INNER CODE UNIT · Python

prepare_seconds

invergent-ai/surogate · surogate/serve/tools/reference/qwen3_5/cli.py:353

            prepare_seconds = time.perf_counter() - prepare_started

            prefill_started = time.perf_counter()
            if vision_output is not None:
                assert prompt.batch is not None
                token = model.prefill_multimodal(
                    prompt.batch,
                    vision_output,
                    sampler=sampler,
                    tap=tap,
                )
            else:
                token = model.prefill(
                    prompt.token_ids,
                    sampler=sampler,
                    tap=tap,
                )
            _synchronize(model.device)

View source record →