Feat: add OCR's muti-gpus and parallel processing support (#5972)

### What problem does this PR solve?

Add OCR's muti-gpus and parallel processing support

### Type of change

- [x] New Feature (non-breaking change which adds functionality)

@yuzhichang I've tried to resolve the comments in #5697. OCR jobs can
now be done on both CPU and GPU. ( By the way, I've encountered a
“Generate embedding error” issue #5954 that might be due to my outdated
GPUs? idk. ) Please review it and give me suggestions.

GPU:

![gpu_ocr](https://github.com/user-attachments/assets/0ee2ecfb-a665-4e50-8bc7-15941b9cd80e)

![smi](https://github.com/user-attachments/assets/a2312f8c-cf24-443d-bf89-bec50503546d)

CPU:

![cpu_ocr](https://github.com/user-attachments/assets/1ba6bb0b-94df-41ea-be79-790096da4bf1)

deepdoc/parser/pdf_parser.py

@@ -20,6 +20,7 @@ import random
 from timeit import default_timer as timer
 import sys
 import threading
+import trio
 
 import xgboost as xgb
 from io import BytesIO
@@ -41,7 +42,7 @@ if LOCK_KEY_pdfplumber not in sys.modules:
     sys.modules[LOCK_KEY_pdfplumber] = threading.Lock()
 
 class RAGFlowPdfParser:
-    def __init__(self):
+    def __init__(self, parallel_devices: int | None = None):
         """
         If you have trouble downloading HuggingFace models, -_^ this might help!!
 
@@ -53,7 +54,13 @@ class RAGFlowPdfParser:
         ^_-
 
         """
-        self.ocr = OCR()
+        
+        self.ocr = OCR(parallel_devices = parallel_devices)
+        self.parallel_devices = parallel_devices
+        self.parallel_limiter = None
+        if parallel_devices is not None and parallel_devices > 1:
+            self.parallel_limiter = [trio.CapacityLimiter(1) for _ in range(parallel_devices)]
+        
         if hasattr(self, "model_speciess"):
             self.layouter = LayoutRecognizer("layout." + self.model_speciess)
         else:
@@ -63,7 +70,7 @@ class RAGFlowPdfParser:
         self.updown_cnt_mdl = xgb.Booster()
         if not settings.LIGHTEN:
             try:
-                import torch
+                import torch.cuda
                 if torch.cuda.is_available():
                     self.updown_cnt_mdl.set_param({"device": "cuda"})
             except Exception:
@@ -283,9 +290,9 @@ class RAGFlowPdfParser:
                 b["H_right"] = spans[ii]["x1"]
                 b["SP"] = ii
 
-    def __ocr(self, pagenum, img, chars, ZM=3):
+    def __ocr(self, pagenum, img, chars, ZM=3, device_id: int | None = None):
         start = timer()
-        bxs = self.ocr.detect(np.array(img))
+        bxs = self.ocr.detect(np.array(img), device_id)
         logging.info(f"__ocr detecting boxes of a image cost ({timer() - start}s)")
 
         start = timer()
@@ -330,7 +337,7 @@ class RAGFlowPdfParser:
                 b["box_image"] = self.ocr.get_rotate_crop_image(img_np, np.array([[left, top], [right, top], [right, bott], [left, bott]], dtype=np.float32))
                 boxes_to_reg.append(b)
             del b["txt"]
-        texts = self.ocr.recognize_batch([b["box_image"] for b in boxes_to_reg])
+        texts = self.ocr.recognize_batch([b["box_image"] for b in boxes_to_reg], device_id)
         for i in range(len(boxes_to_reg)):
             boxes_to_reg[i]["text"] = texts[i]
             del boxes_to_reg[i]["box_image"]
@@ -1022,28 +1029,54 @@ class RAGFlowPdfParser:
         else:
             self.is_english = False
 
