edgerazor 1.2.6__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (44) hide show
  1. edgerazor-1.2.6/LICENSE +201 -0
  2. edgerazor-1.2.6/PKG-INFO +234 -0
  3. edgerazor-1.2.6/README.md +198 -0
  4. edgerazor-1.2.6/pyproject.toml +160 -0
  5. edgerazor-1.2.6/setup.cfg +4 -0
  6. edgerazor-1.2.6/src/edgerazor/__init__.py +57 -0
  7. edgerazor-1.2.6/src/edgerazor/edgerazor.py +173 -0
  8. edgerazor-1.2.6/src/edgerazor/edgerazor_config.py +554 -0
  9. edgerazor-1.2.6/src/edgerazor/kd/__init__.py +41 -0
  10. edgerazor-1.2.6/src/edgerazor/kd/kd.py +933 -0
  11. edgerazor-1.2.6/src/edgerazor/kd/util/__init__.py +29 -0
  12. edgerazor-1.2.6/src/edgerazor/kd/util/distill_config.py +228 -0
  13. edgerazor-1.2.6/src/edgerazor/kd/util/distill_function.py +459 -0
  14. edgerazor-1.2.6/src/edgerazor/kd/util/distill_function_config.py +47 -0
  15. edgerazor-1.2.6/src/edgerazor/kd/util/layer_select.py +88 -0
  16. edgerazor-1.2.6/src/edgerazor/kd/util/moe_loss.py +52 -0
  17. edgerazor-1.2.6/src/edgerazor/log.py +176 -0
  18. edgerazor-1.2.6/src/edgerazor/qat/__init__.py +40 -0
  19. edgerazor-1.2.6/src/edgerazor/qat/block/__init__.py +31 -0
  20. edgerazor-1.2.6/src/edgerazor/qat/block/qattn.py +458 -0
  21. edgerazor-1.2.6/src/edgerazor/qat/block/qkv_cache_llama.py +134 -0
  22. edgerazor-1.2.6/src/edgerazor/qat/block/qkv_cache_olmoe.py +449 -0
  23. edgerazor-1.2.6/src/edgerazor/qat/block/qkv_cache_qwen2_5omni.py +151 -0
  24. edgerazor-1.2.6/src/edgerazor/qat/block/qkv_cache_qwen3.py +142 -0
  25. edgerazor-1.2.6/src/edgerazor/qat/block/qkv_cache_qwen3moe.py +142 -0
  26. edgerazor-1.2.6/src/edgerazor/qat/map.py +606 -0
  27. edgerazor-1.2.6/src/edgerazor/qat/module/__init__.py +15 -0
  28. edgerazor-1.2.6/src/edgerazor/qat/module/qconv1d.py +157 -0
  29. edgerazor-1.2.6/src/edgerazor/qat/module/qconv2d.py +158 -0
  30. edgerazor-1.2.6/src/edgerazor/qat/module/qconv3d.py +159 -0
  31. edgerazor-1.2.6/src/edgerazor/qat/module/qembedding.py +112 -0
  32. edgerazor-1.2.6/src/edgerazor/qat/module/qlinear.py +178 -0
  33. edgerazor-1.2.6/src/edgerazor/qat/qat.py +643 -0
  34. edgerazor-1.2.6/src/edgerazor/qat/quantize.py +274 -0
  35. edgerazor-1.2.6/src/edgerazor/qat/util/__init__.py +5 -0
  36. edgerazor-1.2.6/src/edgerazor/qat/util/quant_config.py +741 -0
  37. edgerazor-1.2.6/src/edgerazor/qat/util/quant_function.py +1377 -0
  38. edgerazor-1.2.6/src/edgerazor/qat/util/quant_function_config.py +13 -0
  39. edgerazor-1.2.6/src/edgerazor/qat/util/quant_selector.py +159 -0
  40. edgerazor-1.2.6/src/edgerazor.egg-info/PKG-INFO +234 -0
  41. edgerazor-1.2.6/src/edgerazor.egg-info/SOURCES.txt +42 -0
  42. edgerazor-1.2.6/src/edgerazor.egg-info/dependency_links.txt +1 -0
  43. edgerazor-1.2.6/src/edgerazor.egg-info/requires.txt +15 -0
  44. edgerazor-1.2.6/src/edgerazor.egg-info/top_level.txt +1 -0
@@ -0,0 +1,201 @@
1
+ Apache License
2
+ Version 2.0, January 2004
3
+ http://www.apache.org/licenses/
4
+
5
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
6
+
7
+ 1. Definitions.
8
+
9
+ "License" shall mean the terms and conditions for use, reproduction,
10
+ and distribution as defined by Sections 1 through 9 of this document.
11
+
12
+ "Licensor" shall mean the copyright owner or entity authorized by
13
+ the copyright owner that is granting the License.
14
+
15
+ "Legal Entity" shall mean the union of the acting entity and all
16
+ other entities that control, are controlled by, or are under common
17
+ control with that entity. For the purposes of this definition,
18
+ "control" means (i) the power, direct or indirect, to cause the
19
+ direction or management of such entity, whether by contract or
20
+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
21
+ outstanding shares, or (iii) beneficial ownership of such entity.
22
+
23
+ "You" (or "Your") shall mean an individual or Legal Entity
24
+ exercising permissions granted by this License.
25
+
26
+ "Source" form shall mean the preferred form for making modifications,
27
+ including but not limited to software source code, documentation
28
+ source, and configuration files.
29
+
30
+ "Object" form shall mean any form resulting from mechanical
31
+ transformation or translation of a Source form, including but
32
+ not limited to compiled object code, generated documentation,
33
+ and conversions to other media types.
34
+
35
+ "Work" shall mean the work of authorship, whether in Source or
36
+ Object form, made available under the License, as indicated by a
37
+ copyright notice that is included in or attached to the work
38
+ (an example is provided in the Appendix below).
39
+
40
+ "Derivative Works" shall mean any work, whether in Source or Object
41
+ form, that is based on (or derived from) the Work and for which the
42
+ editorial revisions, annotations, elaborations, or other modifications
43
+ represent, as a whole, an original work of authorship. For the purposes
44
+ of this License, Derivative Works shall not include works that remain
45
+ separable from, or merely link (or bind by name) to the interfaces of,
46
+ the Work and Derivative Works thereof.
47
+
48
+ "Contribution" shall mean any work of authorship, including
49
+ the original version of the Work and any modifications or additions
50
+ to that Work or Derivative Works thereof, that is intentionally
51
+ submitted to Licensor for inclusion in the Work by the copyright owner
52
+ or by an individual or Legal Entity authorized to submit on behalf of
53
+ the copyright owner. For the purposes of this definition, "submitted"
54
+ means any form of electronic, verbal, or written communication sent
55
+ to the Licensor or its representatives, including but not limited to
56
+ communication on electronic mailing lists, source code control systems,
57
+ and issue tracking systems that are managed by, or on behalf of, the
58
+ Licensor for the purpose of discussing and improving the Work, but
59
+ excluding communication that is conspicuously marked or otherwise
60
+ designated in writing by the copyright owner as "Not a Contribution."
61
+
62
+ "Contributor" shall mean Licensor and any individual or Legal Entity
63
+ on behalf of whom a Contribution has been received by Licensor and
64
+ subsequently incorporated within the Work.
65
+
66
+ 2. Grant of Copyright License. Subject to the terms and conditions of
67
+ this License, each Contributor hereby grants to You a perpetual,
68
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
69
+ copyright license to reproduce, prepare Derivative Works of,
70
+ publicly display, publicly perform, sublicense, and distribute the
71
+ Work and such Derivative Works in Source or Object form.
