MoE-Infinity

MoE-Infinity is a cost-effective, fast, and easy-to-use library for Mixture-of-Experts (MoE) inference.

MoE-Infinity is cost-effective yet fast:

• Offloading MoE's experts to host memory, allowing memory-constrained GPUs to serve MoE models.
• Minimizing the expert offloading overheads through several novel techniques: expert activation tracing, activation-aware expert prefetching, and activation-aware expert caching.
• Supporting LLM acceleration techniques (such as FlashAttention).
• Supporting multi-GPU environments with numeorous OS-level performance optimizations.
• Achieving SOTA latency performance when serving MoEs in a resource-constrained GPU environment (in comparison with vLLM, HuggingFace Accelerate, DeepSpeed, Mixtral-Offloading, and Ollama/LLama.cpp).

MoE-Infinity is easy-to-use:

• HuggingFace model compatible, and HuggingFace programmer friendly.
• Supporting all available MoE checkpoints (including Deepseek-V2, Google Switch Transformers, Meta NLLB-MoE, and Mixtral).

Note that: The open-sourced MoE-Infinity has been redesigned for making it HuggingFace-users friendly. This version is different from the version reported in the paper, which takes extreme performance as the top priority. As a result, distributed inference is currently not supported in this open-sourced version.

Contents

• Performance
• Installation
  • Prerequisites
  • Install from conda environment
  • Install from PyPI
  • Install from Source
  • Enable FlashAttention (Optional)
• Usage and Examples
  • Sample Code of Huggingface LLM Inference
  • Running Inference
• Release Plan
• Citation

Performance

Single GPU A5000 (24GB Memory), per-token-latency (seconds) for generation with a mixed dataset that includes LongBench, GSM8K, FLAN, BIG-Bench and MMLU datasets. Lower per-token-latency is preferable.

	Switch-large-128	NLLB-MoE-54B	Mixtral-8x7b	DeepSeek-V2-Lite
MoE-Infinity	0.130	0.119	0.735	0.155
Accelerate	1.043	3.071	6.633	1.743
DeepSpeed	4.578	8.381	2.486	0.737
Mixtral Offloading	X	X	1.752	X
Ollama	X	X	0.903	1.250
vLLM	X	X	2.137	0.493

Installation

We recommend installing MoE-Infinity in a virtual environment. To install MoE-Infinity, you can either install it from PyPI or build it from source.

conda create -n moe-infinity python=3.9
conda activate moe-infinity
# install from either PyPI or Source will trigger requirements.txt automatically

Install from PyPI

# install stable release
pip install moe-infinity

# install nightly release
pip install -i https://test.pypi.org/simple/ --extra-index-url https://pypi.org/simple/ moe-infinity

Install from Source

git clone https://github.com/EfficientMoE/MoE-Infinity.git
cd MoE-Infinity
pip install -e .
conda install -c conda-forge libstdcxx-ng=12 # assume using conda, otherwise install libstdcxx-ng=12 using your package manager or gcc=12

Enable FlashAttention (Optional)

Install FlashAttention (>=2.5.2) for faster inference with the following command.

FLASH_ATTENTION_FORCE_BUILD=TRUE pip install flash-attn

Post-installation, MoE-Infinity will automatically integrate with FlashAttention to enhance performance.

Usage and Examples

We provide a simple API for diverse setups, including single GPU, multiple GPUs, and multiple nodes. The following examples show how to use MoE-Infinity to run generation on a Huggingface LLM model.

Important Note

• The offload_path must be unique for each MoE model. Reusing the same offload_path for different MoE models will result in unexpected behavior.

Sample Code of Huggingface LLM Inference

import torch
import os
from transformers import AutoTokenizer
from moe_infinity import MoE

user_home = os.path.expanduser('~')

checkpoint = "deepseek-ai/DeepSeek-V2-Lite-Chat"
tokenizer = AutoTokenizer.from_pretrained(checkpoint, trust_remote=True)

config = {
    "offload_path": os.path.join(user_home, "moe-infinity"),
    "device_memory_ratio": 0.75, # 75% of the device memory is used for caching, change the value according to your device memory size on OOM
}

model = MoE(checkpoint, config)

input_text = "translate English to German: How old are you?"
input_ids = tokenizer(input_text, return_tensors="pt").input_ids.to("cuda:0")

output_ids = model.generate(input_ids)
output_text = tokenizer.decode(output_ids[0], skip_special_tokens=True)

print(output_text)

Running Inference

This command runs the script on selected GPUs.

CUDA_VISIBLE_DEVICES=0,1 python script.py

We provide a simple example to run inference on a Huggingface LLM model. The script will download the model checkpoint and run inference on the specified input text. The output will be printed to the console.

CUDA_VISIBLE_DEVICES=0 python examples/interface_example.py --model_name_or_path "deepseek-ai/DeepSeek-V2-Lite-Chat" --offload_dir <your local path on SSD>

OpenAI-Compatible Server

Start the OpenAI-compatible server locally

python -m moe_infinity.entrypoints.openai.api_server --model deepseek-ai/DeepSeek-V2-Lite-Chat --offload-dir ./offload_dir

Query the model via /v1/components/. (We currently only support the required fields, i.e., "model" and "prompt").

curl http://localhost:8000/v1/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "deepseek-ai/DeepSeek-V2-Lite-Chat",
        "prompt": "Hello, my name is"
    }'

You can also use openai python package to query the model.

pip install openai
python tests/test_oai_completions.py

Query the model via /v1/chat/completions. (We currently only support the required fields, i.e., "model" and "messages").

curl http://localhost:8000/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "deepseek-ai/DeepSeek-V2-Lite-Chat",
        "messages": [
            {"role": "system", "content": "You are a helpful assistant."},
            {"role": "user", "content": "Tell me a joke"}
        ]
    }'

You can also use openai python package to query the model.

pip install openai
python tests/test_oai_chat_completions.py

Release Plan

We plan to release two functions in the following months:

• We currently support PyTorch as the default inference engine, and we are in the process of supporting vLLM as another inference runtime, which includes the support of KV cache offloading.
• Supporting expert parallelism for distributed MoE inference.
• More (We welcome contributors to join us!)

Citation

If you use MoE-Inifity for your research, please cite our paper:

@misc{moe-infinity,
  author       = {Leyang Xue and
                  Yao Fu and
                  Zhan Lu and
                  Luo Mai and
                  Mahesh Marina},
  title        = {MoE-Infinity: Efficient MoE Inference on Personal Machines with Sparsity-Aware Expert Cache},
  archivePrefix= {arXiv},
  eprint       = {2401.14361},
  year         = {2024}
}