Upload app.py with huggingface_hub

app.py

@@ -27,13 +27,26 @@ MODEL_NICKNAME = "core"  # ARDY-Core-RP-20FPS-Horizon40, 27-joint skeleton @ 20
 MAX_SEED = 2**31 - 1
 
 # -----------------------------------------------------------------------------
-# Model loading (module scope, eager .to("cuda") per ZeroGPU rules)
+# Model loading (module scope).
+#
+# ZeroGPU has no live GPU at startup: it intercepts .to("cuda") *placement* but
+# NOT arbitrary CUDA compute. LLM2Vec's PEFT load runs a LoRA `merge_and_unload`
+# (real matmuls), so we build everything on CPU first, then .to("cuda") — the
+# placement call is what the ZeroGPU hijack packs to disk and streams into VRAM
+# on the first @spaces.GPU request.
 # -----------------------------------------------------------------------------
-print("Loading ARDY text encoder (LLM2Vec-Llama-3-8B)…", flush=True)
-_text_encoder = load_text_encoder(mode="local", device="cuda")
-
-print("Loading ARDY motion model…", flush=True)
-MODEL = load_model(MODEL_NICKNAME, device="cuda", text_encoder=_text_encoder)
+print("Loading ARDY text encoder (LLM2Vec-Llama-3-8B) on CPU…", flush=True)
+_text_encoder = load_text_encoder(mode="local", device="cpu")
+
+print("Loading ARDY motion model on CPU…", flush=True)
+MODEL = load_model(MODEL_NICKNAME, device="cpu", text_encoder=_text_encoder)
+
+print("Moving models to CUDA (ZeroGPU-intercepted placement)…", flush=True)
+_text_encoder = _text_encoder.to("cuda")
+MODEL = MODEL.to("cuda")
+# self.device was captured at construction (device="cpu"); generation creates
+# many tensors on it, so retarget it to cuda to match the moved weights.
+MODEL.device = "cuda"
 MODEL = MODEL.eval()
 
 FPS = float(MODEL.motion_rep.fps)