RYDE-WORK/ktransformers
1
0
Fork 0
mirror of https://github.com/RYDE-WORK/ktransformers.git synced 2026-09-18 15:43:20 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Initial commit

Browse Source
This commit is contained in:
chenxl 2024-07-27 16:06:58 +08:00
commit 18c42e67df
247 changed files with 53775 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
.gitignore vendored Normal file
View File

@ -0,0 +1,17 @@
__pycache__
build
.vscode
*.so
*.cache
server.db
logs
node_modules
*.nsys-rep
.vs/
*pycache*
*build/
*/third_party/*
.DS_Store
compile_commands.json
*.egg-info*
*dist/

6
.gitmodules vendored Normal file
View File

@ -0,0 +1,6 @@
[submodule "third_party/llama.cpp"]
path = third_party/llama.cpp
url = https://github.com/ggerganov/llama.cpp.git
[submodule "third_party/pybind11"]
path = third_party/pybind11
url = https://github.com/pybind/pybind11.git

6
.pylintrc Normal file
View File

@ -0,0 +1,6 @@
[MASTER]
extension-pkg-whitelist=pydantic
max-line-length=120
[MESSAGES CONTROL]
disable=missing-function-docstring

201
LICENSE Normal file
View File

@ -0,0 +1,201 @@
Apache License
Version 2.0, January 2004
http://www.apache.org/licenses/
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
1. Definitions.
"License" shall mean the terms and conditions for use, reproduction,
and distribution as defined by Sections 1 through 9 of this document.
"Licensor" shall mean the copyright owner or entity authorized by
the copyright owner that is granting the License.
"Legal Entity" shall mean the union of the acting entity and all
other entities that control, are controlled by, or are under common
control with that entity. For the purposes of this definition,
"control" means (i) the power, direct or indirect, to cause the
direction or management of such entity, whether by contract or
otherwise, or (ii) ownership of fifty percent (50%) or more of the
outstanding shares, or (iii) beneficial ownership of such entity.
"You" (or "Your") shall mean an individual or Legal Entity
exercising permissions granted by this License.
"Source" form shall mean the preferred form for making modifications,
including but not limited to software source code, documentation
source, and configuration files.
"Object" form shall mean any form resulting from mechanical
transformation or translation of a Source form, including but
not limited to compiled object code, generated documentation,
and conversions to other media types.
"Work" shall mean the work of authorship, whether in Source or
Object form, made available under the License, as indicated by a
copyright notice that is included in or attached to the work
(an example is provided in the Appendix below).
"Derivative Works" shall mean any work, whether in Source or Object
form, that is based on (or derived from) the Work and for which the
editorial revisions, annotations, elaborations, or other modifications
represent, as a whole, an original work of authorship. For the purposes
of this License, Derivative Works shall not include works that remain
separable from, or merely link (or bind by name) to the interfaces of,
the Work and Derivative Works thereof.
"Contribution" shall mean any work of authorship, including
the original version of the Work and any modifications or additions
to that Work or Derivative Works thereof, that is intentionally
submitted to Licensor for inclusion in the Work by the copyright owner
or by an individual or Legal Entity authorized to submit on behalf of
the copyright owner. For the purposes of this definition, "submitted"
means any form of electronic, verbal, or written communication sent
to the Licensor or its representatives, including but not limited to
communication on electronic mailing lists, source code control systems,
and issue tracking systems that are managed by, or on behalf of, the
Licensor for the purpose of discussing and improving the Work, but
excluding communication that is conspicuously marked or otherwise
designated in writing by the copyright owner as "Not a Contribution."
"Contributor" shall mean Licensor and any individual or Legal Entity
on behalf of whom a Contribution has been received by Licensor and
subsequently incorporated within the Work.
2. Grant of Copyright License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
copyright license to reproduce, prepare Derivative Works of,
publicly display, publicly perform, sublicense, and distribute the
Work and such Derivative Works in Source or Object form.
3. Grant of Patent License. Subject to the terms and conditions of
this License, each Contributor hereby grants to You a perpetual,
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
(except as stated in this section) patent license to make, have made,
use, offer to sell, sell, import, and otherwise transfer the Work,
where such license applies only to those patent claims licensable
by such Contributor that are necessarily infringed by their
Contribution(s) alone or by combination of their Contribution(s)
with the Work to which such Contribution(s) was submitted. If You
institute patent litigation against any entity (including a
cross-claim or counterclaim in a lawsuit) alleging that the Work
or a Contribution incorporated within the Work constitutes direct
or contributory patent infringement, then any patent licenses
granted to You under this License for that Work shall terminate
