trt-llm相比hf跑qwen的forward仅context phrase有加速效果，generation没有加速效果

System Info

• CPU architecture： X86_64
• GPU properties：NVIDIA L40S，46G，8卡
• Libraries
  -TensorRT-LLM 版本:v0.14.0
  -TensorRT版本：10.4.0
  -llm 模型：Qwen2-7B-Instruct
  -使用镜像：nvcr.io/nvidia/tritonserver:24.10-trtllm-python-py3
  -NVIDIA driver info：NVIDIA-SMI 545.23.08 Driver Version: 545.23.08 CUDA Version: 12.3
  参考test_llama.py写test_qwen.py调用模型forward，脚本中build_engine参数：gemm_plugin和gpt_attention_plugin均使用，enable_remove_input_padding为false，dtype为float16，多次调用runtime进行forward和hf进行对比，耗时如下：

context phrase:
Time cost of trt-llm is 0.1882
Time cost of hf-llm is 0.4351
generation phrase:
ime cost of trt-llm is 0.0219
Time cost of hf-llm is 0.0236

Who can help?

No response

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

1.build engine from hf model reference llama：https://github.com/NVIDIA/TensorRT-LLM/blob/main/tests/model/test_llama.py
2.调用前向计算代码，打印时间

Expected behavior

比hf速度有明显提升

actual behavior

仅context phrase有加速效果，generation没有加速效果

additional notes

none

Start with tests/model/test_llama.py and the reported test_qwen.py reproduction, then run the forward benchmark under the listed TensorRT-LLM, TensorRT, driver, and GPU versions. Compare context and generation timings against Hugging Face and document whether the generation gap is reproducible and what component accounts for it.