| Name | liger-kernel JSON |
| Version |
0.5.2
JSON |
| download |
| home_page | None |
| Summary | Efficient Triton kernels for LLM Training |
| upload_time | 2024-12-11 05:58:38 |
| maintainer | None |
| docs_url | None |
| author | None |
| requires_python | None |
| license | BSD 2-CLAUSE LICENSE Copyright 2024 LinkedIn Corporation All Rights Reserved. Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: 1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. 2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
| keywords |
|
| VCS |
 |
| bugtrack_url |
|
| requirements |
No requirements were recorded.
|
| Travis-CI |
No Travis.
|
| coveralls test coverage |
No coveralls.
|
<a name="readme-top"></a>
# Liger Kernel: Efficient Triton Kernels for LLM Training
<table style="width: 100%; text-align: center; border-collapse: collapse;">
<tr>
<th style="padding: 10px;" colspan="2">Stable</th>
<th style="padding: 10px;" colspan="2">Nightly</th>
<th style="padding: 10px;">Discord</th>
<th style="padding: 10px;">Build</th>
</tr>
<tr>
<td style="padding: 10px;">
<a href="https://pepy.tech/project/liger-kernel">
<img src="https://static.pepy.tech/badge/liger-kernel" alt="Downloads (Stable)">
</a>
</td>
<td style="padding: 10px;">
<a href="https://pypi.org/project/liger-kernel">
<img alt="PyPI - Version" src="https://img.shields.io/pypi/v/liger-kernel?color=green">
</a>
</td>
<td style="padding: 10px;">
<a href="https://pepy.tech/project/liger-kernel-nightly">
<img src="https://static.pepy.tech/badge/liger-kernel-nightly" alt="Downloads (Nightly)">
</a>
</td>
<td style="padding: 10px;">
<a href="https://pypi.org/project/liger-kernel-nightly">
<img alt="PyPI - Version" src="https://img.shields.io/pypi/v/liger-kernel-nightly?color=green">
</a>
</td>
<td style="padding: 10px;">
<a href="https://discord.gg/gpumode">
<img src="https://dcbadge.vercel.app/api/server/gpumode?style=flat" alt="Join Our Discord">
</a>
</td>
<td style="padding: 10px;">
<div style="display: block;">
<a href="https://github.com/linkedin/Liger-Kernel/actions/workflows/nvi-ci.yml">
<img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/nvi-ci.yml/badge.svg?event=schedule" alt="Build">
</a>
</div>
<div style="display: block;">
<a href="https://github.com/linkedin/Liger-Kernel/actions/workflows/amd-ci.yml">
<img src="https://github.com/linkedin/Liger-Kernel/actions/workflows/amd-ci.yml/badge.svg?event=schedule" alt="Build">
</a>
</div>
</td>
</tr>
</table>
<img src="https://raw.githubusercontent.com/linkedin/Liger-Kernel/main/docs/images/logo-banner.png">
[Installation](#installation) | [Getting Started](#getting-started) | [Examples](#examples) | [High-level APIs](#high-level-apis) | [Low-level APIs](#low-level-apis) | [Cite our work](#cite-this-work)
<details>
<summary>Latest News 🔥</summary>
- [2024/12/15] We release LinkedIn Engineering Blog - [Liger-Kernel: Empowering an open source ecosystem of Triton Kernels for Efficient LLM Training](https://www.linkedin.com/blog/engineering/open-source/liger-kernel-open-source-ecosystem-for-efficient-llm-training)
- [2024/11/6] We release [v0.4.0](https://github.com/linkedin/Liger-Kernel/releases/tag/v0.4.0): Full AMD support, Tech Report, Modal CI, Llama-3.2-Vision!
- [2024/10/21] We have released the tech report of Liger Kernel on Arxiv: https://arxiv.org/pdf/2410.10989
- [2024/9/6] We release v0.2.1 ([X post](https://x.com/liger_kernel/status/1832168197002510649)). 2500+ Stars, 10+ New Contributors, 50+ PRs, 50k Downloads in two weeks!
- [2024/8/31] CUDA MODE talk, [Liger-Kernel: Real-world Triton kernel for LLM Training](https://youtu.be/gWble4FreV4?si=dxPeIchhkJ36Mbns), [Slides](https://github.com/cuda-mode/lectures?tab=readme-ov-file#lecture-28-liger-kernel)
- [2024/8/23] Official release: check out our [X post](https://x.com/hsu_byron/status/1827072737673982056)
</details>
**Liger Kernel** is a collection of Triton kernels designed specifically for LLM training. It can effectively increase multi-GPU **training throughput by 20%** and reduces **memory usage by 60%**. We have implemented **Hugging Face Compatible** `RMSNorm`, `RoPE`, `SwiGLU`, `CrossEntropy`, `FusedLinearCrossEntropy`, and more to come. The kernel works out of the box with [Flash Attention](https://github.com/Dao-AILab/flash-attention), [PyTorch FSDP](https://pytorch.org/tutorials/intermediate/FSDP_tutorial.html), and [Microsoft DeepSpeed](https://github.com/microsoft/DeepSpeed). We welcome contributions from the community to gather the best kernels for LLM training.
