Text Generation
Transformers
Safetensors
Japanese
English
qwen
custom_code
ishishow commited on
Commit
8322cb1
1 Parent(s): a14e0f2

Upload 12 files

Browse files
.gitattributes CHANGED
@@ -33,4 +33,3 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
- model.safetensors filter=lfs diff=lfs merge=lfs -text
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
LICENSE ADDED
@@ -0,0 +1,53 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Tongyi Qianwen LICENSE AGREEMENT
2
+
3
+ Tongyi Qianwen Release Date: August 3, 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 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 (including Qwen model and Qwen-Chat model), and software and algorithms, consisting of trained model weights, parameters (including optimizer states), machine-learning model code, inference-enabling code, training-enabling code, fine-tuning enabling code and other elements of the foregoing distributed by Us.
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
+
18
+ 2. Grant of Rights
19
+ 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.
20
+
21
+ 3. Redistribution
22
+ 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:
23
+ a. You shall give any other recipients of the Materials or derivative works a copy of this Agreement;
24
+ b. You shall cause any modified files to carry prominent notices stating that You changed the files;
25
+ 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 LICENSE AGREEMENT, Copyright (c) Alibaba Cloud. All Rights Reserved."; and
26
+ 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.
27
+
28
+ 4. Restrictions
29
+ If you are commercially using the Materials, and your product or service has more than 100 million monthly active users, You shall request a license from Us. You cannot exercise your rights under this Agreement without our express authorization.
30
+
31
+ 5. Rules of use
32
+ 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.
33
+ 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).
34
+
35
+ 6. Intellectual Property
36
+ 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.
37
+ 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.
38
+ 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.
39
+
40
+ 7. Disclaimer of Warranty and Limitation of Liability
41
+
42
+ 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.
43
+ 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.
44
+ 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.
45
+ 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.
46
+
47
+ 8. Survival and Termination.
48
+ 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.
49
+ 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 7 and 9 shall survive the termination of this Agreement.
50
+
51
+ 9. Governing Law and Jurisdiction.
52
+ 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.
53
+ b. The People's Courts in Hangzhou City shall have exclusive jurisdiction over any dispute arising out of this Agreement.
NOTICE ADDED
@@ -0,0 +1,77 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+
55
+ ------------- LICENSE FOR PanQiWei AutoGPTQ code --------------
56
+
57
+ MIT License
58
+
59
+ Copyright (c) 2023 潘其威(William)
60
+
61
+ Permission is hereby granted, free of charge, to any person obtaining a copy
62
+ of this software and associated documentation files (the "Software"), to deal
63
+ in the Software without restriction, including without limitation the rights
64
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
65
+ copies of the Software, and to permit persons to whom the Software is
66
+ furnished to do so, subject to the following conditions:
67
+
68
+ The above copyright notice and this permission notice shall be included in all
69
+ copies or substantial portions of the Software.
70
+
71
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
72
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
73
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
74
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
75
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
76
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
77
+ SOFTWARE.
README.md ADDED
@@ -0,0 +1,137 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ thumbnail: https://github.com/rinnakk/japanese-pretrained-models/blob/master/rinna.png
3
+ datasets:
4
+ - databricks/databricks-dolly-15k
5
+ - kunishou/databricks-dolly-15k-ja
6
+ - izumi-lab/llm-japanese-dataset
7
+ language:
8
+ - ja
9
+ - en
10
+ tags:
11
+ - qwen
12
+ inference: false
13
+ ---
14
+
15
+ # `rinna/nekomata-14b-instruction`
16
+
17
+ ![rinna-icon](./rinna.png)
18
+
19
+ # Overview
20
+ The model is the instruction-tuned version of [`rinna/nekomata-14b`](https://huggingface.co/rinna/nekomata-14b). It adopts the Alpaca input format.
21
+
22
+ * **Model architecture**
23
+
24
+ A 40-layer, 5120-hidden-size transformer-based language model. Please refer to the [Qwen paper](https://arxiv.org/abs/2309.16609) for architecture details.
25
+
26
+ * **Fine-tuning**
27
+
28
+ The fine-tuning data is the subset of the following datasets.
29
+ * [Databricks Dolly data](https://huggingface.co/datasets/databricks/databricks-dolly-15k)
30
+ * [Japanese Databricks Dolly data](https://huggingface.co/datasets/kunishou/databricks-dolly-15k-ja)
31
+ * [FLAN Instruction Tuning data](https://github.com/google-research/FLAN) and its Japanese translation
32
+ * [Izumi lab LLM Japanese dataset](https://github.com/masanorihirano/llm-japanese-dataset/tree/main)
33
+ * The following sections are used
34
+ * alt
35
+ * aozora-txt
36
+ * CourseraParallel
37
+ * ParaNatCom
38
+ * Tab-delimited_Bilingual_Sentence_Pairs
39
+ * tanaka-corpus
40
+ * wikinews
41
+ * wordnet
42
+ * yasashi-japanese
43
+ * The [remaining sections](https://github.com/masanorihirano/llm-japanese-dataset/tree/main/datasets-cc-by-sa) contain commonly used evaluation corpora so they are skipped to prevent data leak.
44
+
45
+ * **Authors**
46
+
47
+ - [Tianyu Zhao](https://huggingface.co/tianyuz)
48
+ - [Kei Sawada](https://huggingface.co/keisawada)
49
+
50
+ ---
51
+
52
+ # Benchmarking
53
+ Please refer to [rinna's LM benchmark page](https://rinnakk.github.io/research/benchmarks/lm/index.html).
54
+
55
+ ---
56
+
57
+ # How to use the model
58
+
59
+ ~~~~python
60
+ import torch
61
+ from transformers import AutoTokenizer, AutoModelForCausalLM
62
+
63
+ tokenizer = AutoTokenizer.from_pretrained("rinna/nekomata-14b-instruction", trust_remote_code=True)
64
+
65
+ # Use GPU with bf16
66
+ # model = AutoModelForCausalLM.from_pretrained("rinna/nekomata-14b-instruction", device_map="auto", trust_remote_code=True, bf16=True)
67
+
68
+ # Use GPU with fp16
69
+ # model = AutoModelForCausalLM.from_pretrained("rinna/nekomata-14b-instruction", device_map="auto", trust_remote_code=True, fp16=True)
70
+
71
+ # Use CPU
72
+ # model = AutoModelForCausalLM.from_pretrained("rinna/nekomata-14b-instruction", device_map="cpu", trust_remote_code=True)
73
+
74
+ # Automatically select device and precision
75
+ model = AutoModelForCausalLM.from_pretrained("rinna/nekomata-14b-instruction", device_map="auto", trust_remote_code=True)
76
+
77
+ instruction = "次の日本語を英語に翻訳してください。"
78
+ input = "大規模言語モデル(だいきぼげんごモデル、英: large language model、LLM)は、多数のパラメータ(数千万から数十億)を持つ人工ニューラルネットワークで構成されるコンピュータ言語モデルで、膨大なラベルなしテキストを使用して自己教師あり学習または半教師あり学習によって訓練が行われる。"
79
+ prompt = f"""
80
+ 以下は、タスクを説明する指示と、文脈のある入力の組み合わせです。要求を適切に満たす応答を書きなさい。
81
+
82
+ ### 指示:
83
+ {instruction}
84
+
85
+ ### 入力:
86
+ {input}
87
+
88
+ ### 応答:
89
+ """
90
+ token_ids = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt")
91
+
92
+ with torch.no_grad():
93
+ output_ids = model.generate(
94
+ token_ids.to(model.device),
95
+ max_new_tokens=200,
96
+ do_sample=True,
97
+ temperature=0.5,
98
+ pad_token_id=tokenizer.pad_token_id,
99
+ bos_token_id=tokenizer.bos_token_id,
100
+ eos_token_id=tokenizer.eos_token_id
101
+ )
102
+
103
+ output = tokenizer.decode(output_ids.tolist()[0])
104
+ print(output)
105
+ """
106
+ 以下は、タスクを説明する指示と、文脈のある入力の組み合わせです。要求を適切に満たす応答を書きなさい。
107
+
108
+ ### 指示:
109
+ 次の日本語を英語に翻訳してください。
110
+
111
+ ### 入力:
112
+ 大規模言語モデル(だいきぼげんごモデル、英: large language model、LLM)は、多数のパラメータ(数千万から数十億)を持つ人工ニューラルネットワークで構成されるコンピュータ言語モデルで、膨大なラベルなしテキストを使 用して自己教師あり学習または半教師あり学習によって訓練が行われる。
113
+
114
+ ### 応答:
115
+ A large language model (LLM) is a computer language model composed of artificial neural networks with many parameters (from tens of millions to billions) trained by self-supervised learning or semi-supervised learning using a large amount of unlabeled text.<|endoftext|>
116
+ """
117
+ ~~~~
118
+
119
+ ---
120
+
121
+ # Tokenization
122
+ Please refer to [`rinna/nekomata-14b`](https://huggingface.co/rinna/nekomata-14b) for tokenization details.
123
+
124
+ ---
125
+
126
+ # How to cite
127
+ ~~~
128
+ @misc{RinnaNekomataInstruction14b,
129
+ url={https://huggingface.co/rinna/nekomata-14b-instruction},
130
+ title={rinna/nekomata-14b-instruction},
131
+ author={Zhao, Tianyu and Sawada, Kei}
132
+ }
133
+ ~~~
134
+ ---
135
+
136
+ # License
137
+ [Tongyi Qianwen LICENSE AGREEMENT](https://github.com/QwenLM/Qwen/blob/main/Tongyi%20Qianwen%20LICENSE%20AGREEMENT)
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
+ }
config.json ADDED
@@ -0,0 +1,37 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "QWenLMHeadModel"
4
+ ],
5
+ "auto_map": {
6
+ "AutoConfig": "configuration_qwen.QWenConfig",
7
+ "AutoModelForCausalLM": "modeling_qwen.QWenLMHeadModel"
8
+ },
9
+ "attn_dropout_prob": 0.0,
10
+ "bf16": false,
11
+ "emb_dropout_prob": 0.0,
12
+ "fp16": false,
13
+ "fp32": false,
14
+ "hidden_size": 5120,
15
+ "intermediate_size": 27392,
16
+ "initializer_range": 0.02,
17
+ "kv_channels": 128,
18
+ "layer_norm_epsilon": 1e-06,
19
+ "max_position_embeddings": 8192,
20
+ "model_type": "qwen",
21
+ "no_bias": true,
22
+ "num_attention_heads": 40,
23
+ "num_hidden_layers": 40,
24
+ "onnx_safe": null,
25
+ "rotary_emb_base": 10000,
26
+ "rotary_pct": 1.0,
27
+ "scale_attn_weights": true,
28
+ "seq_length": 2048,
29
+ "tie_word_embeddings": false,
30
+ "tokenizer_class": "QWenTokenizer",
31
+ "transformers_version": "4.32.0",
32
+ "use_cache": true,
33
+ "use_dynamic_ntk": true,
34
+ "use_flash_attn": "auto",
35
+ "use_logn_attn": true,
36
+ "vocab_size": 152064
37
+ }
configuration_qwen.py ADDED
@@ -0,0 +1,71 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Alibaba Cloud.
2
+ #
3
+ # This source code is licensed under the license found in the
4
+ # LICENSE file in the root directory of this source tree.
