fix(extract-core): don't force flash attention by default

Author: ClemDoum

extract-core/extract_core/docling_.py

@@ -205,9 +205,7 @@ def _resolve_pipeline_options(
     accelerator_opts = getattr(pipeline_options, "accelerator_options", None)
     if accelerator_opts is not None:
         accelerator_opts.device = device.to_docling()
-        if device is Device.CUDA:
-            accelerator_opts.cuda_use_flash_attention2 = True
-        pipeline_options.accelerator_options.device = device.to_docling()
+        pipeline_options.accelerator_options = accelerator_opts
     return pipeline_options
 
 

extract-core/tests/test_docling.py

@@ -58,4 +58,3 @@ def test_docling_format_options_should_resolve_gpu_accelerator() -> None:
     # Then
     accelerator_opts = as_docling.pipeline_options.accelerator_options
     assert accelerator_opts.device is AcceleratorDevice.CUDA
-    assert accelerator_opts.cuda_use_flash_attention2 is True