disable memory-efficient attention for magnet

src/audiocraft/models/loaders.py

@@ -136,9 +136,11 @@ def load_lm_model_magnet(file_or_url_or_id: tp.Union[Path, str], compression_mod
     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
 
     model = builders.get_lm_model(cfg)
     model.load_state_dict(pkg['best_state'])