We've also added optimized Post-Training kernels that deliver **up to 80% memory savings** for alignment and distillation tasks. We support losses like DPO, CPO, ORPO, SimPO, JSD, and many more.
## Supercharge Your Model with Liger Kernel
With one line of code, Liger Kernel can increase throughput by more than 20% and reduce memory usage by 60%, thereby enabling longer context lengths, larger batch sizes, and massive vocabularies.
| Speed Up | Memory Reduction |
|--------------------------|-------------------------|
|  |  |
> **Note:**
> - Benchmark conditions: LLaMA 3-8B, Batch Size = 8, Data Type = `bf16`, Optimizer = AdamW, Gradient Checkpointing = True, Distributed Strategy = FSDP1 on 8 A100s.
> - Hugging Face models start to OOM at a 4K context length, whereas Hugging Face + Liger Kernel scales up to 16K.
## Examples
| **Use Case** | **Description** |
|------------------------------------------------|---------------------------------------------------------------------------------------------------|
| [**Hugging Face Trainer**](https://github.com/linkedin/Liger-Kernel/tree/main/examples/huggingface) | Train LLaMA 3-8B ~20% faster with over 40% memory reduction on Alpaca dataset using 4 A100s with FSDP |
| [**Lightning Trainer**](https://github.com/linkedin/Liger-Kernel/tree/main/examples/lightning) | Increase 15% throughput and reduce memory usage by 40% with LLaMA3-8B on MMLU dataset using 8 A100s with DeepSpeed ZeRO3 |
| [**Medusa Multi-head LLM (Retraining Phase)**](https://github.com/linkedin/Liger-Kernel/tree/main/examples/medusa) | Reduce memory usage by 80% with 5 LM heads and improve throughput by 40% using 8 A100s with FSDP |
| [**Vision-Language Model SFT**](https://github.com/linkedin/Liger-Kernel/tree/main/examples/huggingface/run_qwen2_vl.sh) | Finetune Qwen2-VL on image-text data using 4 A100s with FSDP |
| [**Liger ORPO Trainer**](https://github.com/linkedin/Liger-Kernel/blob/main/examples/alignment/run_orpo.py) | Align Llama 3.2 using Liger ORPO Trainer with FSDP with 50% memory reduction |
## Key Features
- **Ease of use:** Simply patch your Hugging Face model with one line of code, or compose your own model using our Liger Kernel modules.
- **Time and memory efficient:** In the same spirit as Flash-Attn, but for layers like **RMSNorm**, **RoPE**, **SwiGLU**, and **CrossEntropy**! Increases multi-GPU training throughput by 20% and reduces memory usage by 60% with **kernel fusion**, **in-place replacement**, and **chunking** techniques.
- **Exact:** Computation is exact—no approximations! Both forward and backward passes are implemented with rigorous unit tests and undergo convergence testing against training runs without Liger Kernel to ensure accuracy.
- **Lightweight:** Liger Kernel has minimal dependencies, requiring only Torch and Triton—no extra libraries needed! Say goodbye to dependency headaches!
- **Multi-GPU supported:** Compatible with multi-GPU setups (PyTorch FSDP, DeepSpeed, DDP, etc.).
- **Trainer Framework Integration**: [Axolotl](https://github.com/axolotl-ai-cloud/axolotl), [LLaMa-Factory](https://github.com/hiyouga/LLaMA-Factory), [SFTTrainer](https://github.com/huggingface/trl/releases/tag/v0.10.1), [Hugging Face Trainer](https://github.com/huggingface/transformers/pull/32860), [SWIFT](https://github.com/modelscope/ms-swift)
## Installation
### Dependencies
#### CUDA
- `torch >= 2.1.2`
- `triton >= 2.3.0`
#### ROCm
- `torch >= 2.5.0` Install according to the instruction in Pytorch official webpage.
- `triton >= 3.0.0` Install from pypi. (e.g. `pip install triton==3.0.0`)
### Optional Dependencies
- `transformers >= 4.x`: Required if you plan to use the transformers models patching APIs. The specific model you are working will dictate the minimum version of transformers.
> **Note:**
> Our kernels inherit the full spectrum of hardware compatibility offered by [Triton](https://github.com/triton-lang/triton).
To install the stable version:
```bash
$ pip install liger-kernel
```
To install the nightly version:
```bash
$ pip install liger-kernel-nightly
```
To install from source:
```bash
git clone https://github.com/linkedin/Liger-Kernel.git
cd Liger-Kernel
pip install -e .
# or if installing on amd platform
pip install -e .[amd] --extra-index-url https://download.pytorch.org/whl/nightly/rocm6.2 # rocm6.2
# or if using transformers
pip install -e .[transformers]
```
## Getting Started
There are a couple of ways to apply Liger kernels, depending on the level of customization required.
