KTransformers, pronounced as Quick Transformers, is designed to enhance your 🤗 <a href="https://github.com/huggingface/transformers">Transformers</a> experience with advanced kernel optimizations and placement/parallelism strategies.
<br/><br/>
KTransformers is a flexible, Python-centric framework designed with extensibility at its core.
By implementing and injecting an optimized module with a single line of code, users gain access to a Transformers-compatible
interface, RESTful APIs compliant with OpenAI and Ollama, and even a simplified ChatGPT-like web UI.
<br/><br/>
Our vision for KTransformers is to serve as a flexible platform for experimenting with innovative LLM inference optimizations. Please let us know if you need any other features.
***Mar 5, 2025**: Support unsloth 1.58/2.51 bits weights and [IQ1_S/FP8 hybrid](./doc/en/fp8_kernel.md) weights. Support 139K [Longer Context](./doc/en/DeepseekR1_V3_tutorial.md#v022-longer-context) for DeepSeek-V3 and R1 in 24GB VRAM.
**[Write With Transformer](https://transformer.huggingface.co)**，由抱抱脸团队打造，是一个文本生成的官方 demo。
***Feb 25, 2025**: Support [FP8 GPU kernel](./doc/en/fp8_kernel.md) for DeepSeek-V3 and R1; [Longer Context](./doc/en/DeepseekR1_V3_tutorial.md#v022-longer-context).
***Feb 15, 2025**: Longer Context (from 4K to 8K for 24GB VRAM) & Slightly Faster Speed （+15%, up to 16 Tokens/s), update [docs](./doc/en/DeepseekR1_V3_tutorial.md) and [online books](https://kvcache-ai.github.io/ktransformers/).
***Feb 10, 2025**: Support Deepseek-R1 and V3 on single (24GB VRAM)/multi gpu and 382G DRAM, up to 3~28x speedup. For detailed show case and reproduction tutorial, see [here](./doc/en/DeepseekR1_V3_tutorial.md).
***Aug 28, 2024**: Decrease DeepseekV2's required VRAM from 21G to 11G.
***Aug 15, 2024**: Update detailed [tutorial](doc/en/injection_tutorial.md) for injection and multi-GPU.
***Aug 14, 2024**: Support llamfile as linear backend.
***Aug 12, 2024**: Support multiple GPU; Support new model: mixtral 8\*7B and 8\*22B; Support q2k, q3k, q5k dequant on gpu.
***Aug 9, 2024**: Support windows native.
<!-- * **Aug 28, 2024**: Support 1M context under the InternLM2.5-7B-Chat-1M model, utilizing 24GB of VRAM and 150GB of DRAM. The detailed tutorial is [here](./doc/en/long_context_tutorial.md). -->
<h2id="show-cases">🌟 Show Cases</h2>
<div>
<h3>GPT-4/o1-level Local VSCode Copilot on a Desktop with only 24GB VRAM</h3>
-**[NEW!!!] Local 671B DeepSeek-Coder-V3/R1:** Running its Q4_K_M version using only 14GB VRAM and 382GB DRAM([Tutorial](./doc/en/DeepseekR1_V3_tutorial.md)).
- Compared to 4.51 tokens/s in llama.cpp with 2×32 cores, achieving up to **3.03× speedup**.
- Upcoming Open Source Release:
- AMX optimizations and selective expert activation will be open-sourced in V0.3.
- Currently available only in preview binary distribution, which can be downloaded [here](./doc/en/DeepseekR1_V3_tutorial.md).
-**Local 236B DeepSeek-Coder-V2:** Running its Q4_K_M version using only 21GB VRAM and 136GB DRAM, attainable on a local desktop machine, which scores even better than GPT4-0613 in [BigCodeBench](https://huggingface.co/blog/leaderboard-bigcodebench).
-**Faster Speed:** Achieving 126 tokens/s for 2K prompt prefill and 13.6 tokens/s for generation through MoE offloading and injecting advanced kernels from [Llamafile](https://github.com/Mozilla-Ocho/llamafile/tree/main) and [Marlin](https://github.com/IST-DASLab/marlin).
-**VSCode Integration:** Wrapped into an OpenAI and Ollama compatible API for seamless integration as a backend for [Tabby](https://github.com/TabbyML/tabby) and various other frontends.
***1M Context InternLM 2.5 7B**: Operates at full bf16 precision, utilizing 24GB VRAM and 150GB DRAM, which is feasible on a local desktop setup. It achieves a 92.88% success rate on the 1M "Needle In a Haystack" test and 100% on the 128K NIAH test.
***Enhanced Speed**: Reaches 16.91 tokens/s for generation with a 1M context using sparse attention, powered by llamafile kernels. This method is over 10 times faster than full attention approach of llama.cpp.
