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>

scripts/quant/plot.py

@@ -0,0 +1,78 @@
+"""Speed-vs-accuracy Pareto scatter and per-method bar charts."""
+from __future__ import annotations
+
+from pathlib import Path
+from typing import Any, Dict, List
+
+import matplotlib
+
+matplotlib.use("Agg")  # headless
+import matplotlib.pyplot as plt  # noqa: E402
+
+
+def pareto_points(rows: List[Dict[str, Any]], x_key: str, y_key: str) -> List[Dict[str, Any]]:
+    points = []
+    for row in rows:
+        if row.get("status") != "ok":
+            continue
+        if row.get(x_key) is None or row.get(y_key) is None:
+            continue
+        points.append({
+            "method": row["method"],
+            "x": row[x_key],
+            "y": row[y_key],
+            "t4_deployable": row.get("t4_deployable"),
+        })
+    return points
+
+
+def _scatter(points, x_label, y_label, title, out_path: Path) -> None:
+    fig, ax = plt.subplots(figsize=(7, 5))
+    for p in points:
+        marker = "o" if p["t4_deployable"] else "x"
+        ax.scatter(p["x"], p["y"], marker=marker, s=80)
+        ax.annotate(p["method"], (p["x"], p["y"]), textcoords="offset points", xytext=(5, 5))
+    ax.set_xlabel(x_label)
+    ax.set_ylabel(y_label)
+    ax.set_title(title + "  (o = T4-deployable, x = A100-only)")
+    fig.tight_layout()
+    fig.savefig(out_path, dpi=120)
+    plt.close(fig)
+
+
+def _bar(rows, metric_key, y_label, out_path: Path) -> None:
+    ok = [r for r in rows if r.get("status") == "ok" and r.get(metric_key) is not None]
+    fig, ax = plt.subplots(figsize=(7, 5))
+    ax.bar([r["method"] for r in ok], [r[metric_key] for r in ok])
+    ax.set_ylabel(y_label)
+    ax.set_title(metric_key)
+    fig.tight_layout()
+    fig.savefig(out_path, dpi=120)
+    plt.close(fig)
+
+
+def render_all(rows: List[Dict[str, Any]], out_dir: Path) -> List[Path]:
+    out_dir = Path(out_dir)
+    out_dir.mkdir(parents=True, exist_ok=True)
+    written: List[Path] = []
+
+    p1 = out_dir / "pareto_latency_cer.png"
+    _scatter(pareto_points(rows, "mean_latency_s", "mean_cer"),
+             "mean latency (s)  [lower=faster]", "CER vs BF16  [lower=better]",
+             "Speed vs recognition accuracy", p1)
+    written.append(p1)
+
+    p2 = out_dir / "pareto_latency_iou.png"
+    _scatter(pareto_points(rows, "mean_latency_s", "mean_bbox_iou"),
+             "mean latency (s)  [lower=faster]", "bbox IoU vs BF16  [higher=better]",
+             "Speed vs detection accuracy", p2)
+    written.append(p2)
+
+    for metric, label in [("mean_latency_s", "mean latency (s)"),
+                          ("mean_cer", "CER vs BF16"),
+                          ("mean_bbox_iou", "bbox IoU vs BF16")]:
+        bp = out_dir / f"bar_{metric}.png"
+        _bar(rows, metric, label, bp)
+        written.append(bp)
+
+    return written