### 1. Use AutoLigerKernelForCausalLM
Using the `AutoLigerKernelForCausalLM` is the simplest approach, as you don't have to import a model-specific patching API. If the model type is supported, the modeling code will be automatically patched using the default settings.
```python
from liger_kernel.transformers import AutoLigerKernelForCausalLM
# This AutoModel wrapper class automatically monkey-patches the
# model with the optimized Liger kernels if the model is supported.
model = AutoLigerKernelForCausalLM.from_pretrained("path/to/some/model")
```
### 2. Apply Model-Specific Patching APIs
Using the [patching APIs](#patching), you can swap Hugging Face models with optimized Liger Kernels.
```python
import transformers
from liger_kernel.transformers import apply_liger_kernel_to_llama
# 1a. Adding this line automatically monkey-patches the model with the optimized Liger kernels
apply_liger_kernel_to_llama()
# 1b. You could alternatively specify exactly which kernels are applied
apply_liger_kernel_to_llama(
rope=True,
swiglu=True,
cross_entropy=True,
fused_linear_cross_entropy=False,
rms_norm=False
)
# 2. Instantiate patched model
model = transformers.AutoModelForCausalLM("path/to/llama/model")
```
### 3. Compose Your Own Model
You can take individual [kernels](#kernels) to compose your models.
```python
from liger_kernel.transformers import LigerFusedLinearCrossEntropyLoss
import torch.nn as nn
import torch
model = nn.Linear(128, 256).cuda()
# fuses linear + cross entropy layers together and performs chunk-by-chunk computation to reduce memory
loss_fn = LigerFusedLinearCrossEntropyLoss()
input = torch.randn(4, 128, requires_grad=True, device="cuda")
target = torch.randint(256, (4, ), device="cuda")
loss = loss_fn(model.weight, input, target)
loss.backward()
```
## High-level APIs
### AutoModel
| **AutoModel Variant** | **API** |
|-----------|---------|
| AutoModelForCausalLM | `liger_kernel.transformers.AutoLigerKernelForCausalLM` |
### Patching
| **Model** | **API** | **Supported Operations** |
|-------------|--------------------------------------------------------------|-------------------------------------------------------------------------|
| LLaMA 2 & 3 | `liger_kernel.transformers.apply_liger_kernel_to_llama` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
| LLaMA 3.2-Vision | `liger_kernel.transformers.apply_liger_kernel_to_mllama` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
| Mistral | `liger_kernel.transformers.apply_liger_kernel_to_mistral` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
| Mixtral | `liger_kernel.transformers.apply_liger_kernel_to_mixtral` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
| Gemma1 | `liger_kernel.transformers.apply_liger_kernel_to_gemma` | RoPE, RMSNorm, GeGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
| Gemma2 | `liger_kernel.transformers.apply_liger_kernel_to_gemma2` | RoPE, RMSNorm, GeGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
| Qwen2, Qwen2.5, & QwQ | `liger_kernel.transformers.apply_liger_kernel_to_qwen2` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
| Qwen2-VL | `liger_kernel.transformers.apply_liger_kernel_to_qwen2_vl` | RMSNorm, LayerNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
| Phi3 & Phi3.5 | `liger_kernel.transformers.apply_liger_kernel_to_phi3` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |
## Low-level APIs
- `Fused Linear` kernels combine linear layers with losses, reducing memory usage by up to 80% - ideal for HBM-constrained workloads.
- Other kernels use fusion and in-place techniques for memory and performance optimization.
### Model Kernels
| **Kernel** | **API** |
|---------------------------------|-------------------------------------------------------------|
| RMSNorm | `liger_kernel.transformers.LigerRMSNorm` |
| LayerNorm | `liger_kernel.transformers.LigerLayerNorm` |
| RoPE | `liger_kernel.transformers.liger_rotary_pos_emb` |
| SwiGLU | `liger_kernel.transformers.LigerSwiGLUMLP` |
| GeGLU | `liger_kernel.transformers.LigerGEGLUMLP` |
| CrossEntropy | `liger_kernel.transformers.LigerCrossEntropyLoss` |
| Fused Linear CrossEntropy | `liger_kernel.transformers.LigerFusedLinearCrossEntropyLoss`|
### Alignment Kernels
| **Kernel** | **API** |
|---------------------------------|-------------------------------------------------------------|
| Fused Linear CPO Loss | `liger_kernel.chunked_loss.LigerFusedLinearCPOLoss` |
| Fused Linear DPO Loss | `liger_kernel.chunked_loss.LigerFusedLinearDPOLoss` |
| Fused Linear ORPO Loss | `liger_kernel.chunked_loss.LigerFusedLinearORPOLoss` |
| Fused Linear SimPO Loss | `liger_kernel.chunked_loss.LigerFusedLinearSimPOLoss` |
### Distillation Kernels
| **Kernel** | **API** |
|---------------------------------|-------------------------------------------------------------|
| KLDivergence | `liger_kernel.transformers.LigerKLDIVLoss` |
| JSD | `liger_kernel.transformers.LigerJSD` |
| Fused Linear JSD | `liger_kernel.transformers.LigerFusedLinearJSD` |
### Experimental Kernels
| **Kernel** | **API** |
|---------------------------------|-------------------------------------------------------------|
| Embedding | `liger_kernel.transformers.experimental.LigerEmbedding` |
| Matmul int2xint8 | `liger_kernel.transformers.experimental.matmul` |
## Contributing, Acknowledgements, and License
- [Contributing Guidelines](https://github.com/linkedin/Liger-Kernel/blob/main/docs/CONTRIBUTING.md)
- [Acknowledgements](https://github.com/linkedin/Liger-Kernel/blob/main/docs/Acknowledgement.md)
- [License Information](https://github.com/linkedin/Liger-Kernel/blob/main/docs/License.md)
## Sponsorship and Collaboration
- [AMD](https://www.amd.com/en.html): Providing AMD GPUs for our AMD CI.
