The inductor backend of torch.compile generates dedicated fused Triton kernels for each unique input tensor shape; in LLM inference deployment, kernel compilation for dozens of different shapes may be triggered, with cumulative overhead exceeding 30 seconds.