Build the model when the worker starts

load_model is cached, but it was first called inside the GPU call, and ZeroGPU runs each request in a process of its own, so every request built the model inside its reservation. pyharp's worker now builds it at startup, where ZeroGPU's emulation records the placement.

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

app.py

@@ -9,7 +9,7 @@ import gradio as gr
 from pyharp import ModelCard, build_endpoint
 from pyharp.tags import Subcategory
 
-from runtime import EXAMPLE, render
+from runtime import EXAMPLE, load_model, render
 
 model_card = ModelCard(
     name="DExter",
@@ -23,6 +23,14 @@ model_card = ModelCard(
 )
 
 
+# pyharp runs process_fn in a worker, which imports this file as __mp_main__ and starts with the
+# server. The model is built and placed there alone, where ZeroGPU's emulation records the CUDA
+# placement. ZeroGPU runs each request in a process of its own, so a model first built inside the
+# GPU call was built again, inside the reservation, every time.
+if __name__ != "__main__":
+    load_model()
+
+
 def process_fn(xml_path: str | None, use_example: bool, seed: int) -> str:
     # Checked before the GPU is reserved, so a request without a score spends none of the
     # caller's ZeroGPU allowance