5
+
6
+ from transformers import PretrainedConfig
7
+
8
+
9
+ class QWenConfig(PretrainedConfig):
10
+ model_type = "qwen"
11
+ keys_to_ignore_at_inference = ["past_key_values"]
12
+
13
+ def __init__(
14
+ self,
15
+ vocab_size=151936,
16
+ hidden_size=4096,
17
+ num_hidden_layers=32,
18
+ num_attention_heads=32,
19
+ emb_dropout_prob=0.0,
20
+ attn_dropout_prob=0.0,
21
+ layer_norm_epsilon=1e-6,
22
+ initializer_range=0.02,
23
+ max_position_embeddings=8192,
24
+ scale_attn_weights=True,
25
+ use_cache=True,
26
+ bf16=False,
27
+ fp16=False,
28
+ fp32=False,
29
+ kv_channels=128,
30
+ rotary_pct=1.0,
31
+ rotary_emb_base=10000,
32
+ use_dynamic_ntk=True,
33
+ use_logn_attn=True,
34
+ use_flash_attn="auto",
35
+ intermediate_size=22016,
36
+ no_bias=True,
37
+ tie_word_embeddings=False,
38
+ use_cache_quantization=False,
39
+ use_cache_kernel=False,
40
+ softmax_in_fp32=False,
41
+ **kwargs,
42
+ ):
43
+ self.vocab_size = vocab_size
44
+ self.hidden_size = hidden_size
45
+ self.intermediate_size = intermediate_size
46
+ self.num_hidden_layers = num_hidden_layers
47
+ self.num_attention_heads = num_attention_heads
48
+ self.emb_dropout_prob = emb_dropout_prob
49
+ self.attn_dropout_prob = attn_dropout_prob
50
+ self.layer_norm_epsilon = layer_norm_epsilon
51
+ self.initializer_range = initializer_range
52
+ self.scale_attn_weights = scale_attn_weights
53
+ self.use_cache = use_cache
54
+ self.max_position_embeddings = max_position_embeddings
55
+ self.bf16 = bf16
56
+ self.fp16 = fp16
57
+ self.fp32 = fp32
58
+ self.kv_channels = kv_channels
59
+ self.rotary_pct = rotary_pct
60
+ self.rotary_emb_base = rotary_emb_base
61
+ self.use_dynamic_ntk = use_dynamic_ntk
62
+ self.use_logn_attn = use_logn_attn
63
+ self.use_flash_attn = use_flash_attn
64
+ self.no_bias = no_bias
65
+ self.use_cache_quantization = use_cache_quantization
66
+ self.use_cache_kernel = use_cache_kernel
67
+ self.softmax_in_fp32 = softmax_in_fp32
68
+ super().__init__(
69
+ tie_word_embeddings=tie_word_embeddings,
70
+ **kwargs
71
+ )
cpp_kernels.py ADDED
@@ -0,0 +1,55 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from torch.utils import cpp_extension
2
+ import pathlib
3
+ import os
4
+ import subprocess
5
+
6
+ def _get_cuda_bare_metal_version(cuda_dir):
7
+ raw_output = subprocess.check_output([cuda_dir + "/bin/nvcc", "-V"],
8
+ universal_newlines=True)
9
+ output = raw_output.split()
10
+ release_idx = output.index("release") + 1
11
+ release = output[release_idx].split(".")
12
+ bare_metal_major = release[0]
13
+ bare_metal_minor = release[1][0]
14
+
15
+ return raw_output, bare_metal_major, bare_metal_minor
16
+
17
+ def _create_build_dir(buildpath):
18
+ try:
19
+ os.mkdir(buildpath)
20
+ except OSError:
21
+ if not os.path.isdir(buildpath):
22
+ print(f"Creation of the build directory {buildpath} failed")
23
+
24
+ # Check if cuda 11 is installed for compute capability 8.0
25
+ cc_flag = []
26
+ _, bare_metal_major, bare_metal_minor = _get_cuda_bare_metal_version(cpp_extension.CUDA_HOME)
27
+ if int(bare_metal_major) >= 11:
28
+ cc_flag.append('-gencode')
29
+ cc_flag.append('arch=compute_80,code=sm_80')
30
+ if int(bare_metal_minor) >= 7:
31
+ cc_flag.append('-gencode')
32
+ cc_flag.append('arch=compute_90,code=sm_90')
33
+
34
+ # Build path
35
+ srcpath = pathlib.Path(__file__).parent.absolute()
36
+ buildpath = srcpath / 'build'
37
+ _create_build_dir(buildpath)
38
+
39
+ def _cpp_extention_load_helper(name, sources, extra_cuda_flags):
40
+ return cpp_extension.load(
41
+ name=name,
42
+ sources=sources,
43
+ build_directory=buildpath,
44
+ extra_cflags=['-O3', ],
45
+ extra_cuda_cflags=['-O3',
46
+ '-gencode', 'arch=compute_70,code=sm_70',
47
+ '--use_fast_math'] + extra_cuda_flags + cc_flag,
48
+ verbose=1
49
+ )
50
+
51
+ extra_flags = []
52
+
53
+ cache_autogptq_cuda_256_sources = ["./cache_autogptq_cuda_256.cpp",
54
+ "./cache_autogptq_cuda_kernel_256.cu"]
55
+ cache_autogptq_cuda_256 = _cpp_extention_load_helper("cache_autogptq_cuda_256", cache_autogptq_cuda_256_sources, extra_flags)
modeling_qwen.py ADDED
@@ -0,0 +1,1363 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Alibaba Cloud.
2
+ #
3
+ # This source code is licensed under the license found in the
4
+ # LICENSE file in the root directory of this source tree.
5
+
6
+ import copy
7
+ import importlib
8
+ import math
9
+ import pathlib
10
+ from typing import TYPE_CHECKING, Optional, Tuple, Union, Callable, List, Any, Generator
11
+
12
+ import torch
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
19
+ from transformers.generation.logits_process import LogitsProcessorList
20
+
21
+ if TYPE_CHECKING:
22
+ from transformers.generation.streamers import BaseStreamer
23
+ from transformers.generation.utils import GenerateOutput
24
+ from transformers.modeling_outputs import (
25
+ BaseModelOutputWithPast,
26
+ CausalLMOutputWithPast,
27
+ )
28
+ from transformers.modeling_utils import PreTrainedModel
29
+ from transformers.utils import logging
30
+
31
+ try:
32
+ from einops import rearrange
33
+ except ImportError:
34
+ rearrange = None
35
+ from torch import nn
36
+
37
+ SUPPORT_CUDA = torch.cuda.is_available()
38
+ SUPPORT_BF16 = SUPPORT_CUDA and torch.cuda.is_bf16_supported()
39
+ SUPPORT_FP16 = SUPPORT_CUDA and torch.cuda.get_device_capability(0)[0] >= 7
40
+ SUPPORT_TORCH2 = hasattr(torch, '__version__') and int(torch.__version__.split(".")[0]) >= 2
41
+
42
+
43
+ from .configuration_qwen import QWenConfig
44
+ from .qwen_generation_utils import (
45
+ HistoryType,
46
+ make_context,
47
+ decode_tokens,
48
+ get_stop_words_ids,
49
+ StopWordsLogitsProcessor,
50
+ )
51
+
52
+
53
+ logger = logging.get_logger(__name__)
54
+
55
+ _CHECKPOINT_FOR_DOC = "qwen"
56
+ _CONFIG_FOR_DOC = "QWenConfig"
57
+
58
+ QWen_PRETRAINED_MODEL_ARCHIVE_LIST = ["qwen-7b"]
59
+
60
+ _ERROR_BAD_CHAT_FORMAT = """\
61
+ We detect you are probably using the pretrained model (rather than chat model) for chatting, since the chat_format in generation_config is not "chatml".
62
+ If you are directly using the model downloaded from Huggingface, please make sure you are using our "Qwen/Qwen-7B-Chat" Huggingface model (rather than "Qwen/Qwen-7B") when you call model.chat().
63
+ 我们检测到您可能在使用预训练模型(而非chat模型)进行多轮chat,因为您当前在generation_config指定的chat_format,并未设置为我们在对话中所支持的"chatml"格式。
64
+ 如果您在直接使用我们从Huggingface提供的模型,请确保您在调用model.chat()时,使用的是"Qwen/Qwen-7B-Chat"模型(而非"Qwen/Qwen-7B"预训练模型)。
65
+ """
66
+
67
+ _SENTINEL = object()
68
+ _ERROR_STREAM_IN_CHAT = """\
69
+ Pass argument `stream` to model.chat() is buggy, deprecated, and marked for removal. Please use model.chat_stream(...) instead of model.chat(..., stream=True).
70
+ 向model.chat()传入参数stream的用法可能存在Bug,该用法已被废弃,将在未来被移除。请使用model.chat_stream(...)代替model.chat(..., stream=True)。
71
+ """
72
+
73
+ _ERROR_INPUT_CPU_QUERY_WITH_FLASH_ATTN_ACTIVATED = """\
74
+ We detect you have activated flash attention support, but running model computation on CPU. Please make sure that your input data has been placed on GPU. If you actually want to run CPU computation, please following the readme and set device_map="cpu" to disable flash attention when loading the model (calling AutoModelForCausalLM.from_pretrained).
