Halve @spaces.GPU duration 60s → 30s to save quota per call

llm.py

@@ -237,14 +237,17 @@ def format_mistral_prompt(
 
 # ──────────────────────────────────────────────────────────── inference
 
-@spaces.GPU(duration=60)
+@spaces.GPU(duration=30)
 def generate(prompt: str, **gen_kwargs) -> str:
     """Run one generation pass and return only the newly-generated text.
 
-    Decorated with @spaces.GPU so ZeroGPU acquires the A100 for the
-    duration of this call. The 60-second budget covers up to ~1500
-    tokens at ~25 tok/s; tune via gen_kwargs['max_new_tokens'] if you
-    need longer answers.
+    Decorated with @spaces.GPU(duration=30) so ZeroGPU acquires the A100
+    for up to 30 seconds. ZeroGPU bills the FULL declared duration
+    against the caller's daily quota regardless of how long the actual
+    call took — so a 60s budget for a 5s response wastes 55s of quota.
+    Most chat answers complete in 5-15s; 30s is comfortably above that
+    while halving the quota cost vs. the old 60s. If responses ever
+    start truncating because they hit the timeout, bump back up.
 
     On a local CPU host the decorator is a no-op (see the spaces shim
     at the top of this file) and the call runs on whatever device the