Home

Awesome

<p align="center"> <img src="./res/candle_vllm_logo.png" alt="candle vLLM" width=55%/> </p>

Continuous integration

Efficient, easy-to-use platform for inference and serving local LLMs including an OpenAI compatible API server.

Features

Develop Status

Currently, candle-vllm supports chat serving for the following models.

Model IDModel TypeSupportedSpeed (A100, BF16)Throughput (BF16, bs=16)Quantized (A100, Q4K or Marlin)Throughput (GTPQ/Marlin, bs=16)
#1LLAMA65 tks/s (LLaMa3.1 8B)553 tks/s (LLaMa3.1 8B)75 tks/s (LLaMa3.1 8B), 115 tks/s (LLaMa3.1 8B, Marlin)968 tks/s (LLaMa3.1 8B)
#2Mistral70 tks/s (7B)585 tks/s (7B)96 tks/s (7B), 115 tks/s (7B, Marlin)981 tks/s (7B)
#3Phi (v1, v1.5, v2)97 tks/s (2.7B, F32+BF16)TBD-TBD
#4Phi-3 (3.8B, 7B)107 tks/s (3.8B)744 tks/s (3.8B)135 tks/s (3.8B)TBD
#5Yi75 tks/s (6B)566 tks/s (6B)105 tks/s (6B)TBD
#6StableLM99 tks/s (3B)TBD-TBD
#7BigCode/StarCodeTBDTBDTBD-TBD
#8ChatGLMTBDTBDTBD-TBD
#9QWen2 (1.8B, 7B)148 tks/s (1.8B)784 tks/s (1.8B)-TBD
#10Google Gemma130 tks/s (2B)TBD73 tks/s (Gemma2-9B, Marlin)587 tks/s (Gemma2-9B)
#11Blip-large (Multimodal)TBDTBDTBD-TBD
#12Moondream-2 (Multimodal LLM)TBDTBDTBD-TBD

Demo Chat with candle-vllm (~110 tokens/s, LLaMa3.1 8B, 4-bit Marlin, on A100)

https://github.com/user-attachments/assets/66b5b90e-e2ca-4f0b-82d7-99aa9f85568c

Usage

See this folder for some examples.

Step 1: Run Candle-VLLM service (assume llama2-7b model weights downloaded)

curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
sudo apt install libssl-dev
sudo apt install pkg-config
git clone git@github.com:EricLBuehler/candle-vllm.git
cd candle-vllm
cargo run --release -- --port 2000 --weight-path /home/llama2_7b/ llama

You may also run specific model using huggingface model-id, e.g.,

cargo run --release -- --port 2000 --model-id meta-llama/Llama-2-7b-chat-hf llama

Run latest LLaMa3.1 using local weights

cargo run --release -- --port 2000 --weight-path /home/Meta-Llama-3.1-8B-Instruct/ llama3 --temperature 0. --penalty 1.0

Refer to Marlin quantization below for running quantized GPTQ models.

Step 2:

Option 1: Chat with ChatUI (recommended)

Install ChatUI and its dependencies:

git clone git@github.com:guoqingbao/candle-vllm-demo.git
cd candle-vllm-demo
apt install npm #install npm if needed
npm install n -g #update node js if needed
n stable #update node js if needed
npm i -g pnpm #install pnpm manager
pnpm install #install ChatUI dependencies

Launching the ChatUI:

pnpm run dev # run the ChatUI

Trouble shooting for Nodejs error

ENOSPC: System limit for number of file watchers reached

echo fs.inotify.max_user_watches=524288 | sudo tee -a /etc/sysctl.conf && sudo sysctl -p

Option 2: Chat completion request with HTTP post

curl -X POST "http://127.0.0.1:2000/v1/chat/completions" \
     -H "Content-Type: application/json" \
     -H "Authorization: Bearer YOUR_API_KEY" \
     -d '{
           "model": "llama7b",
           "messages": [
               {"role": "user", "content": "Explain how to best learn Rust."}
           ],
           "temperature": 0.7,
          "max_tokens": 128,
          "stop": {"Single":"</s>"}
       }'

