Using flash_attention_2?

demo/app_januspro.py

@@ -10,7 +10,9 @@ from transformers import AutoConfig, AutoModelForCausalLM
 model_path = "deepseek-ai/Janus-Pro-7B"
 config = AutoConfig.from_pretrained(model_path)
 language_config = config.language_config
-language_config._attn_implementation = "eager"
+# language_config._attn_implementation = "eager"
+language_config["attn_implementation"] = "flash_attention_2"
+
 vl_gpt = AutoModelForCausalLM.from_pretrained(
     model_path, language_config=language_config, trust_remote_code=True
 )