Flash attn hotfix (#951)

* use previous arg * use eager to use legacy attention that can be patched
axolotl-ai-cloud · Dec 13, 2023 · f1f60cb · f1f60cb
1 parent 450e04d
commit f1f60cb
Showing 1 changed file with 4 additions and 0 deletions.
diff --git a/src/axolotl/utils/models.py b/src/axolotl/utils/models.py
@@ -324,6 +324,10 @@ def load_model(
                 model_config._attn_implementation = (  # pylint: disable=protected-access
                     "flash_attention_2"
                 )
+            else:
+                model_config._attn_implementation = (  # pylint: disable=protected-access
+                    "eager"
+                )
 
     try:
         if cfg.is_llama_derived_model and not cfg.trust_remote_code and not cfg.gptq: