Upload folder using huggingface_hub

app.py

@@ -30,9 +30,6 @@ MODEL_ID = "naver/v-splade-quality"
 # ── Module-scope model load (eager, as required by ZeroGPU) ──────────────────
 processor = AutoProcessor.from_pretrained(MODEL_ID)
 model = build_model(MODEL_ID, mode="inference_only", dtype=torch.bfloat16).to("cuda")
-# Pre-build the Li-LSR vocab lookup table so the first query is fast.
-model.query_encoder.to("cuda", dtype=torch.bfloat16)
-model.query_encoder.build_lookup_table()
 tokenizer = processor.tokenizer
 
 
@@ -53,6 +50,9 @@ def retrieve(image: Image.Image, queries: str, topk: int = 15) -> tuple:
     if not queries.strip():
         return "", "<p style='color:red'>Please enter at least one query.</p>", ""
 
+    # Build lookup table on GPU (lazy — first call builds it).
+    model.query_encoder.build_lookup_table()
+
     # ── Encode image → sparse embedding ──────────────────────────────────
     chat = [{"role": "user",
              "content": [{"type": "image"}, {"type": "text", "text": ""}]}]