Auto-discover InferenceClient endpoints in gr.Workflow
Replace the two hardcoded task tables (`_INFERENCE_ENDPOINT_SCHEMAS`,
`_PIPELINE_TAG_TO_ENDPOINT`) with runtime introspection of the
`InferenceClient` methods. Adds ~50 lines of discovery, removes ~230 of
per-task schemas — net −54 lines — and picks up 3 endpoints that were
missing from the hardcoded table (`audio_to_audio`,
`table_question_answering`, `zero_shot_image_classification`).
Endpoint schemas now derive from method signatures:
- port `id` = param name
- port `label` = titleized param name
- port `type` = name lookup (`image`/`audio`/`video`) or annotation
(`int`/`float` → number, `bool` → boolean, else text). Unwraps
`Optional[X]` so numeric knobs with a default aren't misclassified as text
- port `required` = whether the param has a default
- output `type` = derived from method name (`text_to_image` → image,
`*_classification` → json, else text)
Pipeline-tag → endpoint resolution uses `pipeline_tag.replace("-", "_")`
with a small alias table for tags that don't match that convention
(`text2text-generation`, `conversational`, `text-to-audio`, and the four
vision tags that all route to `chat_completion`).
`chat_completion` stays as a synthetic endpoint (hand-crafted `image +
prompt → text` shape) merged in after auto-discovery, since it doesn't
correspond 1:1 to any InferenceClient method signature. The streaming
dispatch, `max_tokens=16384`, `/gradio_api/file=` handling and
reasoning-aware error messages added in #13666 are preserved.
Also broadens vision routing beyond `image-text-to-text` to also cover
`visual-question-answering`, `document-question-answering`, and
`image-to-text` — none of these are served as task-specific endpoints
by any Inference Provider, so all four now route through
`chat_completion` for consistency.
Zero-day behavior: when huggingface_hub ships a new task method, adds a
param, or changes a param's default, the schema updates automatically
with no code change here. New pipeline_tags that don't match any method
fall through to the raw inference API path (server-side dispatch).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
