t0 = time.perf_counter()

base_out = target.generate(**inputs, max_new_tokens=200, do_sample=False, streamer=streamer)

base_time = time.perf_counter() - t0

print(f"\n⏱ Without Speculative Decoding:")
print(f"  Runtime: {base_time:.2f}s")
print(f"  Token speed: {len(base_out[0]) / base_time:.1f} tok/s")