- [Intel](https://www.intel.com/): Providing Intel GPUs for our Intel CI.
- [Modal](https://modal.com/): Free 3000 credits from GPU MODE IRL for our NVIDIA CI.
- [EmbeddedLLM](https://embeddedllm.com/): Making Liger Kernel run fast and stable on AMD.
- [HuggingFace](https://huggingface.co/): Integrating Liger Kernel into Hugging Face Transformers and TRL.
- [Lightning AI](https://lightning.ai/): Integrating Liger Kernel into Lightning Thunder.
- [Axolotl](https://axolotl.ai/): Integrating Liger Kernel into Axolotl.
- [Llama-Factory](https://github.com/hiyouga/LLaMA-Factory): Integrating Liger Kernel into Llama-Factory.
## Contact
- For issues, create a Github ticket in this repository
- For open discussion, join [our discord channel](https://discord.gg/gpumode)
- For formal collaboration, send an email to byhsu@linkedin.com
## Cite this work
Biblatex entry:
```bib
@article{hsu2024ligerkernelefficienttriton,
title={Liger Kernel: Efficient Triton Kernels for LLM Training},
author={Pin-Lun Hsu and Yun Dai and Vignesh Kothapalli and Qingquan Song and Shao Tang and Siyu Zhu and Steven Shimizu and Shivam Sahni and Haowen Ning and Yanning Chen},
year={2024},
eprint={2410.10989},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2410.10989},
journal={arXiv preprint arXiv:2410.10989},
}
```
## Star History
[](https://star-history.com/#linkedin/Liger-Kernel&Date)
<p align="right" style="font-size: 14px; color: #555; margin-top: 20px;">
<a href="#readme-top" style="text-decoration: none; color: #007bff; font-weight: bold;">
↑ Back to Top ↑
</a>
</p>
Raw data
{
"_id": null,
"home_page": null,
"name": "liger-kernel",
"maintainer": null,
"docs_url": null,
"requires_python": null,
"maintainer_email": null,
"keywords": null,
"author": null,
"author_email": null,
"download_url": "https://files.pythonhosted.org/packages/a6/cb/5b6debc3fc29bb6a034446a0f1e90cfdfa246772edb6484c3c7db6e295d9/liger_kernel-0.5.2.tar.gz",
"platform": null,
"description": "<a name=\"readme-top\"></a>\n\n# Liger Kernel: Efficient Triton Kernels for LLM Training\n\n\n<table style=\"width: 100%; text-align: center; border-collapse: collapse;\">\n <tr>\n <th style=\"padding: 10px;\" colspan=\"2\">Stable</th>\n <th style=\"padding: 10px;\" colspan=\"2\">Nightly</th>\n <th style=\"padding: 10px;\">Discord</th>\n <th style=\"padding: 10px;\">Build</th>\n </tr>\n <tr>\n <td style=\"padding: 10px;\">\n <a href=\"https://pepy.tech/project/liger-kernel\">\n <img src=\"https://static.pepy.tech/badge/liger-kernel\" alt=\"Downloads (Stable)\">\n </a>\n </td>\n <td style=\"padding: 10px;\">\n <a href=\"https://pypi.org/project/liger-kernel\">\n <img alt=\"PyPI - Version\" src=\"https://img.shields.io/pypi/v/liger-kernel?color=green\">\n </a>\n </td>\n <td style=\"padding: 10px;\">\n <a href=\"https://pepy.tech/project/liger-kernel-nightly\">\n <img src=\"https://static.pepy.tech/badge/liger-kernel-nightly\" alt=\"Downloads (Nightly)\">\n </a>\n </td>\n <td style=\"padding: 10px;\">\n <a href=\"https://pypi.org/project/liger-kernel-nightly\">\n <img alt=\"PyPI - Version\" src=\"https://img.shields.io/pypi/v/liger-kernel-nightly?color=green\">\n </a>\n </td>\n <td style=\"padding: 10px;\">\n <a href=\"https://discord.gg/gpumode\">\n <img src=\"https://dcbadge.vercel.app/api/server/gpumode?style=flat\" alt=\"Join Our Discord\">\n </a>\n </td>\n <td style=\"padding: 10px;\">\n <div style=\"display: block;\">\n <a href=\"https://github.com/linkedin/Liger-Kernel/actions/workflows/nvi-ci.yml\">\n <img src=\"https://github.com/linkedin/Liger-Kernel/actions/workflows/nvi-ci.yml/badge.svg?event=schedule\" alt=\"Build\">\n </a>\n </div>\n <div style=\"display: block;\">\n <a href=\"https://github.com/linkedin/Liger-Kernel/actions/workflows/amd-ci.yml\">\n <img src=\"https://github.com/linkedin/Liger-Kernel/actions/workflows/amd-ci.yml/badge.svg?event=schedule\" alt=\"Build\">\n </a>\n </div>\n </td>\n </tr>\n</table>\n\n\n\n<img src=\"https://raw.githubusercontent.com/linkedin/Liger-Kernel/main/docs/images/logo-banner.png\">\n\n[Installation](#installation) | [Getting