| <!--Copyright 2024 The HuggingFace Team. All rights reserved. | |
| 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. | |
| ⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be | |
| rendered properly in your Markdown viewer. | |
| --> | |
| # BitNet | |
| [BitNet](https://arxiv.org/abs/2402.17764) replaces traditional linear layers in Multi-Head Attention and feed-forward networks with specialized BitLinear layers. The BitLinear layers quantize the weights using ternary precision (with values of -1, 0, and 1) and quantize the activations to 8-bit precision. | |
| <figure style="text-align: center;"> | |
| <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/1.58llm_extreme_quantization/bitlinear.png" alt="Alt Text" /> | |
| <figcaption>The architecture of BitNet with BitLinear layers.</figcaption> | |
| </figure> | |
| BitNet models can't be quantized on the fly. They need to be quantized during pretraining or fine-tuning because it is a Quantization-Aware Training (QAT) technique. During training, the weights are quantized to ternary values with symmetric per tensor quantization. | |
| 1. Compute the average of the absolute values of the weight matrix and use as a scale. | |
| 2. Divide the weights by the scale, round the values, constrain them between -1 and 1, and rescale them to continue in full precision. | |
| 3. Activations are quantized to a specified bit-width (8-bit) using [absmax](https://arxiv.org/pdf/2208.07339) quantization (symmetric per channel quantization). This involves scaling the activations into a range of [−128,127]. | |
| Refer to this [PR](https://github.com/huggingface/nanotron/pull/180) to pretrain or fine-tune a 1.58-bit model with [Nanotron](https://github.com/huggingface/nanotron). For fine-tuning, convert a model from the Hugging Face to Nanotron format. Find the conversion steps in this [PR](https://github.com/huggingface/nanotron/pull/174). | |
| Load a BitNet quantized model with [`~PreTrainedModel.from_pretrained`]. | |
| ```py | |
| from transformers import AutoModelForCausalLM | |
| path = "/path/to/model" | |
| model = AutoModelForCausalLM.from_pretrained(path, device_map="auto") | |
| ``` | |
| ## Kernels | |
| `@torch.compile` is used to unpack the weights and perform the forward pass. It’s very straightforward to implement and delivers significant speed improvements. Additional optimized kernels will be integrated in future versions. | |
| ## Resources | |
| Read [Fine-tuning LLMs to 1.58bit: extreme quantization made easy](https://huggingface.co/blog/1_58_llm_extreme_quantization) to learn more about how BitNet models are trained and fine-tuned. | |