Upload app.py with huggingface_hub

app.py

@@ -35,9 +35,15 @@ def load_model_zerogpu():
     """Load the MuScriptor model in a ZeroGPU-compatible way.
 
     On ZeroGPU, safetensors.load_file(device="cuda") fails because there's no
-    real GPU at module scope. We load weights on CPU, build the model on CPU,
-    then call .to("cuda") — which ZeroGPU intercepts to pack weights to disk
-    and stream into VRAM on the first @spaces.GPU call.
+    real GPU at module scope. However, .to("cuda") is intercepted by the
+    spaces hijack. So we:
+      1. Build the model with device="cuda" (conditioners store this as
+         self.device and use it at runtime to route tensors).
+      2. Load safetensors weights on CPU.
+      3. Load_state_dict into the model (weights land on CPU because the
+         model's tensors are still fake-CUDA via the hijack).
+      4. Call model.to("cuda") — ZeroGPU intercepts this and packs weights
+         to disk, streaming them into VRAM on the first @spaces.GPU call.
     """
     device = torch.device("cuda")  # ZeroGPU intercepts this
 
@@ -48,17 +54,20 @@ def load_model_zerogpu():
     ))
     cfg = _resolve_config(source, weights_path)
 
-    # Build the model on CPU first
-    cpu_device = torch.device("cpu")
-    model = _build_model(cpu_device, cfg)
+    # Build the model with device="cuda" — conditioners store this and use it
+    # at runtime to route tensors. ZeroGPU intercepts .to("cuda") calls inside.
+    model = _build_model(device, cfg)
     model.eval()
 
-    # Load weights on CPU
+    # Load weights on CPU (safetensors can't load to fake CUDA)
     state_dict = load_file(str(weights_path), device="cpu")
     state_dict = _remap_single_codebook_keys(state_dict)
+
+    # Load state dict — the model's parameters are fake-CUDA (ZeroGPU),
+    # but load_state_dict copies CPU data into them, which is fine.
     model.load_state_dict(state_dict)
 
-    # Now move to "cuda" — ZeroGPU intercepts this and packs to disk
+    # Move everything to "cuda" — ZeroGPU intercepts this and packs to disk
     model.to("cuda")
 
     tokenizer = MT3Tokenizer(