flash attn

app.py

@@ -7,6 +7,7 @@ import imagehash
 import cv2
 import os
 import spaces
+import subprocess
 
 from transformers import AutoProcessor, AutoModelForCausalLM
 from transformers.image_utils import to_numpy_array, PILImageResampling, ChannelDimension
@@ -19,6 +20,9 @@ from collections import Counter
 from datasets import load_dataset, concatenate_datasets
 
 
+subprocess.run('pip install flash-attn --no-build-isolation', env={'FLASH_ATTENTION_SKIP_CUDA_BUILD': "TRUE"}, shell=True)
+
+
 DEVICE = torch.device("cuda")
 PROCESSOR = AutoProcessor.from_pretrained(
     "HuggingFaceM4/idefics2_raven_finetuned",