75
+ 检测到您的模型已激活了flash attention支持,但正在执行CPU运算任务。如使用flash attention,请您确认模型输入已经传到GPU上。如果您确认要执行CPU运算,请您在载入模型(调用AutoModelForCausalLM.from_pretrained)时,按照readme说法,指定device_map="cpu"以禁用flash attention。
76
+ """
77
+
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
88
+ except ImportError:
89
+ logger.warn(
90
+ "Warning: import flash_attn rotary fail, please install FlashAttention rotary to get higher efficiency "
91
+ "https://github.com/Dao-AILab/flash-attention/tree/main/csrc/rotary"
92
+ )
93
+
94
+ try:
95
+ from flash_attn.ops.rms_norm import rms_norm as __rms_norm
96
+ rms_norm = __rms_norm
97
+ except ImportError:
98
+ logger.warn(
99
+ "Warning: import flash_attn rms_norm fail, please install FlashAttention layer_norm to get higher efficiency "
100
+ "https://github.com/Dao-AILab/flash-attention/tree/main/csrc/layer_norm"
101
+ )
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 "
120
+ "https://github.com/Dao-AILab/flash-attention"
121
+ )
122
+
123
+ def quantize_cache_v(fdata, bits, qmax, qmin):
124
+ # b, s, head, h-dim->b, head, s, h-dim
125
+ qtype = torch.uint8
126
+ device = fdata.device
127
+ shape = fdata.shape
128
+
129
+ fdata_cal = torch.flatten(fdata, 2)
130
+ fmax = torch.amax(fdata_cal, dim=-1, keepdim=True)
131
+ fmin = torch.amin(fdata_cal, dim=-1, keepdim=True)
132
+ # Compute params
133
+ if qmax.device != fmax.device:
134
+ qmax = qmax.to(device)
135
+ qmin = qmin.to(device)
136
+ scale = (fmax - fmin) / (qmax - qmin)
137
+ zero = qmin - fmin / scale
138
+ scale = scale.unsqueeze(-1).repeat(1,1,shape[2],1).contiguous()
139
+ zero = zero.unsqueeze(-1).repeat(1,1,shape[2],1).contiguous()
140
+ # Quantize
141
+ res_data = fdata / scale + zero
142
+ qdata = torch.clamp(res_data, qmin, qmax).to(qtype)
143
+ return qdata.contiguous(), scale, zero
144
+
145
+ def dequantize_cache_torch(qdata, scale, zero):
146
+ data = scale * (qdata - zero)
147
+ return data
148
+
149
+ class FlashSelfAttention(torch.nn.Module):
150
+ def __init__(
151
+ self,
152
+ causal=False,
153
+ softmax_scale=None,
154
+ attention_dropout=0.0,
155
+ ):
156
+ super().__init__()
157
+ assert flash_attn_unpadded_func is not None, (
158
+ "Please install FlashAttention first, " "e.g., with pip install flash-attn"
159
+ )
160
+ assert (
161
+ rearrange is not None
162
+ ), "Please install einops first, e.g., with pip install einops"
163
+ self.causal = causal
164
+ self.softmax_scale = softmax_scale
165
+ self.dropout_p = attention_dropout
166
+
167
+ def unpad_input(self, hidden_states, attention_mask):
168
+ valid_mask = attention_mask.squeeze(1).squeeze(1).eq(0)
169
+ seqlens_in_batch = valid_mask.sum(dim=-1, dtype=torch.int32)
170
+ indices = torch.nonzero(valid_mask.flatten(), as_tuple=False).flatten()
171
+ max_seqlen_in_batch = seqlens_in_batch.max().item()
172
+ cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0))
173
+ hidden_states = hidden_states[indices]
174
+ return hidden_states, indices, cu_seqlens, max_seqlen_in_batch
175
+
176
+ def pad_input(self, hidden_states, indices, batch, seqlen):
177
+ output = torch.zeros(batch * seqlen, *hidden_states.shape[1:], device=hidden_states.device,
178
+ dtype=hidden_states.dtype)
179
+ output[indices] = hidden_states
180
+ return rearrange(output, '(b s) ... -> b s ...', b=batch)
181
+
182
+ def forward(self, q, k, v, attention_mask=None):
183
+ assert all((i.dtype in [torch.float16, torch.bfloat16] for i in (q, k, v)))
184
+ assert all((i.is_cuda for i in (q, k, v)))
185
+ batch_size, seqlen_q = q.shape[0], q.shape[1]
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,
197
+ (batch_size + 1) * seqlen_q,
198
+ step=seqlen_q,
199
+ dtype=torch.int32,
200
+ device=q.device,
201
+ )
202
+
203
+ if batch_size > 1 and attention_mask is not None:
204
+ k, indices_k, cu_seqlens_k, seqlen_k = self.unpad_input(k, attention_mask)
205
+ if q.size(0) == v.size(0):
206
+ q = q[indices_k]
207
+ cu_seqlens_q = cu_seqlens_k
208
+ seqlen_q = seqlen_k
209
+ v = v[indices_k]
210
+ else:
211
+ cu_seqlens_k = torch.arange(
212
+ 0,
213
+ (batch_size + 1) * seqlen_k,
214
+ step=seqlen_k,
215
+ dtype=torch.int32,
216
+ device=q.device,
217
+ )
218
+
219
+ if self.training:
220
+ assert seqlen_k == seqlen_q
221
+ is_causal = self.causal
222
+ dropout_p = self.dropout_p
223
+ else:
224
+ is_causal = seqlen_q == seqlen_k
225
+ dropout_p = 0
226
+
227
+ output = flash_attn_unpadded_func(
228
+ q,
229
+ k,
230
+ v,
231
+ cu_seqlens_q,
232
+ cu_seqlens_k,
233
+ seqlen_q,
234
+ seqlen_k,
235
+ dropout_p,
236
+ softmax_scale=self.softmax_scale,
237
+ causal=is_causal,
238
+ )
239
+ if batch_size > 1 and attention_mask is not None and seqlen_q == seqlen_k:
240
+ output = self.pad_input(output, indices_k, batch_size, seqlen_out)
241
+ else:
242
+ new_shape = (batch_size, output.shape[0] // batch_size) + output.shape[1:]
243
+ output = output.view(new_shape)
244
+ return output
245
+
246
+
247
+ class QWenAttention(nn.Module):
248
+ def __init__(self, config):
249
+ super().__init__()
250
+
251
+ self.register_buffer("masked_bias", torch.tensor(-1e4), persistent=False)
252
+ self.seq_length = config.seq_length
253
+
254
+ self.hidden_size = config.hidden_size
255
+ self.split_size = config.hidden_size
256
+ self.num_heads = config.num_attention_heads
257
+ self.head_dim = self.hidden_size // self.num_heads
258
+
259
+ self.use_flash_attn = config.use_flash_attn
260
+ self.scale_attn_weights = True
261
+
262
+ self.projection_size = config.kv_channels * config.num_attention_heads
263
+
264
+ assert self.projection_size % config.num_attention_heads == 0
265
+ self.hidden_size_per_attention_head = (
266
+ self.projection_size // config.num_attention_heads
267
+ )
268
+
269
+ self.c_attn = nn.Linear(config.hidden_size, 3 * self.projection_size)
270
+
271
+ self.c_proj = nn.Linear(
272
+ config.hidden_size, self.projection_size, bias=not config.no_bias
273
+ )
274
+
275
+ self.is_fp32 = not (config.bf16 or config.fp16)
276
+ if (
277
+ self.use_flash_attn
278
+ and flash_attn_unpadded_func is not None
279
+ and not self.is_fp32
280
+ ):
281
+ self.core_attention_flash = FlashSelfAttention(
282
+ causal=True, attention_dropout=config.attn_dropout_prob
283
+ )
284
+ self.bf16 = config.bf16
285
+
286
+ self.use_dynamic_ntk = config.use_dynamic_ntk
287
+ self.use_logn_attn = config.use_logn_attn
288
+
289
+ logn_list = [
290
+ math.log(i, self.seq_length) if i > self.seq_length else 1
291
+ for i in range(1, 32768)
292
+ ]
293
+ logn_tensor = torch.tensor(logn_list)[None, :, None, None]
294
+ self.register_buffer("logn_tensor", logn_tensor, persistent=False)
295
+
296
+ self.attn_dropout = nn.Dropout(config.attn_dropout_prob)
297
+ self.softmax_in_fp32 = config.softmax_in_fp32 if hasattr(config, 'softmax_in_fp32') else False
298
+ self.use_cache_quantization = config.use_cache_quantization if hasattr(config, 'use_cache_quantization') else False
299
+ self.use_cache_kernel = config.use_cache_kernel if hasattr(config,'use_cache_kernel') else False
300
+ cache_dtype = torch.float
301
+ if self.bf16:
302
+ cache_dtype=torch.bfloat16
303
+ elif config.fp16:
304
+ cache_dtype = torch.float16
305
+ self.cache_qmax = torch.tensor(torch.iinfo(torch.uint8).max, dtype=cache_dtype)
306
+ self.cache_qmin = torch.tensor(torch.iinfo(torch.uint8).min, dtype=cache_dtype)
307
+
308
+ if config.use_cache_quantization and config.use_cache_kernel:
309
+ # pre check if the support files existing
310
+ module_root = pathlib.Path(__file__).parent
311
+ src_files = ("cache_autogptq_cuda_256.cpp", "cache_autogptq_cuda_kernel_256.cu")
312
+ if any(not (module_root/src).is_file() for src in src_files):
313
+ warnings.warn("KV cache kernel source files (.cpp and .cu) not found.")
314
+ self.cache_kernels = None
315
+ else:
316
+ try:
317
+ from .cpp_kernels import cache_autogptq_cuda_256
318
+ self.cache_kernels = cache_autogptq_cuda_256
319
+ except ImportError:
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
327
+ if self.use_cache_kernel and self.cache_kernels is not None:
328
+ shape = query.shape[:-1] + (qk.shape[-2],)
329
+ attn_weights = torch.zeros(shape, dtype=torch.float16, device=device)
330
+ self.cache_kernels.vecquant8matmul_batched_faster_old(
331
+ query.contiguous() if query.dtype == torch.float16 else query.to(torch.float16).contiguous(),
332
+ qk.transpose(-1, -2).contiguous(),
333
+ attn_weights,
334
+ qk_scale.contiguous() if qk_scale.dtype == torch.float16 else qk_scale.to(torch.float16).contiguous(),
335
+ qk_zero.contiguous()if qk_zero.dtype == torch.float16 else qk_zero.to(torch.float16).contiguous())
336
+ # attn_weights = attn_weights.to(query.dtype).contiguous()
337
+ else:
338
+ key = dequantize_cache_torch(qk, qk_scale, qk_zero)
339
+ attn_weights = torch.matmul(query, key.transpose(-1, -2))
340
+ else:
341
+ attn_weights = torch.matmul(query, key.transpose(-1, -2))
342
+
343
+ if self.scale_attn_weights:
344
+ if self.use_cache_quantization:
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
358
+
359
+ if self.softmax_in_fp32:
360
+ attn_weights = nn.functional.softmax(attn_weights.float(), dim=-1)
361
+ else:
362
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1)
363
+
364
+ attn_weights = attn_weights.type(query.dtype)
365
+ attn_weights = self.attn_dropout(attn_weights)
366
+
367
+ if head_mask is not None:
368
