disable memory-efficient attention in loopgen's own magnet loader

src/loopgen.py

@@ -40,10 +40,11 @@ def load_lm_model_loopgen(file_or_url_or_id: tp.Union[Path, str], compression_mo
     cfg.transformer_lm.span_len = cfg.masking.span_len
 
     # MAGNeT models v1 support only xformers backend.
-    from src.audiocraft.modules.transformer import set_efficient_attention_backend
-
-    if cfg.transformer_lm.memory_efficient:
-        set_efficient_attention_backend("xformers")
+    # from src.audiocraft.modules.transformer import set_efficient_attention_backend
+    # if cfg.transformer_lm.memory_efficient:
+    #     set_efficient_attention_backend("xformers")
+    # xformers has no kernel for ZeroGPU's Blackwell hardware; falling back to plain attention.
+    cfg.transformer_lm.memory_efficient = False
 
     kwargs = dict_from_config(getattr(cfg, "transformer_lm"))
     n_q = kwargs["n_q"]