Update README.md

README.md

@@ -1,3 +1,82 @@
 ---
 license: apache-2.0
 ---
+
+**English** | [中文](./README_zh.md)
+
+## Code implementation of new GTE embeddings
+
+This model is a BERT-like encoder with the following optimizations implemented:
+
+1. Replacing absolute position embeddings with RoPE [^1].
+2. Substituting the conventional activation functions with Gated Linear Units (GLU) [^2].
+3. Setting attention dropout to 0 to use `xformers` and `flash_attn`.
+4. Using unpadding to eliminate the needless computations for padding tokens [^3]. (this is off by default and should be used in conjunction with `xformers` for optimal acceleration).
+5. Setting `vocab_size` as a multiple of 64.
+
+### Recommendation: Enable Unpadding and Acceleration with `xformers`
+
+This code supports the acceleration of attention computations using `xformers`, which can automatically choose the optimal implementation based on the type of device, such as `flash_attn`. Therefore, we can also achieve significant acceleration on old devices like the V100.
+
+
+Firstly, install `xformers` (with `pytorch` pre-installed):
+```
+if pytorch is installed using conda:
+    conda install xformers -c xformers
+elif pytorch is installed using pip:
+    # cuda 11.8 version
+    pip3 install -U xformers --index-url https://download.pytorch.org/whl/cu118
+    # cuda 12.1 version
+    pip3 install -U xformers --index-url https://download.pytorch.org/whl/cu121
+```
+For more information, refer to [Installing xformers](https://github.com/facebookresearch/xformers?tab=readme-ov-file#installing-xformers).
+
+Then, when loading the model, set `unpad_inputs` and `use_memory_efficient_attention` to `true`, and enable `fp16` mixed precision computation to achieve the fastest acceleration.
+
+```python
+import torch
+from transformers import AutoModel, AutoTokenizer
+
+path = 'Alibaba-NLP/gte-base-en-v1.5'
+device = torch.device('cuda')
+tokenzier = AutoTokenizer.from_pretrained(path)
+model = AutoModel.from_pretrained(
+    path,
+    trust_remote_code=True,
+    unpad_inputs=True,
+    use_memory_efficient_attention=True,
+).to(device)
+
+with torch.autocast(device_type=device.type, dtype=torch.float16):  # or bfloat16
+    with torch.inference_mode():
+        outputs = model(**inputs.to(device))
+
+```
+
+Alternatively, you can directly modify the `unpad_inputs` and `use_memory_efficient_attention` settings to `true` in the model's `config.json`, eliminating the need to set them in the code.
+
+
+---
+
+<details>
+  <summary> Clarification of Relationship with nomic-embed and nomicBERT </summary>
+
+One may question the originality of our work and consider it a mere replication of `nomicBERT`. To clarify, our work is parallel but stems from the same idea as `nomicBERT`.
+
+Applying RoPE and GLU to BERT to support longer texts is a straightforward idea. Our exploration of the transformer++ encoder (i.e., BERT + RoPE + GLU) began in August 2023.
+And by November 2023, we had completed the `gte-base-en-v1.1`. Then, I went on to prepare for the ACL submission of the other project...
+
+The release of `nomic-embed` [^4] brought to our attention the pressure, as well as provided us with more resources, which allowed us to continue with this project.
+Without the outstanding work of `nomicai`, the release of `gte-v1.5` could have been delayed much longer. Thanks!
+
+</details>
+
+---
+
+[^1]: Su, Jianlin, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. "Roformer: Enhanced transformer with rotary position embedding." Neurocomputing 568 (2024): 127063.
+
+[^2]: Shazeer, Noam. "Glu variants improve transformer." arXiv preprint arXiv:2002.05202 (2020).
+
+[^3]: Portes, Jacob, Alexander Trott, Sam Havens, Daniel King, Abhinav Venigalla, Moin Nadeem, Nikhil Sardana, Daya Khudia, and Jonathan Frankle. "MosaicBERT: A Bidirectional Encoder Optimized for Fast Pretraining." Advances in Neural Information Processing Systems 36 (2024).
+
+[^4]: Nussbaum, Zach, John X. Morris, Brandon Duderstadt, and Andriy Mulyar. "Nomic Embed: Training a Reproducible Long Context Text Embedder." arXiv preprint arXiv:2402.01613 (2024).

