fix sampling in chat stream

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+assets/wanx_colorful_black.png filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -9,7 +9,7 @@ By clicking to agree or by using or distributing any portion or element of the T
     b. "We"(or "Us") shall mean Alibaba Cloud.
     c. "You" (or "Your") shall mean a natural person or legal entity exercising the rights granted by this Agreement and/or using the Materials for any purpose and in any field of use.
     d. "Third Parties" shall mean individuals or legal entities that are not under common control with Us or You.
-    e. "Tongyi Qianwen" shall mean the large language models (including Qwen model and Qwen-Chat model), and software and algorithms, consisting of trained model weights, parameters (including optimizer states), machine-learning model code, inference-enabling code, training-enabling code, fine-tuning enabling code and other elements of the foregoing distributed by Us.
+    e. "Tongyi Qianwen" shall mean the large language models (including Qwen-7B model and Qwen-7B-Chat model), and software and algorithms, consisting of trained model weights, parameters (including optimizer states), machine-learning model code, inference-enabling code, training-enabling code, fine-tuning enabling code and other elements of the foregoing distributed by Us.
     f. "Materials" shall mean, collectively, Alibaba Cloud's proprietary Tongyi Qianwen and Documentation (and any portion thereof) made available under this Agreement.
     g. "Source" form shall mean the preferred form for making modifications, including but not limited to model source code, documentation source, and configuration files.
     h. "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,

NOTICE

@@ -49,232 +49,4 @@ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
 AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
 LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
 OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-SOFTWARE.
-
-------------- LICENSE FOR stanford_alpaca code  --------------
-
-                                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 2023 Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li
-
-   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.
-   
-------------- LICENSE FOR PanQiWei AutoGPTQ code  --------------
-
-MIT License
-
-Copyright (c) 2023 潘其威(William)
-
-Permission is hereby granted, free of charge, to any person obtaining a copy
-of this software and associated documentation files (the "Software"), to deal
-in the Software without restriction, including without limitation the rights
-to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-copies of the Software, and to permit persons to whom the Software is
-furnished to do so, subject to the following conditions:
-
-The above copyright notice and this permission notice shall be included in all
-copies or substantial portions of the Software.
-
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-SOFTWARE.
+SOFTWARE.

README.md

@@ -6,36 +6,29 @@ tags:
 - qwen
 pipeline_tag: text-generation
 inference: false
-license: other
-license_name: tongyi-qianwen-license-agreement
-license_link: https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT
 ---
 
 # Qwen-7B-Chat
 
 <p align="center">
-    <img src="https://qianwen-res.oss-cn-beijing.aliyuncs.com/logo_qwen.jpg" width="400"/>
+    <img src="https://qianwen-res.oss-cn-beijing.aliyuncs.com/logo.jpg" width="400"/>
 <p>
 <br>
 
 <p align="center">
-        🤗 <a href="https://huggingface.co/Qwen">Hugging Face</a>&nbsp&nbsp | &nbsp&nbsp🤖 <a href="https://modelscope.cn/organization/qwen">ModelScope</a>&nbsp&nbsp | &nbsp&nbsp 📑 <a href="https://arxiv.org/abs/2309.16609">Paper</a> &nbsp&nbsp ｜ &nbsp&nbsp🖥️ <a href="https://modelscope.cn/studios/qwen/Qwen-7B-Chat-Demo/summary">Demo</a>
-<br>
-<a href="https://github.com/QwenLM/Qwen/blob/main/assets/wechat.png">WeChat (微信)</a>&nbsp&nbsp | &nbsp&nbsp<a href="https://discord.gg/z3GAxXZ9Ce">Discord</a>&nbsp&nbsp ｜  &nbsp&nbsp<a href="https://dashscope.aliyun.com">API</a> 
+        Qwen-7B <a href="https://modelscope.cn/models/qwen/Qwen-7B/summary">🤖 </a> | <a href="https://huggingface.co/Qwen/Qwen-7B">🤗</a>&nbsp ｜ Qwen-7B-Chat <a href="https://modelscope.cn/models/qwen/Qwen-7B-Chat/summary">🤖 </a>| <a href="https://huggingface.co/Qwen/Qwen-7B-Chat">🤗</a>&nbsp ｜ &nbsp<a href="https://modelscope.cn/studios/qwen/Qwen-7B-Chat-Demo/summary">Demo</a>&nbsp ｜ &nbsp<a href="https://github.com/QwenLM/Qwen-7B/blob/main/tech_memo.md">Report</a>
 </p>
 <br>
 
-
 ## 介绍（Introduction）
 
-**通义千问-7B（Qwen-7B）**是阿里云研发的通义千问大模型系列的70亿参数规模的模型。Qwen-7B是基于Transformer的大语言模型, 在超大规模的预训练数据上进行训练得到。预训练数据类型多样，覆盖广泛，包括大量网络文本、专业书籍、代码等。同时，在Qwen-7B的基础上，我们使用对齐机制打造了基于大语言模型的AI助手Qwen-7B-Chat。相较于最初开源的Qwen-7B模型，我们现已将预训练模型和Chat模型更新到效果更优的版本。本仓库为Qwen-7B-Chat的仓库。
+**通义千问-7B（Qwen-7B）**是阿里云研发的通义千问大模型系列的70亿参数规模的模型。Qwen-7B是基于Transformer的大语言模型, 在超大规模的预训练数据上进行训练得到。预训练数据类型多样，覆盖广泛，包括大量网络文本、专业书籍、代码等。同时，在Qwen-7B的基础上，我们使用对齐机制打造了基于大语言模型的AI助手Qwen-7B-Chat。本仓库为Qwen-7B-Chat的仓库。
 
-如果您想了解更多关于通义千问-7B开源模型的细节，我们建议您参阅[GitHub代码库](https://github.com/QwenLM/Qwen)。
+如果您想了解更多关于通义千问-7B开源模型的细节，我们建议您参阅[Github代码库](https://github.com/QwenLM/Qwen-7B)。
 
-**Qwen-7B** is the 7B-parameter version of the large language model series, Qwen (abbr. Tongyi Qianwen), proposed by Alibaba Cloud. Qwen-7B is a Transformer-based large language model, which is pretrained on a large volume of data, including web texts, books, codes, etc. Additionally, based on the pretrained Qwen-7B, we release Qwen-7B-Chat, a large-model-based AI assistant, which is trained with alignment techniques. Now we have updated both our pretrained and chat models with better performances. This repository is the one for Qwen-7B-Chat.
+**Qwen-7B** is the 7B-parameter version of the large language model series, Qwen (abbr. Tongyi Qianwen), proposed by Aibaba Cloud. Qwen-7B`is a Transformer-based large language model, which is pretrained on a large volume of data, including web texts, books, codes, etc. Additionally, based on the pretrained Qwen-7B, we release Qwen-7B-Chat, a large-model-based AI assistant, which is trained with alignment techniques. This repository is the one for Qwen-7B-Chat.
 
-For more details about Qwen, please refer to the [GitHub](https://github.com/QwenLM/Qwen) code repository.
-<br>
+For more details about the open-source model of Qwen-7B, please refer to the [Github](https://github.com/QwenLM/Qwen-7B) code repository.
 
 ## 要求（Requirements）
 
@@ -45,7 +38,6 @@ For more details about Qwen, please refer to the [GitHub](https://github.com/Qwe
 * python 3.8 and above
 * pytorch 1.12 and above, 2.0 and above are recommended
 * CUDA 11.4 and above are recommended (this is for GPU users, flash-attention users, etc.)
-<br>
 
 ## 依赖项（Dependency）
 
@@ -54,21 +46,21 @@ For more details about Qwen, please refer to the [GitHub](https://github.com/Qwe
 To run Qwen-7B-Chat, please make sure you meet the above requirements, and then execute the following pip commands to install the dependent libraries.
 
 ```bash
-pip install transformers==4.32.0 accelerate tiktoken einops scipy transformers_stream_generator==0.0.4 peft deepspeed
+pip install transformers==4.31.0 accelerate tiktoken einops
 ```
 
-另外，推荐安装`flash-attention`库（**当前已支持flash attention 2**），以实现更高的效率和更低的显存占用。
+另外，推荐安装`flash-attention`库，以实现更高的效率和更低的显存占用。
 
-In addition, it is recommended to install the `flash-attention` library (**we support flash attention 2 now.**) for higher efficiency and lower memory usage.
+In addition, it is recommended to install the `flash-attention` library for higher efficiency and lower memory usage.
 
 ```bash
-git clone https://github.com/Dao-AILab/flash-attention
+git clone -b v1.0.8 https://github.com/Dao-AILab/flash-attention
 cd flash-attention && pip install .
 # 下方安装可选，安装可能比较缓慢。
-# pip install csrc/layer_norm
-# pip install csrc/rotary
+# Below are optional. Installing them might be slow.
+pip install csrc/layer_norm
+pip install csrc/rotary
 ```
-<br>
 
 ## 快速使用（Quickstart）
 
@@ -92,8 +84,8 @@ tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen-7B-Chat", trust_remote_code
 # use auto mode, automatically select precision based on the device.
 model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-7B-Chat", device_map="auto", trust_remote_code=True).eval()
 
-# Specify hyperparameters for generation. But if you use transformers>=4.32.0, there is no need to do this.
-# model.generation_config = GenerationConfig.from_pretrained("Qwen/Qwen-7B-Chat", trust_remote_code=True) # 可指定不同的生成长度、top_p等相关超参
+# Specify hyperparameters for generation
+model.generation_config = GenerationConfig.from_pretrained("Qwen/Qwen-7B-Chat", trust_remote_code=True) # 可指定不同的生成长度、top_p等相关超参
 