72
+
73
+ 3. Grant of Patent License. Subject to the terms and conditions of
74
+ this License, each Contributor hereby grants to You a perpetual,
75
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
76
+ (except as stated in this section) patent license to make, have made,
77
+ use, offer to sell, sell, import, and otherwise transfer the Work,
78
+ where such license applies only to those patent claims licensable
79
+ by such Contributor that are necessarily infringed by their
80
+ Contribution(s) alone or by combination of their Contribution(s)
81
+ with the Work to which such Contribution(s) was submitted. If You
82
+ institute patent litigation against any entity (including a
83
+ cross-claim or counterclaim in a lawsuit) alleging that the Work
84
+ or a Contribution incorporated within the Work constitutes direct
85
+ or contributory patent infringement, then any patent licenses
86
+ granted to You under this License for that Work shall terminate
87
+ as of the date such litigation is filed.
88
+
89
+ 4. Redistribution. You may reproduce and distribute copies of the
90
+ Work or Derivative Works thereof in any medium, with or without
91
+ modifications, and in Source or Object form, provided that You
92
+ meet the following conditions:
93
+
94
+ (a) You must give any other recipients of the Work or
95
+ Derivative Works a copy of this License; and
96
+
97
+ (b) You must cause any modified files to carry prominent notices
98
+ stating that You changed the files; and
99
+
100
+ (c) You must retain, in the Source form of any Derivative Works
101
+ that You distribute, all copyright, patent, trademark, and
102
+ attribution notices from the Source form of the Work,
103
+ excluding those notices that do not pertain to any part of
104
+ the Derivative Works; and
105
+
106
+ (d) If the Work includes a "NOTICE" text file as part of its
107
+ distribution, then any Derivative Works that You distribute must
108
+ include a readable copy of the attribution notices contained
109
+ within such NOTICE file, excluding those notices that do not
110
+ pertain to any part of the Derivative Works, in at least one
111
+ of the following places: within a NOTICE text file distributed
112
+ as part of the Derivative Works; within the Source form or
113
+ documentation, if provided along with the Derivative Works; or,
114
+ within a display generated by the Derivative Works, if and
115
+ wherever such third-party notices normally appear. The contents
116
+ of the NOTICE file are for informational purposes only and
117
+ do not modify the License. You may add Your own attribution
118
+ notices within Derivative Works that You distribute, alongside
119
+ or as an addendum to the NOTICE text from the Work, provided
120
+ that such additional attribution notices cannot be construed
121
+ as modifying the License.
122
+
123
+ You may add Your own copyright statement to Your modifications and
124
+ may provide additional or different license terms and conditions
125
+ for use, reproduction, or distribution of Your modifications, or
126
+ for any such Derivative Works as a whole, provided Your use,
127
+ reproduction, and distribution of the Work otherwise complies with
128
+ the conditions stated in this License.
129
+
130
+ 5. Submission of Contributions. Unless You explicitly state otherwise,
131
+ any Contribution intentionally submitted for inclusion in the Work
132
+ by You to the Licensor shall be under the terms and conditions of
133
+ this License, without any additional terms or conditions.
134
+ Notwithstanding the above, nothing herein shall supersede or modify
135
+ the terms of any separate license agreement you may have executed
136
+ with Licensor regarding such Contributions.
137
+
138
+ 6. Trademarks. This License does not grant permission to use the trade
139
+ names, trademarks, service marks, or product names of the Licensor,
140
+ except as required for reasonable and customary use in describing the
141
+ origin of the Work and reproducing the content of the NOTICE file.
142
+
143
+ 7. Disclaimer of Warranty. Unless required by applicable law or
144
+ agreed to in writing, Licensor provides the Work (and each
145
+ Contributor provides its Contributions) on an "AS IS" BASIS,
146
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
147
+ implied, including, without limitation, any warranties or conditions
148
+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
149
+ PARTICULAR PURPOSE. You are solely responsible for determining the
150
+ appropriateness of using or redistributing the Work and assume any
151
+ risks associated with Your exercise of permissions under this License.
152
+
153
+ 8. Limitation of Liability. In no event and under no legal theory,
154
+ whether in tort (including negligence), contract, or otherwise,
155
+ unless required by applicable law (such as deliberate and grossly
156
+ negligent acts) or agreed to in writing, shall any Contributor be
157
+ liable to You for damages, including any direct, indirect, special,
158
+ incidental, or consequential damages of any character arising as a
159
+ result of this License or out of the use or inability to use the
160
+ Work (including but not limited to damages for loss of goodwill,
161
+ work stoppage, computer failure or malfunction, or any and all
162
+ other commercial damages or losses), even if such Contributor
163
+ has been advised of the possibility of such damages.
164
+
165
+ 9. Accepting Warranty or Additional Liability. While redistributing
166
+ the Work or Derivative Works thereof, You may choose to offer,
167
+ and charge a fee for, acceptance of support, warranty, indemnity,
168
+ or other liability obligations and/or rights consistent with this
169
+ License. However, in accepting such obligations, You may act only
170
+ on Your own behalf and on Your sole responsibility, not on behalf
171
+ of any other Contributor, and only if You agree to indemnify,
172
+ defend, and hold each Contributor harmless for any liability
173
+ incurred by, or claims asserted against, such Contributor by reason
174
+ of your accepting any such warranty or additional liability.
175
+
176
+ END OF TERMS AND CONDITIONS
177
+
178
+ APPENDIX: How to apply the Apache License to your work.
179
+
180
+ To apply the Apache License to your work, attach the following
181
+ boilerplate notice, with the fields enclosed by brackets "[]"
182
+ replaced with your own identifying information. (Don't include
183
+ the brackets!) The text should be enclosed in the appropriate
184
+ comment syntax for the file format. We also recommend that a
185
+ file or class name and description of purpose be included on the
186
+ same "printed page" as the copyright notice for easier
187
+ identification within third-party archives.
188
+
189
+ Copyright 2026 LAMDA
190
+
191
+ Licensed under the Apache License, Version 2.0 (the "License");
192
+ you may not use this file except in compliance with the License.
193
+ You may obtain a copy of the License at
194
+
195
+ http://www.apache.org/licenses/LICENSE-2.0
196
+
197
+ Unless required by applicable law or agreed to in writing, software
198
+ distributed under the License is distributed on an "AS IS" BASIS,
199
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
200
+ See the License for the specific language governing permissions and
201
+ limitations under the License.
@@ -0,0 +1,234 @@
1
+ Metadata-Version: 2.4
2
+ Name: edgerazor
3
+ Version: 1.2.6
4
+ Summary: EdgeRazor: Lightweight framework for Edge AI.
5
+ Author: Shu-Hao Zhang
6
+ Author-email: zhangsh@lamda.nju.edu.cn
7
+ License: MIT
8
+ Project-URL: Homepage, https://github.com/zhangsq-nju/EdgeRazor
9
+ Keywords: Lightweight AI,Edge AI,Model Compression,Quantization-Aware Distillation,Quantization,Distillation,LLM,Lightweight LLM,LLM Compression,LLM Quantization,LLM Distillation
10
+ Classifier: Development Status :: 4 - Beta
11
+ Classifier: Intended Audience :: Developers
12
+ Classifier: Intended Audience :: Science/Research
13
+ Classifier: Programming Language :: Python :: 3
14
+ Classifier: Programming Language :: Python :: 3.10
15
+ Classifier: Operating System :: Unix
16
+ Classifier: Operating System :: POSIX :: Linux
17
+ Classifier: Operating System :: MacOS :: MacOS X
18
+ Classifier: Operating System :: Microsoft :: Windows
19
+ Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
20
+ Requires-Python: ==3.10.20
21
+ Description-Content-Type: text/markdown
22
+ License-File: LICENSE
23
+ Requires-Dist: ruff>=0.13.1
24
+ Requires-Dist: transformers==4.57.1
25
+ Requires-Dist: pyyaml==6.0.3
26
+ Provides-Extra: cpu
27
+ Requires-Dist: torch==2.9.1; extra == "cpu"
28
+ Requires-Dist: torchvision==0.24.1; extra == "cpu"
29
+ Provides-Extra: cu126
30
+ Requires-Dist: torch==2.9.1; extra == "cu126"
31
+ Requires-Dist: torchvision==0.24.1; extra == "cu126"
32
+ Provides-Extra: cu128
33