***Flexible Sparse Attention Framework**: Offers a flexible block sparse attention framework for CPU offloaded decoding. Compatible with SnapKV, Quest, and InfLLm. Further information is available [here](./doc/en/long_context_introduction.md).
-->
<strong>More advanced features will coming soon, so stay tuned!</strong>
<h2 id="quick-start">🚀 Quick Start</h2>
Getting started with KTransformers is simple! Follow the steps below to set up and start using it.
### 📥 Installation
To install KTransformers, follow the official [Installation Guide](https://kvcache-ai.github.io/ktransformers/en/install.html).
<h2 id="tutorial">📃 Brief Injection Tutorial</h2>
At the heart of KTransformers is a user-friendly, template-based injection framework.
This allows researchers to easily replace original torch modules with optimized variants. It also simplifies the process of combining multiple optimizations, allowing the exploration of their synergistic effects.
Given that vLLM already serves as a great framework for large-scale deployment optimizations, KTransformers is particularly focused on local deployments that are constrained by limited resources. We pay special attention to heterogeneous computing opportunities, such as GPU/CPU offloading of quantized models. For example, we support the efficient <a herf="https://github.com/Mozilla-Ocho/llamafile/tree/main">Llamafile</a> and <a herf="https://github.com/IST-DASLab/marlin">Marlin</a> kernels for CPU and GPU, respectively. More details can be found <a herf="doc/en/operators/llamafile.md">here</a>.
<h3>Example Usage</h3>
To utilize the provided kernels, users only need to create a YAML-based injection template and add the call to `optimize_and_load_gguf` before using the Transformers model.
In this example, the AutoModel is first initialized on the meta device to avoid occupying any memory resources. Then, `optimize_and_load_gguf` iterates through all sub-modules of the model, matches rules specified in your YAML rule file, and replaces them with advanced modules as specified.
export USE_FASTPT_CUDA=True
bash install_dcu.sh
After injection, the original `generate` interface is available, but we also provide a compatible `prefill_and_generate` method, which enables further optimizations like CUDAGraph to improve generation speed.
<h3>How to custom your model</h3>
A detailed tutorial of the injection and multi-GPU using DeepSeek-V2 as an example is given [here](doc/en/injection_tutorial.md).
Below is an example of a YAML template for replacing all original Linear modules with Marlin, an advanced 4-bit quantization kernel.
```yaml
-match:
name:"^model\\.layers\\..*$"# regular expression
class:torch.nn.Linear# only match modules matching name and class simultaneously
replace:
class:ktransformers.operators.linear.KTransformerLinear# optimized Kernel on quantized data types
device:"cpu"# which devices to load this module when initializing
Each rule in the YAML file has two parts: `match` and `replace`. The `match` part specifies which module should be replaced, and the `replace` part specifies the module to be injected into the model along with the initialization keywords.
You can find example rule templates for optimizing DeepSeek-V2 and Qwen2-57B-A14, two SOTA MoE models, in the [ktransformers/optimize/optimize_rules](ktransformers/optimize/optimize_rules) directory. These templates are used to power the `local_chat.py` demo.
## 模型架构
If you are interested in our design principles and the implementation of the injection framework, please refer to the [design document](doc/en/deepseek-v2-injection.md).
The development of KTransformer is based on the flexible and versatile framework provided by Transformers. We also benefit from advanced kernels such as GGUF/GGML, Llamafile, Marlin, sglang and flashinfer. We are planning to contribute back to the community by upstreaming our modifications.
KTransformer is actively maintained and developed by contributors from the <a href="https://madsys.cs.tsinghua.edu.cn/">MADSys group</a> at Tsinghua University and members from <a href="http://approaching.ai/">Approaching.AI</a>. We welcome new contributors to join us in making KTransformer faster and easier to use.
If you have any questions, feel free to open an issue. Alternatively, you can join our WeChat group for further discussion. QR Code: [WeChat Group](WeChatGroup.png)
## 了解更多
<h2 id="FAQ">🙋 FAQ</h2>
| 章节 | 描述 |
|-|-|
| [文档](https://huggingface.co/docs/transformers/) | 完整的 API 文档和教程 |
title="Transformers: State-of-the-Art Natural Language Processing",
author="Thomas Wolf and Lysandre Debut and Victor Sanh and Julien Chaumond and Clement Delangue and Anthony Moi and Pierric Cistac and Tim Rault and Rémi Louf and Morgan Funtowicz and Joe Davison and Sam Shleifer and Patrick von Platen and Clara Ma and Yacine Jernite and Julien Plu and Canwen Xu and Teven Le Scao and Sylvain Gugger and Mariama Drame and Quentin Lhoest and Alexander M. Rush",
booktitle="Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations",
month=oct,
year="2020",
address="Online",
publisher="Association for Computational Linguistics",
