TensorRT-LLMs/docs/source/torch.md

PyTorch Backend

Note:
This feature is currently experimental, and the related API is subjected to change in future versions.

To enhance the usability of the system and improve developer efficiency, TensorRT-LLM launches a new experimental backend based on PyTorch.

The PyTorch backend of TensorRT-LLM is available in version 0.17 and later. You can try it via importing tensorrt_llm._torch.

Quick Start

Here is a simple example to show how to use tensorrt_llm._torch.LLM API with Llama model.

    :language: python
    :linenos:

Quantization

The PyTorch backend supports FP8 and NVFP4 quantization. You can pass quantized models in HF model hub, which are generated by TensorRT Model Optimizer.

from tensorrt_llm._torch import LLM
llm = LLM(model='nvidia/Llama-3.1-8B-Instruct-FP8')
llm.generate("Hello, my name is")

Or you can try the following commands to get a quantized model by yourself:

git clone https://github.com/NVIDIA/TensorRT-Model-Optimizer.git
cd TensorRT-Model-Optimizer/examples/llm_ptq
scripts/huggingface_example.sh --model <huggingface_model_card> --quant fp8 --export_fmt hf

Developer Guide

• Architecture Overview
• Adding a New Model
• Examples

Key Components

• Attention
• KV Cache Manager
• Scheduler

Known Issues

• The PyTorch workflow on SBSA is incompatible with bare metal environments like Ubuntu 24.04. Please use the PyTorch NGC Container for optimal support on SBSA platforms.