这是一个本地托管版本的 GitHub Copilot。它在英伟达的 Triton 推理服务器中使用了 SalesForce CodeGen 模型和 FasterTransformer 后端。
前提条件
- Docker
- docker-compose >= 1.28
- 一台计算能力大于 6.0 的英伟达 GPU，以及足够的 VRAM 来运行你想要的模型
- nvidia-docker
- curl 和 zstd，用于下载和解包模型
Copilot 插件
你可以配置官方 VSCode Copilot 插件来使用你的本地服务器。只要编辑你的 settings.json 来添加。
"github.copilot.advanced": {
"debug.overrideEngine": "codegen",
"debug.testOverrideProxyUrl": "http://localhost:5000",
"debug.overrideProxyUrl": "http://localhost:5000"
}
设置
运行设置脚本以选择要使用的模型。 这将从 Huggingface 下载模型，然后将其转换为与 FasterTransformer 一起使用。
$ ./setup.sh
Models available:
[1] codegen-350M-mono (2GB total VRAM required; Python-only)
[2] codegen-350M-multi (2GB total VRAM required; multi-language)
[3] codegen-2B-mono (7GB total VRAM required; Python-only)
[4] codegen-2B-multi (7GB total VRAM required; multi-language)
[5] codegen-6B-mono (13GB total VRAM required; Python-only)
[6] codegen-6B-multi (13GB total VRAM required; multi-language)
[7] codegen-16B-mono (32GB total VRAM required; Python-only)
[8] codegen-16B-multi (32GB total VRAM required; multi-language)
Enter your choice [6]: 2
Enter number of GPUs [1]: 1
Where do you want to save the model [/home/moyix/git/fauxpilot/models]? /fastdata/mymodels
Downloading and converting the model, this will take a while...
Converting model codegen-350M-multi with 1 GPUs
Loading CodeGen model
Downloading config.json: 100%|██████████| 996/996 [00:00<00:00, 1.25MB/s]
Downloading pytorch_model.bin: 100%|██████████| 760M/760M [00:11<00:00, 68.3MB/s]
Creating empty GPTJ model
Converting...
Conversion complete.
Saving model to codegen-350M-multi-hf...
=============== Argument ===============
saved_dir: /models/codegen-350M-multi-1gpu/fastertransformer/1
in_file: codegen-350M-multi-hf
trained_gpu_num: 1
infer_gpu_num: 1
processes: 4
weight_data_type: fp32
========================================
transformer.wte.weight
transformer.h.0.ln_1.weight
[... more conversion output trimmed ...]
transformer.ln_f.weight
transformer.ln_f.bias
lm_head.weight
lm_head.bias
Done! Now run ./launch.sh to start the FauxPilot server.
