GitHub - facebookresearch/fastgen: Simple high-throughput inference library

Fastgen

High-throughput inference from dense transformers.

Fastgen is an LLM inference library featuring state-of-the-art techniques: batched inference, cuda graphs, paged attention, chunked prefills, host-side kv-cache, tensor parallelism, cpu/gpu profiling.

You can see it as a mini (~3kloc) vLLM that is easily imported into larger systems such as RL loops or synthetic data generation pipelines.

The following HuggingFace models are known to be supported: Llama 3.1 8B, Mistral 7B, Ministral 8B, R1-Llama 8B and 70B, R1-Qwen 14B, Qwen2.5 7B and 14B; more may work.

Benchmarks

	Fastgen	vLLM (0.8.4)
Mistral 7B / 1x H100	3430 toks/s	3390 toks/s
Qwen2.5 14B / 1x H100	1700 toks/s	1700 toks/s
Llama 70B / 8x H100	2260 toks/s	2230 toks/s

The throughputs presented in this table were collected using fgserve (see below) and vllm serve with a load of 128 decoding steps for 512 prompts of 1024 tokens. The figure below shows more detailed results for the Mistral 7B model across a range of generation loads.

Note

When running tests against fgserve you might observe reduced performance for short prompts. This lower performance is due to the unoptimized threaded HTTP serving code and causes Fastgen to run on small (inefficient) batches. Production applications using Fastgen should adopt process-based parallelism.

Demo Tools

The library comes with two tools: fgchat and fgserve. The first tool is a basic demo implementing a command-line chat; the second one serves a bare-bones OpenAI chat API.

fgchat

$ # on an RTX A5000
$ # accelerate inference with '--tensor-parallel N' if you have N GPUs
$ fgchat mistralai/Mistral-7B-Instruct-v0.3  
loading: 100%|███████████████████████▉| 14.5G/14.5G [00:00<00:00, 75.9GB/s]
Ready to chat!
  Available commands:
    /r <N> -- reset conversation before round N
    /m     -- input a multiline message, end your input
              with a single line containing '.'
    /q     -- quit

---
1> hello
Hello! How can I help you today? [...]
--- [33 tokens, 26.554toks/s]
2> can you give me some makeup advice?
Of course! Here are some general makeup tips that might be helpful:

[...]
--- [489 tokens, 34.715toks/s]
3> /q

fgserve

$ fgserve mistralai/Mistral-7B-Instruct-v0.3  
...
$ cat test.py
from openai import OpenAI
client = OpenAI(base_url="http://localhost:5678", api_key="foo")

completion = client.chat.completions.create(
    model="gpt-4o-mini",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
    ] + [
        {
            "role": "user",
            "content": "Write a haiku about recursion in programming."
        }
    ]
)
print(completion.choices[0].message.content)

$ python test.py
Code dives within,
Echoes of self repeat,
Solving the unseen.

Acknowledgements

Thanks to:

The vLLM project for inspiration and for most of the innovations (paged attention, chunked prefills, ...) leveraged in Fastgen.
Daniel Haziza (@danthe3rd) for constructive initial design discussions as well as continued advice on achieving high GPU utilization.

License

Fastgen is BSD 3-Clause licensed, as found in the LICENSE file.