+ attn_weights = attn_weights * head_mask
369
+
370
+ if self.use_cache_quantization:
371
+ qv, qv_scale, qv_zero = value
372
+ if self.use_cache_kernel and self.cache_kernels is not None:
373
+ shape = attn_weights.shape[:-1] + (query.shape[-1],)
374
+ attn_output = torch.zeros(shape, dtype=torch.float16, device=device)
375
+ self.cache_kernels.vecquant8matmul_batched_column_compression_faster_old(
376
+ attn_weights.contiguous() if attn_weights.dtype == torch.float16 else attn_weights.to(torch.float16).contiguous(),
377
+ qv.contiguous(), # dtype: int32
378
+ attn_output,
379
+ qv_scale.contiguous() if qv_scale.dtype == torch.float16 else qv_scale.to(torch.float16).contiguous(),
380
+ qv_zero.contiguous() if qv_zero.dtype == torch.float16 else qv_zero.to(torch.float16).contiguous())
381
+ if attn_output.dtype != query.dtype:
382
+ attn_output = attn_output.to(query.dtype)
383
+ attn_weights = attn_weights.to(query.dtype)
384
+ else:
385
+ value = dequantize_cache_torch(qv, qv_scale, qv_zero)
386
+ attn_output = torch.matmul(attn_weights, value)
387
+ else:
388
+ attn_output = torch.matmul(attn_weights, value)
389
+
390
+ attn_output = attn_output.transpose(1, 2)
391
+
392
+ return attn_output, attn_weights
393
+
394
+ def _split_heads(self, tensor, num_heads, attn_head_size):
395
+ new_shape = tensor.size()[:-1] + (num_heads, attn_head_size)
396
+ tensor = tensor.view(new_shape)
397
+ return tensor
398
+
399
+ def _merge_heads(self, tensor, num_heads, attn_head_size):
400
+ tensor = tensor.contiguous()
401
+ new_shape = tensor.size()[:-2] + (num_heads * attn_head_size,)
402
+ return tensor.view(new_shape)
403
+
404
+ def forward(
405
+ self,
406
+ hidden_states: Optional[Tuple[torch.FloatTensor]],
407
+ rotary_pos_emb_list: Optional[List[List[torch.Tensor]]] = None,
408
+ layer_past: Optional[Tuple[torch.Tensor]] = None,
409
+ attention_mask: Optional[torch.FloatTensor] = None,
410
+ head_mask: Optional[torch.FloatTensor] = None,
411
+ encoder_hidden_states: Optional[torch.Tensor] = None,
412
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
413
+ output_attentions: Optional[bool] = False,
414
+ use_cache: Optional[bool] = False,
415
+ ):
416
+ mixed_x_layer = self.c_attn(hidden_states)
417
+
418
+ query, key, value = mixed_x_layer.split(self.split_size, dim=2)
419
+
420
+ query = self._split_heads(query, self.num_heads, self.head_dim)
421
+ key = self._split_heads(key, self.num_heads, self.head_dim)
422
+ value = self._split_heads(value, self.num_heads, self.head_dim)
423
+
424
+ if rotary_pos_emb_list is not None:
425
+ cur_len = query.shape[1]
426
+ if len(rotary_pos_emb_list) == 1:
427
+ rotary_pos_emb = rotary_pos_emb_list[0]
428
+ rotary_pos_emb = [i[:, -cur_len:, :, :] for i in rotary_pos_emb]
429
+ rotary_pos_emb = (rotary_pos_emb,) * 2
430
+ q_pos_emb, k_pos_emb = rotary_pos_emb
431
+ # Slice the pos emb for current inference
432
+ query = apply_rotary_pos_emb(query, q_pos_emb)
433
+ key = apply_rotary_pos_emb(key, k_pos_emb)
434
+ else:
435
+ query_list = []
436
+ key_list = []
437
+ for i, rotary_pos_emb in enumerate(rotary_pos_emb_list):
438
+ rotary_pos_emb = [i[:, -cur_len:, :, :] for i in rotary_pos_emb]
439
+ rotary_pos_emb = (rotary_pos_emb,) * 2
440
+ q_pos_emb, k_pos_emb = rotary_pos_emb
441
+ # Slice the pos emb for current inference
442
+ query_list += [apply_rotary_pos_emb(query[i:i+1, :, :], q_pos_emb)]
443
+ key_list += [apply_rotary_pos_emb(key[i:i+1, :, :], k_pos_emb)]
444
+ query = torch.cat(query_list, dim=0)
445
+ key = torch.cat(key_list, dim=0)
446
+
447
+ if self.use_cache_quantization:
448
+ key = quantize_cache_v(key.permute(0, 2, 1, 3),
449
+ bits=8,
450
+ qmin=self.cache_qmin,
451
+ qmax=self.cache_qmax)
452
+ value = quantize_cache_v(value.permute(0, 2, 1, 3),
453
+ bits=8,
454
+ qmin=self.cache_qmin,
455
+ qmax=self.cache_qmax)
456
+
457
+
458
+ if layer_past is not None:
459
+ past_key, past_value = layer_past[0], layer_past[1]
460
+ if self.use_cache_quantization:
461
+ # use_cache_quantization:
462
+ # present=((q_key,key_scale,key_zero_point),
463
+ # (q_value,value_scale,value_zero_point))
464
+ key = (torch.cat((past_key[0], key[0]), dim=2),
465
+ torch.cat((past_key[1], key[1]), dim=2),
466
+ torch.cat((past_key[2], key[2]), dim=2))
467
+ value = (torch.cat((past_value[0], value[0]), dim=2),
468
+ torch.cat((past_value[1], value[1]), dim=2),
469
+ torch.cat((past_value[2], value[2]), dim=2))
470
+ else:
471
+ # not use_cache_quantization:
472
+ # present=(key,value)
473
+ key = torch.cat((past_key, key), dim=1)
474
+ value = torch.cat((past_value, value), dim=1)
475
+
476
+ if use_cache:
477
+ present = (key, value)
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)
486
+ else:
487
+ seq_start = key.size(1) - query.size(1)
488
+ seq_end = key.size(1)
489
+ logn_tensor = self.logn_tensor[:, seq_start:seq_end, :, :].type_as(query)
490
+ query = query * logn_tensor.expand_as(query)
491
+
492
+ if (
493
+ self.use_flash_attn
494
+ and flash_attn_unpadded_func is not None
495
+ and not self.is_fp32
496
+ and query.is_cuda
497
+ ):
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
517
+ and not query.is_cuda
518
+ ):
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(-1, -1, query.size(2), -1)
524
+ if causal_mask is not None:
525
+ attention_mask = attention_mask.masked_fill(~causal_mask, torch.finfo(query.dtype).min)
526
+ else:
527
+ attention_mask = causal_mask
528
+ attn_output = F.scaled_dot_product_attention(
529
+ query, key, value, attn_mask=attention_mask
530
+ ).transpose(1, 2)
531
+ attn_weight = None
532
+ else:
533
+ attn_output, attn_weight = self._attn(
534
+ query, key, value, causal_mask, attention_mask, head_mask
535
+ )
536
+ context_layer = self._merge_heads(
537
+ attn_output, self.num_heads, self.head_dim
538
+ )
539
+
540
+ attn_output = self.c_proj(context_layer)
541
+
542
+ outputs = (attn_output, present)
543
+ if output_attentions:
544
+ if (
545
+ self.use_flash_attn
546
+ and flash_attn_unpadded_func is not None
547
+ and not self.is_fp32
548
+ ):
549
+ raise ValueError("Cannot output attentions while using flash-attn")
550
+ elif not self.use_cache_quantization and SUPPORT_TORCH2:
551
+ raise ValueError("Cannot output attentions while using scaled_dot_product_attention")
552
+ else:
553
+ outputs += (attn_weight,)
554
+
555
+ return outputs
556
+
557
+
558
+ class QWenMLP(nn.Module):
559
+ def __init__(self, config):
560
+ super().__init__()
561
+ self.w1 = nn.Linear(
562
+ config.hidden_size, config.intermediate_size // 2, bias=not config.no_bias
563
+ )
564
+ self.w2 = nn.Linear(
565
+ config.hidden_size, config.intermediate_size // 2, bias=not config.no_bias
566
+ )
567
+ ff_dim_in = config.intermediate_size // 2
568
+ self.c_proj = nn.Linear(ff_dim_in, config.hidden_size, bias=not config.no_bias)
569
+
570
+ def forward(self, hidden_states):
571
+ a1 = self.w1(hidden_states)
572
+ a2 = self.w2(hidden_states)
573
+ intermediate_parallel = a1 * F.silu(a2)
574
+ output = self.c_proj(intermediate_parallel)
575
+ return output
576
+
577
+
578
+ class QWenBlock(nn.Module):
579
+ def __init__(self, config):
580
+ super().__init__()
581
+ hidden_size = config.hidden_size
582
+ self.bf16 = config.bf16
583
+
584
+ self.ln_1 = RMSNorm(
585
+ hidden_size,
586
+ eps=config.layer_norm_epsilon,
587
+ )
588
+ self.attn = QWenAttention(config)
589
+ self.ln_2 = RMSNorm(
590
+ hidden_size,
591
+ eps=config.layer_norm_epsilon,
592
+ )
593
+
594
+ self.mlp = QWenMLP(config)
595
+
596
+ def forward(
597
+ self,
598
+ hidden_states: Optional[Tuple[torch.FloatTensor]],
599
+ rotary_pos_emb_list: Optional[List[List[torch.Tensor]]] = None,
600
+ layer_past: Optional[Tuple[torch.Tensor]] = None,
601
+ attention_mask: Optional[torch.FloatTensor] = None,
602
+ head_mask: Optional[torch.FloatTensor] = None,
603
+ encoder_hidden_states: Optional[torch.Tensor] = None,
604
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
605
+ use_cache: Optional[bool] = False,
606
+ output_attentions: Optional[bool] = False,
607
+ ):
608
+ layernorm_output = self.ln_1(hidden_states)
609
+
610
+ attn_outputs = self.attn(
611
+ layernorm_output,
612
+ rotary_pos_emb_list,
613
+ layer_past=layer_past,
614
+ attention_mask=attention_mask,
615
+ head_mask=head_mask,
616
+ use_cache=use_cache,
617
+ output_attentions=output_attentions,
618
+ )
619
+ attn_output = attn_outputs[0]
620
+
621
+ outputs = attn_outputs[1:]
622
+
623
+ residual = hidden_states
624
+ layernorm_input = attn_output + residual
625
+
626
+ layernorm_output = self.ln_2(layernorm_input)
627
+
628
+ residual = layernorm_input
629
+ mlp_output = self.mlp(layernorm_output)
630
+ hidden_states = residual + mlp_output
631
+
632
+ if use_cache:
633
+ outputs = (hidden_states,) + outputs
634
+ else:
635
+ outputs = (hidden_states,) + outputs[1:]
636
+
637
+ return outputs
638
+
639
+
640
+ class QWenPreTrainedModel(PreTrainedModel):
641
+ config_class = QWenConfig
642
+ base_model_prefix = "transformer"
643
+ is_parallelizable = False
644
+ supports_gradient_checkpointing = True
645
+ _no_split_modules = ["QWenBlock"]
646
+ _skip_keys_device_placement = "past_key_values"
647
+
648
+ def __init__(self, *inputs, **kwargs):
649
+ super().__init__(*inputs, **kwargs)
650
+
651
+ def _init_weights(self, module):
652
+ """Initialize the weights."""