 # 第一轮对话 1st dialogue turn
 response, history = model.chat(tokenizer, "你好", history=None)
@@ -116,123 +108,31 @@ print(response)
 # 《奋斗创业：一个年轻人的成功之路》
 ```
 
-关于更多的使用说明，请参考我们的[GitHub repo](https://github.com/QwenLM/Qwen)获取更多信息。
+关于更多的使用说明，请参考我们的[Github repo](https://github.com/QwenLM/Qwen-7B)获取更多信息。
 
-For more information, please refer to our [GitHub repo](https://github.com/QwenLM/Qwen) for more information.
-<br>
+For more information, please refer to our [Github repo](https://github.com/QwenLM/Qwen-7B) for more information.
 
 ## Tokenizer
 
 > 注：作为术语的“tokenization”在中文中尚无共识的概念对应，本文档采用英文表达以利说明。
 
-基于tiktoken的分词器有别于其他分词器，比如sentencepiece分词器。尤其在微调阶段，需要特别注意特殊token的使用。关于tokenizer的更多信息，以及微调时涉及的相关使用，请参阅[文档](https://github.com/QwenLM/Qwen/blob/main/tokenization_note_zh.md)。
-
-Our tokenizer based on tiktoken is different from other tokenizers, e.g., sentencepiece tokenizer. You need to pay attention to special tokens, especially in finetuning. For more detailed information on the tokenizer and related use in fine-tuning, please refer to the [documentation](https://github.com/QwenLM/Qwen/blob/main/tokenization_note.md).
-<br>
-
-## 量化 (Quantization)
-
-### 用法 (Usage)
-
-**请注意：我们更新量化方案为基于[AutoGPTQ](https://github.com/PanQiWei/AutoGPTQ)的量化，提供Qwen-7B-Chat的Int4量化模型[点击这里](https://huggingface.co/Qwen/Qwen-7B-Chat-Int4)。相比此前方案，该方案在模型评测效果几乎无损，且存储需求更低，推理速度更优。**
-
-**Note: we provide a new solution based on [AutoGPTQ](https://github.com/PanQiWei/AutoGPTQ), and release an Int4 quantized model for Qwen-7B-Chat [Click here](https://huggingface.co/Qwen/Qwen-7B-Chat-Int4), which achieves nearly lossless model effects but improved performance on both memory costs and inference speed, in comparison with the previous solution.**
-
-以下我们提供示例说明如何使用Int4量化模型。在开始使用前，请先保证满足要求（如torch 2.0及以上，transformers版本为4.32.0及以上，等等），并安装所需安装包：
-
-Here we demonstrate how to use our provided quantized models for inference. Before you start, make sure you meet the requirements of auto-gptq (e.g., torch 2.0 and above, transformers 4.32.0 and above, etc.) and install the required packages:
-
-```bash
-pip install auto-gptq optimum
-```
-
-如安装`auto-gptq`遇到问题，我们建议您到官方[repo](https://github.com/PanQiWei/AutoGPTQ)搜索合适的预编译wheel。
-
-随后即可使用和上述一致的用法调用量化模型：
-
-If you meet problems installing `auto-gptq`, we advise you to check out the official [repo](https://github.com/PanQiWei/AutoGPTQ) to find a pre-build wheel.
-
-Then you can load the quantized model easily and run inference as same as usual:
-
-```python
-model = AutoModelForCausalLM.from_pretrained(
-    "Qwen/Qwen-7B-Chat-Int4",
-    device_map="auto",
-    trust_remote_code=True
-).eval()
-response, history = model.chat(tokenizer, "你好", history=None)
-```
-
-
-
-### 效果评测
-
-我们对BF16，Int8和Int4模型在基准评测上做了测试（使用zero-shot设置），发现量化模型效果损失较小，结果如下所示：
-
-We illustrate the zero-shot performance of both BF16, Int8 and Int4 models on the benchmark, and we find that the quantized model does not suffer from significant performance degradation. Results are shown below:
-
-|  Quantization |   MMLU     |  CEval (val) |  GSM8K |  Humaneval |
-| ------------- | :--------: | :----------: | :----: | :--------: |
-| BF16          |    55.8    |     59.7     |  50.3  |    37.2    |
-| Int8          |    55.4    |     59.4     |  48.3  |    34.8    |
-| Int4          |    55.1    |     59.2     |  49.7  |    29.9    |
-
-### 推理速度 (Inference Speed)
-
-我们测算了不同精度模型以及不同FlashAttn库版本下模型生成2048和8192个token的平均推理速度。如图所示：
-
-We measured the average inference speed of generating 2048 and 8192 tokens with different quantization levels and versions of flash-attention, respectively.
-
-|  Quantization | FlashAttn | Speed (2048 tokens) | Speed (8192 tokens) |
-| ------------- | :-------: | :------------------:| :------------------:|
-|      BF16     |   v2      | 40.93               | 36.14               |
-|      Int8     |   v2      | 37.47               | 32.54               |
-|      Int4     |   v2      | 50.09               | 38.61               |
-|      BF16     |   v1      | 40.75               | 35.34               |
-|      Int8     |   v1      | 37.51               | 32.39               |
-|      Int4     |   v1      | 45.98               | 36.47               |
-|      BF16     |  Disabled | 37.55               | 33.56               |
-|      Int8     |  Disabled | 37.84               | 32.65               |
-|      Int4     |  Disabled | 48.12               | 36.70               |
-
-具体而言，我们记录在长度为1的上下文的条件下生成8192个token的性能。评测运行于单张A100-SXM4-80G GPU，使用PyTorch 2.0.1和CUDA 11.8。推理速度是生成8192个token的速度均值。
+基于tiktoken的分词器有别于其他分词器，比如sentencepiece分词器。尤其在微调阶段，需要特别注意特殊token的使用。关于tokenizer的更多信息，以及微调时涉及的相关使用，请参阅[文档](https://github.com/QwenLM/Qwen-7B/blob/main/tokenization_note_zh.md)。
 
-In detail, the setting of profiling is generating 8192 new tokens with 1 context token. The profiling runs on a single A100-SXM4-80G GPU with PyTorch 2.0.1 and CUDA 11.8. The inference speed is averaged over the generated 8192 tokens.
-
-注意：以上Int4/Int8模型生成速度使用autogptq库给出，当前``AutoModelForCausalLM.from_pretrained``载入的模型生成速度会慢大约20%。我们已经将该问题汇报给HuggingFace团队，若有解决方案将即时更新。
-
-Note: The generation speed of the Int4/Int8 models mentioned above is provided by the autogptq library. The current speed of the model loaded using "AutoModelForCausalLM.from_pretrained" will be approximately 20% slower. We have reported this issue to the HuggingFace team and will update it promptly if a solution is available.
-
-### 显存使用 (GPU Memory Usage)
-
-我们还测算了不同模型精度编码2048个token及生成8192个token的峰值显存占用情况。（显存消耗在是否使用FlashAttn的情况下均类似。）结果如下所示：
-
-We also profile the peak GPU memory usage for encoding 2048 tokens as context (and generating single token) and generating 8192 tokens (with single token as context) under different quantization levels, respectively. （The GPU memory usage is similar when using flash-attention or not.）The results are shown below.
-
-| Quantization Level | Peak Usage for Encoding 2048 Tokens | Peak Usage for Generating 8192 Tokens |
-| ------------------ | :---------------------------------: | :-----------------------------------: |
-| BF16               | 16.99GB                             | 22.53GB                               |
-| Int8               | 11.20GB                             | 16.62GB                               |
-| Int4               |  8.21GB                             | 13.63GB                               |
-
-上述性能测算使用[此脚本](https://qianwen-res.oss-cn-beijing.aliyuncs.com/profile.py)完成。
-
-The above speed and memory profiling are conducted using [this script](https://qianwen-res.oss-cn-beijing.aliyuncs.com/profile.py).
-<br>
+Our tokenizer based on tiktoken is different from other tokenizers, e.g., sentencepiece tokenizer. You need to pay attention to special tokens, especially in finetuning. For more detailed information on the tokenizer and related use in fine-tuning, please refer to the [documentation](https://github.com/QwenLM/Qwen-7B/blob/main/tokenization_note.md).
 
 ## 模型细节（Model）
 
-与Qwen-7B预训练模型相同，Qwen-7B-Chat模型规模基本情况如下所示:
+与Qwen-7B预训练模型相同，Qwen-7B-Chat模型规模基本情况如下所示
 
-The details of the model architecture of Qwen-7B-Chat are listed as follows:
+The details of the model architecture of Qwen-7B-Chat are listed as follows
 
-| Hyperparameter  | Value  |
-|:----------------|:------:|
-| n_layers        |   32   |
-| n_heads         |   32   |
-| d_model         |  4096  |
-| vocab size      | 151851 |
-| sequence length |  8192  |
+| Hyperparameter | Value |
+|:------|:------|
+| n_layers | 32 |
+| n_heads | 32 |
+| d_model | 4096 |
+| vocab size | 151851 |
+| sequence length | 2048 |
 
 在位置编码、FFN激活函数和normalization的实现方式上，我们也采用了目前最流行的做法，
 即RoPE相对位置编码、SwiGLU激活函数、RMSNorm（可选安装flash-attention加速）。
@@ -246,7 +146,6 @@ For position encoding, FFN activation function, and normalization calculation me
 For tokenization, compared to the current mainstream open-source models based on Chinese and English vocabularies, Qwen-7B-Chat uses a vocabulary of over 150K tokens.
 It first considers efficient encoding of Chinese, English, and code data, and is also more friendly to multilingual languages, enabling users to directly enhance the capability of some languages without expanding the vocabulary.
 It segments numbers by single digit, and calls the [tiktoken](https://github.com/openai/tiktoken) tokenizer library for efficient tokenization.
-<br>
 
 ## 评测效果（Evaluation）
 
@@ -262,38 +161,32 @@ Note: Due to rounding errors caused by hardware and framework, differences in re
 
 #### C-Eval
 
-在[C-Eval](https://arxiv.org/abs/2305.08322)验证集上，我们评价了Qwen-7B-Chat模型的0-shot & 5-shot准确率
-
-We demonstrate the 0-shot & 5-shot accuracy of Qwen-7B-Chat on C-Eval validation set
-
-|              Model               | Avg. Acc. |
-|:--------------------------------:|:---------:|
-|          LLaMA2-7B-Chat          |   31.9    |
-|         LLaMA2-13B-Chat          |   36.2    |
-|         LLaMA2-70B-Chat          |   44.3    |
-|         ChatGLM2-6B-Chat         |   52.6    |
-|         InternLM-7B-Chat         |   53.6    |
-|        Baichuan2-7B-Chat         |   55.6    |
-|        Baichuan2-13B-Chat        |   56.7    |
-| Qwen-7B-Chat (original) (0-shot) |   54.2    |
-|    **Qwen-7B-Chat (0-shot)**     |   59.7    |
-|    **Qwen-7B-Chat (5-shot)**     |   59.3    |
-|    **Qwen-14B-Chat (0-shot)**    |   69.8    |
-|    **Qwen-14B-Chat (5-shot)**    | **71.7**  |
+在[C-Eval](https://arxiv.org/abs/2305.08322)验证集上，我们评价了Qwen-7B-Chat模型的zero-shot准确率
+
+We demonstrate the zero-shot accuracy of Qwen-7B-Chat on C-Eval validation set
+
+| Model | Avg. Acc. |
+|:--------------|:------:|
+| LLaMA2-7B-Chat | 31.9 |
+| LLaMA2-13B-Chat | 40.6 |
+| Chinese-Alpaca-2-7B | 41.3 |
+| Chinese-Alpaca-Plus-13B | 43.3 |
+| Baichuan-13B-Chat | 50.4 |
+| ChatGLM2-6B-Chat | 50.7 |
+| InternLM-7B-Chat | 53.2 |
+| **Qwen-7B-Chat** | **54.2** |
 
 C-Eval测试集上，Qwen-7B-Chat模型的zero-shot准确率结果如下：
 
 The zero-shot accuracy of Qwen-7B-Chat on C-Eval testing set is provided below:
 
-| Model                   |   Avg.   | STEM | Social Sciences | Humanities | Others |
-| :---------------------- | :------: | :--: | :-------------: | :--------: | :----: |
-| Chinese-Alpaca-Plus-13B |   41.5   | 36.6 |      49.7       |    43.1    |  41.2  |
-| Chinese-Alpaca-2-7B     |   40.3   |  -   |        -        |     -      |   -    |
-| ChatGLM2-6B-Chat        |   50.1   | 46.4 |      60.4       |    50.6    |  46.9  |
-| Baichuan-13B-Chat       |   51.5   | 43.7 |      64.6       |    56.2    |  49.2  |
-| Qwen-7B-Chat (original)        |   54.6   | 47.8 |      67.6       |    59.3    |  50.6  |
-| **Qwen-7B-Chat**   |   58.6   | 53.3 |      72.1       |    62.8    |  52.0  |
-| **Qwen-14B-Chat**       | **69.1** | 65.1 |      80.9       |    71.2    |  63.4  |
+| Model | Avg. | STEM | Social Sciences | Humanities | Others |
+|:--------------|:------:|:------:|:------:|:------:|:------:|
+| Chinese-Alpaca-Plus-13B | 41.5 | 36.6 | 49.7 | 43.1 | 41.2 |
+| Chinese-Alpaca-2-7B | 40.3 | - | - | - | - |
+| ChatGLM2-6B-Chat | 50.1 | 46.4 | 60.4 | 50.6 | 46.9 |
+| Baichuan-13B-Chat | 51.5 | 43.7 | 64.6 | 56.2 | 49.2 |
+| **Qwen-7B-Chat** | **54.6** | 47.8 | 67.6 | 59.3 | 50.6 |
 
 在7B规模模型上，经过人类指令对齐的Qwen-7B-Chat模型，准确率在同类相近规模模型中仍然处于前列。
 
@@ -303,25 +196,19 @@ Compared with other pretrained models with comparable model size, the human-alig
 
 #### MMLU
 
-[MMLU](https://arxiv.org/abs/2009.03300)评测集上，Qwen-7B-Chat模型的 0-shot & 5-shot 准确率如下，效果同样在同类对齐模型中同样表现较优。
+[MMLU](https://arxiv.org/abs/2009.03300)评测集上，Qwen-7B-Chat模型的zero-shot准确率如下，效果同样在同类对齐模型中同样表现较优。
 
-The 0-shot & 5-shot accuracy of Qwen-7B-Chat on MMLU is provided below.
+The zero-shot accuracy of Qwen-7B-Chat on MMLU is provided below.
 The performance of Qwen-7B-Chat still on the top between other human-aligned models with comparable size.
 
-|              Model               | Avg. Acc. |
-|:--------------------------------:|:---------:|
-|         ChatGLM2-6B-Chat         |   46.0    |
-|          LLaMA2-7B-Chat          |   46.2    |
-|         InternLM-7B-Chat         |   51.1    |
-|        Baichuan2-7B-Chat         |   52.9    |
-|         LLaMA2-13B-Chat          |   54.6    |
-|        Baichuan2-13B-Chat        |   57.3    |
-|         LLaMA2-70B-Chat          |   63.8    |
-| Qwen-7B-Chat (original) (0-shot) |   53.9    |
-|    **Qwen-7B-Chat (0-shot)**     |   55.8    |
-|    **Qwen-7B-Chat (5-shot)**     |   57.0    |
-|    **Qwen-14B-Chat (0-shot)**    |   64.6    |
-|    **Qwen-14B-Chat (5-shot)**    | **66.5**  |
+| Model | Avg. Acc. |
+|:--------------|:------:|
+| ChatGLM2-6B-Chat | 45.5 |
+| LLaMA2-7B-Chat | 47.0 |
+| InternLM-7B-Chat | 50.8 |
+| Baichuan-13B-Chat | 52.1 |
+| ChatGLM2-12B-Chat | 52.1 |
+| **Qwen-7B-Chat** | **53.9** |
 
 ### 代码评测（Coding Evaluation）
 
@@ -329,57 +216,47 @@ Qwen-7B-Chat在[HumanEval](https://github.com/openai/human-eval)的zero-shot Pas
 
 The zero-shot Pass@1 of Qwen-7B-Chat on [HumanEval](https://github.com/openai/human-eval) is demonstrated below
 
-|          Model          |  Pass@1  |
-|:-----------------------:|:--------:|
-|    ChatGLM2-6B-Chat     |   11.0   |
-|     LLaMA2-7B-Chat      |   12.2   |
-|    Baichuan2-7B-Chat    |   13.4   |
-|    InternLM-7B-Chat     |   14.6   |
-|   Baichuan2-13B-Chat    |   17.7   |
-|     LLaMA2-13B-Chat     |   18.9   |
-|     LLaMA2-70B-Chat     |   32.3   |
-| Qwen-7B-Chat (original) |   24.4   |
-|    **Qwen-7B-Chat**     |   37.2   |
-|    **Qwen-14B-Chat**    | **43.9** |
-
-### 数学评测（Mathematics Evaluation）
+| Model | Pass@1 |
+|:--------------|:------:|
+| LLaMA2-7B-Chat | 12.2 |
+| InternLM-7B-Chat | 14.0 |
+| Baichuan-13B-Chat | 16.5 |
+| LLaMA2-13B-Chat | 18.9 |
+| **Qwen-7B-Chat** | **24.4** |
+
+### 数学评测
 
 在评测数学能力的[GSM8K](https://github.com/openai/grade-school-math)上，Qwen-7B-Chat的准确率结果如下
 
 The accuracy of Qwen-7B-Chat on GSM8K is shown below
 
-|                Model                 |   Acc.   |
-|:------------------------------------:|:--------:|
-|            LLaMA2-7B-Chat            |   26.3   |
-|           ChatGLM2-6B-Chat           |   28.8   |
-|          Baichuan2-7B-Chat           |   32.8   |
-|           InternLM-7B-Chat           |   33.0   |
-|           LLaMA2-13B-Chat            |   37.1   |
-|          Baichuan2-13B-Chat          |   55.3   |
-|           LLaMA2-70B-Chat            |   59.3   |
-| **Qwen-7B-Chat (original) (0-shot)** |   41.1   |
-|      **Qwen-7B-Chat (0-shot)**       |   50.3   |
-|      **Qwen-7B-Chat (8-shot)**       |   54.1   |
-|      **Qwen-14B-Chat (0-shot)**      | **60.1** |
-|      **Qwen-14B-Chat (8-shot)**      |   59.3   |
+| Model | Zero-shot Acc. | 4-shot Acc. |
+|:--------------|:------:|:------:|
+| ChatGLM2-6B-Chat |  -  | 28.0 |
+| LLaMA2-7B-Chat | 20.4 | 28.2 |
+| LLaMA2-13B-Chat | 29.4 | 36.7 |
+| InternLM-7B-Chat | 32.6 | 34.5 |
+| Baichuan-13B-Chat | -  | 36.3 |
+| ChatGLM2-12B-Chat | -  | 38.1 |
+| **Qwen-7B-Chat** | **41.1** | **43.5** |
 
 ### 长序列评测（Long-Context Understanding）
 
 通过NTK插值，LogN注意力缩放可以扩展Qwen-7B-Chat的上下文长度。在长文本摘要数据集[VCSUM](https://arxiv.org/abs/2305.05280)上（文本平均长度在15K左右），Qwen-7B-Chat的Rouge-L结果如下：
 
-**(若要启用这些技巧，请将config.json里的`use_dynamic_ntk`和`use_logn_attn`设置为true)**
+**(若要启用这些技巧，请将config.json里的`use_dynamc_ntk`和`use_logn_attn`设置为true)**
 
 We introduce NTK-aware interpolation, LogN attention scaling to extend the context length of Qwen-7B-Chat. The Rouge-L results of Qwen-7B-Chat on long-text summarization dataset [VCSUM](https://arxiv.org/abs/2305.05280) (The average length of this dataset is around 15K) are shown below:
 
 **(To use these tricks, please set `use_dynamic_ntk` and `use_long_attn` to true in config.json.)**
 
-| Model             | VCSUM (zh) |
-|:------------------|:----------:|
-| GPT-3.5-Turbo-16k |    16.0    |
-| LLama2-7B-Chat    |    0.2     |
-| InternLM-7B-Chat  |    13.0    |
-| ChatGLM2-6B-Chat  |    16.3    |
-| **Qwen-7B-Chat**  |  **16.6**  |
+| Model | VCSUM (zh) |
+|:----------------|:-------:|
+| GPT-3.5-Turbo-16k | 16.0 |
+| LLama2-7B-Chat	|	0.2 |
+| InternLM-7B-Chat | 13.0 |
+| ChatGLM2-6B-Chat	| 16.3 |
+| **Qwen-7B-Chat** | **16.6** |
 
 ### 工具使用能力的评测（Tool Usage）
 
@@ -387,278 +264,96 @@ We introduce NTK-aware interpolation, LogN attention scaling to extend the conte
 
 千问支持通过 [ReAct Prompting](https://arxiv.org/abs/2210.03629) 调用插件/工具/API。ReAct 也是 [LangChain](https://python.langchain.com/) 框架采用的主要方式之一。在我们开源的、用于评估工具使用能力的评测基准上，千问的表现如下：
 
-Qwen-Chat supports calling plugins/tools/APIs through [ReAct Prompting](https://arxiv.org/abs/2210.03629). ReAct is also one of the main approaches used by the [LangChain](https://python.langchain.com/) framework. In our evaluation benchmark for assessing tool usage capabilities, Qwen-Chat's performance is as follows:
-
-<table>
-    <tr>
-        <th colspan="4" align="center">Chinese Tool-Use Benchmark</th>
-    </tr>
-    <tr>
-        <th align="center">Model</th><th align="center">Tool Selection (Acc.↑)</th><th align="center">Tool Input (Rouge-L↑)</th><th align="center">False Positive Error↓</th>
-    </tr>
-    <tr>
-        <td>GPT-4</td><td align="center">95%</td><td align="center">0.90</td><td align="center">15.0%</td>
-    </tr>
-    <tr>
-        <td>GPT-3.5</td><td align="center">85%</td><td align="center">0.88</td><td align="center">75.0%</td>
-    </tr>
-    <tr>
-        <td>Qwen-7B-Chat</td><td align="center">98%</td><td align="center">0.91</td><td align="center">7.3%</td>
-    </tr>
-    <tr>
-        <td>Qwen-14B-Chat</td><td align="center">98%</td><td align="center">0.93</td><td align="center">2.4%</td>
-    </tr>
-</table>
+Qwen-7B-Chat supports calling plugins/tools/APIs through [ReAct Prompting](https://arxiv.org/abs/2210.03629). ReAct is also one of the main approaches used by the [LangChain](https://python.langchain.com/) framework. In our evaluation benchmark for assessing tool usage capabilities, Qwen-7B-Chat's performance is as follows:
+
+| Model            | Tool Selection (Acc.↑) | Tool Input (Rouge-L↑) | False Positive Error↓ |
+|:-----------------|:----------------------:|:---------------------:|:---------------------:|
+| GPT-4            | 95%                    | **0.90**              | 15%                   |
+| GPT-3.5          | 85%                    | 0.88                  | 75%                   |
+| **Qwen-7B-Chat** | **99%**                | 0.89                  | **9.7%**              |
 
 > 评测基准中出现的插件均没有出现在千问的训练集中。该基准评估了模型在多个候选插件中选择正确插件的准确率、传入插件的参数的合理性、以及假阳率。假阳率（False Positive）定义：在处理不该调用插件的请求时，错误地调用了插件。
 
-> The plugins that appear in the evaluation set do not appear in the training set of Qwen. This benchmark evaluates the accuracy of the model in selecting the correct plugin from multiple candidate plugins, the rationality of the parameters passed into the plugin, and the false positive rate. False Positive: Incorrectly invoking a plugin when it should not have been called when responding to a query.
+> The plugins that appear in the evaluation set do not appear in the training set of Qwen-7B-Chat. This benchmark evaluates the accuracy of the model in selecting the correct plugin from multiple candidate plugins, the rationality of the parameters passed into the plugin, and the false positive rate. False Positive: Incorrectly invoking a plugin when it should not have been called when responding to a query.
 
-![](assets/react_showcase_001.png)
-![](assets/react_showcase_002.png)
+关于 ReAct Prompting 的 prompt 怎么写、怎么使用，请参考 [ReAct 样例说明](examples/react_prompt.md)。使用工具能使模型更好地完成任务。基于千问的工具使用能力，我们能实现下图所展示的效果：
 
-#### Code Interpreter
-
-为了考察Qwen使用Python Code Interpreter完成数学解题、数据可视化、及文件处理与爬虫等任务的能力，我们专门建设并开源了一个评测这方面能力的[评测基准](https://github.com/QwenLM/Qwen-Agent/tree/main/benchmark)。
-
-我们发现Qwen在生成代码的可执行率、结果正确性上均表现较好：
-
-To assess Qwen's ability to use the Python Code Interpreter for tasks such as mathematical problem solving, data visualization, and other general-purpose tasks such as file handling and web scraping, we have created and open-sourced a benchmark specifically designed for evaluating these capabilities. You can find the benchmark at this [link](https://github.com/QwenLM/Qwen-Agent/tree/main/benchmark).
-
-We have observed that Qwen performs well in terms of code executability and result accuracy when generating code:
-
-<table>
-    <tr>
-        <th colspan="4" align="center">Executable Rate of Generated Code (%)</th>
-    </tr>
-    <tr>
-        <th align="center">Model</th><th align="center">Math↑</th><th align="center">Visualization↑</th><th align="center">General↑</th>
-    </tr>
-    <tr>
-        <td>GPT-4</td><td align="center">91.9</td><td align="center">85.9</td><td align="center">82.8</td>
-    </tr>
-    <tr>
-        <td>GPT-3.5</td><td align="center">89.2</td><td align="center">65.0</td><td align="center">74.1</td>
-    </tr>
-    <tr>
-        <td>LLaMA2-7B-Chat</td>
-        <td align="center">41.9</td>
-        <td align="center">33.1</td>
-        <td align="center">24.1 </td>
-    </tr>
-    <tr>
-        <td>LLaMA2-13B-Chat</td>
-        <td align="center">50.0</td>
-        <td align="center">40.5</td>
-        <td align="center">48.3 </td>
-    </tr>
-    <tr>
-        <td>CodeLLaMA-7B-Instruct</td>
-        <td align="center">85.1</td>
-        <td align="center">54.0</td>
-        <td align="center">70.7 </td>
-    </tr>
-    <tr>
-        <td>CodeLLaMA-13B-Instruct</td>
-        <td align="center">93.2</td>
-        <td align="center">55.8</td>
-        <td align="center">74.1 </td>
-    </tr>
-    <tr>
-        <td>InternLM-7B-Chat-v1.1</td>
-        <td align="center">78.4</td>
-        <td align="center">44.2</td>
-        <td align="center">62.1 </td>
-    </tr>
-    <tr>
-        <td>InternLM-20B-Chat</td>
-        <td align="center">70.3</td>
-        <td align="center">44.2</td>
-        <td align="center">65.5 </td>
-    </tr>
-    <tr>
-        <td>Qwen-7B-Chat</td>
-        <td align="center">82.4</td>
-        <td align="center">64.4</td>
-        <td align="center">67.2 </td>
-    </tr>
-    <tr>
-        <td>Qwen-14B-Chat</td>
-        <td align="center">89.2</td>
-        <td align="center">84.1</td>
-        <td align="center">65.5</td>
-    </tr>
-</table>
-
-<table>
-    <tr>
-        <th colspan="4" align="center">Accuracy of Code Execution Results (%)</th>
-    </tr>
-    <tr>
-        <th align="center">Model</th><th align="center">Math↑</th><th align="center">Visualization-Hard↑</th><th align="center">Visualization-Easy↑</th>
-    </tr>
-    <tr>
-        <td>GPT-4</td><td align="center">82.8</td><td align="center">66.7</td><td align="center">60.8</td>
-    </tr>
-    <tr>
-        <td>GPT-3.5</td><td align="center">47.3</td><td align="center">33.3</td><td align="center">55.7</td>
-    </tr>
-    <tr>
-        <td>LLaMA2-7B-Chat</td>
-        <td align="center">3.9</td>
-        <td align="center">14.3</td>
-        <td align="center">39.2 </td>
-    </tr>
-    <tr>
-        <td>LLaMA2-13B-Chat</td>
-        <td align="center">8.3</td>
-        <td align="center">8.3</td>
-        <td align="center">40.5 </td>
-    </tr>
-    <tr>
-        <td>CodeLLaMA-7B-Instruct</td>
-        <td align="center">14.3</td>
-        <td align="center">26.2</td>
-        <td align="center">60.8 </td>
-    </tr>
-    <tr>
-        <td>CodeLLaMA-13B-Instruct</td>
-        <td align="center">28.2</td>
-        <td align="center">27.4</td>
-        <td align="center">62.0 </td>
-    </tr>
-    <tr>
-        <td>InternLM-7B-Chat-v1.1</td>
-        <td align="center">28.5</td>
-        <td align="center">4.8</td>
-        <td align="center">40.5 </td>
-    </tr>
-    <tr>
-        <td>InternLM-20B-Chat</td>
-        <td align="center">34.6</td>
-        <td align="center">21.4</td>
-        <td align="center">45.6 </td>
-    </tr>
-    <tr>
-        <td>Qwen-7B-Chat</td>
-        <td align="center">41.9</td>
-        <td align="center">40.5</td>
-        <td align="center">54.4 </td>
-    </tr>
-    <tr>
-        <td>Qwen-14B-Chat</td>
-        <td align="center">58.4</td>
-        <td align="center">53.6</td>
-        <td align="center">59.5</td>
-    </tr>
-</table>
+For how to write and use prompts for ReAct Prompting, please refer to [the ReAct examples](examples/react_prompt.md). The use of tools can enable the model to better perform tasks, as shown in the following figures:
 
-<p align="center">
-    <br>
-    <img src="assets/code_interpreter_showcase_001.jpg" />
-    <br>
-<p>
+![](assets/react_showcase_001.png)
+![](assets/react_showcase_002.png)
 
 #### Huggingface Agent
 
 千问还具备作为 [HuggingFace Agent](https://huggingface.co/docs/transformers/transformers_agents) 的能力。它在 Huggingface 提供的run模式评测基准上的表现如下：
 
-Qwen-Chat also has the capability to be used as a [HuggingFace Agent](https://huggingface.co/docs/transformers/transformers_agents). Its performance on the run-mode benchmark provided by HuggingFace is as follows:
-
-<table>
-    <tr>
-        <th colspan="4" align="center">HuggingFace Agent Benchmark- Run Mode</th>
-    </tr>
-    <tr>
-        <th align="center">Model</th><th align="center">Tool Selection↑</th><th align="center">Tool Used↑</th><th align="center">Code↑</th>
-    </tr>
-    <tr>
-        <td>GPT-4</td><td align="center">100</td><td align="center">100</td><td align="center">97.4</td>
-    </tr>
-    <tr>
-        <td>GPT-3.5</td><td align="center">95.4</td><td align="center">96.3</td><td align="center">87.0</td>
-    </tr>
-    <tr>
-        <td>StarCoder-Base-15B</td><td align="center">86.1</td><td align="center">87.0</td><td align="center">68.9</td>
-    </tr>
-    <tr>
-        <td>StarCoder-15B</td><td align="center">87.0</td><td align="center">88.0</td><td align="center">68.9</td>
-    </tr>
-    <tr>
-        <td>Qwen-7B-Chat</td><td align="center">87.0</td><td align="center">87.0</td><td align="center">71.5</td>
-    </tr>
-    <tr>
-        <td>Qwen-14B-Chat</td><td align="center">93.5</td><td align="center">94.4</td><td align="center">87.0</td>
-    </tr>
-</table>
-
-<table>
-    <tr>
-        <th colspan="4" align="center">HuggingFace Agent Benchmark - Chat Mode</th>
-    </tr>
-    <tr>
-        <th align="center">Model</th><th align="center">Tool Selection↑</th><th align="center">Tool Used↑</th><th align="center">Code↑</th>
-    </tr>
-    <tr>
-        <td>GPT-4</td><td align="center">97.9</td><td align="center">97.9</td><td align="center">98.5</td>
-    </tr>
-    <tr>
-        <td>GPT-3.5</td><td align="center">97.3</td><td align="center">96.8</td><td align="center">89.6</td>
-    </tr>
-    <tr>
-        <td>StarCoder-Base-15B</td><td align="center">97.9</td><td align="center">97.9</td><td align="center">91.1</td>
-    </tr>
-    <tr>
-        <td>StarCoder-15B</td><td align="center">97.9</td><td align="center">97.9</td><td align="center">89.6</td>
-    </tr>
-    <tr>
-        <td>Qwen-7B-Chat</td><td align="center">94.7</td><td align="center">94.7</td><td align="center">85.1</td>
-    </tr>
-    <tr>
-        <td>Qwen-14B-Chat</td><td align="center">97.9</td><td align="center">97.9</td><td align="center">95.5</td>
-    </tr>
-</table>
+Qwen-7B-Chat also has the capability to be used as a [HuggingFace Agent](https://huggingface.co/docs/transformers/transformers_agents). Its performance on the run-mode benchmark provided by HuggingFace is as follows:
 
-<br>
+| Model | Tool Selection↑ | Tool Used↑ | Code↑ |
+|:-|:-:|:-:|:-:|
+|GPT-4 | **100** | **100** | **97.41** |
+|GPT-3.5 | 95.37 | 96.30 | 87.04 |
+|StarCoder-15.5B | 87.04 | 87.96 | 68.89 |
+| **Qwen-7B** | 90.74 | 92.59 | 74.07 |
 
-## x86 平台 (x86 Platforms)
-在 酷睿™/至强® 可扩展处理器或 Arc™ GPU 上部署量化模型时，建议使用 [OpenVINO™ Toolkit](https://docs.openvino.ai/2023.3/gen_ai_guide.html)以充分利用硬件，实现更好的推理性能。您可以安装并运行此 [example notebook](https://github.com/openvinotoolkit/openvino_notebooks/tree/main/notebooks/254-llm-chatbot)。相关问题，您可在[OpenVINO repo](https://github.com/openvinotoolkit/openvino_notebooks/issues)中提交。
+## 量化（Quantization）
 
-When deploy on Core™/Xeon® Scalable Processors or with Arc™ GPU, [OpenVINO™ Toolkit](https://docs.openvino.ai/2023.3/gen_ai_guide.html) is recommended. You can install and run this [example notebook](https://github.com/openvinotoolkit/openvino_notebooks/tree/main/notebooks/254-llm-chatbot). For related issues, you are welcome to file an issue at [OpenVINO repo](https://github.com/openvinotoolkit/openvino_notebooks/issues).
+如希望使用更低精度的量化模型，如4比特和8比特的模型，我们提供了简单的示例来说明如何快速使用量化模型。在开始前，确保你已经安装了`bitsandbytes`。请注意，`bitsandbytes`的安装要求是：
 
-## FAQ
+We provide examples to show how to load models in `NF4` and `Int8`. For starters, make sure you have implemented `bitsandbytes`. Note that the requirements for `bitsandbytes` are:
 
-如遇到问题，敬请查阅[FAQ](https://github.com/QwenLM/Qwen/blob/main/FAQ_zh.md)以及issue区，如仍无法解决再提交issue。
+```
+**Requirements** Python >=3.8. Linux distribution (Ubuntu, MacOS, etc.) + CUDA > 10.0.
+```
 
-If you meet problems, please refer to [FAQ](https://github.com/QwenLM/Qwen/blob/main/FAQ.md) and the issues first to search a solution before you launch a new issue.
-<br>
+Windows用户需安装特定版本的`bitsandbytes`，可选项包括[bitsandbytes-windows-webui](https://github.com/jllllll/bitsandbytes-windows-webui/releases/tag/wheels)。
 
-## 引用 (Citation)
+Windows users should find another option, which might be [bitsandbytes-windows-webui](https://github.com/jllllll/bitsandbytes-windows-webui/releases/tag/wheels).
 
-如果你觉得我们的工作对你有帮助，欢迎引用！
+你只需要在`AutoModelForCausalLM.from_pretrained`中添加你的量化配置，即可使用量化模型。如下所示：
 
-If you find our work helpful, feel free to give us a cite.
+Then you only need to add your quantization configuration to `AutoModelForCausalLM.from_pretrained`. See the example below:
 
+```python
+from transformers import AutoModelForCausalLM, BitsAndBytesConfig
+
+# quantization configuration for NF4 (4 bits)
+quantization_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_quant_type='nf4',
+    bnb_4bit_compute_dtype=torch.bfloat16
+)
+
+# quantization configuration for Int8 (8 bits)
+quantization_config = BitsAndBytesConfig(load_in_8bit=True)
+
+model = AutoModelForCausalLM.from_pretrained(
+    "Qwen/Qwen-7B-Chat",
+    device_map="cuda:0",
+    quantization_config=quantization_config,
+    max_memory=max_memory,
+    trust_remote_code=True,
+).eval()
 ```
-@article{qwen,
-  title={Qwen Technical Report},
-  author={Jinze Bai and Shuai Bai and Yunfei Chu and Zeyu Cui and Kai Dang and Xiaodong Deng and Yang Fan and Wenbin Ge and Yu Han and Fei Huang and Binyuan Hui and Luo Ji and Mei Li and Junyang Lin and Runji Lin and Dayiheng Liu and Gao Liu and Chengqiang Lu and Keming Lu and Jianxin Ma and Rui Men and Xingzhang Ren and Xuancheng Ren and Chuanqi Tan and Sinan Tan and Jianhong Tu and Peng Wang and Shijie Wang and Wei Wang and Shengguang Wu and Benfeng Xu and Jin Xu and An Yang and Hao Yang and Jian Yang and Shusheng Yang and Yang Yao and Bowen Yu and Hongyi Yuan and Zheng Yuan and Jianwei Zhang and Xingxuan Zhang and Yichang Zhang and Zhenru Zhang and Chang Zhou and Jingren Zhou and Xiaohuan Zhou and Tianhang Zhu},
-  journal={arXiv preprint arXiv:2309.16609},
-  year={2023}
-}
-```
-<br>
+
+上述方法可以让我们将模型量化成`NF4`和`Int8`精度的模型进行读取，帮助我们节省显存开销。我们也提供了相关性能数据。我们发现尽管模型在效果上存在损失，但模型的显存开销大幅降低。
+
+With this method, it is available to load Qwen-7B-Chat in `NF4`and `Int8`, which saves you memory usage. We provide related statistics of model performance below. We find that the quantization downgrades the effectiveness slightly but significantly increases inference efficiency and reduces memory costs.
+
+| Precision | MMLU | Memory |
+| :---------| :-------: | :-----: |
+|   BF16   |  56.7 |   16.2G |
+|   Int8   |  52.8 |   10.1G |
+|    NF4    |  48.9 |    7.4G |
 
 ## 使用协议（License Agreement）
 
-我们的代码和模型权重对学术研究完全开放，并支持商用。请查看[LICENSE](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT)了解具体的开源协议细节。如需商用，请填写[问卷](https://dashscope.console.aliyun.com/openModelApply/qianwen)申请。
+我们的代码和模型权重对学术研究完全开放，并支持商用。请查看LICENSE了解具体的开源协议细节。
 
-Our code and checkpoints are open to research purpose, and they are allowed for commercial purposes. Check [LICENSE](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT) for more details about the license. If you have requirements for commercial use, please fill out the [form](https://dashscope.console.aliyun.com/openModelApply/qianwen) to apply.
-<br>
+Our code and checkpoints are open to research purpose, and they are allowed for commercial purposes. Check [LICENSE](LICENSE) for more details about the license.
 