-        start = timer()
-        for i, img in enumerate(self.page_images):
-            chars = self.page_chars[i] if not self.is_english else []
-            self.mean_height.append(
-                np.median(sorted([c["height"] for c in chars])) if chars else 0
-            )
-            self.mean_width.append(
-                np.median(sorted([c["width"] for c in chars])) if chars else 8
-            )
-            self.page_cum_height.append(img.size[1] / zoomin)
+        async def __img_ocr(i, id, img, chars, limiter):
             j = 0
             while j + 1 < len(chars):
                 if chars[j]["text"] and chars[j + 1]["text"] \
                         and re.match(r"[0-9a-zA-Z,.:;!%]+", chars[j]["text"] + chars[j + 1]["text"]) \
                         and chars[j + 1]["x0"] - chars[j]["x1"] >= min(chars[j + 1]["width"],
-                                                                       chars[j]["width"]) / 2:
+                                                                    chars[j]["width"]) / 2:
                     chars[j]["text"] += " "
                 j += 1
 
-            self.__ocr(i + 1, img, chars, zoomin)
+            if limiter:
+                async with limiter:
+                    await trio.to_thread.run_sync(lambda: self.__ocr(i + 1, img, chars, zoomin, id))
+            else:
+                self.__ocr(i + 1, img, chars, zoomin, id)
+                
             if callback and i % 6 == 5:
                 callback(prog=(i + 1) * 0.6 / len(self.page_images), msg="")
+
+        async def __img_ocr_launcher():
+            def __ocr_preprocess():
+                chars = self.page_chars[i] if not self.is_english else []
+                self.mean_height.append(
+                    np.median(sorted([c["height"] for c in chars])) if chars else 0
+                )
+                self.mean_width.append(
+                    np.median(sorted([c["width"] for c in chars])) if chars else 8
+                )
+                self.page_cum_height.append(img.size[1] / zoomin)
+                return chars
+            
+            if self.parallel_limiter:
+                async with trio.open_nursery() as nursery:
+                    for i, img in enumerate(self.page_images):
+                        chars = __ocr_preprocess()
+
+                        nursery.start_soon(__img_ocr, i, i % self.parallel_devices, img, chars,
+                                        self.parallel_limiter[i % self.parallel_devices])
+                        await trio.sleep(0.1)
+            else:
+                for i, img in enumerate(self.page_images):
+                    chars = __ocr_preprocess()
+                    await __img_ocr(i, 0, img, chars, None)
+
+        start = timer()
+        
+        trio.run(__img_ocr_launcher)
+            
         logging.info(f"__images__ {len(self.page_images)} pages cost {timer() - start}s")
 
         if not self.is_english and not any(

deepdoc/vision/ocr.py

@@ -66,10 +66,12 @@ def create_operators(op_param_list, global_config=None):
     return ops
 
 
-def load_model(model_dir, nm):
+def load_model(model_dir, nm, device_id: int | None = None):
     model_file_path = os.path.join(model_dir, nm + ".onnx")
+    model_cached_tag = model_file_path + str(device_id) if device_id is not None else model_file_path
+
     global loaded_models
-    loaded_model = loaded_models.get(model_file_path)
+    loaded_model = loaded_models.get(model_cached_tag)
     if loaded_model:
         logging.info(f"load_model {model_file_path} reuses cached model")
         return loaded_model
@@ -81,7 +83,7 @@ def load_model(model_dir, nm):
     def cuda_is_available():
         try:
             import torch
-            if torch.cuda.is_available():
+            if torch.cuda.is_available() and torch.cuda.device_count() > device_id:
                 return True
         except Exception:
             return False
@@ -98,7 +100,7 @@ def load_model(model_dir, nm):
     run_options = ort.RunOptions()
     if cuda_is_available():
         cuda_provider_options = {
-            "device_id": 0, # Use specific GPU
+            "device_id": device_id, # Use specific GPU
             "gpu_mem_limit": 512 * 1024 * 1024, # Limit gpu memory
             "arena_extend_strategy": "kNextPowerOfTwo",  # gpu memory allocation strategy
         }
@@ -108,7 +110,7 @@ def load_model(model_dir, nm):
             providers=['CUDAExecutionProvider'],
             provider_options=[cuda_provider_options]
             )
-        run_options.add_run_config_entry("memory.enable_memory_arena_shrinkage", "gpu:0")
+        run_options.add_run_config_entry("memory.enable_memory_arena_shrinkage", "gpu:" + str(device_id))
         logging.info(f"load_model {model_file_path} uses GPU")
     else:
         sess = ort.InferenceSession(
@@ -118,12 +120,12 @@ def load_model(model_dir, nm):
         run_options.add_run_config_entry("memory.enable_memory_arena_shrinkage", "cpu")
         logging.info(f"load_model {model_file_path} uses CPU")
     loaded_model = (sess, run_options)
-    loaded_models[model_file_path] = loaded_model
+    loaded_models[model_cached_tag] = loaded_model
     return loaded_model
 