as of the date such litigation is filed.
4. Redistribution. You may reproduce and distribute copies of the
Work or Derivative Works thereof in any medium, with or without
modifications, and in Source or Object form, provided that You
meet the following conditions:
(a) You must give any other recipients of the Work or
Derivative Works a copy of this License; and
(b) You must cause any modified files to carry prominent notices
stating that You changed the files; and
(c) You must retain, in the Source form of any Derivative Works
that You distribute, all copyright, patent, trademark, and
attribution notices from the Source form of the Work,
excluding those notices that do not pertain to any part of
the Derivative Works; and
(d) If the Work includes a "NOTICE" text file as part of its
distribution, then any Derivative Works that You distribute must
include a readable copy of the attribution notices contained
within such NOTICE file, excluding those notices that do not
pertain to any part of the Derivative Works, in at least one
of the following places: within a NOTICE text file distributed
as part of the Derivative Works; within the Source form or
documentation, if provided along with the Derivative Works; or,
within a display generated by the Derivative Works, if and
wherever such third-party notices normally appear. The contents
of the NOTICE file are for informational purposes only and
do not modify the License. You may add Your own attribution
notices within Derivative Works that You distribute, alongside
or as an addendum to the NOTICE text from the Work, provided
that such additional attribution notices cannot be construed
as modifying the License.
You may add Your own copyright statement to Your modifications and
may provide additional or different license terms and conditions
for use, reproduction, or distribution of Your modifications, or
for any such Derivative Works as a whole, provided Your use,
reproduction, and distribution of the Work otherwise complies with
the conditions stated in this License.
5. Submission of Contributions. Unless You explicitly state otherwise,
any Contribution intentionally submitted for inclusion in the Work
by You to the Licensor shall be under the terms and conditions of
this License, without any additional terms or conditions.
Notwithstanding the above, nothing herein shall supersede or modify
the terms of any separate license agreement you may have executed
with Licensor regarding such Contributions.
6. Trademarks. This License does not grant permission to use the trade
names, trademarks, service marks, or product names of the Licensor,
except as required for reasonable and customary use in describing the
origin of the Work and reproducing the content of the NOTICE file.
7. Disclaimer of Warranty. Unless required by applicable law or
agreed to in writing, Licensor provides the Work (and each
Contributor provides its Contributions) on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
implied, including, without limitation, any warranties or conditions
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
PARTICULAR PURPOSE. You are solely responsible for determining the
appropriateness of using or redistributing the Work and assume any
risks associated with Your exercise of permissions under this License.
8. Limitation of Liability. In no event and under no legal theory,
whether in tort (including negligence), contract, or otherwise,
unless required by applicable law (such as deliberate and grossly
negligent acts) or agreed to in writing, shall any Contributor be
liable to You for damages, including any direct, indirect, special,
incidental, or consequential damages of any character arising as a
result of this License or out of the use or inability to use the
Work (including but not limited to damages for loss of goodwill,
work stoppage, computer failure or malfunction, or any and all
other commercial damages or losses), even if such Contributor
has been advised of the possibility of such damages.
9. Accepting Warranty or Additional Liability. While redistributing
the Work or Derivative Works thereof, You may choose to offer,
and charge a fee for, acceptance of support, warranty, indemnity,
or other liability obligations and/or rights consistent with this
License. However, in accepting such obligations, You may act only
on Your own behalf and on Your sole responsibility, not on behalf
of any other Contributor, and only if You agree to indemnify,
defend, and hold each Contributor harmless for any liability
incurred by, or claims asserted against, such Contributor by reason
of your accepting any such warranty or additional liability.
END OF TERMS AND CONDITIONS
APPENDIX: How to apply the Apache License to your work.
To apply the Apache License to your work, attach the following
boilerplate notice, with the fields enclosed by brackets "[]"
replaced with your own identifying information. (Don't include
the brackets!) The text should be enclosed in the appropriate
comment syntax for the file format. We also recommend that a
file or class name and description of purpose be included on the
same "printed page" as the copyright notice for easier
identification within third-party archives.
Copyright [yyyy] [name of copyright owner]
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.

11
MANIFEST.in Normal file
View File

@ -0,0 +1,11 @@
graft third_party
graft ktransformers
graft local_chat.py
include LICENSE README.md
prune ktransformers/website
prune ktransformers/logs
prune ktransformers.egg-info
prune third_party/llama.cpp/models
graft ktransformers/website/dist
global-exclude __pycache__
include KTransformersOps.*.so

284
README.md Normal file
View File

@ -0,0 +1,284 @@
<div align="center">
<!-- <h1>KTransformers</h1> -->
<p align="center">
<picture>
<img alt="DeepSeek-Coder-V2 Score" src="https://github.com/user-attachments/assets/9fa710bf-1389-46b7-b9d2-3f67b98bd7a6" width=50%>
</picture>
</p>
<h3>A Flexible Framework for Experiencing Cutting-edge LLM Inference Optimizations</h3>
<strong><a href="#show-cases">🔥 Show Cases</a> | <a href="#quick-start">🚀 Quick Start</a> | <a href="#tutorial">📃 Tutorial</a> | <a href="https://github.com/kvcache-ai/ktransformers/discussions">💬 Discussion </a> </strong>
</div>
<h2 id="intro">🎉 Introduction</h2>
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.
<h2 id="show-cases">🔥 Show Cases</h2>
<h3>GPT-4-level Local VSCode Copilot on a Desktop with only 24GB VRAM</h3>
<p align="center">
https://github.com/user-attachments/assets/3f85780e-aa53-4d2f-91b2-5585c8dade85
</p>
- **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).
<p align="center">
<picture>
<img alt="DeepSeek-Coder-V2 Score" src="https://github.com/user-attachments/assets/81efb94f-f859-4413-b6e0-d986508ad667" width=80%>
</picture>
</p>
- **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.
<p align="center">
<!-- <img alt="Tabby integration" src="https://XXXX.png" width=55%> -->
https://github.com/user-attachments/assets/e6e27cb3-8372-44e6-8f1f-34402eae56c1
</p>
<strong>More advanced features will coming soon, so stay tuned!</strong>
<h2 id="quick-start">🚀 Quick Start</h2>
<h3>Preparation</h3>
Some preparation:
- CUDA 12.1 and above, if you didn't have it yet, you may install from [here](https://developer.nvidia.com/cuda-downloads).
<!-- ```
export PATH=/usr/local/cuda/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
export CUDA_PATH=/usr/local/cuda
``` -->
- Linux-x86_64 with gcc, g++ and cmake
```sh
sudo apt-get update
sudo apt-get install gcc g++ cmake ninja-build
```
- We recommend using [Conda](https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh) to create a virtual environment with Python=3.11 to run our program.
```sh
conda create --name ktransformers python=3.11
conda activate ktransformers # you may need to run ‘conda init’ and reopen shell first
```
Download source code:
```sh
git clone https://github.com/kvcache-ai/ktransformers.git
cd ktransformers
git submodule init
git submodule update
```
<h3>Local Chat</h3>
We provide a simple command-line local chat Python script that you can run for testing.
> Note that this is a very simple test tool only support one round chat without any memory about last input, if you want to try full ability of the model, you may go to [RESTful API and Web UI](#id_666). We use the DeepSeek-V2-Lite-Chat-GGUF model as an example here. But we alse support other models, you can replace it with any other model that you want to test.
<h4>Install</h4>
```sh
bash install.sh
```
<h4>Run Example</h4>
```shell
# Begin from root of your cloned repo!
# Begin from root of your cloned repo!!
# Begin from root of your cloned repo!!!
# Download mzwing/DeepSeek-V2-Lite-Chat-GGUF from huggingface
mkdir DeepSeek-V2-Lite-Chat-GGUF
cd DeepSeek-V2-Lite-Chat-GGUF
wget https://huggingface.co/mzwing/DeepSeek-V2-Lite-Chat-GGUF/resolve/main/DeepSeek-V2-Lite-Chat.Q4_K_M.gguf -O DeepSeek-V2-Lite-Chat.Q4_K_M.gguf
cd .. # Move to repo's root dir
# Start local chat
python ktransformers/local_chat.py --model_path deepseek-ai/DeepSeek-V2-Lite-Chat --gguf_path ./DeepSeek-V2-Lite-Chat-GGUF
# If you see “OSError: We couldn't connect to 'https://huggingface.co' to load this file”, try：
# GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite
# python ktransformers/local_chat.py --model_path ./DeepSeek-V2-Lite --gguf_path ./DeepSeek-V2-Lite-Chat-GGUF
```
It features the following arguments:
- `--model_path` (required): Name of the model (such as "deepseek-ai/DeepSeek-V2-Lite-Chat" which will automatically download configs from [Hugging Face](https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite)). Or if you already got local files you may directly use that path to initialize the model.
>Note: <strong>.safetensors</strong> files are not required in the directory. We only need config files to build model and tokenizer.
- `--gguf_path` (required): Path of a directory containing GGUF files which could that can be downloaded from [Hugging Face](https://huggingface.co/mzwing/DeepSeek-V2-Lite-Chat-GGUF/tree/main) (we only support q4_k_m and q8_0 for now, more formats are coming soon).
- `--optimize_rule_path` (required except for Qwen2Moe and DeepSeek-V2): Path of YAML file containing optimize rules. There are two rule files pre-written in the [ktransformers/optimize/optimize_rules](ktransformers/optimize/optimize_rules) directory for optimizing DeepSeek-V2 and Qwen2-57B-A14, two SOTA MoE models.
- `--max_new_tokens`: Int (default=1000). Maximum number of new tokens to generate.
- `--cpu_infer`: Int (default=10). The number of CPUs used for inference. Should ideally be set to the (total number of cores - 2).
<h3 id="supported-model"> Supported Model</h3>
| Model Name | Model Size | VRAM | Minimum DRAM | Recommended DRAM |
| ---- | ---- | ---- | ---- | ---- |
| DeepSeek-V2-q4_k_m | 133G | 24G | 136G | 192G |
| Qwen2-57B-A14B-Instruct-q4_k_m | 33G | 8G | 34G | 64G |
