Analyze Document Layout (Donut)

How it fits

Typical backends

Caveats

• I/O contractThe output arm is fixed by the downstream connection once the graph is deployed: the String arm emits the raw decoded sequence, while the [(String, String)] arm emits the parse flattened to dotted key/value pairs whose key set is determined by the checkpoint, not a fixed schema.
• Parameter interactionmodel and task_prompt must come from the same fine-tuned head — the decoder start token selects which parser runs, and a mismatch still produces text but the content is meaningless, with no error raised.
• Accuracymax_length caps the decoded sequence; documents with many fields can be silently truncated at the tail, so raise it when the end of a long parse is missing.
• Hard constraintA zero-size input image short-circuits to an empty result without ever contacting the model; very large or extreme-aspect pages are clamped per axis before encoding to avoid a memory spike.
• Fallbackdevice requesting a GPU on a host without one falls back to CPU, where per-page latency climbs sharply.
• LimitationOne page is parsed per call; multi-page PDFs and tiled contact sheets must be split into individual page images upstream.
• LatencyWeights load once at startup, so first-call latency on a fresh container is dominated by the model download and load rather than inference.