 
 class TextRecognizer:
-    def __init__(self, model_dir):
+    def __init__(self, model_dir, device_id: int | None = None):
         self.rec_image_shape = [int(v) for v in "3, 48, 320".split(",")]
         self.rec_batch_num = 16
         postprocess_params = {
@@ -132,7 +134,7 @@ class TextRecognizer:
             "use_space_char": True
         }
         self.postprocess_op = build_post_process(postprocess_params)
-        self.predictor, self.run_options = load_model(model_dir, 'rec')
+        self.predictor, self.run_options = load_model(model_dir, 'rec', device_id)
         self.input_tensor = self.predictor.get_inputs()[0]
 
     def resize_norm_img(self, img, max_wh_ratio):
@@ -394,7 +396,7 @@ class TextRecognizer:
 
 
 class TextDetector:
-    def __init__(self, model_dir):
+    def __init__(self, model_dir, device_id: int | None = None):
         pre_process_list = [{
             'DetResizeForTest': {
                 'limit_side_len': 960,
@@ -418,7 +420,7 @@ class TextDetector:
                               "unclip_ratio": 1.5, "use_dilation": False, "score_mode": "fast", "box_type": "quad"}
 
         self.postprocess_op = build_post_process(postprocess_params)
-        self.predictor, self.run_options = load_model(model_dir, 'det')
+        self.predictor, self.run_options = load_model(model_dir, 'det', device_id)
         self.input_tensor = self.predictor.get_inputs()[0]
 
         img_h, img_w = self.input_tensor.shape[2:]
@@ -507,7 +509,7 @@ class TextDetector:
 
 
 class OCR:
-    def __init__(self, model_dir=None):
+    def __init__(self, model_dir=None, parallel_devices: int | None = None):
         """
         If you have trouble downloading HuggingFace models, -_^ this might help!!
 
@@ -524,14 +526,33 @@ class OCR:
                 model_dir = os.path.join(
                         get_project_base_directory(),
                         "rag/res/deepdoc")
-                self.text_detector = TextDetector(model_dir)
-                self.text_recognizer = TextRecognizer(model_dir)
+                
+                # Append muti-gpus task to the list
+                if parallel_devices is not None and parallel_devices > 0:
+                    self.text_detector = []
+                    self.text_recognizer = []
+                    for device_id in range(parallel_devices):
+                        self.text_detector.append(TextDetector(model_dir, device_id))
+                        self.text_recognizer.append(TextRecognizer(model_dir, device_id))
+                else:
+                    self.text_detector = [TextDetector(model_dir, 0)]
+                    self.text_recognizer = [TextRecognizer(model_dir, 0)]
+
             except Exception:
                 model_dir = snapshot_download(repo_id="InfiniFlow/deepdoc",
                                               local_dir=os.path.join(get_project_base_directory(), "rag/res/deepdoc"),
                                               local_dir_use_symlinks=False)
-                self.text_detector = TextDetector(model_dir)
-                self.text_recognizer = TextRecognizer(model_dir)
+                
+                if parallel_devices is not None:
+                    assert parallel_devices > 0 , "Number of devices must be >= 1"
+                    self.text_detector = []
+                    self.text_recognizer = []
+                    for device_id in range(parallel_devices):
+                        self.text_detector.append(TextDetector(model_dir, device_id))
+                        self.text_recognizer.append(TextRecognizer(model_dir, device_id))
+                else:
+                    self.text_detector = [TextDetector(model_dir, 0)]
+                    self.text_recognizer = [TextRecognizer(model_dir, 0)]
 
         self.drop_score = 0.5
         self.crop_image_res_index = 0
@@ -593,14 +614,17 @@ class OCR:
                     break
         return _boxes
 