| DeepSeek-V2-Lite-q4_k_m | 9.7G | 3G | 13G | 16G |
More will come soon. Please let us know which models you are most interested in.
Be aware that you need to be subject to their corresponding model licenses when using [DeepSeek](https://huggingface.co/deepseek-ai/DeepSeek-V2/blob/main/LICENSE) and [QWen](https://huggingface.co/Qwen/Qwen2-72B-Instruct/blob/main/LICENSE).
<details>
<summary>Click To Show how to run other examples</summary>
* Qwen2-57B
```sh
pip install flash_attn # For Qwen2
mkdir Qwen2-57B-GGUF && cd Qwen2-57B-GGUF
wget https://huggingface.co/Qwen/Qwen2-57B-A14B-Instruct-GGUF/resolve/main/qwen2-57b-a14b-instruct-q4_k_m.gguf?download=true -O qwen2-57b-a14b-instruct-q4_k_m.gguf
cd ..
python ktransformers/local_chat.py --model_name Qwen/Qwen2-57B-A14B-Instruct --gguf_path ./Qwen2-57B-GGUF
# If you see “OSError: We couldn't connect to 'https://huggingface.co' to load this file”, try：
# GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/Qwen/Qwen2-57B-A14B-Instruct
# python ktransformers/local_chat.py --model_path ./Qwen2-57B-A14B-Instruct --gguf_path ./DeepSeek-V2-Lite-Chat-GGUF
```
* DeepseekV2
```sh
mkdir DeepSeek-V2-Chat-0628-GGUF && cd DeepSeek-V2-Chat-0628-GGUF
# Download weights
wget https://huggingface.co/bartowski/DeepSeek-V2-Chat-0628-GGUF/resolve/main/DeepSeek-V2-Chat-0628-Q4_K_M/DeepSeek-V2-Chat-0628-Q4_K_M-00001-of-00004.gguf -o DeepSeek-V2-Chat-0628-Q4_K_M-00001-of-00004.gguf
wget https://huggingface.co/bartowski/DeepSeek-V2-Chat-0628-GGUF/resolve/main/DeepSeek-V2-Chat-0628-Q4_K_M/DeepSeek-V2-Chat-0628-Q4_K_M-00002-of-00004.gguf -o DeepSeek-V2-Chat-0628-Q4_K_M-00002-of-00004.gguf
wget https://huggingface.co/bartowski/DeepSeek-V2-Chat-0628-GGUF/resolve/main/DeepSeek-V2-Chat-0628-Q4_K_M/DeepSeek-V2-Chat-0628-Q4_K_M-00003-of-00004.gguf -o DeepSeek-V2-Chat-0628-Q4_K_M-00003-of-00004.gguf
wget https://huggingface.co/bartowski/DeepSeek-V2-Chat-0628-GGUF/resolve/main/DeepSeek-V2-Chat-0628-Q4_K_M/DeepSeek-V2-Chat-0628-Q4_K_M-00004-of-00004.gguf -o DeepSeek-V2-Chat-0628-Q4_K_M-00004-of-00004.gguf
cd ..
python ktransformers/local_chat.py --model_name deepseek-ai/DeepSeek-V2-Chat-0628 --gguf_path ./DeepSeek-V2-Chat-0628-GGUF
# If you see “OSError: We couldn't connect to 'https://huggingface.co' to load this file”, try：
# GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/deepseek-ai/DeepSeek-V2-Chat-0628
# python ktransformers/local_chat.py --model_path ./DeepSeek-V2-Chat-0628 --gguf_path ./DeepSeek-V2-Chat-0628-GGUF
```
| model name | weights download link |
|----------|----------|
| Qwen2-57B | [Qwen2-57B-A14B-gguf-Q4K-M](https://huggingface.co/Qwen/Qwen2-57B-A14B-Instruct-GGUF/tree/main) |
| DeepseekV2-coder |[DeepSeek-Coder-V2-Instruct-gguf-Q4K-M](https://huggingface.co/LoneStriker/DeepSeek-Coder-V2-Instruct-GGUF/tree/main) |
| DeepseekV2-chat |[DeepSeek-V2-Chat-gguf-Q4K-M](https://huggingface.co/bullerwins/DeepSeek-V2-Chat-0628-GGUF/tree/main) |
| DeepseekV2-lite | [DeepSeek-V2-Lite-Chat-GGUF-Q4K-M](https://huggingface.co/mzwing/DeepSeek-V2-Lite-Chat-GGUF/tree/main) |
</details>
<!-- pin block for jump -->
<span id='id_666'>
<h3>RESTful API and Web UI</h3>
<h4>Install</h4>
[Optional] If you want to run with website, please [compile the website](./doc/en/api/server/website.md) before execute ```pip install .```
Install ktransformers with source.
```
pip install -r requirements-local_chat.txt
pip install . --no-build-isolation
```
Start without website:
```sh
ktransformers --model_path deepseek-ai/DeepSeek-V2-Lite-Chat --gguf_path /path/to/DeepSeek-V2-Lite-Chat-GGUF --port 10002
```
Start with website:
```sh
ktransformers --model_path deepseek-ai/DeepSeek-V2-Lite-Chat --gguf_path /path/to/DeepSeek-V2-Lite-Chat-GGUF --port 10002 --web True
```
Or you want to start server with transformers, the model_path should include safetensors
```bash
ktransformers --type transformers --model_path /mnt/data/model/Qwen2-0.5B-Instruct --port 10002 --web True
```
Access website with url [http://localhost:10002/web/index.html#/chat](http://localhost:10002/web/index.html#/chat) :
<p align="center">
<picture>
<img alt="Web UI" src="https://github.com/user-attachments/assets/a8eca392-e948-4706-ba9c-743142d8a464" width=80%>
</picture>
</p>
More information about the RESTful API server can be found [here](doc/en/api/server/server.md). You can also find an example of integrating with Tabby [here](doc/en/api/server/tabby.md).
<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.
</br>
<p align="center">
<picture>
<img alt="Inject-Struction" src="https://github.com/user-attachments/assets/b922180e-3e73-4b62-b5a0-5ac98d7052c5" width=50%>
</picture>
</p>
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.
```python
with torch.device("meta"):
model = AutoModelForCausalLM.from_config(config, trust_remote_code=True)
optimize_and_load_gguf(model, optimize_rule_path, gguf_path, config)
...
generated = prefill_and_generate(model, tokenizer, input_tensor.cuda(), max_new_tokens=1000)
```
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.
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>YAML Template</h3>
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
kwargs:
generate_device: "cuda"
generate_linear_type: "QuantizedLinearMarlin"
```
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.
A detailed description of the injection using DeepSeek-V2 as an example is given [here](doc/en/deepseek-v2-injection.md).
<h2 id="ack">Acknowledgment and Contributors</h2>
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, and Marlin. 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.

