Large Model（大模型）#

cv 大模型#

• facebookresearch/segment-anything
• UX-Decoder/Segment-Everything-Everywhere-All-At-Once
  • SEEM: Segment Everything Everywhere All at Once
• baaivision/Painter
  • Painter & SegGPT Series: Vision Foundation Models from BAAI

大模型训练框架#

• deepspeedai/DeepSpeed
• NVIDIA/Megatron-LM
• hpcaitech/ColossalAI

大模型推理框架#

• sgl-project/sglang
• vllm-project/vllm
• huggingface/text-generation-inference
  • TGI
• InternLM/lmdeploy
• ggerganov/llama.cpp
• ollama/ollama
• gpustack/gpustack
• NVIDIA/TensorRT-LLM
  • NVIDIA/FasterTransformer
    • 废弃了，改用 TensorRT-LLM
• deepspeedai/DeepSpeed-MII
• microsoft/onnxruntime
• ModelTC/lightllm

教程#

• rasbt/LLMs-from-scratch

调度#

• ray-project/ray
• kubeflow/kubeflow
• mlflow/mlflow
• sql-machine-learning/sqlflow
  • Brings SQL and AI together.

加速 Pandas#

• dask/dask
• modin-project/modin
• vaexio/vaex
• rapidsai/cudf

部署模型推理#

• SeldonIO/seldon-core
• mlflow/mlflow
• combust/mleap
• awslabs/multi-model-server

大模型训练#

目前训练超大规模语言模型主要有两条技术路线：

• TPU + XLA + TensorFlow/JAX
  • tensorflow/tensorflow
  • openxla/xla
  • google/jax
  • 由 Google 主导，由于 TPU 和自家云平台 GCP 深度绑定，对于非 Googler 来说，只可远观而不可把玩，
• GPU + PyTorch + Megatron-LM + DeepSpeed
  • pytorch/pytorch
  • NVIDIA/Megatron-LM
  • microsoft/DeepSpeed
  • 背后则有 NVIDIA、Meta、MS 大厂加持，社区氛围活跃，也更受到群众欢迎
• horovod/horovod
• bytedance/byteps