-    def detect(self, img):
+    def detect(self, img, device_id: int | None = None):
+        if device_id is None:
+            device_id = 0
+
         time_dict = {'det': 0, 'rec': 0, 'cls': 0, 'all': 0}
 
         if img is None:
             return None, None, time_dict
 
         start = time.time()
-        dt_boxes, elapse = self.text_detector(img)
+        dt_boxes, elapse = self.text_detector[device_id](img)
         time_dict['det'] = elapse
 
         if dt_boxes is None:
@@ -611,17 +635,22 @@ class OCR:
         return zip(self.sorted_boxes(dt_boxes), [
                    ("", 0) for _ in range(len(dt_boxes))])
 
-    def recognize(self, ori_im, box):
+    def recognize(self, ori_im, box, device_id: int | None = None):
+        if device_id is None:
+            device_id = 0
+
         img_crop = self.get_rotate_crop_image(ori_im, box)
 
-        rec_res, elapse = self.text_recognizer([img_crop])
+        rec_res, elapse = self.text_recognizer[device_id]([img_crop])
         text, score = rec_res[0]
         if score < self.drop_score:
             return ""
         return text
 
-    def recognize_batch(self, img_list):
-        rec_res, elapse = self.text_recognizer(img_list)
+    def recognize_batch(self, img_list, device_id: int | None = None):
+        if device_id is None:
+            device_id = 0
+        rec_res, elapse = self.text_recognizer[device_id](img_list)
         texts = []
         for i in range(len(rec_res)):
             text, score = rec_res[i]
@@ -630,15 +659,17 @@ class OCR:
             texts.append(text)
         return texts
 
-    def __call__(self, img, cls=True):
+    def __call__(self, img, device_id = 0, cls=True):
         time_dict = {'det': 0, 'rec': 0, 'cls': 0, 'all': 0}
+        if device_id is None:
+            device_id = 0
 
         if img is None:
             return None, None, time_dict
 
         start = time.time()
         ori_im = img.copy()
-        dt_boxes, elapse = self.text_detector(img)
+        dt_boxes, elapse = self.text_detector[device_id](img)
         time_dict['det'] = elapse
 
         if dt_boxes is None:
@@ -655,7 +686,7 @@ class OCR:
             img_crop = self.get_rotate_crop_image(ori_im, tmp_box)
             img_crop_list.append(img_crop)
 
-        rec_res, elapse = self.text_recognizer(img_crop_list)
+        rec_res, elapse = self.text_recognizer[device_id](img_crop_list)
 
         time_dict['rec'] = elapse
 

deepdoc/vision/t_ocr.py

@@ -28,14 +28,24 @@ from deepdoc.vision.seeit import draw_box
 from deepdoc.vision import OCR, init_in_out
 import argparse
 import numpy as np
+import trio
 
+# os.environ['CUDA_VISIBLE_DEVICES'] = '0,2' #2 gpus, uncontinuous
+os.environ['CUDA_VISIBLE_DEVICES'] = '0' #1 gpu
+# os.environ['CUDA_VISIBLE_DEVICES'] = '' #cpu
 
 def main(args):
-    ocr = OCR()
+    import torch.cuda
+
+    cuda_devices = torch.cuda.device_count()
+    limiter = [trio.CapacityLimiter(1) for _ in range(cuda_devices)] if cuda_devices > 1 else None
+    ocr = OCR(parallel_devices = cuda_devices)
     images, outputs = init_in_out(args)
 
-    for i, img in enumerate(images):
-        bxs = ocr(np.array(img))
+
+    def __ocr(i, id, img):
+        print("Task {} start".format(i))
+        bxs = ocr(np.array(img), id)
         bxs = [(line[0], line[1][0]) for line in bxs]
         bxs = [{
             "text": t,
@@ -47,6 +57,30 @@ def main(args):
         with open(outputs[i] + ".txt", "w+", encoding='utf-8') as f:
             f.write("\n".join([o["text"] for o in bxs]))
 