BIN
doc/assets/BigCodeBench.png Normal file
View File

Binary file not shown.

BIN
doc/assets/DeepSeek-on-KTransformers.PNG Normal file
View File

Binary file not shown.

BIN
doc/assets/InjectStruction.png Normal file
View File

Binary file not shown.

BIN
doc/assets/KTransformers.png Normal file
View File

Binary file not shown.

BIN
doc/assets/cpuinfer.png Normal file
View File

Binary file not shown.

BIN
doc/assets/website.png Normal file
View File

Binary file not shown.

108
doc/en/api/server/api.md Normal file
View File

@ -0,0 +1,108 @@
# API
- [OpenAI ChatCompletion](#openai-chatcompletion)
- [Ollama ChatCompletion](#ollama-chatcompletion)
- [OpenAI Assistant](#openai-assistant)
## OpenAI ChatCompletion
```bash
POST /v1/chat/completions
```
Generate responses based on the selected model.
### Parameters
- `messages`: An array of `message` representing all historical messages. A `message` can be from a user or model (assistant) and includes:
- `role`: Either `user` or `assistant`, indicating the creator of this message.
- `content`: The message from the user or model.
- `model`: The name of the selected model
- `stream`: Either true or false. Indicates whether to use streaming response. If true, model inference results are returned via HTTP event stream.
### Response
- Streaming response: An event stream, each event contains a `chat.completion.chunk`. `chunk.choices[0].delta.content` is the incremental output returned by the model each time.
- Non-streaming response: Not supported yet.
### Example
```bash
curl -X 'POST' \
'http://localhost:9112/v1/chat/completions' \
-H 'accept: application/json' \
-H 'Content-Type: application/json' \
-d '{
"messages": [
{
"content": "tell a joke",
"role": "user"
}
],
"model": "Meta-Llama-3-8B-Instruct",
"stream": true
}'
```
```bash
data:{"id":"c30445e8-1061-4149-a101-39b8222e79e1","object":"chat.completion.chunk","created":1720511671,"model":"not implmented","system_fingerprint":"not implmented","usage":null,"choices":[{"index":0,"delta":{"content":"Why ","role":"assistant","name":null},"logprobs":null,"finish_reason":null}]}
data:{"id":"c30445e8-1061-4149-a101-39b8222e79e1","object":"chat.completion.chunk","created":1720511671,"model":"not implmented","system_fingerprint":"not implmented","usage":null,"choices":[{"index":0,"delta":{"content":"","role":"assistant","name":null},"logprobs":null,"finish_reason":null}]}
data:{"id":"c30445e8-1061-4149-a101-39b8222e79e1","object":"chat.completion.chunk","created":1720511671,"model":"not implmented","system_fingerprint":"not implmented","usage":null,"choices":[{"index":0,"delta":{"content":"couldn't ","role":"assistant","name":null},"logprobs":null,"finish_reason":null}]}
...
data:{"id":"c30445e8-1061-4149-a101-39b8222e79e1","object":"chat.completion.chunk","created":1720511671,"model":"not implmented","system_fingerprint":"not implmented","usage":null,"choices":[{"index":0,"delta":{"content":"two-tired!","role":"assistant","name":null},"logprobs":null,"finish_reason":null}]}
event: done
data: [DONE]
```
## Ollama ChatCompletion
```bash
POST /api/generate
```
Generate responses using the selected model.
### Parameters
- `prompt`: A string representing the input prompt.
- `model`: The name of the selected model
- `stream`: Either true or false. Indicates whether to use streaming responses. If true, returns the model inference results in the form of an HTTP event stream.
### Response
- Streaming response: A stream of JSON responses, each line is a JSON.
- `response`: The incremental result of the model completion.
- `done`: Whether the inference has finished.
- Non-streaming response: Not yet supported.
### 例子
```bash
curl -X 'POST' \
'http://localhost:9112/api/generate' \
-H 'accept: application/json' \
-H 'Content-Type: application/json' \
-d '{
"model": "Meta-Llama-3-8B-Instruct",
"prompt": "tell me a joke",
"stream": true
}'
```
```bash
{"model":"Meta-Llama-3-8B-Instruct","created_at":"2024-07-09 08:13:11.686513","response":"I'll ","done":false}
{"model":"Meta-Llama-3-8B-Instruct","created_at":"2024-07-09 08:13:11.729214","response":"give ","done":false}
...
{"model":"Meta-Llama-3-8B-Instruct","created_at":"2024-07-09 08:13:33.955475","response":"for","done":false}
{"model":"Meta-Llama-3-8B-Instruct","created_at":"2024-07-09 08:13:33.956795","response":"","done":true}
```

BIN
doc/en/api/server/run-tabby.png Normal file
View File

Binary file not shown.

BIN
doc/en/api/server/server-arch.png Normal file
View File

Binary file not shown.

