fix sdp attention to use the flash/mem-efficient context manaager

src/axolotl/monkeypatch/llama_attn_hijack_xformers.py

@@ -184,6 +184,7 @@ def sdp_attention_forward(
 
     # We only apply sdp attention if we don't need to output the whole attention matrix
     if not output_attentions:
+        with torch.backends.cuda.sdp_kernel():
             attn_output = torch.nn.functional.scaled_dot_product_attention(
                 query_states,
                 key_states,