Started](#getting-started) | [Examples](#examples) | [High-level APIs](#high-level-apis) | [Low-level APIs](#low-level-apis) | [Cite our work](#cite-this-work)\n\n<details>\n <summary>Latest News \ud83d\udd25</summary>\n\n - [2024/12/15] We release LinkedIn Engineering Blog - [Liger-Kernel: Empowering an open source ecosystem of Triton Kernels for Efficient LLM Training](https://www.linkedin.com/blog/engineering/open-source/liger-kernel-open-source-ecosystem-for-efficient-llm-training)\n - [2024/11/6] We release [v0.4.0](https://github.com/linkedin/Liger-Kernel/releases/tag/v0.4.0): Full AMD support, Tech Report, Modal CI, Llama-3.2-Vision!\n - [2024/10/21] We have released the tech report of Liger Kernel on Arxiv: https://arxiv.org/pdf/2410.10989\n - [2024/9/6] We release v0.2.1 ([X post](https://x.com/liger_kernel/status/1832168197002510649)). 2500+ Stars, 10+ New Contributors, 50+ PRs, 50k Downloads in two weeks!\n - [2024/8/31] CUDA MODE talk, [Liger-Kernel: Real-world Triton kernel for LLM Training](https://youtu.be/gWble4FreV4?si=dxPeIchhkJ36Mbns), [Slides](https://github.com/cuda-mode/lectures?tab=readme-ov-file#lecture-28-liger-kernel)\n - [2024/8/23] Official release: check out our [X post](https://x.com/hsu_byron/status/1827072737673982056)\n\n</details>\n\n\n**Liger Kernel** is a collection of Triton kernels designed specifically for LLM training. It can effectively increase multi-GPU **training throughput by 20%** and reduces **memory usage by 60%**. We have implemented **Hugging Face Compatible** `RMSNorm`, `RoPE`, `SwiGLU`, `CrossEntropy`, `FusedLinearCrossEntropy`, and more to come. The kernel works out of the box with [Flash Attention](https://github.com/Dao-AILab/flash-attention), [PyTorch FSDP](https://pytorch.org/tutorials/intermediate/FSDP_tutorial.html), and [Microsoft DeepSpeed](https://github.com/microsoft/DeepSpeed). We welcome contributions from the community to gather the best kernels for LLM training.\n\nWe've also added optimized Post-Training kernels that deliver **up to 80% memory savings** for alignment and distillation tasks. We support losses like DPO, CPO, ORPO, SimPO, JSD, and many more. \n\n## Supercharge Your Model with Liger Kernel\n\n\n\nWith one line of code, Liger Kernel can increase throughput by more than 20% and reduce memory usage by 60%, thereby enabling longer context lengths, larger batch sizes, and massive vocabularies.\n\n\n| Speed Up | Memory Reduction |\n|--------------------------|-------------------------|\n|  |  |\n\n> **Note:**\n> - Benchmark conditions: LLaMA 3-8B, Batch Size = 8, Data Type = `bf16`, Optimizer = AdamW, Gradient Checkpointing = True, Distributed Strategy = FSDP1 on 8 A100s.\n> - Hugging Face models start to OOM at a 4K context length, whereas Hugging Face + Liger Kernel scales up to 16K.\n\n## Examples\n\n| **Use Case** | **Description** |\n|------------------------------------------------|---------------------------------------------------------------------------------------------------|\n| [**Hugging Face Trainer**](https://github.com/linkedin/Liger-Kernel/tree/main/examples/huggingface) | Train LLaMA 3-8B ~20% faster with over 40% memory reduction on Alpaca dataset using 4 A100s with FSDP |\n| [**Lightning Trainer**](https://github.com/linkedin/Liger-Kernel/tree/main/examples/lightning) | Increase 15% throughput and reduce memory usage by 40% with LLaMA3-8B on MMLU dataset using 8 A100s with DeepSpeed ZeRO3 |\n| [**Medusa Multi-head LLM (Retraining Phase)**](https://github.com/linkedin/Liger-Kernel/tree/main/examples/medusa) | Reduce memory usage by 80% with 5 LM heads and improve throughput by 40% using 8 A100s with FSDP |\n| [**Vision-Language Model SFT**](https://github.com/linkedin/Liger-Kernel/tree/main/examples/huggingface/run_qwen2_vl.sh) | Finetune Qwen2-VL on image-text data using 4 A100s with FSDP |\n| [**Liger ORPO Trainer**](https://github.com/linkedin/Liger-Kernel/blob/main/examples/alignment/run_orpo.py) | Align Llama 3.2 using Liger ORPO Trainer with FSDP with 50% memory reduction |\n\n## Key Features\n\n- **Ease of use:** Simply patch your Hugging Face model with one line of code, or compose your own model using our Liger Kernel modules.