Skip to content

ContextualAI/fastertransformer_backend

 
 

Repository files navigation

NOTE: Fastertransformer backend is currently undergoing restructuring. Build instructions are only tested with Triton container versions <= 23.04.

1. FasterTransformer Backend

The triton faster transformer backend works as an interface to call FasterTransformer in triton.

All necessary implements are actually in FasterTransformer repository.

The CMakeList.txt will fetch relative repository to organize and compile the project with:

  • this repository itself
  • Faster Transformer repository
  • 3rdparty
    • cutlass
    • Megatron
    • etc...

2. LLaMa support

To check how faster transformer support LLaMa, and how triton support LLaMa, here is the structure:

Faster Transformer Library
├── examples
│   └── cpp
│       └── llama
│           ├── CMakeList.txt
│           ├── llama_config.ini
│           ├── llama_example.cc
│           └── llama_triton_example.cc
└── src
    └── fastertransformer
        ├── models
        │   └── llama
        │       ├── CMakeList.txt
        │       ├── Llama.h
        │       ├── LlamaContextDecoder.h
        │       ├── LlamaDecoder.h
        │       ├── LlamaDecoderLayerWeight.h
        │       └── LlamaWeight.h
        └── triton_backend
            └── llama
                ├── CMakeList.txt
                ├── LlamaTritonModel.h
                └── LlamaTritonModelInstance.h

Faster Transformer Backend
├── all_models
│   └── llama
│       ├── ensemble
│       ├── fastertransformer
│       ├── postprocessing
│       └── preprocessing
└── src
    └── libfastertransformer.cc

2.1 build your faster transformer library

The faster transformer repository work as a library to support different models.

2.1.1 (essential) faster transformer library for your_model

examples/cpp/your_model is essential if you want to run your model on faster transformer.

2.1.2 (essential) examples for your_model

src/fastertransformer/models/your_model is essential because it stores your_model_config.ini, and other files (bad_words.csv) to ensure your model to work well.

2.1.3 (optional) triton backend

src/triton_backend/your/model is optional.

Only when you want to deploy your model on triton server with faster transformer backend, you need to implement this part.

3. Quick Start

We have deployed llama-7b to triton inference server, see the llama_guide to boost your deploying work and get familiar with NVIDIA Triton Inference Server

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 74.3%
  • C++ 11.8%
  • Shell 10.7%
  • CMake 2.8%
  • Dockerfile 0.4%