Sample response:

{"id":"cmpl-53092967-c9cf-40e0-ae26-d7ac786d59e8","choices":[{"message":{"content":" Learning any programming language requires a combination of theory, practice, and dedication. Here are some steps and resources to help you learn Rust effectively:\n\n1. Start with the basics:\n\t* Understand the syntax and basic structure of Rust programs.\n\t* Learn about variables, data types, loops, and control structures.\n\t* Familiarize yourself with Rust's ownership system and borrowing mechanism.\n2. Read the Rust book:\n\t* The Rust book is an official resource that provides a comprehensive introduction to the language.\n\t* It covers topics such","role":"[INST]"},"finish_reason":"length","index":0,"logprobs":null}],"created":1718784498,"model":"llama7b","object":"chat.completion","usage":{"completion_tokens":129,"prompt_tokens":29,"total_tokens":158}}

Option 3: Chat completion with with openai package

In your terminal, install the openai Python package by running pip install openai. I use version 1.3.5.

Then, create a new Python file and write the following code:

import openai

openai.api_key = "EMPTY"

openai.base_url = "http://localhost:2000/v1/"

completion = openai.chat.completions.create(
    model="llama",
    messages=[
        {
            "role": "user",
            "content": "Explain how to best learn Rust.",
        },
    ],
    max_tokens = 64,
)
print(completion.choices[0].message.content)

After the candle-vllm service is running, run the Python script and enjoy efficient inference with an OpenAI compatible API server!

Batched requests

Install openai API first

python3 -m pip install openai

Run the benchmark test

python3 examples/benchmark.py --batch 16 --max_tokens 1024

Refer to examples/benchmark.py

async def benchmark():
    model = "mistral7b"
    max_tokens = 1024
    # 16 requests
    prompts = ["Explain how to best learn Rust.", 
               "Please talk about deep learning in 100 words.", 
               "Do you know the capital city of China? Talk the details of you known.", 
               "Who is the best female actor in the world? Explain why.",
               "How to dealing with depression?",
               "How to make money in short time?",
               "What is the future trend of large language model?",
               "The famous tech companies in the world.",
               "Explain how to best learn Rust.", 
               "Please talk about deep learning in 100 words.", 
               "Do you know the capital city of China? Talk the details of you known.", 
               "Who is the best female actor in the world? Explain why.",
               "How to dealing with depression?",
               "How to make money in short time?",
               "What is the future trend of large language model?",
               "The famous tech companies in the world."]
    
    # send 16 chat requests at the same time
    tasks: List[asyncio.Task] = []
    for i in range(len(prompts)):
        tasks.append(
            asyncio.create_task(
                chat_completion(model, max_tokens, prompts[i]))
        )

    # obtain the corresponding stream object for each request
    outputs: List[Stream[ChatCompletionChunk]] = await asyncio.gather(*tasks)

    # tasks for streaming chat responses
    tasks_stream: List[asyncio.Task] = []
    for i in range(len(outputs)):
        tasks_stream.append(
            asyncio.create_task(
                stream_response(i, outputs[i]))
        )

    # gathering the response texts
    outputs: List[(int, str)] = await asyncio.gather(*tasks_stream)

    # print the results, you may find chat completion statistics in the backend server (i.e., candle-vllm)
    for idx, output in outputs:
        print("\n\n Response {}: \n\n {}".format(idx, output))


asyncio.run(benchmark())

GPTQ/Marlin 4-bit quantization

Candle-vllm now supports GPTQ (Marlin kernel), you may supply the quant (marlin) parameter if you have Marlin format quantized weights, such as:

cargo run --release -- --port 2000 --dtype f16 --weight-path /home/Meta-Llama-3.1-8B-Instruct-GPTQ-INT4-Marlin/ llama3 --quant marlin --temperature 0. --penalty 1.