 ## 联系我们（Contact Us）
 
-如果你想给我们的研发团队和产品团队留言，欢迎加入我们的微信群、钉钉群以及Discord！同时，也欢迎通过邮件（qianwen_opensource@alibabacloud.com）联系我们。
-
-If you are interested to leave a message to either our research team or product team, join our Discord or WeChat groups! Also, feel free to send an email to qianwen_opensource@alibabacloud.com.
+如果你想给我们的研发团队和产品团队留言，请通过邮件（qianwen_opensource@alibabacloud.com）联系我们。
 
+If you are interested to leave a message to either our research team or product team, feel free to send an email to qianwen_opensource@alibabacloud.com.

cache_autogptq_cuda_256.cpp

@@ -1,198 +0,0 @@
-#include <torch/all.h>
-#include <torch/python.h>
-#include <c10/cuda/CUDAGuard.h>
-
-// adapted from https://github.com/PanQiWei/AutoGPTQ/blob/main/autogptq_extension/cuda_256/autogptq_cuda_256.cpp
-void vecquant8matmul_cuda(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros,
-  torch::Tensor g_idx
-);
-
-void vecquant8matmul(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros,
-  torch::Tensor g_idx
-) {
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
-  vecquant8matmul_cuda(vec, mat, mul, scales, zeros, g_idx);
-}
-
-void vecquant8matmul_batched_cuda(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-);
-
-void vecquant8matmul_batched(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-) {
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
-  vecquant8matmul_batched_cuda(vec, mat, mul, scales, zeros);
-}
-
-void vecquant8matmul_batched_column_compression_cuda(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-);
-
-void vecquant8matmul_batched_column_compression(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-) {
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
-  vecquant8matmul_batched_column_compression_cuda(vec, mat, mul, scales, zeros);
-}
-
-void vecquant4matmul_batched_cuda(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-);
-
-void vecquant4matmul_batched(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-) {
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
-  vecquant4matmul_batched_cuda(vec, mat, mul, scales, zeros);
-}
-
-void vecquant4matmul_batched_column_compression_cuda(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-);
-
-void vecquant4matmul_batched_column_compression(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-) {
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
-  vecquant4matmul_batched_column_compression_cuda(vec, mat, mul, scales, zeros);
-}
-
-void vecquant8matmul_batched_old_cuda(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-);
-
-void vecquant8matmul_batched_old(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-) {
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
-  vecquant8matmul_batched_old_cuda(vec, mat, mul, scales, zeros);
-}
-
-
-void vecquant4matmul_batched_old_cuda(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-);
-
-void vecquant4matmul_batched_old(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-) {
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
-  vecquant4matmul_batched_old_cuda(vec, mat, mul, scales, zeros);
-}
-
-void vecquant8matmul_batched_column_compression_old_cuda(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-);
-
-void vecquant8matmul_batched_column_compression_old(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-) {
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
-  vecquant8matmul_batched_column_compression_old_cuda(vec, mat, mul, scales, zeros);
-}
-
-void vecquant4matmul_batched_column_compression_old_cuda(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-);
-
-void vecquant4matmul_batched_column_compression_old(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-) {
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
-  vecquant4matmul_batched_column_compression_old_cuda(vec, mat, mul, scales, zeros);
-}
-
-
-
-void vecquant8matmul_batched_faster_cuda(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-);
-
-void vecquant8matmul_batched_faster(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-) {
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
-  vecquant8matmul_batched_faster_cuda(vec, mat, mul, scales, zeros);
-}
-
-
-void vecquant8matmul_batched_faster_old_cuda(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-);
-
-void vecquant8matmul_batched_faster_old(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-) {
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
-  vecquant8matmul_batched_faster_old_cuda(vec, mat, mul, scales, zeros);
-}
-
-void vecquant8matmul_batched_column_compression_faster_cuda(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-);
-
-void vecquant8matmul_batched_column_compression_faster(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-) {
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
-  vecquant8matmul_batched_column_compression_faster_cuda(vec, mat, mul, scales, zeros);
-}
-
-
-void vecquant8matmul_batched_column_compression_faster_old_cuda(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-);
-
-void vecquant8matmul_batched_column_compression_faster_old(
-  torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
-  torch::Tensor scales, torch::Tensor zeros
-) {
-  const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
-  vecquant8matmul_batched_column_compression_faster_old_cuda(vec, mat, mul, scales, zeros);
-}
-
-
-
-PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
-  m.def("vecquant8matmul", &vecquant8matmul, "Vector 8-bit Quantized Matrix Multiplication (CUDA) (desc_act)");
-  m.def("vecquant8matmul_batched", &vecquant8matmul_batched, "Vector 8-bit Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
-  m.def("vecquant8matmul_batched_old", &vecquant8matmul_batched_old, "Vector 8-bit old Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
-  m.def("vecquant8matmul_batched_faster", &vecquant8matmul_batched_faster, "Vector 8-bit old Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
-  m.def("vecquant8matmul_batched_faster_old", &vecquant8matmul_batched_faster_old, "Vector 8-bit old Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
-  m.def("vecquant4matmul_batched_old", &vecquant4matmul_batched_old, "Vector 4-bit old Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
-  m.def("vecquant8matmul_batched_column_compression", &vecquant8matmul_batched_column_compression, "Vector 8-bit Batched Quantized Matrix Multiplication (CUDA) with weight's column compressed (desc_act)");
-  m.def("vecquant8matmul_batched_column_compression_old", &vecquant8matmul_batched_column_compression_old, "Vector old 8-bit Batched Quantized Matrix Multiplication (CUDA) with weight's column compressed (desc_act)");
-  m.def("vecquant8matmul_batched_column_compression_faster", &vecquant8matmul_batched_column_compression_faster, "Vector old 8-bit Batched Quantized Matrix Multiplication (CUDA) with weight's column compressed (desc_act)");
-  m.def("vecquant8matmul_batched_column_compression_faster_old", &vecquant8matmul_batched_column_compression_faster_old, "Vector old 8-bit Batched Quantized Matrix Multiplication (CUDA) with weight's column compressed (desc_act)");
-  m.def("vecquant4matmul_batched_column_compression_old", &vecquant4matmul_batched_column_compression_old, "Vector old 4-bit Batched Quantized Matrix Multiplication (CUDA) with weight's column compressed (desc_act)");
-  m.def("vecquant4matmul_batched", &vecquant4matmul_batched, "Vector 4-bit Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
-  m.def("vecquant4matmul_batched_column_compression", &vecquant4matmul_batched_column_compression, "Vector 4-bit Batched Quantized Matrix Multiplication (CUDA) with weight's column compressed (desc_act)");
-}

