Lower @spaces.GPU duration from 180/120 to 60 (ZeroGPU max-duration cap)

zsgdp/benchmarks/embedding_retriever.py

@@ -41,7 +41,7 @@ from zsgdp.schema import Chunk
 Embedder = Callable[[list[str]], list[list[float]]]
 
 
-@zero_gpu_slot(duration=180)
+@zero_gpu_slot(duration=60)
 def _gpu_encode_batch(model_id: str, task: str | None, texts: list[str]) -> list[list[float]]:
     """Load a sentence-transformers model and encode `texts` under a ZeroGPU slot.
 

zsgdp/gpu/transformers_client.py

@@ -19,7 +19,7 @@ from zsgdp.gpu.worker_prompts import prompt_for_task
 from zsgdp.gpu.zero_gpu import gpu as zero_gpu_slot
 
 
-@zero_gpu_slot(duration=120)
+@zero_gpu_slot(duration=60)
 def _gpu_run_pipeline(
     model_id: str,
     pipeline_task: str,