Fix @spaces.GPU decorator, stable search timing via averaging

app.py

@@ -24,10 +24,7 @@ import matplotlib.pyplot as plt
 import numpy as np
 
 
-try:
-    import spaces
-except ImportError:
-    spaces = None
+import spaces
 
 from backend import get_model_and_tokenizer, run_all_sizes, run_all_sizes_indexed, SIZES, DEVICE
 from examples import HOW_TO_USE_CONTENT
@@ -1104,7 +1101,7 @@ def on_size_change(results, size_idx: int, corpus_name: str, active_qrels):
            page_indicator_html, gr.update(interactive=False), gr.update(interactive=total_pages > 1), 0
 
 
-@spaces.GPU if spaces else lambda f: f
+@spaces.GPU
 def run_demo(query: str, corpus_name: str, size_idx: int, active_qrels):
     """Run all 6 Starbucks sizes and return results for the selected size."""
     import torch

backend.py

@@ -254,13 +254,18 @@ def compute_ranking(
     query_emb: np.ndarray,
     doc_embs: np.ndarray,
     documents: list[str],
+    n_timing_runs: int = 50,
 ) -> tuple[list[tuple[str, float]], float]:
     """Rank documents via dot-product similarity (numpy fallback)."""
-    start = time.perf_counter()
+    # Warm up + actual run for ranking
     scores = np.dot(doc_embs, query_emb)
     ranked_idx = np.argsort(scores)[::-1]
-    search_time_ms = (time.perf_counter() - start) * 1000
     ranked = [(documents[i], float(scores[i])) for i in ranked_idx]
+    # Stable timing: average over multiple runs (sub-ms operations are noisy)
+    start = time.perf_counter()
+    for _ in range(n_timing_runs):
+        np.dot(doc_embs, query_emb)
+    search_time_ms = (time.perf_counter() - start) * 1000 / n_timing_runs
     return ranked, search_time_ms