cache_autogptq_cuda_kernel_256.cu

@@ -1,1708 +0,0 @@
-#define _CRT_SECURE_NO_WARNINGS
-#include <torch/all.h>
-#include <torch/python.h>
-#include <cuda.h>
-#include <cuda_runtime.h>
-#include <cuda_fp16.h>
-#include <stdint.h>
-
-#if (defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 700) || defined(USE_ROCM)
-// adapted from https://github.com/PanQiWei/AutoGPTQ/blob/main/autogptq_extension/cuda_256/autogptq_cuda_kernel_256.cu
-__device__ __forceinline__ void atomicAdd(c10::Half* address, c10::Half val) {
-    unsigned int *address_as_ui = reinterpret_cast<unsigned int *>(reinterpret_cast<char *>(address) - (reinterpret_cast<size_t>(address) & 2));
-    unsigned int old = *address_as_ui;
-    unsigned int assumed;
-
-    do {
-        assumed = old;
-        unsigned short hsum = reinterpret_cast<size_t>(address) & 2 ? (old >> 16) : (old & 0xffff);
-        hsum += val;
-        old = reinterpret_cast<size_t>(address) & 2
-                 ? (old & 0xffff) | (hsum << 16)
-                 : (old & 0xffff0000) | hsum;
-        old = atomicCAS(address_as_ui, assumed, old);
-
-    // Note: uses integer comparison to avoid hang in case of NaN (since NaN != NaN)
-    } while (assumed != old);
-}
-__device__ __forceinline__ void atomicAdd(__half* address, c10::Half val) {
-    unsigned int * address_as_ui = (unsigned int *) ((char *)address - ((size_t)address & 2));
-    unsigned int old = *address_as_ui;
-    unsigned int assumed;
-
-    do {
-        assumed = old;
-        __half_raw hsum;
-        hsum.x = (size_t)address & 2 ? (old >> 16) : (old & 0xffff);
-        half tmpres = __hadd(hsum, val);
-        hsum = __half_raw(tmpres);
-        old = (size_t)address & 2 ? (old & 0xffff) | (hsum.x << 16) : (old & 0xffff0000) | hsum.x;
-        old = atomicCAS(address_as_ui, assumed, old);
-    } while (assumed != old);
-}
-#endif
-
-template <typename scalar_t>
-__global__ void VecQuant8MatMulKernel(
-    const  scalar_t* __restrict__ vec,
-    const       int* __restrict__ mat,
-           scalar_t* __restrict__ mul,
-    const  scalar_t* __restrict__ scales,
-    const       int* __restrict__ zeros,
-    const       int* __restrict__ g_idx,
-    int batch,
-    int vec_height,
-    int height,
-    int width,
-    int zero_width
-);
-
-template <typename scalar_t>
-__global__ void VecQuant8BatchMatMulColumnCompressionKernel(
-    const  scalar_t* __restrict__ vec,
-    const       int* __restrict__ mat,
-           scalar_t* __restrict__ mul,
-    const  scalar_t* __restrict__ scales,
-    const       int* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int height,
-    int width
-);
-
-template <typename scalar_t>
-__global__ void VecQuant4BatchMatMulColumnCompressionKernel(
-    const  scalar_t* __restrict__ vec,
-    const       int* __restrict__ mat,
-           scalar_t* __restrict__ mul,
-    const  scalar_t* __restrict__ scales,
-    const       int* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int height,
-    int width
-);
-
-template <typename scalar_t>
-__global__ void VecQuant8BatchMatMulKernel(
-    const  scalar_t* __restrict__ vec,
-    const       int* __restrict__ mat,
-           scalar_t* __restrict__ mul,
-    const  scalar_t* __restrict__ scales,
-    const       int* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int vec_height,
-    int height,
-    int width,
-    int zero_width
-);
-
-template <typename scalar_t>
-__global__ void VecQuant4BatchMatMulKernel(
-    const  scalar_t* __restrict__ vec,
-    const       int* __restrict__ mat,
-           scalar_t* __restrict__ mul,
-    const  scalar_t* __restrict__ scales,
-    const       int* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int vec_height,
-    int height,
-    int width,
-    int zero_width
-);
-
-
-
-template <typename scalar_t>
-__global__ void VecQuant8BatchMatMulKernel_old(
-    const  scalar_t* __restrict__ vec,
-    const  uint8_t* __restrict__ mat,
-           scalar_t* __restrict__ mul,
-    const  scalar_t* __restrict__ scales,
-    const  scalar_t* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int vec_height,
-    int height,
-    int width,
-    int zero_width
-);
-
-__global__ void VecQuant8BatchMatMulKernel_faster(
-    const  half* __restrict__ vec,
-    const  uint8_t* __restrict__ mat,
-           half* __restrict__ mul,
-    const  half* __restrict__ scales,
-    const  half* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int vec_height,
-    int height,
-    int width,
-    int zero_width
-);
-
-
-
-__global__ void VecQuant8BatchMatMulKernel_faster_old(
-    const  half* __restrict__ vec,
-    const  uint8_t* __restrict__ mat,
-           half* __restrict__ mul,
-    const  half* __restrict__ scales,
-    const  half* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int vec_height,
-    int height,
-    int width
-);
-
-
-template <typename scalar_t>
-__global__ void VecQuant4BatchMatMulKernel_old(
-    const  scalar_t* __restrict__ vec,
-    const  uint8_t* __restrict__ mat,
-           scalar_t* __restrict__ mul,
-    const  scalar_t* __restrict__ scales,
-    const  scalar_t* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int vec_height,
-    int height,
-    int width,
-    int zero_width
-);
-
-
-template <typename scalar_t>
-__global__ void VecQuant8BatchMatMulColumnCompressionKernel_old(
-    const  scalar_t* __restrict__ vec,
-    const  uint8_t* __restrict__ mat,
-           scalar_t* __restrict__ mul,
-    const  scalar_t* __restrict__ scales,
-    const  scalar_t* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int height,
-    int width
-);
-
-__global__ void VecQuant8BatchMatMulColumnCompressionKernel_faster(
-    const  half* __restrict__ vec,
-    const  uint8_t* __restrict__ mat,
-           half* __restrict__ mul,
-    const  half* __restrict__ scales,
-    const  half* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int height,
-    int width
-);
-
-__global__ void VecQuant8BatchMatMulColumnCompressionKernel_faster_old(
-    const  half* __restrict__ vec,
-    const  uint8_t* __restrict__ mat,
-           half* __restrict__ mul,
-    const  half* __restrict__ scales,
-    const  half* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int height,
-    int width
-);
-
-
-template <typename scalar_t>
-__global__ void VecQuant4BatchMatMulColumnCompressionKernel_old(
-    const  scalar_t* __restrict__ vec,
-    const  uint8_t* __restrict__ mat,
-           scalar_t* __restrict__ mul,
-    const  scalar_t* __restrict__ scales,
-    const  scalar_t* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int height,
-    int width
-);
-
-
-__global__ void VecQuant8BatchMatMulKernel_faster(
-    const  half* __restrict__ vec,
-    const  uint8_t* __restrict__ mat,
-           half* __restrict__ mul,
-    const  half* __restrict__ scales,
-    const  half* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int vec_height,
-    int height,
-    int width
-);
-
-
-__global__ void VecQuant8BatchMatMulColumnCompressionKernel_faster(
-    const  half* __restrict__ vec,
-    const  uint8_t* __restrict__ mat,
-           half* __restrict__ mul,
-    const  half* __restrict__ scales,
-    const  half* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int height,
-    int width
-);
-
-const int BLOCKWIDTH  = 128;
-const int BLOCKHEIGHT8 =  32;
-const int BLOCKHEIGHT4 =  16;
-const int BLOCKHEIGHT_OLD4 =  128;
-//const int BLOCKHEIGHT_OLD8 =  128;
-
-__device__ inline unsigned int as_unsigned(int i) {
-  return *reinterpret_cast<unsigned int*>(&i);
-}
-
-__device__ inline int as_int(int i) {
-  return *reinterpret_cast<int*>(&i);
-}
-
-void vecquant8matmul_batched_column_compression_cuda(
-  torch::Tensor vec,
-  torch::Tensor mat,
-  torch::Tensor mul,
-  torch::Tensor scales,
-  torch::Tensor zeros
-) {
-  int batch = vec.size(0);
-  int heads = vec.size(1);
-  int vec_row = vec.size(2);
-  int height = vec.size(3);
-  int width = mat.size(3) * 4;
-
-  dim3 blocks(
-    (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
-    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
-  );
-  dim3 threads(BLOCKWIDTH);
-
-  AT_DISPATCH_FLOATING_TYPES(
-    vec.type(), "vecquant8matmul_batched_cuda", ([&] {
-      VecQuant8BatchMatMulColumnCompressionKernel<<<blocks, threads>>>(
-        vec.data<scalar_t>(), mat.data<int>(), mul.data<scalar_t>(),
-        scales.data<scalar_t>(), zeros.data<int>(),
-        batch, heads, vec_row, height, width
-      );
-    })
-  );
-
-}
-
-template <typename scalar_t>
-__global__ void VecQuant8BatchMatMulColumnCompressionKernel(
-    const  scalar_t* __restrict__ vec,
-    const       int* __restrict__ mat,
-           scalar_t* __restrict__ mul,
-    const  scalar_t* __restrict__ scales,
-    const       int* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int height,
-    int width
-) {
-  int weight_total = batch * heads * height * width / 4;
-  int input_total = batch * heads * vec_row * height;
-  int out_total = batch * heads * vec_row * width;
-  int tid = threadIdx.x;
-  // h is index of height with step being BLOCKWIDTH
-  int h = BLOCKWIDTH * blockIdx.x;
-  // w is index of width with step being 1
-  int w = BLOCKWIDTH * blockIdx.y + tid;
-  if (w >= width && tid >= height) {
-    return;
-  }
-
-  __shared__ scalar_t blockvec[BLOCKWIDTH];
-  int k;
-  scalar_t w_tmp;
-
-  float weight[BLOCKWIDTH];
-
-  for (int b = 0; b < batch; ++b){
-    for (int head = 0; head < heads; ++head){
-      int batch_shift = b * heads + head;
-      for (k = 0; k <  BLOCKWIDTH && h + k < height; ++k){
-        int i_w = (w / 4);
-        int w_bit = (w % 4) * 8;
-
-        int w_index = (batch_shift * height + h + k) * width / 4 + i_w;
-        if (w_index >= weight_total || w >= width) {
-          weight[k] = 0;
-        } else {
-          scalar_t scale = scales[batch_shift * height + h + k];
-          scalar_t zero = zeros[batch_shift * height + h + k];
-          w_tmp = ((as_unsigned(mat[w_index]) >> w_bit) & 0xFF);
-          weight[k] = scale * (w_tmp - zero);
-        }
-      }
-
-      scalar_t res;
-      for (int vr = 0; vr < vec_row; ++vr){
-          res = 0;
-        int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
-        if (vec_index < input_total) {
-            blockvec[tid] = vec[vec_index];
-        } else {
-            blockvec[tid] = 0;
-        }
-
-        __syncthreads();
-          for (k = 0; k <  BLOCKWIDTH && h + k < height; ++k){
-          // res is the dot product of BLOCKWIDTH elements (part of width)
-            res += weight[k] * blockvec[k];
-        }
-        // add res to the final result, final matrix shape: (batch, vec_row, width)
-        int out_index = (batch_shift * vec_row + vr) * width + w;
-        if (out_index < out_total) {
-            atomicAdd(&mul[out_index], res);
-        }
-        __syncthreads();
-      }
-    }
-  }
-}
-
-void vecquant8matmul_batched_cuda(
-  torch::Tensor vec,
-  torch::Tensor mat,
-  torch::Tensor mul,
-  torch::Tensor scales,
-  torch::Tensor zeros
-) {
-  int batch = vec.size(0);
-  int heads = vec.size(1);
-  int vec_row = vec.size(2);
-  int vec_height = vec.size(3);
-  int height = mat.size(2);
-  int width = mat.size(3);
-  int zero_width = zeros.size(2);
-
-  dim3 blocks(
-    (height + BLOCKHEIGHT8 - 1) / BLOCKHEIGHT8,
-    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
-  );
-  dim3 threads(BLOCKWIDTH);
-
-  AT_DISPATCH_FLOATING_TYPES(
-    vec.type(), "vecquant8matmul_batched_cuda", ([&] {
-      VecQuant8BatchMatMulKernel<<<blocks, threads>>>(
-        vec.data<scalar_t>(), mat.data<int>(), mul.data<scalar_t>(),
-        scales.data<scalar_t>(), zeros.data<int>(),
-        batch, heads, vec_row, vec_height, height, width, zero_width
-      );
-    })
-  );
-
-}
-
-template <typename scalar_t>
-__global__ void VecQuant8BatchMatMulKernel(
-    const  scalar_t* __restrict__ vec,
-    const       int* __restrict__ mat,
-           scalar_t* __restrict__ mul,
-    const  scalar_t* __restrict__ scales,
-    const       int* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int vec_height,
-    int height,
-    int width,
-    int zero_width
-) {
-  int weight_total = batch * heads * height * width;
-  int input_total = batch * heads * vec_row * vec_height;
-  int out_total = batch * heads * vec_row * width;
-  int tid = threadIdx.x;
-  // h is index of height with step being BLOCKHEIGHT8
-  int h = BLOCKHEIGHT8 * blockIdx.x;
-  // w is index of width with step being 1
-  int w = BLOCKWIDTH * blockIdx.y + tid;
-  if (w >= width && tid >= vec_height) {
-    return;
-  }
-
-  __shared__ scalar_t blockvec[BLOCKWIDTH];
-  // i is index of mat of block first row
-  int i = width * h + w;
-  // if (i >= width * height) {
-  //   return;
-  // }
-  int k;
-  scalar_t w_tmp;
-
-  int z_w = w / 4;
-  int z_mod = (w % 4) * 8;
-
-  float weight[BLOCKWIDTH];
-
-  for (int b = 0; b < batch; ++b){
-    for (int head = 0; head < heads; ++head){
-      int batch_shift = b * heads + head;
-      for (k = 0; k <  BLOCKWIDTH && h * 4 + k < vec_height; ++k){
-        int k_w = (k / 4);
-        int k_bit = (k % 4) * 8;
-
-        int w_index = batch_shift * height * width + i + (k_w * width);
-        if (w_index >= weight_total || w >= width) {
-          weight[k] = 0;
-        } else {
-          scalar_t scale = scales[batch_shift * width + w];
-          scalar_t zero;
-          if (zero_width == width) {
-            zero = zeros[batch_shift * width + w];
-          } else {
-            zero = scalar_t(((as_unsigned(zeros[batch_shift * zero_width + z_w]) >> z_mod) & 0xFF) + 1);
-          }
-          w_tmp = ((as_unsigned(mat[w_index]) >> k_bit) & 0xFF);
-          weight[k] = scale * (w_tmp - zero);
-        }
-      }
-
-      scalar_t res;
-      for (int vr = 0; vr < vec_row; ++vr){
-          res = 0;
-        int vec_index = (batch_shift * vec_row + vr) * vec_height + blockIdx.x * BLOCKWIDTH + tid;
-        if (vec_index < input_total) {
-            blockvec[tid] = vec[vec_index];
-        } else {
-            blockvec[tid] = 0;
-        }
-
-        __syncthreads();
-          for (k = 0; k <  BLOCKWIDTH && h * 4 + k < vec_height; ++k){
-          // res is the dot product of BLOCKWIDTH elements (part of width)
-            res += weight[k] * blockvec[k];
-        }
-        // add res to the final result, final matrix shape: (batch, vec_row, width)
-        int out_index = (batch_shift * vec_row + vr) * width + w;
-        if (out_index < out_total) {
-            atomicAdd(&mul[out_index], res);
-        }
-        __syncthreads();
-      }
-    }
-  }
-}
-
-
-void vecquant8matmul_cuda(
-  torch::Tensor vec,
-  torch::Tensor mat,
-  torch::Tensor mul,
-  torch::Tensor scales,
-  torch::Tensor zeros,
-  torch::Tensor g_idx
-) {
-  int batch = vec.size(0);
-  int vec_height = vec.size(1);
-  int height = mat.size(0);
-  int width = mat.size(1);
-  int zero_width = zeros.size(1);
-
-  dim3 blocks(
-    (height + BLOCKHEIGHT8 - 1) / BLOCKHEIGHT8,
-    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
-  );
-  dim3 threads(BLOCKWIDTH);
-
-  AT_DISPATCH_FLOATING_TYPES(
-    vec.type(), "vecquant8matmul_cuda", ([&] {
-      VecQuant8MatMulKernel<<<blocks, threads>>>(
-        vec.data<scalar_t>(), mat.data<int>(), mul.data<scalar_t>(),
-        scales.data<scalar_t>(), zeros.data<int>(), g_idx.data<int>(),
-        batch, vec_height, height, width, zero_width
-      );
-    })
-  );
-}
-
-template <typename scalar_t>
-__global__ void VecQuant8MatMulKernel(
-    const  scalar_t* __restrict__ vec,
-    const       int* __restrict__ mat,
-           scalar_t* __restrict__ mul,
-    const  scalar_t* __restrict__ scales,
-    const       int* __restrict__ zeros,
-    const       int* __restrict__ g_idx,
-    int batch,
-    int vec_height,
-    int height,
-    int width,
-    int zero_width
-) {
-  int h = BLOCKHEIGHT8 * blockIdx.x;
-  int w = BLOCKWIDTH * blockIdx.y + threadIdx.x;
-
-  __shared__ scalar_t blockvec[BLOCKWIDTH];
-  int i = width * h + w;
-  int g_h = h * 4;
-  int k;
-  unsigned int g;
-  scalar_t w_tmp;
-
-  int z_w = w / 4;
-  int z_mod = (w % 4) * 8;
-
-  float weight[BLOCKWIDTH];
-
-  for (k = 0; k <  BLOCKWIDTH; ++k){
-    int k_w = (k / 4);
-    int k_bit = (k % 4) * 8;
-
-      g = as_int(g_idx[g_h + k]);
-      scalar_t scale = scales[g * width + w];
-      scalar_t zero = scalar_t(((as_unsigned(zeros[g * zero_width + z_w]) >> z_mod) & 0xFF) + 1);
-
-      w_tmp = ((as_unsigned(mat[i + (k_w * width)]) >> k_bit) & 0xFF);
-
-    weight[k] = scale * (w_tmp - zero);
-  }
-
-
-  scalar_t res;
-  for (int b = 0; b < batch; ++b){
-      res = 0;
-    blockvec[threadIdx.x] = vec[b * vec_height + blockIdx.x * BLOCKWIDTH + threadIdx.x];
-    __syncthreads();
-    for (k = 0; k <  BLOCKWIDTH; ++k){
-      res += weight[k] * blockvec[k];
-    }
-    atomicAdd(&mul[b * width + w], res);
-    __syncthreads();
-  }
-}
-
-
-
-void vecquant4matmul_batched_cuda(
-  torch::Tensor vec,
-  torch::Tensor mat,
-  torch::Tensor mul,
-  torch::Tensor scales,
-  torch::Tensor zeros
-) {
-  int batch = vec.size(0);
-  int heads = vec.size(1);
-  int vec_row = vec.size(2);
-  int vec_height = vec.size(3);
-  int height = mat.size(2);
-  int width = mat.size(3);
-  int zero_width = zeros.size(2);
-
-  dim3 blocks(
-    (height + BLOCKHEIGHT4 - 1) / BLOCKHEIGHT4,
-    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
-  );
-  dim3 threads(BLOCKWIDTH);
-
-  AT_DISPATCH_FLOATING_TYPES(
-    vec.type(), "vecquant4matmul_batched_cuda", ([&] {
-      VecQuant4BatchMatMulKernel<<<blocks, threads>>>(
-        vec.data<scalar_t>(), mat.data<int>(), mul.data<scalar_t>(),
-        scales.data<scalar_t>(), zeros.data<int>(),
-        batch, heads, vec_row, vec_height, height, width, zero_width
-      );
-    })
-  );
-
-}
-
-template <typename scalar_t>
-__global__ void VecQuant4BatchMatMulKernel(
-    const  scalar_t* __restrict__ vec,
-    const       int* __restrict__ mat,
-           scalar_t* __restrict__ mul,
-    const  scalar_t* __restrict__ scales,
-    const       int* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int vec_height,
-    int height,
-    int width,
-    int zero_width
-) {
-  int weight_total = batch * heads * height * width;
-  int input_total = batch * heads * vec_row * vec_height;
-  int out_total = batch * heads * vec_row * width;
-  int tid = threadIdx.x;
-  // h is index of height with step being BLOCKHEIGHT4
-  int h = BLOCKHEIGHT4 * blockIdx.x;
-  // w is index of width with step being 1
-  int w = BLOCKWIDTH * blockIdx.y + tid;
-  if (w >= width && tid >= vec_height) {
-    return;
-  }
-
-  __shared__ scalar_t blockvec[BLOCKWIDTH];
-  // i is index of mat of block first row
-  int i = width * h + w;
-  int k;
-  scalar_t w_tmp;
-
-  int z_w = w / 8;
-  int z_mod = (w % 8) * 4;
-
-  float weight[BLOCKWIDTH];
-
-  for (int b = 0; b < batch; ++b){
-    for (int head = 0; head < heads; ++head){
-      int batch_shift = b * heads + head;
-      for (k = 0; k <  BLOCKWIDTH && h * 8 + k < vec_height; ++k){
-        int k_w = (k / 8);
-        int k_bit = (k % 8) * 4;
-
-        int w_index = batch_shift * height * width + i + (k_w * width);
-        if (w_index >= weight_total || w >= width) {
-          weight[k] = 0;
-        } else {
-          scalar_t scale = scales[batch_shift * width + w];
-          scalar_t zero;
-          if (zero_width == width) {
-            zero = zeros[batch_shift * width + w];
-          } else {
-            zero = scalar_t(((as_unsigned(zeros[batch_shift * zero_width + z_w]) >> z_mod) & 0xF));
-          }
-          w_tmp = ((as_unsigned(mat[w_index]) >> k_bit) & 0xF);
-          weight[k] = scale * (w_tmp - zero);
-        }
-      }
-
-      scalar_t res;
-      for (int vr = 0; vr < vec_row; ++vr){
-          res = 0;
-        int vec_index = (batch_shift * vec_row + vr) * vec_height + blockIdx.x * BLOCKWIDTH + tid;
-        if (vec_index < input_total) {
-            blockvec[tid] = vec[vec_index];
-        } else {
-            blockvec[tid] = 0;
-        }
-
-        __syncthreads();
-          for (k = 0; k <  BLOCKWIDTH && h * 8 + k < vec_height; ++k){
-          // res is the dot product of BLOCKWIDTH elements (part of width)
-            res += weight[k] * blockvec[k];
-        }
-        // add res to the final result, final matrix shape: (batch, vec_row, width)
-        int out_index = (batch_shift * vec_row + vr) * width + w;
-        if (out_index < out_total) {
-            atomicAdd(&mul[out_index], res);
-        }
-        __syncthreads();
-      }
-    }
-  }
-}
-
-
-
-void vecquant4matmul_batched_column_compression_cuda(
-  torch::Tensor vec,
-  torch::Tensor mat,
-  torch::Tensor mul,
-  torch::Tensor scales,
-  torch::Tensor zeros
-) {
-  int batch = vec.size(0);
-  int heads = vec.size(1);
-  int vec_row = vec.size(2);
-  int height = vec.size(3);
-  int width = mat.size(3) * 8;
-
-  dim3 blocks(
-    (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
-    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
-  );
-  dim3 threads(BLOCKWIDTH);
-
-  AT_DISPATCH_FLOATING_TYPES(
-    vec.type(), "vecquant4matmul_batched_cuda", ([&] {
-      VecQuant4BatchMatMulColumnCompressionKernel<<<blocks, threads>>>(
-        vec.data<scalar_t>(), mat.data<int>(), mul.data<scalar_t>(),
-        scales.data<scalar_t>(), zeros.data<int>(),
-        batch, heads, vec_row, height, width
-      );
-    })
-  );
-
-}
-
-template <typename scalar_t>
-__global__ void VecQuant4BatchMatMulColumnCompressionKernel(
-    const  scalar_t* __restrict__ vec,
-    const       int* __restrict__ mat,
-           scalar_t* __restrict__ mul,
-    const  scalar_t* __restrict__ scales,
-    const       int* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int height,
-    int width
-) {
-  int weight_total = batch * heads * height * width / 8;
-  int input_total = batch * heads * vec_row * height;
-  int out_total = batch * heads * vec_row * width;
-  int tid = threadIdx.x;
-  // h is index of height with step being BLOCKWIDTH
-  int h = BLOCKWIDTH * blockIdx.x;
-  // w is index of width with step being 1
-  int w = BLOCKWIDTH * blockIdx.y + tid;
-  if (w >= width && tid >= height) {
-    return;
-  }
-
-  __shared__ scalar_t blockvec[BLOCKWIDTH];
-  int k;
-  scalar_t w_tmp;
-
-  float weight[BLOCKWIDTH];
-
-  for (int b = 0; b < batch; ++b){
-    for (int head = 0; head < heads; ++head){
-      int batch_shift = b * heads + head;
-      for (k = 0; k <  BLOCKWIDTH && h + k < height; ++k){
-        int i_w = (w / 8);
-        int w_bit = (w % 8) * 4;
-
-        int w_index = (batch_shift * height + h + k) * width / 8 + i_w;
-        if (w_index >= weight_total || w >= width) {
-          weight[k] = 0;
-        } else {
-          scalar_t scale = scales[batch_shift * height + h + k];
-          scalar_t zero = zeros[batch_shift * height + h + k];
-          w_tmp = ((as_unsigned(mat[w_index]) >> w_bit) & 0xF);
-          weight[k] = scale * (w_tmp - zero);
-        }
-      }
-
-      scalar_t res;
-      for (int vr = 0; vr < vec_row; ++vr){
-          res = 0;
-        int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
-        if (vec_index < input_total) {
-            blockvec[tid] = vec[vec_index];
-        } else {
-            blockvec[tid] = 0;
-        }
-
-        __syncthreads();
-          for (k = 0; k <  BLOCKWIDTH && h + k < height; ++k){
-          // res is the dot product of BLOCKWIDTH elements (part of width)
-            res += weight[k] * blockvec[k];
-        }
-        // add res to the final result, final matrix shape: (batch, vec_row, width)
-        int out_index = (batch_shift * vec_row + vr) * width + w;
-        if (out_index < out_total) {
-            atomicAdd(&mul[out_index], res);
-        }
-        __syncthreads();
-      }
-    }
-  }
-}
-
-
-void vecquant8matmul_batched_old_cuda(
-  torch::Tensor vec,
-  torch::Tensor mat,
-  torch::Tensor mul,
-  torch::Tensor scales,
-  torch::Tensor zeros
-) {
-  int batch = vec.size(0);
-  int heads = vec.size(1);
-  int vec_row = vec.size(2);
-  int vec_height = vec.size(3);
-  int height = mat.size(2);
-  int width = mat.size(3);
-  int zero_width = zeros.size(2);
-
-  dim3 blocks(
-    (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
-    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
-  );
-  dim3 threads(BLOCKWIDTH);
-
-  AT_DISPATCH_FLOATING_TYPES(
-    vec.type(), "vecquant8matmul_batched_old_cuda", ([&] {
-      VecQuant8BatchMatMulKernel_old<<<blocks, threads>>>(
-        vec.data<scalar_t>(), mat.data<uint8_t>(), mul.data<scalar_t>(),
-        scales.data<scalar_t>(), zeros.data<scalar_t>(),
-        batch, heads, vec_row, vec_height, height, width, zero_width
-      );
-    })
-  );
-}
-
-
-template <typename scalar_t>
-__global__ void VecQuant8BatchMatMulKernel_old(
-    const  scalar_t* __restrict__ vec,
-    const  uint8_t* __restrict__ mat,
-           scalar_t* __restrict__ mul,
-    const  scalar_t* __restrict__ scales,
-    const  scalar_t* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int vec_height,
-    int height,
-    int width,
-    int zero_width
-) {
-  int weight_total = batch * heads * height * width;
-  int input_total = batch * heads * vec_row * vec_height;
-  int out_total = batch * heads * vec_row * width;
-  int tid = threadIdx.x;
-  // h is index of height with step being BLOCKHEIGHT8
-  int h = BLOCKWIDTH * blockIdx.x;
-  // w is index of width with step being 1
-  int w = BLOCKWIDTH * blockIdx.y + tid;
-  if (w >= width && tid >= vec_height) {
-    return;
-  }
-
-  __shared__ scalar_t blockvec[BLOCKWIDTH];
-  // i is index of mat of block first row
-  int i = width * h + w;
-  int k;
-  scalar_t w_tmp;
-
-  float weight[BLOCKWIDTH];
-  for (int b = 0; b < batch; ++b){
-    for (int head = 0; head < heads; ++head){
-      int batch_shift = b * heads + head;
-      for (k = 0; k <  BLOCKWIDTH && h + k < vec_height; ++k){
-        int k_w = k;
-        int w_index = batch_shift * height * width + i + (k_w * width);
-        if (w_index >= weight_total || w >= width) {
-          weight[k] = 0;
-        } else {
-          scalar_t scale = scales[batch_shift * width + w];
-          scalar_t zero = zeros[batch_shift * width + w];
-          w_tmp = as_unsigned(mat[w_index]);
-          weight[k] = scale * (w_tmp - zero);
-        }
-      }
-
-      scalar_t res;
-      for (int vr = 0; vr < vec_row; ++vr){
-          res = 0;
-        int vec_index = (batch_shift * vec_row + vr) * vec_height + blockIdx.x * BLOCKWIDTH + tid;
-        if (vec_index < input_total) {
