Suya OCR API — vLLM-backed, OpenAI-compatible OCR service

FastAPI service wrapping the Surya-OCR-2 model (datalab-to) served through vLLM:
legacy /v1/api/ai/* endpoints, an OpenAI-compatible /v1/chat/completions endpoint,
a coalescing request batcher, a local OCR CLI, Docker packaging, multilingual
example outputs, and quantization/concurrency benchmarks.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

surya/ocr_error/__init__.py

@@ -0,0 +1,52 @@
+import math
+from typing import List, Optional
+
+from tqdm import tqdm
+
+from surya.common.predictor import BasePredictor
+from surya.ocr_error.loader import OCRErrorModelLoader
+from surya.ocr_error.model.config import ID2LABEL
+from surya.ocr_error.schema import OCRErrorDetectionResult
+from surya.settings import settings
+
+
+class OCRErrorPredictor(BasePredictor):
+    model_loader_cls = OCRErrorModelLoader
+    batch_size = settings.OCR_ERROR_BATCH_SIZE
+    default_batch_sizes = {"cpu": 8, "mps": 8, "cuda": 64}
+
+    def __call__(self, texts: List[str], batch_size: Optional[int] = None):
+        return self.batch_ocr_error_detection(texts, batch_size)
+
+    def batch_ocr_error_detection(
+        self, texts: List[str], batch_size: Optional[int] = None
+    ):
+        if batch_size is None:
+            batch_size = self.get_batch_size()
+
+        num_batches = math.ceil(len(texts) / batch_size)
+        texts_processed = self.processor(
+            texts, padding="longest", truncation=True, return_tensors="pt"
+        )
+        predictions = []
+        for batch_idx in tqdm(
+            range(num_batches),
+            desc="Running OCR Error Detection",
+            disable=self.disable_tqdm,
+        ):
+            start_idx, end_idx = batch_idx * batch_size, (batch_idx + 1) * batch_size
+            batch_input_ids = texts_processed.input_ids[start_idx:end_idx].to(
+                self.model.device
+            )
+            batch_attention_mask = texts_processed.attention_mask[start_idx:end_idx].to(
+                self.model.device
+            )
+
+            with settings.INFERENCE_MODE():
+                pred = self.model(batch_input_ids, attention_mask=batch_attention_mask)
+                logits = pred.logits.argmax(dim=1).cpu().tolist()
+                predictions.extend(logits)
+
+        return OCRErrorDetectionResult(
+            texts=texts, labels=[ID2LABEL[p] for p in predictions]
+        )