653
+ if isinstance(module, nn.Linear):
654
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
655
+ if module.bias is not None:
656
+ module.bias.data.zero_()
657
+ elif isinstance(module, nn.Embedding):
658
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
659
+ if module.padding_idx is not None:
660
+ module.weight.data[module.padding_idx].zero_()
661
+ elif isinstance(module, RMSNorm):
662
+ module.weight.data.fill_(1.0)
663
+
664
+ for name, p in module.named_parameters():
665
+ if name == "c_proj.weight":
666
+ p.data.normal_(
667
+ mean=0.0,
668
+ std=(
669
+ self.config.initializer_range
670
+ / math.sqrt(2 * self.config.num_hidden_layers)
671
+ ),
672
+ )
673
+
674
+ def _set_gradient_checkpointing(self, module, value=False):
675
+ if isinstance(module, QWenModel):
676
+ module.gradient_checkpointing = value
677
+
678
+
679
+ class QWenModel(QWenPreTrainedModel):
680
+ _keys_to_ignore_on_load_missing = ["attn.masked_bias"]
681
+
682
+ def __init__(self, config):
683
+ super().__init__(config)
684
+ self.vocab_size = config.vocab_size
685
+ self.num_hidden_layers = config.num_hidden_layers
686
+ self.embed_dim = config.hidden_size
687
+ self.use_cache_quantization = self.config.use_cache_quantization if hasattr(self.config, 'use_cache_quantization') else False
688
+
689
+ self.gradient_checkpointing = False
690
+ self.use_dynamic_ntk = config.use_dynamic_ntk
691
+ self.seq_length = config.seq_length
692
+
693
+ self.wte = nn.Embedding(self.vocab_size, self.embed_dim)
694
+
695
+ self.drop = nn.Dropout(config.emb_dropout_prob)
696
+
697
+ if config.rotary_pct == 1.0:
698
+ self.rotary_ndims = None
699
+ else:
700
+ assert config.rotary_pct < 1
701
+ self.rotary_ndims = int(
702
+ config.kv_channels * config.rotary_pct
703
+ )
704
+ dim = (
705
+ self.rotary_ndims
706
+ if self.rotary_ndims is not None
707
+ else config.kv_channels
708
+ )
709
+ self.rotary_emb = RotaryEmbedding(dim, base=config.rotary_emb_base)
710
+
711
+ self.use_flash_attn = config.use_flash_attn
712
+ self.is_fp32 = not (config.bf16 or config.fp16)
713
+
714
+ self.h = nn.ModuleList(
715
+ [
716
+ QWenBlock(
717
+ config
718
+ )
719
+ for i in range(config.num_hidden_layers)
720
+ ]
721
+ )
722
+ self.ln_f = RMSNorm(
723
+ self.embed_dim,
724
+ eps=config.layer_norm_epsilon,
725
+ )
726
+
727
+ self.post_init()
728
+
729
+ def get_input_embeddings(self):
730
+ return self.wte
731
+
732
+ def set_input_embeddings(self, new_embeddings):
733
+ self.wte = new_embeddings
734
+
735
+ def get_ntk_alpha(self, true_seq_len):
736
+ context_value = math.log(true_seq_len / self.seq_length, 2) + 1
737
+ ntk_alpha = 2 ** math.ceil(context_value) - 1
738
+ ntk_alpha = max(ntk_alpha, 1)
739
+ return ntk_alpha
740
+
741
+ def forward(
742
+ self,
743
+ input_ids: Optional[torch.LongTensor] = None,
744
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
745
+ attention_mask: Optional[torch.FloatTensor] = None,
746
+ token_type_ids: Optional[torch.LongTensor] = None,
747
+ position_ids: Optional[torch.LongTensor] = None,
748
+ head_mask: Optional[torch.FloatTensor] = None,
749
+ inputs_embeds: Optional[torch.FloatTensor] = None,
750
+ encoder_hidden_states: Optional[torch.Tensor] = None,
751
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
752
+ use_cache: Optional[bool] = None,
753
+ output_attentions: Optional[bool] = None,
754
+ output_hidden_states: Optional[bool] = None,
755
+ return_dict: Optional[bool] = None,
756
+ ):
757
+ output_attentions = (
758
+ output_attentions
759
+ if output_attentions is not None
760
+ else self.config.output_attentions
761
+ )
762
+ output_hidden_states = (
763
+ output_hidden_states
764
+ if output_hidden_states is not None
765
+ else self.config.output_hidden_states
766
+ )
767
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
768
+ return_dict = (
769
+ return_dict if return_dict is not None else self.config.use_return_dict
770
+ )
771
+
772
+ if input_ids is not None and inputs_embeds is not None:
773
+ raise ValueError(
774
+ "You cannot specify both input_ids and inputs_embeds at the same time"
775
+ )
776
+ elif input_ids is not None:
777
+ input_shape = input_ids.size()
778
+ input_ids = input_ids.view(-1, input_shape[-1])
779
+ batch_size = input_ids.shape[0]
780
+ elif inputs_embeds is not None:
781
+ input_shape = inputs_embeds.size()[:-1]
782
+ batch_size = inputs_embeds.shape[0]
783
+ else:
784
+ raise ValueError("You have to specify either input_ids or inputs_embeds")
785
+
786
+ device = input_ids.device if input_ids is not None else inputs_embeds.device
787
+
788
+ if token_type_ids is not None:
789
+ token_type_ids = token_type_ids.view(-1, input_shape[-1])
790
+ if position_ids is not None:
791
+ position_ids = position_ids.view(-1, input_shape[-1])
792
+
793
+ if past_key_values is None:
794
+ past_length = 0
795
+ past_key_values = tuple([None] * len(self.h))
796
+ else:
797
+ if self.use_cache_quantization:
798
+ past_length = past_key_values[0][0][0].size(2)
799
+ else:
800
+ past_length = past_key_values[0][0].size(-2)
801
+ if position_ids is None:
802
+ position_ids = torch.arange(
803
+ past_length,
804
+ input_shape[-1] + past_length,
805
+ dtype=torch.long,
806
+ device=device,
807
+ )
808
+ position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
809
+
810
+ if attention_mask is not None:
811
+ if batch_size <= 0:
812
+ raise ValueError("batch_size has to be defined and > 0")
813
+ attention_mask = attention_mask.view(batch_size, -1)
814
+ attention_mask = attention_mask[:, None, None, :]
815
+ attention_mask = attention_mask.to(dtype=self.dtype)
816
+ attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min
817
+
818
+ encoder_attention_mask = None
819
+ head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
820
+
821
+ if inputs_embeds is None:
822
+ inputs_embeds = self.wte(input_ids)
823
+ hidden_states = inputs_embeds
824
+
825
+ kv_seq_len = hidden_states.size()[1]
826
+ if past_key_values[0] is not None:
827
+ # past key values[0][0] shape: bs * seq_len * head_num * dim
828
+ if self.use_cache_quantization:
829
+ kv_seq_len += past_key_values[0][0][0].shape[2]
830
+ else:
831
+ kv_seq_len += past_key_values[0][0].shape[1]
832
+
833
+ if self.training or not self.use_dynamic_ntk:
834
+ ntk_alpha_list = [1.0]
835
+ elif kv_seq_len != hidden_states.size()[1]:
836
+ ntk_alpha_list = self.rotary_emb._ntk_alpha_cached_list
837
+ else:
838
+ ntk_alpha_list = []
839
+ if attention_mask is not None and kv_seq_len > self.seq_length:
840
+ true_seq_lens = attention_mask.squeeze(1).squeeze(1).eq(0).sum(dim=-1, dtype=torch.int32)
841
+ for i in range(hidden_states.size()[0]):
842
+ true_seq_len = true_seq_lens[i].item()
843
+ ntk_alpha = self.get_ntk_alpha(true_seq_len)
844
+ ntk_alpha_list.append(ntk_alpha)
845
+ else:
846
+ ntk_alpha = self.get_ntk_alpha(kv_seq_len)
847
+ ntk_alpha_list.append(ntk_alpha)
848
+ self.rotary_emb._ntk_alpha_cached_list = ntk_alpha_list
849
+ rotary_pos_emb_list = [
850
+ self.rotary_emb(kv_seq_len, ntk_alpha=ntk_alpha) for ntk_alpha in ntk_alpha_list
851
+ ]
852
+
853
+ hidden_states = self.drop(hidden_states)
854
+ output_shape = input_shape + (hidden_states.size(-1),)
855
+
856
+ if self.gradient_checkpointing and self.training:
857
+ if use_cache:
858
+ logger.warning_once(
859
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
860
+ )
861
+ use_cache = False
862
+
863
+ presents = () if use_cache else None
864
+ all_self_attentions = () if output_attentions else None
865
+ all_hidden_states = () if output_hidden_states else None
866
+ for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
867
+
868
+ if output_hidden_states:
869
+ all_hidden_states = all_hidden_states + (hidden_states,)
870
+
871
+ if self.gradient_checkpointing and self.training:
872
+
873
+ def create_custom_forward(module):
874
+ def custom_forward(*inputs):
875
+ # None for past_key_value
876
+ return module(*inputs, use_cache, output_attentions)
877
+
878
+ return custom_forward
879
+
880
+ outputs = torch.utils.checkpoint.checkpoint(
881
+ create_custom_forward(block),
882
+ hidden_states,
883
+ rotary_pos_emb_list,
884
+ None,
885
+ attention_mask,
886
+ head_mask[i],
887
+ encoder_hidden_states,
888
+ encoder_attention_mask,
889
+ )
890
+ else:
891
+ outputs = block(
892
+ hidden_states,
893
+ layer_past=layer_past,
894
+ rotary_pos_emb_list=rotary_pos_emb_list,
895
+ attention_mask=attention_mask,
896
+ head_mask=head_mask[i],
897
+ encoder_hidden_states=encoder_hidden_states,
898
+ encoder_attention_mask=encoder_attention_mask,
899
+ use_cache=use_cache,
900
+ output_attentions=output_attentions,
901
+ )
902
+
903
+ hidden_states = outputs[0]
904
+ if use_cache is True:
905
+ presents = presents + (outputs[1],)
906
+
907
+ if output_attentions:
908
+ all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
909
+
910
+ hidden_states = self.ln_f(hidden_states)
911
+ hidden_states = hidden_states.view(output_shape)
912
+ # Add last hidden state
913
+ if output_hidden_states:
914
+ all_hidden_states = all_hidden_states + (hidden_states,)
915
+
916
+ if not return_dict:
917
+ return tuple(
918
+ v for v in [hidden_states, presents, all_hidden_states] if v is not None
919
+ )
920
+
921
+ return BaseModelOutputWithPast(
922
+ last_hidden_state=hidden_states,
923
+ past_key_values=presents,
924
+ hidden_states=all_hidden_states,
925
+ attentions=all_self_attentions,
926
+ )
927
+
928
+
929
+ class QWenLMHeadModel(QWenPreTrainedModel):
930
+ _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.rotary_emb\.inv_freq"]
931
+ _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.masked_bias"]
932
+
933
+ def __init__(self, config):
934
+ super().__init__(config)
935
+ assert (
936
+ config.bf16 + config.fp16 + config.fp32 <= 1
937
+ ), "Only one of \"bf16\", \"fp16\", \"fp32\" can be true"
938
+
939
+ autoset_precision = config.bf16 + config.fp16 + config.fp32 == 0
940
+
941
+ if autoset_precision:
942
+ if SUPPORT_BF16:
943
+ logger.warn(
944
+ "The model is automatically converting to bf16 for faster inference. "
945
+ "If you want to disable the automatic precision, please manually add bf16/fp16/fp32=True to \"AutoModelForCausalLM.from_pretrained\"."
946
+ )
947
+ config.bf16 = True
948
+ elif SUPPORT_FP16:
949
+ logger.warn(
950
+ "The model is automatically converting to fp16 for faster inference. "
951
+ "If you want to disable the automatic precision, please manually add bf16/fp16/fp32=True to \"AutoModelForCausalLM.from_pretrained\"."
952
+ )
953
+ config.fp16 = True
954
+ else:
955
+ config.fp32 = True
956
+
957
+ if config.bf16 and SUPPORT_CUDA and not SUPPORT_BF16:
958
+ logger.warn("Your device does NOT seem to support bf16, you can switch to fp16 or fp32 by by passing fp16/fp32=True in \"AutoModelForCausalLM.from_pretrained\".")
959
+ if config.fp16 and SUPPORT_CUDA and not SUPPORT_FP16:
960
+ logger.warn("Your device does NOT support faster inference with fp16, please switch to fp32 which is likely to be faster")
961
+ if config.fp32:
962
+ if SUPPORT_BF16:
963
+ logger.warn("Your device support faster inference by passing bf16=True in \"AutoModelForCausalLM.from_pretrained\".")
964
+ elif SUPPORT_FP16:
965
+ logger.warn("Your device support faster inference by passing fp16=True in \"AutoModelForCausalLM.from_pretrained\".")
966
+
967
+ if config.use_flash_attn == "auto":
968
+ if config.bf16 or config.fp16:
969
+ logger.warn("Try importing flash-attention for faster inference...")
970
+ config.use_flash_attn = True
971
+ else:
972
+ config.use_flash_attn = False
973
+ if config.use_flash_attn and config.fp32:
974
+ logger.warn("Flash attention will be disabled because it does NOT support fp32.")
975
+
976
+ if config.use_flash_attn:
977
+ _import_flash_attn()
978
+
979
+ self.transformer = QWenModel(config)
980
+ self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
981
+
982
+ if config.bf16:
983
+ self.transformer.bfloat16()
984
+ self.lm_head.bfloat16()
985
+ if config.fp16:
986
+ self.transformer.half()
987
+ self.lm_head.half()
988
+ self.post_init()
989
+
990
+ def get_output_embeddings(self):
991
+ return self.lm_head
992
+
993
+ def set_output_embeddings(self, new_embeddings):
994
+ self.lm_head = new_embeddings
995
+
996
+ def prepare_inputs_for_generation(
997
+ self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs
998
+ ):
999
+ if past_key_values:
1000
+ input_ids = input_ids[:, -1].unsqueeze(-1)
1001
+
1002
+ if input_ids.size(0) == 1:
1003
+ attention_mask = None
1004
+ else:
1005
+ attention_mask = kwargs.get("attention_mask", None)
1006
+
1007
+ if inputs_embeds is not None and past_key_values is None:
1008
+ model_inputs = {"inputs_embeds": inputs_embeds}
1009
+ else:
1010
+ model_inputs = {"input_ids": input_ids}
1011
+
1012
+ model_inputs.update(
1013
+ {
1014
+ "past_key_values": past_key_values,
1015
+ "use_cache": kwargs.get("use_cache"),
1016
+ "attention_mask": attention_mask,
1017
+ }
1018
+ )
1019
+ return model_inputs
1020
+
1021
+ def forward(
1022
+ self,
1023
+ input_ids: Optional[torch.LongTensor] = None,
1024
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
1025
+ attention_mask: Optional[torch.FloatTensor] = None,
1026
+ token_type_ids: Optional[torch.LongTensor] = None,
1027
+ position_ids: Optional[torch.LongTensor] = None,
1028
+ head_mask: Optional[torch.FloatTensor] = None,
1029
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1030
+ encoder_hidden_states: Optional[torch.Tensor] = None,
1031
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
1032
+ labels: Optional[torch.LongTensor] = None,
1033
+ use_cache: Optional[bool] = None,
1034
+ output_attentions: Optional[bool] = None,
1035
+ output_hidden_states: Optional[bool] = None,
1036
+ return_dict: Optional[bool] = None,
1037
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
1038
+
1039
+ return_dict = (
1040
+ return_dict if return_dict is not None else self.config.use_return_dict
1041
+ )
1042
+
1043
+ transformer_outputs = self.transformer(
1044
+ input_ids,
1045
+ past_key_values=past_key_values,
1046
+ attention_mask=attention_mask,
1047
+ token_type_ids=token_type_ids,
1048
+ position_ids=position_ids,
1049
+ head_mask=head_mask,
1050
+ inputs_embeds=inputs_embeds,
1051
+ encoder_hidden_states=encoder_hidden_states,
1052
+ encoder_attention_mask=encoder_attention_mask,
1053
+ use_cache=use_cache,
1054
+ output_attentions=output_attentions,
1055
+ output_hidden_states=output_hidden_states,
1056
+ return_dict=return_dict,
1057
+ )
1058
+ hidden_states = transformer_outputs[0]
1059
+
1060
+ lm_logits = self.lm_head(hidden_states)
1061
+
1062
+ loss = None
1063
+ if labels is not None:
1064
+ labels = labels.to(lm_logits.device)
1065
+ shift_logits = lm_logits[..., :-1, :].contiguous()
1066
+ shift_labels = labels[..., 1:].contiguous()
1067
+ loss_fct = CrossEntropyLoss()
1068
+ loss = loss_fct(
1069
+ shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)
1070
+ )
1071
+
1072
+ if not return_dict:
1073
+ output = (lm_logits,) + transformer_outputs[1:]
1074
+ return ((loss,) + output) if loss is not None else output
1075
+
1076
+ return CausalLMOutputWithPast(
1077
+ loss=loss,
1078
+ logits=lm_logits,
1079
+ past_key_values=transformer_outputs.past_key_values,
1080
+ hidden_states=transformer_outputs.hidden_states,
1081
+ attentions=transformer_outputs.attentions,
1082
+ )
1083
+
1084
+ @staticmethod
1085
+ def _reorder_cache(
1086
+ past_key_values: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
1087
+ ) -> Tuple[Tuple[torch.Tensor]]:
1088
+
1089
+ return tuple(
1090
+ tuple(
1091
+ past_state.index_select(0, beam_idx.to(past_state.device))
1092
+ for past_state in layer_past
1093
+ )
1094
+ for layer_past in past_key_values
1095
+ )
1096
+
1097
+ def chat(
1098
+ self,
1099
+ tokenizer: PreTrainedTokenizer,
1100
+ query: str,
1101
+ history: Optional[HistoryType],
1102
+ system: str = "You are a helpful assistant.",
1103
+ stream: Optional[bool] = _SENTINEL,
1104
+ stop_words_ids: Optional[List[List[int]]] = None,
1105
+ generation_config: Optional[GenerationConfig] = None,
1106
+ **kwargs,
1107
+ ) -> Tuple[str, HistoryType]:
1108
+ generation_config = generation_config if generation_config is not None else self.generation_config
1109
+
1110
+ assert stream is _SENTINEL, _ERROR_STREAM_IN_CHAT
1111
+ assert generation_config.chat_format == 'chatml', _ERROR_BAD_CHAT_FORMAT
1112
+ if history is None:
1113
+ history = []
1114
+ else:
1115
+ # make a copy of the user's input such that is is left untouched
1116
+ history = copy.deepcopy(history)
1117
+
1118
+ if stop_words_ids is None:
1119
+ stop_words_ids = []
1120
+
1121
+ max_window_size = kwargs.get('max_window_size', None)
1122
+ if max_window_size is None:
1123
+ max_window_size = generation_config.max_window_size
1124
+ raw_text, context_tokens = make_context(
1125
+ tokenizer,
1126
+ query,
1127
+ history=history,
1128
+ system=system,
1129
+ max_window_size=max_window_size,
1130
+ chat_format=generation_config.chat_format,
1131
+ )
1132
+
1133
+ stop_words_ids.extend(get_stop_words_ids(
1134
+ generation_config.chat_format, tokenizer
1135
+ ))
1136
+ input_ids = torch.tensor([context_tokens]).to(self.device)
1137
+ outputs = self.generate(
1138
+ input_ids,
1139
+ stop_words_ids=stop_words_ids,
1140
+ return_dict_in_generate=False,
1141
+ generation_config=generation_config,
1142
+ **kwargs,
1143
+ )
1144
+
1145
+ response = decode_tokens(
1146
+ outputs[0],
1147
+ tokenizer,
1148
+ raw_text_len=len(raw_text),
1149
+ context_length=len(context_tokens),
1150
+ chat_format=generation_config.chat_format,
1151
+ verbose=False,
1152
+ errors='replace'
1153
+ )
1154
+
1155
+ # as history is a copy of the user inputs,
1156
+ # we can always return the new turn to the user.
1157
+ # separating input history and output history also enables the user
1158
+ # to implement more complex history management
1159
+ history.append((query, response))
1160
+
1161
+ return response, history
1162
+
1163
+ def chat_stream(
1164
+ self,
1165
+ tokenizer: PreTrainedTokenizer,
1166
+ query: str,
1167
+ history: Optional[HistoryType],
1168
+ system: str = "You are a helpful assistant.",
1169
+ stop_words_ids: Optional[List[List[int]]] = None,
1170
+ logits_processor: Optional[LogitsProcessorList] = None,
1171
+ generation_config: Optional[GenerationConfig] = None,
1172
+ **kwargs,
1173
+ ) -> Generator[str, Any, None]:
1174
+ generation_config = generation_config if generation_config is not None else self.generation_config
1175
+ assert generation_config.chat_format == 'chatml', _ERROR_BAD_CHAT_FORMAT
1176
+ if history is None:
1177
+ history = []
1178
+ if stop_words_ids is None:
1179
+ stop_words_ids = []
1180
+
1181
+ max_window_size = kwargs.get('max_window_size', None)
1182
+ if max_window_size is None:
1183
+ max_window_size = generation_config.max_window_size
1184
+ raw_text, context_tokens = make_context(
1185
+ tokenizer,
1186
+ query,
1187
+ history=history,
1188
+ system=system,
1189
+ max_window_size=max_window_size,
1190
+ chat_format=generation_config.chat_format,
1191
+ )
1192
+
1193
+ stop_words_ids.extend(get_stop_words_ids(
1194
+ generation_config.chat_format, tokenizer
1195
+ ))
1196
+ if stop_words_ids is not None:
1197
+ stop_words_logits_processor = StopWordsLogitsProcessor(
1198
+ stop_words_ids=stop_words_ids,
1199
+ eos_token_id=generation_config.eos_token_id,
1200
+ )
1201
+ if logits_processor is None:
1202
+ logits_processor = LogitsProcessorList([stop_words_logits_processor])
1203
+ else:
1204
+ logits_processor.append(stop_words_logits_processor)
1205
+ input_ids = torch.tensor([context_tokens]).to(self.device)
1206
+
1207
+ from transformers_stream_generator.main import NewGenerationMixin, StreamGenerationConfig
1208
+ self.__class__.generate_stream = NewGenerationMixin.generate
1209
+ self.__class__.sample_stream = NewGenerationMixin.sample_stream
1210
+ stream_config = StreamGenerationConfig(**generation_config.to_dict(), do_stream=True)
1211
+
1212
+ def stream_generator():
1213
+ outputs = []
1214
+ for token in self.generate_stream(
1215
+ input_ids,
1216
+ return_dict_in_generate=False,
1217
+ generation_config=stream_config,
1218
+ logits_processor=logits_processor,
1219
+ seed=-1,
1220
+ **kwargs):
1221
+ outputs.append(token.item())
1222
+ yield tokenizer.decode(outputs, skip_special_tokens=True, errors='ignore')
1223
+
1224
+ return stream_generator()
1225
+
1226
+ def generate(
1227
+ self,
1228
+ inputs: Optional[torch.Tensor] = None,
1229
+ generation_config: Optional[GenerationConfig] = None,
1230
+ logits_processor: Optional[LogitsProcessorList] = None,
1231
+ stopping_criteria: Optional[StoppingCriteriaList] = None,
1232
+ prefix_allowed_tokens_fn: Optional[
1233
+ Callable[[int, torch.Tensor], List[int]]
1234
+ ] = None,
1235
+ synced_gpus: Optional[bool] = None,
1236
+ assistant_model: Optional["PreTrainedModel"] = None,
1237
+ streamer: Optional["BaseStreamer"] = None,
1238
+ **kwargs,
1239
+ ) -> Union[GenerateOutput, torch.LongTensor]:
1240
+ generation_config = generation_config if generation_config is not None else self.generation_config
1241
+
1242
+ # Process stop_words_ids.
