up

main.py

@@ -44,7 +44,8 @@ LANGUAGE = os.getenv("LANGUAGE", "english").lower()
 device = get_device(force_cpu=False)
 use_device_map = True if device == "cuda" else False
 try_compile_model = True if device == "cuda" or (device == "mps" and torch.__version__ >= "2.7.0") else False
-try_use_flash_attention = True if device == "cuda" and is_flash_attn_2_available() else False
+try_use_flash_attention = False
+#try_use_flash_attention = True if device == "cuda" and is_flash_attn_2_available() else False
 
 torch_dtype, np_dtype = get_torch_and_np_dtypes(device, use_bfloat16=False)
 logger.info(f"Using device: {device}, torch_dtype: {torch_dtype}, np_dtype: {np_dtype}")