Problem

The inference-providers doc build prerenders an <InferenceSnippet> for every (provider, task) section. The doc-builder generates those snippets via @huggingface/inference's getProviderHelper(provider, task), which throws an InferenceClientInputError when the provider has no helper for that task. The uncaught error surfaces as a SvelteKit 500 and fails the whole build:

Failed to get provider helper for together (automatic-speech-recognition)
  InferenceClientInputError: Task 'automatic-speech-recognition' not supported for provider 'together'.
  Available tasks: text-to-image, conversational, text-generation
TypeError: Cannot read properties of undefined (reading 'length')
Error: 500 /docs/inference-providers/.../providers/together

This happens whenever a provider's partner API (/api/partners/<provider>/models) advertises a live model for a task that @huggingface/inference doesn't (yet) expose as a snippet. The generator happily emits the section, and the build crashes on it.

Concretely, the Automatic Speech Recognition section for together (model nvidia/parakeet-tdt-0.6b-v3) was added by the automated generator in #2507 (2026-05-28), and the Build Inference Providers Documentation workflow has been red on main every day since 2026-05-29 — and on every PR (e.g. #2526) since.

Fix

Guard generate.ts so it never emits a section the doc-builder can't render. We mirror the doc-builder's exact lookup — resolving conversational text-generation / image-text-to-text to the "conversational" task — and skip any (provider, task) combo getProviderHelper rejects, logging each skip so the bot's CI output stays transparent.

The check lives at the single chokepoint fetchWarmModels(), through which all perProviderWarmModels (and therefore both provider pages and task pages) flow.

Verification

Ran npm run generate locally — exits 0, and the guard skips exactly the unsupported combos:

⏭️  Skipping together for automatic-speech-recognition: ...
⏭️  Skipping together for feature-extraction / image-to-image / text-to-video: ...
⏭️  Skipping hyperbolic for image-text-to-text: ...
⏭️  Skipping hf-inference for text-to-video / image-text-to-text: ...

The regenerated together.md no longer contains the ASR section. (hf-inference + non-conversational image-text-to-text is dropped harmlessly — that snippet variant is never rendered; the task page uses the conversational one.)

Note

This PR is code-only. The committed provider/task docs are owned by the automated Update Inference Providers documentation job — once this guard lands, the next regeneration will drop the broken sections and turn the build green. (A one-off regeneration could be committed to unblock immediately if desired.)

Split out of #2526, which is unaffected by its own changes but inherits this pre-existing breakage.

🤖 Generated with Claude Code