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“Generate embedding error” issue #5954 that might be due to my outdated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)

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: