今天主要是结合理论进一步熟悉TensorRT-LLM的内容
从下面的分享可以看出,TensorRT-LLM是在TensorRT的基础上进行了进一步封装,提供拼batch,量化等推理加速实现方式。
下面的图片更好的展示了TensorRT-LLM的流程,包含权重转换,构建Engine,以及推理,评估等内容。总结一下就是三步。
不想看图的话,可以看看AI的总结,我放在附录中。
下图也很好的展示的trt-llm推理的全流程。
多卡并行
值得注意的是,trt-llm特意考虑了多卡部署的使用场景。通过tp-size参数来控制张量并行的程度,pp-size来控制溧水县并行的程度。
流水线并行
量化
权重&激活值量化
KV Cache量化
量化精度影响
从下图可以看出,使用FP8进行量化,量化精度较高。
性能调优
关于性能调优,trt-llm中也使用了类似于vllm中xontinuous batching的策略。
附录
The image describes an overview of the TensorRT-LLM (Large Language Model) workflow. Here's a summary of the key steps and elements involved:
- Input Models:
- Various external models from frameworks like HuggingFace, NeMo, AMMO, and Jax can be used as inputs.
- TRT-LLM Checkpoint:
- These external models are converted into a format defined by TRT-LLM using scripts like convert_checkpoint.py or quantize.py.
- This conversion determines several key backward layer parameters, including:
- Quantization method
- Parallelization method
- And more...
- TRT-LLM Engines:
- After converting to the checkpoint format, the trtllm-build command is used to further convert and optimize the checkpoint into TensorRT Engines.
- During this step, important inference parameters are set, such as:
- Max batch size
- Max input length
- Max output length
- Max beam width
- Plugin configuration
- And others...
- Most of the automatic optimizations occur at this stage.
- Application Development:
- Using C++/Python APIs, developers can build applications with these optimized engines.
- TensorRT-LLM comes with several built-in tools to help with secondary development:
- summarize.py for text summarization
- mmlu.py for accuracy testing
- run.py for a dry run to verify the model
- benchmark for benchmarking
- The runtime options include:
- Temperature (for sampling)
- Top K (for top K sampling)
- Top P (for nucleus sampling)
This workflow outlines how to integrate and optimize models for efficient inference with TensorRT-LLM and leverage its tools for application development and performance testing.
NVIDIA AI 加速精讲堂-TensorRT-LLM 应用与部署_哔哩哔哩_bilibili
版权归原作者 Undefined游侠 所有, 如有侵权,请联系我们删除。