Should the HTML build run on a GPU runner (RunsOn) for accurate horse-race timings?

[mmcky]

@Smit-create — a question for you on infra.

Context

The final lecture, schelling_jax_parallel.md, ends with a timed "horse race" between NumPy, JAX-sequential, and JAX-parallel implementations. Its whole conclusion is that the parallel JAX algorithm wins on a GPU ("shines on GPUs", "algorithm structure matters").

The participant experience is fine: the Colab notebooks in notebooks/ are committed unexecuted (no baked-in outputs), so participants generate their own GPU timings live on Colab. ✅

The gap is only in the published website HTML:

• lectures/_config.yml uses execute_notebooks: "cache", so code runs at build time and the printed Converged in … seconds numbers get baked into the rendered site.
• Both ci.yml and publish.yml use runs-on: ubuntu-latest — GitHub-hosted CPU, no GPU.

On CPU the parallel algorithm is expected to be the slowest (it deliberately does uniform/redundant work per agent for SIMT efficiency, and needs more iterations) — so the static site can show timing numbers that contradict the lecture's conclusion.

The question

We have a RunsOn image already defined in .github/runs-on.yml:

images:
  quantecon_ubuntu2404:
    platform: "linux"
    arch: "x64"
    ami: "ami-0edec81935264b6d3"
    region: "us-west-2"

…but the workflows don't reference it — they still use ubuntu-latest.

Smit: do you want this repo's build to run on a GPU instance via RunsOn so the published HTML horse-race timings are accurate? If so:

1. Is the existing quantecon_ubuntu2404 image GPU-capable, or do we need a GPU instance type + a CUDA/GPU-enabled AMI?
2. Which workflow(s) should move to the GPU runner — just publish.yml (the site that gets the baked timings), or ci.yml too?
3. Are the RunsOn credentials/app already enabled on QuantEcon/scipy_tutorial_2026?

Alternatives if GPU CI isn't worth it

• Add a one-line admonition on the lecture page noting the displayed timings are from a CPU build and that Colab GPU gives representative numbers; or
• Leave as-is, since the live workshop runs on Colab GPU and the static numbers are a minor footnote.

Related: #4 (CI currently red on the Netlify preview step — separate infra issue).

[Smit-create]

Thanks for bringing this up @mmcky. I tried initially to use runs-on and use a GPU runner to build the lectures to show accurate timing and optimization that we can get from JAX. But the job was placed into a queue and never ran, see: https://github.com/QuantEcon/scipy_tutorial_2026/actions/runs/22375132817. Please let me know how I can fix the runs-on to use a GPU?

[Smit-create]

Is the existing quantecon_ubuntu2404 image GPU-capable, or do we need a GPU instance type + a CUDA/GPU-enabled AMI?

yeah, this should be okay.

Which workflow(s) should move to the GPU runner — just publish.yml (the site that gets the baked timings), or ci.yml too?

We can just keep it on publish.yml and skip CI.

Are the RunsOn credentials/app already enabled on QuantEcon/scipy_tutorial_2026?

Oh, I see, this is the issue. I think they are not set up for this repo.