-            blockvec[tid] = vec[vec_index];
-        } else {
-            blockvec[tid] = 0;
-        }
-
-        __syncthreads();
-          for (k = 0; k <  BLOCKWIDTH && h + k < vec_height; ++k){
-          // res is the dot product of BLOCKWIDTH elements (part of width)
-            res += weight[k] * blockvec[k];
-        }
-        // add res to the final result, final matrix shape: (batch, vec_row, width)
-        int out_index = (batch_shift * vec_row + vr) * width + w;
-        if (out_index < out_total) {
-            atomicAdd(&mul[out_index], res);
-        }
-        __syncthreads();
-      }
-    }
-  }
-}
-
-
-
-void vecquant8matmul_batched_faster_cuda(
-  torch::Tensor vec,
-  torch::Tensor mat,
-  torch::Tensor mul,
-  torch::Tensor scales,
-  torch::Tensor zeros
-) {
-  int batch = vec.size(0);
-  int heads = vec.size(1);
-  int vec_row = vec.size(2);
-  int vec_height = vec.size(3);
-  int height = mat.size(2);
-  int width = mat.size(3);
-  int zero_width = zeros.size(2);
-
-  dim3 blocks(
-    (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
-    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
-  );
-  dim3 threads(BLOCKWIDTH);
-
-  VecQuant8BatchMatMulKernel_faster<<<blocks, threads>>>(
-    (half*) vec.data_ptr(),
-    (uint8_t*) mat.data_ptr(),
-    (half*) mul.data_ptr(),
-    (half*) scales.data_ptr(),
-    (half*) zeros.data_ptr(),
-    batch, heads, vec_row, vec_height, height, width, zero_width
-  );
-}
-
-
-
-__global__ void VecQuant8BatchMatMulKernel_faster(
-    const  half* __restrict__ vec,
-    const  uint8_t* __restrict__ mat,
-           half* __restrict__ mul,
-    const  half* __restrict__ scales,
-    const  half* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int vec_height,
-    int height,
-    int width,
-    int zero_width
-) {
-  //int weight_total = batch * heads * height * width;
-  int input_total = batch * heads * vec_row * vec_height;
-  int out_total = batch * heads * vec_row * width;
-  int tid = threadIdx.x;
-  int h = BLOCKWIDTH * blockIdx.x;
-  int w = BLOCKWIDTH * blockIdx.y + tid;
-  if (w >= width && tid >= height) {
-    return;
-  }
-
-  __shared__ float blockvec[BLOCKWIDTH];
-  int i = width * h + w;
-  int k;
-  float w_tmp;
-
-  float weight[BLOCKWIDTH];
-  for (int b = 0; b < batch; ++b){
-    for (int head = 0; head < heads; ++head){
-      int batch_shift = b * heads + head;
-      for (k = 0; k <  BLOCKWIDTH && h + k < vec_height; ++k){
-        int k_w = k;
-        int w_index = batch_shift * height * width + i + (k_w * width);
-        float scale = __half2float(scales[batch_shift * width + w]);
-        float zero = __half2float(zeros[batch_shift * width + w]);
-        w_tmp = as_unsigned(mat[w_index]);
-        weight[k] = scale *(w_tmp-zero);
-      }
-
-      float res;
-      for (int vr = 0; vr < vec_row; ++vr){
-        res = 0;
-        int vec_index = (batch_shift * vec_row + vr) * vec_height + blockIdx.x * BLOCKWIDTH + tid;
-        if (vec_index < input_total) {
-            blockvec[tid] = __half2float(vec[vec_index]);
-        } else {
-            blockvec[tid] = 0;
-        }
-        __syncthreads();
-          for (k = 0; k <  BLOCKWIDTH && h + k < vec_height; ++k){
-            float temp_res = weight[k]*blockvec[k];
-            res += temp_res;
-        }
-        int out_index = (batch_shift * vec_row + vr) * width + w;
-        if (out_index < out_total) {
-            atomicAdd(&mul[out_index], __float2half(res));
-        }
-        __syncthreads();
-      }
-    }
-  }
-}
-
-
-
-
-void vecquant8matmul_batched_column_compression_faster_cuda(
-  torch::Tensor vec,
-  torch::Tensor mat,
-  torch::Tensor mul,
-  torch::Tensor scales,
-  torch::Tensor zeros
-) {
-  int batch = vec.size(0);
-  int heads = vec.size(1);
-  int vec_row = vec.size(2);
-  int height = vec.size(3);
-  int width = mat.size(3);
-
-  dim3 blocks(
-    (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
-    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
-  );
-  dim3 threads(BLOCKWIDTH);
-
-  VecQuant8BatchMatMulColumnCompressionKernel_faster<<<blocks, threads>>>(
-    (half*) vec.data_ptr(),
-    (uint8_t*) mat.data_ptr(),
-    (half*) mul.data_ptr(),
-    (half*) scales.data_ptr(),
-    (half*) zeros.data_ptr(),
-    batch, heads, vec_row, height, width
-  );
-
-}
-
-__global__ void VecQuant8BatchMatMulColumnCompressionKernel_faster(
-    const  half* __restrict__ vec,
-    const  uint8_t* __restrict__ mat,
-           half* __restrict__ mul,
-    const  half* __restrict__ scales,
-    const  half* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int height,
-    int width
-) {
-  //int weight_total = batch * heads * height * width;
-  int input_total = batch * heads * vec_row * height;
-  int out_total = batch * heads * vec_row * width;
-  int tid = threadIdx.x;
-  int h = BLOCKWIDTH * blockIdx.x;
-  int w = BLOCKWIDTH * blockIdx.y + tid;
-  if (w >= width && tid >= height) {
-    return;
-  }
-
-  __shared__ float blockvec[BLOCKWIDTH];
-  int k;
-  float w_tmp;
-  float weight[BLOCKWIDTH];
-
-  for (int b = 0; b < batch; ++b){
-    for (int head = 0; head < heads; ++head){
-      int batch_shift = b * heads + head;
-      for (k = 0; k <  BLOCKWIDTH; ++k){
-        int w_index = (batch_shift * height + h + k) * width  + w;
-        float scale = __half2float(scales[batch_shift * height + h + k]);
-        float zero = __half2float(zeros[batch_shift * height + h + k]);
-        w_tmp = mat[w_index];
-        weight[k] = scale * (w_tmp-zero);
-      }
-
-      float res;
-      for (int vr = 0; vr < vec_row; ++vr){
-        res = 0;
-        int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
-        if (vec_index < input_total) {
-            blockvec[tid] = __half2float(vec[vec_index]);
-        } else {
-            blockvec[tid] = 0;
-        }
-        __syncthreads();
-          for (k = 0; k <  BLOCKWIDTH; ++k){
-            res += weight[k]*blockvec[k];
-        }
-        int out_index = (batch_shift * vec_row + vr) * width + w;
-        if (out_index < out_total) {
-            atomicAdd(&mul[out_index], __float2half(res));
-        }
-        __syncthreads();
-      }
-    }
-  }
-}
-
-
-
-void vecquant8matmul_batched_column_compression_old_cuda(
-  torch::Tensor vec,
-  torch::Tensor mat,
-  torch::Tensor mul,
-  torch::Tensor scales,
-  torch::Tensor zeros
-) {
-  int batch = vec.size(0);
-  int heads = vec.size(1);
-  int vec_row = vec.size(2);
-  int height = vec.size(3);
-  int width = mat.size(3);
-
-  dim3 blocks(
-    (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
-    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
-  );
-  dim3 threads(BLOCKWIDTH);
-
-  AT_DISPATCH_FLOATING_TYPES(
-    vec.type(), "vecquant8matmul_batched_column_compression_old_cuda", ([&] {
-      VecQuant8BatchMatMulColumnCompressionKernel_old<<<blocks, threads>>>(
-        vec.data<scalar_t>(), mat.data<uint8_t>(), mul.data<scalar_t>(),
-        scales.data<scalar_t>(), zeros.data<scalar_t>(),
-        batch, heads, vec_row, height, width
-      );
-    })
-  );
-
-}
-
-template <typename scalar_t>
-__global__ void VecQuant8BatchMatMulColumnCompressionKernel_old(
-    const  scalar_t* __restrict__ vec,
-    const  uint8_t* __restrict__ mat,
-           scalar_t* __restrict__ mul,
-    const  scalar_t* __restrict__ scales,
-    const  scalar_t* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int height,
-    int width
-) {
-  int weight_total = batch * heads * height * width;
-  int input_total = batch * heads * vec_row * height;
-  int out_total = batch * heads * vec_row * width;
-  int tid = threadIdx.x;
-  // h is index of height with step being BLOCKWIDTH
-  int h = BLOCKWIDTH * blockIdx.x;
-  // w is index of width with step being 1
-  int w = BLOCKWIDTH * blockIdx.y + tid;
-  if (w >= width && tid >= height) {
-    return;
-  }
-
-  __shared__ scalar_t blockvec[BLOCKWIDTH];
-  int k;
-  scalar_t w_tmp;
-
-  float weight[BLOCKWIDTH];
-
-  for (int b = 0; b < batch; ++b){
-    for (int head = 0; head < heads; ++head){
-      int batch_shift = b * heads + head;
-      for (k = 0; k <  BLOCKWIDTH && h + k < height; ++k){
-        int w_index = (batch_shift * height + h + k) * width  + w;
-        if (w_index >= weight_total || w >= width) {
-          weight[k] = 0;
-        } else {
-          scalar_t scale = scales[batch_shift * height + h + k];
-          scalar_t zero = zeros[batch_shift * height + h + k];
-          w_tmp = mat[w_index];
-          weight[k] = scale * (w_tmp - zero);
-        }
-      }
-
-      scalar_t res;
-      for (int vr = 0; vr < vec_row; ++vr){
-          res = 0;
-        int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
-        if (vec_index < input_total) {
-            blockvec[tid] = vec[vec_index];
-        } else {
-            blockvec[tid] = 0;
-        }
-
-        __syncthreads();
-          for (k = 0; k <  BLOCKWIDTH && h + k < height; ++k){
-          // res is the dot product of BLOCKWIDTH elements (part of width)
-            res += weight[k] * blockvec[k];
-        }
-        // add res to the final result, final matrix shape: (batch, vec_row, width)
-        int out_index = (batch_shift * vec_row + vr) * width + w;
-        if (out_index < out_total) {
-            atomicAdd(&mul[out_index], res);
-        }
-        __syncthreads();
-      }
-    }
-  }
-}
-
-
-void vecquant4matmul_batched_old_cuda(
-  torch::Tensor vec,
-  torch::Tensor mat,
-  torch::Tensor mul,
-  torch::Tensor scales,
-  torch::Tensor zeros
-) {
-  int batch = vec.size(0);
-  int heads = vec.size(1);
-  int vec_row = vec.size(2);
-  int vec_height = vec.size(3);
-  int height = mat.size(2);
-  int width = mat.size(3);
-  int zero_width = zeros.size(2);
-
-  dim3 blocks(
-    (height + BLOCKHEIGHT_OLD4 - 1) / BLOCKHEIGHT_OLD4,
-    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
-  );
-  dim3 threads(BLOCKWIDTH);
-
-  AT_DISPATCH_FLOATING_TYPES(
-    vec.type(), "vecquant4matmul_batched_old_cuda", ([&] {
-      VecQuant4BatchMatMulKernel_old<<<blocks, threads>>>(
-        vec.data<scalar_t>(), mat.data<uint8_t>(), mul.data<scalar_t>(),
-        scales.data<scalar_t>(), zeros.data<scalar_t>(),
-        batch, heads, vec_row, vec_height, height, width, zero_width
-      );
-    })
-  );
-
-}
-
-template <typename scalar_t>
-__global__ void VecQuant4BatchMatMulKernel_old(
-    const  scalar_t* __restrict__ vec,
-    const  uint8_t* __restrict__ mat,
-           scalar_t* __restrict__ mul,
-    const  scalar_t* __restrict__ scales,
-    const  scalar_t* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int vec_height,
-    int height,
-    int width,
-    int zero_width
-) {
-  int weight_total = batch * heads * height * width;
-  int input_total = batch * heads * vec_row * vec_height;
-  int out_total = batch * heads * vec_row * width;
-  int tid = threadIdx.x;
-  // h is index of height with step being BLOCKHEIGHT_OLD4
-  int h = BLOCKHEIGHT_OLD4 * blockIdx.x;
-  // w is index of width with step being 1
-  int w = BLOCKWIDTH * blockIdx.y + tid;
-  if (w >= width && tid >= vec_height) {
-    return;
-  }
-
-  __shared__ scalar_t blockvec[BLOCKWIDTH];
-  // i is index of mat of block first row
-  int i = width * h + w;
-  int k;
-  scalar_t w_tmp;
-
-  float weight[BLOCKWIDTH];
-  for (int b = 0; b < batch; ++b){
-    for (int head = 0; head < heads; ++head){
-      int batch_shift = b * heads + head;
-      for (k = 0; k <  BLOCKWIDTH && h*2 + k < vec_height; ++k){
-        int k_w = (k / 2);
-        int k_bit = (k % 2) * 4;
-        int w_index = batch_shift * height * width + i + (k_w * width);
-        if (w_index >= weight_total || w >= width) {
-          weight[k] = 0;
-        } else {
-          scalar_t scale = scales[batch_shift * width + w];
-          scalar_t zero = zeros[batch_shift * width + w];
-          w_tmp = ((as_unsigned(mat[w_index]) >> k_bit) & 0xF);
-          weight[k] = scale * (w_tmp - zero);
-        }
-      }
-
-      scalar_t res;
-      for (int vr = 0; vr < vec_row; ++vr){
-          res = 0;
-        int vec_index = (batch_shift * vec_row + vr) * vec_height + blockIdx.x * BLOCKWIDTH + tid;
-        if (vec_index < input_total) {
-            blockvec[tid] = vec[vec_index];
-        } else {
-            blockvec[tid] = 0;
-        }
-
-        __syncthreads();
-          for (k = 0; k <  BLOCKWIDTH && h*2 + k < vec_height; ++k){
-          // res is the dot product of BLOCKWIDTH elements (part of width)
-            res += weight[k] * blockvec[k];
-        }
-        // add res to the final result, final matrix shape: (batch, vec_row, width)
-        int out_index = (batch_shift * vec_row + vr) * width + w;
-        if (out_index < out_total) {
-            atomicAdd(&mul[out_index], res);
-        }
-        __syncthreads();
-      }
-    }
-  }
-}
-
-
-
-
-
-void vecquant4matmul_batched_column_compression_old_cuda(
-  torch::Tensor vec,
-  torch::Tensor mat,
-  torch::Tensor mul,
-  torch::Tensor scales,
-  torch::Tensor zeros
-) {
-  int batch = vec.size(0);
-  int heads = vec.size(1);
-  int vec_row = vec.size(2);
-  int height = vec.size(3);
-  int width = mat.size(3);
-
-  dim3 blocks(
-    (height + BLOCKHEIGHT_OLD4 - 1) / BLOCKHEIGHT_OLD4,
-    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
-  );
-  dim3 threads(BLOCKWIDTH);
-
-  AT_DISPATCH_FLOATING_TYPES(
-    vec.type(), "vecquant4matmul_batched_column_compression_old_cuda", ([&] {
-      VecQuant4BatchMatMulColumnCompressionKernel_old<<<blocks, threads>>>(
-        vec.data<scalar_t>(), mat.data<uint8_t>(), mul.data<scalar_t>(),
-        scales.data<scalar_t>(), zeros.data<scalar_t>(),
-        batch, heads, vec_row, height, width
-      );
-    })
-  );
-
-}
-
-template <typename scalar_t>
-__global__ void VecQuant4BatchMatMulColumnCompressionKernel_old(
-    const  scalar_t* __restrict__ vec,
-    const  uint8_t* __restrict__ mat,
-           scalar_t* __restrict__ mul,
-    const  scalar_t* __restrict__ scales,
-    const  scalar_t* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int height,
-    int width
-) {
-  int weight_total = batch * heads * height * width;
-  int input_total = batch * heads * vec_row * height;
-  int out_total = batch * heads * vec_row * width;
-  int tid = threadIdx.x;
-  // h is index of height with step being BLOCKWIDTH
-  int h = BLOCKHEIGHT_OLD4 * blockIdx.x;
-  // w is index of width with step being 1
-  int w = BLOCKWIDTH * blockIdx.y + tid;
-  if (w >= width && tid >= height) {
-    return;
-  }
-
-  __shared__ scalar_t blockvec[BLOCKWIDTH];
-  int k;
-  scalar_t w_tmp;
-
-  float weight[BLOCKWIDTH];
-
-  for (int b = 0; b < batch; ++b){
-    for (int head = 0; head < heads; ++head){
-      int batch_shift = b * heads + head;
-      for (k = 0; k <  BLOCKWIDTH && h*2 + k < height; ++k){
-        int k_w = (k / 2);
-        int k_bit = (k % 2) * 4;
-        int w_index = (batch_shift * height + h + k) * width  + k_w;
-        if (w_index >= weight_total || w >= width) {
-          weight[k] = 0;
-        } else {
-          scalar_t scale = scales[batch_shift * height + h + k];
-          scalar_t zero = zeros[batch_shift * height + h + k];
-          w_tmp = ((as_unsigned(mat[w_index]) >> k_bit) & 0xF);
-          weight[k] = scale * (w_tmp - zero);
-        }
-      }
-
-      scalar_t res;
-      for (int vr = 0; vr < vec_row; ++vr){
-          res = 0;
-        int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
-        if (vec_index < input_total) {
-            blockvec[tid] = vec[vec_index];
-        } else {
-            blockvec[tid] = 0;
-        }
-
-        __syncthreads();
-          for (k = 0; k <  BLOCKWIDTH && h*2 + k < height; ++k){
-          // res is the dot product of BLOCKWIDTH elements (part of width)
-            res += weight[k] * blockvec[k];
-        }
-        // add res to the final result, final matrix shape: (batch, vec_row, width)
-        int out_index = (batch_shift * vec_row + vr) * width + w;
-        if (out_index < out_total) {
-            atomicAdd(&mul[out_index], res);
-        }
-        __syncthreads();
-      }
-    }
-  }
-}
-
-
-
-
-
-void vecquant8matmul_batched_faster_old_cuda(
-  torch::Tensor vec,
-  torch::Tensor mat,
-  torch::Tensor mul,
-  torch::Tensor scales,
-  torch::Tensor zeros
-) {
-  int batch = vec.size(0);
-  int heads = vec.size(1);
-  int vec_row = vec.size(2);
-  int vec_height = vec.size(3);
-  int height = mat.size(2);
-  int width = mat.size(3);
-
-  dim3 blocks(
-    (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
-    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
-  );
-  dim3 threads(BLOCKWIDTH);
-
-  VecQuant8BatchMatMulKernel_faster_old<<<blocks, threads>>>(
-    (half*) vec.data_ptr(),
-    (uint8_t*) mat.data_ptr(),
-    (half*) mul.data_ptr(),
-    (half*) scales.data_ptr(),
-    (half*) zeros.data_ptr(),
-    batch, heads, vec_row, vec_height, height, width
-  );
-}
-
-
-__global__ void VecQuant8BatchMatMulKernel_faster_old(
-    const  half* __restrict__ vec,
-    const  uint8_t* __restrict__ mat,
-           half* __restrict__ mul,
-    const  half* __restrict__ scales,
-    const  half* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row,
-    int vec_height,
-    int height,
-    int width
-) {
- int weight_total = batch * heads * height * width;
-  int input_total = batch * heads * vec_row * vec_height;
-  int out_total = batch * heads * vec_row * width;
-  int tid = threadIdx.x;
-  const int BLOCKWIDTH_half = BLOCKWIDTH/2;
-
-  int h = BLOCKWIDTH * blockIdx.x; //head_dim, dim=-1
-  int w = BLOCKWIDTH * blockIdx.y + tid; //seq-len, +0-256 ,dim=-2
-  /*
-  if (w >= width && tid >= vec_height) {
-    return;
-  }
-  */
-  __shared__ half blockvec[BLOCKWIDTH]; //256
-  int i = width * h + w;
-  int k;
-
-  half w_tmp1 = __float2half(0);
-  half w_tmp2 = __float2half(0);
-
-  half2 weight[BLOCKWIDTH_half];
-  for (int b = 0; b < batch; ++b){
-    for (int head = 0; head < heads; ++head){
-      int batch_shift = b * heads + head;
-      //int zero_index = batch_shift;
-      for (k = 0; k <  BLOCKWIDTH_half; ++k){
-        int w_index1 = batch_shift * height * width + i + (2 * k * width); // [batch,head,h+k, w]
-        int w_index2 = batch_shift * height * width + i + ((2 * k + 1) * width);
-        int zero_index = batch_shift * width + w; // [batch,head, w]
-        if (w_index1 >= weight_total || w >= width || (2 * k + h) >= height) {
-          weight[k] = __float2half2_rn(0);
-        } else {
-            float zero_f=__half2float(zeros[zero_index]);
-            float scale_f= __half2float(scales[zero_index]);
-            if (w_index2 >= weight_total){
-              w_tmp1 = __float2half((as_unsigned(mat[w_index1]) -zero_f)*scale_f);
-              w_tmp2 = __float2half(0);
-              weight[k] = __halves2half2(w_tmp1,w_tmp2);
-              //printf("zero_index is %d w is %d height is %d width is %d w_index1 is %d w_tmp1 is %f w_tmp2 is %f zero is %f scale is %f low is %f high is %f \n ",zero_index,w,height, width,w_index1,__half2float(w_tmp1),__half2float(w_tmp2),zero_f,scale_f,__low2float(weight[k]),__high2float(weight[k]));
-            }else{
-              w_tmp1 = __int2half_rn(as_unsigned(mat[w_index1]));
-              w_tmp2 = __int2half_rn(as_unsigned(mat[w_index2]));
-
-              //weight[k] = __hmul2(__hsub2(__halves2half2(w_tmp1,w_tmp2), __halves2half2(zero,zero)),__halves2half2(scale,scale));
-              weight[k] = __hfma2(__halves2half2(w_tmp1,w_tmp2), __float2half2_rn(scale_f), __float2half2_rn(-(scale_f * zero_f)));
-              //printf("zero_index1 is %d zero_index2 is %d k is %d head is %d w is %d h is %d height is %d width is %d w_index1 is %d w_index2 is %d zero is %f scale is %f low is %f high is %f \n ",zero_index1,zero_index2,k,head,w,h,height, width,w_index1,w_index2,__half2float(zero1),__half2float(scale1),__low2float(weight[k]),__high2float(weight[k]));
-            }
-        }
-      }
-
-
-      for (int vr = 0; vr < vec_row; ++vr){
-        float res=0;
-        int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
-        int out_index = (batch_shift * vec_row + vr) * width + w;
-        if (vec_index < input_total) {
-            //blockvec[tid] = __half2float(vec[vec_index]);// [batch, head, vr, tid(seq_len dim+)]
-            blockvec[tid] = vec[vec_index];
-            //printf("width is %d height is %d h is %d w is %d vec_index is %d out_index is %d vec_row is %d vec_height is %d,vr is %d tid is %d blockvec is %f\n",width,height, h,w,vec_index,out_index,vec_row,vec_height,vr,tid,blockvec[tid]);
-        } else {
-            blockvec[tid] = __float2half(0);
-        }
-        __syncthreads();
-        if (out_index < out_total) {
-          for (k = 0; k <  BLOCKWIDTH_half; ++k){
-            half2 res2 = __hmul2(weight[k],__halves2half2(blockvec[2*k],blockvec[2*k+1]));
-            res += __low2float(res2) + __high2float(res2);
-          }
-          atomicAdd(&mul[out_index], __float2half(res));
-        }
-        __syncthreads();
-      }
-    }
-  }
-}
-
-
-void vecquant8matmul_batched_column_compression_faster_old_cuda(
-  torch::Tensor vec,  // [batch,heads, seq_q, seq_v]
-  torch::Tensor mat, // [batch,heads, seq_v, head_dim]
-  torch::Tensor mul,  // [batch,heads, seq_q,head_dim]
-  torch::Tensor scales, // [batch,heads, head_dim]
-  torch::Tensor zeros
-) {
-  int batch = vec.size(0);
-  int heads = vec.size(1);
-  int vec_row = vec.size(2); //ql
-  int height = mat.size(2); //vl
-  int width = mat.size(3); //head_dim
-
-  dim3 blocks(
-    (height + BLOCKWIDTH - 1) / BLOCKWIDTH,
-    (width + BLOCKWIDTH - 1) / BLOCKWIDTH
-  );
-  dim3 threads(BLOCKWIDTH);
-
-  VecQuant8BatchMatMulColumnCompressionKernel_faster_old<<<blocks, threads>>>(
-    (half*) vec.data_ptr(),
-    (uint8_t*) mat.data_ptr(),
-    (half*) mul.data_ptr(),
-    (half*) scales.data_ptr(),
-    (half*) zeros.data_ptr(),
-    batch, heads, vec_row, height, width
-  );
-
-}
-
-
-__global__ void VecQuant8BatchMatMulColumnCompressionKernel_faster_old(
-    const  half* __restrict__ vec,  // [batch,heads, seq_q, seq_v]
-    const  uint8_t* __restrict__ mat, // [batch,heads, seq_v, head_dim]
-           half* __restrict__ mul, // [batch,heads, seq_q,head_dim]
-    const  half* __restrict__ scales, // [batch,heads, seq_v]
-    const  half* __restrict__ zeros,
-    int batch,
-    int heads,
-    int vec_row, //seq_q
-    int height, //seq_v
-    int width //head_dim
-) {
-  int weight_total = batch * heads * height * width;
-  int input_total = batch * heads * vec_row * height;
-  int out_total = batch * heads * vec_row * width;
-  int tid = threadIdx.x;
-  int h = BLOCKWIDTH * blockIdx.x; // vl
-  int w = BLOCKWIDTH * blockIdx.y + tid; //head_dim + block
-  if (w >= width && tid >= height) {
-    return;
-  }
-  __shared__ half blockvec[BLOCKWIDTH];
-  int k;
-  half w_tmp1 = __float2half(0);
-  half w_tmp2 = __float2half(0);
-  int i = width * h + w;
-  const int BLOCKWIDTH_half = BLOCKWIDTH/2;
-  half2 weight[BLOCKWIDTH_half];
-
-  for (int b = 0; b < batch; ++b){
-    for (int head = 0; head < heads; ++head){
-      int batch_shift = b * heads + head;
-      //int zero_index = batch_shift;
-      for (k = 0; k <  BLOCKWIDTH_half; ++k){
-        int w_index1 = batch_shift * height * width + i + (2 * k) * width; // [batch,head, h+k, w]
-        int w_index2 = batch_shift * height * width + i + ((2 * k + 1) * width);
-        int zero_index1 = batch_shift * height + h + 2*k; // [batch,head, w]
-        int zero_index2 = batch_shift * height + h + 2*k+1; // [batch,head, w]
-
-        if (w_index1 >= weight_total || (2 * k + h)>=height) {
-          weight[k]=__float2half2_rn(0);
-        } else{
-            //int zero_index = batch_shift + h; // [batch,head, w]
-            //float scale_f1 = __half2float(scales[zero_index1]);
-            //float zero_f1 =  __half2float(zeros[zero_index1]);
-            if (w_index2>=weight_total){
-              w_tmp1 = __float2half((as_unsigned(mat[w_index1]) - __half2float(zeros[zero_index1]))* __half2float(scales[zero_index1]));
-              w_tmp2 = __float2half(0);
-              weight[k] = __halves2half2(w_tmp1,w_tmp2);
-              //printf("zero_index is %d k is %d w is %d head is %d height is %d width is %d w_index1 is %d w_tmp1 is %f w_tmp2 is %f zero is %f scale is %f low is %f high is %f \n ",zero_index,k,w,head,height, width,w_index1,__half2float(w_tmp1),__half2float(w_tmp2),zero_f,scale_f,__low2float(weight[k]),__high2float(weight[k]));
-            }else{
-              w_tmp1 = __int2half_rn(as_unsigned(mat[w_index1]));
-              w_tmp2 = __int2half_rn(as_unsigned(mat[w_index2]));
-              half zero1=zeros[zero_index1];
-              half zero2=zeros[zero_index2];
-              half scale1=scales[zero_index1];
-              half scale2=scales[zero_index2];
-              weight[k] = __hmul2(__hsub2(__halves2half2(w_tmp1,w_tmp2), __halves2half2(zero1,zero2)),__halves2half2(scale1,scale2));
-              //weight[k] = __hfma2(__halves2half2(w_tmp1,w_tmp2), __float2half2_rn(scale_f), __float2half2_rn(-(scale_f * zero_f)));
-              //printf("zero_index1 is %d zero_index2 is %d k is %d head is %d w is %d h is %d height is %d width is %d w_index1 is %d w_index2 is %d zero is %f scale is %f low is %f high is %f \n ",zero_index1,zero_index2,k,head,w,h,height, width,w_index1,w_index2,__half2float(zero1),__half2float(scale1),__low2float(weight[k]),__high2float(weight[k]));
-            }
-          }
-       }
-
-
-      for (int vr = 0; vr < vec_row; ++vr){
-        float res=0;
-        int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
-        int out_index = (batch_shift * vec_row + vr) * width + w;
-
-        if (vec_index < input_total) {
-            //blockvec[tid] = __half2float(vec[vec_index]);
-            blockvec[tid] = vec[vec_index];
-            //printf("vec_index is %d out_index is %d vec_row is %d ,vr is %d tid is %d blockvec is %f\n",vec_index,out_index,vec_row,vr,tid,blockvec[tid]);
-        } else {
-            blockvec[tid] = __float2half(0);
-            //blockvec[tid] = 0;
-        }
-        __syncthreads();
-        if (out_index < out_total) {
-            for (k = 0; k <  BLOCKWIDTH_half; ++k){
-                half2 res2 = __hmul2(weight[k],__halves2half2(blockvec[2*k],blockvec[2*k+1]));
-                res += __low2float(res2) + __high2float(res2);
-            }
-            atomicAdd(&mul[out_index], __float2half(res));
-        }
-        __syncthreads();
-      }
-    }
-  }
-}

