thudm/fastldm

Inference speed-up for stable-diffusion (ldm) with TensorRT.

FastLDM

We focus on inference speed-up for ldm.

TensorRT

We are now trying to convert model of stable-diffusion into an optimized TensorRT engine. See examples/ folder for use cases.

When we transform a PyTorch model to TensorRT engine, we usually follow "torch->onnx->trt" footprint. It is convenient for most cases, except when we want to use plugins. If we want to use plugins, we usually need to modify the exported onnx file based on TensorRT official demo. The core idea of this repo is to provide a shortcut. We provide many plugin-friendly PyTorch modules, which can plug into your normal PyTorch model and export onnx directly able to be optimized by TensorRT with plugins.

The advantages of this repo are:

• Seamlessly transform from PyTorch to TensorRT with plugins.
• Seamlessly export onnx to run directly in onnxruntime.

More specifically, we provide:

• parameter mapping functions from a module to another one in utils.mapping and utils.modifier
• TensorRT-friendly modules in utils.plugins and utils.modules
• benchmarking functions in utils.benchmark
• experiment functions to compare different implemented models with same interface

Install

pip install -e .

or

pip install git+https://github.com/THUDM/FastLDM.git

Environment

We rely on the docker environment nvcr.io/nvidia/pytorch:22.12-py3. (For newer version of TensorRT, this repo may not work because of removement of plugins.)

fastldm's Issues

is support lora

lora

the trt can be used in c++?

where is the cpp code?