+        print("Task {} done".format(i))
+
+    async def __ocr_thread(i, id, img, limiter = None):
+        if limiter:
+            async with limiter:
+                print("Task {} use device {}".format(i, id))
+                await trio.to_thread.run_sync(lambda: __ocr(i, id, img))
+        else:
+            __ocr(i, id, img)
+
+    async def __ocr_launcher():
+        if cuda_devices > 1:
+            async with trio.open_nursery() as nursery:
+                for i, img in enumerate(images):
+                    nursery.start_soon(__ocr_thread, i, i % cuda_devices, img, limiter[i % cuda_devices])
+                    await trio.sleep(0.1)
+        else:
+            for i, img in enumerate(images):
+                await __ocr_thread(i, 0, img)
+
+    trio.run(__ocr_launcher)
+
+    print("OCR tasks are all done")
+
 
 if __name__ == "__main__":
     parser = argparse.ArgumentParser()

rag/app/naive.py

@@ -128,6 +128,9 @@ class Docx(DocxParser):
 
 
 class Pdf(PdfParser):
+    def __init__(self, parallel_devices = None):
+        super().__init__(parallel_devices)
+
     def __call__(self, filename, binary=None, from_page=0,
                  to_page=100000, zoomin=3, callback=None):
         start = timer()
@@ -194,7 +197,7 @@ class Markdown(MarkdownParser):
 
 
 def chunk(filename, binary=None, from_page=0, to_page=100000,
-          lang="Chinese", callback=None, **kwargs):
+          lang="Chinese", parallel_devices=None, callback=None, **kwargs):
     """
         Supported file formats are docx, pdf, excel, txt.
         This method apply the naive ways to chunk files.
@@ -234,7 +237,7 @@ def chunk(filename, binary=None, from_page=0, to_page=100000,
         return res
 
     elif re.search(r"\.pdf$", filename, re.IGNORECASE):
-        pdf_parser = Pdf()
+        pdf_parser = Pdf(parallel_devices)
         if parser_config.get("layout_recognize", "DeepDOC") == "Plain Text":
             pdf_parser = PlainParser()
         sections, tables = pdf_parser(filename if not binary else binary, from_page=from_page, to_page=to_page,

rag/svr/task_executor.py

@@ -100,6 +100,14 @@ MAX_CONCURRENT_CHUNK_BUILDERS = int(os.environ.get('MAX_CONCURRENT_CHUNK_BUILDER
 task_limiter = trio.CapacityLimiter(MAX_CONCURRENT_TASKS)
 chunk_limiter = trio.CapacityLimiter(MAX_CONCURRENT_CHUNK_BUILDERS)
 
+PARALLEL_DEVICES = None
+try:
+    import torch.cuda
+    PARALLEL_DEVICES = torch.cuda.device_count()
+    logging.info(f"found {PARALLEL_DEVICES} gpus")
+except Exception:
+    logging.info("can't import package 'torch'")
+
 # SIGUSR1 handler: start tracemalloc and take snapshot
 def start_tracemalloc_and_snapshot(signum, frame):
     if not tracemalloc.is_tracing():
@@ -241,7 +249,7 @@ async def build_chunks(task, progress_callback):
     try:
         async with chunk_limiter:
             cks = await trio.to_thread.run_sync(lambda: chunker.chunk(task["name"], binary=binary, from_page=task["from_page"],
-                                to_page=task["to_page"], lang=task["language"], callback=progress_callback,
+                                to_page=task["to_page"], lang=task["language"], parallel_devices = PARALLEL_DEVICES, callback=progress_callback,
                                 kb_id=task["kb_id"], parser_config=task["parser_config"], tenant_id=task["tenant_id"]))
         logging.info("Chunking({}) {}/{} done".format(timer() - st, task["location"], task["name"]))
     except TaskCanceledException: