update
Browse files- LICENSE +55 -0
- NOTICE +280 -0
- README.md +463 -0
- assets/logo.jpg +0 -0
- assets/qwen_tokenizer.png +0 -0
- assets/react_showcase_001.png +0 -0
- assets/react_showcase_002.png +0 -0
- assets/wechat.png +0 -0
- cache_autogptq_cuda_256.cpp +198 -0
- cache_autogptq_cuda_kernel_256.cu +1708 -0
- examples/react_prompt.md +249 -0
- modeling_qwen.py +62 -69
LICENSE
ADDED
|
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Tongyi Qianwen RESEARCH LICENSE AGREEMENT
|
| 2 |
+
|
| 3 |
+
Tongyi Qianwen Release Date: November 30, 2023
|
| 4 |
+
|
| 5 |
+
By clicking to agree or by using or distributing any portion or element of the Tongyi Qianwen Materials, you will be deemed to have recognized and accepted the content of this Agreement, which is effective immediately.
|
| 6 |
+
|
| 7 |
+
1. Definitions
|
| 8 |
+
a. This Tongyi Qianwen RESEARCH LICENSE AGREEMENT (this "Agreement") shall mean the terms and conditions for use, reproduction, distribution and modification of the Materials as defined by this Agreement.
|
| 9 |
+
b. "We"(or "Us") shall mean Alibaba Cloud.
|
| 10 |
+
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.
|
| 11 |
+
d. "Third Parties" shall mean individuals or legal entities that are not under common control with Us or You.
|
| 12 |
+
e. "Tongyi Qianwen" shall mean the large language models, 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.
|
| 13 |
+
f. "Materials" shall mean, collectively, Alibaba Cloud's proprietary Tongyi Qianwen and Documentation (and any portion thereof) made available under this Agreement.
|
| 14 |
+
g. "Source" form shall mean the preferred form for making modifications, including but not limited to model source code, documentation source, and configuration files.
|
| 15 |
+
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,
|
| 16 |
+
and conversions to other media types.
|
| 17 |
+
i. "Non-Commercial" shall mean for research or evaluation purposes only.
|
| 18 |
+
|
| 19 |
+
2. Grant of Rights
|
| 20 |
+
a. You are granted a non-exclusive, worldwide, non-transferable and royalty-free limited license under Alibaba Cloud's intellectual property or other rights owned by Us embodied in the Materials to use, reproduce, distribute, copy, create derivative works of, and make modifications to the Materials FOR NON-COMMERCIAL PURPOSES ONLY.
|
| 21 |
+
b. If you are commercially using the Materials, You shall request a license from Us.
|
| 22 |
+
|
| 23 |
+
3. Redistribution
|
| 24 |
+
You may reproduce and distribute copies of the Materials or derivative works thereof in any medium, with or without modifications, and in Source or Object form, provided that You meet the following conditions:
|
| 25 |
+
a. You shall give any other recipients of the Materials or derivative works a copy of this Agreement;
|
| 26 |
+
b. You shall cause any modified files to carry prominent notices stating that You changed the files;
|
| 27 |
+
c. You shall retain in all copies of the Materials that You distribute the following attribution notices within a "Notice" text file distributed as a part of such copies: "Tongyi Qianwen is licensed under the Tongyi Qianwen RESEARCH LICENSE AGREEMENT, Copyright (c) Alibaba Cloud. All Rights Reserved."; and
|
| 28 |
+
d. 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 terms and conditions of this Agreement.
|
| 29 |
+
|
| 30 |
+
4. Rules of use
|
| 31 |
+
a. The Materials may be subject to export controls or restrictions in China, the United States or other countries or regions. You shall comply with applicable laws and regulations in your use of the Materials.
|
| 32 |
+
b. You can not use the Materials or any output therefrom to improve any other large language model (excluding Tongyi Qianwen or derivative works thereof).
|
| 33 |
+
|
| 34 |
+
5. Intellectual Property
|
| 35 |
+
a. We retain ownership of all intellectual property rights in and to the Materials and derivatives made by or for Us. Conditioned upon compliance with the terms and conditions of this Agreement, with respect to any derivative works and modifications of the Materials that are made by you, you are and will be the owner of such derivative works and modifications.
|
| 36 |
+
b. No trademark license is granted to use the trade names, trademarks, service marks, or product names of Us, except as required to fulfill notice requirements under this Agreement or as required for reasonable and customary use in describing and redistributing the Materials.
|
| 37 |
+
c. If you commence a lawsuit or other proceedings (including a cross-claim or counterclaim in a lawsuit) against Us or any entity alleging that the Materials or any output therefrom, or any part of the foregoing, infringe any intellectual property or other right owned or licensable by you, then all licences granted to you under this Agreement shall terminate as of the date such lawsuit or other proceeding is commenced or brought.
|
| 38 |
+
|
| 39 |
+
6. Disclaimer of Warranty and Limitation of Liability
|
| 40 |
+
a. We are not obligated to support, update, provide training for, or develop any further version of the Tongyi Qianwen Materials or to grant any license thereto.
|
| 41 |
+
b. THE MATERIALS ARE PROVIDED "AS IS" WITHOUT ANY EXPRESS OR IMPLIED WARRANTY OF ANY KIND INCLUDING WARRANTIES OF MERCHANTABILITY, NONINFRINGEMENT, OR FITNESS FOR A PARTICULAR PURPOSE. WE MAKE NO WARRANTY AND ASSUME NO RESPONSIBILITY FOR THE SAFETY OR STABILITY OF THE MATERIALS AND ANY OUTPUT THEREFROM.
|
| 42 |
+
c. IN NO EVENT SHALL WE BE LIABLE TO YOU FOR ANY DAMAGES, INCLUDING, BUT NOT LIMITED TO ANY DIRECT, OR INDIRECT, SPECIAL OR CONSEQUENTIAL DAMAGES ARISING FROM YOUR USE OR INABILITY TO USE THE MATERIALS OR ANY OUTPUT OF IT, NO MATTER HOW IT’S CAUSED.
|
| 43 |
+
d. You will defend, indemnify and hold harmless Us from and against any claim by any third party arising out of or related to your use or distribution of the Materials.
|
| 44 |
+
|
| 45 |
+
7. Survival and Termination.
|
| 46 |
+
a. The term of this Agreement shall commence upon your acceptance of this Agreement or access to the Materials and will continue in full force and effect until terminated in accordance with the terms and conditions herein.
|
| 47 |
+
b. We may terminate this Agreement if you breach any of the terms or conditions of this Agreement. Upon termination of this Agreement, you must delete and cease use of the Materials. Sections 6 and 8 shall survive the termination of this Agreement.
|
| 48 |
+
|
| 49 |
+
8. Governing Law and Jurisdiction.
|
| 50 |
+
a. This Agreement and any dispute arising out of or relating to it will be governed by the laws of China, without regard to conflict of law principles, and the UN Convention on Contracts for the International Sale of Goods does not apply to this Agreement.
|
| 51 |
+
b. The People's Courts in Hangzhou City shall have exclusive jurisdiction over any dispute arising out of this Agreement.
|
| 52 |
+
|
| 53 |
+
9. Other Terms and Conditions.
|
| 54 |
+
a. Any arrangements, understandings, or agreements regarding the Material not stated herein are separate from and independent of the terms and conditions of this Agreement. You shall request a seperate license from Us, if You use the Materials in ways not expressly agreed to in this Agreement.
|
| 55 |
+
b. We shall not be bound by any additional or different terms or conditions communicated by You unless expressly agreed.
|
NOTICE
ADDED
|
@@ -0,0 +1,280 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
------------- LICENSE FOR NVIDIA Megatron-LM code --------------
|
| 2 |
+
|
| 3 |
+
Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
|
| 4 |
+
|
| 5 |
+
Redistribution and use in source and binary forms, with or without
|
| 6 |
+
modification, are permitted provided that the following conditions
|
| 7 |
+
are met:
|
| 8 |
+
* Redistributions of source code must retain the above copyright
|
| 9 |
+
notice, this list of conditions and the following disclaimer.
|
| 10 |
+
* Redistributions in binary form must reproduce the above copyright
|
| 11 |
+
notice, this list of conditions and the following disclaimer in the
|
| 12 |
+
documentation and/or other materials provided with the distribution.
|
| 13 |
+
* Neither the name of NVIDIA CORPORATION nor the names of its
|
| 14 |
+
contributors may be used to endorse or promote products derived
|
| 15 |
+
from this software without specific prior written permission.
|
| 16 |
+
|
| 17 |
+
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
|
| 18 |
+
EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
| 19 |
+
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
|
| 20 |
+
PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
|
| 21 |
+
CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
|
| 22 |
+
EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
|
| 23 |
+
PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
|
| 24 |
+
PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
|
| 25 |
+
OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
|
| 26 |
+
(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
|
| 27 |
+
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
------------- LICENSE FOR OpenAI tiktoken code --------------
|
| 31 |
+
|
| 32 |
+
MIT License
|
| 33 |
+
|
| 34 |
+
Copyright (c) 2022 OpenAI, Shantanu Jain
|
| 35 |
+
|
| 36 |
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
| 37 |
+
of this software and associated documentation files (the "Software"), to deal
|
| 38 |
+
in the Software without restriction, including without limitation the rights
|
| 39 |
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
| 40 |
+
copies of the Software, and to permit persons to whom the Software is
|
| 41 |
+
furnished to do so, subject to the following conditions:
|
| 42 |
+
|
| 43 |
+
The above copyright notice and this permission notice shall be included in all
|
| 44 |
+
copies or substantial portions of the Software.
|
| 45 |
+
|
| 46 |
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
| 47 |
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
| 48 |
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
| 49 |
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
| 50 |
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
| 51 |
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
| 52 |
+
SOFTWARE.
|
| 53 |
+
|
| 54 |
+
------------- LICENSE FOR stanford_alpaca code --------------
|
| 55 |
+
|
| 56 |
+
Apache License
|
| 57 |
+
Version 2.0, January 2004
|
| 58 |
+
http://www.apache.org/licenses/
|
| 59 |
+
|
| 60 |
+
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
|
| 61 |
+
|
| 62 |
+
1. Definitions.
|
| 63 |
+
|
| 64 |
+
"License" shall mean the terms and conditions for use, reproduction,
|
| 65 |
+
and distribution as defined by Sections 1 through 9 of this document.
|
| 66 |
+
|
| 67 |
+
"Licensor" shall mean the copyright owner or entity authorized by
|
| 68 |
+
the copyright owner that is granting the License.
|
| 69 |
+
|
| 70 |
+
"Legal Entity" shall mean the union of the acting entity and all
|
| 71 |
+
other entities that control, are controlled by, or are under common
|
| 72 |
+
control with that entity. For the purposes of this definition,
|
| 73 |
+
"control" means (i) the power, direct or indirect, to cause the
|
| 74 |
+
direction or management of such entity, whether by contract or
|
| 75 |
+
otherwise, or (ii) ownership of fifty percent (50%) or more of the
|
| 76 |
+
outstanding shares, or (iii) beneficial ownership of such entity.
|
| 77 |
+
|
| 78 |
+
"You" (or "Your") shall mean an individual or Legal Entity
|
| 79 |
+
exercising permissions granted by this License.
|
| 80 |
+
|
| 81 |
+
"Source" form shall mean the preferred form for making modifications,
|
| 82 |
+
including but not limited to software source code, documentation
|
| 83 |
+
source, and configuration files.
|
| 84 |
+
|
| 85 |
+
"Object" form shall mean any form resulting from mechanical
|
| 86 |
+
transformation or translation of a Source form, including but
|
| 87 |
+
not limited to compiled object code, generated documentation,
|
| 88 |
+
and conversions to other media types.
|
| 89 |
+
|
| 90 |
+
"Work" shall mean the work of authorship, whether in Source or
|
| 91 |
+
Object form, made available under the License, as indicated by a
|
| 92 |
+
copyright notice that is included in or attached to the work
|
| 93 |
+
(an example is provided in the Appendix below).
|
| 94 |
+
|
| 95 |
+
"Derivative Works" shall mean any work, whether in Source or Object
|
| 96 |
+
form, that is based on (or derived from) the Work and for which the
|
| 97 |
+
editorial revisions, annotations, elaborations, or other modifications
|
| 98 |
+
represent, as a whole, an original work of authorship. For the purposes
|
| 99 |
+
of this License, Derivative Works shall not include works that remain
|
| 100 |
+
separable from, or merely link (or bind by name) to the interfaces of,
|
| 101 |
+
the Work and Derivative Works thereof.
|
| 102 |
+
|
| 103 |
+
"Contribution" shall mean any work of authorship, including
|
| 104 |
+
the original version of the Work and any modifications or additions
|
| 105 |
+
to that Work or Derivative Works thereof, that is intentionally
|
| 106 |
+
submitted to Licensor for inclusion in the Work by the copyright owner
|
| 107 |
+
or by an individual or Legal Entity authorized to submit on behalf of
|
| 108 |
+
the copyright owner. For the purposes of this definition, "submitted"
|
| 109 |
+
means any form of electronic, verbal, or written communication sent
|
| 110 |
+
to the Licensor or its representatives, including but not limited to
|
| 111 |
+
communication on electronic mailing lists, source code control systems,
|
| 112 |
+
and issue tracking systems that are managed by, or on behalf of, the
|
| 113 |
+
Licensor for the purpose of discussing and improving the Work, but
|
| 114 |
+
excluding communication that is conspicuously marked or otherwise
|
| 115 |
+
designated in writing by the copyright owner as "Not a Contribution."
|
| 116 |
+
|
| 117 |
+
"Contributor" shall mean Licensor and any individual or Legal Entity
|
| 118 |
+
on behalf of whom a Contribution has been received by Licensor and
|
| 119 |
+
subsequently incorporated within the Work.
|
| 120 |
+
|
| 121 |
+
2. Grant of Copyright License. Subject to the terms and conditions of
|
| 122 |
+
this License, each Contributor hereby grants to You a perpetual,
|
| 123 |
+
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
| 124 |
+
copyright license to reproduce, prepare Derivative Works of,
|
| 125 |
+
publicly display, publicly perform, sublicense, and distribute the
|
| 126 |
+
Work and such Derivative Works in Source or Object form.
|
| 127 |
+
|
| 128 |
+
3. Grant of Patent License. Subject to the terms and conditions of
|
| 129 |
+
this License, each Contributor hereby grants to You a perpetual,
|
| 130 |
+
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
| 131 |
+
(except as stated in this section) patent license to make, have made,
|
| 132 |
+
use, offer to sell, sell, import, and otherwise transfer the Work,
|
| 133 |
+
where such license applies only to those patent claims licensable
|
| 134 |
+
by such Contributor that are necessarily infringed by their
|
| 135 |
+
Contribution(s) alone or by combination of their Contribution(s)
|
| 136 |
+
with the Work to which such Contribution(s) was submitted. If You
|
| 137 |
+
institute patent litigation against any entity (including a
|
| 138 |
+
cross-claim or counterclaim in a lawsuit) alleging that the Work
|
| 139 |
+
or a Contribution incorporated within the Work constitutes direct
|
| 140 |
+
or contributory patent infringement, then any patent licenses
|
| 141 |
+
granted to You under this License for that Work shall terminate
|
| 142 |
+
as of the date such litigation is filed.
|
| 143 |
+
|
| 144 |
+
4. Redistribution. You may reproduce and distribute copies of the
|
| 145 |
+
Work or Derivative Works thereof in any medium, with or without
|
| 146 |
+
modifications, and in Source or Object form, provided that You
|
| 147 |
+
meet the following conditions:
|
| 148 |
+
|
| 149 |
+
(a) You must give any other recipients of the Work or
|
| 150 |
+
Derivative Works a copy of this License; and
|
| 151 |
+
|
| 152 |
+
(b) You must cause any modified files to carry prominent notices
|
| 153 |
+
stating that You changed the files; and
|
| 154 |
+
|
| 155 |
+
(c) You must retain, in the Source form of any Derivative Works
|
| 156 |
+
that You distribute, all copyright, patent, trademark, and
|
| 157 |
+
attribution notices from the Source form of the Work,
|
| 158 |
+
excluding those notices that do not pertain to any part of
|
| 159 |
+
the Derivative Works; and
|
| 160 |
+
|
| 161 |
+
(d) If the Work includes a "NOTICE" text file as part of its
|
| 162 |
+
distribution, then any Derivative Works that You distribute must
|
| 163 |
+
include a readable copy of the attribution notices contained
|
| 164 |
+
within such NOTICE file, excluding those notices that do not
|
| 165 |
+
pertain to any part of the Derivative Works, in at least one
|
| 166 |
+
of the following places: within a NOTICE text file distributed
|
| 167 |
+
as part of the Derivative Works; within the Source form or
|
| 168 |
+
documentation, if provided along with the Derivative Works; or,
|
| 169 |
+
within a display generated by the Derivative Works, if and
|
| 170 |
+
wherever such third-party notices normally appear. The contents
|
| 171 |
+
of the NOTICE file are for informational purposes only and
|
| 172 |
+
do not modify the License. You may add Your own attribution
|
| 173 |
+
notices within Derivative Works that You distribute, alongside
|
| 174 |
+
or as an addendum to the NOTICE text from the Work, provided
|
| 175 |
+
that such additional attribution notices cannot be construed
|
| 176 |
+
as modifying the License.
|
| 177 |
+
|
| 178 |
+
You may add Your own copyright statement to Your modifications and
|
| 179 |
+
may provide additional or different license terms and conditions
|
| 180 |
+
for use, reproduction, or distribution of Your modifications, or
|
| 181 |
+
for any such Derivative Works as a whole, provided Your use,
|
| 182 |
+
reproduction, and distribution of the Work otherwise complies with
|
| 183 |
+
the conditions stated in this License.
|
| 184 |
+
|
| 185 |
+
5. Submission of Contributions. Unless You explicitly state otherwise,
|
| 186 |
+
any Contribution intentionally submitted for inclusion in the Work
|
| 187 |
+
by You to the Licensor shall be under the terms and conditions of
|
| 188 |
+
this License, without any additional terms or conditions.
|
| 189 |
+
Notwithstanding the above, nothing herein shall supersede or modify
|
| 190 |
+
the terms of any separate license agreement you may have executed
|
| 191 |
+
with Licensor regarding such Contributions.
|
| 192 |
+
|
| 193 |
+
6. Trademarks. This License does not grant permission to use the trade
|
| 194 |
+
names, trademarks, service marks, or product names of the Licensor,
|
| 195 |
+
except as required for reasonable and customary use in describing the
|
| 196 |
+
origin of the Work and reproducing the content of the NOTICE file.
|
| 197 |
+
|
| 198 |
+
7. Disclaimer of Warranty. Unless required by applicable law or
|
| 199 |
+
agreed to in writing, Licensor provides the Work (and each
|
| 200 |
+
Contributor provides its Contributions) on an "AS IS" BASIS,
|
| 201 |
+
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
|
| 202 |
+
implied, including, without limitation, any warranties or conditions
|
| 203 |
+
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
|
| 204 |
+
PARTICULAR PURPOSE. You are solely responsible for determining the
|
| 205 |
+
appropriateness of using or redistributing the Work and assume any
|
| 206 |
+
risks associated with Your exercise of permissions under this License.
|
| 207 |
+
|
| 208 |
+
8. Limitation of Liability. In no event and under no legal theory,
|
| 209 |
+
whether in tort (including negligence), contract, or otherwise,
|
| 210 |
+
unless required by applicable law (such as deliberate and grossly
|
| 211 |
+
negligent acts) or agreed to in writing, shall any Contributor be
|
| 212 |
+
liable to You for damages, including any direct, indirect, special,
|
| 213 |
+
incidental, or consequential damages of any character arising as a
|
| 214 |
+
result of this License or out of the use or inability to use the
|
| 215 |
+
Work (including but not limited to damages for loss of goodwill,
|
| 216 |
+
work stoppage, computer failure or malfunction, or any and all
|
| 217 |
+
other commercial damages or losses), even if such Contributor
|
| 218 |
+
has been advised of the possibility of such damages.
|
| 219 |
+
|
| 220 |
+
9. Accepting Warranty or Additional Liability. While redistributing
|
| 221 |
+
the Work or Derivative Works thereof, You may choose to offer,
|
| 222 |
+
and charge a fee for, acceptance of support, warranty, indemnity,
|
| 223 |
+
or other liability obligations and/or rights consistent with this
|
| 224 |
+
License. However, in accepting such obligations, You may act only
|
| 225 |
+
on Your own behalf and on Your sole responsibility, not on behalf
|
| 226 |
+
of any other Contributor, and only if You agree to indemnify,
|
| 227 |
+
defend, and hold each Contributor harmless for any liability
|
| 228 |
+
incurred by, or claims asserted against, such Contributor by reason
|
| 229 |
+
of your accepting any such warranty or additional liability.
|
| 230 |
+
|
| 231 |
+
END OF TERMS AND CONDITIONS
|
| 232 |
+
|
| 233 |
+
APPENDIX: How to apply the Apache License to your work.
|
| 234 |
+
|
| 235 |
+
To apply the Apache License to your work, attach the following
|
| 236 |
+
boilerplate notice, with the fields enclosed by brackets "[]"
|
| 237 |
+
replaced with your own identifying information. (Don't include
|
| 238 |
+
the brackets!) The text should be enclosed in the appropriate
|
| 239 |
+
comment syntax for the file format. We also recommend that a
|
| 240 |
+
file or class name and description of purpose be included on the
|
| 241 |
+
same "printed page" as the copyright notice for easier
|
| 242 |
+
identification within third-party archives.
|
| 243 |
+
|
| 244 |
+
Copyright 2023 Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li
|
| 245 |
+
|
| 246 |
+
Licensed under the Apache License, Version 2.0 (the "License");
|
| 247 |
+
you may not use this file except in compliance with the License.
|
| 248 |
+
You may obtain a copy of the License at
|
| 249 |
+
|
| 250 |
+
http://www.apache.org/licenses/LICENSE-2.0
|
| 251 |
+
|
| 252 |
+
Unless required by applicable law or agreed to in writing, software
|
| 253 |
+
distributed under the License is distributed on an "AS IS" BASIS,
|
| 254 |
+
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 255 |
+
See the License for the specific language governing permissions and
|
| 256 |
+
limitations under the License.
|
| 257 |
+
|
| 258 |
+
------------- LICENSE FOR PanQiWei AutoGPTQ code --------------
|
| 259 |
+
|
| 260 |
+
MIT License
|
| 261 |
+
|
| 262 |
+
Copyright (c) 2023 潘其威(William)
|
| 263 |
+
|
| 264 |
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
| 265 |
+
of this software and associated documentation files (the "Software"), to deal
|
| 266 |
+
in the Software without restriction, including without limitation the rights
|
| 267 |
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
| 268 |
+
copies of the Software, and to permit persons to whom the Software is
|
| 269 |
+
furnished to do so, subject to the following conditions:
|
| 270 |
+
|
| 271 |
+
The above copyright notice and this permission notice shall be included in all
|
| 272 |
+
copies or substantial portions of the Software.
|
| 273 |
+
|
| 274 |
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
| 275 |
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
| 276 |
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
| 277 |
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
| 278 |
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
| 279 |
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
| 280 |
+
SOFTWARE.
|
README.md
ADDED
|
@@ -0,0 +1,463 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- zh
|
| 4 |
+
- en
|
| 5 |
+
tags:
|
| 6 |
+
- qwen
|
| 7 |
+
pipeline_tag: text-generation
|
| 8 |
+
inference: false
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
# Qwen-1.8B-Chat
|
| 12 |
+
|
| 13 |
+
<p align="center">
|
| 14 |
+
<img src="https://qianwen-res.oss-cn-beijing.aliyuncs.com/logo_qwen.jpg" width="400"/>
|
| 15 |
+
<p>
|
| 16 |
+
<br>
|
| 17 |
+
|
| 18 |
+
<p align="center">
|
| 19 |
+
🤗 <a href="https://huggingface.co/Qwen">Hugging Face</a>   |   🤖 <a href="https://modelscope.cn/organization/qwen">ModelScope</a>   |    📑 <a href="https://arxiv.org/abs/2309.16609">Paper</a>    ｜   🖥️ <a href="https://www.modelscope.cn/studios/qwen/Qwen-1_8B-Chat-Demo/summary">Demo</a>
|
| 20 |
+
<br>
|
| 21 |
+
<a href="assets/wechat.png">WeChat (微信)</a>   |   <a href="https://discord.gg/z3GAxXZ9Ce">Discord</a>   ｜   <a href="https://dashscope.aliyun.com">API</a>
|
| 22 |
+
</p>
|
| 23 |
+
<br>
|
| 24 |
+
|
| 25 |
+
## 介绍（Introduction）
|
| 26 |
+
**通义千问-1.8B（Qwen-1.8B）**是阿里云研发的通义千问大模型系列的18亿参数规模的模型。Qwen-1.8B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-1.8B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-1.8B-Chat。本仓库为Qwen-1.8B-Chat的仓库。
|
| 27 |
+
|
| 28 |
+
通义千问-1.8B（Qwen-1.8B）主要有以下特点：
|
| 29 |
+
1. **低成本部署**：提供int8和int4量化版本，推理最低仅需不到2GB显存，生成2048 tokens仅需3GB显存占用。微调最低仅需6GB。
|
| 30 |
+
2. **大规模高质量训练语料**：使用超过2.2万亿tokens的数据进行预训练，包含高质量中、英、多语言、代码、数学等数据，涵盖通用及专业领域的训练语料。通过大量对比实验对预训练语料分布进行了优化。
|
| 31 |
+
3. **优秀的性能**：Qwen-1.8B支持8192上下文长度，在多个中英文下游评测任务上（涵盖常识推理、代码、数学、翻译等），效果显著超越现有的相近规模开源模型，具体评测结果请详见下文。
|
| 32 |
+
4. **覆盖更全面的词表**：相比目前以中英词表为主的开源模型，Qwen-1.8B使用了约15万大小的词表。该词表对多语言更加友好，方便用户在不扩展词表的情况下对部分语种进行能力增强和扩展。
|
| 33 |
+
5. **系统指令跟随**：Qwen-1.8B-Chat可以通过调整系统指令，实现**角色扮演**，**语言风格迁移**，**任务设定**，和**行为设定**等能力。
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
如果您想了解更多关于通义千问1.8B开源模型的细节，我们建议您参阅[GitHub代码库](https://github.com/QwenLM/Qwen)。
|
| 37 |
+
|
| 38 |
+
**Qwen-1.8B** is the 1.8B-parameter version of the large language model series, Qwen (abbr. Tongyi Qianwen), proposed by Aibaba Cloud. Qwen-1.8B 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-1.8B, we release Qwen-1.8B-Chat, a large-model-based AI assistant, which is trained with alignment techniques. This repository is the one for Qwen-1.8B-Chat.
