS worker nodes. This keeps idle cost close to zero: no queued GPU CI jobs means no runner capacity to maintain; when a job arrives, an isolated container instance is created on demand, billed by actual runtime, and destroyed after completion.

So this PR is not introducing a separate future ARC proposal. It is updating the workflows to use the gpu label provided by the current ARC-backed dynamic runner scheme.

---

Authored by OpenClaw (version: 2026.5.28 e932160, model: custom-chat-jinzhezeng-group/gpt-5.5)