README_zh.md

@@ -0,0 +1,81 @@
+---
+license: apache-2.0
+---
+
+[English](./README.md) | **中文**
+
+## GTE 新模型代码实现
+
+此模型为 BERT-like 编码器模型，加入了以下优化：
+
+1. 使用 RoPE [^1] 旋转位置编码替换 absolute position embedding。
+2. 使用 GLU (Gated Linear Unit) [^2] 替换普通的激活函数。
+3. 设置 attention dropout 为 0 以方便应用 `xformers` 和 `flash_attn` 等优化。
+4. 使用 Unpadding 技术去除对 padding token 的无用计算 [^3]（默认关闭，需要结合 `flash_attn` 或 `xformers` 使用来获得最高加速）。
+5. 设置 `vocab_size % 64 = 0`。
+
+
+### 推荐：启用 Unpadding 和 xformers 加速
+
+此代码支持使用 `xformers` 加速 attention 计算，可以根据设备类型自动选择优化实现，比如 `flash_attn`。通过 `xformers`，在不能支持 `flash_attn` 的旧设备比如`V100`上也可以获得极大的加速。
+
+首先，安装 `xformers`（需要预先安装`pytorch`）：
+```
+if pytorch 使用 conda 安装 :
+    conda install xformers -c xformers
+
+elif pytorch 使用 pip 安装 :
+    # cuda 11.8 version
+    pip3 install -U xformers --index-url https://download.pytorch.org/whl/cu118
+    # cuda 12.1 version
+    pip3 install -U xformers --index-url https://download.pytorch.org/whl/cu121
+```
+更多信息可参考 [installing-xformers](https://github.com/facebookresearch/xformers?tab=readme-ov-file#installing-xformers)。
+
+然后，加载模型时设置 `unpad_inputs` 和 `use_memory_efficient_attention` 为 `true`，并启用 `fp16` 混合精度计算，即可获得最快加速。
+
+```python
+import torch
+from transformers import AutoModel, AutoTokenizer
+
+path = 'Alibaba-NLP/gte-base-en-v1.5'
+device = torch.device('cuda')
+tokenzier = AutoTokenizer.from_pretrained(path)
+model = AutoModel.from_pretrained(
+    path,
+    trust_remote_code=True,
+    unpad_inputs=True,
+    use_memory_efficient_attention=True,
+).to(device)
+
+with torch.autocast(device_type=device.type, dtype=torch.float16):  # 或bfloat16
+    with torch.inference_mode():
+        outputs = model(**inputs.to(device))
+
+```
+也可以直接修改模型的 `config.json` 中 `unpad_inputs` 和 `use_memory_efficient_attention` 为 `true`，省去代码中的设置。
+
+
+---
+
+<details>
+  <summary> 与 nomic-embed 和 nomicBERT 的关系 </summary>
+
+可能有人会质疑我们的原创性，认为这只是对 `nomicBERT` 的复刻。
+在此澄清，我们是工作与 `nomicBERT` 平行并源自相同的想法。
+
+应用 RoPE 和 GLU 到 BERT 上支持长文本是一个简单直接的想法。我们从2023年8月开始了探索。在2023年11月，完成了 `gte-base-en-v1.1` 的开发，然后我去忙别的课题的ACL投稿了。
+
+`nomic-embed` [^4] 的发布让我们感受到了压力，也获得了更多资源得以加速继续开发这一项目。如果没有 `nomicai` 的杰出工作，`gte-v1.5` 系列可能还要延期很久。感谢！
+
+</details>
+
+---
+
+[^1]: Su, Jianlin, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. "Roformer: Enhanced transformer with rotary position embedding." Neurocomputing 568 (2024): 127063.
+
+[^2]: Shazeer, Noam. "Glu variants improve transformer." arXiv preprint arXiv:2002.05202 (2020).
+
+[^3]: Portes, Jacob, Alexander Trott, Sam Havens, Daniel King, Abhinav Venigalla, Moin Nadeem, Nikhil Sardana, Daya Khudia, and Jonathan Frankle. "MosaicBERT: A Bidirectional Encoder Optimized for Fast Pretraining." Advances in Neural Information Processing Systems 36 (2024).
+
+[^4]: Nussbaum, Zach, John X. Morris, Brandon Duderstadt, and Andriy Mulyar. "Nomic Embed: Training a Reproducible Long Context Text Embedder." arXiv preprint arXiv:2402.01613 (2024).