\n- **Time and memory efficient:** In the same spirit as Flash-Attn, but for layers like **RMSNorm**, **RoPE**, **SwiGLU**, and **CrossEntropy**! Increases multi-GPU training throughput by 20% and reduces memory usage by 60% with **kernel fusion**, **in-place replacement**, and **chunking** techniques.\n- **Exact:** Computation is exact\u2014no approximations! Both forward and backward passes are implemented with rigorous unit tests and undergo convergence testing against training runs without Liger Kernel to ensure accuracy.\n- **Lightweight:** Liger Kernel has minimal dependencies, requiring only Torch and Triton\u2014no extra libraries needed! Say goodbye to dependency headaches!\n- **Multi-GPU supported:** Compatible with multi-GPU setups (PyTorch FSDP, DeepSpeed, DDP, etc.).\n- **Trainer Framework Integration**: [Axolotl](https://github.com/axolotl-ai-cloud/axolotl), [LLaMa-Factory](https://github.com/hiyouga/LLaMA-Factory), [SFTTrainer](https://github.com/huggingface/trl/releases/tag/v0.10.1), [Hugging Face Trainer](https://github.com/huggingface/transformers/pull/32860), [SWIFT](https://github.com/modelscope/ms-swift)\n\n## Installation\n\n### Dependencies\n\n#### CUDA\n\n- `torch >= 2.1.2`\n- `triton >= 2.3.0`\n\n#### ROCm\n\n- `torch >= 2.5.0` Install according to the instruction in Pytorch official webpage.\n- `triton >= 3.0.0` Install from pypi. (e.g. `pip install triton==3.0.0`)\n\n### Optional Dependencies\n\n- `transformers >= 4.x`: Required if you plan to use the transformers models patching APIs. The specific model you are working will dictate the minimum version of transformers.\n\n> **Note:**\n> Our kernels inherit the full spectrum of hardware compatibility offered by [Triton](https://github.com/triton-lang/triton).\n\nTo install the stable version:\n\n```bash\n$ pip install liger-kernel\n```\n\nTo install the nightly version:\n\n```bash\n$ pip install liger-kernel-nightly\n```\n\nTo install from source:\n\n```bash\ngit clone https://github.com/linkedin/Liger-Kernel.git\ncd Liger-Kernel\npip install -e .\n# or if installing on amd platform\npip install -e .[amd] --extra-index-url https://download.pytorch.org/whl/nightly/rocm6.2 # rocm6.2\n# or if using transformers\npip install -e .[transformers]\n```\n\n\n## Getting Started\n\nThere are a couple of ways to apply Liger kernels, depending on the level of customization required.\n\n### 1. Use AutoLigerKernelForCausalLM\n\nUsing the `AutoLigerKernelForCausalLM` is the simplest approach, as you don't have to import a model-specific patching API. If the model type is supported, the modeling code will be automatically patched using the default settings.\n\n```python\nfrom liger_kernel.transformers import AutoLigerKernelForCausalLM\n\n# This AutoModel wrapper class automatically monkey-patches the\n# model with the optimized Liger kernels if the model is supported.\nmodel = AutoLigerKernelForCausalLM.from_pretrained(\"path/to/some/model\")\n```\n\n### 2. Apply Model-Specific Patching APIs\n\nUsing the [patching APIs](#patching), you can swap Hugging Face models with optimized Liger Kernels.\n\n```python\nimport transformers\nfrom liger_kernel.transformers import apply_liger_kernel_to_llama\n\n# 1a. Adding this line automatically monkey-patches the model with the optimized Liger kernels\napply_liger_kernel_to_llama()\n\n# 1b. You could alternatively specify exactly which kernels are applied\napply_liger_kernel_to_llama(\n rope=True,\n swiglu=True,\n cross_entropy=True,\n fused_linear_cross_entropy=False,\n rms_norm=False\n)\n\n# 2. Instantiate patched model\nmodel = transformers.AutoModelForCausalLM(\"path/to/llama/model\")\n```\n\n### 3. Compose Your Own Model\n\nYou can take individual [kernels](#kernels) to compose your models.