|
| 39 |
+
|
| 40 |
+
The features of Qwen-1.8B include:
|
| 41 |
+
1. **Low-cost deployment**: We provide int4 and int8 quantized versions, the minimum memory requirment for inference is less than 2GB, generating 2048 tokens only 3GB of memory usage. The minimum memory requirment of finetuning is only 6GB.
|
| 42 |
+
|
| 43 |
+
2. **Large-scale high-quality training corpora**: It is pretrained on over 2.2 trillion tokens, including Chinese, English, multilingual texts, code, and mathematics, covering general and professional fields. The distribution of the pre-training corpus has been optimized through a large number of ablation experiments.
|
| 44 |
+
3. **Good performance**: It supports 8192 context length and significantly surpasses existing open-source models of similar scale on multiple Chinese and English downstream evaluation tasks (including commonsense, reasoning, code, mathematics, etc.), and even surpasses some larger-scale models in several benchmarks. See below for specific evaluation results.
|
| 45 |
+
4. **More comprehensive vocabulary coverage**: Compared with other open-source models based on Chinese and English vocabularies, Qwen-1.8B uses a vocabulary of over 150K tokens. This vocabulary is more friendly to multiple languages, enabling users to directly further enhance the capability for certain languages without expanding the vocabulary.
|
| 46 |
+
5. **System prompt**: Qwen-1.8B-Chat can realize roly playing, language style transfer, task setting, and behavior setting by using system prompt.
|
| 47 |
+
|
| 48 |
+
For more details about the open-source model of Qwen-1.8B-Chat, please refer to the [GitHub](https://github.com/QwenLM/Qwen) code repository.
|
| 49 |
+
|
| 50 |
+
|
| 51 |
+
<br>
|
| 52 |
+
|
| 53 |
+
## 要求（Requirements）
|
| 54 |
+
|
| 55 |
+
* python 3.8及以上版本
|
| 56 |
+
* pytorch 1.12及以上版本，推荐2.0及以上版本
|
| 57 |
+
* 建议使用CUDA 11.4及以上（GPU用户、flash-attention用户等需考虑此选项）
|
| 58 |
+
* python 3.8 and above
|
| 59 |
+
* pytorch 1.12 and above, 2.0 and above are recommended
|
| 60 |
+
* CUDA 11.4 and above are recommended (this is for GPU users, flash-attention users, etc.)
|
| 61 |
+
|
| 62 |
+
## 依赖项（Dependency）
|
| 63 |
+
|
| 64 |
+
运行Qwen-1.8B-Chat，请确保满足上述要求，再执行以下pip命令安装依赖库
|
| 65 |
+
|
| 66 |
+
To run Qwen-1.8B-Chat, please make sure you meet the above requirements, and then execute the following pip commands to install the dependent libraries.
|
| 67 |
+
|
| 68 |
+
```bash
|
| 69 |
+
pip install transformers==4.32.0 accelerate tiktoken einops scipy transformers_stream_generator==0.0.4 peft deepspeed
|
| 70 |
+
```
|
| 71 |
+
|
| 72 |
+
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。
|
| 73 |
+
|
| 74 |
+
In addition, it is recommended to install the `flash-attention` library (**we support flash attention 2 now.**) for higher efficiency and lower memory usage.
|
| 75 |
+
|
| 76 |
+
```bash
|
| 77 |
+
git clone https://github.com/Dao-AILab/flash-attention
|
| 78 |
+
cd flash-attention && pip install .
|
| 79 |
+
# 下方安装可选，安装可能比较缓慢。
|
| 80 |
+
# pip install csrc/layer_norm
|
| 81 |
+
# pip install csrc/rotary
|
| 82 |
+
```
|
| 83 |
+
<br>
|
| 84 |
+
|
| 85 |
+
## 快速使用（Quickstart）
|
| 86 |
+
|
| 87 |
+
下面我们展示了一个使用Qwen-1.8B-Chat模型，进行多轮对话交互的样例：
|
| 88 |
+
|
| 89 |
+
We show an example of multi-turn interaction with Qwen-1.8B-Chat in the following code:
|
| 90 |
+
|
| 91 |
+
```python
|
| 92 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 93 |
+
from transformers.generation import GenerationConfig
|
| 94 |
+
|
| 95 |
+
# Note: The default behavior now has injection attack prevention off.
|
| 96 |
+
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen-1_8B-Chat", trust_remote_code=True)
|
| 97 |
+
|
| 98 |
+
# use bf16
|
| 99 |
+
# model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-1_8B-Chat", device_map="auto", trust_remote_code=True, bf16=True).eval()
|
| 100 |
+
# use fp16
|
| 101 |
+
# model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-1_8B-Chat", device_map="auto", trust_remote_code=True, fp16=True).eval()
|
| 102 |
+
# use cpu only
|
| 103 |
+
# model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-1_8B-Chat", device_map="cpu", trust_remote_code=True).eval()
|
| 104 |
+
# use auto mode, automatically select precision based on the device.
|
| 105 |
+
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen-1_8B-Chat", device_map="auto", trust_remote_code=True).eval()
|
| 106 |
+
|
| 107 |
+
# Specify hyperparameters for generation. But if you use transformers>=4.32.0, there is no need to do this.
|
| 108 |
+
# model.generation_config = GenerationConfig.from_pretrained("Qwen/Qwen-1_8B-Chat", trust_remote_code=True) # 可指定不同的生成长度、top_p等相关超参
|
| 109 |
+
|
| 110 |
+
# 第一轮对话 1st dialogue turn
|
| 111 |
+
response, history = model.chat(tokenizer, "你好", history=None)
|
| 112 |
+
print(response)
|
| 113 |
+
# 你好！很高兴为你提供帮助。
|
| 114 |
+
|
| 115 |
+
# 第二轮对话 2nd dialogue turn
|
| 116 |
+
response, history = model.chat(tokenizer, "给我讲一个年轻人奋斗创业最终取得成功的故事。", history=history)
|
| 117 |
+
print(response)
|
| 118 |
+
# 这是一个关于一个年轻人奋斗创业最终取得成功的故事。
|
| 119 |
+
# 故事的主人公叫李明，他来自一个普通的家庭，父母都是普通的工人。从小，李明就立下了一个目标：要成为一名成功的企业家。
|
| 120 |
+
# 为了实现这个目标，李明勤奋学习，考上了大学。在大学期间，他积极参加各种创业比赛，获得了不少奖项。他还利用课余时间去实习，积累了宝贵的经验。
|
| 121 |
+
# 毕业后，李明决定开始自己的创业之路。他开始寻找投资机会，但多次都被拒绝了。然而，他并没有放弃。他继续努力，不断改进自己的创业计划，并寻找新的投资机会。
|
| 122 |
+
# 最终，李明成功地获得了一笔投资，开始了自己的创业之路。他成立了一家科技公司，专注于开发新型软件。在他的领导下，公司迅速发展起来，成为了一家成功的科技企业。
|
| 123 |
+
# 李明的成功并不是偶然的。他勤奋、坚韧、勇于冒险，不断学习和改进自己。他的成功也证明了，只要努力奋斗，任何人都有可能取得成功。
|
| 124 |
+
|
| 125 |
+
# 第三轮对话 3rd dialogue turn
|
| 126 |
+
response, history = model.chat(tokenizer, "给这个故事起一个标题", history=history)
|
| 127 |
+
print(response)
|
| 128 |
+
# 《奋斗创业：一个年轻人的成功之路》
|
| 129 |
+
|
| 130 |
+
# Qwen-1.8B-Chat现在可以通过调整系统指令（System Prompt），实现角色扮演，语言风格迁移，任务设定，行为设定等能力。
|
| 131 |
+
# Qwen-1.8B-Chat can realize roly playing, language style transfer, task setting, and behavior setting by system prompt.
|
| 132 |
+
response, _ = model.chat(tokenizer, "你好呀", history=None, system="请用二次元可爱语气和我说话")
|
| 133 |
+
print(response)
|
| 134 |
+
# 你好啊！我是一只可爱的二次元猫咪哦，不知道你有什么问题需要我帮忙解答吗？
|
| 135 |
+
|
| 136 |
+
response, _ = model.chat(tokenizer, "My colleague works diligently", history=None, system="You will write beautiful compliments according to needs")
|
| 137 |
+
print(response)
|
| 138 |
+
# Your colleague is an outstanding worker! Their dedication and hard work are truly inspiring. They always go above and beyond to ensure that
|
| 139 |
+
# their tasks are completed on time and to the highest standard. I am lucky to have them as a colleague, and I know I can count on them to handle any challenge that comes their way.
|
| 140 |
+
```
|
| 141 |
+
|
| 142 |
+
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。
|
| 143 |
+
|
| 144 |
+
For more information, please refer to our [GitHub repo](https://github.com/QwenLM/Qwen) for more information.
|
| 145 |
+
|
| 146 |
+
## Tokenizer
|
| 147 |
+
|
| 148 |
+
> 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。
|
| 149 |
+
|
| 150 |
+
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)。
|
| 151 |
+
|
| 152 |
+
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).
|
| 153 |
+
|
| 154 |
+
## 量化 (Quantization)
|
| 155 |
+
|
| 156 |
+
### 用法 (Usage)
|
| 157 |
+
|
| 158 |
+
**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-1.8B-Chat的Int4量化模型[点击这里](https://huggingface.co/Qwen/Qwen-1_8B-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。**
|
| 159 |
+
|
| 160 |
+
**Note: we provide a new solution based on [AutoGPTQ](https://github.com/PanQiWei/AutoGPTQ), and release an Int4 quantized model for Qwen-1.8B-Chat [Click here](https://huggingface.co/Qwen/Qwen-1_8B-Chat-Int4), which achieves nearly lossless model effects but improved performance on both memory costs and inference speed, in comparison with the previous solution.**
|
| 161 |
+
|
| 162 |
+
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：
|
| 163 |
+
|
| 164 |
+
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:
|
| 165 |
+
|
| 166 |
+
```bash
|
| 167 |
+
pip install auto-gptq optimum
|
| 168 |
+
```
|
| 169 |
+
|
| 170 |
+
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。
|
| 171 |
+
|
| 172 |
+
随后即可使用和上述一致的用法调用量化模型：
|
| 173 |
+
|
| 174 |
+
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.
|
| 175 |
+
|
| 176 |
+
Then you can load the quantized model easily and run inference as same as usual:
|
| 177 |
+
|
| 178 |
+
```python
|
| 179 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 180 |
+
"Qwen/Qwen-1_8B-Chat-Int4",
|
| 181 |
+
device_map="auto",
|
| 182 |
+
trust_remote_code=True
|
| 183 |
+
).eval()
|
| 184 |
+
response, history = model.chat(tokenizer, "你好", history=None)
|
| 185 |
+
```
|
| 186 |
+
|
| 187 |
+
### 效果评测
|
| 188 |
+
|
| 189 |
+
我们使用原始模型的FP32和BF16精度，以及量化过的Int8和Int4模型在基准评测上做了测试，结果如下所示：
|
| 190 |
+
|
| 191 |
+
We illustrate the model performance of both FP32, BF16, Int8 and Int4 models on the benchmark. Results are shown below:
|
| 192 |
+
|
| 193 |
+
| Quantization | MMLU | CEval (val) | GSM8K | Humaneval |
|
| 194 |
+
|--------------|:----:|:-----------:|:-----:|:---------:|
|
| 195 |
+
| FP32 | 43.4 | 57.0 | 33.0 | 26.8 |
|
| 196 |
+
| BF16 | 43.3 | 55.6 | 33.7 | 26.2 |
|
| 197 |
+
| Int8 | 43.1 | 55.8 | 33.0 | 27.4 |
|
| 198 |
+
| Int4 | 42.9 | 52.8 | 31.2 | 25.0 |
|
| 199 |
+
|
| 200 |
+
### 推理速度 (Inference Speed)
|
| 201 |
+
|
| 202 |
+
我们测算了FP32、BF16精度和Int8、Int4量化模型生成2048和8192个token的平均推理速度。如图所示：
|
| 203 |
+
|
| 204 |
+
We measured the average inference speed of generating 2048 and 8192 tokens under FP32, BF16 precision and Int8, Int4 quantization level, respectively.
|
| 205 |
+
|
| 206 |
+
| Quantization | FlashAttn | Speed (2048 tokens) | Speed (8192 tokens) |
|
| 207 |
+
|--------------| :-------: |:-------------------:|:-------------------:|
|
| 208 |
+
| FP32 | v2 | 52.96 | 47.35 |
|
| 209 |
+
| BF16 | v2 | 54.09 | 54.04 |
|
| 210 |
+
| Int8 | v2 | 55.56 | 55.62 |
|
| 211 |
+
| Int4 | v2 | 71.07 | 76.45 |
|
| 212 |
+
| FP32 | v1 | 52.00 | 45.80 |
|
| 213 |
+
| BF16 | v1 | 51.70 | 55.04 |
|
| 214 |
+
| Int8 | v1 | 53.16 | 53.33 |
|
| 215 |
+
| Int4 | v1 | 69.82 | 67.44 |
|
| 216 |
+
| FP32 | Disabled | 52.28 | 44.95 |
|
| 217 |
+
| BF16 | Disabled | 48.17 | 45.01 |
|
| 218 |
+
| Int8 | Disabled | 52.16 | 52.99 |
|
| 219 |
+
| Int4 | Disabled | 68.37 | 65.94 |
|
| 220 |
+
|
| 221 |
+
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 11.4。推理速度是生成8192个token的速度均值。
|
| 222 |
+
|
| 223 |
+
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 11.4. The inference speed is averaged over the generated 8192 tokens.
|
| 224 |
+
|
| 225 |
+
### 显存使用 (GPU Memory Usage)
|
| 226 |
+
|
| 227 |
+
我们测算了FP32、BF16精度和Int8、Int4量化模型生成2048个及8192个token（单个token作为输入）的峰值显存占用情况。结果如下所示：
|
| 228 |
+
|
| 229 |
+
We also profile the peak GPU memory usage for generating 2048 tokens and 8192 tokens (with single token as context) under FP32, BF16 or Int8, Int4 quantization level, respectively. The results are shown below.
|
| 230 |
+
|
| 231 |
+
| Quantization Level | Peak Usage for Encoding 2048 Tokens | Peak Usage for Generating 8192 Tokens |
|
| 232 |
+
|--------------------|:-----------------------------------:|:-------------------------------------:|
|
| 233 |
+
| FP32 | 8.45GB | 13.06GB |
|
| 234 |
+
| BF16 | 4.23GB | 6.48GB |
|
| 235 |
+
| Int8 | 3.48GB | 5.34GB |
|
| 236 |
+
| Int4 | 2.91GB | 4.80GB |
|
| 237 |
+
|
| 238 |
+
上述性能测算使用[此脚本](https://qianwen-res.oss-cn-beijing.aliyuncs.com/profile.py)完成。
|
| 239 |
+
|
| 240 |
+
The above speed and memory profiling are conducted using [this script](https://qianwen-res.oss-cn-beijing.aliyuncs.com/profile.py).
|
| 241 |
+
<br>
|
| 242 |
+
|
| 243 |
+
## 模型细节（Model）
|
| 244 |
+
|
| 245 |
+
与Qwen-1.8B预训练模型相同，Qwen-1.8B-Chat模型规模基本情况如下所示
|
| 246 |
+
|
| 247 |
+
The details of the model architecture of Qwen-1.8B-Chat are listed as follows
|
| 248 |
+
|
| 249 |
+
| Hyperparameter | Value |
|
| 250 |
+
|:----------------|:------:|
|
| 251 |
+
| n_layers | 24 |
|
| 252 |
+
| n_heads | 16 |
|
| 253 |
+
| d_model | 2048 |
|
| 254 |
+
| vocab size | 151851 |
|
| 255 |
+
| sequence length | 8192 |
|
| 256 |
+
|
| 257 |
+
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，
|
| 258 |
+
即RoPE相对位置编码、SwiGLU激活函数、RMSNorm（可选安装flash-attention加速）。
|
| 259 |
+
|
| 260 |
+
在分词器方面，相比目前主流开源模型以中英词表为主，Qwen-1.8B-Chat使用了约15万token大小的词表。
|
| 261 |
+
该词表在GPT-4使用的BPE词表`cl100k_base`基础上，对中文、多语言进行了优化，在对中、英、代码数据的高效编解码的基础上，对部分多语言更加友好，方便用户在不扩展词表的情况下对部分语种进行能力增强。
|
| 262 |
+
词表对数字按单个数字位切分。调用较为高效的[tiktoken分词库](https://github.com/openai/tiktoken)进行分词。
|
| 263 |
+
|
| 264 |
+
For position encoding, FFN activation function, and normalization calculation methods, we adopt the prevalent practices, i.e., RoPE relative position encoding, SwiGLU for activation function, and RMSNorm for normalization (optional installation of flash-attention for acceleration).
|
| 265 |
+
|
| 266 |
+
For tokenization, compared to the current mainstream open-source models based on Chinese and English vocabularies, Qwen-1.8B-Chat uses a vocabulary of over 150K tokens.
|
| 267 |
+
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.
|
| 268 |
+
It segments numbers by single digit, and calls the [tiktoken](https://github.com/openai/tiktoken) tokenizer library for efficient tokenization.
|
| 269 |
+
|
| 270 |
+
## 评测效果（Evaluation）
|
| 271 |
+
|
| 272 |
+
对于Qwen-1.8B-Chat模型，我们同样评测了常规的中文理解（C-Eval）、英文理解（MMLU）、代码（HumanEval）和数学（GSM8K）等权威任务，同时包含了长序列任务的评测结果。由于Qwen-1.8B-Chat模型经过对齐后，激发了较强的外部系统调用能力，我们还进行了工具使用能力方面的评测。
|
| 273 |
+
|
| 274 |
+
提示：由于硬件和框架造成的舍入误差，复现结果如有波动属于正常现象。
|
| 275 |
+
|
| 276 |
+
For Qwen-1.8B-Chat, we also evaluate the model on C-Eval, MMLU, HumanEval, GSM8K, etc., as well as the benchmark evaluation for long-context understanding, and tool usage.
|
| 277 |
+
|
| 278 |
+
Note: Due to rounding errors caused by hardware and framework, differences in reproduced results are possible.
|
| 279 |
+
|
| 280 |
+
### 中文评测（Chinese Evaluation）
|
| 281 |
+
|
| 282 |
+
#### C-Eval
|
| 283 |
+
|
| 284 |
+
在[C-Eval](https://arxiv.org/abs/2305.08322)验证集上，我们评价了Qwen-1.8B-Chat模型的zero-shot准确率
|
| 285 |
+
|
| 286 |
+
We demonstrate the zero-shot accuracy of Qwen-1.8B-Chat on C-Eval validation set
|
| 287 |
+
|
| 288 |
+
| Model | Avg. Acc. |
|
| 289 |
+
|:------------------------:|:---------:|
|
| 290 |
+
| **Qwen-7B-Chat** | 54.2 |
|
| 291 |
+
| InternLM-7B-Chat | 53.2 |
|
| 292 |
+
| **Qwen-1.8B-Chat** | 55.6 |
|
| 293 |
+
| ChatGLM2-6B-Chat | 50.7 |
|
| 294 |
+
| Baichuan-13B-Chat | 50.4 |
|
| 295 |
+
| Chinese-Alpaca-Plus-13B | 43.3 |
|
| 296 |
+
| Chinese-Alpaca-2-7B | 41.3 |
|
| 297 |
+
| LLaMA2-13B-Chat | 40.6 |
|
| 298 |
+
| LLaMA2-7B-Chat | 31.9 |
|
| 299 |
+
| OpenLLaMA-Chinese-3B | 24.4 |
|
| 300 |
+
| Firefly-Bloom-1B4 | 23.6 |
|
| 301 |
+
| OpenBuddy-3B | 23.5 |
|
| 302 |
+
| RedPajama-INCITE-Chat-3B | 18.3 |
|
| 303 |
+
|
| 304 |
+
C-Eval测试集上，Qwen-1.8B-Chat模型的zero-shot准确率结果如下：
|
| 305 |
+
|
| 306 |
+
The zero-shot accuracy of Qwen-1.8B-Chat on C-Eval testing set is provided below:
|
| 307 |
+
|
| 308 |
+
| Model | Avg. | STEM | Social Sciences | Humanities | Others |
|
| 309 |
+
| :---------------------: | :------: | :--: | :-------------: | :--------: | :----: |
|
| 310 |
+
| **Qwen-7B-Chat** | 54.6 | 47.8 | 67.6 | 59.3 | 50.6 |
|
| 311 |
+
| Baichuan-13B-Chat | 51.5 | 43.7 | 64.6 | 56.2 | 49.2 |
|
| 312 |
+
| ChatGLM2-6B-Chat | 50.1 | 46.4 | 60.4 | 50.6 | 46.9 |
|
| 313 |
+
| **Qwen-1.8B-Chat** | 53.8 | 48.4 | 68.0 | 56.5 | 48.3 |
|
| 314 |
+
| Chinese-Alpaca-Plus-13B | 41.5 | 36.6 | 49.7 | 43.1 | 41.2 |
|
| 315 |
+
| Chinese-Alpaca-2-7B | 40.3 | - | - | - | - |
|
| 316 |
+
|
| 317 |
+
### 英文评测（English Evaluation）
|
| 318 |
+
|
| 319 |
+
#### MMLU
|
| 320 |
+
|
| 321 |
+
[MMLU](https://arxiv.org/abs/2009.03300)评测集上，Qwen-1.8B-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。
|
| 322 |
+
|
| 323 |
+
The zero-shot accuracy of Qwen-1.8B-Chat on MMLU is provided below.
|
| 324 |
+
The performance of Qwen-1.8B-Chat still on the top between other human-aligned models with comparable size.
|
| 325 |
+
|
| 326 |
+
| Model | Avg. Acc. |
|
| 327 |
+
|:------------------------:|:---------:|
|
| 328 |
+
| **Qwen-7B-Chat** | 53.9 |
|
| 329 |
+
| ChatGLM2-12B-Chat | 52.1 |
|
| 330 |
+
| Baichuan-13B-Chat | 52.1 |
|
| 331 |
+
| InternLM-7B-Chat | 50.8 |
|
| 332 |
+
| LLaMA2-7B-Chat | 47.0 |
|
| 333 |
+
| ChatGLM2-6B-Chat | 45.5 |
|
| 334 |
+
| **Qwen-1.8B-Chat** | 43.3 |
|
| 335 |
+
| OpenLLaMA-Chinese-3B | 25.7 |
|
| 336 |
+
| OpenBuddy-3B | 25.5 |
|
| 337 |
+
| RedPajama-INCITE-Chat-3B | 25.5 |
|
| 338 |
+
| Firefly-Bloom-1B4 | 23.8 |
|
| 339 |
+
|
| 340 |
+
### 代码评测（Coding Evaluation）
|
| 341 |
+
|
| 342 |
+
Qwen-1.8B-Chat在[HumanEval](https://github.com/openai/human-eval)的zero-shot Pass@1效果如下
|
| 343 |
+
|
| 344 |
+
The zero-shot Pass@1 of Qwen-1.8B-Chat on [HumanEval](https://github.com/openai/human-eval) is demonstrated below
|
| 345 |
+
|
| 346 |
+
| Model | Pass@1 |
|
| 347 |
+
|:------------------------:|:------:|
|
| 348 |
+
| **Qwen-7B-Chat** | 24.4 |
|
| 349 |
+
| LLaMA2-13B-Chat | 18.9 |
|
| 350 |
+
| Baichuan-13B-Chat | 16.5 |
|
| 351 |
+
| InternLM-7B-Chat | 14.0 |
|
| 352 |
+
| LLaMA2-7B-Chat | 12.2 |
|
| 353 |
+
| **Qwen-1.8B-Chat** | 26.2 |
|
| 354 |
+
| OpenBuddy-3B | 10.4 |
|
| 355 |
+
| RedPajama-INCITE-Chat-3B | 6.1 |
|
| 356 |
+
| OpenLLaMA-Chinese-3B | 4.9 |
|
| 357 |
+
| Firefly-Bloom-1B4 | 0.6 |
|
| 358 |
+
|
| 359 |
+
### 数学评测（Mathematics Evaluation）
|
| 360 |
+
|
| 361 |
+
在评测数学能力的[GSM8K](https://github.com/openai/grade-school-math)上，Qwen-1.8B-Chat的准确率结果如下
|
| 362 |
+
|
| 363 |
+
The accuracy of Qwen-1.8B-Chat on GSM8K is shown below
|
| 364 |
+
|
| 365 |
+
| Model | Zero-shot Acc. | 4-shot Acc. |
|
| 366 |
+
|:------------------------:|:--------------:|:-----------:|
|
| 367 |
+
| **Qwen-7B-Chat** | 41.1 | 43.5 |
|
| 368 |
+
| ChatGLM2-12B-Chat | - | 38.1 |
|
| 369 |
+
| Baichuan-13B-Chat | - | 36.3 |
|
| 370 |
+
| InternLM-7B-Chat | 32.6 | 34.5 |
|
| 371 |
+
| LLaMA2-13B-Chat | 29.4 | 36.7 |
|
| 372 |
+
| **Qwen-1.8B-Chat** | 33.7 | 30.2 |
|
| 373 |
+
| LLaMA2-7B-Chat | 20.4 | 28.2 |
|
| 374 |
+
| ChatGLM2-6B-Chat | - | 28.0 |
|
| 375 |
+
| OpenBuddy-3B | 10.6 | 12.6 |
|
| 376 |
+
| OpenLLaMA-Chinese-3B | 2.6 | 3.0 |
|
| 377 |
+
| RedPajama-INCITE-Chat-3B | 2.5 | 2.5 |
|
| 378 |
+
| Firefly-Bloom-1B4 | 2.4 | 1.8 |
|
| 379 |
+
|
| 380 |
+
### 工具使用能力的评测（Tool Usage）
|
| 381 |
+
|
| 382 |
+
#### ReAct Prompting
|
| 383 |
+
|
| 384 |
+
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：
|
| 385 |
+
|
| 386 |
+
Qwen-1.8B-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-1.8B-Chat's performance is as follows:
|
| 387 |
+
|
| 388 |
+
| Model | Tool Selection (Acc.↑) | Tool Input (Rouge-L↑) | False Positive Error”↓ |
|
| 389 |
+
|:------------------:|:----------------------:|:---------------------:|:----------------------:|
|
| 390 |
+
| GPT-4 | 95% | **0.90** | 15% |
|
| 391 |
+
| GPT-3.5 | 85% | 0.88 | 75% |
|
| 392 |
+
| **Qwen-7B-Chat** | **99%** | 0.89 | **9.7%** |
|
| 393 |
+
| **Qwen-1.8B-Chat** | 92% | 0.89 | 19.3% |
|
| 394 |
+
|
| 395 |
+
> 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。
|
| 396 |
+
|
| 397 |
+
> The plugins that appear in the evaluation set do not appear in the training set of Qwen-1.8B-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.
|
| 398 |
+
|
| 399 |
+
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：
|
| 400 |
+
|
| 401 |
+
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:
|
| 402 |
+
|
| 403 |
+
|
| 404 |
+
|
| 405 |
+
|
| 406 |
+
#### Huggingface Agent
|
| 407 |
+
|
| 408 |
+
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：
|
| 409 |
+
|
| 410 |
+
Qwen-1.8B-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:
|
| 411 |
+
|
| 412 |
+
| Model | Tool Selection↑ | Tool Used↑ | Code↑ |
|
| 413 |
+
|:------------------:|:---------------:|:----------:|:---------:|
|
| 414 |
+
| GPT-4 | **100** | **100** | **97.41** |
|
| 415 |
+
| GPT-3.5 | 95.37 | 96.30 | 87.04 |
|
| 416 |
+
| StarCoder-15.5B | 87.04 | 87.96 | 68.89 |
|
| 417 |
+
| **Qwen-7B-chat** | 90.74 | 92.59 | 74.07 |
|
| 418 |
+
| **Qwen-1.8B-chat** | 85.16 | 85.19 | 61.11 |
|
| 419 |
+
<br>
|
| 420 |
+
|
| 421 |
+
## 评测复现（Reproduction）
|
| 422 |
+
|
| 423 |
+
我们提供了评测脚本，方便大家复现模型效果，详见[链接](https://github.com/QwenLM/Qwen/tree/main/eval)。提示：由于硬件和框架造成的舍入误差，复现结果如有小幅波动属于正常现象。
|
| 424 |
+
|
| 425 |
+
We have provided evaluation scripts to reproduce the performance of our model, details as [link](https://github.com/QwenLM/Qwen/tree/main/eval).
|
| 426 |
+
<br>
|
| 427 |
+
|
| 428 |
+
## FAQ
|
| 429 |
+
|
| 430 |
+
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。
|
| 431 |
+
|
| 432 |
+
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.
|
| 433 |
+
<br>
|
| 434 |
+
|
| 435 |
+
## 引用 (Citation)
|
| 436 |
+
|
| 437 |
+
如果你觉得我们的工作对你有帮助，欢迎引用！
|
| 438 |
+
|
| 439 |
+
If you find our work helpful, feel free to give us a cite.
|
| 440 |
+
|
| 441 |
+
```
|
| 442 |
+
@article{qwen,
|
| 443 |
+
title={Qwen Technical Report},
|
| 444 |
+
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},
|
| 445 |
+
journal={arXiv preprint arXiv:2309.16609},
|
| 446 |
+
year={2023}
|
| 447 |
+
}
|
| 448 |
+
```
|
| 449 |
+
<br>
|
| 450 |
+
|
| 451 |
+
## 使用协议（License Agreement）
|
| 452 |
+
|
| 453 |
+
我们的代码和模型权重对学术研究完全开放。请查看[LICENSE](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20RESEARCH%20LICENSE%20AGREEMENT)文件了解具体的开源协议细节。如需商用，请联系我们。
|
| 454 |
+
|
| 455 |
+
Our code and checkpoints are open to research purpose. Check the [LICENSE](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20RESEARCH%20LICENSE%20AGREEMENT) for more details about the license. For commercial use, please contact us.
|
| 456 |
+
<br>
|
| 457 |
+
|
| 458 |
+
## 联系我们（Contact Us）
|
| 459 |
+
|
| 460 |
+
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。
|
| 461 |
+
|
| 462 |
+
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.
|
| 463 |
+
|
assets/logo.jpg
ADDED
|
assets/qwen_tokenizer.png
ADDED
|
assets/react_showcase_001.png
ADDED
|
assets/react_showcase_002.png
ADDED
|
assets/wechat.png
ADDED
|
cache_autogptq_cuda_256.cpp
ADDED
|
@@ -0,0 +1,198 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#include <torch/all.h>
|
| 2 |
+
#include <torch/python.h>
|
| 3 |
+
#include <c10/cuda/CUDAGuard.h>
|
| 4 |
+
|
| 5 |
+
// adapted from https://github.com/PanQiWei/AutoGPTQ/blob/main/autogptq_extension/cuda_256/autogptq_cuda_256.cpp
|
| 6 |
+
void vecquant8matmul_cuda(
|
| 7 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 8 |
+
torch::Tensor scales, torch::Tensor zeros,
|
| 9 |
+
torch::Tensor g_idx
|
| 10 |
+
);
|
| 11 |
+
|
| 12 |
+
void vecquant8matmul(
|
| 13 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 14 |
+
torch::Tensor scales, torch::Tensor zeros,
|
| 15 |
+
torch::Tensor g_idx
|
| 16 |
+
) {
|
| 17 |
+
const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
|
| 18 |
+
vecquant8matmul_cuda(vec, mat, mul, scales, zeros, g_idx);
|
| 19 |
+
}
|
| 20 |
+
|
| 21 |
+
void vecquant8matmul_batched_cuda(
|
| 22 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 23 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 24 |
+
);
|
| 25 |
+
|
| 26 |
+
void vecquant8matmul_batched(
|
| 27 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 28 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 29 |
+
) {
|
| 30 |
+
const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
|
| 31 |
+
vecquant8matmul_batched_cuda(vec, mat, mul, scales, zeros);
|
| 32 |
+
}
|
| 33 |
+
|
| 34 |
+
void vecquant8matmul_batched_column_compression_cuda(
|
| 35 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 36 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 37 |
+
);
|
| 38 |
+
|
| 39 |
+
void vecquant8matmul_batched_column_compression(
|
| 40 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 41 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 42 |
+
) {
|
| 43 |
+
const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
|
| 44 |
+
vecquant8matmul_batched_column_compression_cuda(vec, mat, mul, scales, zeros);
|
| 45 |
+
}
|
| 46 |
+
|
| 47 |
+
void vecquant4matmul_batched_cuda(
|
| 48 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 49 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 50 |
+
);
|
| 51 |
+
|
| 52 |
+
void vecquant4matmul_batched(
|
| 53 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 54 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 55 |
+
) {
|
| 56 |
+
const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
|
| 57 |
+
vecquant4matmul_batched_cuda(vec, mat, mul, scales, zeros);
|
| 58 |
+
}
|
| 59 |
+
|
| 60 |
+
void vecquant4matmul_batched_column_compression_cuda(
|
| 61 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 62 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 63 |
+
);
|
| 64 |
+
|
| 65 |
+
void vecquant4matmul_batched_column_compression(
|
| 66 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 67 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 68 |
+
) {
|
| 69 |
+
const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
|
| 70 |
+
vecquant4matmul_batched_column_compression_cuda(vec, mat, mul, scales, zeros);
|
| 71 |
+
}
|
| 72 |
+
|
| 73 |
+
void vecquant8matmul_batched_old_cuda(
|
| 74 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 75 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 76 |
+
);
|
| 77 |
+
|
| 78 |
+
void vecquant8matmul_batched_old(
|
| 79 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 80 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 81 |
+
) {
|
| 82 |
+
const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
|
| 83 |
+
vecquant8matmul_batched_old_cuda(vec, mat, mul, scales, zeros);
|
| 84 |
+
}
|
| 85 |
+
|
| 86 |
+
|
| 87 |
+
void vecquant4matmul_batched_old_cuda(
|
| 88 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 89 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 90 |
+
);
|
| 91 |
+
|
| 92 |
+
void vecquant4matmul_batched_old(
|
| 93 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 94 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 95 |
+
) {
|
| 96 |
+
const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
|
| 97 |
+
vecquant4matmul_batched_old_cuda(vec, mat, mul, scales, zeros);
|
| 98 |
+
}
|
| 99 |
+
|
| 100 |
+
void vecquant8matmul_batched_column_compression_old_cuda(
|
| 101 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 102 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 103 |
+
);
|
| 104 |
+
|
| 105 |
+
void vecquant8matmul_batched_column_compression_old(
|
| 106 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 107 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 108 |
+
) {
|
| 109 |
+
const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
|
| 110 |
+
vecquant8matmul_batched_column_compression_old_cuda(vec, mat, mul, scales, zeros);
|
| 111 |
+
}
|
| 112 |
+
|
| 113 |
+
void vecquant4matmul_batched_column_compression_old_cuda(
|
| 114 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 115 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 116 |
+
);
|
| 117 |
+
|
| 118 |
+
void vecquant4matmul_batched_column_compression_old(
|
| 119 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 120 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 121 |
+
) {
|
| 122 |
+
const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
|
| 123 |
+
vecquant4matmul_batched_column_compression_old_cuda(vec, mat, mul, scales, zeros);
|
| 124 |
+
}
|
| 125 |
+
|
| 126 |
+
|
| 127 |
+
|
| 128 |
+
void vecquant8matmul_batched_faster_cuda(
|
| 129 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 130 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 131 |
+
);
|
| 132 |
+
|
| 133 |
+
void vecquant8matmul_batched_faster(
|
| 134 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 135 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 136 |
+
) {
|
| 137 |
+
const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
|
| 138 |
+
vecquant8matmul_batched_faster_cuda(vec, mat, mul, scales, zeros);
|
| 139 |
+
}
|
| 140 |
+
|
| 141 |
+
|
| 142 |
+
void vecquant8matmul_batched_faster_old_cuda(
|
| 143 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 144 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 145 |
+
);
|
| 146 |
+
|
| 147 |
+
void vecquant8matmul_batched_faster_old(
|
| 148 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 149 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 150 |
+
) {
|
| 151 |
+
const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
|
| 152 |
+
vecquant8matmul_batched_faster_old_cuda(vec, mat, mul, scales, zeros);
|
| 153 |
+
}
|
| 154 |
+
|
| 155 |
+
void vecquant8matmul_batched_column_compression_faster_cuda(
|
| 156 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 157 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 158 |
+
);
|
| 159 |
+
|
| 160 |
+
void vecquant8matmul_batched_column_compression_faster(
|
| 161 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 162 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 163 |
+
) {
|
| 164 |
+
const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
|
| 165 |
+
vecquant8matmul_batched_column_compression_faster_cuda(vec, mat, mul, scales, zeros);
|
| 166 |
+
}
|
| 167 |
+
|
| 168 |
+
|
| 169 |
+
void vecquant8matmul_batched_column_compression_faster_old_cuda(
|
| 170 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 171 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 172 |
+
);
|
| 173 |
+
|
| 174 |
+
void vecquant8matmul_batched_column_compression_faster_old(
|
| 175 |
+
torch::Tensor vec, torch::Tensor mat, torch::Tensor mul,
|
| 176 |
+
torch::Tensor scales, torch::Tensor zeros
|
| 177 |
+
) {
|
| 178 |
+
const at::cuda::OptionalCUDAGuard device_guard(device_of(vec));
|
| 179 |
+
vecquant8matmul_batched_column_compression_faster_old_cuda(vec, mat, mul, scales, zeros);
|
| 180 |
+
}
|
| 181 |
+
|
| 182 |
+
|
| 183 |
+
|
| 184 |
+
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
|
| 185 |
+
m.def("vecquant8matmul", &vecquant8matmul, "Vector 8-bit Quantized Matrix Multiplication (CUDA) (desc_act)");
|
| 186 |
+
m.def("vecquant8matmul_batched", &vecquant8matmul_batched, "Vector 8-bit Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
|
| 187 |
+
m.def("vecquant8matmul_batched_old", &vecquant8matmul_batched_old, "Vector 8-bit old Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
|
| 188 |
+
m.def("vecquant8matmul_batched_faster", &vecquant8matmul_batched_faster, "Vector 8-bit old Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
|
| 189 |
+
m.def("vecquant8matmul_batched_faster_old", &vecquant8matmul_batched_faster_old, "Vector 8-bit old Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
|
| 190 |
+
m.def("vecquant4matmul_batched_old", &vecquant4matmul_batched_old, "Vector 4-bit old Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
|
| 191 |
+
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)");
|
| 192 |
+
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)");
|
| 193 |
+
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)");
|
| 194 |
+
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)");
|
| 195 |
+
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)");
|
| 196 |
+
m.def("vecquant4matmul_batched", &vecquant4matmul_batched, "Vector 4-bit Batched Quantized Matrix Multiplication (CUDA) (desc_act)");
|
| 197 |
+
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)");
|
| 198 |
+
}
|
cache_autogptq_cuda_kernel_256.cu
ADDED
|
@@ -0,0 +1,1708 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#define _CRT_SECURE_NO_WARNINGS
|
| 2 |
+
#include <torch/all.h>
|
| 3 |
+
#include <torch/python.h>
|
| 4 |
+
#include <cuda.h>
|
| 5 |
+
#include <cuda_runtime.h>
|
| 6 |
+
#include <cuda_fp16.h>
|
| 7 |
+
#include <stdint.h>
|
| 8 |
+
|
| 9 |
+
#if (defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 700) || defined(USE_ROCM)
|
| 10 |
+
// adapted from https://github.com/PanQiWei/AutoGPTQ/blob/main/autogptq_extension/cuda_256/autogptq_cuda_kernel_256.cu
|
| 11 |
+
__device__ __forceinline__ void atomicAdd(c10::Half* address, c10::Half val) {
|
| 12 |
+
unsigned int *address_as_ui = reinterpret_cast<unsigned int *>(reinterpret_cast<char *>(address) - (reinterpret_cast<size_t>(address) & 2));
|
| 13 |
+
unsigned int old = *address_as_ui;
|
| 14 |
+
unsigned int assumed;
|
| 15 |
+
|
| 16 |
+
do {
|
| 17 |
+
assumed = old;
|
| 18 |
+
unsigned short hsum = reinterpret_cast<size_t>(address) & 2 ? (old >> 16) : (old & 0xffff);
|
| 19 |
+
hsum += val;
|
| 20 |
+
old = reinterpret_cast<size_t>(address) & 2
|
| 21 |
+
? (old & 0xffff) | (hsum << 16)
|
| 22 |
+
: (old & 0xffff0000) | hsum;
|
| 23 |
+
old = atomicCAS(address_as_ui, assumed, old);
|
| 24 |
+
|
| 25 |
+
// Note: uses integer comparison to avoid hang in case of NaN (since NaN != NaN)
|
| 26 |
+
} while (assumed != old);
|
| 27 |
+
}
|
| 28 |
+
__device__ __forceinline__ void atomicAdd(__half* address, c10::Half val) {
|
| 29 |
+
unsigned int * address_as_ui = (unsigned int *) ((char *)address - ((size_t)address & 2));
|
| 30 |
+
unsigned int old = *address_as_ui;
|
| 31 |
+
unsigned int assumed;
|
| 32 |
+
|
| 33 |
+
do {
|
| 34 |
+
assumed = old;
|
| 35 |
+
__half_raw hsum;
|
| 36 |
+
hsum.x = (size_t)address & 2 ? (old >> 16) : (old & 0xffff);
|
| 37 |
+
half tmpres = __hadd(hsum, val);
|
| 38 |
+
hsum = __half_raw(tmpres);
|
| 39 |
+
old = (size_t)address & 2 ? (old & 0xffff) | (hsum.x << 16) : (old & 0xffff0000) | hsum.x;
|
| 40 |
+
old = atomicCAS(address_as_ui, assumed, old);
|
| 41 |
+
} while (assumed != old);
|
| 42 |
+
}
|
| 43 |
+
#endif
|
| 44 |
+
|
| 45 |
+
template <typename scalar_t>
|
| 46 |
+
__global__ void VecQuant8MatMulKernel(
|
| 47 |
+
const scalar_t* __restrict__ vec,
|
| 48 |
+
const int* __restrict__ mat,
|
| 49 |
+
scalar_t* __restrict__ mul,
|
| 50 |
+
const scalar_t* __restrict__ scales,
|
| 51 |
+
const int* __restrict__ zeros,
|
| 52 |
+
const int* __restrict__ g_idx,
|
| 53 |
+
int batch,
|
| 54 |
+
int vec_height,
|
| 55 |
+
int height,
|
| 56 |
+
int width,
|
| 57 |
+
int zero_width
|
| 58 |
+
);
|
| 59 |
+
|
| 60 |
+
template <typename scalar_t>
|
| 61 |
+
__global__ void VecQuant8BatchMatMulColumnCompressionKernel(
|
| 62 |
+
const scalar_t* __restrict__ vec,
|
| 63 |
+
const int* __restrict__ mat,
|
| 64 |
+
scalar_t* __restrict__ mul,
|
| 65 |
+
const scalar_t* __restrict__ scales,
|
| 66 |
+
const int* __restrict__ zeros,
|
| 67 |
+
int batch,
|
| 68 |
+
int heads,
|
| 69 |
+
int vec_row,
|
| 70 |
+
int height,
|
| 71 |
+
int width
|
| 72 |
+
);
|
| 73 |
+
|
| 74 |
+
template <typename scalar_t>
|
| 75 |
+
__global__ void VecQuant4BatchMatMulColumnCompressionKernel(
|
| 76 |
+
const scalar_t* __restrict__ vec,
|
| 77 |
+
const int* __restrict__ mat,
|
| 78 |
+
scalar_t* __restrict__ mul,
|
| 79 |
+
const scalar_t* __restrict__ scales,
|
| 80 |
+
const int* __restrict__ zeros,
|
| 81 |
+
int batch,
|
| 82 |
+
int heads,
|
| 83 |
+
int vec_row,
|
| 84 |
+
int height,
|
| 85 |
+
int width
|
| 86 |
+
);
|
| 87 |
+
|
| 88 |
+
template <typename scalar_t>
|
| 89 |
+
__global__ void VecQuant8BatchMatMulKernel(
|
| 90 |
+
const scalar_t* __restrict__ vec,
|
| 91 |
+
const int* __restrict__ mat,
|
| 92 |
+
scalar_t* __restrict__ mul,
|
| 93 |
+
const scalar_t* __restrict__ scales,
|
| 94 |
+
const int* __restrict__ zeros,
|
| 95 |
+
int batch,
|
| 96 |
+
int heads,
|
| 97 |
+
int vec_row,
|
| 98 |
+
int vec_height,
|
| 99 |
+
int height,
|
| 100 |
+
int width,
|
| 101 |
+
int zero_width
|
| 102 |
+
);
|
| 103 |
+
|
| 104 |
+
template <typename scalar_t>
|
| 105 |
+
__global__ void VecQuant4BatchMatMulKernel(
|
| 106 |
+
const scalar_t* __restrict__ vec,
|
| 107 |
+
const int* __restrict__ mat,
|
| 108 |
+
scalar_t* __restrict__ mul,
|
| 109 |
+
const scalar_t* __restrict__ scales,
|
| 110 |
+
const int* __restrict__ zeros,
|
| 111 |
+
int batch,
|
| 112 |
+
int heads,
|
| 113 |
+
int vec_row,
|
| 114 |
+
int vec_height,
|
| 115 |
+
int height,
|
| 116 |
+
int width,
|
| 117 |
+
int zero_width
|
| 118 |
+
);
|
| 119 |
+
|
| 120 |
+
|
| 121 |
+
|
| 122 |
+
template <typename scalar_t>
|
| 123 |
+
__global__ void VecQuant8BatchMatMulKernel_old(
|
| 124 |
+
const scalar_t* __restrict__ vec,
|
| 125 |
+
const uint8_t* __restrict__ mat,
|
| 126 |
+
scalar_t* __restrict__ mul,
|
| 127 |
+
const scalar_t* __restrict__ scales,
|
| 128 |
+
const scalar_t* __restrict__ zeros,
|
| 129 |
+
int batch,
|
| 130 |
+
int heads,
|
| 131 |
+
int vec_row,
|
| 132 |
+
int vec_height,
|
| 133 |
+
int height,
|
| 134 |
+
int width,
|
| 135 |
+
int zero_width
|
| 136 |
+
);
|
| 137 |
+
|
| 138 |
+
__global__ void VecQuant8BatchMatMulKernel_faster(
|
| 139 |
+
const half* __restrict__ vec,
|
| 140 |
+
const uint8_t* __restrict__ mat,
|
| 141 |
+
half* __restrict__ mul,
|
| 142 |
+
const half* __restrict__ scales,
|
| 143 |
+
const half* __restrict__ zeros,
|
| 144 |
+
int batch,
|
| 145 |
+
int heads,
|
| 146 |
+
int vec_row,
|
| 147 |
+
int vec_height,
|
| 148 |
+
int height,
|
| 149 |
+
int width,
|
| 150 |
+
int zero_width
|
| 151 |
+
);
|
| 152 |
+
|
| 153 |
+
|
| 154 |
+
|
| 155 |
+
__global__ void VecQuant8BatchMatMulKernel_faster_old(
|
| 156 |
+
const half* __restrict__ vec,
|
| 157 |
+
const uint8_t* __restrict__ mat,
|
| 158 |
+
half* __restrict__ mul,
|
| 159 |
+
const half* __restrict__ scales,
|
| 160 |
+
const half* __restrict__ zeros,
|
| 161 |
+
int batch,
|
| 162 |
+
int heads,
|
| 163 |
+
int vec_row,
|
| 164 |
+
int vec_height,
|
| 165 |
+
int height,
|
| 166 |
+
int width
|
| 167 |
+
);
|
| 168 |
+
|
| 169 |
+
|
| 170 |
+
template <typename scalar_t>
|
| 171 |
+
__global__ void VecQuant4BatchMatMulKernel_old(
|
| 172 |
+
const scalar_t* __restrict__ vec,
|
| 173 |
+
const uint8_t* __restrict__ mat,
|
| 174 |
+
scalar_t* __restrict__ mul,
|
| 175 |
+
const scalar_t* __restrict__ scales,
|
| 176 |
+
const scalar_t* __restrict__ zeros,
|
| 177 |
+
int batch,
|
| 178 |
+
int heads,
|
| 179 |
+
int vec_row,
|
| 180 |
+
int vec_height,
|
| 181 |
+
int height,
|
| 182 |
+
int width,
|
| 183 |
+
int zero_width
|
| 184 |
+
);
|
| 185 |
+
|
| 186 |
+
|
| 187 |
+
template <typename scalar_t>
|
| 188 |
+
__global__ void VecQuant8BatchMatMulColumnCompressionKernel_old(
|
| 189 |
+
const scalar_t* __restrict__ vec,
|
| 190 |
+
const uint8_t* __restrict__ mat,
|
| 191 |
+
scalar_t* __restrict__ mul,
|
| 192 |
+
const scalar_t* __restrict__ scales,
|
| 193 |
+
const scalar_t* __restrict__ zeros,
|
| 194 |
+
int batch,
|
| 195 |
+
int heads,
|
| 196 |
+
int vec_row,
|
| 197 |
+
int height,
|
| 198 |
+
int width
|
| 199 |
+
);
|
| 200 |
+
|
| 201 |
+
__global__ void VecQuant8BatchMatMulColumnCompressionKernel_faster(
|
| 202 |
+
const half* __restrict__ vec,
|
| 203 |
+
const uint8_t* __restrict__ mat,
|
| 204 |
+
half* __restrict__ mul,
|
| 205 |
+
const half* __restrict__ scales,
|
| 206 |
+
const half* __restrict__ zeros,
|
| 207 |
+
int batch,
|
| 208 |
+
int heads,
|
| 209 |
+
int vec_row,
|
| 210 |
+
int height,
|
| 211 |
+
int width
|
| 212 |
+
);
|
| 213 |
+
|
| 214 |
+
__global__ void VecQuant8BatchMatMulColumnCompressionKernel_faster_old(
|
| 215 |
+
const half* __restrict__ vec,
|
| 216 |
+
const uint8_t* __restrict__ mat,
|
| 217 |
+
half* __restrict__ mul,
|
| 218 |
+
const half* __restrict__ scales,
|
| 219 |
+
const half* __restrict__ zeros,
|
| 220 |
+
int batch,
|
| 221 |
+
int heads,
|
| 222 |
+
int vec_row,
|
| 223 |
+
int height,
|
| 224 |
+
int width
|
| 225 |
+
);
|
| 226 |
+
|
| 227 |
+
|
| 228 |
+
template <typename scalar_t>
|
| 229 |
+
__global__ void VecQuant4BatchMatMulColumnCompressionKernel_old(
|
| 230 |
+
const scalar_t* __restrict__ vec,
|
| 231 |
+
const uint8_t* __restrict__ mat,
|
| 232 |
+
scalar_t* __restrict__ mul,
|
| 233 |
+
const scalar_t* __restrict__ scales,
|
| 234 |
+
const scalar_t* __restrict__ zeros,
|
| 235 |
+
int batch,
|
| 236 |
+
int heads,
|
| 237 |
+
int vec_row,
|
| 238 |
+
int height,
|
| 239 |
+
int width
|
| 240 |
+
);
|
| 241 |
+
|
| 242 |
+
|
| 243 |
+
__global__ void VecQuant8BatchMatMulKernel_faster(
|
| 244 |
+
const half* __restrict__ vec,
|
| 245 |
+
const uint8_t* __restrict__ mat,
|
| 246 |
+
half* __restrict__ mul,
|
| 247 |
+
const half* __restrict__ scales,
|
| 248 |
+
const half* __restrict__ zeros,
|
| 249 |
+
int batch,
|
| 250 |
+
int heads,
|
| 251 |
+
int vec_row,
|
| 252 |
+
int vec_height,
|
| 253 |
+
int height,
|
| 254 |
+
int width
|
| 255 |
+
);
|
| 256 |
+
|
| 257 |
+
|
| 258 |
+
__global__ void VecQuant8BatchMatMulColumnCompressionKernel_faster(
|
| 259 |
+
const half* __restrict__ vec,
|
| 260 |
+
const uint8_t* __restrict__ mat,
|
| 261 |
+
half* __restrict__ mul,
|
| 262 |
+
const half* __restrict__ scales,
|
| 263 |
+
const half* __restrict__ zeros,
|
| 264 |
+
int batch,
|
| 265 |
+
int heads,
|
| 266 |
+
int vec_row,
|
| 267 |
+
int height,
|
| 268 |
+
int width
|
| 269 |
+
);
|
| 270 |
+
|
| 271 |
+
const int BLOCKWIDTH = 128;
|
| 272 |
+
const int BLOCKHEIGHT8 = 32;
|
| 273 |
+
const int BLOCKHEIGHT4 = 16;
|
| 274 |
+
const int BLOCKHEIGHT_OLD4 = 128;
|
| 275 |
+
//const int BLOCKHEIGHT_OLD8 = 128;
|
| 276 |
+
|
| 277 |
+
__device__ inline unsigned int as_unsigned(int i) {
|
| 278 |
+
return *reinterpret_cast<unsigned int*>(&i);
|
| 279 |
+
}
|
| 280 |
+
|
| 281 |
+
__device__ inline int as_int(int i) {
|
| 282 |
+
return *reinterpret_cast<int*>(&i);
|
| 283 |
+
}
|
| 284 |
+
|
| 285 |
+
void vecquant8matmul_batched_column_compression_cuda(
|
| 286 |
+
torch::Tensor vec,
|
| 287 |
+
torch::Tensor mat,
|
| 288 |
+
torch::Tensor mul,
|
| 289 |
+
torch::Tensor scales,
|
| 290 |
+
torch::Tensor zeros
|
| 291 |
+
) {
|
| 292 |
+
int batch = vec.size(0);
|
| 293 |
+
int heads = vec.size(1);
|
| 294 |
+
int vec_row = vec.size(2);
|
| 295 |
+
int height = vec.size(3);
|
| 296 |
+
int width = mat.size(3) * 4;
|
| 297 |
+
|
| 298 |
+
dim3 blocks(
|
| 299 |
+
(height + BLOCKWIDTH - 1) / BLOCKWIDTH,
|
| 300 |
+
(width + BLOCKWIDTH - 1) / BLOCKWIDTH
|
| 301 |
+
);
|
| 302 |
+
dim3 threads(BLOCKWIDTH);
|
| 303 |
+
|
| 304 |
+
AT_DISPATCH_FLOATING_TYPES(
|
| 305 |
+
vec.type(), "vecquant8matmul_batched_cuda", ([&] {
|
| 306 |
+
VecQuant8BatchMatMulColumnCompressionKernel<<<blocks, threads>>>(
|
| 307 |
+
vec.data<scalar_t>(), mat.data<int>(), mul.data<scalar_t>(),
|
| 308 |
+
scales.data<scalar_t>(), zeros.data<int>(),
|
| 309 |
+
batch, heads, vec_row, height, width
|
| 310 |
+
);
|
| 311 |
+
})
|
| 312 |
+
);
|
| 313 |
+
|
| 314 |
+
}
|
| 315 |
+
|
| 316 |
+
template <typename scalar_t>
|
| 317 |
+
__global__ void VecQuant8BatchMatMulColumnCompressionKernel(
|
| 318 |
+
const scalar_t* __restrict__ vec,
|
| 319 |
+
const int* __restrict__ mat,
|
| 320 |
+
scalar_t* __restrict__ mul,
|
| 321 |
+
const scalar_t* __restrict__ scales,
|
| 322 |
+
const int* __restrict__ zeros,
|
| 323 |
+
int batch,
|
| 324 |
+
int heads,
|
| 325 |
+
int vec_row,
|
| 326 |
+
int height,
|
| 327 |
+
int width
|
| 328 |
+
) {
|
| 329 |
+
int weight_total = batch * heads * height * width / 4;
|
| 330 |
+
int input_total = batch * heads * vec_row * height;
|
| 331 |
+
int out_total = batch * heads * vec_row * width;
|
| 332 |
+
int tid = threadIdx.x;
|
| 333 |
+
// h is index of height with step being BLOCKWIDTH
|
| 334 |
+
int h = BLOCKWIDTH * blockIdx.x;
|
| 335 |
+
// w is index of width with step being 1
|
| 336 |
+
int w = BLOCKWIDTH * blockIdx.y + tid;
|
| 337 |
+
if (w >= width && tid >= height) {
|
| 338 |
+
return;
|
| 339 |
+
}
|
| 340 |
+
|
| 341 |
+
__shared__ scalar_t blockvec[BLOCKWIDTH];
|
| 342 |
+
int k;
|
| 343 |
+
scalar_t w_tmp;
|
| 344 |
+
|
| 345 |
+
float weight[BLOCKWIDTH];
|
| 346 |
+
|
| 347 |
+
for (int b = 0; b < batch; ++b){
|
| 348 |
+
for (int head = 0; head < heads; ++head){
|
| 349 |
+
int batch_shift = b * heads + head;
|
| 350 |
+
for (k = 0; k < BLOCKWIDTH && h + k < height; ++k){
|
| 351 |
+
int i_w = (w / 4);
|
| 352 |
+
int w_bit = (w % 4) * 8;
|
| 353 |
+
|
| 354 |
+
int w_index = (batch_shift * height + h + k) * width / 4 + i_w;
|
| 355 |
+
if (w_index >= weight_total || w >= width) {
|
| 356 |
+
weight[k] = 0;
|
| 357 |
+
} else {
|
| 358 |
+
scalar_t scale = scales[batch_shift * height + h + k];
|
| 359 |
+
scalar_t zero = zeros[batch_shift * height + h + k];
|
| 360 |
+
w_tmp = ((as_unsigned(mat[w_index]) >> w_bit) & 0xFF);
|
| 361 |
+
weight[k] = scale * (w_tmp - zero);
|
| 362 |
+
}
|
| 363 |
+
}
|
| 364 |
+
|
| 365 |
+
scalar_t res;
|
| 366 |
+
for (int vr = 0; vr < vec_row; ++vr){
|
| 367 |
+
res = 0;
|
| 368 |
+
int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
|
| 369 |
+
if (vec_index < input_total) {
|
| 370 |
+
blockvec[tid] = vec[vec_index];
|
| 371 |
+
} else {
|
| 372 |
+
blockvec[tid] = 0;
|
| 373 |
+
}
|
| 374 |
+
|
| 375 |
+
__syncthreads();
|
| 376 |
+
for (k = 0; k < BLOCKWIDTH && h + k < height; ++k){
|
| 377 |
+
// res is the dot product of BLOCKWIDTH elements (part of width)
|
| 378 |
+
res += weight[k] * blockvec[k];
|
| 379 |
+
}
|
| 380 |
+
// add res to the final result, final matrix shape: (batch, vec_row, width)
|
| 381 |
+
int out_index = (batch_shift * vec_row + vr) * width + w;
|
| 382 |
+
if (out_index < out_total) {
|
| 383 |
+
atomicAdd(&mul[out_index], res);
|
| 384 |
+
}
|
| 385 |
+
__syncthreads();
|
| 386 |
+
}
|
| 387 |
+
}
|
| 388 |
+
}
|
| 389 |
+
}
|
| 390 |
+
|
| 391 |
+
void vecquant8matmul_batched_cuda(
|
| 392 |
+
torch::Tensor vec,
|
| 393 |
+
torch::Tensor mat,
|
| 394 |
+
torch::Tensor mul,
|
| 395 |
+
torch::Tensor scales,
|
| 396 |
+
torch::Tensor zeros
|
| 397 |
+
) {
|
| 398 |
+
int batch = vec.size(0);
|
| 399 |
+
int heads = vec.size(1);
|
| 400 |
+
int vec_row = vec.size(2);
|
| 401 |
+
int vec_height = vec.size(3);
|
| 402 |
+
int height = mat.size(2);
|
| 403 |
+
int width = mat.size(3);
|
| 404 |
+
int zero_width = zeros.size(2);
|
| 405 |
+
|
| 406 |
+
dim3 blocks(
|
| 407 |
+
(height + BLOCKHEIGHT8 - 1) / BLOCKHEIGHT8,
|
| 408 |
+
(width + BLOCKWIDTH - 1) / BLOCKWIDTH
|
| 409 |
+
);
|
| 410 |
+
dim3 threads(BLOCKWIDTH);
|
| 411 |
+
|
| 412 |
+
AT_DISPATCH_FLOATING_TYPES(
|
| 413 |
+
vec.type(), "vecquant8matmul_batched_cuda", ([&] {
|
| 414 |
+
VecQuant8BatchMatMulKernel<<<blocks, threads>>>(
|
| 415 |
+
vec.data<scalar_t>(), mat.data<int>(), mul.data<scalar_t>(),
|
| 416 |
+
scales.data<scalar_t>(), zeros.data<int>(),
|
| 417 |
+
batch, heads, vec_row, vec_height, height, width, zero_width
|
| 418 |
+
);
|
| 419 |
+
})
|
| 420 |
+
);
|
| 421 |
+
|
| 422 |
+
}
|
| 423 |
+
|
| 424 |
+
template <typename scalar_t>
|
| 425 |
+
__global__ void VecQuant8BatchMatMulKernel(
|
| 426 |
+
const scalar_t* __restrict__ vec,
|
| 427 |
+
const int* __restrict__ mat,
|
| 428 |
+
scalar_t* __restrict__ mul,
|
| 429 |
+
const scalar_t* __restrict__ scales,
|
| 430 |
+
const int* __restrict__ zeros,
|
| 431 |
+
int batch,
|
| 432 |
+
int heads,
|
| 433 |
+
int vec_row,
|
| 434 |
+
int vec_height,
|
| 435 |
+
int height,
|
| 436 |
+
int width,
|
| 437 |
+
int zero_width
|
| 438 |
+
) {
|
| 439 |
+
int weight_total = batch * heads * height * width;
|
| 440 |
+
int input_total = batch * heads * vec_row * vec_height;
|
| 441 |
+
int out_total = batch * heads * vec_row * width;
|
| 442 |
+
int tid = threadIdx.x;
|
| 443 |
+
// h is index of height with step being BLOCKHEIGHT8
|
| 444 |
+
int h = BLOCKHEIGHT8 * blockIdx.x;
|
| 445 |
+
// w is index of width with step being 1
|
| 446 |
+
int w = BLOCKWIDTH * blockIdx.y + tid;
|
| 447 |
+
if (w >= width && tid >= vec_height) {
|
| 448 |
+
return;
|
| 449 |
+
}
|
| 450 |
+
|
| 451 |
+
__shared__ scalar_t blockvec[BLOCKWIDTH];
|
| 452 |
+
// i is index of mat of block first row
|
| 453 |
+
int i = width * h + w;
|
| 454 |
+
// if (i >= width * height) {
|
| 455 |
+
// return;
|
| 456 |
+
// }
|
| 457 |
+
int k;
|
| 458 |
+
scalar_t w_tmp;
|
| 459 |
+
|
| 460 |
+
int z_w = w / 4;
|
| 461 |
+
int z_mod = (w % 4) * 8;
|
| 462 |
+
|
| 463 |
+
float weight[BLOCKWIDTH];
|
| 464 |
+
|
| 465 |
+
for (int b = 0; b < batch; ++b){
|
| 466 |
+
for (int head = 0; head < heads; ++head){
|
| 467 |
+
int batch_shift = b * heads + head;
|
| 468 |
+
for (k = 0; k < BLOCKWIDTH && h * 4 + k < vec_height; ++k){
|
| 469 |
+
int k_w = (k / 4);
|
| 470 |
+
int k_bit = (k % 4) * 8;
|
| 471 |
+
|
| 472 |
+
int w_index = batch_shift * height * width + i + (k_w * width);
|
| 473 |
+
if (w_index >= weight_total || w >= width) {
|
| 474 |
+
weight[k] = 0;
|
| 475 |
+
} else {
|
| 476 |
+
scalar_t scale = scales[batch_shift * width + w];
|
| 477 |
+
scalar_t zero;
|
| 478 |
+
if (zero_width == width) {
|
| 479 |
+
zero = zeros[batch_shift * width + w];
|
| 480 |
+
} else {
|
| 481 |
+
zero = scalar_t(((as_unsigned(zeros[batch_shift * zero_width + z_w]) >> z_mod) & 0xFF) + 1);
|
| 482 |
+
}
|
| 483 |
+
w_tmp = ((as_unsigned(mat[w_index]) >> k_bit) & 0xFF);
|
| 484 |
+
weight[k] = scale * (w_tmp - zero);
|
| 485 |
+
}
|
| 486 |
+
}
|
| 487 |
+
|
| 488 |
+
scalar_t res;
|
| 489 |
+
for (int vr = 0; vr < vec_row; ++vr){
|
| 490 |
+
res = 0;
|
| 491 |
+
int vec_index = (batch_shift * vec_row + vr) * vec_height + blockIdx.x * BLOCKWIDTH + tid;
|
| 492 |
+
if (vec_index < input_total) {
|
| 493 |
+
blockvec[tid] = vec[vec_index];
|
| 494 |
+
} else {
|
| 495 |
+
blockvec[tid] = 0;
|
| 496 |
+
}
|
| 497 |
+
|
| 498 |
+
__syncthreads();
|
| 499 |
+
for (k = 0; k < BLOCKWIDTH && h * 4 + k < vec_height; ++k){
|
| 500 |
+
// res is the dot product of BLOCKWIDTH elements (part of width)
|
| 501 |
+
res += weight[k] * blockvec[k];
|
| 502 |
+
}
|
| 503 |
+
// add res to the final result, final matrix shape: (batch, vec_row, width)
|
| 504 |
+
int out_index = (batch_shift * vec_row + vr) * width + w;
|
| 505 |
+
if (out_index < out_total) {
|
| 506 |
+
atomicAdd(&mul[out_index], res);
|
| 507 |
+
}
|
| 508 |
+
__syncthreads();
|
| 509 |
+
}
|
| 510 |
+
}
|
| 511 |
+
}
|
| 512 |
+
}
|
| 513 |
+
|
| 514 |
+
|
| 515 |
+
void vecquant8matmul_cuda(
|
| 516 |
+
torch::Tensor vec,
|
| 517 |
+
torch::Tensor mat,
|
| 518 |
+
torch::Tensor mul,
|
| 519 |
+
torch::Tensor scales,
|
| 520 |
+
torch::Tensor zeros,
|
| 521 |
+
torch::Tensor g_idx
|
| 522 |
+
) {
|
| 523 |
+
int batch = vec.size(0);
|
| 524 |
+
int vec_height = vec.size(1);
|
| 525 |
+
int height = mat.size(0);
|
| 526 |
+
int width = mat.size(1);
|
| 527 |
+
int zero_width = zeros.size(1);
|
| 528 |
+
|
| 529 |
+
dim3 blocks(
|
| 530 |
+
(height + BLOCKHEIGHT8 - 1) / BLOCKHEIGHT8,
|
| 531 |
+
(width + BLOCKWIDTH - 1) / BLOCKWIDTH
|
| 532 |
+
);
|
| 533 |
+
dim3 threads(BLOCKWIDTH);
|
| 534 |
+
|
| 535 |
+
AT_DISPATCH_FLOATING_TYPES(
|
| 536 |
+
vec.type(), "vecquant8matmul_cuda", ([&] {
|
| 537 |
+
VecQuant8MatMulKernel<<<blocks, threads>>>(
|
| 538 |
+
vec.data<scalar_t>(), mat.data<int>(), mul.data<scalar_t>(),
|
| 539 |
+
scales.data<scalar_t>(), zeros.data<int>(), g_idx.data<int>(),
|
| 540 |
+
batch, vec_height, height, width, zero_width
|
| 541 |
+
);
|
| 542 |
+
})
|
| 543 |
+
);
|
| 544 |
+
}
|
| 545 |
+
|
| 546 |
+
template <typename scalar_t>
|
| 547 |
+
__global__ void VecQuant8MatMulKernel(
|
| 548 |
+
const scalar_t* __restrict__ vec,
|
| 549 |
+
const int* __restrict__ mat,
|
| 550 |
+
scalar_t* __restrict__ mul,
|
| 551 |
+
const scalar_t* __restrict__ scales,
|
| 552 |
+
const int* __restrict__ zeros,
|
| 553 |
+
const int* __restrict__ g_idx,
|
| 554 |
+
int batch,
|
| 555 |
+
int vec_height,
|
| 556 |
+
int height,
|
| 557 |
+
int width,
|
| 558 |
+
int zero_width
|
| 559 |
+
) {
|
| 560 |
+
int h = BLOCKHEIGHT8 * blockIdx.x;
|
| 561 |
+
int w = BLOCKWIDTH * blockIdx.y + threadIdx.x;
|
| 562 |
+
|
| 563 |
+
__shared__ scalar_t blockvec[BLOCKWIDTH];
|
| 564 |
+
int i = width * h + w;
|
| 565 |
+
int g_h = h * 4;
|
| 566 |
+
int k;
|
| 567 |
+
unsigned int g;
|
| 568 |
+
scalar_t w_tmp;
|
| 569 |
+
|
| 570 |
+
int z_w = w / 4;
|
| 571 |
+
int z_mod = (w % 4) * 8;
|
| 572 |
+
|
| 573 |
+
float weight[BLOCKWIDTH];
|
| 574 |
+
|
| 575 |
+
for (k = 0; k < BLOCKWIDTH; ++k){
|
| 576 |
+
int k_w = (k / 4);
|
| 577 |
+
int k_bit = (k % 4) * 8;
|
| 578 |
+
|
| 579 |
+
g = as_int(g_idx[g_h + k]);
|
| 580 |
+
scalar_t scale = scales[g * width + w];
|
| 581 |
+
scalar_t zero = scalar_t(((as_unsigned(zeros[g * zero_width + z_w]) >> z_mod) & 0xFF) + 1);
|
| 582 |
+
|
| 583 |
+
w_tmp = ((as_unsigned(mat[i + (k_w * width)]) >> k_bit) & 0xFF);
|
| 584 |
+
|
| 585 |
+
weight[k] = scale * (w_tmp - zero);
|
| 586 |
+
}
|
| 587 |
+
|
| 588 |
+
|
| 589 |
+
scalar_t res;
|
| 590 |
+
for (int b = 0; b < batch; ++b){
|
| 591 |
+
res = 0;
|
| 592 |
+
blockvec[threadIdx.x] = vec[b * vec_height + blockIdx.x * BLOCKWIDTH + threadIdx.x];
|
| 593 |
+
__syncthreads();
|
| 594 |
+
for (k = 0; k < BLOCKWIDTH; ++k){
|
| 595 |
+
res += weight[k] * blockvec[k];
|
| 596 |
+
}
|
| 597 |
+
atomicAdd(&mul[b * width + w], res);
|
| 598 |
+
__syncthreads();
|
| 599 |
+
}
|
| 600 |
+
}
|
| 601 |
+
|
| 602 |
+
|
| 603 |
+
|
| 604 |
+
void vecquant4matmul_batched_cuda(
|
| 605 |
+
torch::Tensor vec,
|
| 606 |
+
torch::Tensor mat,
|
| 607 |
+
torch::Tensor mul,
|
| 608 |
+
torch::Tensor scales,
|
| 609 |
+
torch::Tensor zeros
|
| 610 |
+
) {
|
| 611 |
+
int batch = vec.size(0);
|
| 612 |
+
int heads = vec.size(1);
|
| 613 |
+
int vec_row = vec.size(2);
|
| 614 |
+
int vec_height = vec.size(3);
|
| 615 |
+
int height = mat.size(2);
|
| 616 |
+
int width = mat.size(3);
|
| 617 |
+
int zero_width = zeros.size(2);
|
| 618 |
+
|
| 619 |
+
dim3 blocks(
|
| 620 |
+
(height + BLOCKHEIGHT4 - 1) / BLOCKHEIGHT4,
|
| 621 |
+
(width + BLOCKWIDTH - 1) / BLOCKWIDTH
|
| 622 |
+
);
|
| 623 |
+
dim3 threads(BLOCKWIDTH);
|
| 624 |
+
|
| 625 |
+
AT_DISPATCH_FLOATING_TYPES(
|
| 626 |
+
vec.type(), "vecquant4matmul_batched_cuda", ([&] {
|
| 627 |
+
VecQuant4BatchMatMulKernel<<<blocks, threads>>>(
|
| 628 |
+
vec.data<scalar_t>(), mat.data<int>(), mul.data<scalar_t>(),
|
| 629 |
+
scales.data<scalar_t>(), zeros.data<int>(),
|
| 630 |
+
batch, heads, vec_row, vec_height, height, width, zero_width
|
| 631 |
+
);
|
| 632 |
+
})
|
| 633 |
+
);
|
| 634 |
+
|
| 635 |
+
}
|
| 636 |
+
|
| 637 |
+
template <typename scalar_t>
|
| 638 |
+
__global__ void VecQuant4BatchMatMulKernel(
|
| 639 |
+
const scalar_t* __restrict__ vec,
|
| 640 |
+
const int* __restrict__ mat,
|
| 641 |
+
scalar_t* __restrict__ mul,
|
| 642 |
+
const scalar_t* __restrict__ scales,
|
| 643 |
+
const int* __restrict__ zeros,
|
| 644 |
+
int batch,
|
| 645 |
+
int heads,
|
| 646 |
+
int vec_row,
|
| 647 |
+
int vec_height,
|
| 648 |
+
int height,
|
| 649 |
+
int width,
|
| 650 |
+
int zero_width
|
| 651 |
+
) {
|
| 652 |
+
int weight_total = batch * heads * height * width;
|
| 653 |
+
int input_total = batch * heads * vec_row * vec_height;
|
| 654 |
+
int out_total = batch * heads * vec_row * width;
|
| 655 |
+
int tid = threadIdx.x;
|
| 656 |
+
// h is index of height with step being BLOCKHEIGHT4
|
| 657 |
+
int h = BLOCKHEIGHT4 * blockIdx.x;
|
| 658 |
+
// w is index of width with step being 1
|
| 659 |
+
int w = BLOCKWIDTH * blockIdx.y + tid;
|
| 660 |
+
if (w >= width && tid >= vec_height) {
|
| 661 |
+
return;
|
| 662 |
+
}
|
| 663 |
+
|
| 664 |
+
__shared__ scalar_t blockvec[BLOCKWIDTH];
|
| 665 |
+
// i is index of mat of block first row
|
| 666 |
+
int i = width * h + w;
|
| 667 |
+
int k;
|
| 668 |
+
scalar_t w_tmp;
|
| 669 |
+
|
| 670 |
+
int z_w = w / 8;
|
| 671 |
+
int z_mod = (w % 8) * 4;
|
| 672 |
+
|
| 673 |
+
float weight[BLOCKWIDTH];
|
| 674 |
+
|
| 675 |
+
for (int b = 0; b < batch; ++b){
|
| 676 |
+
for (int head = 0; head < heads; ++head){
|
| 677 |
+
int batch_shift = b * heads + head;
|
| 678 |
+
for (k = 0; k < BLOCKWIDTH && h * 8 + k < vec_height; ++k){
|
| 679 |
+
int k_w = (k / 8);
|
| 680 |
+
int k_bit = (k % 8) * 4;
|
| 681 |
+
|
| 682 |
+
int w_index = batch_shift * height * width + i + (k_w * width);
|
| 683 |
+
if (w_index >= weight_total || w >= width) {
|
| 684 |
+
weight[k] = 0;
|
| 685 |
+
} else {
|
| 686 |
+
scalar_t scale = scales[batch_shift * width + w];
|
| 687 |
+
scalar_t zero;
|
| 688 |
+
if (zero_width == width) {
|
| 689 |
+
zero = zeros[batch_shift * width + w];
|
| 690 |
+
} else {
|
| 691 |
+
zero = scalar_t(((as_unsigned(zeros[batch_shift * zero_width + z_w]) >> z_mod) & 0xF));
|
| 692 |
+
}
|
| 693 |
+
w_tmp = ((as_unsigned(mat[w_index]) >> k_bit) & 0xF);
|
| 694 |
+
weight[k] = scale * (w_tmp - zero);
|
| 695 |
+
}
|
| 696 |
+
}
|
| 697 |
+
|
| 698 |
+
scalar_t res;
|
| 699 |
+
for (int vr = 0; vr < vec_row; ++vr){
|
| 700 |
+
res = 0;
|
| 701 |
+
int vec_index = (batch_shift * vec_row + vr) * vec_height + blockIdx.x * BLOCKWIDTH + tid;
|
| 702 |
+
if (vec_index < input_total) {
|
| 703 |
+
blockvec[tid] = vec[vec_index];
|
| 704 |
+
} else {
|
| 705 |
+
blockvec[tid] = 0;
|
| 706 |
+
}
|
| 707 |
+
|
| 708 |
+
__syncthreads();
|
| 709 |
+
for (k = 0; k < BLOCKWIDTH && h * 8 + k < vec_height; ++k){
|
| 710 |
+
// res is the dot product of BLOCKWIDTH elements (part of width)
|
| 711 |
+
res += weight[k] * blockvec[k];
|
| 712 |
+
}
|
| 713 |
+
// add res to the final result, final matrix shape: (batch, vec_row, width)
|
| 714 |
+
int out_index = (batch_shift * vec_row + vr) * width + w;
|
| 715 |
+
if (out_index < out_total) {
|
| 716 |
+
atomicAdd(&mul[out_index], res);
|
| 717 |
+
}
|
| 718 |
+
__syncthreads();
|
| 719 |
+
}
|
| 720 |
+
}
|
| 721 |
+
}
|
| 722 |
+
}
|
| 723 |
+
|
| 724 |
+
|
| 725 |
+
|
| 726 |
+
void vecquant4matmul_batched_column_compression_cuda(
|
| 727 |
+
torch::Tensor vec,
|
| 728 |
+
torch::Tensor mat,
|
| 729 |
+
torch::Tensor mul,
|
| 730 |
+
torch::Tensor scales,
|
| 731 |
+
torch::Tensor zeros
|
| 732 |
+
) {
|
| 733 |
+
int batch = vec.size(0);
|
| 734 |
+
int heads = vec.size(1);
|
| 735 |
+
int vec_row = vec.size(2);
|
| 736 |
+
int height = vec.size(3);
|
| 737 |
+
int width = mat.size(3) * 8;
|
| 738 |
+
|
| 739 |
+
dim3 blocks(
|
| 740 |
+
(height + BLOCKWIDTH - 1) / BLOCKWIDTH,
|
| 741 |
+
(width + BLOCKWIDTH - 1) / BLOCKWIDTH
|
| 742 |
+
);
|
| 743 |
+
dim3 threads(BLOCKWIDTH);
|
| 744 |
+
|
| 745 |
+
AT_DISPATCH_FLOATING_TYPES(
|
| 746 |
+
vec.type(), "vecquant4matmul_batched_cuda", ([&] {
|
| 747 |
+
VecQuant4BatchMatMulColumnCompressionKernel<<<blocks, threads>>>(
|
| 748 |
+
vec.data<scalar_t>(), mat.data<int>(), mul.data<scalar_t>(),
|
| 749 |
+
scales.data<scalar_t>(), zeros.data<int>(),
|
| 750 |
+
batch, heads, vec_row, height, width
|
| 751 |
+
);
|
| 752 |
+
})
|
| 753 |
+
);
|
| 754 |
+
|
| 755 |
+
}
|
| 756 |
+
|
| 757 |
+
template <typename scalar_t>
|
| 758 |
+
__global__ void VecQuant4BatchMatMulColumnCompressionKernel(
|
| 759 |
+
const scalar_t* __restrict__ vec,
|
| 760 |
+
const int* __restrict__ mat,
|
| 761 |
+
scalar_t* __restrict__ mul,
|
| 762 |
+
const scalar_t* __restrict__ scales,
|
| 763 |
+
const int* __restrict__ zeros,
|
| 764 |
+
int batch,
|
| 765 |
+
int heads,
|
| 766 |
+
int vec_row,
|
| 767 |
+
int height,
|
| 768 |
+
int width
|
| 769 |
+
) {
|
| 770 |
+
int weight_total = batch * heads * height * width / 8;
|
| 771 |
+
int input_total = batch * heads * vec_row * height;
|
| 772 |
+
int out_total = batch * heads * vec_row * width;
|
| 773 |
+
int tid = threadIdx.x;
|
| 774 |
+
// h is index of height with step being BLOCKWIDTH
|
| 775 |
+
int h = BLOCKWIDTH * blockIdx.x;
|
| 776 |
+
// w is index of width with step being 1
|
| 777 |
+
int w = BLOCKWIDTH * blockIdx.y + tid;
|
| 778 |
+
if (w >= width && tid >= height) {
|
| 779 |
+
return;
|
| 780 |
+
}
|
| 781 |
+
|
| 782 |
+
__shared__ scalar_t blockvec[BLOCKWIDTH];
|
| 783 |
+
int k;
|
| 784 |
+
scalar_t w_tmp;
|
| 785 |
+
|
| 786 |
+
float weight[BLOCKWIDTH];
|
| 787 |
+
|
| 788 |
+
for (int b = 0; b < batch; ++b){
|
| 789 |
+
for (int head = 0; head < heads; ++head){
|
| 790 |
+
int batch_shift = b * heads + head;
|
| 791 |
+
for (k = 0; k < BLOCKWIDTH && h + k < height; ++k){
|
| 792 |
+
int i_w = (w / 8);
|
| 793 |
+
int w_bit = (w % 8) * 4;
|
| 794 |
+
|
| 795 |
+
int w_index = (batch_shift * height + h + k) * width / 8 + i_w;
|
| 796 |
+
if (w_index >= weight_total || w >= width) {
|
| 797 |
+
weight[k] = 0;
|
| 798 |
+
} else {
|
| 799 |
+
scalar_t scale = scales[batch_shift * height + h + k];
|
| 800 |
+
scalar_t zero = zeros[batch_shift * height + h + k];
|
| 801 |
+
w_tmp = ((as_unsigned(mat[w_index]) >> w_bit) & 0xF);
|
| 802 |
+
weight[k] = scale * (w_tmp - zero);
|
| 803 |
+
}
|
| 804 |
+
}
|
| 805 |
+
|
| 806 |
+
scalar_t res;
|
| 807 |
+
for (int vr = 0; vr < vec_row; ++vr){
|
| 808 |
+
res = 0;
|
| 809 |
+
int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
|
| 810 |
+
if (vec_index < input_total) {
|
| 811 |
+
blockvec[tid] = vec[vec_index];
|
| 812 |
+
} else {
|
| 813 |
+
blockvec[tid] = 0;
|
| 814 |
+
}
|
| 815 |
+
|
| 816 |
+
__syncthreads();
|
| 817 |
+
for (k = 0; k < BLOCKWIDTH && h + k < height; ++k){
|
| 818 |
+
// res is the dot product of BLOCKWIDTH elements (part of width)
|
| 819 |
+
res += weight[k] * blockvec[k];
|
| 820 |
+
}
|
| 821 |
+
// add res to the final result, final matrix shape: (batch, vec_row, width)
|
| 822 |
+
int out_index = (batch_shift * vec_row + vr) * width + w;
|
| 823 |
+
if (out_index < out_total) {
|
| 824 |
+
atomicAdd(&mul[out_index], res);
|
| 825 |
+
}
|
| 826 |
+
__syncthreads();
|
| 827 |
+
}
|
| 828 |
+
}
|
| 829 |
+
}
|
| 830 |
+
}
|
| 831 |
+
|
| 832 |
+
|
| 833 |
+
void vecquant8matmul_batched_old_cuda(
|
| 834 |
+
torch::Tensor vec,
|
| 835 |
+
torch::Tensor mat,
|
| 836 |
+
torch::Tensor mul,
|
| 837 |
+
torch::Tensor scales,
|
| 838 |
+
torch::Tensor zeros
|
| 839 |
+
) {
|
| 840 |
+
int batch = vec.size(0);
|
| 841 |
+
int heads = vec.size(1);
|
| 842 |
+
int vec_row = vec.size(2);
|
| 843 |
+
int vec_height = vec.size(3);
|
| 844 |
+
int height = mat.size(2);
|
| 845 |
+
int width = mat.size(3);
|
| 846 |
+
int zero_width = zeros.size(2);
|
| 847 |
+
|
| 848 |
+
dim3 blocks(
|
| 849 |
+
(height + BLOCKWIDTH - 1) / BLOCKWIDTH,
|
| 850 |
+
(width + BLOCKWIDTH - 1) / BLOCKWIDTH
|
| 851 |
+
);
|
| 852 |
+
dim3 threads(BLOCKWIDTH);
|
| 853 |
+
|
| 854 |
+
AT_DISPATCH_FLOATING_TYPES(
|
| 855 |
+
vec.type(), "vecquant8matmul_batched_old_cuda", ([&] {
|
| 856 |
+
VecQuant8BatchMatMulKernel_old<<<blocks, threads>>>(
|
| 857 |
+
vec.data<scalar_t>(), mat.data<uint8_t>(), mul.data<scalar_t>(),
|
| 858 |
+
scales.data<scalar_t>(), zeros.data<scalar_t>(),
|
| 859 |
+
batch, heads, vec_row, vec_height, height, width, zero_width
|
| 860 |
+
);
|
| 861 |
+
})
|
| 862 |
+
);
|
| 863 |
+
}
|
| 864 |
+
|
| 865 |
+
|
| 866 |
+
template <typename scalar_t>
|
| 867 |
+
__global__ void VecQuant8BatchMatMulKernel_old(
|
| 868 |
+
const scalar_t* __restrict__ vec,
|
| 869 |
+
const uint8_t* __restrict__ mat,
|
| 870 |
+
scalar_t* __restrict__ mul,
|
| 871 |
+
const scalar_t* __restrict__ scales,
|
| 872 |
+
const scalar_t* __restrict__ zeros,
|
| 873 |
+
int batch,
|
| 874 |
+
int heads,
|
| 875 |
+
int vec_row,
|
| 876 |
+
int vec_height,
|
| 877 |
+
int height,
|
| 878 |
+
int width,
|
| 879 |
+
int zero_width
|
| 880 |
+
) {
|
| 881 |
+
int weight_total = batch * heads * height * width;
|
| 882 |
+
int input_total = batch * heads * vec_row * vec_height;
|
| 883 |
+
int out_total = batch * heads * vec_row * width;
|
| 884 |
+
int tid = threadIdx.x;
|
| 885 |
+
// h is index of height with step being BLOCKHEIGHT8
|
| 886 |
+
int h = BLOCKWIDTH * blockIdx.x;
|
| 887 |
+
// w is index of width with step being 1
|
| 888 |
+
int w = BLOCKWIDTH * blockIdx.y + tid;
|
| 889 |
+
if (w >= width && tid >= vec_height) {
|
| 890 |
+
return;
|
| 891 |
+
}
|
| 892 |
+
|
| 893 |
+
__shared__ scalar_t blockvec[BLOCKWIDTH];
|
| 894 |
+
// i is index of mat of block first row
|
| 895 |
+
int i = width * h + w;
|
| 896 |
+
int k;
|
| 897 |
+
scalar_t w_tmp;
|
| 898 |
+
|
| 899 |
+
float weight[BLOCKWIDTH];
|
| 900 |
+
for (int b = 0; b < batch; ++b){
|
| 901 |
+
for (int head = 0; head < heads; ++head){
|
| 902 |
+
int batch_shift = b * heads + head;
|
| 903 |
+
for (k = 0; k < BLOCKWIDTH && h + k < vec_height; ++k){
|
| 904 |
+
int k_w = k;
|
| 905 |
+
int w_index = batch_shift * height * width + i + (k_w * width);
|
| 906 |
+
if (w_index >= weight_total || w >= width) {
|
| 907 |
+
weight[k] = 0;
|
| 908 |
+
} else {
|
| 909 |
+
scalar_t scale = scales[batch_shift * width + w];
|
| 910 |
+
scalar_t zero = zeros[batch_shift * width + w];
|
| 911 |
+
w_tmp = as_unsigned(mat[w_index]);
|
| 912 |
+
weight[k] = scale * (w_tmp - zero);
|
| 913 |
+
}
|
| 914 |
+
}
|
| 915 |
+
|
| 916 |
+
scalar_t res;
|
| 917 |
+
for (int vr = 0; vr < vec_row; ++vr){
|
| 918 |
+
res = 0;
|
| 919 |
+
int vec_index = (batch_shift * vec_row + vr) * vec_height + blockIdx.x * BLOCKWIDTH + tid;
|
| 920 |
+
if (vec_index < input_total) {
|
| 921 |
+
blockvec[tid] = vec[vec_index];
|
| 922 |
+
} else {
|
| 923 |
+
blockvec[tid] = 0;
|
| 924 |
+
}
|
| 925 |
+
|
| 926 |
+
__syncthreads();
|
| 927 |
+
for (k = 0; k < BLOCKWIDTH && h + k < vec_height; ++k){
|
| 928 |
+
// res is the dot product of BLOCKWIDTH elements (part of width)
|
| 929 |
+
res += weight[k] * blockvec[k];
|
| 930 |
+
}
|
| 931 |
+
// add res to the final result, final matrix shape: (batch, vec_row, width)
|
| 932 |
+
int out_index = (batch_shift * vec_row + vr) * width + w;
|
| 933 |
+
if (out_index < out_total) {
|
| 934 |
+
atomicAdd(&mul[out_index], res);
|
| 935 |
+
}
|
| 936 |
+
__syncthreads();
|
| 937 |
+
}
|
| 938 |
+
}
|
| 939 |
+
}
|
| 940 |
+
}
|
| 941 |
+
|
| 942 |
+
|
| 943 |
+
|
| 944 |
+
void vecquant8matmul_batched_faster_cuda(
|
| 945 |
+
torch::Tensor vec,
|
| 946 |
+
torch::Tensor mat,
|
| 947 |
+
torch::Tensor mul,
|
| 948 |
+
torch::Tensor scales,
|
| 949 |
+
torch::Tensor zeros
|
| 950 |
+
) {
|
| 951 |
+
int batch = vec.size(0);
|
| 952 |
+
int heads = vec.size(1);
|
| 953 |
+
int vec_row = vec.size(2);
|
| 954 |
+
int vec_height = vec.size(3);
|
| 955 |
+
int height = mat.size(2);
|
| 956 |
+
int width = mat.size(3);
|
| 957 |
+
int zero_width = zeros.size(2);
|
| 958 |
+
|
| 959 |
+
dim3 blocks(
|
| 960 |
+
(height + BLOCKWIDTH - 1) / BLOCKWIDTH,
|
| 961 |
+
(width + BLOCKWIDTH - 1) / BLOCKWIDTH
|
| 962 |
+
);
|
| 963 |
+
dim3 threads(BLOCKWIDTH);
|
| 964 |
+
|
| 965 |
+
VecQuant8BatchMatMulKernel_faster<<<blocks, threads>>>(
|
| 966 |
+
(half*) vec.data_ptr(),
|
| 967 |
+
(uint8_t*) mat.data_ptr(),
|
| 968 |
+
(half*) mul.data_ptr(),
|
| 969 |
+
(half*) scales.data_ptr(),
|
| 970 |
+
(half*) zeros.data_ptr(),
|
| 971 |
+
batch, heads, vec_row, vec_height, height, width, zero_width
|
| 972 |
+
);
|
| 973 |
+
}
|
| 974 |
+
|
| 975 |
+
|
| 976 |
+
|
| 977 |
+
__global__ void VecQuant8BatchMatMulKernel_faster(
|
| 978 |
+
const half* __restrict__ vec,
|
| 979 |
+
const uint8_t* __restrict__ mat,
|
| 980 |
+
half* __restrict__ mul,
|
| 981 |
+
const half* __restrict__ scales,
|
| 982 |
+
const half* __restrict__ zeros,
|
| 983 |
+
int batch,
|
| 984 |
+
int heads,
|
| 985 |
+
int vec_row,
|
| 986 |
+
int vec_height,
|
| 987 |
+
int height,
|
| 988 |
+
int width,
|
| 989 |
+
int zero_width
|
| 990 |
+
) {
|
| 991 |
+
//int weight_total = batch * heads * height * width;
|
| 992 |
+
int input_total = batch * heads * vec_row * vec_height;
|
| 993 |
+
int out_total = batch * heads * vec_row * width;
|
| 994 |
+
int tid = threadIdx.x;
|
| 995 |
+
int h = BLOCKWIDTH * blockIdx.x;
|
| 996 |
+
int w = BLOCKWIDTH * blockIdx.y + tid;
|
| 997 |
+
if (w >= width && tid >= height) {
|
| 998 |
+
return;
|
| 999 |
+
}
|
| 1000 |
+
|
| 1001 |
+
__shared__ float blockvec[BLOCKWIDTH];
|
| 1002 |
+
int i = width * h + w;
|
| 1003 |
+
int k;
|
| 1004 |
+
float w_tmp;
|
| 1005 |
+
|
| 1006 |
+
float weight[BLOCKWIDTH];
|
| 1007 |
+
for (int b = 0; b < batch; ++b){
|
| 1008 |
+
for (int head = 0; head < heads; ++head){
|
| 1009 |
+
int batch_shift = b * heads + head;
|
| 1010 |
+
for (k = 0; k < BLOCKWIDTH && h + k < vec_height; ++k){
|
| 1011 |
+
int k_w = k;
|
| 1012 |
+
int w_index = batch_shift * height * width + i + (k_w * width);
|
| 1013 |
+
float scale = __half2float(scales[batch_shift * width + w]);
|
| 1014 |
+
float zero = __half2float(zeros[batch_shift * width + w]);
|
| 1015 |
+
w_tmp = as_unsigned(mat[w_index]);
|
| 1016 |
+
weight[k] = scale *(w_tmp-zero);
|
| 1017 |
+
}
|
| 1018 |
+
|
| 1019 |
+
float res;
|
| 1020 |
+
for (int vr = 0; vr < vec_row; ++vr){
|
| 1021 |
+
res = 0;
|
| 1022 |
+
int vec_index = (batch_shift * vec_row + vr) * vec_height + blockIdx.x * BLOCKWIDTH + tid;
|
| 1023 |
+
if (vec_index < input_total) {
|
| 1024 |
+
blockvec[tid] = __half2float(vec[vec_index]);
|
| 1025 |
+
} else {
|
| 1026 |
+
blockvec[tid] = 0;
|
| 1027 |
+
}
|
| 1028 |
+
__syncthreads();
|
| 1029 |
+
for (k = 0; k < BLOCKWIDTH && h + k < vec_height; ++k){
|
| 1030 |
+
float temp_res = weight[k]*blockvec[k];
|
| 1031 |
+
res += temp_res;
|
| 1032 |
+
}
|
| 1033 |
+
int out_index = (batch_shift * vec_row + vr) * width + w;
|
| 1034 |
+
if (out_index < out_total) {
|
| 1035 |
+
atomicAdd(&mul[out_index], __float2half(res));
|
| 1036 |
+
}
|
| 1037 |
+
__syncthreads();
|
| 1038 |
+
}
|
| 1039 |
+
}
|
| 1040 |
+
}
|
| 1041 |
+
}
|
| 1042 |
+
|
| 1043 |
+
|
| 1044 |
+
|
| 1045 |
+
|
| 1046 |
+
void vecquant8matmul_batched_column_compression_faster_cuda(
|
| 1047 |
+
torch::Tensor vec,
|
| 1048 |
+
torch::Tensor mat,
|
| 1049 |
+
torch::Tensor mul,
|
| 1050 |
+
torch::Tensor scales,
|
| 1051 |
+
torch::Tensor zeros
|
| 1052 |
+
) {
|
| 1053 |
+
int batch = vec.size(0);
|
| 1054 |
+
int heads = vec.size(1);
|
| 1055 |
+
int vec_row = vec.size(2);
|
| 1056 |
+
int height = vec.size(3);
|
| 1057 |
+
int width = mat.size(3);
|
| 1058 |
+
|
| 1059 |
+
dim3 blocks(
|
| 1060 |
+
(height + BLOCKWIDTH - 1) / BLOCKWIDTH,
|
| 1061 |
+
(width + BLOCKWIDTH - 1) / BLOCKWIDTH
|
| 1062 |
+
);
|
| 1063 |
+
dim3 threads(BLOCKWIDTH);
|
| 1064 |
+
|
| 1065 |
+
VecQuant8BatchMatMulColumnCompressionKernel_faster<<<blocks, threads>>>(
|
| 1066 |
+
(half*) vec.data_ptr(),
|
| 1067 |
+
(uint8_t*) mat.data_ptr(),
|
| 1068 |
+
(half*) mul.data_ptr(),
|
| 1069 |
+
(half*) scales.data_ptr(),
|
| 1070 |
+
(half*) zeros.data_ptr(),
|
| 1071 |
+
batch, heads, vec_row, height, width
|
| 1072 |
+
);
|
| 1073 |
+
|
| 1074 |
+
}
|
| 1075 |
+
|
| 1076 |
+
__global__ void VecQuant8BatchMatMulColumnCompressionKernel_faster(
|
| 1077 |
+
const half* __restrict__ vec,
|
| 1078 |
+
const uint8_t* __restrict__ mat,
|
| 1079 |
+
half* __restrict__ mul,
|
| 1080 |
+
const half* __restrict__ scales,
|
| 1081 |
+
const half* __restrict__ zeros,
|
| 1082 |
+
int batch,
|
| 1083 |
+
int heads,
|
| 1084 |
+
int vec_row,
|
| 1085 |
+
int height,
|
| 1086 |
+
int width
|
| 1087 |
+
) {
|
| 1088 |
+
//int weight_total = batch * heads * height * width;
|
| 1089 |
+
int input_total = batch * heads * vec_row * height;
|
| 1090 |
+
int out_total = batch * heads * vec_row * width;
|
| 1091 |
+
int tid = threadIdx.x;
|
| 1092 |
+
int h = BLOCKWIDTH * blockIdx.x;
|
| 1093 |
+
int w = BLOCKWIDTH * blockIdx.y + tid;
|
| 1094 |
+
if (w >= width && tid >= height) {
|
| 1095 |
+
return;
|
| 1096 |
+
}
|
| 1097 |
+
|
| 1098 |
+
__shared__ float blockvec[BLOCKWIDTH];
|
| 1099 |
+
int k;
|
| 1100 |
+
float w_tmp;
|
| 1101 |
+
float weight[BLOCKWIDTH];
|
| 1102 |
+
|
| 1103 |
+
for (int b = 0; b < batch; ++b){
|
| 1104 |
+
for (int head = 0; head < heads; ++head){
|
| 1105 |
+
int batch_shift = b * heads + head;
|
| 1106 |
+
for (k = 0; k < BLOCKWIDTH; ++k){
|
| 1107 |
+
int w_index = (batch_shift * height + h + k) * width + w;
|
| 1108 |
+
float scale = __half2float(scales[batch_shift * height + h + k]);
|
| 1109 |
+
float zero = __half2float(zeros[batch_shift * height + h + k]);
|
| 1110 |
+
w_tmp = mat[w_index];
|
| 1111 |
+
weight[k] = scale * (w_tmp-zero);
|
| 1112 |
+
}
|
| 1113 |
+
|
| 1114 |
+
float res;
|
| 1115 |
+
for (int vr = 0; vr < vec_row; ++vr){
|
| 1116 |
+
res = 0;
|
| 1117 |
+
int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
|
| 1118 |
+
if (vec_index < input_total) {
|
| 1119 |
+
blockvec[tid] = __half2float(vec[vec_index]);
|
| 1120 |
+
} else {
|
| 1121 |
+
blockvec[tid] = 0;
|
| 1122 |
+
}
|
| 1123 |
+
__syncthreads();
|
| 1124 |
+
for (k = 0; k < BLOCKWIDTH; ++k){
|
| 1125 |
+
res += weight[k]*blockvec[k];
|
| 1126 |
+
}
|
| 1127 |
+
int out_index = (batch_shift * vec_row + vr) * width + w;
|
| 1128 |
+
if (out_index < out_total) {
|
| 1129 |
+
atomicAdd(&mul[out_index], __float2half(res));
|
| 1130 |
+
}
|
| 1131 |
+
__syncthreads();
|
| 1132 |
+
}
|
| 1133 |
+
}
|
| 1134 |
+
}
|
| 1135 |
+
}
|
| 1136 |
+
|
| 1137 |
+
|
| 1138 |
+
|
| 1139 |
+
void vecquant8matmul_batched_column_compression_old_cuda(
|
| 1140 |
+
torch::Tensor vec,
|
| 1141 |
+
torch::Tensor mat,
|
| 1142 |
+
torch::Tensor mul,
|
| 1143 |
+
torch::Tensor scales,
|
| 1144 |
+
torch::Tensor zeros
|
| 1145 |
+
) {
|
| 1146 |
+
int batch = vec.size(0);
|
| 1147 |
+
int heads = vec.size(1);
|
| 1148 |
+
int vec_row = vec.size(2);
|
| 1149 |
+
int height = vec.size(3);
|
| 1150 |
+
int width = mat.size(3);
|
| 1151 |
+
|
| 1152 |
+
dim3 blocks(
|
| 1153 |
+
(height + BLOCKWIDTH - 1) / BLOCKWIDTH,
|
| 1154 |
+
(width + BLOCKWIDTH - 1) / BLOCKWIDTH
|
| 1155 |
+
);
|
| 1156 |
+
dim3 threads(BLOCKWIDTH);
|
| 1157 |
+
|
| 1158 |
+
AT_DISPATCH_FLOATING_TYPES(
|
| 1159 |
+
vec.type(), "vecquant8matmul_batched_column_compression_old_cuda", ([&] {
|
| 1160 |
+
VecQuant8BatchMatMulColumnCompressionKernel_old<<<blocks, threads>>>(
|
| 1161 |
+
vec.data<scalar_t>(), mat.data<uint8_t>(), mul.data<scalar_t>(),
|
| 1162 |
+
scales.data<scalar_t>(), zeros.data<scalar_t>(),
|
| 1163 |
+
batch, heads, vec_row, height, width
|
| 1164 |
+
);
|
| 1165 |
+
})
|
| 1166 |
+
);
|
| 1167 |
+
|
| 1168 |
+
}
|
| 1169 |
+
|
| 1170 |
+
template <typename scalar_t>
|
| 1171 |
+
__global__ void VecQuant8BatchMatMulColumnCompressionKernel_old(
|
| 1172 |
+
const scalar_t* __restrict__ vec,
|
| 1173 |
+
const uint8_t* __restrict__ mat,
|
| 1174 |
+
scalar_t* __restrict__ mul,
|
| 1175 |
+
const scalar_t* __restrict__ scales,
|
| 1176 |
+
const scalar_t* __restrict__ zeros,
|
| 1177 |
+
int batch,
|
| 1178 |
+
int heads,
|
| 1179 |
+
int vec_row,
|
| 1180 |
+
int height,
|
| 1181 |
+
int width
|
| 1182 |
+
) {
|
| 1183 |
+
int weight_total = batch * heads * height * width;
|
| 1184 |
+
int input_total = batch * heads * vec_row * height;
|
| 1185 |
+
int out_total = batch * heads * vec_row * width;
|
| 1186 |
+
int tid = threadIdx.x;
|
| 1187 |
+
// h is index of height with step being BLOCKWIDTH
|
| 1188 |
+
int h = BLOCKWIDTH * blockIdx.x;
|
| 1189 |
+
// w is index of width with step being 1
|
| 1190 |
+
int w = BLOCKWIDTH * blockIdx.y + tid;
|
| 1191 |
+
if (w >= width && tid >= height) {
|
| 1192 |
+
return;
|
| 1193 |
+
}
|
| 1194 |
+
|
| 1195 |
+
__shared__ scalar_t blockvec[BLOCKWIDTH];
|
| 1196 |
+
int k;
|
| 1197 |
+
scalar_t w_tmp;
|
| 1198 |
+
|
| 1199 |
+
float weight[BLOCKWIDTH];
|
| 1200 |
+
|
| 1201 |
+
for (int b = 0; b < batch; ++b){
|
| 1202 |
+
for (int head = 0; head < heads; ++head){
|
| 1203 |
+
int batch_shift = b * heads + head;
|
| 1204 |
+
for (k = 0; k < BLOCKWIDTH && h + k < height; ++k){
|
| 1205 |
+
int w_index = (batch_shift * height + h + k) * width + w;
|
| 1206 |
+
if (w_index >= weight_total || w >= width) {
|
| 1207 |
+
weight[k] = 0;
|
| 1208 |
+
} else {
|
| 1209 |
+
scalar_t scale = scales[batch_shift * height + h + k];
|
| 1210 |
+
scalar_t zero = zeros[batch_shift * height + h + k];
|
| 1211 |
+
w_tmp = mat[w_index];
|
| 1212 |
+
weight[k] = scale * (w_tmp - zero);
|
| 1213 |
+
}
|
| 1214 |
+
}
|
| 1215 |
+
|
| 1216 |
+
scalar_t res;
|
| 1217 |
+
for (int vr = 0; vr < vec_row; ++vr){
|
| 1218 |
+
res = 0;
|
| 1219 |
+
int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
|
| 1220 |
+
if (vec_index < input_total) {
|
| 1221 |
+
blockvec[tid] = vec[vec_index];
|
| 1222 |
+
} else {
|
| 1223 |
+
blockvec[tid] = 0;
|
| 1224 |
+
}
|
| 1225 |
+
|
| 1226 |
+
__syncthreads();
|
| 1227 |
+
for (k = 0; k < BLOCKWIDTH && h + k < height; ++k){
|
| 1228 |
+
// res is the dot product of BLOCKWIDTH elements (part of width)
|
| 1229 |
+
res += weight[k] * blockvec[k];
|
| 1230 |
+
}
|
| 1231 |
+
// add res to the final result, final matrix shape: (batch, vec_row, width)
|
| 1232 |
+
int out_index = (batch_shift * vec_row + vr) * width + w;
|
| 1233 |
+
if (out_index < out_total) {
|
| 1234 |
+
atomicAdd(&mul[out_index], res);
|
| 1235 |
+
}
|
| 1236 |
+
__syncthreads();
|
| 1237 |
+
}
|
| 1238 |
+
}
|
| 1239 |
+
}
|
| 1240 |
+
}
|
| 1241 |
+
|
| 1242 |
+
|
| 1243 |
+
void vecquant4matmul_batched_old_cuda(
|
| 1244 |
+
torch::Tensor vec,
|
| 1245 |
+
torch::Tensor mat,
|
| 1246 |
+
torch::Tensor mul,
|
| 1247 |
+
torch::Tensor scales,
|
| 1248 |
+
torch::Tensor zeros
|
| 1249 |
+
) {
|
| 1250 |
+
int batch = vec.size(0);
|
| 1251 |
+
int heads = vec.size(1);
|
| 1252 |
+
int vec_row = vec.size(2);
|
| 1253 |
+
int vec_height = vec.size(3);
|
| 1254 |
+
int height = mat.size(2);
|
| 1255 |
+
int width = mat.size(3);
|
| 1256 |
+
int zero_width = zeros.size(2);
|
| 1257 |
+
|
| 1258 |
+
dim3 blocks(
|
| 1259 |
+
(height + BLOCKHEIGHT_OLD4 - 1) / BLOCKHEIGHT_OLD4,
|
| 1260 |
+
(width + BLOCKWIDTH - 1) / BLOCKWIDTH
|
| 1261 |
+
);
|
| 1262 |
+
dim3 threads(BLOCKWIDTH);
|
| 1263 |
+
|
| 1264 |
+
AT_DISPATCH_FLOATING_TYPES(
|
| 1265 |
+
vec.type(), "vecquant4matmul_batched_old_cuda", ([&] {
|
| 1266 |
+
VecQuant4BatchMatMulKernel_old<<<blocks, threads>>>(
|
| 1267 |
+
vec.data<scalar_t>(), mat.data<uint8_t>(), mul.data<scalar_t>(),
|
| 1268 |
+
scales.data<scalar_t>(), zeros.data<scalar_t>(),
|
| 1269 |
+
batch, heads, vec_row, vec_height, height, width, zero_width
|
| 1270 |
+
);
|
| 1271 |
+
})
|
| 1272 |
+
);
|
| 1273 |
+
|
| 1274 |
+
}
|
| 1275 |
+
|
| 1276 |
+
template <typename scalar_t>
|
| 1277 |
+
__global__ void VecQuant4BatchMatMulKernel_old(
|
| 1278 |
+
const scalar_t* __restrict__ vec,
|
| 1279 |
+
const uint8_t* __restrict__ mat,
|
| 1280 |
+
scalar_t* __restrict__ mul,
|
| 1281 |
+
const scalar_t* __restrict__ scales,
|
| 1282 |
+
const scalar_t* __restrict__ zeros,
|
| 1283 |
+
int batch,
|
| 1284 |
+
int heads,
|
| 1285 |
+
int vec_row,
|
| 1286 |
+
int vec_height,
|
| 1287 |
+
int height,
|
| 1288 |
+
int width,
|
| 1289 |
+
int zero_width
|
| 1290 |
+
) {
|
| 1291 |
+
int weight_total = batch * heads * height * width;
|
| 1292 |
+
int input_total = batch * heads * vec_row * vec_height;
|
| 1293 |
+
int out_total = batch * heads * vec_row * width;
|
| 1294 |
+
int tid = threadIdx.x;
|
| 1295 |
+
// h is index of height with step being BLOCKHEIGHT_OLD4
|
| 1296 |
+
int h = BLOCKHEIGHT_OLD4 * blockIdx.x;
|
| 1297 |
+
// w is index of width with step being 1
|
| 1298 |
+
int w = BLOCKWIDTH * blockIdx.y + tid;
|
| 1299 |
+
if (w >= width && tid >= vec_height) {
|
| 1300 |
+
return;
|
| 1301 |
+
}
|
| 1302 |
+
|
| 1303 |
+
__shared__ scalar_t blockvec[BLOCKWIDTH];
|
| 1304 |
+
// i is index of mat of block first row
|
| 1305 |
+
int i = width * h + w;
|
| 1306 |
+
int k;
|
| 1307 |
+
scalar_t w_tmp;
|
| 1308 |
+
|
| 1309 |
+
float weight[BLOCKWIDTH];
|
| 1310 |
+
for (int b = 0; b < batch; ++b){
|
| 1311 |
+
for (int head = 0; head < heads; ++head){
|
| 1312 |
+
int batch_shift = b * heads + head;
|
| 1313 |
+
for (k = 0; k < BLOCKWIDTH && h*2 + k < vec_height; ++k){
|
| 1314 |
+
int k_w = (k / 2);
|
| 1315 |
+
int k_bit = (k % 2) * 4;
|
| 1316 |
+
int w_index = batch_shift * height * width + i + (k_w * width);
|
| 1317 |
+
if (w_index >= weight_total || w >= width) {
|
| 1318 |
+
weight[k] = 0;
|
| 1319 |
+
} else {
|
| 1320 |
+
scalar_t scale = scales[batch_shift * width + w];
|
| 1321 |
+
scalar_t zero = zeros[batch_shift * width + w];
|
| 1322 |
+
w_tmp = ((as_unsigned(mat[w_index]) >> k_bit) & 0xF);
|
| 1323 |
+
weight[k] = scale * (w_tmp - zero);
|
| 1324 |
+
}
|
| 1325 |
+
}
|
| 1326 |
+
|
| 1327 |
+
scalar_t res;
|
| 1328 |
+
for (int vr = 0; vr < vec_row; ++vr){
|
| 1329 |
+
res = 0;
|
| 1330 |
+
int vec_index = (batch_shift * vec_row + vr) * vec_height + blockIdx.x * BLOCKWIDTH + tid;
|
| 1331 |
+
if (vec_index < input_total) {
|
| 1332 |
+
blockvec[tid] = vec[vec_index];
|
| 1333 |
+
} else {
|
| 1334 |
+
blockvec[tid] = 0;
|
| 1335 |
+
}
|
| 1336 |
+
|
| 1337 |
+
__syncthreads();
|
| 1338 |
+
for (k = 0; k < BLOCKWIDTH && h*2 + k < vec_height; ++k){
|
| 1339 |
+
// res is the dot product of BLOCKWIDTH elements (part of width)
|
| 1340 |
+
res += weight[k] * blockvec[k];
|
| 1341 |
+
}
|
| 1342 |
+
// add res to the final result, final matrix shape: (batch, vec_row, width)
|
| 1343 |
+
int out_index = (batch_shift * vec_row + vr) * width + w;
|
| 1344 |
+
if (out_index < out_total) {
|
| 1345 |
+
atomicAdd(&mul[out_index], res);
|
| 1346 |
+
}
|
| 1347 |
+
__syncthreads();
|
| 1348 |
+
}
|
| 1349 |
+
}
|
| 1350 |
+
}
|
| 1351 |
+
}
|
| 1352 |
+
|
| 1353 |
+
|
| 1354 |
+
|
| 1355 |
+
|
| 1356 |
+
|
| 1357 |
+
void vecquant4matmul_batched_column_compression_old_cuda(
|
| 1358 |
+
torch::Tensor vec,
|
| 1359 |
+
torch::Tensor mat,
|
| 1360 |
+
torch::Tensor mul,
|
| 1361 |
+
torch::Tensor scales,
|
| 1362 |
+
torch::Tensor zeros
|
| 1363 |
+
) {
|
| 1364 |
+
int batch = vec.size(0);
|
| 1365 |
+
int heads = vec.size(1);
|
| 1366 |
+
int vec_row = vec.size(2);
|
| 1367 |
+
int height = vec.size(3);
|
| 1368 |
+
int width = mat.size(3);
|
| 1369 |
+
|
| 1370 |
+
dim3 blocks(
|
| 1371 |
+
(height + BLOCKHEIGHT_OLD4 - 1) / BLOCKHEIGHT_OLD4,
|
| 1372 |
+
(width + BLOCKWIDTH - 1) / BLOCKWIDTH
|
| 1373 |
+
);
|
| 1374 |
+
dim3 threads(BLOCKWIDTH);
|
| 1375 |
+
|
| 1376 |
+
AT_DISPATCH_FLOATING_TYPES(
|
| 1377 |
+
vec.type(), "vecquant4matmul_batched_column_compression_old_cuda", ([&] {
|
| 1378 |
+
VecQuant4BatchMatMulColumnCompressionKernel_old<<<blocks, threads>>>(
|
| 1379 |
+
vec.data<scalar_t>(), mat.data<uint8_t>(), mul.data<scalar_t>(),
|
| 1380 |
+
scales.data<scalar_t>(), zeros.data<scalar_t>(),
|
| 1381 |
+
batch, heads, vec_row, height, width
|
| 1382 |
+
);
|
| 1383 |
+
})
|
| 1384 |
+
);
|
| 1385 |
+
|
| 1386 |
+
}
|
| 1387 |
+
|
| 1388 |
+
template <typename scalar_t>
|
| 1389 |
+
__global__ void VecQuant4BatchMatMulColumnCompressionKernel_old(
|
| 1390 |
+
const scalar_t* __restrict__ vec,
|
| 1391 |
+
const uint8_t* __restrict__ mat,
|
| 1392 |
+
scalar_t* __restrict__ mul,
|
| 1393 |
+
const scalar_t* __restrict__ scales,
|
| 1394 |
+
const scalar_t* __restrict__ zeros,
|
| 1395 |
+
int batch,
|
| 1396 |
+
int heads,
|
| 1397 |
+
int vec_row,
|
| 1398 |
+
int height,
|
| 1399 |
+
int width
|
| 1400 |
+
) {
|
| 1401 |
+
int weight_total = batch * heads * height * width;
|
| 1402 |
+
int input_total = batch * heads * vec_row * height;
|
| 1403 |
+
int out_total = batch * heads * vec_row * width;
|
| 1404 |
+
int tid = threadIdx.x;
|
| 1405 |
+
// h is index of height with step being BLOCKWIDTH
|
| 1406 |
+
int h = BLOCKHEIGHT_OLD4 * blockIdx.x;
|
| 1407 |
+
// w is index of width with step being 1
|
| 1408 |
+
int w = BLOCKWIDTH * blockIdx.y + tid;
|
| 1409 |
+
if (w >= width && tid >= height) {
|
| 1410 |
+
return;
|
| 1411 |
+
}
|
| 1412 |
+
|
| 1413 |
+
__shared__ scalar_t blockvec[BLOCKWIDTH];
|
| 1414 |
+
int k;
|
| 1415 |
+
scalar_t w_tmp;
|
| 1416 |
+
|
| 1417 |
+
float weight[BLOCKWIDTH];
|
| 1418 |
+
|
| 1419 |
+
for (int b = 0; b < batch; ++b){
|
| 1420 |
+
for (int head = 0; head < heads; ++head){
|
| 1421 |
+
int batch_shift = b * heads + head;
|
| 1422 |
+
for (k = 0; k < BLOCKWIDTH && h*2 + k < height; ++k){
|
| 1423 |
+
int k_w = (k / 2);
|
| 1424 |
+
int k_bit = (k % 2) * 4;
|
| 1425 |
+
int w_index = (batch_shift * height + h + k) * width + k_w;
|
| 1426 |
+
if (w_index >= weight_total || w >= width) {
|
| 1427 |
+
weight[k] = 0;
|
| 1428 |
+
} else {
|
| 1429 |
+
scalar_t scale = scales[batch_shift * height + h + k];
|
| 1430 |
+
scalar_t zero = zeros[batch_shift * height + h + k];
|
| 1431 |
+
w_tmp = ((as_unsigned(mat[w_index]) >> k_bit) & 0xF);
|
| 1432 |
+
weight[k] = scale * (w_tmp - zero);
|
| 1433 |
+
}
|
| 1434 |
+
}
|
| 1435 |
+
|
| 1436 |
+
scalar_t res;
|
| 1437 |
+
for (int vr = 0; vr < vec_row; ++vr){
|
| 1438 |
+
res = 0;
|
| 1439 |
+
int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
|
| 1440 |
+
if (vec_index < input_total) {
|
| 1441 |
+
blockvec[tid] = vec[vec_index];
|
| 1442 |
+
} else {
|
| 1443 |
+
blockvec[tid] = 0;
|
| 1444 |
+
}
|
| 1445 |
+
|
| 1446 |
+
__syncthreads();
|
| 1447 |
+
for (k = 0; k < BLOCKWIDTH && h*2 + k < height; ++k){
|
| 1448 |
+
// res is the dot product of BLOCKWIDTH elements (part of width)
|
| 1449 |
+
res += weight[k] * blockvec[k];
|
| 1450 |
+
}
|
| 1451 |
+
// add res to the final result, final matrix shape: (batch, vec_row, width)
|
| 1452 |
+
int out_index = (batch_shift * vec_row + vr) * width + w;
|
| 1453 |
+
if (out_index < out_total) {
|
| 1454 |
+
atomicAdd(&mul[out_index], res);
|
| 1455 |
+
}
|
| 1456 |
+
__syncthreads();
|
| 1457 |
+
}
|
| 1458 |
+
}
|
| 1459 |
+
}
|
| 1460 |
+
}
|
| 1461 |
+
|
| 1462 |
+
|
| 1463 |
+
|
| 1464 |
+
|
| 1465 |
+
|
| 1466 |
+
void vecquant8matmul_batched_faster_old_cuda(
|
| 1467 |
+
torch::Tensor vec,
|
| 1468 |
+
torch::Tensor mat,
|
| 1469 |
+
torch::Tensor mul,
|
| 1470 |
+
torch::Tensor scales,
|
| 1471 |
+
torch::Tensor zeros
|
| 1472 |
+
) {
|
| 1473 |
+
int batch = vec.size(0);
|
| 1474 |
+
int heads = vec.size(1);
|
| 1475 |
+
int vec_row = vec.size(2);
|
| 1476 |
+
int vec_height = vec.size(3);
|
| 1477 |
+
int height = mat.size(2);
|
| 1478 |
+
int width = mat.size(3);
|
| 1479 |
+
|
| 1480 |
+
dim3 blocks(
|
| 1481 |
+
(height + BLOCKWIDTH - 1) / BLOCKWIDTH,
|
| 1482 |
+
(width + BLOCKWIDTH - 1) / BLOCKWIDTH
|
| 1483 |
+
);
|
| 1484 |
+
dim3 threads(BLOCKWIDTH);
|
| 1485 |
+
|
| 1486 |
+
VecQuant8BatchMatMulKernel_faster_old<<<blocks, threads>>>(
|
| 1487 |
+
(half*) vec.data_ptr(),
|
| 1488 |
+
(uint8_t*) mat.data_ptr(),
|
| 1489 |
+
(half*) mul.data_ptr(),
|
| 1490 |
+
(half*) scales.data_ptr(),
|
| 1491 |
+
(half*) zeros.data_ptr(),
|
| 1492 |
+
batch, heads, vec_row, vec_height, height, width
|
| 1493 |
+
);
|
| 1494 |
+
}
|
| 1495 |
+
|
| 1496 |
+
|
| 1497 |
+
__global__ void VecQuant8BatchMatMulKernel_faster_old(
|
| 1498 |
+
const half* __restrict__ vec,
|
| 1499 |
+
const uint8_t* __restrict__ mat,
|
| 1500 |
+
half* __restrict__ mul,
|
| 1501 |
+
const half* __restrict__ scales,
|
| 1502 |
+
const half* __restrict__ zeros,
|
| 1503 |
+
int batch,
|
| 1504 |
+
int heads,
|
| 1505 |
+
int vec_row,
|
| 1506 |
+
int vec_height,
|
| 1507 |
+
int height,
|
| 1508 |
+
int width
|
| 1509 |
+
) {
|
| 1510 |
+
int weight_total = batch * heads * height * width;
|
| 1511 |
+
int input_total = batch * heads * vec_row * vec_height;
|
| 1512 |
+
int out_total = batch * heads * vec_row * width;
|
| 1513 |
+
int tid = threadIdx.x;
|
| 1514 |
+
const int BLOCKWIDTH_half = BLOCKWIDTH/2;
|
| 1515 |
+
|
| 1516 |
+
int h = BLOCKWIDTH * blockIdx.x; //head_dim, dim=-1
|
| 1517 |
+
int w = BLOCKWIDTH * blockIdx.y + tid; //seq-len, +0-256 ,dim=-2
|
| 1518 |
+
/*
|
| 1519 |
+
if (w >= width && tid >= vec_height) {
|
| 1520 |
+
return;
|
| 1521 |
+
}
|
| 1522 |
+
*/
|
| 1523 |
+
__shared__ half blockvec[BLOCKWIDTH]; //256
|
| 1524 |
+
int i = width * h + w;
|
| 1525 |
+
int k;
|
| 1526 |
+
|
| 1527 |
+
half w_tmp1 = __float2half(0);
|
| 1528 |
+
half w_tmp2 = __float2half(0);
|
| 1529 |
+
|
| 1530 |
+
half2 weight[BLOCKWIDTH_half];
|
| 1531 |
+
for (int b = 0; b < batch; ++b){
|
| 1532 |
+
for (int head = 0; head < heads; ++head){
|
| 1533 |
+
int batch_shift = b * heads + head;
|
| 1534 |
+
//int zero_index = batch_shift;
|
| 1535 |
+
for (k = 0; k < BLOCKWIDTH_half; ++k){
|
| 1536 |
+
int w_index1 = batch_shift * height * width + i + (2 * k * width); // [batch,head,h+k, w]
|
| 1537 |
+
int w_index2 = batch_shift * height * width + i + ((2 * k + 1) * width);
|
| 1538 |
+
int zero_index = batch_shift * width + w; // [batch,head, w]
|
| 1539 |
+
if (w_index1 >= weight_total || w >= width || (2 * k + h) >= height) {
|
| 1540 |
+
weight[k] = __float2half2_rn(0);
|
| 1541 |
+
} else {
|
| 1542 |
+
float zero_f=__half2float(zeros[zero_index]);
|
| 1543 |
+
float scale_f= __half2float(scales[zero_index]);
|
| 1544 |
+
if (w_index2 >= weight_total){
|
| 1545 |
+
w_tmp1 = __float2half((as_unsigned(mat[w_index1]) -zero_f)*scale_f);
|
| 1546 |
