# ztxz16/fastllm

fastllm是后端无依赖的高性能大模型推理库。同时支持张量并行推理稠密模型和混合模式推理MOE模型，任意10G以上显卡即可推理满血DeepSeek。双路9004/9005服务器+单显卡部署DeepSeek满血满精度原版模型，单并发20tps；INT4量化模型单并发30tps，多并发可达60+。

Repository: https://github.com/ztxz16/fastllm
Canonical: https://ross.abutalabs.com/products/fastllm
Language: C++
License: Apache-2.0
License Family: permissive
Last push: 2026-08-26T06:22:32+00:00

## Health v2 (maintenance only)
Score: 74/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 99, release rhythm 35, longevity 86
- inputs: {"age_days": 1208, "days_push": 7, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 4945, forks 485 (observed 2026-08-28T04:09:03.624006+00:00)

## What it is
fastllm is a high-performance C++ LLM inference library with its own custom operators, requiring no PyTorch dependency. It supports tensor-parallel inference of dense models and hybrid CPU+GPU inference of MOE models, enabling full DeepSeek 671B inference on a single 10GB+ GPU.

## Use cases
- run DeepSeek R1 671B on a single GPU with limited VRAM
- serve an OpenAI-compatible API for local LLMs
- chat with Qwen or Llama models from the terminal
- deploy MOE models with CPU+GPU hybrid tensor parallelism
- run FP8 inference on older or domestic GPUs
- quantize models to INT4 or dynamic quantization for faster inference
- deploy LLM inference on AMD ROCm or Chinese domestic accelerators

## When to choose
- you need to run very large MOE models like DeepSeek 671B with limited GPU memory
- you want a PyTorch-free inference engine with broad GPU compatibility including older cards
- you need tensor parallelism across an arbitrary number of GPUs including odd counts
- you target domestic Chinese accelerators or AMD GPUs

## When to avoid
- you need the broadest ecosystem integration and community tooling of vLLM or SGLang
- you require macOS or Apple Silicon support
- you need extensive fine-tuning or training features rather than inference only

## Facets
- artifact type: library
- maturity: active
- function: llm-inference, gpu-computing, cli, http-server, chat-interface
- domain: large-language-models, artificial-intelligence, developer-tools
- platform: windows, python, cpp, cli
- tags: moe-inference, tensor-parallelism, cpu-gpu-hybrid, fp8, quantization, deepseek, openai-compatible-api, self-contained-operators, no-pytorch-dependency, command-line, linux, gpu, web-server, android

## Member repositories
- ztxz16/fastllm (main) score 74

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:09:03.624006+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T18:17:54.046777+00:00, confidence not recorded.
  - readme: https://github.com/ztxz16/fastllm (fetched 2026-08-28T04:09:03.624006+00:00, sha fd3d330cb892)
- Data as of 2026-08-30T08:39:29.467469+00:00.