\n\n```python\nfrom liger_kernel.transformers import LigerFusedLinearCrossEntropyLoss\nimport torch.nn as nn\nimport torch\n\nmodel = nn.Linear(128, 256).cuda()\n\n# fuses linear + cross entropy layers together and performs chunk-by-chunk computation to reduce memory\nloss_fn = LigerFusedLinearCrossEntropyLoss()\n\ninput = torch.randn(4, 128, requires_grad=True, device=\"cuda\")\ntarget = torch.randint(256, (4, ), device=\"cuda\")\n\nloss = loss_fn(model.weight, input, target)\nloss.backward()\n```\n\n## High-level APIs\n\n### AutoModel\n\n| **AutoModel Variant** | **API** |\n|-----------|---------|\n| AutoModelForCausalLM | `liger_kernel.transformers.AutoLigerKernelForCausalLM` |\n\n\n### Patching\n\n| **Model** | **API** | **Supported Operations** |\n|-------------|--------------------------------------------------------------|-------------------------------------------------------------------------|\n| LLaMA 2 & 3 | `liger_kernel.transformers.apply_liger_kernel_to_llama` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |\n| LLaMA 3.2-Vision | `liger_kernel.transformers.apply_liger_kernel_to_mllama` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |\n| Mistral | `liger_kernel.transformers.apply_liger_kernel_to_mistral` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |\n| Mixtral | `liger_kernel.transformers.apply_liger_kernel_to_mixtral` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |\n| Gemma1 | `liger_kernel.transformers.apply_liger_kernel_to_gemma` | RoPE, RMSNorm, GeGLU, CrossEntropyLoss, FusedLinearCrossEntropy |\n| Gemma2 | `liger_kernel.transformers.apply_liger_kernel_to_gemma2` | RoPE, RMSNorm, GeGLU, CrossEntropyLoss, FusedLinearCrossEntropy |\n| Qwen2, Qwen2.5, & QwQ | `liger_kernel.transformers.apply_liger_kernel_to_qwen2` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |\n| Qwen2-VL | `liger_kernel.transformers.apply_liger_kernel_to_qwen2_vl` | RMSNorm, LayerNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |\n| Phi3 & Phi3.5 | `liger_kernel.transformers.apply_liger_kernel_to_phi3` | RoPE, RMSNorm, SwiGLU, CrossEntropyLoss, FusedLinearCrossEntropy |\n\n\n## Low-level APIs\n\n- `Fused Linear` kernels combine linear layers with losses, reducing memory usage by up to 80% - ideal for HBM-constrained workloads.\n- Other kernels use fusion and in-place techniques for memory and performance optimization.\n\n### Model Kernels\n\n| **Kernel** | **API** |\n|---------------------------------|-------------------------------------------------------------|\n| RMSNorm | `liger_kernel.transformers.LigerRMSNorm` |\n| LayerNorm | `liger_kernel.transformers.LigerLayerNorm` |\n| RoPE | `liger_kernel.transformers.liger_rotary_pos_emb` |\n| SwiGLU | `liger_kernel.transformers.LigerSwiGLUMLP` |\n| GeGLU | `liger_kernel.transformers.LigerGEGLUMLP` |\n| CrossEntropy | `liger_kernel.transformers.LigerCrossEntropyLoss` |\n| Fused Linear CrossEntropy | `liger_kernel.transformers.LigerFusedLinearCrossEntropyLoss`|\n\n\n### Alignment Kernels\n\n| **Kernel** | **API** |\n|---------------------------------|-------------------------------------------------------------|\n| Fused Linear CPO Loss | `liger_kernel.chunked_loss.LigerFusedLinearCPOLoss` |\n| Fused Linear DPO Loss | `liger_kernel.chunked_loss.LigerFusedLinearDPOLoss` |\n| Fused Linear ORPO Loss | `liger_kernel.chunked_loss.LigerFusedLinearORPOLoss` |\n| Fused Linear SimPO Loss | `liger_kernel.chunked_loss.LigerFusedLinearSimPOLoss` |\n\n### Distillation Kernels\n\n| **Kernel** | **API** |\n|---------------------------------|-------------------------------------------------------------|\n| KLDivergence | `liger_kernel.transformers.LigerKLDIVLoss` |\n| JSD | `liger_kernel.transformers.LigerJSD` |\n| Fused Linear JSD | `liger_kernel.transformers.LigerFusedLinearJSD` |\n\n### Experimental Kernels\n\n| **Kernel** | **API** |\n|---------------------------------|-------------------------------------------------------------|\n| Embedding | `liger_kernel.transformers.experimental.LigerEmbedding` |\n| Matmul int2xint8 | `liger_kernel.transformers.experimental.matmul` |\n\n\n## Contributing, Acknowledgements, and License\n\n- [Contributing Guidelines](https://github.com/linkedin/Liger-Kernel/blob/main/docs/CONTRIBUTING.md)\n- [Acknowledgements](https://github.com/linkedin/Liger-Kernel/blob/main/docs/Acknowledgement.md)\n- [License Information](https://github.com/linkedin/Liger-Kernel/blob/main/docs/License.md)\n\n## Sponsorship and Collaboration\n \n- [AMD](https://www.amd.com/en.html): Providing AMD GPUs for our AMD CI.