37
doc/en/api/server/server.md Normal file
View File

@ -0,0 +1,37 @@
# Backend Services (Server)
The Server offers fast heterogeneous inference capabilities of ktransformers through an API for external usage.
<img src="server-arch.png" height="600" alt="Server architecture">
## API
The Server provides model inference services externally through a RESTful API, with two methods of interaction: ChatCompletion and Assistant.
- The ChatCompletion interface requires users to provide all historical dialogues at once, after which the model responds. AI service providers (such as [OpenAI](https://platform.openai.com/docs/api-reference/chat/create)) and local inference frameworks (such as [Ollama](https://github.com/ollama/ollama/blob/main/docs/api.md)) both offer the ChatCompletion interface. To ensure compatibility with OpenAI and Ollama, the Server offers APIs that are consistent with theirs. Therefore, applications currently using OpenAI and Ollama can seamlessly switch to our Server. For example: [How to use Tabby and ktransformers locally with a 236B model for code completion?](tabby.md).
- The Assistant is suitable for applications that need to reuse a series of resources and call the model. For instance, in educational applications, developers can create an Assistant named "Second Grade Math Teacher" and set an initial prompt ("You are an experienced second-grade math teacher..."), and upload relevant materials (second grade math textbooks). After creating the Assistant, the application needs to create a Thread to store the dialogues between the user and the model (Message). When calling the model, the application creates a Run to obtain the Assistant's response. Compared to ChatCompletion, the Assistant-enabled Server handles the reuse of conversational contexts and multi-turn dialogues, making model calls in complex scenarios more convenient. The [OpenAI Assistant API](https://platform.openai.com/docs/api-reference/assistants/createAssistant) introduces such an Assistant interface, and the Server provides a consistent API.
These API definitions are located in `server/api`, and their specific usage can be seen [here](api.md).
## Integrating Model Inference Frameworks
The Server uses ktransformers for model calling and inference. It also supports other inference frameworks, such as the already supported [transformers](https://huggingface.co/docs/transformers/index), and plans to support [exllamav2](https://github.com/turboderp/exllamav2). These functionalities are implemented in `server/backend`.
The model inference functionalities of the frameworks are abstracted into a base class `BackendInterfaceBase`. This class includes a function: inference. It takes historical dialogue information messages as input and returns the text result from the model. The inference function adopts an async generator design, allowing the Server to return model responses in a streaming manner.
```python
class BackendInterfaceBase:
async def inference(self, messages, **kwargs)->AsyncIterator[str]:
...
```
This inference function naturally implements the functionality of ChatCompletion because its inputs and outputs are historical dialogues and model responses, respectively. Thus, the ChatCompletion API can directly call the inference function to complete model inference.
Assistant is more complex than ChatCompletion, requiring the Server to store the related state of the Assistant and call the inference function appropriately. The Server maintains a set of Assistant logic in the database, storing the Assistants, Threads, and Messages created by applications. In memory, the Server maintains a `ThreadContext` for each Thread, gathering information related to each Thread's Assistant, etc. When a user sends a new Message, the Server calls the get_local_messages function of ThreadContext to obtain messages and then calls the inference function to get the inference results.
```python
class MyThreadContext(ThreadContext):
def get_local_messages(self):
...
```
Since different model inference frameworks have different historical dialogue input formats, `ThreadContext` and `BackendInterface` need to be used in pairs. Besides its own ktransformers, the Server also supports transformers. For integrating other model inference frameworks, refer to the implementations of `TransformersInterface` and `TransformersThreadContext` in [transformers.py](https://github.com/kvcache-ai/ktransformers-dev/blob/main/ktransformers/server/backend/interfaces/transformers.py).

33
doc/en/api/server/tabby.md Normal file
View File

@ -0,0 +1,33 @@
# How to Use Tabby and ktransformers Locally with 236B Large Models for Code Completion?
[Tabby](https://tabby.tabbyml.com/docs/welcome/) is an open-source code assistant that allows users to manually configure the backend framework and model, and use it across multiple IDEs/editors, such as VSCode and IntelliJ. Since Tabby can interface with Ollama on the framework side, and the ktransformers server provides a consistent API with Ollama, we can connect Tabby to the ktransformers server. This setup allows us to experience fast, heterogeneous inference in code completion scenarios.
1. Start ktransformers.
```bash
./ktransformers --port 9112
```
2. Install Tabby: Follow the official tutorial to install Tabby on a Linux server or Windows PC with an NVIDIA GPU [here](https://tabby.tabbyml.com/docs/quick-start/installation/linux/).
3. Configure Tabby: Create `~/.tabby/config.toml` and add the following configuration.
```toml
[model.completion.http]
kind = "ollama/completion"
api_endpoint = "http://127.0.0.1:9112/"
model_name = "DeepSeek-Coder-V2-Instruct"
prompt_template = "<｜fim▁begin｜>{prefix}<｜fim▁hole｜>{suffix}<｜fim▁end｜>" # Prompt Template
```
In this configuration, `kind` specifies that ktransformers uses the standard Ollama API to serve Tabby; `api_endpoint` matches the interface bound when launching ktransformers; `model_name` is set to the model used by ktransformers, here `DeepSeek-Coder-V2-Instruct` is the backend inference model; `prompt_template` is the model's prompt template, which requires a corresponding template for different models to use the Fill In the Middle feature properly.
Here we demonstrate the relevant configuration for Tabby using the Ollama API to provide the Completion feature. For configuration information about other functions available in Tabby, refer to [here](https://tabby.tabbyml.com/docs/administration/model/).
4. Start the Tabby service: `./tabby serve`.
<img src="run-tabby.png" alt="image-20240709112329577" style="zoom:50%;" />
After launching, you should see access to the `/api/tags` interface in the ktransformers command line (in version v0.13.0 of Tabby, this changes to access to the `/api/show/` interface).
<img src="visit-api-tags.png" alt="image-20240709111648215" style="zoom:67%;" />
6. Register a Tabby account, obtain a Token: After starting the Tabby service, open the corresponding link in a browser (as shown above at 0.0.0.0:8080), and follow the [tutorial](https://tabby.tabbyml.com/docs/quick-start/register-account/) to create a user and get a Token.
7. Start VSCode, install the Tabby extension plugin, and use the Token obtained in the previous step to connect to the Tabby Server, following [here](https://tabby.tabbyml.com/docs/extensions/installation/vscode/).
8. Open any code file and experience the fast heterogeneous inference of ktransformers.

BIN
doc/en/api/server/visit-api-tags.png Normal file
View File

Binary file not shown.

