aphrodite-engine

Breathing Life into Language

Aphrodite is the official backend engine for PygmalionAI. It is designed to serve as the inference endpoint for the PygmalionAI website, and to allow serving the Pygmalion models to a large number of users with blazing fast speeds (thanks to FasterTransformer and vLLM).

Aphrodite builds upon and integrates the exceptional work from various projects.

The compute necessary for Aphrodite's development is provided by Arc Compute.

Features

• Continuous Batching
• Efficient K/V management with PagedAttention
• Optimized CUDA kernels for improved inference
• Quantization support via GPTQ, GGUF, AWQ, QuIP#, and SqueezeLLM.
• Distributed inference
• Variety of sampling methods (Mirostat, Locally Typical Sampling, Tail-Free Sampling, etc)
• 8-bit KV Cache for higher context lengths and throughput.

Quickstart

pip install aphrodite-engine

python -m aphrodite.endpoints.openai.api_server --model PygmalionAI/pygmalion-2-7b

[!CAUTION] If the installation reports CUDA kernel errors, please run pip install aphrodite-engine=0.4.5 instead.

This will create a OpenAI-compatible API server that can be accessed at port 2242 of the localhost. You can plug in the API into a UI that supports Kobold, such as SillyTavern.

Docker

sudo mkdir ~/workspace
sudo docker run --gpus '"all"' --shm-size 10g -p 2242:2242 -p 7650:7650 -p 8080:8080 --rm -it -v '/home/workspace' alignmentlabai/aphrodite-engine:latest

Performance

Speeds vary with different GPUs, model sizes, quantization schemes, batch sizes, etc. Here are some baseline benchmarks conducted by requesting as many completions as possible from the API server. Keep in mind that these are the theoritical peak throughput with parallel decoding, with as high a batch size as possible. Per-request generation speed is a fraction of this, at 30-40 t/s.

High Batch Size Performance

[!NOTE]
The numbers below are the theoritical peak achieved by only requesting output tokens at very high batch sizes. At lower batch sizes with much larger prompts, the results will be vastly different. Throughput refers to output tokens per second.

Model	Quantization	bits	GPU	Throughput (T/s)
Mistral 7B	None	16	RTX 4090	5489.3
	AWQ	4	RTX 4090	4078.8
	GPTQ	4	RTX 4090	7850.4
		8	RTX 4090	7658.0
	GGUF	Q8	RTX 4090	5141.2
		Q6KM	RTX 4090	5791.7
		Q5KM	RTX 4090	5786.2
		Q4KM	RTX 4090	5815.8
	SqueezeLLM	4	RTX 4090	549.5
Llama-2 7B	None	16	RTX 4090	2576.2
	AWQ	4	RTX 4090	3551.3
	GPTQ	4	RTX 4090	2919.1
	GGUF	Q4KM	RTX 4090	2726.6
		Q5KM	RTX 4090	2763.4
		Q6KM	RTX 4090	2694.7
		Q8	RTX 4090	2647.0
	SqueezeLLM	4	RTX 4090	580.3

Batch Size 1

These are the speeds a user would normally get if they request a single output with a sizable prompt and output length. Essentially, normal chatting experience.

The following results were gathered by sending a request with 2000 prompt tokens and requesting 1024 tokens with ignore_eos=True.

Model	Quantization	bits	GPU	Throughput (T/s)
Mistral 7B	None	16	RTX 4090	54.0
	AWQ	4	RTX 4090	128.2
	GPTQ	8	RTX 4090	92.8
		4	RTX 4090	146.8
	GGUF	Q8	RTX 4090	91.0
		Q6KM	RTX 4090	105.4
		Q5KM	RTX 4090	117.8
		Q4KM	RTX 4090	128.9
Llama-2 7B	None	16	RTX 4090	55.2
	GPTQ	8	RTX 4090	90.2
		4	RTX 4090	128.0
	AWQ	4	RTX 4090	116.3
	GGUF	Q8	RTX 4090	88.1
		Q6KM	RTX 4090	99.4
		Q5KM	RTX 4090	109.9
		Q4KM	RTX 4090	118.9

Requirements

• Operating System: Linux (or WSL for Windows)
• Python: at least 3.8

Build Requirements:

• CUDA >=12

For supported GPUs, see here.

Installation

• From PyPi
• Build from source

Usage

For usage, please refer to the wiki page for detailed instructions. Aphrodite provides many different options for LLM inference, so please read through the list of options here.

Notes

1. By design, Aphrodite takes up 90% of your GPU's VRAM. If you're not serving an LLM at scale, you may want to limit the amount of memory it takes up. You can do this in the API example by launching the server with the --gpu-memory-utilization 0.6 (0.6 means 60%).

2. You can view the full list of commands by running python -m aphrodite.endpoints.openai.api_server --help.

3. Context Length extension via the RoPE method is supported for most models. Use the command-line flag --max-model-len to specify a desired context length and the engine will adjust the RoPE scaling accordingly.

4. Please refer to the FAQ & Issues if you run into problems. If you don't find an answer there, please make an issue.

Acknowledgements

Aphrodite Engine would have not been possible without the phenomenal work of other open-source projects. Credits go to:

• vLLM (CacheFlow)
• FasterTransformer
• xFormers
• AutoAWQ
• AutoGPTQ
• SqueezeLLM
• ExLlama
• Exllamav2
• KoboldAI
• Text Generation WebUI
• Megatron-LM
• FastChat
• Ray
• SkyPilot
• OpenAI Python Library

Contributing

Everyone is welcome to contribute. You can support the project by opening Pull Requests for new features, fixes, or general UX improvements.