Merge pull request #46 from datalab-to/dev

Hide imports

chandra/model/hf.py

@@ -1,7 +1,6 @@
 from typing import List
 
 from qwen_vl_utils import process_vision_info
-from transformers import Qwen3VLForConditionalGeneration, Qwen3VLProcessor
 
 from chandra.model.schema import BatchInputItem, GenerationResult
 from chandra.model.util import scale_to_fit
@@ -71,12 +70,15 @@ def process_batch_element(item: BatchInputItem, processor, bbox_scale: int):
 
 
 def load_model():
+    import torch
+    from transformers import Qwen3VLForConditionalGeneration, Qwen3VLProcessor
+
     device_map = "auto"
     if settings.TORCH_DEVICE:
         device_map = {"": settings.TORCH_DEVICE}
 
     kwargs = {
-        "dtype": settings.TORCH_DTYPE,
+        "dtype": torch.bfloat16,
         "device_map": device_map,
     }
     if settings.TORCH_ATTN:

chandra/output.py

@@ -6,7 +6,7 @@ from functools import lru_cache
 
 import six
 from PIL import Image
-from bs4 import BeautifulSoup, NavigableString
+from bs4 import BeautifulSoup
 from markdownify import MarkdownConverter, re_whitespace
 
 from chandra.settings import settings
@@ -89,39 +89,6 @@ def parse_html(
     return out_html
 
 
-def escape_dollars(text):
-    return text.replace("$", r"\$")
-
-
-def get_formatted_table_text(element):
-    text = []
-    for content in element.contents:
-        if content is None:
-            continue
-
-        if isinstance(content, NavigableString):
-            stripped = content.strip()
-            if stripped:
-                text.append(escape_dollars(stripped))
-        elif content.name == "br":
-            text.append("<br>")
-        elif content.name == "math":
-            text.append("$" + content.text + "$")
-        else:
-            content_str = escape_dollars(str(content))
-            text.append(content_str)
-
-    full_text = ""
-    for i, t in enumerate(text):
-        if t == "<br>":
-            full_text += t
-        elif i > 0 and text[i - 1] != "<br>":
-            full_text += " " + t
-        else:
-            full_text += t
-    return full_text
-
-
 class Markdownify(MarkdownConverter):
     def __init__(
         self,

chandra/settings.py

@@ -1,7 +1,5 @@
 from dotenv import find_dotenv
-from pydantic import computed_field
 from pydantic_settings import BaseSettings
-import torch
 import os
 
 
@@ -24,11 +22,6 @@ class Settings(BaseSettings):
     VLLM_GPUS: str = "0"
     MAX_VLLM_RETRIES: int = 6
 
-    @computed_field
-    @property
-    def TORCH_DTYPE(self) -> torch.dtype:
-        return torch.bfloat16
-
     class Config:
         env_file = find_dotenv("local.env")
         extra = "ignore"