fix: update flash_attn

app.py

@@ -28,9 +28,15 @@ MAX_PAGES = int(os.getenv("MAX_PAGES", "2"))
 
 # Install poppler-utils
 import os
+import subprocess
 
 os.system("apt-get update")
 os.system("apt-get install --no-install-recommends -y poppler-utils tesseract-ocr")
+subprocess.run(
+    "pip3 install flash-attn --no-build-isolation",
+    env={"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"},
+    shell=True,
+)
 
 
 def convert_document(path, method, start_page=0, enabled=True):

backends/smoldocling.py

@@ -19,7 +19,7 @@ processor = AutoProcessor.from_pretrained("ds4sd/SmolDocling-256M-preview")
 model = AutoModelForVision2Seq.from_pretrained(
     "ds4sd/SmolDocling-256M-preview",
     torch_dtype=torch.bfloat16,
-    _attn_implementation="sdpa",
+    _attn_implementation="flash_attention_2",
 ).to(DEVICE)
 
 # Create input messages

requirements.txt

@@ -5,7 +5,7 @@ unstructured[pdf]
 openai
 img2table
 gmft
-transformers<5.0.0,>=4.45.2
+transformers
 pypdf
 docling_core
 opencv-contrib-python