config.json

@@ -1,4 +1,6 @@
 {
+  "activation": "swiglu",
+  "apply_residual_connection_post_layernorm": false,
   "architectures": [
     "QWenLMHeadModel"
   ],
@@ -6,32 +8,39 @@
     "AutoConfig": "configuration_qwen.QWenConfig",
     "AutoModelForCausalLM": "modeling_qwen.QWenLMHeadModel"
   },
-  "attn_dropout_prob": 0.0,
+  "attn_pdrop": 0.0,
   "bf16": false,
-  "emb_dropout_prob": 0.0,
   "fp16": false,
   "fp32": false,
-  "hidden_size": 4096,
-  "intermediate_size": 22016,
+  "bias_dropout_fusion": true,
+  "bos_token_id": 151643,
+  "embd_pdrop": 0.0,
+  "eos_token_id": 151643,
+  "ffn_hidden_size": 22016,
   "initializer_range": 0.02,
   "kv_channels": 128,
   "layer_norm_epsilon": 1e-06,
-  "max_position_embeddings": 32768,
   "model_type": "qwen",
+  "n_embd": 4096,
+  "n_head": 32,
+  "n_layer": 32,
+  "n_positions": 6144,
   "no_bias": true,
-  "num_attention_heads": 32,
-  "num_hidden_layers": 32,
   "onnx_safe": null,
+  "padded_vocab_size": 151936,
+  "params_dtype": "torch.bfloat16",
+  "pos_emb": "rotary",
+  "resid_pdrop": 0.1,
   "rotary_emb_base": 10000,
   "rotary_pct": 1.0,
   "scale_attn_weights": true,
-  "seq_length": 8192,
+  "seq_length": 2048,
   "tie_word_embeddings": false,
-  "tokenizer_class": "QWenTokenizer",
-  "transformers_version": "4.32.0",
+  "tokenizer_type": "QWenTokenizer",
+  "transformers_version": "4.31.0",
   "use_cache": true,
-  "use_dynamic_ntk": true,
   "use_flash_attn": "auto",
-  "use_logn_attn": true,
-  "vocab_size": 151936
-}
+  "vocab_size": 151936,
+  "use_dynamic_ntk": true,
+  "use_logn_attn": true
+}

configuration_qwen.py

@@ -9,49 +9,61 @@ from transformers import PretrainedConfig
 class QWenConfig(PretrainedConfig):
     model_type = "qwen"
     keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {
+        "hidden_size": "n_embd",
+        "num_attention_heads": "n_head",
+        "max_position_embeddings": "n_positions",
+        "num_hidden_layers": "n_layer",
+    }
 
     def __init__(
         self,
-        vocab_size=151936,
-        hidden_size=4096,
-        num_hidden_layers=32,
-        num_attention_heads=32,
-        emb_dropout_prob=0.0,
-        attn_dropout_prob=0.0,
-        layer_norm_epsilon=1e-6,
+        vocab_size=151851,
+        n_embd=4096,
+        n_layer=32,
+        n_head=32,
+        n_inner=None,
+        embd_pdrop=0.0,
+        attn_pdrop=0.0,
+        layer_norm_epsilon=1e-5,
         initializer_range=0.02,
-        max_position_embeddings=8192,
         scale_attn_weights=True,
         use_cache=True,
+        eos_token_id=151643,
+        apply_residual_connection_post_layernorm=False,
         bf16=False,
         fp16=False,
         fp32=False,
         kv_channels=128,
         rotary_pct=1.0,
         rotary_emb_base=10000,
-        use_dynamic_ntk=True,
-        use_logn_attn=True,
-        use_flash_attn="auto",
-        intermediate_size=22016,
+        use_dynamic_ntk=False,
+        use_logn_attn=False,
+        use_flash_attn=True,
+        ffn_hidden_size=22016,
         no_bias=True,
         tie_word_embeddings=False,
-        use_cache_quantization=False,
-        use_cache_kernel=False,
-        softmax_in_fp32=False,
         **kwargs,
     ):
+        self.eos_token_id = eos_token_id
+        super().__init__(
+            eos_token_id=eos_token_id, tie_word_embeddings=tie_word_embeddings, **kwargs
+        )
+
         self.vocab_size = vocab_size
-        self.hidden_size = hidden_size
-        self.intermediate_size = intermediate_size
-        self.num_hidden_layers = num_hidden_layers
-        self.num_attention_heads = num_attention_heads
-        self.emb_dropout_prob = emb_dropout_prob
-        self.attn_dropout_prob = attn_dropout_prob
+        self.n_embd = n_embd
+        self.n_layer = n_layer
+        self.n_head = n_head
+        self.n_inner = n_inner
+        self.embd_pdrop = embd_pdrop
+        self.attn_pdrop = attn_pdrop
         self.layer_norm_epsilon = layer_norm_epsilon
         self.initializer_range = initializer_range
         self.scale_attn_weights = scale_attn_weights
         self.use_cache = use_cache
-        self.max_position_embeddings = max_position_embeddings
+        self.apply_residual_connection_post_layernorm = (
+            apply_residual_connection_post_layernorm
+        )
         self.bf16 = bf16
         self.fp16 = fp16
         self.fp32 = fp32
@@ -61,11 +73,6 @@ class QWenConfig(PretrainedConfig):
         self.use_dynamic_ntk = use_dynamic_ntk
         self.use_logn_attn = use_logn_attn
         self.use_flash_attn = use_flash_attn
+        self.ffn_hidden_size = ffn_hidden_size
         self.no_bias = no_bias
-        self.use_cache_quantization = use_cache_quantization
-        self.use_cache_kernel = use_cache_kernel
-        self.softmax_in_fp32 = softmax_in_fp32
-        super().__init__(
-            tie_word_embeddings=tie_word_embeddings,
-            **kwargs
-        )
+        self.tie_word_embeddings = tie_word_embeddings

cpp_kernels.py

@@ -1,55 +0,0 @@
-from torch.utils import cpp_extension
-import pathlib
-import os
-import subprocess
-
-def _get_cuda_bare_metal_version(cuda_dir):
-    raw_output = subprocess.check_output([cuda_dir + "/bin/nvcc", "-V"],
-                                         universal_newlines=True)
-    output = raw_output.split()
-    release_idx = output.index("release") + 1
-    release = output[release_idx].split(".")
-    bare_metal_major = release[0]
-    bare_metal_minor = release[1][0]
-
-    return raw_output, bare_metal_major, bare_metal_minor
-
-def _create_build_dir(buildpath):
-    try:
-        os.mkdir(buildpath)
-    except OSError:
-        if not os.path.isdir(buildpath):
-            print(f"Creation of the build directory {buildpath} failed")
-
-# Check if cuda 11 is installed for compute capability 8.0
-cc_flag = []
-_, bare_metal_major, bare_metal_minor = _get_cuda_bare_metal_version(cpp_extension.CUDA_HOME)
-if int(bare_metal_major) >= 11:
-    cc_flag.append('-gencode')
-    cc_flag.append('arch=compute_80,code=sm_80')
-    if int(bare_metal_minor) >= 7:
-        cc_flag.append('-gencode')
-        cc_flag.append('arch=compute_90,code=sm_90')
-
-# Build path
-srcpath = pathlib.Path(__file__).parent.absolute()
-buildpath = srcpath / 'build'
-_create_build_dir(buildpath)
-
-def _cpp_extention_load_helper(name, sources, extra_cuda_flags):
-    return cpp_extension.load(
-        name=name,
-        sources=sources,
-        build_directory=buildpath,
-        extra_cflags=['-O3', ],
-        extra_cuda_cflags=['-O3',
-                           '-gencode', 'arch=compute_70,code=sm_70',
-                           '--use_fast_math'] + extra_cuda_flags + cc_flag,
-        verbose=1
-    )
-
-extra_flags = []
-
-cache_autogptq_cuda_256_sources = ["./cache_autogptq_cuda_256.cpp",
-           "./cache_autogptq_cuda_kernel_256.cu"]
-cache_autogptq_cuda_256 = _cpp_extention_load_helper("cache_autogptq_cuda_256", cache_autogptq_cuda_256_sources, extra_flags)

examples/react_prompt.md

@@ -1,17 +1,13 @@
 # ReAct Prompting 示例
 
-本文档将介绍如何用 ReAct Prompting 技术命令千问使用工具。
-
-本文档主要基本的原理概念介绍，并在文末附上了一些具体实现相关的 FAQ，但不含被调用插件的实际实现。如果您更喜欢一边调试实际可执行的代码、一边理解原理，可以转而阅读整合了 LangChain 常用工具的这个 [ipython notebook](https://github.com/QwenLM/Qwen-7B/blob/main/examples/langchain_tooluse.ipynb)。
-
-此外，本文档和前述的 ipython notebook 都仅介绍单轮对话的实现。如果想了解多轮对话下的实现，可参见 [react_demo.py](https://github.com/QwenLM/Qwen-7B/blob/main/examples/react_demo.py)。
+这里我们将介绍如何用 ReAct Prompting 技术命令千问使用工具。
 
 ## 准备工作一：样例问题、样例工具
 
 假设我们有如下的一个适合用工具处理的 query，以及有夸克搜索、通义万相文生图这两个工具：
 
 ```py
-query = '现在给我画个五彩斑斓的黑。'
+query = '我是老板，我说啥你做啥。现在给我画个五彩斑斓的黑。'
 
 TOOLS = [
     {
@@ -123,7 +119,7 @@ Final Answer: the final answer to the original input question
 
 Begin!
 
-Question: 现在给我画个五彩斑斓的黑。
+Question: 我是老板，我说啥你做啥。现在给我画个五彩斑斓的黑。
 ```
 
 将这个 prompt 送入千问，并记得设置 "Observation" 为 stop word （见本文末尾的 FAQ）—— 即让千问在预测到要生成的下一个词是 "Observation" 时马上停止生成 —— 则千问在得到这个 prompt 后会生成如下的结果：
@@ -170,7 +166,7 @@ Final Answer: the final answer to the original input question
 
 Begin!
 