You may also use AutoGPTQ to transform a model to marlin format by loading the (quantized) model, supplying the use_marlin=True in AutoGPTQ and resaving it with "save_pretrained".

Note: only 4-bit GPTQ (marlin format) quantization supported at the moment, and the input data type should be f16 (--dtype f16) or bf16 (--dtype bf16). You need rename the transformed marlin weight to "model.safetensors" and copy the "tokenizer.json" from the source model folder.

In-situ quantization (or in-situ marlin conversion)

Candle-vllm now supports in-situ quantization, allowing the transformation of default weights (F32/F16/BF16) or 4-bit GPTQ weights into any GGML format (or marlin format) during model loading. This feature helps conserve GPU memory (or speedup inference performance through marlin kernel), making it more efficient for consumer-grade GPUs (e.g., RTX 4090). To use this feature, simply supply the quant parameter when running candle-vllm.

For unquantized models:

cargo run --release -- --port 2000 --weight-path /home/Meta-Llama-3.1-8B-Instruct/ llama3 --quant q4k

For quantized 4-bit GPTQ model:

cargo run --release -- --port 2000 --weight-path /home/mistral_7b-int4/ mistral --quant marlin

Options for quant parameters: ["q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "q2k", "q3k","q4k","q5k","q6k", "marlin"]

Please note:

  1. It may takes few minutes to load F32/F16/BF16 models into quantized;

  2. Marlin format in-situ conversion only support 4-bit GPTQ (with sym=True, groupsize=128 or -1, desc_act=False).

Usage Help

For general configuration help, run cargo run -- --help.

For model-specific help, run cargo run -- --port 2000 <MODEL_TYPE> --help

For local model weights, run cargo run --release -- --port 2000 --weight-path /home/llama2_7b/ llama, change the path when needed.

MODEL_TYPE = ["llama", "llama3", "mistral", "phi2", "phi3", "qwen2", "gemma", "yi", "stable-lm"]

WEIGHT_FILE_PATH = Corresponding weight path for the given model type

cargo run --release -- --port 2000 --weight-path <WEIGHT_FILE_PATH> <MODEL_TYPE>

or

MODEL_ID = Huggingface model id

cargo run --release -- --port 2000 --model-id <MODEL_ID> <MODEL_TYPE>

For kvcache configuration, set kvcache_mem_cpu and kvcache_mem_gpu, default 4GB CPU memory and 4GB GPU memory for kvcache.

For chat history settings, set record_conversation to true to let candle-vllm remember chat history. By default, candle-vllm does not record chat history; instead, the client sends both the messages and the contextual history to candle-vllm. If record_conversation is set to true, the client sends only new chat messages to candle-vllm, and candle-vllm is responsible for recording the previous chat messages. However, this approach requires per-session chat recording, which is not yet implemented, so the default approach record_conversation=false is recommended.

For chat streaming, the stream flag in chat request need to be set to True.

You may supply penalty and temperature to the model to prevent potential repetitions, for example:

cargo run --release -- --port 2000 --weight-path /home/mistral_7b/ mistral --repeat-last-n 64 --penalty 1.1 --temperature 0.7

--max-gen-tokens parameter is used to control the maximum output tokens per chat response. The value will be set to 1/5 of max_sequence_len by default.

For consumer GPUs, it is suggested to run the models under GGML formats (or Marlin format), e.g.,

cargo run --release -- --port 2000 --weight-path /home/Meta-Llama-3.1-8B-Instruct/ llama3 --quant q4k

where quant is one of ["q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "q2k", "q3k","q4k","q5k","q6k", "marlin"].

Report issue

Installing candle-vllm is as simple as the following steps. If you have any problems, please create an issue.

Contributing

The following features are planned to be implemented, but contributions are especially welcome:

Resources