Build the model inside the GPU call again

Building the default size in the worker failed at startup with No CUDA GPUs are available: muscriptor loads its safetensors straight onto CUDA, which ZeroGPU emulation does not answer outside a call, and its conditioners keep the device they were built for. The model is built inside the call again, as before, from weights the worker downloads.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

app.py

@@ -48,11 +48,13 @@ DEFAULT_SIZE = "medium"
 
 
 def _get_model(variant: str) -> TranscriptionModel:
-    """Builds the TranscriptionModel of one size, once.
+    """Builds the TranscriptionModel of one size, on the GPU the call holds.
 
-    The default size is built in the worker at startup. Holding all three would keep about 7 GB
-    there, so another size is built when a request picks it, inside the call, where ZeroGPU builds
-    it again on every call, which takes seconds from local weights.
+    Built inside the call, so ZeroGPU builds it again on every call, which takes seconds from local
+    weights. It cannot be built in the worker at startup, since muscriptor loads its safetensors
+    straight onto CUDA, which ZeroGPU's emulation does not answer outside a call, and its
+    conditioners keep the device they were built for, so building on the CPU and moving it is not
+    enough either.
     """
     if variant not in _models:
         _models[variant] = TranscriptionModel.load_model(_weights[variant], device=DEVICE)
@@ -60,9 +62,7 @@ def _get_model(variant: str) -> TranscriptionModel:
 
 
 # Fetched in the worker, which runs process_fn and starts as the server comes up, so that no
-# request downloads inside its GPU reservation, and the default size is built and placed there, on
-# its main thread, where ZeroGPU's emulation records the placement. The server never runs
-# process_fn.
+# request downloads inside its GPU reservation. The server never runs process_fn.
 if __name__ != "__main__":
     for _size, _revision in _WEIGHT_REVISIONS.items():
         _weights[_size] = Path(snapshot_download(
@@ -81,8 +81,6 @@ if __name__ != "__main__":
                 revision="4afbe376535aaab71b76ae2ca2d7c16b08877321",
             ))
 
-    _get_model(DEFAULT_SIZE)
-
 
 def _resolve_instruments(text: str) -> tuple[list[str] | None, str]:
     """Resolve the comma-separated instruments box into exact group names.