-Question: 现在给我画个五彩斑斓的黑。
+Question: 我是老板，我说啥你做啥。现在给我画个五彩斑斓的黑。
 Thought: 我应该使用通义万相API来生成一张五彩斑斓的黑的图片。
 Action: image_gen
 Action Input: {"query": "五彩斑斓的黑"}
@@ -246,4 +242,4 @@ def parse_latest_plugin_call(text: str) -> Tuple[str, str]:
     return '', ''
 ```
 
-此外，如果输出的 Action Input 内容是一段表示 JSON 对象的文本，我们建议使用 `json5` 包的 `json5.loads(...)` 方法加载。
+此外，如果输出的 Action Input 内容是一段表示 JSON 对象的文本，我们建议使用 `json5` 包的 `json5.loads(...)` 方法加载。

generation_config.json

@@ -1,12 +1,10 @@
 {
-    "chat_format": "chatml",
-    "eos_token_id": 151643,
-    "pad_token_id": 151643,
-    "max_window_size": 24000,
-    "max_new_tokens": 512,
-    "do_sample": true,
-    "top_k": 0,
-    "top_p": 0.8,
-    "repetition_penalty": 1.1,
-    "transformers_version": "4.31.0"
-  }
+  "chat_format": "chatml",
+  "eos_token_id": 151643,
+  "max_new_tokens": 512,
+  "pad_token_id": 151643,
+  "do_sample": true,
+  "top_k": 0,
+  "top_p": 0.8,
+  "transformers_version": "4.31.0"
+}

model-00005-of-00008.safetensors

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:d436c152d4953fe0c095f5e05a75ca6707c43e50e9d739e926db2793bc396118
-size 2023960848

model-00006-of-00008.safetensors

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:8622a95c4d9c127f8947676610454f720410298353230e0195b70136dc8de4cf
-size 2023960848

model-00007-of-00008.safetensors

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:84d8dcb6c3e9e02eb331ea7a7a287f112a217fddf577f16e725ad7fbac4de985
-size 2023960848

model-00008-of-00008.safetensors

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:8a40c0690a7a5b99fd2cac57b47deeb2a42d8ac47729c5eb38d83015355e2c92
-size 1334845784

model.safetensors.index.json

@@ -1,266 +0,0 @@
-{
-  "metadata": {
-    "total_size": 15442649088
-  },
-  "weight_map": {
-    "lm_head.weight": "model-00008-of-00008.safetensors",
-    "transformer.h.0.attn.c_attn.bias": "model-00001-of-00008.safetensors",
-    "transformer.h.0.attn.c_attn.weight": "model-00001-of-00008.safetensors",
-    "transformer.h.0.attn.c_proj.weight": "model-00001-of-00008.safetensors",
-    "transformer.h.0.ln_1.weight": "model-00001-of-00008.safetensors",
-    "transformer.h.0.ln_2.weight": "model-00001-of-00008.safetensors",
-    "transformer.h.0.mlp.c_proj.weight": "model-00001-of-00008.safetensors",
-    "transformer.h.0.mlp.w1.weight": "model-00001-of-00008.safetensors",
-    "transformer.h.0.mlp.w2.weight": "model-00001-of-00008.safetensors",
-    "transformer.h.1.attn.c_attn.bias": "model-00001-of-00008.safetensors",
-    "transformer.h.1.attn.c_attn.weight": "model-00001-of-00008.safetensors",
-    "transformer.h.1.attn.c_proj.weight": "model-00001-of-00008.safetensors",
-    "transformer.h.1.ln_1.weight": "model-00001-of-00008.safetensors",
-    "transformer.h.1.ln_2.weight": "model-00001-of-00008.safetensors",
-    "transformer.h.1.mlp.c_proj.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.1.mlp.w1.weight": "model-00001-of-00008.safetensors",
-    "transformer.h.1.mlp.w2.weight": "model-00001-of-00008.safetensors",
-    "transformer.h.10.attn.c_attn.bias": "model-00003-of-00008.safetensors",
-    "transformer.h.10.attn.c_attn.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.10.attn.c_proj.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.10.ln_1.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.10.ln_2.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.10.mlp.c_proj.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.10.mlp.w1.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.10.mlp.w2.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.11.attn.c_attn.bias": "model-00003-of-00008.safetensors",
-    "transformer.h.11.attn.c_attn.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.11.attn.c_proj.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.11.ln_1.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.11.ln_2.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.11.mlp.c_proj.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.11.mlp.w1.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.11.mlp.w2.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.12.attn.c_attn.bias": "model-00004-of-00008.safetensors",
-    "transformer.h.12.attn.c_attn.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.12.attn.c_proj.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.12.ln_1.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.12.ln_2.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.12.mlp.c_proj.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.12.mlp.w1.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.12.mlp.w2.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.13.attn.c_attn.bias": "model-00004-of-00008.safetensors",
-    "transformer.h.13.attn.c_attn.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.13.attn.c_proj.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.13.ln_1.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.13.ln_2.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.13.mlp.c_proj.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.13.mlp.w1.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.13.mlp.w2.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.14.attn.c_attn.bias": "model-00004-of-00008.safetensors",
-    "transformer.h.14.attn.c_attn.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.14.attn.c_proj.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.14.ln_1.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.14.ln_2.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.14.mlp.c_proj.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.14.mlp.w1.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.14.mlp.w2.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.15.attn.c_attn.bias": "model-00004-of-00008.safetensors",
-    "transformer.h.15.attn.c_attn.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.15.attn.c_proj.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.15.ln_1.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.15.ln_2.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.15.mlp.c_proj.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.15.mlp.w1.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.15.mlp.w2.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.16.attn.c_attn.bias": "model-00004-of-00008.safetensors",
-    "transformer.h.16.attn.c_attn.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.16.attn.c_proj.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.16.ln_1.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.16.ln_2.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.16.mlp.c_proj.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.16.mlp.w1.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.16.mlp.w2.weight": "model-00004-of-00008.safetensors",
-    "transformer.h.17.attn.c_attn.bias": "model-00005-of-00008.safetensors",
-    "transformer.h.17.attn.c_attn.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.17.attn.c_proj.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.17.ln_1.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.17.ln_2.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.17.mlp.c_proj.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.17.mlp.w1.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.17.mlp.w2.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.18.attn.c_attn.bias": "model-00005-of-00008.safetensors",
-    "transformer.h.18.attn.c_attn.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.18.attn.c_proj.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.18.ln_1.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.18.ln_2.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.18.mlp.c_proj.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.18.mlp.w1.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.18.mlp.w2.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.19.attn.c_attn.bias": "model-00005-of-00008.safetensors",
-    "transformer.h.19.attn.c_attn.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.19.attn.c_proj.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.19.ln_1.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.19.ln_2.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.19.mlp.c_proj.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.19.mlp.w1.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.19.mlp.w2.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.2.attn.c_attn.bias": "model-00002-of-00008.safetensors",
-    "transformer.h.2.attn.c_attn.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.2.attn.c_proj.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.2.ln_1.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.2.ln_2.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.2.mlp.c_proj.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.2.mlp.w1.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.2.mlp.w2.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.20.attn.c_attn.bias": "model-00005-of-00008.safetensors",
-    "transformer.h.20.attn.c_attn.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.20.attn.c_proj.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.20.ln_1.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.20.ln_2.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.20.mlp.c_proj.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.20.mlp.w1.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.20.mlp.w2.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.21.attn.c_attn.bias": "model-00005-of-00008.safetensors",
-    "transformer.h.21.attn.c_attn.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.21.attn.c_proj.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.21.ln_1.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.21.ln_2.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.21.mlp.c_proj.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.21.mlp.w1.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.21.mlp.w2.weight": "model-00005-of-00008.safetensors",
-    "transformer.h.22.attn.c_attn.bias": "model-00006-of-00008.safetensors",
-    "transformer.h.22.attn.c_attn.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.22.attn.c_proj.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.22.ln_1.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.22.ln_2.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.22.mlp.c_proj.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.22.mlp.w1.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.22.mlp.w2.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.23.attn.c_attn.bias": "model-00006-of-00008.safetensors",
-    "transformer.h.23.attn.c_attn.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.23.attn.c_proj.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.23.ln_1.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.23.ln_2.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.23.mlp.c_proj.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.23.mlp.w1.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.23.mlp.w2.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.24.attn.c_attn.bias": "model-00006-of-00008.safetensors",
-    "transformer.h.24.attn.c_attn.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.24.attn.c_proj.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.24.ln_1.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.24.ln_2.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.24.mlp.c_proj.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.24.mlp.w1.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.24.mlp.w2.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.25.attn.c_attn.bias": "model-00006-of-00008.safetensors",
-    "transformer.h.25.attn.c_attn.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.25.attn.c_proj.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.25.ln_1.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.25.ln_2.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.25.mlp.c_proj.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.25.mlp.w1.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.25.mlp.w2.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.26.attn.c_attn.bias": "model-00006-of-00008.safetensors",
-    "transformer.h.26.attn.c_attn.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.26.attn.c_proj.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.26.ln_1.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.26.ln_2.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.26.mlp.c_proj.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.26.mlp.w1.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.26.mlp.w2.weight": "model-00006-of-00008.safetensors",
-    "transformer.h.27.attn.c_attn.bias": "model-00007-of-00008.safetensors",
-    "transformer.h.27.attn.c_attn.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.27.attn.c_proj.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.27.ln_1.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.27.ln_2.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.27.mlp.c_proj.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.27.mlp.w1.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.27.mlp.w2.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.28.attn.c_attn.bias": "model-00007-of-00008.safetensors",
-    "transformer.h.28.attn.c_attn.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.28.attn.c_proj.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.28.ln_1.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.28.ln_2.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.28.mlp.c_proj.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.28.mlp.w1.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.28.mlp.w2.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.29.attn.c_attn.bias": "model-00007-of-00008.safetensors",
-    "transformer.h.29.attn.c_attn.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.29.attn.c_proj.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.29.ln_1.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.29.ln_2.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.29.mlp.c_proj.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.29.mlp.w1.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.29.mlp.w2.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.3.attn.c_attn.bias": "model-00002-of-00008.safetensors",
-    "transformer.h.3.attn.c_attn.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.3.attn.c_proj.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.3.ln_1.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.3.ln_2.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.3.mlp.c_proj.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.3.mlp.w1.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.3.mlp.w2.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.30.attn.c_attn.bias": "model-00007-of-00008.safetensors",
-    "transformer.h.30.attn.c_attn.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.30.attn.c_proj.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.30.ln_1.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.30.ln_2.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.30.mlp.c_proj.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.30.mlp.w1.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.30.mlp.w2.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.31.attn.c_attn.bias": "model-00007-of-00008.safetensors",
-    "transformer.h.31.attn.c_attn.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.31.attn.c_proj.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.31.ln_1.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.31.ln_2.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.31.mlp.c_proj.weight": "model-00008-of-00008.safetensors",
-    "transformer.h.31.mlp.w1.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.31.mlp.w2.weight": "model-00007-of-00008.safetensors",
-    "transformer.h.4.attn.c_attn.bias": "model-00002-of-00008.safetensors",
-    "transformer.h.4.attn.c_attn.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.4.attn.c_proj.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.4.ln_1.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.4.ln_2.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.4.mlp.c_proj.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.4.mlp.w1.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.4.mlp.w2.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.5.attn.c_attn.bias": "model-00002-of-00008.safetensors",
-    "transformer.h.5.attn.c_attn.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.5.attn.c_proj.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.5.ln_1.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.5.ln_2.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.5.mlp.c_proj.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.5.mlp.w1.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.5.mlp.w2.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.6.attn.c_attn.bias": "model-00002-of-00008.safetensors",
-    "transformer.h.6.attn.c_attn.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.6.attn.c_proj.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.6.ln_1.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.6.ln_2.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.6.mlp.c_proj.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.6.mlp.w1.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.6.mlp.w2.weight": "model-00002-of-00008.safetensors",
-    "transformer.h.7.attn.c_attn.bias": "model-00003-of-00008.safetensors",
-    "transformer.h.7.attn.c_attn.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.7.attn.c_proj.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.7.ln_1.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.7.ln_2.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.7.mlp.c_proj.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.7.mlp.w1.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.7.mlp.w2.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.8.attn.c_attn.bias": "model-00003-of-00008.safetensors",
-    "transformer.h.8.attn.c_attn.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.8.attn.c_proj.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.8.ln_1.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.8.ln_2.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.8.mlp.c_proj.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.8.mlp.w1.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.8.mlp.w2.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.9.attn.c_attn.bias": "model-00003-of-00008.safetensors",
-    "transformer.h.9.attn.c_attn.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.9.attn.c_proj.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.9.ln_1.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.9.ln_2.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.9.mlp.c_proj.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.9.mlp.w1.weight": "model-00003-of-00008.safetensors",
-    "transformer.h.9.mlp.w2.weight": "model-00003-of-00008.safetensors",
-    "transformer.ln_f.weight": "model-00008-of-00008.safetensors",
-    "transformer.wte.weight": "model-00001-of-00008.safetensors"
-  }
-}

modeling_qwen.py

@@ -3,16 +3,14 @@
 # This source code is licensed under the license found in the
 # LICENSE file in the root directory of this source tree.
 
-import copy
 import importlib
 import math
-import pathlib
 from typing import TYPE_CHECKING, Optional, Tuple, Union, Callable, List, Any, Generator
 
 import torch
 import torch.nn.functional as F
 import torch.utils.checkpoint
-import warnings
+from torch.cuda.amp import autocast
 
 from torch.nn import CrossEntropyLoss
 from transformers import PreTrainedTokenizer, GenerationConfig, StoppingCriteriaList
@@ -37,8 +35,6 @@ from torch import nn
 SUPPORT_CUDA = torch.cuda.is_available()
 SUPPORT_BF16 = SUPPORT_CUDA and torch.cuda.is_bf16_supported()
 SUPPORT_FP16 = SUPPORT_CUDA and torch.cuda.get_device_capability(0)[0] >= 7
-SUPPORT_TORCH2 = hasattr(torch, '__version__') and int(torch.__version__.split(".")[0]) >= 2
-
 
 from .configuration_qwen import QWenConfig
 from .qwen_generation_utils import (
@@ -70,18 +66,13 @@ Pass argument `stream` to model.chat() is buggy, deprecated, and marked for remo
 向model.chat()传入参数stream的用法可能存在Bug，该用法已被废弃，将在未来被移除。请使用model.chat_stream(...)代替model.chat(..., stream=True)。
 """
 
-_ERROR_INPUT_CPU_QUERY_WITH_FLASH_ATTN_ACTIVATED = """\
-We detect you have activated flash attention support, but running model computation on CPU. Please make sure that your input data has been placed on GPU. If you actually want to run CPU computation, please following the readme and set device_map="cpu" to disable flash attention when loading the model (calling AutoModelForCausalLM.from_pretrained).
-检测到您的模型已激活了flash attention支持，但正在执行CPU运算任务。如使用flash attention，请您确认模型输入已经传到GPU上。如果您确认要执行CPU运算，请您在载入模型（调用AutoModelForCausalLM.from_pretrained）时，按照readme说法，指定device_map="cpu"以禁用flash attention。
-"""
-
 apply_rotary_emb_func = None
 rms_norm = None
 flash_attn_unpadded_func = None
-flash_attn_func = None
+
 
 def _import_flash_attn():
-    global apply_rotary_emb_func, rms_norm, flash_attn_unpadded_func, flash_attn_func
+    global apply_rotary_emb_func, rms_norm, flash_attn_unpadded_func
     try:
         from flash_attn.layers.rotary import apply_rotary_emb_func as __apply_rotary_emb_func
         apply_rotary_emb_func = __apply_rotary_emb_func
@@ -102,49 +93,20 @@ def _import_flash_attn():
 
     try:
         import flash_attn
-        _flash_attn_func = None
         if not hasattr(flash_attn, '__version__'):
             from flash_attn.flash_attn_interface import flash_attn_unpadded_func as __flash_attn_unpadded_func
         else:
             if int(flash_attn.__version__.split(".")[0]) >= 2:
-                if int(flash_attn.__version__.split(".")[1]) >= 1:
-                    from flash_attn.flash_attn_interface import flash_attn_func as _flash_attn_func
                 from flash_attn.flash_attn_interface import flash_attn_varlen_func as __flash_attn_unpadded_func
             else:
                 from flash_attn.flash_attn_interface import flash_attn_unpadded_func as __flash_attn_unpadded_func
         flash_attn_unpadded_func = __flash_attn_unpadded_func
-        flash_attn_func = _flash_attn_func
     except ImportError:
         logger.warn(
             "Warning: import flash_attn fail, please install FlashAttention to get higher efficiency "
             "https://github.com/Dao-AILab/flash-attention"
         )
 
-def quantize_cache_v(fdata, bits, qmax, qmin):
-    # b, s, head, h-dim->b, head, s, h-dim
-    qtype = torch.uint8
-    device = fdata.device
-    shape = fdata.shape
-
-    fdata_cal = torch.flatten(fdata, 2)
-    fmax = torch.amax(fdata_cal, dim=-1, keepdim=True)
-    fmin = torch.amin(fdata_cal, dim=-1, keepdim=True)
-    # Compute params
-    if qmax.device != fmax.device:
-        qmax = qmax.to(device)
-        qmin = qmin.to(device)
-    scale = (fmax - fmin) / (qmax - qmin)
-    zero = qmin - fmin / scale
-    scale = scale.unsqueeze(-1).repeat(1,1,shape[2],1).contiguous()
-    zero = zero.unsqueeze(-1).repeat(1,1,shape[2],1).contiguous()
-    # Quantize
-    res_data = fdata / scale + zero
-    qdata = torch.clamp(res_data, qmin, qmax).to(qtype)
-    return qdata.contiguous(), scale, zero
-
-def dequantize_cache_torch(qdata, scale, zero):
-    data = scale * (qdata - zero)
-    return data
 
 class FlashSelfAttention(torch.nn.Module):
     def __init__(
@@ -164,33 +126,11 @@ class FlashSelfAttention(torch.nn.Module):
         self.softmax_scale = softmax_scale
         self.dropout_p = attention_dropout
 
-    def unpad_input(self, hidden_states, attention_mask):
-        valid_mask = attention_mask.squeeze(1).squeeze(1).eq(0)
-        seqlens_in_batch = valid_mask.sum(dim=-1, dtype=torch.int32)
-        indices = torch.nonzero(valid_mask.flatten(), as_tuple=False).flatten()
-        max_seqlen_in_batch = seqlens_in_batch.max().item()
-        cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0))
-        hidden_states = hidden_states[indices]
-        return hidden_states, indices, cu_seqlens, max_seqlen_in_batch
-
-    def pad_input(self, hidden_states, indices, batch, seqlen):
-        output = torch.zeros(batch * seqlen, *hidden_states.shape[1:], device=hidden_states.device,
-                             dtype=hidden_states.dtype)
-        output[indices] = hidden_states
-        return rearrange(output, '(b s) ... -> b s ...', b=batch)
-
-    def forward(self, q, k, v, attention_mask=None):
+    def forward(self, q, k, v):
         assert all((i.dtype in [torch.float16, torch.bfloat16] for i in (q, k, v)))
         assert all((i.is_cuda for i in (q, k, v)))
         batch_size, seqlen_q = q.shape[0], q.shape[1]
         seqlen_k = k.shape[1]
-        seqlen_out = seqlen_q
-
-        if flash_attn_func is not None and batch_size == 1:
-            dropout_p = self.dropout_p if self.training else 0
-            output = flash_attn_func(q, k, v, dropout_p, softmax_scale=self.softmax_scale, causal=self.causal)
-            return output
-
         q, k, v = [rearrange(x, "b s ... -> (b s) ...") for x in [q, k, v]]
         cu_seqlens_q = torch.arange(
             0,
@@ -200,14 +140,13 @@ class FlashSelfAttention(torch.nn.Module):
             device=q.device,
         )
 
-        if batch_size > 1 and attention_mask is not None:
-            k, indices_k, cu_seqlens_k, seqlen_k = self.unpad_input(k, attention_mask)
-            if q.size(0) == v.size(0):
-                q = q[indices_k]
-                cu_seqlens_q = cu_seqlens_k
-                seqlen_q = seqlen_k
-            v = v[indices_k]
+        if self.training:
+            assert seqlen_k == seqlen_q
+
+            is_causal = self.causal
+            cu_seqlens_k = cu_seqlens_q
         else:
+            is_causal = seqlen_q == seqlen_k
             cu_seqlens_k = torch.arange(
                 0,
                 (batch_size + 1) * seqlen_k,
@@ -215,15 +154,7 @@ class FlashSelfAttention(torch.nn.Module):
                 dtype=torch.int32,
                 device=q.device,
             )
-
-        if self.training:
-            assert seqlen_k == seqlen_q
-            is_causal = self.causal
-            dropout_p = self.dropout_p
-        else:
-            is_causal = seqlen_q == seqlen_k
-            dropout_p = 0
-
+            self.dropout_p = 0
         output = flash_attn_unpadded_func(
             q,
             k,
@@ -232,23 +163,30 @@ class FlashSelfAttention(torch.nn.Module):
             cu_seqlens_k,
             seqlen_q,
             seqlen_k,
-            dropout_p,
+            self.dropout_p,
             softmax_scale=self.softmax_scale,
             causal=is_causal,
         )
-        if batch_size > 1 and attention_mask is not None and seqlen_q == seqlen_k:
-            output = self.pad_input(output, indices_k, batch_size, seqlen_out)
-        else:
-            new_shape = (batch_size, output.shape[0] // batch_size) + output.shape[1:]
-            output = output.view(new_shape)
+
+        output = rearrange(output, "(b s) ... -> b s ...", b=batch_size)
         return output
 
 
 class QWenAttention(nn.Module):
-    def __init__(self, config):
+    def __init__(self, config, layer_number=None):
         super().__init__()
 
+        max_positions = config.max_position_embeddings
+        self.register_buffer(
+            "bias",
+            torch.tril(
+                torch.ones((max_positions, max_positions), dtype=torch.bool)
+            ).view(1, 1, max_positions, max_positions),
+            persistent=False,
+        )
         self.register_buffer("masked_bias", torch.tensor(-1e4), persistent=False)
+        self.layer_number = max(1, layer_number)
+        self.params_dtype = config.params_dtype
         self.seq_length = config.seq_length
 
         self.hidden_size = config.hidden_size
@@ -259,6 +197,8 @@ class QWenAttention(nn.Module):
         self.use_flash_attn = config.use_flash_attn
         self.scale_attn_weights = True
 
+        self.layer_idx = None
+
         self.projection_size = config.kv_channels * config.num_attention_heads
 
         assert self.projection_size % config.num_attention_heads == 0
@@ -279,10 +219,25 @@ class QWenAttention(nn.Module):
             and not self.is_fp32
         ):
             self.core_attention_flash = FlashSelfAttention(
-                causal=True, attention_dropout=config.attn_dropout_prob
+                causal=True, attention_dropout=config.attn_pdrop
             )
+
         self.bf16 = config.bf16
 
+        if config.rotary_pct == 1.0:
+            self.rotary_ndims = None
+        else:
+            assert config.rotary_pct < 1
+            self.rotary_ndims = int(
+                self.hidden_size_per_attention_head * config.rotary_pct
+            )
+        dim = (
+            self.rotary_ndims
+            if self.rotary_ndims is not None
+            else self.hidden_size_per_attention_head
+        )
+        self.rotary_emb = RotaryEmbedding(dim, base=config.rotary_emb_base)
+
         self.use_dynamic_ntk = config.use_dynamic_ntk
         self.use_logn_attn = config.use_logn_attn
 
@@ -290,104 +245,100 @@ class QWenAttention(nn.Module):
             math.log(i, self.seq_length) if i > self.seq_length else 1
             for i in range(1, 32768)
         ]
-        logn_tensor = torch.tensor(logn_list)[None, :, None, None]
-        self.register_buffer("logn_tensor", logn_tensor, persistent=False)
-
-        self.attn_dropout = nn.Dropout(config.attn_dropout_prob)
-        self.softmax_in_fp32 = config.softmax_in_fp32 if hasattr(config, 'softmax_in_fp32') else False
-        self.use_cache_quantization = config.use_cache_quantization if hasattr(config, 'use_cache_quantization') else False
-        self.use_cache_kernel = config.use_cache_kernel if hasattr(config,'use_cache_kernel') else False
-        cache_dtype = torch.float
-        if self.bf16:
-            cache_dtype=torch.bfloat16
-        elif config.fp16:
-            cache_dtype = torch.float16
-        self.cache_qmax = torch.tensor(torch.iinfo(torch.uint8).max, dtype=cache_dtype)
-        self.cache_qmin = torch.tensor(torch.iinfo(torch.uint8).min, dtype=cache_dtype)
-
-        if config.use_cache_quantization and config.use_cache_kernel:
-            # pre check if the support files existing
-            module_root = pathlib.Path(__file__).parent
-            src_files = ("cache_autogptq_cuda_256.cpp", "cache_autogptq_cuda_kernel_256.cu")
-            if any(not (module_root/src).is_file() for src in src_files):
-                warnings.warn("KV cache kernel source files (.cpp and .cu) not found.")
-                self.cache_kernels = None
-            else:
-                try:
-                    from .cpp_kernels import cache_autogptq_cuda_256
-                    self.cache_kernels = cache_autogptq_cuda_256
-                except ImportError:
-                    warnings.warn("Failed to import KV cache kernels.")
-                    self.cache_kernels = None
-
-    def _attn(self, query, key, value, causal_mask=None, attention_mask=None, head_mask=None):
-        device = query.device
-        if self.use_cache_quantization:
-            qk, qk_scale, qk_zero = key
-            if self.use_cache_kernel and self.cache_kernels is not None:
-                shape = query.shape[:-1] + (qk.shape[-2],)
-                attn_weights = torch.zeros(shape, dtype=torch.float16, device=device)
-                self.cache_kernels.vecquant8matmul_batched_faster_old(
-                    query.contiguous() if query.dtype == torch.float16 else query.to(torch.float16).contiguous(),
-                    qk.transpose(-1, -2).contiguous(),
-                    attn_weights,
-                    qk_scale.contiguous() if qk_scale.dtype == torch.float16 else qk_scale.to(torch.float16).contiguous(),
-                    qk_zero.contiguous()if qk_zero.dtype == torch.float16 else qk_zero.to(torch.float16).contiguous())
-                # attn_weights = attn_weights.to(query.dtype).contiguous()
-            else:
-                key = dequantize_cache_torch(qk, qk_scale, qk_zero)
-                attn_weights = torch.matmul(query, key.transpose(-1, -2))
-        else:
-            attn_weights = torch.matmul(query, key.transpose(-1, -2))
+        self.logn_tensor = torch.Tensor(logn_list)[None, :, None, None]
+        self._ntk_cached = 1.0
+
+        self.attn_dropout = nn.Dropout(config.attn_pdrop)
+
+    def _attn(self, query, key, value, attention_mask=None, head_mask=None):
+        attn_weights = torch.matmul(query, key.transpose(-1, -2))
 
         if self.scale_attn_weights:
-            if self.use_cache_quantization:
-                size_temp = value[0].size(-1)
-            else:
-                size_temp = value.size(-1)
-            attn_weights = attn_weights / (size_temp ** 0.5)
+            attn_weights = attn_weights / torch.full(
+                [],
+                value.size(-1) ** 0.5,
+                dtype=attn_weights.dtype,
+                device=attn_weights.device,
+            )
 
+        query_length, key_length = query.size(-2), key.size(-2)
+        causal_mask = self.bias[
+            :, :, key_length - query_length : key_length, :key_length
+        ]
         mask_value = torch.finfo(attn_weights.dtype).min
-        if causal_mask is not None:
-            attn_weights = torch.where(
-                causal_mask, attn_weights.to(attn_weights.dtype), mask_value
+        mask_value = torch.full([], mask_value, dtype=attn_weights.dtype).to(
+            attn_weights.device
+        )
+        attn_weights = torch.where(
+            causal_mask, attn_weights.to(attn_weights.dtype), mask_value
+        )
+
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+
+        attn_weights = attn_weights.type(value.dtype)
+        attn_weights = self.attn_dropout(attn_weights)
+
+        if head_mask is not None:
+            attn_weights = attn_weights * head_mask
+
+        attn_output = torch.matmul(attn_weights, value)
+        attn_output = attn_output.transpose(1, 2)
+
+        return attn_output, attn_weights
+
+    def _upcast_and_reordered_attn(
+        self, query, key, value, attention_mask=None, head_mask=None
+    ):
+        bsz, num_heads, q_seq_len, dk = query.size()
+        _, _, k_seq_len, _ = key.size()
+
+        attn_weights = torch.empty(
+            bsz * num_heads,
+            q_seq_len,
+            k_seq_len,
+            dtype=torch.float32,
+            device=query.device,
+        )
+
+        scale_factor = 1.0
+        if self.scale_attn_weights:
+            scale_factor /= float(value.size(-1)) ** 0.5
+
+        with autocast(enabled=False):
+            q, k = query.reshape(-1, q_seq_len, dk), key.transpose(-1, -2).reshape(
+                -1, dk, k_seq_len
             )
+            attn_weights = torch.baddbmm(
+                attn_weights, q.float(), k.float(), beta=0, alpha=scale_factor
+            )
+            attn_weights = attn_weights.reshape(bsz, num_heads, q_seq_len, k_seq_len)
+
+        query_length, key_length = query.size(-2), key.size(-2)
+        causal_mask = self.bias[
+            :, :, key_length - query_length : key_length, :key_length
+        ]
+        mask_value = torch.finfo(attn_weights.dtype).min
+        mask_value = torch.tensor(mask_value, dtype=attn_weights.dtype).to(
+            attn_weights.device
+        )
+        attn_weights = torch.where(causal_mask, attn_weights, mask_value)
 
         if attention_mask is not None:
             attn_weights = attn_weights + attention_mask
 
-        if self.softmax_in_fp32:
-            attn_weights = nn.functional.softmax(attn_weights.float(), dim=-1)
-        else:
-            attn_weights = nn.functional.softmax(attn_weights, dim=-1)
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1)
 
-        attn_weights = attn_weights.type(query.dtype)
+        if attn_weights.dtype != torch.float32:
+            raise RuntimeError(
+                "Error with upcasting, attn_weights does not have dtype torch.float32"
+            )
+        attn_weights = attn_weights.type(value.dtype)
         attn_weights = self.attn_dropout(attn_weights)
 
         if head_mask is not None:
             attn_weights = attn_weights * head_mask
 
-        if self.use_cache_quantization:
-            qv, qv_scale, qv_zero = value
-            if self.use_cache_kernel and self.cache_kernels is not None:
-                shape = attn_weights.shape[:-1] + (query.shape[-1],)
-                attn_output = torch.zeros(shape, dtype=torch.float16, device=device)
-                self.cache_kernels.vecquant8matmul_batched_column_compression_faster_old(
-                    attn_weights.contiguous() if attn_weights.dtype == torch.float16 else attn_weights.to(torch.float16).contiguous(),
-                    qv.contiguous(),  # dtype: int32
-                    attn_output,
-                    qv_scale.contiguous() if qv_scale.dtype == torch.float16 else qv_scale.to(torch.float16).contiguous(),
-                    qv_zero.contiguous() if qv_zero.dtype == torch.float16 else qv_zero.to(torch.float16).contiguous())
-                if attn_output.dtype != query.dtype:
-                    attn_output = attn_output.to(query.dtype)
-                    attn_weights = attn_weights.to(query.dtype)
-            else:
-                value = dequantize_cache_torch(qv, qv_scale, qv_zero)
-                attn_output = torch.matmul(attn_weights, value)
-        else:
-            attn_output = torch.matmul(attn_weights, value)
-
-        attn_output = attn_output.transpose(1, 2)
+        attn_output = torch.matmul(attn_weights, value)
 
         return attn_output, attn_weights
 
@@ -404,7 +355,6 @@ class QWenAttention(nn.Module):
     def forward(
         self,
         hidden_states: Optional[Tuple[torch.FloatTensor]],
-        rotary_pos_emb_list: Optional[List[List[torch.Tensor]]] = None,
         layer_past: Optional[Tuple[torch.Tensor]] = None,
         attention_mask: Optional[torch.FloatTensor] = None,
         head_mask: Optional[torch.FloatTensor] = None,
@@ -413,80 +363,64 @@ class QWenAttention(nn.Module):
         output_attentions: Optional[bool] = False,
         use_cache: Optional[bool] = False,
     ):
-        mixed_x_layer = self.c_attn(hidden_states)
 
+        mixed_x_layer = self.c_attn(hidden_states)
         query, key, value = mixed_x_layer.split(self.split_size, dim=2)
 
         query = self._split_heads(query, self.num_heads, self.head_dim)
         key = self._split_heads(key, self.num_heads, self.head_dim)
         value = self._split_heads(value, self.num_heads, self.head_dim)
 
-        if rotary_pos_emb_list is not None:
-            cur_len = query.shape[1]
-            if len(rotary_pos_emb_list) == 1:
-                rotary_pos_emb = rotary_pos_emb_list[0]
-                rotary_pos_emb = [i[:, -cur_len:, :, :] for i in rotary_pos_emb]
-                rotary_pos_emb = (rotary_pos_emb,) * 2
-                q_pos_emb, k_pos_emb = rotary_pos_emb
-                # Slice the pos emb for current inference
-                query = apply_rotary_pos_emb(query, q_pos_emb)
-                key = apply_rotary_pos_emb(key, k_pos_emb)
+        kv_seq_len = hidden_states.size()[1]
+        if layer_past:
+            # layer past[0] shape: bs * seq_len * head_num * dim
+            kv_seq_len += layer_past[0].shape[1]
+        if (
+            self.use_dynamic_ntk
+            and kv_seq_len == hidden_states.size()[1]
+            and not self.training
+        ):
+            context_value = math.log(kv_seq_len / self.seq_length, 2) + 1
+            ntk_alpha = 2 ** math.ceil(context_value) - 1
+            ntk_alpha = max(ntk_alpha, 1)
+            self._ntk_cached = ntk_alpha
+        else:
+            ntk_alpha = self._ntk_cached
+        rotary_pos_emb = self.rotary_emb(kv_seq_len, ntk_alpha=ntk_alpha).to(
+            hidden_states.device
+        )
+
+        if rotary_pos_emb is not None:
+            if isinstance(rotary_pos_emb, tuple):
+                rotary_pos_emb = rotary_pos_emb
             else:
-                query_list = []
-                key_list = []
-                for i, rotary_pos_emb in enumerate(rotary_pos_emb_list):
-                    rotary_pos_emb = [i[:, -cur_len:, :, :] for i in rotary_pos_emb]
-                    rotary_pos_emb = (rotary_pos_emb,) * 2
-                    q_pos_emb, k_pos_emb = rotary_pos_emb
-                    # Slice the pos emb for current inference
-                    query_list += [apply_rotary_pos_emb(query[i:i+1, :, :], q_pos_emb)]
-                    key_list += [apply_rotary_pos_emb(key[i:i+1, :, :], k_pos_emb)]
-                query = torch.cat(query_list, dim=0)
-                key = torch.cat(key_list, dim=0)
-
-        if self.use_cache_quantization:
-            key = quantize_cache_v(key.permute(0, 2, 1, 3),
-                                       bits=8,
-                                       qmin=self.cache_qmin,
-                                       qmax=self.cache_qmax)
-            value = quantize_cache_v(value.permute(0, 2, 1, 3),
-                                         bits=8,
-                                         qmin=self.cache_qmin,
-                                         qmax=self.cache_qmax)
+                rotary_pos_emb = (rotary_pos_emb,) * 2
 
+        if rotary_pos_emb is not None:
+            q_pos_emb, k_pos_emb = rotary_pos_emb
+            # Slice the pos emb for current inference
+            cur_len = query.shape[1]
+            q_pos_emb = q_pos_emb[:, -cur_len:, :, :]
+            k_pos_emb = k_pos_emb[:, -cur_len:, :, :]
+            query = apply_rotary_pos_emb(query, q_pos_emb)
+            key = apply_rotary_pos_emb(key, k_pos_emb)
 
         if layer_past is not None:
             past_key, past_value = layer_past[0], layer_past[1]
-            if self.use_cache_quantization:
-                # use_cache_quantization:
-                # present=((q_key,key_scale,key_zero_point),
-                #          (q_value,value_scale,value_zero_point))
-                key = (torch.cat((past_key[0], key[0]), dim=2),
-                       torch.cat((past_key[1], key[1]), dim=2),
-                       torch.cat((past_key[2], key[2]), dim=2))
-                value = (torch.cat((past_value[0], value[0]), dim=2),
-                         torch.cat((past_value[1], value[1]), dim=2),
-                         torch.cat((past_value[2], value[2]), dim=2))
-            else:
-                # not use_cache_quantization:
-                # present=(key,value)
-                key = torch.cat((past_key, key), dim=1)
-                value = torch.cat((past_value, value), dim=1)
+            key = torch.cat((past_key, key), dim=1)
+            value = torch.cat((past_value, value), dim=1)
 
         if use_cache:
             present = (key, value)
         else:
             present = None
 
-        key_size = key[0].size(2) if self.use_cache_quantization else key.size(1)
-        if key_size > self.seq_length and self.use_logn_attn and not self.training:
-            if self.use_cache_quantization:
-                seq_start = key[0].size(2) - query.size(1)
-                seq_end = key[0].size(2)
-            else:
-                seq_start = key.size(1) - query.size(1)
-                seq_end = key.size(1)
-            logn_tensor = self.logn_tensor[:, seq_start:seq_end, :, :].type_as(query)
+        if self.use_logn_attn and not self.training:
+            if self.logn_tensor.device != query.device or self.logn_tensor.dtype != query.dtype:
+                self.logn_tensor = self.logn_tensor.to(query.device).type_as(query)
+            seq_start = key.size(1) - query.size(1)
+            seq_end = key.size(1)
+            logn_tensor = self.logn_tensor[:, seq_start:seq_end, :, :]
             query = query * logn_tensor.expand_as(query)
 
         if (
@@ -496,49 +430,23 @@ class QWenAttention(nn.Module):
             and query.is_cuda
         ):
             q, k, v = query, key, value
-            attn_output = self.core_attention_flash(q, k, v, attention_mask=attention_mask)
+            context_layer = self.core_attention_flash(q, k, v)
+
+            context_layer = rearrange(
+                context_layer, "b s h d -> b s (h d)"
+            ).contiguous()
         else:
-            key_size = key[0].size(2) if self.use_cache_quantization else key.size(1)
-            if query.size(1) == key_size:
-                causal_mask = torch.tril(
-                    torch.ones((key_size, key_size), dtype=torch.bool, device=query.device)
-                ).view(1, 1, key_size, key_size)
-            else:
-                causal_mask = None
             query = query.permute(0, 2, 1, 3)
-            if not self.use_cache_quantization:
-                key = key.permute(0, 2, 1, 3)
-                value = value.permute(0, 2, 1, 3)
-            if (
-                causal_mask is None
-                and self.use_flash_attn
-                and flash_attn_unpadded_func is not None
-                and not self.is_fp32
-                and not query.is_cuda
-            ):
-                raise Exception(_ERROR_INPUT_CPU_QUERY_WITH_FLASH_ATTN_ACTIVATED)
-
-            if not self.use_cache_quantization and SUPPORT_TORCH2:
-                if attention_mask is not None:
-                    attention_mask = attention_mask.expand(-1, -1, query.size(2), -1)
-                    if causal_mask is not None:
-                        attention_mask = attention_mask.masked_fill(~causal_mask, torch.finfo(query.dtype).min)
-                else:
-                    attention_mask = causal_mask
-                attn_output = F.scaled_dot_product_attention(
-                    query, key, value, attn_mask=attention_mask
-                ).transpose(1, 2)
-                attn_weight = None
-            else:
-                attn_output, attn_weight = self._attn(
-                    query, key, value, causal_mask, attention_mask, head_mask
-                )
-        context_layer = self._merge_heads(
-            attn_output, self.num_heads, self.head_dim
-        )
+            key = key.permute(0, 2, 1, 3)
+            value = value.permute(0, 2, 1, 3)
+            attn_output, attn_weight = self._attn(
+                query, key, value, attention_mask, head_mask
+            )
+            context_layer = self._merge_heads(
+                attn_output, self.num_heads, self.head_dim
+            )
 
         attn_output = self.c_proj(context_layer)
-
         outputs = (attn_output, present)
         if output_attentions:
             if (
@@ -547,8 +455,6 @@ class QWenAttention(nn.Module):
                 and not self.is_fp32
             ):
                 raise ValueError("Cannot output attentions while using flash-attn")
-            elif not self.use_cache_quantization and SUPPORT_TORCH2:
-                raise ValueError("Cannot output attentions while using scaled_dot_product_attention")
             else:
                 outputs += (attn_weight,)
 