32
doc/en/api/server/website.md Normal file
View File

@ -0,0 +1,32 @@
# Start with website
This document provides the necessary steps to set up and run the web service for this project.
## 1. Starting the Web Service
### 1.1. Compiling the Web Code
Before you can compile the web code, make sure you have installed [Node.js](https://nodejs.org) version 18.3 or higher
Once npm is installed, navigate to the `ktransformers/website` directory:
```bash
cd ktransformers/website
```
Next, install the Vue CLI with the following command:
```bash
npm install @vue/cli
```
Now you can build the project:
```bash
npm run build
```
Finally you can build ktransformers with website:
```
cd ../../
pip install .
```

161
doc/en/deepseek-v2-injection.md Normal file
View File

@ -0,0 +1,166 @@
# Tutorial: Heterogeneous and Local DeepSeek-V2 Inference
DeepSeek-(Code)-V2 is a series of strong mixture-of-experts (MoE) models, featuring a total of 236 billion parameters, with 21 billion parameters activated per token. This model has demonstrated remarkable reasoning capabilities across various benchmarks, positioning it as one of the SOTA open models and nearly comparable in performance to GPT-4.
<p align="center">
<picture>
<img alt="DeepSeek-Coder-V2 Score" src="../assets/BigCodeBench.png" width=80%>
</picture>
</p>
Moreover, unlike previous models that employed traditional attention mechanisms like Grouped-Query Attention (GQA), DeepSeek-V2 incorporates a novel Multi-head Latent Attention (MLA). This innovation significantly reduces the size of the KV cache required during inference, enhancing efficiency.
However, despite its efficiency, the practicality of running such a large model on personal computing setups seems impractical. Official documentation for DeepSeek-V2 indicates that eight 80GB GPUs are necessary for standard inference operations, and even the scaled-down Q4_k_m version requires at least two 80GB GPUs. These requirements are beyond the reach of most individual researchers and small teams.
Nonetheless, by employing several cutting-edge optimization techniques, we have successfully operated this colossal model on a desktop computer with only 21GB of VRAM and 136GB of DRAM. In this document, we outline the specific optimizations utilized and provide a detailed tutorial on how to implement these strategies using KTransformers.
## Applied Optimizations
### Optimized MLA Operator
The following figure provides a brief overview of DeepSeek-V2 architecture. At the heart of its attention layer, DeepSeek-V2 introduces a novel MLA operator that represents the heads of key-value pairs using a common, joint compressed representation, which holds significant potential for efficiency improvements. However, the official open-source implementation of the MLA operator explicitly decompresses this compressed representation and caches the decompressed key-value pairs. This process not only enlarges the KV cache size but also diminishes inference performance.
<p align="center">
<picture>
<img alt="DeepSeek on KTransformers" src="../assets/DeepSeek-on-KTransformers.PNG" width=80%>
</picture>
</p>
To truly capitalize on the benefits of MLA, we have implemented an optimized version for inference. According to its original paper, we absorb the decompression matrices directly into the q_proj and out_proj weights. Consequently, the compressed representation does not need to be decompressed to compute the attention. This adjustment significantly reduces the KV cache size and increases the arithmetic intensity of this operator, which greatly optimizes the utilization of GPU computational power.
### Advanced Quantization Kernels
The original DeepSeek-V2 model stores its parameters in BF16 format, consuming approximately 470GB of raw storage. This exceeds the RAM capacity available on mainstream desktop computers. To address this, we leverage the well-established GGUF community's quantized weights to simplify the process for users.
However, quantized data types are not typically supported by highly-optimized BLAS packages. As a result, the original HuggingFace Transformers' Torch implementation must dequantize these tensors to supported data types before processing, which introduces unnecessary computational overhead and increases memory traffic. To overcome this, we have incorporated advanced kernels that operate directly on quantized data types, thereby optimizing inference performance.
In the current version of KTransformers, we utilize Marlin for GPU kernels and llamafile for CPU kernels. These kerenls are specially designed to benefit from modern GPU architecture and modern CPU instruction extensions such as AVX512-BF16 (AMD Zen4 or newer) and AVX-VNNI (Intel Alder Lake or newer), that are tailored for quantized data types and machine learning workloads. We also use expert parallelism and other optimization for MOE inferencem on CPU based on llamafile, and call them as CPUInfer. As demonstrated in Figure 2(cite from Marlin), Marlin can achieve near ideal 3.87x speedup compare to corresponding Torch counterparts. As demonstrated in the following figure, our micro benchmarks show that inference using CPUInfer performs several times faster than Torch in low bits representation. Note that in practical inference such as using transformers, the Torch baseline use BF16 or FP16 as linear weights, and will occupy more memory resources, or it will be more slower due to dequantization when using quanted weights.
<p align="center">
<picture>
<img alt="CPUInfer Performance" src="../assets/cpuinfer.png" width=80%>
</picture>
</p>
### Arithmetic Intensity Guided Offloading
Storing all 236 billion parameters of a model in GPU VRAM is clearly impractical for local users. Therefore, we strategically store only the most computationally intensive parameters on the GPU. For instance, after our optimizations, the MLA operator, which contains 128 heads with a shared compressed key-value representation, shows an arithmetic intensity of 512. This makes it the most intensive operator, particularly during smaller inference batch sizes. Hence, it is allocated to the GPU to leverage the power of tensor cores.
On the other hand, as shown in Figure 1, each transformer block in DeepSeek-V2 includes 160 mixture-of-experts (MoE) experts, comprising 96% of the total parameters. However, the MoE router activates only 6 out of these 160 experts for each token, which means that only 3.75% of the MoE parameters are utilized during the decoding phase. With a batch size of one, the arithmetic intensity of the MoE operation is roughly 0.075. This operation, primarily involving a batched General Matrix-Vector Multiplication (GEMV), can thus be efficiently handled by the CPU.