\n- [Intel](https://www.intel.com/): Providing Intel GPUs for our Intel CI.\n- [Modal](https://modal.com/): Free 3000 credits from GPU MODE IRL for our NVIDIA CI.\n- [EmbeddedLLM](https://embeddedllm.com/): Making Liger Kernel run fast and stable on AMD. \n- [HuggingFace](https://huggingface.co/): Integrating Liger Kernel into Hugging Face Transformers and TRL.\n- [Lightning AI](https://lightning.ai/): Integrating Liger Kernel into Lightning Thunder.\n- [Axolotl](https://axolotl.ai/): Integrating Liger Kernel into Axolotl.\n- [Llama-Factory](https://github.com/hiyouga/LLaMA-Factory): Integrating Liger Kernel into Llama-Factory.\n\n## Contact\n\n- For issues, create a Github ticket in this repository\n- For open discussion, join [our discord channel](https://discord.gg/gpumode)\n- For formal collaboration, send an email to byhsu@linkedin.com\n\n## Cite this work\n\nBiblatex entry:\n```bib\n@article{hsu2024ligerkernelefficienttriton,\n title={Liger Kernel: Efficient Triton Kernels for LLM Training},\n author={Pin-Lun Hsu and Yun Dai and Vignesh Kothapalli and Qingquan Song and Shao Tang and Siyu Zhu and Steven Shimizu and Shivam Sahni and Haowen Ning and Yanning Chen},\n year={2024},\n eprint={2410.10989},\n archivePrefix={arXiv},\n primaryClass={cs.LG},\n url={https://arxiv.org/abs/2410.10989},\n journal={arXiv preprint arXiv:2410.10989},\n}\n```\n\n## Star History\n[](https://star-history.com/#linkedin/Liger-Kernel&Date)\n\n<p align=\"right\" style=\"font-size: 14px; color: #555; margin-top: 20px;\">\n <a href=\"#readme-top\" style=\"text-decoration: none; color: #007bff; font-weight: bold;\">\n \u2191 Back to Top \u2191\n </a>\n</p>\n",
"bugtrack_url": null,
"license": "BSD 2-CLAUSE LICENSE Copyright 2024 LinkedIn Corporation All Rights Reserved. Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met: 1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. 2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS \"AS IS\" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. ",
"summary": "Efficient Triton kernels for LLM Training",
"version": "0.5.2",
"project_urls": {
"Homepage": "https://github.com/linkedin/Liger-Kernel"
},
"split_keywords": [],
"urls": [
{
"comment_text": "",
"digests": {
"blake2b_256": "fc9eb05809867d937dbcba4d8717197f69d05a1d41087e7a5c2fdbd5cae79c33",
"md5": "70433d5cc79caf7dbed70e51d23657dd",
"sha256": "31a7409ac9db20fc77e7caf8ab146f56d37d4a1d22e3cefd08661560a99d0c46"
},
"downloads": -1,
"filename": "liger_kernel-0.5.2-py3-none-any.whl",
"has_sig": false,
"md5_digest": "70433d5cc79caf7dbed70e51d23657dd",
"packagetype": "bdist_wheel",
"python_version": "py3",
"requires_python": null,
"size": 104184,
"upload_time": "2024-12-11T05:58:36",
"upload_time_iso_8601": "2024-12-11T05:58:36.575019Z",
"url": "https://files.pythonhosted.org/packages/fc/9e/b05809867d937dbcba4d8717197f69d05a1d41087e7a5c2fdbd5cae79c33/liger_kernel-0.5.2-py3-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "a6cb5b6debc3fc29bb6a034446a0f1e90cfdfa246772edb6484c3c7db6e295d9",
"md5": "53609f0d3dea67ae6e09f88866033d75",
"sha256": "784b5e74848d6802f52e38766bdc0241bf723040f8bf82b153485dd2c9ba0054"
},
"downloads": -1,
"filename": "liger_kernel-0.5.2.tar.gz",
"has_sig": false,
"md5_digest": "53609f0d3dea67ae6e09f88866033d75",
"packagetype": "sdist",
"python_version": "source",
"requires_python": null,
"size": 70272,
"upload_time": "2024-12-11T05:58:38",
"upload_time_iso_8601": "2024-12-11T05:58:38.769067Z",
"url": "https://files.pythonhosted.org/packages/a6/cb/5b6debc3fc29bb6a034446a0f1e90cfdfa246772edb6484c3c7db6e295d9/liger_kernel-0.5.2.tar.gz",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2024-12-11 05:58:38",
"github": true,
"gitlab": false,
"bitbucket": false,
"codeberg": false,
"github_user": "linkedin",
"github_project": "Liger-Kernel",
"travis_ci": false,
"coveralls": false,
"github_actions": true,
"lcname": "liger-kernel"
}