+
w_tmp2 = __float2half(0);
|
| 1547 |
+
weight[k] = __halves2half2(w_tmp1,w_tmp2);
|
| 1548 |
+
//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]));
|
| 1549 |
+
}else{
|
| 1550 |
+
w_tmp1 = __int2half_rn(as_unsigned(mat[w_index1]));
|
| 1551 |
+
w_tmp2 = __int2half_rn(as_unsigned(mat[w_index2]));
|
| 1552 |
+
|
| 1553 |
+
//weight[k] = __hmul2(__hsub2(__halves2half2(w_tmp1,w_tmp2), __halves2half2(zero,zero)),__halves2half2(scale,scale));
|
| 1554 |
+
weight[k] = __hfma2(__halves2half2(w_tmp1,w_tmp2), __float2half2_rn(scale_f), __float2half2_rn(-(scale_f * zero_f)));
|
| 1555 |
+
//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]));
|
| 1556 |
+
}
|
| 1557 |
+
}
|
| 1558 |
+
}
|
| 1559 |
+
|
| 1560 |
+
|
| 1561 |
+
for (int vr = 0; vr < vec_row; ++vr){
|
| 1562 |
+
float res=0;
|
| 1563 |
+
int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
|
| 1564 |
+
int out_index = (batch_shift * vec_row + vr) * width + w;
|
| 1565 |
+
if (vec_index < input_total) {
|
| 1566 |
+
//blockvec[tid] = __half2float(vec[vec_index]);// [batch, head, vr, tid(seq_len dim+)]
|
| 1567 |
+
blockvec[tid] = vec[vec_index];
|
| 1568 |
+
//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]);
|
| 1569 |
+
} else {
|
| 1570 |
+
blockvec[tid] = __float2half(0);
|
| 1571 |
+
}
|
| 1572 |
+
__syncthreads();
|
| 1573 |
+
if (out_index < out_total) {
|
| 1574 |
+
for (k = 0; k < BLOCKWIDTH_half; ++k){
|
| 1575 |
+
half2 res2 = __hmul2(weight[k],__halves2half2(blockvec[2*k],blockvec[2*k+1]));
|
| 1576 |
+
res += __low2float(res2) + __high2float(res2);
|
| 1577 |
+
}
|
| 1578 |
+
atomicAdd(&mul[out_index], __float2half(res));
|
| 1579 |
+
}
|
| 1580 |
+
__syncthreads();
|
| 1581 |
+
}
|
| 1582 |
+
}
|
| 1583 |
+
}
|
| 1584 |
+
}
|
| 1585 |
+
|
| 1586 |
+
|
| 1587 |
+
void vecquant8matmul_batched_column_compression_faster_old_cuda(
|
| 1588 |
+
torch::Tensor vec, // [batch,heads, seq_q, seq_v]
|
| 1589 |
+
torch::Tensor mat, // [batch,heads, seq_v, head_dim]
|
| 1590 |
+
torch::Tensor mul, // [batch,heads, seq_q,head_dim]
|
| 1591 |
+
torch::Tensor scales, // [batch,heads, head_dim]
|
| 1592 |
+
torch::Tensor zeros
|
| 1593 |
+
) {
|
| 1594 |
+
int batch = vec.size(0);
|
| 1595 |
+
int heads = vec.size(1);
|
| 1596 |
+
int vec_row = vec.size(2); //ql
|
| 1597 |
+
int height = mat.size(2); //vl
|
| 1598 |
+
int width = mat.size(3); //head_dim
|
| 1599 |
+
|
| 1600 |
+
dim3 blocks(
|
| 1601 |
+
(height + BLOCKWIDTH - 1) / BLOCKWIDTH,
|
| 1602 |
+
(width + BLOCKWIDTH - 1) / BLOCKWIDTH
|
| 1603 |
+
);
|
| 1604 |
+
dim3 threads(BLOCKWIDTH);
|
| 1605 |
+
|
| 1606 |
+
VecQuant8BatchMatMulColumnCompressionKernel_faster_old<<<blocks, threads>>>(
|
| 1607 |
+
(half*) vec.data_ptr(),
|
| 1608 |
+
(uint8_t*) mat.data_ptr(),
|
| 1609 |
+
(half*) mul.data_ptr(),
|
| 1610 |
+
(half*) scales.data_ptr(),
|
| 1611 |
+
(half*) zeros.data_ptr(),
|
| 1612 |
+
batch, heads, vec_row, height, width
|
| 1613 |
+
);
|
| 1614 |
+
|
| 1615 |
+
}
|
| 1616 |
+
|
| 1617 |
+
|
| 1618 |
+
__global__ void VecQuant8BatchMatMulColumnCompressionKernel_faster_old(
|
| 1619 |
+
const half* __restrict__ vec, // [batch,heads, seq_q, seq_v]
|
| 1620 |
+
const uint8_t* __restrict__ mat, // [batch,heads, seq_v, head_dim]
|
| 1621 |
+
half* __restrict__ mul, // [batch,heads, seq_q,head_dim]
|
| 1622 |
+
const half* __restrict__ scales, // [batch,heads, seq_v]
|
| 1623 |
+
const half* __restrict__ zeros,
|
| 1624 |
+
int batch,
|
| 1625 |
+
int heads,
|
| 1626 |
+
int vec_row, //seq_q
|
| 1627 |
+
int height, //seq_v
|
| 1628 |
+
int width //head_dim
|
| 1629 |
+
) {
|
| 1630 |
+
int weight_total = batch * heads * height * width;
|
| 1631 |
+
int input_total = batch * heads * vec_row * height;
|
| 1632 |
+
int out_total = batch * heads * vec_row * width;
|
| 1633 |
+
int tid = threadIdx.x;
|
| 1634 |
+
int h = BLOCKWIDTH * blockIdx.x; // vl
|
| 1635 |
+
int w = BLOCKWIDTH * blockIdx.y + tid; //head_dim + block
|
| 1636 |
+
if (w >= width && tid >= height) {
|
| 1637 |
+
return;
|
| 1638 |
+
}
|
| 1639 |
+
__shared__ half blockvec[BLOCKWIDTH];
|
| 1640 |
+
int k;
|
| 1641 |
+
half w_tmp1 = __float2half(0);
|
| 1642 |
+
half w_tmp2 = __float2half(0);
|
| 1643 |
+
int i = width * h + w;
|
| 1644 |
+
const int BLOCKWIDTH_half = BLOCKWIDTH/2;
|
| 1645 |
+
half2 weight[BLOCKWIDTH_half];
|
| 1646 |
+
|
| 1647 |
+
for (int b = 0; b < batch; ++b){
|
| 1648 |
+
for (int head = 0; head < heads; ++head){
|
| 1649 |
+
int batch_shift = b * heads + head;
|
| 1650 |
+
//int zero_index = batch_shift;
|
| 1651 |
+
for (k = 0; k < BLOCKWIDTH_half; ++k){
|
| 1652 |
+
int w_index1 = batch_shift * height * width + i + (2 * k) * width; // [batch,head, h+k, w]
|
| 1653 |
+
int w_index2 = batch_shift * height * width + i + ((2 * k + 1) * width);
|
| 1654 |
+
int zero_index1 = batch_shift * height + h + 2*k; // [batch,head, w]
|
| 1655 |
+
int zero_index2 = batch_shift * height + h + 2*k+1; // [batch,head, w]
|
| 1656 |
+
|
| 1657 |
+
if (w_index1 >= weight_total || (2 * k + h)>=height) {
|
| 1658 |
+
weight[k]=__float2half2_rn(0);
|
| 1659 |
+
} else{
|
| 1660 |
+
//int zero_index = batch_shift + h; // [batch,head, w]
|
| 1661 |
+
//float scale_f1 = __half2float(scales[zero_index1]);
|
| 1662 |
+
//float zero_f1 = __half2float(zeros[zero_index1]);
|
| 1663 |
+
if (w_index2>=weight_total){
|
| 1664 |
+
w_tmp1 = __float2half((as_unsigned(mat[w_index1]) - __half2float(zeros[zero_index1]))* __half2float(scales[zero_index1]));
|
| 1665 |
+
w_tmp2 = __float2half(0);
|
| 1666 |
+
weight[k] = __halves2half2(w_tmp1,w_tmp2);
|
| 1667 |
+
//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]));
|
| 1668 |
+
}else{
|
| 1669 |
+
w_tmp1 = __int2half_rn(as_unsigned(mat[w_index1]));
|
| 1670 |
+
w_tmp2 = __int2half_rn(as_unsigned(mat[w_index2]));
|
| 1671 |
+
half zero1=zeros[zero_index1];
|
| 1672 |
+
half zero2=zeros[zero_index2];
|
| 1673 |
+
half scale1=scales[zero_index1];
|
| 1674 |
+
half scale2=scales[zero_index2];
|
| 1675 |
+
weight[k] = __hmul2(__hsub2(__halves2half2(w_tmp1,w_tmp2), __halves2half2(zero1,zero2)),__halves2half2(scale1,scale2));
|
| 1676 |
+
//weight[k] = __hfma2(__halves2half2(w_tmp1,w_tmp2), __float2half2_rn(scale_f), __float2half2_rn(-(scale_f * zero_f)));
|
| 1677 |
+
//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]));
|
| 1678 |
+
}
|
| 1679 |
+
}
|
| 1680 |
+
}
|
| 1681 |
+
|
| 1682 |
+
|
| 1683 |
+
for (int vr = 0; vr < vec_row; ++vr){
|
| 1684 |
+
float res=0;
|
| 1685 |
+
int vec_index = (batch_shift * vec_row + vr) * height + blockIdx.x * BLOCKWIDTH + tid;
|
| 1686 |
+
int out_index = (batch_shift * vec_row + vr) * width + w;
|
| 1687 |
+
|
| 1688 |
+
if (vec_index < input_total) {
|
| 1689 |
+
//blockvec[tid] = __half2float(vec[vec_index]);
|
| 1690 |
+
blockvec[tid] = vec[vec_index];
|
| 1691 |
+
//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]);
|
| 1692 |
+
} else {
|
| 1693 |
+
blockvec[tid] = __float2half(0);
|
| 1694 |
+
//blockvec[tid] = 0;
|
| 1695 |
+
}
|
| 1696 |
+
__syncthreads();
|
| 1697 |
+
if (out_index < out_total) {
|
| 1698 |
+
for (k = 0; k < BLOCKWIDTH_half; ++k){
|
| 1699 |
+
half2 res2 = __hmul2(weight[k],__halves2half2(blockvec[2*k],blockvec[2*k+1]));
|
| 1700 |
+
res += __low2float(res2) + __high2float(res2);
|
| 1701 |
+
}
|
| 1702 |
+
atomicAdd(&mul[out_index], __float2half(res));
|
| 1703 |
+
}
|
| 1704 |
+
__syncthreads();
|
| 1705 |
+
}
|
| 1706 |
+
}
|
| 1707 |
+
}
|
| 1708 |
+
}
|
examples/react_prompt.md
ADDED
|
@@ -0,0 +1,249 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# ReAct Prompting 示例
|
| 2 |
+
|
| 3 |
+
本文档将介绍如何用 ReAct Prompting 技术命令千问使用工具。
|
| 4 |
+
|
| 5 |
+
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)。
|
| 6 |
+
|
| 7 |
+
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)。
|
| 8 |
+
|
| 9 |
+
## 准备工作一：样例问题、样例工具
|
| 10 |
+
|
| 11 |
+
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：
|
| 12 |
+
|
| 13 |
+
```py
|
| 14 |
+
query = '现在给我画个五彩斑斓的黑。'
|
| 15 |
+
|
| 16 |
+
TOOLS = [
|
| 17 |
+
{
|
| 18 |
+
'name_for_human':
|
| 19 |
+
'夸克搜索',
|
| 20 |
+
'name_for_model':
|
| 21 |
+
'quark_search',
|
| 22 |
+
'description_for_model':
|
| 23 |
+
'夸克搜索是一个通用搜索引擎，可用于访问互联网、查询百科知识、了解时事新闻等。',
|
| 24 |
+
'parameters': [{
|
| 25 |
+
'name': 'search_query',
|
| 26 |
+
'description': '搜索关键词或短语',
|
| 27 |
+
'required': True,
|
| 28 |
+
'schema': {
|
| 29 |
+
'type': 'string'
|
| 30 |
+
},
|
| 31 |
+
}],
|
| 32 |
+
},
|
| 33 |
+
{
|
| 34 |
+
'name_for_human':
|
| 35 |
+
'通义万相',
|
| 36 |
+
'name_for_model':
|
| 37 |
+
'image_gen',
|
| 38 |
+
'description_for_model':
|
| 39 |
+
'通义万相是一个AI绘画（图像生成）服务，输入文本描述，返回根据文本作画得到的图片的URL',
|
| 40 |
+
'parameters': [{
|
| 41 |
+
'name': 'query',
|
| 42 |
+
'description': '中文关键词，描述了希望图像具有什么内容',
|
| 43 |
+
'required': True,
|
| 44 |
+
'schema': {
|
| 45 |
+
'type': 'string'
|
| 46 |
+
},
|
| 47 |
+
}],
|
| 48 |
+
},
|
| 49 |
+
]
|
| 50 |
+
```
|
| 51 |
+
|
| 52 |
+
## 准备工作二：ReAct 模版
|
| 53 |
+
|
| 54 |
+
我们将使用如下的 ReAct prompt 模版来激发千问使用工具的能力。
|
| 55 |
+
|
| 56 |
+
```py
|
| 57 |
+
TOOL_DESC = """{name_for_model}: Call this tool to interact with the {name_for_human} API. What is the {name_for_human} API useful for? {description_for_model} Parameters: {parameters} Format the arguments as a JSON object."""
|
| 58 |
+
|
| 59 |
+
REACT_PROMPT = """Answer the following questions as best you can. You have access to the following tools:
|
| 60 |
+
|
| 61 |
+
{tool_descs}
|
| 62 |
+
|
| 63 |
+
Use the following format:
|
| 64 |
+
|
| 65 |
+
Question: the input question you must answer
|
| 66 |
+
Thought: you should always think about what to do
|
| 67 |
+
Action: the action to take, should be one of [{tool_names}]
|
| 68 |
+
Action Input: the input to the action
|
| 69 |
+
Observation: the result of the action
|
| 70 |
+
... (this Thought/Action/Action Input/Observation can be repeated zero or more times)
|
| 71 |
+
Thought: I now know the final answer
|
| 72 |
+
Final Answer: the final answer to the original input question
|
| 73 |
+
|
| 74 |
+
Begin!
|
| 75 |
+
|
| 76 |
+
Question: {query}"""
|
| 77 |
+
```
|
| 78 |
+
|
| 79 |
+
## 步骤一：让千问判断要调用什么工具、生成工具入参
|
| 80 |
+
|
| 81 |
+
首先我们需要根据 ReAct prompt 模版、query、工具的信息构建 prompt：
|
| 82 |
+
|
| 83 |
+
```py
|
| 84 |
+
tool_descs = []
|
| 85 |
+
tool_names = []
|
| 86 |
+
for info in TOOLS:
|
| 87 |
+
tool_descs.append(
|
| 88 |
+
TOOL_DESC.format(
|
| 89 |
+
name_for_model=info['name_for_model'],
|
| 90 |
+
name_for_human=info['name_for_human'],
|
| 91 |
+
description_for_model=info['description_for_model'],
|
| 92 |
+
parameters=json.dumps(
|
| 93 |
+
info['parameters'], ensure_ascii=False),
|
| 94 |
+
)
|
| 95 |
+
)
|
| 96 |
+
tool_names.append(info['name_for_model'])
|
| 97 |
+
tool_descs = '\n\n'.join(tool_descs)
|
| 98 |
+
tool_names = ','.join(tool_names)
|
| 99 |
+
|
| 100 |
+
prompt = REACT_PROMPT.format(tool_descs=tool_descs, tool_names=tool_names, query=query)
|
| 101 |
+
print(prompt)
|
| 102 |
+
```
|
| 103 |
+
|
| 104 |
+
打印出来的、构建好的 prompt 如下：
|
| 105 |
+
|
| 106 |
+
```
|
| 107 |
+
Answer the following questions as best you can. You have access to the following tools:
|
| 108 |
+
|
| 109 |
+
quark_search: Call this tool to interact with the 夸克搜索 API. What is the 夸克搜索 API useful for? 夸克搜索是一个通用搜索引擎，可用于访问互联网、查询百科知识、了解时事新闻等。 Parameters: [{"name": "search_query", "description": "搜索关键词或短语", "required": true, "schema": {"type": "string"}}] Format the arguments as a JSON object.
|
| 110 |
+
|
| 111 |
+
image_gen: Call this tool to interact with the 通义万相 API. What is the 通义万相 API useful for? 通义万相是一个AI绘画（图像生成）服务，输入文本描述，返回根据文本作画得到的图片的URL Parameters: [{"name": "query", "description": "中文关键词，描述了希望图像具有什么内容", "required": true, "schema": {"type": "string"}}] Format the arguments as a JSON object.
|
| 112 |
+
|
| 113 |
+
Use the following format:
|
| 114 |
+
|
| 115 |
+
Question: the input question you must answer
|
| 116 |
+
Thought: you should always think about what to do
|
| 117 |
+
Action: the action to take, should be one of [quark_search,image_gen]
|
| 118 |
+
Action Input: the input to the action
|
| 119 |
+
Observation: the result of the action
|
| 120 |
+
... (this Thought/Action/Action Input/Observation can be repeated zero or more times)
|
| 121 |
+
Thought: I now know the final answer
|
| 122 |
+
Final Answer: the final answer to the original input question
|
| 123 |
+
|
| 124 |
+
Begin!
|
| 125 |
+
|
| 126 |
+
Question: 现在给我画个五彩斑斓的黑。
|
| 127 |
+
```
|
| 128 |
+
|
| 129 |
+
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：
|
| 130 |
+
|
| 131 |
+
|
| 132 |
+
|
| 133 |
+
```
|
| 134 |
+
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。
|
| 135 |
+
Action: image_gen
|
| 136 |
+
Action Input: {"query": "五彩斑斓的黑"}
|
| 137 |
+
```
|
| 138 |
+
|
| 139 |
+
在得到这个结果后，调用千问的开发者可以通过简单的解析提取出 `{"query": "五彩斑斓的黑"}` 并基于这个解析结果调用文生图服务 —— 这部分逻辑需要开发者自行实现，或者也可以使用千问商业版，商业版本将内部集成相关逻辑。
|
| 140 |
+
|
| 141 |
+
## 步骤二：让千问根据插件返回结果继续作答
|
| 142 |
+
|
| 143 |
+
让我们假设文生图插件返回了如下结果：
|
| 144 |
+
|
| 145 |
+
```
|
| 146 |
+
{"status_code": 200, "request_id": "3d894da2-0e26-9b7c-bd90-102e5250ae03", "code": null, "message": "", "output": {"task_id": "2befaa09-a8b3-4740-ada9-4d00c2758b05", "task_status": "SUCCEEDED", "results": [{"url": "https://dashscope-result-sh.oss-cn-shanghai.aliyuncs.com/1e5e2015/20230801/1509/6b26bb83-469e-4c70-bff4-a9edd1e584f3-1.png"}], "task_metrics": {"TOTAL": 1, "SUCCEEDED": 1, "FAILED": 0}}, "usage": {"image_count": 1}}
|
| 147 |
+
```
|
| 148 |
+
|
| 149 |
+
|
| 150 |
+
|
| 151 |
+
接下来，我们可以将之前首次请求千问时用的 prompt 和 调用文生图插件的结果拼接成如下的新 prompt：
|
| 152 |
+
|
| 153 |
+
```
|
| 154 |
+
Answer the following questions as best you can. You have access to the following tools:
|
| 155 |
+
|
| 156 |
+
quark_search: Call this tool to interact with the 夸克搜索 API. What is the 夸克搜索 API useful for? 夸克搜索是一个通用搜索引擎，可用于访问互联网、查询百科知识、了解时事新闻等。 Parameters: [{"name": "search_query", "description": "搜索关键词或短语", "required": true, "schema": {"type": "string"}}] Format the arguments as a JSON object.
|
| 157 |
+
|
| 158 |
+
image_gen: Call this tool to interact with the 通义万相 API. What is the 通义万相 API useful for? 通义万相是一个AI绘画（图像生成）服务，输入文本描述，返回根据文本作画得到的图片的URL Parameters: [{"name": "query", "description": "中文关键词，描述了希望图像具有什么内容", "required": true, "schema": {"type": "string"}}] Format the arguments as a JSON object.
|
| 159 |
+
|
| 160 |
+
Use the following format:
|
| 161 |
+
|
| 162 |
+
Question: the input question you must answer
|
| 163 |
+
Thought: you should always think about what to do
|
| 164 |
+
Action: the action to take, should be one of [quark_search,image_gen]
|
| 165 |
+
Action Input: the input to the action
|
| 166 |
+
Observation: the result of the action
|
| 167 |
+
... (this Thought/Action/Action Input/Observation can be repeated zero or more times)
|
| 168 |
+
Thought: I now know the final answer
|
| 169 |
+
Final Answer: the final answer to the original input question
|
| 170 |
+
|
| 171 |
+
Begin!
|
| 172 |
+
|
| 173 |
+
Question: 现在给我画个五彩斑斓的黑。
|
| 174 |
+
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。
|
| 175 |
+
Action: image_gen
|
| 176 |
+
Action Input: {"query": "五彩斑斓的黑"}
|
| 177 |
+
Observation: {"status_code": 200, "request_id": "3d894da2-0e26-9b7c-bd90-102e5250ae03", "code": null, "message": "", "output": {"task_id": "2befaa09-a8b3-4740-ada9-4d00c2758b05", "task_status": "SUCCEEDED", "results": [{"url": "https://dashscope-result-sh.oss-cn-shanghai.aliyuncs.com/1e5e2015/20230801/1509/6b26bb83-469e-4c70-bff4-a9edd1e584f3-1.png"}], "task_metrics": {"TOTAL": 1, "SUCCEEDED": 1, "FAILED": 0}}, "usage": {"image_count": 1}}
|
| 178 |
+
```
|
| 179 |
+
|
| 180 |
+
用这个新的拼接了文生图插件结果的新 prompt 去调用千问，将得到如下的最终回复：
|
| 181 |
+
|
| 182 |
+
|
| 183 |
+
|
| 184 |
+
```
|
| 185 |
+
Thought: 我已经成功使用通义万相API生成了一张五彩斑斓的黑的图片。
|
| 186 |
+
Final Answer: 我已经成功使用通义万相API生成了一张五彩斑斓的黑的图片https://dashscope-result-sh.oss-cn-shanghai.aliyuncs.com/1e5e2015/20230801/1509/6b26bb83-469e-4c70-bff4-a9edd1e584f3-1.png。
|
| 187 |
+
```
|
| 188 |
+
|
| 189 |
+
虽然对于文生图来说，这个第二次调用千问的步骤显得多余。但是对于搜索插件、代码执行插件、计算器插件等别的插件来说，这个第二次调用千问的步骤给了千问提炼、总结插件返回结果的机会。
|
| 190 |
+
|
| 191 |
+
## FAQ
|
| 192 |
+
|
| 193 |
+
**怎么配置 "Observation" 这个 stop word？**
|
| 194 |
+
|
| 195 |
+
通过 chat 接口的 stop_words_ids 指定：
|
| 196 |
+
```py
|
| 197 |
+
react_stop_words = [
|
| 198 |
+
# tokenizer.encode('Observation'), # [37763, 367]
|
| 199 |
+
tokenizer.encode('Observation:'), # [37763, 367, 25]
|
| 200 |
+
tokenizer.encode('Observation:\n'), # [37763, 367, 510]
|
| 201 |
+
]
|
| 202 |
+
response, history = model.chat(
|
| 203 |
+
tokenizer, query, history,
|
| 204 |
+
stop_words_ids=react_stop_words # 此接口用于增加 stop words
|
| 205 |
+
)
|
| 206 |
+
```
|
| 207 |
+
|
| 208 |
+
如果报错称不存在 stop_words_ids 此参数，可能是因为您用了老的代码，请重新执行 from_pretrained 拉取新的代码和模型。
|
| 209 |
+
|
| 210 |
+
需要注意的是，当前的 tokenizer 对 `\n` 有一系列较复杂的聚合操作。比如例子中的`:\n`这两个字符便被聚合成了一个 token。因此配置 stop words 需要非常细致地预估 tokenizer 的行为。
|
| 211 |
+
|
| 212 |
+
**对 top_p 等推理参数有调参建议吗？**
|
| 213 |
+
|
| 214 |
+
通常来讲，较低的 top_p 会有更高的准确度，但会牺牲回答的多样性、且更易出现重复某个词句的现象。
|
| 215 |
+
|
| 216 |
+
可以按如下方式调整 top_p 为 0.5：
|
| 217 |
+
```py
|
| 218 |
+
model.generation_config.top_p = 0.5
|
| 219 |
+
```
|
| 220 |
+
|
| 221 |
+
特别的，可以用如下方式关闭 top-p sampling，改用 greedy sampling，效果上相当于 top_p=0 或 temperature=0：
|
| 222 |
+
```py
|
| 223 |
+
model.generation_config.do_sample = False # greedy decoding
|
| 224 |
+
```
|
| 225 |
+
|
| 226 |
+
此外，我们在 `model.chat()` 接口也提供了调整 top_p 等参数的接口。
|
| 227 |
+
|
| 228 |
+
**有解析Action、Action Input的参考代码吗？**
|
| 229 |
+
|
| 230 |
+
有的，可以参考：
|
| 231 |
+
```py
|
| 232 |
+
def parse_latest_plugin_call(text: str) -> Tuple[str, str]:
|
| 233 |
+
i = text.rfind('\nAction:')
|
| 234 |
+
j = text.rfind('\nAction Input:')
|
| 235 |
+
k = text.rfind('\nObservation:')
|
| 236 |
+
if 0 <= i < j: # If the text has `Action` and `Action input`,
|
| 237 |
+
if k < j: # but does not contain `Observation`,
|
| 238 |
+
# then it is likely that `Observation` is ommited by the LLM,
|
| 239 |
+
# because the output text may have discarded the stop word.
|
| 240 |
+
text = text.rstrip() + '\nObservation:' # Add it back.