1243
+ stop_words_ids = kwargs.pop("stop_words_ids", None)
1244
+ if stop_words_ids is None and generation_config is not None:
1245
+ stop_words_ids = getattr(generation_config, "stop_words_ids", None)
1246
+ if stop_words_ids is None:
1247
+ stop_words_ids = getattr(generation_config, "stop_words_ids", None)
1248
+
1249
+ if stop_words_ids is not None:
1250
+ stop_words_logits_processor = StopWordsLogitsProcessor(
1251
+ stop_words_ids=stop_words_ids,
1252
+ eos_token_id=generation_config.eos_token_id,
1253
+ )
1254
+ if logits_processor is None:
1255
+ logits_processor = LogitsProcessorList([stop_words_logits_processor])
1256
+ else:
1257
+ logits_processor.append(stop_words_logits_processor)
1258
+
1259
+ return super().generate(
1260
+ inputs,
1261
+ generation_config=generation_config,
1262
+ logits_processor=logits_processor,
1263
+ stopping_criteria=stopping_criteria,
1264
+ prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
1265
+ synced_gpus=synced_gpus,
1266
+ assistant_model=assistant_model,
1267
+ streamer=streamer,
1268
+ **kwargs,
1269
+ )
1270
+
1271
+
1272
+ class RotaryEmbedding(torch.nn.Module):
1273
+ def __init__(self, dim, base=10000):
1274
+ super().__init__()
1275
+ self.dim = dim
1276
+ self.base = base
1277
+ inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
1278
+ self.register_buffer("inv_freq", inv_freq, persistent=False)
1279
+ if importlib.util.find_spec("einops") is None:
1280
+ raise RuntimeError("einops is required for Rotary Embedding")
1281
+
1282
+ self._rotary_pos_emb_cache = None
1283
+ self._seq_len_cached = 0
1284
+ self._ntk_alpha_cached = 1.0
1285
+ self._ntk_alpha_cached_list = [1.0]
1286
+
1287
+ def update_rotary_pos_emb_cache(self, seqlen, ntk_alpha=1.0):
1288
+ if seqlen > self._seq_len_cached or ntk_alpha != self._ntk_alpha_cached:
1289
+ base = self.base * ntk_alpha ** (self.dim / (self.dim - 2))
1290
+ self.inv_freq = 1.0 / (
1291
+ base
1292
+ ** (
1293
+ torch.arange(0, self.dim, 2, device=self.inv_freq.device).float()
1294
+ / self.dim
1295
+ )
1296
+ )
1297
+ self._seq_len_cached = max(2 * seqlen, 16)
1298
+ self._ntk_alpha_cached = ntk_alpha
1299
+ seq = torch.arange(self._seq_len_cached, device=self.inv_freq.device)
1300
+ freqs = torch.outer(seq.type_as(self.inv_freq), self.inv_freq)
1301
+
1302
+ emb = torch.cat((freqs, freqs), dim=-1)
1303
+ from einops import rearrange
1304
+
1305
+ emb = rearrange(emb, "n d -> 1 n 1 d")
1306
+
1307
+ cos, sin = emb.cos(), emb.sin()
1308
+ self._rotary_pos_emb_cache = [cos, sin]
1309
+
1310
+ def forward(self, max_seq_len, ntk_alpha=1.0):
1311
+ self.update_rotary_pos_emb_cache(max_seq_len, ntk_alpha)
1312
+ cos, sin = self._rotary_pos_emb_cache
1313
+ return [cos[:, :max_seq_len], sin[:, :max_seq_len]]
1314
+
1315
+
1316
+ def _rotate_half(x):
1317
+ from einops import rearrange
1318
+
1319
+ x = rearrange(x, "... (j d) -> ... j d", j=2)
1320
+ x1, x2 = x.unbind(dim=-2)
1321
+ return torch.cat((-x2, x1), dim=-1)
1322
+
1323
+
1324
+ def apply_rotary_pos_emb(t, freqs):
1325
+ """ Apply rotary embedding to the first rotary_dim of the iput
1326
+
1327
+ Arguments:
1328
+ t (tensor(batch_size, seq_len, n_head, head_dim)):
1329
+ the input embedding/hidden states
1330
+ freqs (list[tensor(1, seq_len, 1, rotary_dim), tensor(1, seq_len, 1, rotary_dim)]):
1331
+ the cached cos/sin position embeddings
1332
+ """
1333
+ rot_dim = freqs[0].shape[-1]
1334
+ cos, sin = freqs
1335
+ t_float = t.float()
1336
+ if apply_rotary_emb_func is not None and t.is_cuda:
1337
+ # apply_rotary_emb in flash_attn requires cos/sin to be of
1338
+ # shape (seqlen, rotary_dim / 2) and apply rotary embedding
1339
+ # to the first rotary_dim of the input
1340
+ cos = cos.squeeze(0).squeeze(1)[:, : rot_dim // 2]
1341
+ sin = sin.squeeze(0).squeeze(1)[:, : rot_dim // 2]
1342
+ return apply_rotary_emb_func(t_float, cos, sin).type_as(t)
1343
+ else:
1344
+ t_rot, t_pass = t_float[..., :rot_dim], t_float[..., rot_dim:]
1345
+ t_rot = (t_rot * cos) + (_rotate_half(t_rot) * sin)
1346
+ return torch.cat((t_rot, t_pass), dim=-1).type_as(t)
1347
+
1348
+
1349
+ class RMSNorm(torch.nn.Module):
1350
+ def __init__(self, dim: int, eps: float = 1e-6):
1351
+ super().__init__()
1352
+ self.eps = eps
1353
+ self.weight = nn.Parameter(torch.ones(dim))
1354
+
1355
+ def _norm(self, x):
1356
+ return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
1357
+
1358
+ def forward(self, x):
1359
+ if rms_norm is not None and x.is_cuda:
1360
+ return rms_norm(x, self.weight, self.eps)
1361
+ else:
1362
+ output = self._norm(x.float()).type_as(x)
1363
+ return output * self.weight
qwen.tiktoken ADDED
The diff for this file is too large to render. See raw diff
 
qwen_generation_utils.py ADDED
@@ -0,0 +1,416 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Alibaba Cloud.
2
+ #
3
+ # This source code is licensed under the license found in the
4
+ # LICENSE file in the root directory of this source tree.
5
+
6
+ """Generation support."""
7
+
8
+ from typing import Tuple, List, Union, Iterable
9
+
10
+ import numpy as np
11
+ import torch
12
+ import torch.nn.functional as F
13
+ from transformers import PreTrainedTokenizer
14
+ from transformers import logging
15
+ from transformers.generation import LogitsProcessor
16
+
17
+ logger = logging.get_logger(__name__)
18
+
19
+ # Types.
20
+ HistoryType = List[Tuple[str, str]]
21
+ TokensType = List[int]
22
+ BatchTokensType = List[List[int]]
23
+
24
+
25
+ def pad_batch(batch: BatchTokensType, pad_id: int, seq_length: int) -> BatchTokensType:
26
+ for tokens in batch:
27
+ context_length = len(tokens)
28
+ if context_length < seq_length:
29
+ tokens.extend([pad_id] * (seq_length - context_length))
30
+ return batch
31
+
32
+
33
+ def get_ltor_masks_and_position_ids(
34
+ data,
35
+ eod_token,
36
+ reset_position_ids,
37
+ reset_attention_mask,
38
+ eod_mask_loss,
39
+ ):
40
+ """Build masks and position id for left to right model."""
41
+
42
+ # Extract batch size and sequence length.
43
+ micro_batch_size, seq_length = data.size()
44
+
45
+ # Attention mask (lower triangular).
46
+ if reset_attention_mask:
47
+ att_mask_batch = micro_batch_size
48
+ else:
49
+ att_mask_batch = 1
50
+ attention_mask = torch.tril(
51
+ torch.ones((att_mask_batch, seq_length, seq_length), device=data.device)
52
+ ).view(att_mask_batch, 1, seq_length, seq_length)
53
+
54
+ # Loss mask.
55
+ loss_mask = torch.ones(data.size(), dtype=torch.float, device=data.device)
56
+ if eod_mask_loss:
57
+ loss_mask[data == eod_token] = 0.0
58
+
59
+ # Position ids.
60
+ position_ids = torch.arange(seq_length, dtype=torch.long, device=data.device)
61
+ position_ids = position_ids.unsqueeze(0).expand_as(data)
62
+ # We need to clone as the ids will be modifed based on batch index.
63
+ if reset_position_ids:
64
+ position_ids = position_ids.clone()
65
+
66
+ if reset_position_ids or reset_attention_mask:
67
+ # Loop through the batches:
68
+ for b in range(micro_batch_size):
69
+
70
+ # Find indecies where EOD token is.
71
+ eod_index = position_ids[b, data[b] == eod_token]
72
+ # Detach indecies from positions if going to modify positions.
73
+ if reset_position_ids:
74
+ eod_index = eod_index.clone()
75
+
76
+ # Loop through EOD indecies:
77
+ prev_index = 0
78
+ for j in range(eod_index.size()[0]):
79
+ i = eod_index[j]
80
+ # Mask attention loss.
81
+ if reset_attention_mask:
82
+ attention_mask[b, 0, (i + 1) :, : (i + 1)] = 0
83
+ # Reset positions.
84
+ if reset_position_ids:
85
+ position_ids[b, (i + 1) :] -= i + 1 - prev_index
86
+ prev_index = i + 1
87
+
88
+ # Convert attention mask to binary:
89
+ attention_mask = attention_mask < 0.5
90
+
91
+ return attention_mask, loss_mask, position_ids
92
+
93
+
94
+ def get_batch(context_tokens: torch.LongTensor, eod_id: int):
95
+ """Generate batch from context tokens."""
96
+ # Move to GPU.
97
+ tokens = context_tokens.contiguous().to(context_tokens.device)
98
+ # Get the attention mask and postition ids.
