From 447ffcd32f425872ed36fcbb1acc818f9f8033b8 Mon Sep 17 00:00:00 2001
From: myhloli <moe@myhloli.com>
Date: Tue, 23 Dec 2025 17:00:02 +0800
Subject: [PATCH] refactor: implement dynamic batch ratio based on GPU memory
 and environment variable

---
 mineru/backend/hybrid/hybrid_analyze.py | 50 ++++++++++++++++++++++++-
 1 file changed, 48 insertions(+), 2 deletions(-)

diff --git a/mineru/backend/hybrid/hybrid_analyze.py b/mineru/backend/hybrid/hybrid_analyze.py
index 0a605c1c..014283d0 100644
--- a/mineru/backend/hybrid/hybrid_analyze.py
+++ b/mineru/backend/hybrid/hybrid_analyze.py
@@ -4,6 +4,7 @@ from collections import defaultdict
 
 import cv2
 import numpy as np
+from loguru import logger
 from mineru_vl_utils import MinerUClient
 from mineru_vl_utils.structs import BlockType
 from tqdm import tqdm
@@ -12,8 +13,9 @@ from mineru.backend.hybrid.hybrid_model_output_to_middle_json import result_to_m
 from mineru.backend.pipeline.model_init import HybridModelSingleton
 from mineru.backend.vlm.vlm_analyze import ModelSingleton
 from mineru.data.data_reader_writer import DataWriter
+from mineru.utils.config_reader import get_device
 from mineru.utils.enum_class import ImageType, NotExtractType
-from mineru.utils.model_utils import crop_img
+from mineru.utils.model_utils import crop_img, get_vram, clean_memory
 from mineru.utils.ocr_utils import get_adjusted_mfdetrec_res, get_ocr_result_list, sorted_boxes, merge_det_boxes, \
     update_det_boxes, OcrConfidence
 from mineru.utils.pdf_classify import classify
@@ -203,7 +205,7 @@ def _process_ocr_and_formulas(
     inline_formula_enable,
     _ocr_enable,
     _vlm_ocr_enable,
-    batch_radio: int = 2,
+    batch_radio: int = 1,
 ):
     """处理OCR和公式识别"""
 
@@ -317,6 +319,36 @@ def _normalize_bbox(
                 normalize_poly_to_bbox(ocr_res, page_width, page_height)
 
 
+def get_batch_ratio(device):
+    """
+    根据显存大小或环境变量获取 batch ratio
+    """
+    # 1. 优先尝试从环境变量获取
+    env_val = os.getenv("MINERU_HYBRID_BATCH_RATIO")
+    if env_val:
+        try:
+            batch_ratio = int(env_val)
+            logger.info(f"hybrid batch ratio (from env): {batch_ratio}")
+            return batch_ratio
+        except ValueError as e:
+            logger.warning(f"Invalid MINERU_HYBRID_BATCH_RATIO value: {env_val}, switching to auto mode. Error: {e}")
+
+    # 2. 根据显存自动推断
+    gpu_memory = get_vram(device)
+
+    if gpu_memory >= 24:
+        batch_ratio = 8
+    elif gpu_memory >= 16:
+        batch_ratio = 4
+    elif gpu_memory >= 8:
+        batch_ratio = 2
+    else:
+        batch_ratio = 1
+
+    logger.info(f"hybrid batch ratio (auto, vram={gpu_memory}GB): {batch_ratio}")
+    return batch_ratio
+
+
 def doc_analyze(
     pdf_bytes,
     image_writer: DataWriter | None,
@@ -335,6 +367,9 @@ def doc_analyze(
     images_list, pdf_doc = load_images_from_pdf(pdf_bytes, image_type=ImageType.PIL)
     images_pil_list = [image_dict["img_pil"] for image_dict in images_list]
 
+    device = get_device()
+    batch_ratio = get_batch_ratio(device)
+
     _ocr_enable = ocr_classify(pdf_bytes, parse_method=parse_method)
     _vlm_ocr_enable = False
     if _ocr_enable and language in ["ch", "en"] and inline_formula_enable:
@@ -350,6 +385,7 @@ def doc_analyze(
         inline_formula_enable,
         _ocr_enable,
         _vlm_ocr_enable,
+        batch_radio=batch_ratio,
     )
 
     _normalize_bbox(
@@ -369,6 +405,9 @@ def doc_analyze(
         _vlm_ocr_enable,
         hybrid_pipeline_model,
     )
+
+    clean_memory(device)
+
     return middle_json, results
 
 
@@ -390,6 +429,9 @@ async def aio_doc_analyze(
     images_list, pdf_doc = load_images_from_pdf(pdf_bytes, image_type=ImageType.PIL)
     images_pil_list = [image_dict["img_pil"] for image_dict in images_list]
 
+    device = get_device()
+    batch_ratio = get_batch_ratio(device)
+
     _ocr_enable = ocr_classify(pdf_bytes, parse_method=parse_method)
     _vlm_ocr_enable = False
     if _ocr_enable and language in ["ch", "en"] and inline_formula_enable:
@@ -405,6 +447,7 @@ async def aio_doc_analyze(
         inline_formula_enable,
         _ocr_enable,
         _vlm_ocr_enable,
+        batch_radio=batch_ratio,
     )
 
     _normalize_bbox(
@@ -424,5 +467,8 @@ async def aio_doc_analyze(
         _vlm_ocr_enable,
         hybrid_pipeline_model,
     )
+
+    clean_memory(device)
+
     return middle_json, results