+ Requires-Dist: torch==2.9.1; extra == "cu128"
34
+ Requires-Dist: torchvision==0.24.1; extra == "cu128"
35
+ Dynamic: license-file
36
+
37
+ <div align="center">
38
+ <br/>
39
+ <img src="./asset/Logo-full.png" alt="EdgeRazor Logo" width="60%">
40
+ <h3>
41
+ Lightweight Framework for Edge AI
42
+ </h3>
43
+
44
+ <p>
45
+ <!-- <a href="https://arxiv.org/abs/2604.xxxxx" target="_blank">
46
+ <img src="https://img.shields.io/badge/arXiv-EdgeRazor-b31b1b?style=flat&logo=arxiv" alt="arXiv EdgeRazor">
47
+ </a> -->
48
+ <a href="https://huggingface.co/collections/zhangsq-nju/edgerazor-nbit" target="_blank">
49
+ <img src="https://img.shields.io/badge/HuggingFace-Collection-FFD21F?style=flat&logo=huggingface&logoColor=FFD21F" alt="Hugging Face Collection">
50
+ </a>
51
+ <a href="./README_ZH.md">
52
+ <img src="https://img.shields.io/badge/README-ZH-blue?style=flat&logo=readme" alt="README ZH">
53
+ </a>
54
+ <a href="https://github.com/zhangsq-nju/EdgeRazor/blob/main/LICENSE">
55
+ <img src="https://img.shields.io/badge/License-Apache_2.0-green?logo=opensourceinitiative&logoColor=green" alt="License: Apache 2.0">
56
+ </a>
57
+ </p>
58
+
59
+ <h5>
60
+ ✨ If you like our project, please give us a star ⭐️ for the latest update.
61
+ </h5>
62
+
63
+ <img src="./asset/Demo.gif" alt="EdgeRazor Demo" width="75%">
64
+ </div>
65
+
66
+ ---
67
+
68
+ **EdgeRazor** is a unified and lightweight framework for edge AI, designed to train models that are smaller, faster, and deployable across diverse hardware, ranging from mobile and edge endpoints to latency-sensitive clouds. The EdgeRazor framework **seamlessly integrates** model compression techniques into existing full-precision training pipelines with **minimal code modification**, preserving promising task performance and enabling low-cost and high-efficiency computations.
69
+
70
+ EdgeRazor currently focuses on low-bit LLM compression via configurable quantization-aware distillation. In terms of **quantization**, EdgeRazor supports quantizing weights (including embedding and lm_head layers), activations, and KV cache. Quantized bit-widths include the uniform 1.58-bit and 4-bit, as well as matrix-wise mixed-precision, such as 2.79-bit (50% 4-bit + 50% 1.58-bit) and 1.88-bit (12.5% 4-bit + 87.5% 1.58-bit). In terms of **distillation**, EdgeRazor offers the logits, features, and attention distillation, all of which can be flexibly combined within a unified configuration interface.
71
+
72
+ EdgeRazor achieves the state-of-the-art performance across a range of models, including base LLMs, instruction-tuned LLMs, and multimodal LLMs. For W-A8-KV8 quantization, **Qwen3-0.6B-EdgeRazor** attains average scores of **47.80** / **44.10** / **41.76** / **39.81** at 4-bit / 2.79-bit / 1.88-bit / 1.58-bit, corresponding to compression ratios of **3.94×** / **5.05×** / **6.40×** / **7.03×**, respectively. In comparison, the best prior methods achieve <u>45.74</u> / <u>37.38</u> / <u>30.49</u> at 4-bit / 3-bit / 2-bit with compression ratios of <u>2.21×</u> / <u>2.47×</u> / <u>2.78×</u>.
73
+
74
+ <p align="center">
75
+ <img src="./asset/Architeacture.png" alt="EdgeRazor Architecture">
76
+ <br> Figure: The EdgeRazor framework with lightweight model training pipeline.
77
+ </p>
78
+
79
+ ## News
80
+
81
+ - 🔥 **[2026-04]**: 🏆 Low-bit LLMs by EdgeRazor is released! Check our Hugging Face collection: [zhangsq-nju/edgerazor-nbit](https://huggingface.co/collections/zhangsq-nju/edgerazor-nbit).
