Log generation time and token count so the GPU duration can be measured, not guessed

app.py

@@ -10,6 +10,7 @@ fenced ```sql block (extract_sql_from_completion), not the raw completion.
 """
 
 import re
+import time
 
 import spaces
 import gradio as gr
@@ -132,12 +133,27 @@ def generate_sql(question: str) -> tuple[str, str]:
     inputs = tokenizer(prompt_text, return_tensors="pt", add_special_tokens=False).to("cuda")
     prompt_len = inputs["input_ids"].shape[1]
 
+    started = time.perf_counter()
     output_ids = model.generate(
         **inputs,
         max_new_tokens=MAX_NEW_TOKENS,
         do_sample=False,
         pad_token_id=tokenizer.eos_token_id,
     )
+    elapsed = time.perf_counter() - started
+    new_tokens = output_ids.shape[1] - prompt_len
+
+    # The decorator's `duration` above has to exceed this, or ZeroGPU kills
+    # the request mid-generation and the caller sees only "GPU task
+    # aborted". Logging it turns that budget into something measured
+    # rather than guessed: read these lines from the Space logs and set
+    # `duration` from the worst observed generation, plus headroom.
+    print(
+        f"[timing] generated {new_tokens} tokens in {elapsed:.1f}s "
+        f"({new_tokens / elapsed:.1f} tok/s), prompt {prompt_len} tokens",
+        flush=True,
+    )
+
     completion = tokenizer.decode(
         output_ids[0][prompt_len:], skip_special_tokens=True
     ).strip()