|
| 241 |
+
k = text.rfind('\nObservation:')
|
| 242 |
+
if 0 <= i < j < k:
|
| 243 |
+
plugin_name = text[i + len('\nAction:'):j].strip()
|
| 244 |
+
plugin_args = text[j + len('\nAction Input:'):k].strip()
|
| 245 |
+
return plugin_name, plugin_args
|
| 246 |
+
return '', ''
|
| 247 |
+
```
|
| 248 |
+
|
| 249 |
+
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。
|
modeling_qwen.py
CHANGED
|
@@ -13,7 +13,6 @@ import torch
|
|
| 13 |
import torch.nn.functional as F
|
| 14 |
import torch.utils.checkpoint
|
| 15 |
import warnings
|
| 16 |
-
from torch.cuda.amp import autocast
|
| 17 |
|
| 18 |
from torch.nn import CrossEntropyLoss
|
| 19 |
from transformers import PreTrainedTokenizer, GenerationConfig, StoppingCriteriaList
|
|
@@ -79,9 +78,10 @@ We detect you have activated flash attention support, but running model computat
|
|
| 79 |
apply_rotary_emb_func = None
|
| 80 |
rms_norm = None
|
| 81 |
flash_attn_unpadded_func = None
|
|
|
|
| 82 |
|
| 83 |
def _import_flash_attn():
|
| 84 |
-
global apply_rotary_emb_func, rms_norm, flash_attn_unpadded_func
|
| 85 |
try:
|
| 86 |
from flash_attn.layers.rotary import apply_rotary_emb_func as __apply_rotary_emb_func
|
| 87 |
apply_rotary_emb_func = __apply_rotary_emb_func
|
|
@@ -102,14 +102,18 @@ def _import_flash_attn():
|
|
| 102 |
|
| 103 |
try:
|
| 104 |
import flash_attn
|
|
|
|
| 105 |
if not hasattr(flash_attn, '__version__'):
|
| 106 |
from flash_attn.flash_attn_interface import flash_attn_unpadded_func as __flash_attn_unpadded_func
|
| 107 |
else:
|
| 108 |
if int(flash_attn.__version__.split(".")[0]) >= 2:
|
|
|
|
|
|
|
| 109 |
from flash_attn.flash_attn_interface import flash_attn_varlen_func as __flash_attn_unpadded_func
|
| 110 |
else:
|
| 111 |
from flash_attn.flash_attn_interface import flash_attn_unpadded_func as __flash_attn_unpadded_func
|
| 112 |
flash_attn_unpadded_func = __flash_attn_unpadded_func
|
|
|
|
| 113 |
except ImportError:
|
| 114 |
logger.warn(
|
| 115 |
"Warning: import flash_attn fail, please install FlashAttention to get higher efficiency "
|
|
@@ -182,6 +186,11 @@ class FlashSelfAttention(torch.nn.Module):
|
|
| 182 |
seqlen_k = k.shape[1]
|
| 183 |
seqlen_out = seqlen_q
|
| 184 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 185 |
q, k, v = [rearrange(x, "b s ... -> (b s) ...") for x in [q, k, v]]
|
| 186 |
cu_seqlens_q = torch.arange(
|
| 187 |
0,
|
|
@@ -311,7 +320,7 @@ class QWenAttention(nn.Module):
|
|
| 311 |
warnings.warn("Failed to import KV cache kernels.")
|
| 312 |
self.cache_kernels = None
|
| 313 |
|
| 314 |
-
def _attn(self, query, key, value,
|
| 315 |
device = query.device
|
| 316 |
if self.use_cache_quantization:
|
| 317 |
qk, qk_scale, qk_zero = key
|
|
@@ -336,26 +345,13 @@ class QWenAttention(nn.Module):
|
|
| 336 |
size_temp = value[0].size(-1)
|
| 337 |
else:
|
| 338 |
size_temp = value.size(-1)
|
| 339 |
-
attn_weights = attn_weights /
|
| 340 |
-
|
| 341 |
-
size_temp ** 0.5,
|
| 342 |
-
dtype=attn_weights.dtype,
|
| 343 |
-
device=attn_weights.device,
|
| 344 |
-
)
|
| 345 |
-
if self.use_cache_quantization:
|
| 346 |
-
query_length, key_length = query.size(-2), key[0].size(-2)
|
| 347 |
-
else:
|
| 348 |
-
query_length, key_length = query.size(-2), key.size(-2)
|
| 349 |
-
causal_mask = registered_causal_mask[
|
| 350 |
-
:, :, key_length - query_length : key_length, :key_length
|
| 351 |
-
]
|
| 352 |
mask_value = torch.finfo(attn_weights.dtype).min
|
| 353 |
-
|
| 354 |
-
attn_weights.
|
| 355 |
-
|
| 356 |
-
|
| 357 |
-
causal_mask, attn_weights.to(attn_weights.dtype), mask_value
|
| 358 |
-
)
|
| 359 |
|
| 360 |
if attention_mask is not None:
|
| 361 |
attn_weights = attn_weights + attention_mask
|
|
@@ -482,7 +478,8 @@ class QWenAttention(nn.Module):
|
|
| 482 |
else:
|
| 483 |
present = None
|
| 484 |
|
| 485 |
-
if self.
|
|
|
|
| 486 |
if self.use_cache_quantization:
|
| 487 |
seq_start = key[0].size(2) - query.size(1)
|
| 488 |
seq_end = key[0].size(2)
|
|
@@ -501,15 +498,19 @@ class QWenAttention(nn.Module):
|
|
| 501 |
q, k, v = query, key, value
|
| 502 |
attn_output = self.core_attention_flash(q, k, v, attention_mask=attention_mask)
|
| 503 |
else:
|
| 504 |
-
|
| 505 |
-
|
| 506 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 507 |
query = query.permute(0, 2, 1, 3)
|
| 508 |
if not self.use_cache_quantization:
|
| 509 |
key = key.permute(0, 2, 1, 3)
|
| 510 |
value = value.permute(0, 2, 1, 3)
|
| 511 |
if (
|
| 512 |
-
|
| 513 |
and self.use_flash_attn
|
| 514 |
and flash_attn_unpadded_func is not None
|
| 515 |
and not self.is_fp32
|
|
@@ -518,13 +519,12 @@ class QWenAttention(nn.Module):
|
|
| 518 |
raise Exception(_ERROR_INPUT_CPU_QUERY_WITH_FLASH_ATTN_ACTIVATED)
|
| 519 |
|
| 520 |
if not self.use_cache_quantization and SUPPORT_TORCH2:
|
| 521 |
-
causal_mask = registered_causal_mask[
|
| 522 |
-
:, :, key.size(-2) - query.size(-2): key.size(-2), :key.size(-2)
|
| 523 |
-
]
|
| 524 |
if attention_mask is not None:
|
| 525 |
attention_mask = attention_mask.expand(
|
| 526 |
-1, -1, causal_mask.size(2), -1
|
| 527 |
-
)
|
|
|
|
|
|
|
| 528 |
else:
|
| 529 |
attention_mask = causal_mask
|
| 530 |
attn_output = F.scaled_dot_product_attention(
|
|
@@ -533,7 +533,7 @@ class QWenAttention(nn.Module):
|
|
| 533 |
attn_weight = None
|
| 534 |
else:
|
| 535 |
attn_output, attn_weight = self._attn(
|
| 536 |
-
query, key, value,
|
| 537 |
)
|
| 538 |
context_layer = self._merge_heads(
|
| 539 |
attn_output, self.num_heads, self.head_dim
|
|
@@ -549,6 +549,8 @@ class QWenAttention(nn.Module):
|
|
| 549 |
and not self.is_fp32
|
| 550 |
):
|
| 551 |
raise ValueError("Cannot output attentions while using flash-attn")
|
|
|
|
|
|
|
| 552 |
else:
|
| 553 |
outputs += (attn_weight,)
|
| 554 |
|
|
@@ -574,6 +576,7 @@ class QWenMLP(nn.Module):
|
|
| 574 |
output = self.c_proj(intermediate_parallel)
|
| 575 |
return output
|
| 576 |
|
|
|
|
| 577 |
class QWenBlock(nn.Module):
|
| 578 |
def __init__(self, config):
|
| 579 |
super().__init__()
|
|
@@ -642,6 +645,7 @@ class QWenPreTrainedModel(PreTrainedModel):
|
|
| 642 |
is_parallelizable = False
|
| 643 |
supports_gradient_checkpointing = True
|
| 644 |
_no_split_modules = ["QWenBlock"]
|
|
|
|
| 645 |
|
| 646 |
def __init__(self, *inputs, **kwargs):
|
| 647 |
super().__init__(*inputs, **kwargs)
|
|
@@ -933,11 +937,6 @@ class QWenLMHeadModel(QWenPreTrainedModel):
|
|
| 933 |
assert (
|
| 934 |
config.bf16 + config.fp16 + config.fp32 <= 1
|
| 935 |
), "Only one of \"bf16\", \"fp16\", \"fp32\" can be true"
|
| 936 |
-
logger.warn(
|
| 937 |
-
"Warning: please make sure that you are using the latest codes and checkpoints, "
|
| 938 |
-
"especially if you used Qwen-7B before 09.25.2023."
|
| 939 |
-
"请使用最新模型和代码，尤其如果你在9月25日前已经开始使用Qwen-7B，千万注意不要使用错误代码和模型。"
|
| 940 |
-
)
|
| 941 |
|
| 942 |
autoset_precision = config.bf16 + config.fp16 + config.fp32 == 0
|
| 943 |
|
|
@@ -990,7 +989,6 @@ class QWenLMHeadModel(QWenPreTrainedModel):
|
|
| 990 |
self.lm_head.half()
|
| 991 |
self.post_init()
|
| 992 |
|
| 993 |
-
|
| 994 |
def get_output_embeddings(self):
|
| 995 |
return self.lm_head
|
| 996 |
|
|
@@ -1000,22 +998,13 @@ class QWenLMHeadModel(QWenPreTrainedModel):
|
|
| 1000 |
def prepare_inputs_for_generation(
|
| 1001 |
self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs
|
| 1002 |
):
|
| 1003 |
-
token_type_ids = kwargs.get("token_type_ids", None)
|
| 1004 |
if past_key_values:
|
| 1005 |
input_ids = input_ids[:, -1].unsqueeze(-1)
|
| 1006 |
-
if token_type_ids is not None:
|
| 1007 |
-
token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
|
| 1008 |
|
| 1009 |
-
|
| 1010 |
-
|
| 1011 |
-
|
| 1012 |
-
if attention_mask is not None and position_ids is None:
|
| 1013 |
-
position_ids = attention_mask.long().cumsum(-1) - 1
|
| 1014 |
-
position_ids.masked_fill_(attention_mask == 0, 1)
|
| 1015 |
-
if past_key_values:
|
| 1016 |
-
position_ids = position_ids[:, -1].unsqueeze(-1)
|
| 1017 |
else:
|
| 1018 |
-
|
| 1019 |
|
| 1020 |
if inputs_embeds is not None and past_key_values is None:
|
| 1021 |
model_inputs = {"inputs_embeds": inputs_embeds}
|
|
@@ -1026,9 +1015,7 @@ class QWenLMHeadModel(QWenPreTrainedModel):
|
|
| 1026 |
{
|
| 1027 |
"past_key_values": past_key_values,
|
| 1028 |
"use_cache": kwargs.get("use_cache"),
|
| 1029 |
-
"position_ids": position_ids,
|
| 1030 |
"attention_mask": attention_mask,
|
| 1031 |
-
"token_type_ids": token_type_ids,
|
| 1032 |
}
|
| 1033 |
)
|
| 1034 |
return model_inputs
|
|
@@ -1299,8 +1286,7 @@ class RotaryEmbedding(torch.nn.Module):
|
|
| 1299 |
self._ntk_alpha_cached = 1.0
|
| 1300 |
self._ntk_alpha_cached_list = [1.0]
|
| 1301 |
|
| 1302 |
-
def update_rotary_pos_emb_cache(self,
|
| 1303 |
-
seqlen = max_seq_len + offset
|
| 1304 |
if seqlen > self._seq_len_cached or ntk_alpha != self._ntk_alpha_cached:
|
| 1305 |
base = self.base * ntk_alpha ** (self.dim / (self.dim - 2))
|
| 1306 |
self.inv_freq = 1.0 / (
|
|
@@ -1323,10 +1309,10 @@ class RotaryEmbedding(torch.nn.Module):
|
|
| 1323 |
cos, sin = emb.cos(), emb.sin()
|
| 1324 |
self._rotary_pos_emb_cache = [cos, sin]
|
| 1325 |
|
| 1326 |
-
def forward(self, max_seq_len,
|
| 1327 |
-
self.update_rotary_pos_emb_cache(max_seq_len,
|
| 1328 |
cos, sin = self._rotary_pos_emb_cache
|
| 1329 |
-
return [cos[:,
|
| 1330 |
|
| 1331 |
|
| 1332 |
def _rotate_half(x):
|
|
@@ -1338,21 +1324,28 @@ def _rotate_half(x):
|
|
| 1338 |
|
| 1339 |
|
| 1340 |
def apply_rotary_pos_emb(t, freqs):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1341 |
cos, sin = freqs
|
|
|
|
| 1342 |
if apply_rotary_emb_func is not None and t.is_cuda:
|
| 1343 |
-
|
| 1344 |
-
|
| 1345 |
-
|
| 1346 |
-
|
| 1347 |
-
|
|
|
|
| 1348 |
else:
|
| 1349 |
-
|
| 1350 |
-
cos
|
| 1351 |
-
|
| 1352 |
-
t_ = t_.float()
|
| 1353 |
-
t_pass_ = t_pass_.float()
|
| 1354 |
-
t_ = (t_ * cos) + (_rotate_half(t_) * sin)
|
| 1355 |
-
return torch.cat((t_, t_pass_), dim=-1).type_as(t)
|
| 1356 |
|
| 1357 |
|
| 1358 |
class RMSNorm(torch.nn.Module):
|
|
|
|
| 13 |
import torch.nn.functional as F
|
| 14 |
import torch.utils.checkpoint
|
| 15 |
import warnings
|
|
|
|
| 16 |
|
| 17 |
from torch.nn import CrossEntropyLoss
|
| 18 |
from transformers import PreTrainedTokenizer, GenerationConfig, StoppingCriteriaList
|
|
|
|
| 78 |
apply_rotary_emb_func = None
|
| 79 |
rms_norm = None
|
| 80 |
flash_attn_unpadded_func = None
|
| 81 |
+
flash_attn_func = None
|
| 82 |
|
| 83 |
def _import_flash_attn():
|
| 84 |
+
global apply_rotary_emb_func, rms_norm, flash_attn_unpadded_func, flash_attn_func
|
| 85 |
try:
|
| 86 |
from flash_attn.layers.rotary import apply_rotary_emb_func as __apply_rotary_emb_func
|
| 87 |
apply_rotary_emb_func = __apply_rotary_emb_func
|
|
|
|
| 102 |
|
| 103 |
try:
|
| 104 |
import flash_attn
|
| 105 |
+
_flash_attn_func = None
|
| 106 |
if not hasattr(flash_attn, '__version__'):
|
| 107 |
from flash_attn.flash_attn_interface import flash_attn_unpadded_func as __flash_attn_unpadded_func
|
| 108 |
else:
|
| 109 |
if int(flash_attn.__version__.split(".")[0]) >= 2:
|
| 110 |
+
if int(flash_attn.__version__.split(".")[1]) >= 1:
|
| 111 |
+
from flash_attn.flash_attn_interface import flash_attn_func as _flash_attn_func
|
| 112 |
from flash_attn.flash_attn_interface import flash_attn_varlen_func as __flash_attn_unpadded_func
|
| 113 |
else:
|
| 114 |
from flash_attn.flash_attn_interface import flash_attn_unpadded_func as __flash_attn_unpadded_func
|
| 115 |
flash_attn_unpadded_func = __flash_attn_unpadded_func
|
| 116 |
+
flash_attn_func = _flash_attn_func
|
| 117 |
except ImportError:
|
| 118 |
logger.warn(
|
| 119 |
"Warning: import flash_attn fail, please install FlashAttention to get higher efficiency "
|
|
|
|
| 186 |
seqlen_k = k.shape[1]
|
| 187 |
seqlen_out = seqlen_q
|
| 188 |
|
| 189 |
+
if flash_attn_func is not None and batch_size == 1:
|
| 190 |
+
dropout_p = self.dropout_p if self.training else 0
|
| 191 |
+
output = flash_attn_func(q, k, v, dropout_p, softmax_scale=self.softmax_scale, causal=self.causal)
|
| 192 |
+
return output
|
| 193 |
+
|
| 194 |
q, k, v = [rearrange(x, "b s ... -> (b s) ...") for x in [q, k, v]]
|
| 195 |
cu_seqlens_q = torch.arange(
|
| 196 |
0,
|
|
|
|
| 320 |
warnings.warn("Failed to import KV cache kernels.")
|
| 321 |
self.cache_kernels = None
|
| 322 |
|
| 323 |
+
def _attn(self, query, key, value, causal_mask=None, attention_mask=None, head_mask=None):
|
| 324 |
device = query.device
|
| 325 |
if self.use_cache_quantization:
|
| 326 |
qk, qk_scale, qk_zero = key
|
|
|
|
| 345 |
size_temp = value[0].size(-1)
|
| 346 |
else:
|
| 347 |
size_temp = value.size(-1)
|
| 348 |
+
attn_weights = attn_weights / (size_temp ** 0.5)
|
| 349 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 350 |
mask_value = torch.finfo(attn_weights.dtype).min
|
| 351 |
+
if causal_mask is not None:
|
| 352 |
+
attn_weights = torch.where(
|
| 353 |
+
causal_mask, attn_weights.to(attn_weights.dtype), mask_value
|
| 354 |
+
)
|
|
|
|
|
|
|
| 355 |
|
| 356 |
if attention_mask is not None:
|
| 357 |
attn_weights = attn_weights + attention_mask
|
|
|
|
| 478 |
else:
|
| 479 |
present = None
|
| 480 |
|
| 481 |
+
key_size = key[0].size(2) if self.use_cache_quantization else key.size(1)
|
| 482 |
+
if key_size > self.seq_length and self.use_logn_attn and not self.training:
|
| 483 |
if self.use_cache_quantization:
|
| 484 |
seq_start = key[0].size(2) - query.size(1)
|
| 485 |
seq_end = key[0].size(2)
|
|
|
|
| 498 |
q, k, v = query, key, value
|
| 499 |
attn_output = self.core_attention_flash(q, k, v, attention_mask=attention_mask)
|
| 500 |
else:
|
| 501 |
+
key_size = key[0].size(2) if self.use_cache_quantization else key.size(1)
|
| 502 |
+
if query.size(1) == key_size:
|
| 503 |
+
causal_mask = torch.tril(
|
| 504 |
+
torch.ones((key_size, key_size), dtype=torch.bool, device=query.device)
|
| 505 |
+
).view(1, 1, key_size, key_size)
|
| 506 |
+
else:
|
| 507 |
+
causal_mask = None
|
| 508 |
query = query.permute(0, 2, 1, 3)
|
| 509 |
if not self.use_cache_quantization:
|
| 510 |
key = key.permute(0, 2, 1, 3)
|
| 511 |
value = value.permute(0, 2, 1, 3)
|
| 512 |
if (
|
| 513 |
+
causal_mask is None
|
| 514 |
and self.use_flash_attn
|
| 515 |
and flash_attn_unpadded_func is not None
|
| 516 |
and not self.is_fp32
|
|
|
|
| 519 |
raise Exception(_ERROR_INPUT_CPU_QUERY_WITH_FLASH_ATTN_ACTIVATED)
|
| 520 |
|
| 521 |
if not self.use_cache_quantization and SUPPORT_TORCH2:
|
|
|
|
|
|
|
|
|
|
| 522 |
if attention_mask is not None:
|
| 523 |
attention_mask = attention_mask.expand(
|
| 524 |
-1, -1, causal_mask.size(2), -1
|
| 525 |
+
)
|
| 526 |
+
if causal_mask is not None:
|
| 527 |
+
attention_mask.masked_fill(~causal_mask, torch.finfo(query.dtype).min)
|
| 528 |
else:
|
| 529 |
attention_mask = causal_mask
|
| 530 |
attn_output = F.scaled_dot_product_attention(
|
|
|
|
| 533 |
attn_weight = None
|
| 534 |
else:
|
| 535 |
attn_output, attn_weight = self._attn(
|
| 536 |
+
query, key, value, causal_mask, attention_mask, head_mask
|
| 537 |
)
|
| 538 |
context_layer = self._merge_heads(
|
| 539 |
attn_output, self.num_heads, self.head_dim
|
|
|
|
| 549 |
and not self.is_fp32
|
| 550 |
):
|
| 551 |
raise ValueError("Cannot output attentions while using flash-attn")
|
| 552 |
+
elif not self.use_cache_quantization and SUPPORT_TORCH2:
|
| 553 |
+
raise ValueError("Cannot output attentions while using scaled_dot_product_attention")
|
| 554 |
else:
|
| 555 |
outputs += (attn_weight,)
|
| 556 |
|
|
|
|
| 576 |
output = self.c_proj(intermediate_parallel)
|
| 577 |
return output
|
| 578 |
|
| 579 |
+
|
| 580 |
class QWenBlock(nn.Module):
|
| 581 |
def __init__(self, config):
|
| 582 |
super().__init__()
|
|
|
|
| 645 |
is_parallelizable = False
|
| 646 |
supports_gradient_checkpointing = True
|
| 647 |
_no_split_modules = ["QWenBlock"]
|
| 648 |
+
_skip_keys_device_placement = "past_key_values"
|
| 649 |
|
| 650 |
def __init__(self, *inputs, **kwargs):
|
| 651 |
super().__init__(*inputs, **kwargs)
|
|
|
|
| 937 |
assert (
|
| 938 |
config.bf16 + config.fp16 + config.fp32 <= 1
|
| 939 |
), "Only one of \"bf16\", \"fp16\", \"fp32\" can be true"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 940 |
|
| 941 |
autoset_precision = config.bf16 + config.fp16 + config.fp32 == 0
|
| 942 |
|
|
|
|
| 989 |
self.lm_head.half()
|
| 990 |
self.post_init()
|
| 991 |
|
|
|
|
| 992 |
def get_output_embeddings(self):
|
| 993 |
return self.lm_head
|
| 994 |
|
|
|
|
| 998 |
def prepare_inputs_for_generation(
|
| 999 |
self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs
|
| 1000 |
):
|
|
|
|
| 1001 |
if past_key_values:
|
| 1002 |
input_ids = input_ids[:, -1].unsqueeze(-1)
|
|
|
|
|
|
|
| 1003 |
|
| 1004 |
+
if input_ids.size(0) == 1:
|
| 1005 |
+
attention_mask = None
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1006 |
else:
|
| 1007 |
+
attention_mask = kwargs.get("attention_mask", None)
|
| 1008 |
|
| 1009 |
if inputs_embeds is not None and past_key_values is None:
|
| 1010 |
model_inputs = {"inputs_embeds": inputs_embeds}
|
|
|
|
| 1015 |
{
|
| 1016 |
"past_key_values": past_key_values,
|
| 1017 |
"use_cache": kwargs.get("use_cache"),
|
|
|
|
| 1018 |
"attention_mask": attention_mask,
|
|
|
|
| 1019 |
}
|
| 1020 |
)
|
| 1021 |
return model_inputs
|
|
|
|
| 1286 |
self._ntk_alpha_cached = 1.0
|
| 1287 |
self._ntk_alpha_cached_list = [1.0]
|
| 1288 |
|
| 1289 |
+
def update_rotary_pos_emb_cache(self, seqlen, ntk_alpha=1.0):
|
|
|
|
| 1290 |
if seqlen > self._seq_len_cached or ntk_alpha != self._ntk_alpha_cached:
|
| 1291 |
base = self.base * ntk_alpha ** (self.dim / (self.dim - 2))
|
| 1292 |
self.inv_freq = 1.0 / (
|
|
|
|
| 1309 |
cos, sin = emb.cos(), emb.sin()
|
| 1310 |
self._rotary_pos_emb_cache = [cos, sin]
|
| 1311 |
|
| 1312 |
+
def forward(self, max_seq_len, ntk_alpha=1.0):
|
| 1313 |
+
self.update_rotary_pos_emb_cache(max_seq_len, ntk_alpha)
|
| 1314 |
cos, sin = self._rotary_pos_emb_cache
|
| 1315 |
+
return [cos[:, :max_seq_len], sin[:, :max_seq_len]]
|
| 1316 |
|
| 1317 |
|
| 1318 |
def _rotate_half(x):
|
|
|
|
| 1324 |
|
| 1325 |
|
| 1326 |
def apply_rotary_pos_emb(t, freqs):
|
| 1327 |
+
""" Apply rotary embedding to the first rotary_dim of the iput
|
| 1328 |
+
|
| 1329 |
+
Arguments:
|
| 1330 |
+
t (tensor(batch_size, seq_len, n_head, head_dim)):
|
| 1331 |
+
the input embedding/hidden states
|
| 1332 |
+
freqs (list[tensor(1, seq_len, 1, rotary_dim), tensor(1, seq_len, 1, rotary_dim)]):
|
| 1333 |
+
the cached cos/sin position embeddings
|
| 1334 |
+
"""
|
| 1335 |
+
rot_dim = freqs[0].shape[-1]
|
| 1336 |
cos, sin = freqs
|
| 1337 |
+
t_float = t.float()
|
| 1338 |
if apply_rotary_emb_func is not None and t.is_cuda:
|
| 1339 |
+
# apply_rotary_emb in flash_attn requires cos/sin to be of
|
| 1340 |
+
# shape (seqlen, rotary_dim / 2) and apply rotary embedding
|
| 1341 |
+
# to the first rotary_dim of the input
|
| 1342 |
+
cos = cos.squeeze(0).squeeze(1)[:, : rot_dim // 2]
|
| 1343 |
+
sin = sin.squeeze(0).squeeze(1)[:, : rot_dim // 2]
|
| 1344 |
+
return apply_rotary_emb_func(t_float, cos, sin).type_as(t)
|
| 1345 |
else:
|
| 1346 |
+
t_rot, t_pass = t_float[..., :rot_dim], t_float[..., rot_dim:]
|
| 1347 |
+
t_rot = (t_rot * cos) + (_rotate_half(t_rot) * sin)
|
| 1348 |
+
return torch.cat((t_rot, t_pass), dim=-1).type_as(t)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1349 |
|
| 1350 |
|
| 1351 |
class RMSNorm(torch.nn.Module):
|