99
+ attention_mask, _, position_ids = get_ltor_masks_and_position_ids(
100
+ tokens,
101
+ eod_id,
102
+ reset_position_ids=False,
103
+ reset_attention_mask=False,
104
+ eod_mask_loss=False,
105
+ )
106
+ return tokens, attention_mask, position_ids
107
+
108
+
109
+ def get_stop_words_ids(chat_format, tokenizer):
110
+ if chat_format == "raw":
111
+ stop_words_ids = [tokenizer.encode("Human:"), [tokenizer.eod_id]]
112
+ elif chat_format == "chatml":
113
+ stop_words_ids = [[tokenizer.im_end_id], [tokenizer.im_start_id]]
114
+ else:
115
+ raise NotImplementedError(f"Unknown chat format {chat_format!r}")
116
+ return stop_words_ids
117
+
118
+
119
+ def make_context(
120
+ tokenizer: PreTrainedTokenizer,
121
+ query: str,
122
+ history: List[Tuple[str, str]] = None,
123
+ system: str = "",
124
+ max_window_size: int = 6144,
125
+ chat_format: str = "chatml",
126
+ ):
127
+ if history is None:
128
+ history = []
129
+
130
+ if chat_format == "chatml":
131
+ im_start, im_end = "<|im_start|>", "<|im_end|>"
132
+ im_start_tokens = [tokenizer.im_start_id]
133
+ im_end_tokens = [tokenizer.im_end_id]
134
+ nl_tokens = tokenizer.encode("\n")
135
+
136
+ def _tokenize_str(role, content):
137
+ return f"{role}\n{content}", tokenizer.encode(
138
+ role, allowed_special=set()
139
+ ) + nl_tokens + tokenizer.encode(content, allowed_special=set())
140
+
141
+ system_text, system_tokens_part = _tokenize_str("system", system)
142
+ system_tokens = im_start_tokens + system_tokens_part + im_end_tokens
143
+
144
+ raw_text = ""
145
+ context_tokens = []
146
+
147
+ for turn_query, turn_response in reversed(history):
148
+ query_text, query_tokens_part = _tokenize_str("user", turn_query)
149
+ query_tokens = im_start_tokens + query_tokens_part + im_end_tokens
150
+ response_text, response_tokens_part = _tokenize_str(
151
+ "assistant", turn_response
152
+ )
153
+ response_tokens = im_start_tokens + response_tokens_part + im_end_tokens
154
+
155
+ next_context_tokens = nl_tokens + query_tokens + nl_tokens + response_tokens
156
+ prev_chat = (
157
+ f"\n{im_start}{query_text}{im_end}\n{im_start}{response_text}{im_end}"
158
+ )
159
+
160
+ current_context_size = (
161
+ len(system_tokens) + len(next_context_tokens) + len(context_tokens)
162
+ )
163
+ if current_context_size < max_window_size:
164
+ context_tokens = next_context_tokens + context_tokens
165
+ raw_text = prev_chat + raw_text
166
+ else:
167
+ break
168
+
169
+ context_tokens = system_tokens + context_tokens
170
+ raw_text = f"{im_start}{system_text}{im_end}" + raw_text
171
+ context_tokens += (
172
+ nl_tokens
173
+ + im_start_tokens
174
+ + _tokenize_str("user", query)[1]
175
+ + im_end_tokens
176
+ + nl_tokens
177
+ + im_start_tokens
178
+ + tokenizer.encode("assistant")
179
+ + nl_tokens
180
+ )
181
+ raw_text += f"\n{im_start}user\n{query}{im_end}\n{im_start}assistant\n"
182
+
183
+ elif chat_format == "raw":
184
+ raw_text = query
185
+ context_tokens = tokenizer.encode(raw_text)
186
+ else:
187
+ raise NotImplementedError(f"Unknown chat format {chat_format!r}")
188
+
189
+ return raw_text, context_tokens
190
+
191
+
192
+ def _decode_default(
193
+ tokens: List[int],
194
+ *,
195
+ stop_words: List[str],
196
+ eod_words: List[str],
197
+ tokenizer: PreTrainedTokenizer,
198
+ raw_text_len: int,
199
+ verbose: bool = False,
200
+ return_end_reason: bool = False,
201
+ errors: str='replace',
202
+ ):
203
+ trim_decode_tokens = tokenizer.decode(tokens, errors=errors)[raw_text_len:]
204
+ if verbose:
205
+ print("\nRaw Generate: ", trim_decode_tokens)
206
+
207
+ end_reason = f"Gen length {len(tokens)}"
208
+ for stop_word in stop_words:
209
+ trim_decode_tokens = trim_decode_tokens.replace(stop_word, "").strip()
210
+ for eod_word in eod_words:
211
+ if eod_word in trim_decode_tokens:
212
+ end_reason = f"Gen {eod_word!r}"
213
+ trim_decode_tokens = trim_decode_tokens.split(eod_word)[0]
214
+ trim_decode_tokens = trim_decode_tokens.strip()
215
+ if verbose:
216
+ print("\nEnd Reason:", end_reason)
217
+ print("\nGenerate: ", trim_decode_tokens)
218
+
219
+ if return_end_reason:
220
+ return trim_decode_tokens, end_reason
221
+ else:
222
+ return trim_decode_tokens
223
+
224
+
225
+ def _decode_chatml(
226
+ tokens: List[int],
227
+ *,
228
+ stop_words: List[str],
229
+ eod_token_ids: List[int],
230
+ tokenizer: PreTrainedTokenizer,
231
+ raw_text_len: int,
232
+ context_length: int,
233
+ verbose: bool = False,
234
+ return_end_reason: bool = False,
235
+ errors: str='replace'
236
+ ):
237
+ end_reason = f"Gen length {len(tokens)}"
238
+ eod_token_idx = context_length
239
+ for eod_token_idx in range(context_length, len(tokens)):
240
+ if tokens[eod_token_idx] in eod_token_ids:
241
+ end_reason = f"Gen {tokenizer.decode([tokens[eod_token_idx]])!r}"
242
+ break
243
+
244
+ trim_decode_tokens = tokenizer.decode(tokens[:eod_token_idx], errors=errors)[raw_text_len:]
245
+ if verbose:
246
+ print("\nRaw Generate w/o EOD:", tokenizer.decode(tokens, errors=errors)[raw_text_len:])
247
+ print("\nRaw Generate:", trim_decode_tokens)
248
+ print("\nEnd Reason:", end_reason)
249
+ for stop_word in stop_words:
250
+ trim_decode_tokens = trim_decode_tokens.replace(stop_word, "").strip()
251
+ trim_decode_tokens = trim_decode_tokens.strip()
252
+ if verbose:
253
+ print("\nGenerate:", trim_decode_tokens)
254
+
255
+ if return_end_reason:
256
+ return trim_decode_tokens, end_reason
257
+ else:
258
+ return trim_decode_tokens
259
+
260
+
261
+ def decode_tokens(
262
+ tokens: Union[torch.LongTensor, TokensType],
263
+ tokenizer: PreTrainedTokenizer,
264
+ raw_text_len: int,
265
+ context_length: int,
266
+ chat_format: str,
267
+ verbose: bool = False,
268
+ return_end_reason: bool = False,
269
+ errors: str="replace",
270
+ ) -> str:
271
+ if torch.is_tensor(tokens):
272
+ tokens = tokens.cpu().numpy().tolist()
273
+
274
+ if chat_format == "chatml":
275
+ return _decode_chatml(
276
+ tokens,
277
+ stop_words=[],
278
+ eod_token_ids=[tokenizer.im_start_id, tokenizer.im_end_id],
279
+ tokenizer=tokenizer,
280
+ raw_text_len=raw_text_len,
281
+ context_length=context_length,
282
+ verbose=verbose,
283
+ return_end_reason=return_end_reason,
284
+ errors=errors,
285
+ )
286
+ elif chat_format == "raw":
287
+ return _decode_default(
288
+ tokens,
289
+ stop_words=["<|endoftext|>"],
290
+ eod_words=["<|endoftext|>"],
291
+ tokenizer=tokenizer,
292
+ raw_text_len=raw_text_len,
293
+ verbose=verbose,
294
+ return_end_reason=return_end_reason,
295
+ errors=errors,
296
+ )
297
+ else:
298
+ raise NotImplementedError(f"Unknown chat format {chat_format!r}")
299
+
300
+
301
+ class StopWordsLogitsProcessor(LogitsProcessor):
302
+ """
303
+ :class:`transformers.LogitsProcessor` that enforces that when specified sequences appear, stop geration.
304
+
305
+ Args:
306
+ stop_words_ids (:obj:`List[List[int]]`):
307
+ List of list of token ids of stop ids. In order to get the tokens of the words
308
+ that should not appear in the generated text, use :obj:`tokenizer(bad_word,
309
+ add_prefix_space=True).input_ids`.
310
+ eos_token_id (:obj:`int`):
311
+ The id of the `end-of-sequence` token.
312
+ """
313
+
314
+ def __init__(self, stop_words_ids: Iterable[Iterable[int]], eos_token_id: int):
315
+
316
+ if not isinstance(stop_words_ids, List) or len(stop_words_ids) == 0:
317
+ raise ValueError(
318
+ f"`stop_words_ids` has to be a non-emtpy list, but is {stop_words_ids}."
319
+ )
320
+ if any(not isinstance(bad_word_ids, list) for bad_word_ids in stop_words_ids):
321
+ raise ValueError(
322
+ f"`stop_words_ids` has to be a list of lists, but is {stop_words_ids}."
323
+ )
324
+ if any(
325
+ any(
326
+ (not isinstance(token_id, (int, np.integer)) or token_id < 0)
327
+ for token_id in stop_word_ids
328
+ )
329
+ for stop_word_ids in stop_words_ids
330
+ ):
331
+ raise ValueError(
332
+ f"Each list in `stop_words_ids` has to be a list of positive integers, but is {stop_words_ids}."
333
+ )
334
+
335
+ self.stop_words_ids = list(
336
+ filter(
337
+ lambda bad_token_seq: bad_token_seq != [eos_token_id], stop_words_ids
338
+ )
339
+ )
340
+ self.eos_token_id = eos_token_id
341
+ for stop_token_seq in self.stop_words_ids:
342
+ assert (
343
+ len(stop_token_seq) > 0
344
+ ), "Stop words token sequences {} cannot have an empty list".format(
345
+ stop_words_ids
346
+ )
347
+
348
+ def __call__(
349
+ self, input_ids: torch.LongTensor, scores: torch.FloatTensor
350
+ ) -> torch.FloatTensor:
351
+ stopped_samples = self._calc_stopped_samples(input_ids)
352
+ for i, should_stop in enumerate(stopped_samples):
353
+ if should_stop:
354
+ scores[i, self.eos_token_id] = float(2**15)
355
+ return scores
356
+
357
+ def _tokens_match(self, prev_tokens: torch.LongTensor, tokens: List[int]) -> bool:
358
+ if len(tokens) == 0:
359
+ # if bad word tokens is just one token always ban it
360
+ return True
361
+ elif len(tokens) > len(prev_tokens):
362
+ # if bad word tokens are longer then prev input_ids they can't be equal
363
+ return False
364
+ elif prev_tokens[-len(tokens) :].tolist() == tokens:
365
+ # if tokens match
366
+ return True
367
+ else:
368
+ return False
369
+
370
+ def _calc_stopped_samples(self, prev_input_ids: Iterable[int]) -> Iterable[int]:
371
+ stopped_samples = []
372
+ for prev_input_ids_slice in prev_input_ids:
373
+ match = False
374
+ for stop_token_seq in self.stop_words_ids:
375
+ if self._tokens_match(prev_input_ids_slice, stop_token_seq):
376
+ # if tokens do not match continue
377
+ match = True
378
+ break
379
+ stopped_samples.append(match)
380
+
381
+ return stopped_samples
382
+
383
+
384
+ def top_k_logits(logits, top_k=0, top_p=0.0, filter_value=-float("Inf")):
385
+ """This function has been mostly taken from huggingface conversational
386
+ ai code at
387
+ https://medium.com/huggingface/how-to-build-a-state-of-the-art-
388
+ conversational-ai-with-transfer-learning-2d818ac26313"""
389
+
390
+ if top_k > 0:
391
+ # Remove all tokens with a probability less than the
392
+ # last token of the top-k
393
+ indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
394
+ logits[indices_to_remove] = filter_value
395
+
396
+ if top_p > 0.0:
397
+ # Cconvert to 1D
398
+ sorted_logits, sorted_indices = torch.sort(logits, descending=True, dim=-1)
399
+ cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
400
+
401
+ # Remove tokens with cumulative probability above the threshold
402
+ sorted_indices_to_remove = cumulative_probs > top_p
403
+ # Shift the indices to the right to keep also the first token
404
+ # above the threshold
405
+ sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
406
+ sorted_indices_to_remove[..., 0] = 0
407
+ for i in range(sorted_indices.size(0)):
408
+ indices_to_remove = sorted_indices[i][sorted_indices_to_remove[i]]
409
+ logits[i][indices_to_remove] = filter_value
410
+
411
+ return logits
412
+
413
+
414
+ def switch(val1, val2, boolean):
415
+ boolean = boolean.type_as(val1)
416
+ return (1 - boolean) * val1 + boolean * val2