nm-vllm 0.4.0

A high-throughput and memory-efficient inference and serving engine for LLMs

  nm-vllm

Overview

vLLM is a fast and easy-to-use library for LLM inference that Neural Magic regularly contributes upstream improvements to. This fork, nm-vllm is our opinionated focus on incorporating the latest LLM optimizations like quantization and sparsity for enhanced performance.

Installation

The nm-vllm PyPi package includes pre-compiled binaries for CUDA (version 12.1) kernels, streamlining the setup process. For other PyTorch or CUDA versions, please compile the package from source.

Install it using pip:

pip install nm-vllm --extra-index-url https://pypi.neuralmagic.com/simple

For utilizing weight-sparsity kernels, such as through sparsity="sparse_w16a16", you can extend the installation with the sparsity extras:

pip install nm-vllm[sparse] --extra-index-url https://pypi.neuralmagic.com/simple

You can also build and install nm-vllm from source (this will take ~10 minutes):

git clone https://github.com/neuralmagic/nm-vllm.git
cd nm-vllm
pip install -e .

Quickstart

Neural Magic maintains a variety of sparse models on our Hugging Face organization profiles, neuralmagic and nm-testing.

A collection of ready-to-use SparseGPT and GPTQ models in inference optimized marlin format are available on Hugging Face

Model Inference with Marlin (4-bit Quantization)

Marlin is an extremely optimized FP16xINT4 matmul kernel aimed at LLM inference that can deliver close to ideal (4x) speedups up to batchsizes of 16-32 tokens. To use Marlin within nm-vllm, simply pass the Marlin quantized directly to the engine. It will detect the quantization from the model's config.

Here is a demonstraiton with a 4-bit quantized OpenHermes Mistral model:

from vllm import LLM, SamplingParams
from transformers import AutoTokenizer

model_id = "neuralmagic/OpenHermes-2.5-Mistral-7B-marlin"
model = LLM(model_id, max_model_len=4096)
tokenizer = AutoTokenizer.from_pretrained(model_id)
sampling_params = SamplingParams(max_tokens=100, temperature=0.8, top_p=0.95)

messages = [
    {"role": "user", "content": "What is synthetic data in machine learning?"},
]
formatted_prompt =  tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
outputs = model.generate(formatted_prompt, sampling_params=sampling_params)
print(outputs[0].outputs[0].text)

Model Inference with Weight Sparsity

For a quick demonstration, here's how to run a small 50% sparse llama2-110M model trained on storytelling:

from vllm import LLM, SamplingParams

model = LLM(
    "neuralmagic/llama2.c-stories110M-pruned50",
    sparsity="sparse_w16a16",   # If left off, model will be loaded as dense
)

sampling_params = SamplingParams(max_tokens=100, temperature=0)
outputs = model.generate("Hello my name is", sampling_params=sampling_params)
print(outputs[0].outputs[0].text)

Here is a more realistic example of running a 50% sparse OpenHermes 2.5 Mistral 7B model finetuned for instruction-following:

from vllm import LLM, SamplingParams
from transformers import AutoTokenizer

model_id = "neuralmagic/OpenHermes-2.5-Mistral-7B-pruned50"
model = LLM(model_id, sparsity="sparse_w16a16", max_model_len=4096)
tokenizer = AutoTokenizer.from_pretrained(model_id)
sampling_params = SamplingParams(max_tokens=100, temperature=0.8, top_p=0.95)

messages = [
    {"role": "user", "content": "What is sparsity in deep learning?"},
]
formatted_prompt =  tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
outputs = model.generate(formatted_prompt, sampling_params=sampling_params)
print(outputs[0].outputs[0].text)

There is also support for semi-structured 2:4 sparsity using the sparsity="semi_structured_sparse_w16a16" argument:

from vllm import LLM, SamplingParams

model = LLM("neuralmagic/llama2.c-stories110M-pruned2.4", sparsity="semi_structured_sparse_w16a16")
sampling_params = SamplingParams(max_tokens=100, temperature=0)
outputs = model.generate("Once upon a time, ", sampling_params=sampling_params)
print(outputs[0].outputs[0].text)

Integration with OpenAI-Compatible Server

You can also quickly use the same flow with an OpenAI-compatible model server:

python -m vllm.entrypoints.openai.api_server \
    --model neuralmagic/OpenHermes-2.5-Mistral-7B-pruned50 \
    --sparsity sparse_w16a16 \
    --max-model-len 4096

Quantized Inference Performance

Developed in collaboration with IST-Austria, GPTQ is the leading quantization algorithm for LLMs, which enables compressing the model weights from 16 bits to 4 bits with limited impact on accuracy. nm-vllm includes support for the recently-developed Marlin kernels for accelerating GPTQ models. Prior to Marlin, the existing kernels for INT4 inference failed to scale in scenarios with multiple concurrent users.

Sparse Inference Performance

Developed in collaboration with IST-Austria, SparseGPT and Sparse Fine-tuning are the leading algorithms for pruning LLMs, which enables removing at least half of model weights with limited impact on accuracy.

nm-vllm includes support for newly-developed sparse inference kernels, which provides both memory reduction and acceleration of sparse models leveraging sparsity.