disable memory-efficient attention, xformers has no blackwell kernel

src/audiocraft/models/loaders.py

@@ -135,10 +135,10 @@ def load_lm_model_magnet(file_or_url_or_id: tp.Union[Path, str], compression_mod
     cfg.transformer_lm.segment_duration = cfg.dataset.segment_duration
     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")
+    # Deviation: xformers has no kernel for ZeroGPU's current hardware
+    # (Blackwell, capability 12.0 > xformers' 9.0 cap). Disable memory_efficient
+    # rather than switch to 'torch' backend, which ignores custom attn_mask.
+    cfg.transformer_lm.memory_efficient = False
 
     model = builders.get_lm_model(cfg)
     model.load_state_dict(pkg['best_state'])