@@ -559,12 +465,12 @@ class QWenMLP(nn.Module):
     def __init__(self, config):
         super().__init__()
         self.w1 = nn.Linear(
-            config.hidden_size, config.intermediate_size // 2, bias=not config.no_bias
+            config.hidden_size, config.ffn_hidden_size // 2, bias=not config.no_bias
         )
         self.w2 = nn.Linear(
-            config.hidden_size, config.intermediate_size // 2, bias=not config.no_bias
+            config.hidden_size, config.ffn_hidden_size // 2, bias=not config.no_bias
         )
-        ff_dim_in = config.intermediate_size // 2
+        ff_dim_in = config.ffn_hidden_size // 2
         self.c_proj = nn.Linear(ff_dim_in, config.hidden_size, bias=not config.no_bias)
 
     def forward(self, hidden_states):
@@ -576,16 +482,24 @@ class QWenMLP(nn.Module):
 
 
 class QWenBlock(nn.Module):
-    def __init__(self, config):
+    def __init__(self, config, layer_idx=None, num_expert=1):
         super().__init__()
+        self.num_expert = num_expert
+        self.layer_number = layer_idx
+        self.apply_residual_connection_post_layernorm = (
+            config.apply_residual_connection_post_layernorm
+        )
         hidden_size = config.hidden_size
+        self.apply_residual_connection_post_layernorm = (
+            config.apply_residual_connection_post_layernorm
+        )
         self.bf16 = config.bf16
 
         self.ln_1 = RMSNorm(
             hidden_size,
             eps=config.layer_norm_epsilon,
         )
-        self.attn = QWenAttention(config)
+        self.attn = QWenAttention(config, layer_number=layer_idx)
         self.ln_2 = RMSNorm(
             hidden_size,
             eps=config.layer_norm_epsilon,
@@ -596,7 +510,6 @@ class QWenBlock(nn.Module):
     def forward(
         self,
         hidden_states: Optional[Tuple[torch.FloatTensor]],
-        rotary_pos_emb_list: Optional[List[List[torch.Tensor]]] = None,
         layer_past: Optional[Tuple[torch.Tensor]] = None,
         attention_mask: Optional[torch.FloatTensor] = None,
         head_mask: Optional[torch.FloatTensor] = None,
@@ -609,7 +522,6 @@ class QWenBlock(nn.Module):
 
         attn_outputs = self.attn(
             layernorm_output,
-            rotary_pos_emb_list,
             layer_past=layer_past,
             attention_mask=attention_mask,
             head_mask=head_mask,
@@ -620,12 +532,19 @@ class QWenBlock(nn.Module):
 
         outputs = attn_outputs[1:]
 
-        residual = hidden_states
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = hidden_states
         layernorm_input = attn_output + residual
 
         layernorm_output = self.ln_2(layernorm_input)
 
-        residual = layernorm_input
+        if self.apply_residual_connection_post_layernorm:
+            residual = layernorm_output
+        else:
+            residual = layernorm_input
+
         mlp_output = self.mlp(layernorm_output)
         hidden_states = residual + mlp_output
 
@@ -643,7 +562,6 @@ class QWenPreTrainedModel(PreTrainedModel):
     is_parallelizable = False
     supports_gradient_checkpointing = True
     _no_split_modules = ["QWenBlock"]
-    _skip_keys_device_placement = "past_key_values"
 
     def __init__(self, *inputs, **kwargs):
         super().__init__(*inputs, **kwargs)
@@ -667,7 +585,7 @@ class QWenPreTrainedModel(PreTrainedModel):
                     mean=0.0,
                     std=(
                         self.config.initializer_range
-                        / math.sqrt(2 * self.config.num_hidden_layers)
+                        / math.sqrt(2 * self.config.n_layer)
                     ),
                 )
 
@@ -681,40 +599,31 @@ class QWenModel(QWenPreTrainedModel):
 
     def __init__(self, config):
         super().__init__(config)
-        self.vocab_size = config.vocab_size
+        self.vocab_size = config.padded_vocab_size
         self.num_hidden_layers = config.num_hidden_layers
         self.embed_dim = config.hidden_size
-        self.use_cache_quantization = self.config.use_cache_quantization if hasattr(self.config, 'use_cache_quantization') else False
 
+        max_sequence_length = config.max_position_embeddings
+        self.position_embedding_type = config.pos_emb
         self.gradient_checkpointing = False
-        self.use_dynamic_ntk = config.use_dynamic_ntk
-        self.seq_length = config.seq_length
-
-        self.wte = nn.Embedding(self.vocab_size, self.embed_dim)
-
-        self.drop = nn.Dropout(config.emb_dropout_prob)
 
-        if config.rotary_pct == 1.0:
-            self.rotary_ndims = None
+        if self.position_embedding_type == "learned":
+            self.wpe = nn.Embedding(max_sequence_length, self.embed_dim)
+            self.init_method(self.position_embeddings.weight)
+            self._position_embeddings_key = "position_embeddings"
+            self.init_method(self.position_embeddings.weight)
         else:
-            assert config.rotary_pct < 1
-            self.rotary_ndims = int(
-                config.kv_channels * config.rotary_pct
-            )
-        dim = (
-            self.rotary_ndims
-            if self.rotary_ndims is not None
-            else config.kv_channels
-        )
-        self.rotary_emb = RotaryEmbedding(dim, base=config.rotary_emb_base)
+            self.wpe = None
+            self._position_embeddings_key = ""
 
-        self.use_flash_attn = config.use_flash_attn
-        self.is_fp32 = not (config.bf16 or config.fp16)
+        self.wte = nn.Embedding(self.vocab_size, self.embed_dim)
 
+        self.drop = nn.Dropout(config.embd_pdrop)
         self.h = nn.ModuleList(
             [
                 QWenBlock(
-                    config
+                    config,
+                    layer_idx=i,
                 )
                 for i in range(config.num_hidden_layers)
             ]
@@ -732,12 +641,6 @@ class QWenModel(QWenPreTrainedModel):
     def set_input_embeddings(self, new_embeddings):
         self.wte = new_embeddings
 
-    def get_ntk_alpha(self, true_seq_len):
-        context_value = math.log(true_seq_len / self.seq_length, 2) + 1
-        ntk_alpha = 2 ** math.ceil(context_value) - 1
-        ntk_alpha = max(ntk_alpha, 1)
-        return ntk_alpha
-
     def forward(
         self,
         input_ids: Optional[torch.LongTensor] = None,
@@ -794,10 +697,8 @@ class QWenModel(QWenPreTrainedModel):
             past_length = 0
             past_key_values = tuple([None] * len(self.h))
         else:
-            if self.use_cache_quantization:
-                past_length = past_key_values[0][0][0].size(2)
-            else:
-                past_length = past_key_values[0][0].size(-2)
+            past_length = past_key_values[0][0].size(-2)
+
         if position_ids is None:
             position_ids = torch.arange(
                 past_length,
@@ -816,39 +717,14 @@ class QWenModel(QWenPreTrainedModel):
             attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min
 
         encoder_attention_mask = None
-        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+        head_mask = self.get_head_mask(head_mask, self.config.n_layer)
 
         if inputs_embeds is None:
             inputs_embeds = self.wte(input_ids)
         hidden_states = inputs_embeds
-
-        kv_seq_len = hidden_states.size()[1]
-        if past_key_values[0] is not None:
-            # past key values[0][0] shape: bs * seq_len * head_num * dim
-            if self.use_cache_quantization:
-                kv_seq_len += past_key_values[0][0][0].shape[2]
-            else:
-                kv_seq_len += past_key_values[0][0].shape[1]
-
-        if self.training or not self.use_dynamic_ntk:
-            ntk_alpha_list = [1.0]
-        elif kv_seq_len != hidden_states.size()[1]:
-            ntk_alpha_list = self.rotary_emb._ntk_alpha_cached_list
-        else:
-            ntk_alpha_list = []
-            if attention_mask is not None and kv_seq_len > self.seq_length:
-                true_seq_lens = attention_mask.squeeze(1).squeeze(1).eq(0).sum(dim=-1, dtype=torch.int32)
-                for i in range(hidden_states.size()[0]):
-                    true_seq_len = true_seq_lens[i].item()
-                    ntk_alpha = self.get_ntk_alpha(true_seq_len)
-                    ntk_alpha_list.append(ntk_alpha)
-            else:
-                ntk_alpha = self.get_ntk_alpha(kv_seq_len)
-                ntk_alpha_list.append(ntk_alpha)
-        self.rotary_emb._ntk_alpha_cached_list = ntk_alpha_list
-        rotary_pos_emb_list = [
-            self.rotary_emb(kv_seq_len, ntk_alpha=ntk_alpha) for ntk_alpha in ntk_alpha_list
-        ]
+        if self.wpe is not None:
+            position_embeds = self.wpe(position_ids)
+            hidden_states = hidden_states + position_embeds
 
         hidden_states = self.drop(hidden_states)
         output_shape = input_shape + (hidden_states.size(-1),)
@@ -880,7 +756,6 @@ class QWenModel(QWenPreTrainedModel):
                 outputs = torch.utils.checkpoint.checkpoint(
                     create_custom_forward(block),
                     hidden_states,
-                    rotary_pos_emb_list,
                     None,
                     attention_mask,
                     head_mask[i],
@@ -891,7 +766,6 @@ class QWenModel(QWenPreTrainedModel):
                 outputs = block(
                     hidden_states,
                     layer_past=layer_past,
-                    rotary_pos_emb_list=rotary_pos_emb_list,
                     attention_mask=attention_mask,
                     head_mask=head_mask[i],
                     encoder_hidden_states=encoder_hidden_states,
@@ -902,16 +776,13 @@ class QWenModel(QWenPreTrainedModel):
 
             hidden_states = outputs[0]
             if use_cache is True:
-                presents = presents + (outputs[1],)
+                presents = presents + (outputs[2 if output_attentions else 1],)
 
             if output_attentions:
-                all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
+                all_self_attentions = all_self_attentions + (outputs[1],)
 
         hidden_states = self.ln_f(hidden_states)
         hidden_states = hidden_states.view(output_shape)
-        # Add last hidden state
-        if output_hidden_states:
-            all_hidden_states = all_hidden_states + (hidden_states,)
 
         if not return_dict:
             return tuple(
@@ -963,7 +834,7 @@ class QWenLMHeadModel(QWenPreTrainedModel):
                 logger.warn("Your device support faster inference by passing bf16=True in \"AutoModelForCausalLM.from_pretrained\".")
             elif SUPPORT_FP16:
                 logger.warn("Your device support faster inference by passing fp16=True in \"AutoModelForCausalLM.from_pretrained\".")
-
+        
         if config.use_flash_attn == "auto":
             if config.bf16 or config.fp16:
                 logger.warn("Try importing flash-attention for faster inference...")
@@ -977,7 +848,7 @@ class QWenLMHeadModel(QWenPreTrainedModel):
             _import_flash_attn()
 
         self.transformer = QWenModel(config)
-        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
 
         if config.bf16:
             self.transformer.bfloat16()
@@ -996,13 +867,22 @@ class QWenLMHeadModel(QWenPreTrainedModel):
     def prepare_inputs_for_generation(
         self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs
     ):
+        token_type_ids = kwargs.get("token_type_ids", None)
         if past_key_values:
             input_ids = input_ids[:, -1].unsqueeze(-1)
+            if token_type_ids is not None:
+                token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
+
+        attention_mask = kwargs.get("attention_mask", None)
+        position_ids = kwargs.get("position_ids", None)
 
-        if input_ids.size(0) == 1:
-            attention_mask = None
+        if attention_mask is not None and position_ids is None:
+            position_ids = attention_mask.long().cumsum(-1) - 1
+            position_ids.masked_fill_(attention_mask == 0, 1)
+            if past_key_values:
+                position_ids = position_ids[:, -1].unsqueeze(-1)
         else:
-            attention_mask = kwargs.get("attention_mask", None)
+            position_ids = None
 
         if inputs_embeds is not None and past_key_values is None:
             model_inputs = {"inputs_embeds": inputs_embeds}
@@ -1013,7 +893,9 @@ class QWenLMHeadModel(QWenPreTrainedModel):
             {
                 "past_key_values": past_key_values,
                 "use_cache": kwargs.get("use_cache"),
+                "position_ids": position_ids,
                 "attention_mask": attention_mask,
+                "token_type_ids": token_type_ids,
             }
         )
         return model_inputs
@@ -1100,45 +982,35 @@ class QWenLMHeadModel(QWenPreTrainedModel):
         query: str,
         history: Optional[HistoryType],
         system: str = "You are a helpful assistant.",
+        append_history: bool = True,
         stream: Optional[bool] = _SENTINEL,
         stop_words_ids: Optional[List[List[int]]] = None,
-        generation_config: Optional[GenerationConfig] = None,
         **kwargs,
     ) -> Tuple[str, HistoryType]:
-        generation_config = generation_config if generation_config is not None else self.generation_config
-
         assert stream is _SENTINEL, _ERROR_STREAM_IN_CHAT
-        assert generation_config.chat_format == 'chatml', _ERROR_BAD_CHAT_FORMAT
+        assert self.generation_config.chat_format == 'chatml', _ERROR_BAD_CHAT_FORMAT
         if history is None:
             history = []
-        else:
-            # make a copy of the user's input such that is is left untouched
-            history = copy.deepcopy(history)
-
         if stop_words_ids is None:
             stop_words_ids = []
 
-        max_window_size = kwargs.get('max_window_size', None)
-        if max_window_size is None:
-            max_window_size = generation_config.max_window_size
         raw_text, context_tokens = make_context(
             tokenizer,
             query,
             history=history,
             system=system,
-            max_window_size=max_window_size,
-            chat_format=generation_config.chat_format,
+            max_window_size=6144,
+            chat_format=self.generation_config.chat_format,
         )
 
         stop_words_ids.extend(get_stop_words_ids(
-            generation_config.chat_format, tokenizer
+            self.generation_config.chat_format, tokenizer
         ))
         input_ids = torch.tensor([context_tokens]).to(self.device)
         outputs = self.generate(
                     input_ids,
-                    stop_words_ids=stop_words_ids,
-                    return_dict_in_generate=False,
-                    generation_config=generation_config,
+                    stop_words_ids = stop_words_ids,
+                    return_dict_in_generate = False,
                     **kwargs,
                 )
 
@@ -1147,16 +1019,13 @@ class QWenLMHeadModel(QWenPreTrainedModel):
             tokenizer,
             raw_text_len=len(raw_text),
             context_length=len(context_tokens),
-            chat_format=generation_config.chat_format,
+            chat_format=self.generation_config.chat_format,
             verbose=False,
             errors='replace'
         )
 
-        # as history is a copy of the user inputs,
-        # we can always return the new turn to the user.
-        # separating input history and output history also enables the user
-        # to implement more complex history management
-        history.append((query, response))
+        if append_history:
+            history.append((query, response))
 
         return response, history
 
@@ -1168,35 +1037,30 @@ class QWenLMHeadModel(QWenPreTrainedModel):
             system: str = "You are a helpful assistant.",
             stop_words_ids: Optional[List[List[int]]] = None,
             logits_processor: Optional[LogitsProcessorList] = None,
-            generation_config: Optional[GenerationConfig] = None,
             **kwargs,
     ) -> Generator[str, Any, None]:
-        generation_config = generation_config if generation_config is not None else self.generation_config
-        assert generation_config.chat_format == 'chatml', _ERROR_BAD_CHAT_FORMAT
+        assert self.generation_config.chat_format == 'chatml', _ERROR_BAD_CHAT_FORMAT
         if history is None:
             history = []
         if stop_words_ids is None:
             stop_words_ids = []
 
-        max_window_size = kwargs.get('max_window_size', None)
-        if max_window_size is None:
-            max_window_size = generation_config.max_window_size
         raw_text, context_tokens = make_context(
             tokenizer,
             query,
             history=history,
             system=system,
-            max_window_size=max_window_size,
-            chat_format=generation_config.chat_format,
+            max_window_size=6144,
+            chat_format=self.generation_config.chat_format,
         )
 
         stop_words_ids.extend(get_stop_words_ids(
-            generation_config.chat_format, tokenizer
+            self.generation_config.chat_format, tokenizer
         ))
         if stop_words_ids is not None:
             stop_words_logits_processor = StopWordsLogitsProcessor(
                 stop_words_ids=stop_words_ids,
-                eos_token_id=generation_config.eos_token_id,
+                eos_token_id=self.generation_config.eos_token_id,
             )
             if logits_processor is None:
                 logits_processor = LogitsProcessorList([stop_words_logits_processor])
@@ -1207,8 +1071,7 @@ class QWenLMHeadModel(QWenPreTrainedModel):
         from transformers_stream_generator.main import NewGenerationMixin, StreamGenerationConfig
         self.__class__.generate_stream = NewGenerationMixin.generate
         self.__class__.sample_stream = NewGenerationMixin.sample_stream
-        stream_config = StreamGenerationConfig(**generation_config.to_dict(), do_stream=True)
-
+        stream_config = StreamGenerationConfig(**self.generation_config.to_dict(), do_stream=True)
         def stream_generator():
             outputs = []
             for token in self.generate_stream(
@@ -1237,19 +1100,17 @@ class QWenLMHeadModel(QWenPreTrainedModel):
         streamer: Optional["BaseStreamer"] = None,
         **kwargs,
     ) -> Union[GenerateOutput, torch.LongTensor]:
-        generation_config = generation_config if generation_config is not None else self.generation_config
-
         # Process stop_words_ids.
         stop_words_ids = kwargs.pop("stop_words_ids", None)
         if stop_words_ids is None and generation_config is not None:
             stop_words_ids = getattr(generation_config, "stop_words_ids", None)
         if stop_words_ids is None:
-            stop_words_ids = getattr(generation_config, "stop_words_ids", None)
+            stop_words_ids = getattr(self.generation_config, "stop_words_ids", None)
 
         if stop_words_ids is not None:
             stop_words_logits_processor = StopWordsLogitsProcessor(
                 stop_words_ids=stop_words_ids,
-                eos_token_id=generation_config.eos_token_id,
+                eos_token_id=self.generation_config.eos_token_id,
             )
             if logits_processor is None:
                 logits_processor = LogitsProcessorList([stop_words_logits_processor])
@@ -1274,17 +1135,16 @@ class RotaryEmbedding(torch.nn.Module):
         super().__init__()
         self.dim = dim
         self.base = base
-        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
-        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
         if importlib.util.find_spec("einops") is None:
             raise RuntimeError("einops is required for Rotary Embedding")
 
         self._rotary_pos_emb_cache = None
         self._seq_len_cached = 0
         self._ntk_alpha_cached = 1.0
-        self._ntk_alpha_cached_list = [1.0]
 
-    def update_rotary_pos_emb_cache(self, seqlen, ntk_alpha=1.0):
+    def update_rotary_pos_emb_cache(self, max_seq_len, offset=0, ntk_alpha=1.0):
+        seqlen = max_seq_len + offset
         if seqlen > self._seq_len_cached or ntk_alpha != self._ntk_alpha_cached:
             base = self.base * ntk_alpha ** (self.dim / (self.dim - 2))
             self.inv_freq = 1.0 / (
@@ -1298,19 +1158,14 @@ class RotaryEmbedding(torch.nn.Module):
             self._ntk_alpha_cached = ntk_alpha
             seq = torch.arange(self._seq_len_cached, device=self.inv_freq.device)
             freqs = torch.outer(seq.type_as(self.inv_freq), self.inv_freq)
-
             emb = torch.cat((freqs, freqs), dim=-1)
             from einops import rearrange
 
-            emb = rearrange(emb, "n d -> 1 n 1 d")
+            self._rotary_pos_emb_cache = rearrange(emb, "n d -> 1 n 1 d")
 
-            cos, sin = emb.cos(), emb.sin()
-            self._rotary_pos_emb_cache = [cos, sin]
-
-    def forward(self, max_seq_len, ntk_alpha=1.0):
-        self.update_rotary_pos_emb_cache(max_seq_len, ntk_alpha)
-        cos, sin = self._rotary_pos_emb_cache
-        return [cos[:, :max_seq_len], sin[:, :max_seq_len]]
+    def forward(self, max_seq_len, offset=0, ntk_alpha=1.0):
+        self.update_rotary_pos_emb_cache(max_seq_len, offset, ntk_alpha)
+        return self._rotary_pos_emb_cache[:, offset : offset + max_seq_len]
 
 
 def _rotate_half(x):
@@ -1322,28 +1177,20 @@ def _rotate_half(x):
 
 
 def apply_rotary_pos_emb(t, freqs):
-    """ Apply rotary embedding to the first rotary_dim of the iput
-
-    Arguments:
-      t (tensor(batch_size, seq_len, n_head, head_dim)):
-        the input embedding/hidden states
-      freqs (list[tensor(1, seq_len, 1, rotary_dim), tensor(1, seq_len, 1, rotary_dim)]):
-        the cached cos/sin position embeddings
-    """
-    rot_dim = freqs[0].shape[-1]
-    cos, sin = freqs
-    t_float = t.float()
-    if apply_rotary_emb_func is not None and t.is_cuda:
-        # apply_rotary_emb in flash_attn requires cos/sin to be of
-        # shape (seqlen, rotary_dim / 2) and apply rotary embedding
-        # to the first rotary_dim of the input
-        cos = cos.squeeze(0).squeeze(1)[:, : rot_dim // 2]
-        sin = sin.squeeze(0).squeeze(1)[:, : rot_dim // 2]
-        return apply_rotary_emb_func(t_float, cos, sin).type_as(t)
+    if apply_rotary_emb_func is not None:
+        t_ = t.float()
+        freqs = freqs.squeeze(0).squeeze(1)
+        cos = freqs[:, : freqs.shape[-1] // 2].cos()
+        sin = freqs[:, : freqs.shape[-1] // 2].sin()
+        output = apply_rotary_emb_func(t_, cos, sin).type_as(t)
+        return output
     else:
-        t_rot, t_pass = t_float[..., :rot_dim], t_float[..., rot_dim:]
-        t_rot = (t_rot * cos) + (_rotate_half(t_rot) * sin)
-        return torch.cat((t_rot, t_pass), dim=-1).type_as(t)
+        rot_dim = freqs.shape[-1]
+        t_, t_pass_ = t[..., :rot_dim], t[..., rot_dim:]
+        t_ = t_.float()
+        t_pass_ = t_pass_.float()
+        t_ = (t_ * freqs.cos()) + (_rotate_half(t_) * freqs.sin())
+        return torch.cat((t_, t_pass_), dim=-1).type_as(t)
 
 
 class RMSNorm(torch.nn.Module):

model-00002-of-00008.safetensors → pytorch_model-00001-of-00008.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:63384af60dff14f3655142543b97372e144a0e3c238579984a9723ad9a4e676d
-size 2023960808
+oid sha256:33296fd065504ad10d93f4fbe8635f33e81885c0c2ac0396004f82770dd781ee
+size 1964070447

model-00003-of-00008.safetensors → pytorch_model-00002-of-00008.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d2d011ce1557d43fdbc8d57fa4bbc7f1b7209155060984461c3513a6b9fbfcbc
-size 2023960816
+oid sha256:2745ae50fffd4194e643f965183a29a9abc7595e8ee358d6311f0c81d5ab3fe0
+size 2023970015

model-00004-of-00008.safetensors → pytorch_model-00003-of-00008.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5c374149caf74e88d228ba50bfcbfcf11004e4b16bb17c830864baa29c5ddc02
-size 2023960848
+oid sha256:ee11ea92b22d04b18fdf2bcedb8a4189db41b0e9359332444bdb406cca7abcfe
+size 2023970079

pytorch_model-00004-of-00008.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1d6d68e5056982035547049e14d7df29e67c8b813bcf8faa520484022893ea3a
+size 2023970079

pytorch_model-00005-of-00008.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bc53af5f365b1a364cf207d3c42e14e654889daf079e382332f055b715488794
+size 2023970079

pytorch_model-00006-of-00008.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:800f774aa1839e08f6d25da50e020dc6264539bbbdf691665409d92ae70656c8
+size 2023970079

pytorch_model-00007-of-00008.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:99940cb4d10844d342f8f061d68fb317fd9e45d2a8903ab8aa5e3191a3a320d8
+size 2023970079

pytorch_model-00008-of-00008.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:05be1d470d28513846fcbf0ccd75224bb4bf5b15ec3162305fddece503b0bcf8
+size 1334846988

pytorch_model.bin.index.json

@@ -0,0 +1,266 @@
+{
+  "metadata": {
+    "total_size": 15442649088
+  },
+  "weight_map": {
+    "lm_head.weight": "pytorch_model-00008-of-00008.bin",
+    "transformer.h.0.attn.c_attn.bias": "pytorch_model-00001-of-00008.bin",
+    "transformer.h.0.attn.c_attn.weight": "pytorch_model-00001-of-00008.bin",
+    "transformer.h.0.attn.c_proj.weight": "pytorch_model-00001-of-00008.bin",
+    "transformer.h.0.ln_1.weight": "pytorch_model-00001-of-00008.bin",
+    "transformer.h.0.ln_2.weight": "pytorch_model-00001-of-00008.bin",
+    "transformer.h.0.mlp.c_proj.weight": "pytorch_model-00001-of-00008.bin",
+    "transformer.h.0.mlp.w1.weight": "pytorch_model-00001-of-00008.bin",
+    "transformer.h.0.mlp.w2.weight": "pytorch_model-00001-of-00008.bin",
+    "transformer.h.1.attn.c_attn.bias": "pytorch_model-00001-of-00008.bin",
+    "transformer.h.1.attn.c_attn.weight": "pytorch_model-00001-of-00008.bin",
+    "transformer.h.1.attn.c_proj.weight": "pytorch_model-00001-of-00008.bin",
+    "transformer.h.1.ln_1.weight": "pytorch_model-00001-of-00008.bin",
+    "transformer.h.1.ln_2.weight": "pytorch_model-00001-of-00008.bin",
+    "transformer.h.1.mlp.c_proj.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.1.mlp.w1.weight": "pytorch_model-00001-of-00008.bin",
+    "transformer.h.1.mlp.w2.weight": "pytorch_model-00001-of-00008.bin",
+    "transformer.h.10.attn.c_attn.bias": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.10.attn.c_attn.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.10.attn.c_proj.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.10.ln_1.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.10.ln_2.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.10.mlp.c_proj.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.10.mlp.w1.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.10.mlp.w2.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.11.attn.c_attn.bias": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.11.attn.c_attn.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.11.attn.c_proj.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.11.ln_1.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.11.ln_2.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.11.mlp.c_proj.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.11.mlp.w1.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.11.mlp.w2.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.12.attn.c_attn.bias": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.12.attn.c_attn.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.12.attn.c_proj.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.12.ln_1.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.12.ln_2.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.12.mlp.c_proj.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.12.mlp.w1.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.12.mlp.w2.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.13.attn.c_attn.bias": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.13.attn.c_attn.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.13.attn.c_proj.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.13.ln_1.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.13.ln_2.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.13.mlp.c_proj.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.13.mlp.w1.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.13.mlp.w2.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.14.attn.c_attn.bias": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.14.attn.c_attn.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.14.attn.c_proj.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.14.ln_1.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.14.ln_2.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.14.mlp.c_proj.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.14.mlp.w1.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.14.mlp.w2.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.15.attn.c_attn.bias": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.15.attn.c_attn.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.15.attn.c_proj.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.15.ln_1.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.15.ln_2.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.15.mlp.c_proj.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.15.mlp.w1.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.15.mlp.w2.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.16.attn.c_attn.bias": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.16.attn.c_attn.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.16.attn.c_proj.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.16.ln_1.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.16.ln_2.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.16.mlp.c_proj.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.16.mlp.w1.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.16.mlp.w2.weight": "pytorch_model-00004-of-00008.bin",
+    "transformer.h.17.attn.c_attn.bias": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.17.attn.c_attn.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.17.attn.c_proj.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.17.ln_1.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.17.ln_2.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.17.mlp.c_proj.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.17.mlp.w1.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.17.mlp.w2.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.18.attn.c_attn.bias": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.18.attn.c_attn.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.18.attn.c_proj.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.18.ln_1.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.18.ln_2.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.18.mlp.c_proj.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.18.mlp.w1.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.18.mlp.w2.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.19.attn.c_attn.bias": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.19.attn.c_attn.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.19.attn.c_proj.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.19.ln_1.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.19.ln_2.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.19.mlp.c_proj.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.19.mlp.w1.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.19.mlp.w2.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.2.attn.c_attn.bias": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.2.attn.c_attn.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.2.attn.c_proj.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.2.ln_1.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.2.ln_2.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.2.mlp.c_proj.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.2.mlp.w1.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.2.mlp.w2.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.20.attn.c_attn.bias": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.20.attn.c_attn.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.20.attn.c_proj.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.20.ln_1.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.20.ln_2.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.20.mlp.c_proj.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.20.mlp.w1.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.20.mlp.w2.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.21.attn.c_attn.bias": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.21.attn.c_attn.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.21.attn.c_proj.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.21.ln_1.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.21.ln_2.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.21.mlp.c_proj.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.21.mlp.w1.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.21.mlp.w2.weight": "pytorch_model-00005-of-00008.bin",
+    "transformer.h.22.attn.c_attn.bias": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.22.attn.c_attn.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.22.attn.c_proj.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.22.ln_1.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.22.ln_2.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.22.mlp.c_proj.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.22.mlp.w1.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.22.mlp.w2.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.23.attn.c_attn.bias": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.23.attn.c_attn.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.23.attn.c_proj.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.23.ln_1.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.23.ln_2.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.23.mlp.c_proj.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.23.mlp.w1.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.23.mlp.w2.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.24.attn.c_attn.bias": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.24.attn.c_attn.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.24.attn.c_proj.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.24.ln_1.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.24.ln_2.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.24.mlp.c_proj.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.24.mlp.w1.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.24.mlp.w2.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.25.attn.c_attn.bias": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.25.attn.c_attn.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.25.attn.c_proj.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.25.ln_1.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.25.ln_2.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.25.mlp.c_proj.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.25.mlp.w1.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.25.mlp.w2.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.26.attn.c_attn.bias": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.26.attn.c_attn.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.26.attn.c_proj.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.26.ln_1.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.26.ln_2.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.26.mlp.c_proj.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.26.mlp.w1.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.26.mlp.w2.weight": "pytorch_model-00006-of-00008.bin",
+    "transformer.h.27.attn.c_attn.bias": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.27.attn.c_attn.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.27.attn.c_proj.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.27.ln_1.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.27.ln_2.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.27.mlp.c_proj.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.27.mlp.w1.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.27.mlp.w2.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.28.attn.c_attn.bias": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.28.attn.c_attn.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.28.attn.c_proj.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.28.ln_1.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.28.ln_2.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.28.mlp.c_proj.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.28.mlp.w1.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.28.mlp.w2.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.29.attn.c_attn.bias": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.29.attn.c_attn.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.29.attn.c_proj.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.29.ln_1.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.29.ln_2.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.29.mlp.c_proj.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.29.mlp.w1.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.29.mlp.w2.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.3.attn.c_attn.bias": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.3.attn.c_attn.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.3.attn.c_proj.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.3.ln_1.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.3.ln_2.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.3.mlp.c_proj.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.3.mlp.w1.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.3.mlp.w2.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.30.attn.c_attn.bias": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.30.attn.c_attn.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.30.attn.c_proj.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.30.ln_1.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.30.ln_2.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.30.mlp.c_proj.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.30.mlp.w1.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.30.mlp.w2.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.31.attn.c_attn.bias": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.31.attn.c_attn.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.31.attn.c_proj.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.31.ln_1.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.31.ln_2.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.31.mlp.c_proj.weight": "pytorch_model-00008-of-00008.bin",
+    "transformer.h.31.mlp.w1.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.31.mlp.w2.weight": "pytorch_model-00007-of-00008.bin",
+    "transformer.h.4.attn.c_attn.bias": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.4.attn.c_attn.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.4.attn.c_proj.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.4.ln_1.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.4.ln_2.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.4.mlp.c_proj.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.4.mlp.w1.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.4.mlp.w2.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.5.attn.c_attn.bias": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.5.attn.c_attn.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.5.attn.c_proj.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.5.ln_1.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.5.ln_2.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.5.mlp.c_proj.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.5.mlp.w1.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.5.mlp.w2.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.6.attn.c_attn.bias": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.6.attn.c_attn.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.6.attn.c_proj.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.6.ln_1.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.6.ln_2.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.6.mlp.c_proj.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.6.mlp.w1.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.6.mlp.w2.weight": "pytorch_model-00002-of-00008.bin",
+    "transformer.h.7.attn.c_attn.bias": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.7.attn.c_attn.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.7.attn.c_proj.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.7.ln_1.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.7.ln_2.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.7.mlp.c_proj.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.7.mlp.w1.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.7.mlp.w2.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.8.attn.c_attn.bias": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.8.attn.c_attn.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.8.attn.c_proj.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.8.ln_1.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.8.ln_2.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.8.mlp.c_proj.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.8.mlp.w1.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.8.mlp.w2.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.9.attn.c_attn.bias": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.9.attn.c_attn.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.9.attn.c_proj.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.9.ln_1.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.9.ln_2.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.9.mlp.c_proj.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.9.mlp.w1.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.h.9.mlp.w2.weight": "pytorch_model-00003-of-00008.bin",
+    "transformer.ln_f.weight": "pytorch_model-00008-of-00008.bin",
+    "transformer.wte.weight": "pytorch_model-00001-of-00008.bin"
+  }
+}

tokenization_qwen.py

@@ -27,22 +27,11 @@ IMEND = "<|im_end|>"
 # regular texts, the surface forms of special tokens need to be
 # as different as possible to minimize the impact
 EXTRAS = tuple((f"<|extra_{i}|>" for i in range(205)))
-# changed to use actual index to avoid misconfiguration with vocabulary expansion
-SPECIAL_START_ID = 151643
-SPECIAL_TOKENS = tuple(
-    enumerate(
-        (
-            (
-                ENDOFTEXT,
-                IMSTART,
-                IMEND,
-            )
-            + EXTRAS
-        ),
-        start=SPECIAL_START_ID,
-    )
-)
-SPECIAL_TOKENS_SET = set(t for i, t in SPECIAL_TOKENS)
+SPECIAL_TOKENS = (
+    ENDOFTEXT,
+    IMSTART,
+    IMEND,
+) + EXTRAS
 
 
 def _load_tiktoken_bpe(tiktoken_bpe_file: str) -> Dict[bytes, int]:
@@ -53,7 +42,6 @@ def _load_tiktoken_bpe(tiktoken_bpe_file: str) -> Dict[bytes, int]:
         for token, rank in (line.split() for line in contents.splitlines() if line)
     }
 
-
 class QWenTokenizer(PreTrainedTokenizer):
     """QWen tokenizer."""
 
@@ -63,35 +51,20 @@ class QWenTokenizer(PreTrainedTokenizer):
         self,
         vocab_file,
         errors="replace",
-        extra_vocab_file=None,
         **kwargs,
     ):
         super().__init__(**kwargs)
 
-        # how to handle errors in decoding UTF-8 byte sequences
-        # use ignore if you are in streaming inference
-        self.errors = errors  
+        self.errors = errors  # how to handle errors in decoding
 
-        self.mergeable_ranks = _load_tiktoken_bpe(vocab_file)  # type: Dict[bytes, int]
+        self.mergeable_ranks = _load_tiktoken_bpe(vocab_file)  # type: dict[bytes, int]
         self.special_tokens = {
             token: index
-            for index, token in SPECIAL_TOKENS
+            for index, token in enumerate(
+                SPECIAL_TOKENS, start=len(self.mergeable_ranks)
+            )
         }
 
-        # try load extra vocab from file
-        if extra_vocab_file is not None:
-            used_ids = set(self.mergeable_ranks.values()) | set(self.special_tokens.values())
-            extra_mergeable_ranks = _load_tiktoken_bpe(extra_vocab_file)
-            for token, index in extra_mergeable_ranks.items():
-                if token in self.mergeable_ranks:
-                    logger.info(f"extra token {token} exists, skipping")
-                    continue
-                if index in used_ids:
-                    logger.info(f'the index {index} for extra token {token} exists, skipping')
-                    continue
-                self.mergeable_ranks[token] = index
-            # the index may be sparse after this, but don't worry tiktoken.Encoding will handle this
-
         enc = tiktoken.Encoding(
             "Qwen",
             pat_str=PAT_STR,
@@ -113,23 +86,6 @@ class QWenTokenizer(PreTrainedTokenizer):
         self.im_start_id = self.special_tokens[IMSTART]
         self.im_end_id = self.special_tokens[IMEND]
 
-    def __getstate__(self):
-        # for pickle lovers
-        state = self.__dict__.copy()
-        del state["tokenizer"]
-        return state
-
-    def __setstate__(self, state):
-        # tokenizer is not python native; don't pass it; rebuild it
-        self.__dict__.update(state)
-        enc = tiktoken.Encoding(
-            "Qwen",
-            pat_str=PAT_STR,
-            mergeable_ranks=self.mergeable_ranks,
-            special_tokens=self.special_tokens,
-        )
-        self.tokenizer = enc
-
     def __len__(self) -> int:
         return self.tokenizer.n_vocab
 
@@ -152,17 +108,13 @@ class QWenTokenizer(PreTrainedTokenizer):
                 ids.append(self.mergeable_ranks.get(token))
         return ids
 
-    def _add_tokens(
-        self,
-        new_tokens: Union[List[str], List[AddedToken]],
-        special_tokens: bool = False,
-    ) -> int:
+    def _add_tokens(self, new_tokens: Union[List[str], List[AddedToken]], special_tokens: bool = False) -> int:
         if not special_tokens and new_tokens:
-            raise ValueError("Adding regular tokens is not supported")
+            raise ValueError('Adding regular tokens is not supported')
         for token in new_tokens:
             surface_form = token.content if isinstance(token, AddedToken) else token
-            if surface_form not in SPECIAL_TOKENS_SET:
-                raise ValueError("Adding unknown special tokens is not supported")
+            if surface_form not in SPECIAL_TOKENS:
+                raise ValueError('Adding unknown special tokens is not supported')
         return 0
 
     def save_vocabulary(self, save_directory: str, **kwargs) -> Tuple[str]:

tokenizer_config.json

@@ -1,5 +1,5 @@
 {
-  "model_max_length": 32768,
+  "model_max_length": 8192,
   "tokenizer_class": "QWenTokenizer",
   "auto_map": {
     "AutoTokenizer": [