Use dynamic ZeroGPU duration for PDF batches

README.md

@@ -27,9 +27,11 @@ visual-region crops. Model tokens stream into the interface as they are generate
 PDFs are rasterized locally and processed one page at a time in Base mode. Up to
 four pages share one streamed GPU request by default, while each page is still
 inferred independently and sequentially. The browser merges the results, so long
-documents do not have to fit inside one ZeroGPU lease. If a Gradio stream closes
-without an explicit backend error, the browser resumes once from the first
-unfinished page while retaining every completed page. This demo
+documents do not have to fit inside one ZeroGPU lease. GPU duration is estimated
+from the number of pages that actually remain in each group, so a short second
+group does not reserve the same ZeroGPU quota as a four-page group. If a Gradio
+stream closes without an explicit backend error, the browser reconnects from the
+first unfinished page while retaining every completed page. This demo
 intentionally does not expose a separate Long mode.
 
 The five packaged examples cover two financial tables, two Chinese handwriting
@@ -107,7 +109,10 @@ completed pages.
 - `OVISOCR_PDF_RENDER_SCALE`: PDF rasterization scale, defaults to `2.0`.
 - `OVISOCR_STREAM_MIN_CHARS`: minimum new characters between stream updates, defaults to `64`.
 - `OVISOCR_STREAM_MAX_INTERVAL`: maximum seconds between available stream updates, defaults to `0.25`.
-- `OVISOCR_GPU_DURATION`: per-group ZeroGPU reservation in seconds, defaults to `300`.
+- `OVISOCR_GPU_SECONDS_PER_PAGE`: duration-estimation budget, defaults to `30` seconds per page.
+- `OVISOCR_GPU_DURATION_FLOOR`: minimum per-group ZeroGPU reservation, defaults to `45` seconds.
+- `OVISOCR_GPU_DURATION_CEILING`: maximum estimated per-group reservation, defaults to `120` seconds.
+- `OVISOCR_GPU_DURATION`: optional fixed per-group override; unset by default so dynamic duration is used.
 - `OVISOCR_ATTN_IMPLEMENTATION`: defaults to `sdpa`.
 - `OVISOCR_DSW_ID`: generates the Hangzhou DSW proxy root using the active `PORT`.
 - `OVISOCR_ROOT_PATH`: explicit proxy URL or ASGI path prefix; overrides `OVISOCR_DSW_ID`.

app.py

@@ -56,6 +56,12 @@ MAX_PDF_PAGES = int(os.getenv("OVISOCR_MAX_PDF_PAGES", "50"))
 PAGES_PER_GPU_REQUEST = max(
     1, min(5, int(os.getenv("OVISOCR_PAGES_PER_GPU_REQUEST", "4")))
 )
+GPU_SECONDS_PER_PAGE = max(15, int(os.getenv("OVISOCR_GPU_SECONDS_PER_PAGE", "30")))
+GPU_DURATION_FLOOR = max(15, int(os.getenv("OVISOCR_GPU_DURATION_FLOOR", "45")))
+GPU_DURATION_CEILING = max(
+    GPU_DURATION_FLOOR,
+    int(os.getenv("OVISOCR_GPU_DURATION_CEILING", "120")),
+)
 PDF_RENDER_SCALE = float(os.getenv("OVISOCR_PDF_RENDER_SCALE", "2.0"))
 STREAM_MIN_CHARS = int(os.getenv("OVISOCR_STREAM_MIN_CHARS", "64"))
 STREAM_MAX_INTERVAL = float(os.getenv("OVISOCR_STREAM_MAX_INTERVAL", "0.25"))
@@ -461,7 +467,25 @@ def _gpu_duration(
     image_path: FileData | dict[str, Any], page_index: int = 0,
     page_count: int = PAGES_PER_GPU_REQUEST,
 ) -> int:
-    return int(os.getenv("OVISOCR_GPU_DURATION", "300"))
+    configured_duration = os.getenv("OVISOCR_GPU_DURATION", "").strip()
+    if configured_duration:
+        return int(configured_duration)
+
+    requested_count = max(1, min(PAGES_PER_GPU_REQUEST, int(page_count)))
+    try:
+        path = _file_path(image_path)
+        _, total_pages = document_info(path)
+        remaining_pages = max(1, total_pages - int(page_index))
+        requested_count = min(requested_count, remaining_pages)
+    except Exception:
+        # Duration estimation must never prevent a request from reaching the
+        # endpoint. The endpoint performs the authoritative file validation.
+        pass
+
+    return max(
+        GPU_DURATION_FLOOR,
+        min(GPU_DURATION_CEILING, requested_count * GPU_SECONDS_PER_PAGE),
+    )
 
 
 _load_model()
@@ -606,6 +630,9 @@ def healthz() -> JSONResponse:
             "loaded": TEST_MODE or (processor is not None and model is not None),
             "max_pdf_pages": MAX_PDF_PAGES,
             "pages_per_gpu_request": PAGES_PER_GPU_REQUEST,
+            "gpu_seconds_per_page": GPU_SECONDS_PER_PAGE,
+            "gpu_duration_floor": GPU_DURATION_FLOOR,
+            "gpu_duration_ceiling": GPU_DURATION_CEILING,
             "root_path": ROOT_PATH,
             "public_url": PUBLIC_URL,
         }

dist/index.html

@@ -22,7 +22,7 @@
       };
     </script>
     <script defer src="./vendor/mathjax/tex-chtml-full.js"></script>
-    <script type="module" crossorigin src="./assets/index-BuDpKXhW.js"></script>
+    <script type="module" crossorigin src="./assets/index-uj7s2XQX.js"></script>
     <link rel="stylesheet" crossorigin href="./assets/index-CQtDzYS-.css">
   </head>
   <body>