Merge branch 'opendatalab:dev' into dev

Author: myhloli

mineru_vl_utils/mineru_client.py

@@ -121,7 +121,7 @@ def resize_by_need(self, image: Image.Image) -> Image.Image:
             image = new_image
         if min(image.size) < self.min_image_edge:
             scale = self.min_image_edge / min(image.size)
-            new_w, new_h = round(image.width * scale), round(image.height * scale)
+            new_w, new_h = math.ceil(image.width * scale), math.ceil(image.height * scale)
             image = image.resize((new_w, new_h), Image.Resampling.BICUBIC)
         return image
 
@@ -315,6 +315,10 @@ def __init__(
         executor: Executor | None = None,
         batch_size: int = 0,  # for transformers and vllm-engine
         http_timeout: int = 600,  # for http-client backend only
+        connect_timeout: int = 10,  # for http-client backend only
+        max_connections: int | None = None,  # for http-client backend only
+        max_keepalive_connections: int | None = 20,  # for http-client backend only
+        keepalive_expiry: float | None = 5,  # for http-client backend only
         use_tqdm: bool = True,
         debug: bool = False,
         max_retries: int = 3,  # for http-client backend only
@@ -421,6 +425,10 @@ def __init__(
             max_concurrency=max_concurrency,
             batch_size=batch_size,
             http_timeout=http_timeout,
+            connect_timeout=connect_timeout,
+            max_connections=max_connections,
+            max_keepalive_connections=max_keepalive_connections,
+            keepalive_expiry=keepalive_expiry,
             use_tqdm=use_tqdm,
             debug=debug,
             max_retries=max_retries,

mineru_vl_utils/version.py

@@ -1 +1 @@
-__version__ = "0.1.19.1"
+__version__ = "0.1.21"

mineru_vl_utils/vlm_client/base_client.py

@@ -165,12 +165,15 @@ def new_vlm_client(
     max_concurrency: int = 100,
     batch_size: int = 0,
     http_timeout: int = 600,
+    connect_timeout: int = 10,
+    max_connections: int | None = None,
+    max_keepalive_connections: int | None = 20,
+    keepalive_expiry: float | None = 5,
     use_tqdm: bool = True,
     debug: bool = False,
     max_retries: int = 3,
     retry_backoff_factor: float = 0.5,
 ) -> VlmClient:
-
     if backend == "http-client":
         from .http_client import HttpVlmClient
 
@@ -185,6 +188,10 @@ def new_vlm_client(
             allow_truncated_content=allow_truncated_content,
             max_concurrency=max_concurrency,
             http_timeout=http_timeout,
+            connect_timeout=connect_timeout,
+            max_connections=max_connections,
+            max_keepalive_connections=max_keepalive_connections,
+            keepalive_expiry=keepalive_expiry,
             debug=debug,
             max_retries=max_retries,
             retry_backoff_factor=retry_backoff_factor,

mineru_vl_utils/vlm_client/http_client.py

@@ -6,8 +6,8 @@
 
 import httpx
 from httpx_retries import Retry, RetryTransport
-from PIL import Image
 from loguru import logger
+from PIL import Image
 
 from .base_client import (
     DEFAULT_SYSTEM_PROMPT,
@@ -48,6 +48,10 @@ def __init__(
         allow_truncated_content: bool = False,
         max_concurrency: int = 100,
         http_timeout: int = 600,
+        connect_timeout: int = 10,
+        max_connections: int | None = None,
+        max_keepalive_connections: int | None = 20,
+        keepalive_expiry: float | None = 5,
         debug: bool = False,
         max_retries: int = 3,
         retry_backoff_factor: float = 0.5,
@@ -81,6 +85,10 @@ def __init__(
             self.server_headers = server_headers
 
         self.http_timeout = http_timeout
+        self.connect_timeout = connect_timeout
+        self.max_connections = max_connections
+        self.max_keepalive_connections = max_keepalive_connections
+        self.keepalive_expiry = keepalive_expiry
         self.max_retries = max_retries
         self.retry_backoff_factor = retry_backoff_factor
 
@@ -102,23 +110,47 @@ def chat_url(self) -> str:
     def _new_client(self) -> httpx.Client:
         return httpx.Client(
             headers=self.server_headers,
-            timeout=httpx.Timeout(connect=10.0, read=self.http_timeout, write=self.http_timeout, pool=None),
+            timeout=httpx.Timeout(
+                connect=self.connect_timeout,
+                read=self.http_timeout,
+                write=self.http_timeout,
+                pool=None,
+            ),
             transport=RetryTransport(
-                retry=Retry(total=self.max_retries, backoff_factor=self.retry_backoff_factor),
+                retry=Retry(
+                    total=self.max_retries,
+                    backoff_factor=self.retry_backoff_factor,
+                ),
                 transport=httpx.HTTPTransport(
-                    limits=httpx.Limits(max_connections=None, max_keepalive_connections=20),
+                    limits=httpx.Limits(
+                        max_connections=self.max_connections,
+                        max_keepalive_connections=self.max_keepalive_connections,
+                        keepalive_expiry=self.keepalive_expiry,
+                    ),
                 ),
             ),
         )
 
     async def _new_aio_client(self) -> httpx.AsyncClient:
         return httpx.AsyncClient(
             headers=self.server_headers,
-            timeout=httpx.Timeout(connect=10.0, read=self.http_timeout, write=self.http_timeout, pool=None),
+            timeout=httpx.Timeout(
+                connect=self.connect_timeout,
+                read=self.http_timeout,
+                write=self.http_timeout,
+                pool=None,
+            ),
             transport=RetryTransport(
-                retry=Retry(total=self.max_retries, backoff_factor=self.retry_backoff_factor),
+                retry=Retry(
+                    total=self.max_retries,
+                    backoff_factor=self.retry_backoff_factor,
+                ),
                 transport=httpx.AsyncHTTPTransport(
-                    limits=httpx.Limits(max_connections=None, max_keepalive_connections=20),
+                    limits=httpx.Limits(
+                        max_connections=self.max_connections,
+                        max_keepalive_connections=self.max_keepalive_connections,
+                        keepalive_expiry=self.keepalive_expiry,
+                    ),
                 ),
             ),
         )
@@ -290,7 +322,7 @@ def get_response_content(self, response_data: dict) -> str:
         # Set MINERU_VLM_END_TOKEN to override or disable stripping (e.g., set to an empty string).
         end_token = os.getenv("MINERU_VLM_END_TOKEN", "<|im_end|>")
         if end_token and isinstance(content, str) and content.endswith(end_token):
-            content = content[:-len(end_token)]
+            content = content[: -len(end_token)]
         return content or ""
 
     def predict(