webgpu-bench / js /run /config.js
GitHub Actions
sync from abhijitramesh/webgpu-bench@35f91a3bac
4aa88a2
// Public configuration for the one-click bench page.
// Dataset name is NOT a secret. OAuth sign-in uses HF Spaces built-in OAuth
// (declared via `hf_oauth: true` in spaces/README.md) so there's no separate
// OAuth app to register.
// Single source of truth for the consistency-phase prompt. Both the
// interactive Run page (controller.js) and the runner.js-driven harness
// (harness.js) import this so CPU baselines and GPU forced-decode passes
// are always taken against the same input. Plain text on purpose — no
// chat-template wrapping, since this codebase benchmarks 10 different
// model families and each has its own template.
export const CONSISTENCY_PROMPT = 'Explain quantum computing in simple terms.';
export const HF_DATASET_REPO = 'abhijitramesh/webgpu-bench-leaderboard';
// Scopes must match the ones declared in spaces/README.md frontmatter.
// `write-discussions` is the minimum needed to open a community PR against
// the leaderboard dataset on behalf of the signed-in user. `openid` and
// `profile` are auto-included for Space OAuth, so we don't need to ask for
// `read-repos`/`write-repos` (which would grant write to ALL the user's
// repos — far more than needed for a single PR).
export const HF_OAUTH_SCOPES = ['write-discussions'];
export function isHubConfigured() {
// Space-OAuth works whenever the dataset is set. The OAuth client ID is
// injected by HF at Space runtime and read by @huggingface/hub at sign-in.
return HF_DATASET_REPO.length > 0;
}