Following this principle of arranging all operators by their arithmetic intensity and placing the most intensive ones in the GPU as much as possible, we prioritize positioning the MoE parameters and word embeddings computations on the CPU side to utilize its larger memory capacity. Meanwhile, the remaining parameters, including shared experts, projections in the attention module, and MLA, are stored in the GPU VRAM. As these parameters are accessed by every token, their placement on the GPU maximizes the benefits of high memory bandwidth. This configuration leads to approximately 20.7 GB of VRAM usage and 136GB DRAM memory requests if the Q4_K_M version is used, which is feasible even on a local desktop. Additionally, the placement can be adjusted according to the actual configuration, adhering to the same principle.
Moreover, as an extensible framework, KTransformers is set to support more advanced operators in future releases, continually enhancing its capability to handle diverse workloads efficiently.
## YAML Template
To implement the above optimizations in KTransformers, users need to write a YAML file containing the optimized rules.
KTransformers will iterate through all sub-modules of the model, match rules specified in the YAML rule file, and replace them with advanced modules as specified.
<p align="center">
<picture>
<img alt="Inject-Struction" src="../assets/InjectStruction.png" width=80%>
</picture>
</p>
Specifically, the following rules are used:
- Replace the Attention module with our [optimized MLA Operator](#mla).
- Replace routed experts with [CPUInfer kernels](#experts) that use Llamafile.
- Replace all Linear modules not belonging to attention with [Marlin](#linear) kernels.
<h3 id="mla">MLA</h3>
For attention module injection, we only need to match the module name used in Transformers using a regular expression and replace it with our pre-implemented module.
The YAML rule is listed below.
```yaml
- match:
name: "^model\\.layers\\..*\\.self_attn$" # regular expression
replace:
class: ktransformers.operators.attention.DeepseekV2AttentionInjected # optimized MLA implementation
```
As we can see, each rule in the YAML file has two parts: `match` and `replace`.
The match part specifies which module should be replaced, and the replace part specifies the module to be injected into the model along with the initialization keywords.
<h3 id="experts">Routed Experts </h3>
For routed experts, the module we inject is a wrapper of CPUInfer, KTransformersMLPExpert. There are several implementations within a wrapper, and we need to specify keywords to tell the wrapper which implementation we want to use and how we intend to use it.
In KTransformers, some models exhibit different behaviors during prefilling and generation for better performance. KTransformersMLPExpert is one of them. All these special modules have a `device` keyword describing which device the module should be initialized on. Other keywords specify the behaviors during prefilling and generation and may be differ when using different injection modules. Here, we specify which implementation on which device we want to use during prefilling and generation, and which device the output should be on.
Note that we only use these parameters when layer-wise prefilling is enabled; otherwise, prefilling is conducted with the same configuration as generation.
In the original implementation of Transformers, MoE is implemented using `nn.ModuleList`. We don't want KTransformers to iterate through all the sub-modules in the list, so we set `recursive: False` in this rule to prevent recursive injection into submodules of the current module. Here is the YAML rule:
```yaml
- match:
name: "^model\\.layers\\..*\\.mlp\\.experts$"
replace:
class: ktransformers.operators.experts.KTransformersMLPExpert # custom MoE Kernel with expert parallelism
device: "cpu" # device to load this module on initialization
kwargs:
prefill_device: "cuda"
prefill_mlp_type: "MLPExpertsTorch"
generate_device: "cpu"
generate_mlp_type: "MLPCPUExperts"
out_device: "cuda"
recursive: False # don't recursively inject submodules of this module
```
If we inject the expert list as a custom module, we can't use the interface in `nn.ModuleList` as default. We need to change the forward function in the FFN module. The simplest way is implementing a new module using custom forward function and inject it. We have implemented the new module, and the injection can be done by simply adding an injection rule. We can use the `class` instead of `name` to match a module that will be replaced. Here is the YAML rule:
```yaml
- match:
class: ktransformers.models.modeling_deepseek.DeepseekV2MoE
replace:
class: ktransformers.operators.experts.DeepseekV2MoEInjected # MLP module with custom forward function
```
<h3 id="linear">Other Linear Modules</h3>
For the remained linear modules, we want to use our quantization kernels. However, we don't want to inject linear in the MLA operator because we currently don't know the effect of using quantization in MLA.
So, we can change our regular expression and add a class check in the match part of the rule. Only modules matching both name and class simultaneously will be injected.
We also need to transfer some keywords similar to the injection of experts. Here is the YAML rule:
```yaml
- match:
name: "^model\\.layers\\.(?!.*self_attn).*$" # 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
kwargs:
generate_device: "cuda"
prefill_device: "cuda"
generate_op: "QuantizedLinearMarlin"
prefill_op: "QuantizedLinearTorch"
```
<h3 id="Pre-compute Buffers">Pre-compute Buffers </h3>
The original model is initialized on the meta device. The rotary embedding module pre-computes some buffers when initializing, which has no effect and doesn't compute anything when using the meta device. Therefore, we need to compute the buffers when loading the model. For convenience, we inject the rotary embedding module with our custom module, which performs pre-computations when loading. Here is the YAML rule:
```yaml
- match:
class: ktransformers.models.modeling_deepseek.DeepseekV2YarnRotaryEmbedding
replace:
class: ktransformers.operators.RoPE.YarnRotaryEmbedding
```
## Wrap Your Custom Module
We have implemented some modules, but you may need to inject your custom module using KTransformers.
The only thing you need to do is wrap your custom module and write YAML files. We provide a base operator specifying interfaces an injection module should have. You only need to inherit from that module and change the `__init__`, `forward`, or `load` function as needed.