82
+ - 🔥 **[2026-04]**: 🛠️ Open-sourced EdgeRazor-V1 is released! Now configurable on diverse models for seamless integration and customization!
83
+ - 🔥 **[2025-10]**: 📄 Paper-TernaryCLIP is available on arXiv: [2510.21879](https://arxiv.org/abs/2510.21879)!
84
+ <!-- - 🔥 **[2026-04]**: 📄 Paper-EdgeRazor is available on arXiv: [2604.xxxxx](https://arxiv.org/abs/2604.xxxxx)! -->
85
+
86
+ ## Contents
87
+
88
+ - [News](#news)
89
+ - [Contents](#contents)
90
+ - [Getting Started](#getting-started)
91
+ - [Installation](#installation)
92
+ - [Usage](#usage)
93
+ - [Low-bit LLMs Serving on Docker](#low-bit-llms-serving-on-docker)
94
+ - [Main Techniques](#main-techniques)
95
+ - [Applications](#applications)
96
+ - [Model Zoo](#model-zoo)
97
+ - [LLMs](#llms)
98
+ - [MLLMs](#mllms)
99
+ - [Todo List](#todo-list)
100
+ - [Citation](#citation)
101
+ - [Contributor List](#contributor-list)
102
+
103
+ ## Getting Started
104
+
105
+ ### Installation
106
+
107
+ - Download from GitHub (latest version)
108
+
109
+ ```
110
+ git clone https://github.com/zhangsq-nju/EdgeRazor.git && cd EdgeRazor
111
+ conda create -n edgerazor python=3.10.20 -y
112
+ conda activate edgerazor
113
+ pip install -e .[cu128]
114
+ ```
115
+
116
+ ### Usage
117
+
118
+ After installation, you can integrate EdgeRazor into your existing training pipeline to build lightweight models.
119
+
120
+ 1. Use unified configuration by [yaml](./example/configs/qad/qat_w4_a8_kd_fd.yaml), [json](./example/configs/qad/qat_w4_a8_kd_fd.json) or [dict](./example/configs/qad/qat_w4_a8_kd_fd.py).
121
+
122
+ 2. Seamlessly integrate EdgeRazor into your FULL-PRECISION model training and enjoy your lightweight journey!
123
+
124
+ ```python
125
+ # Init EdgeRazor for lightweight model
126
+ edgerazor = EdgeRazor(config="/path/to/config.yaml")
127
+ student = edgerazor.quantize(student)
128
+ # Training loop
129
+ student_outputs = student(inputs)
130
+ teacher_outputs = teacher(inputs)
131
+ # Calculate loss
132
+ loss, loss_dict = edgerazor.compute_loss(student_outputs, teacher_outputs, labels)
133
+ ```
134
+
135
+ ### Low-bit LLMs Serving on Docker
136
+
137
+ You can obtain lightweight models from checkpoints trained with EdgeRazor. For example, you can convert Qwen3-EdgeRazor-4bit checkpoints to Q4_0 GGUF models. We also provide ready-to-use quantized models in our [collection](https://huggingface.co/collections/zhangsq-nju/edgerazor-nbit), including [Qwen3-0.6B-EdgeRazor-GGUF](https://huggingface.co/zhangsq-nju/Qwen3-0.6B-EdgeRazor-GGUF) and [Qwen3-1.7B-EdgeRazor-GGUF](https://huggingface.co/zhangsq-nju/Qwen3-1.7B-EdgeRazor-GGUF).
138
+
139
+ ```bash
140
+ # Serve quantized LLMs under CPU-only environments:
141
+ docker pull ghcr.io/ggml-org/llama.cpp:server
142
+ hf download zhangsq-nju/Qwen3-1.7B-EdgeRazor-GGUF Qwen3-1.7B-EdgeRazor-TQ2_0.gguf --local-dir /path/to/Qwen3-1.7B-EdgeRazor-GGUF
143
+ cd ./docker && bash local_server_tq2_0.sh
144
+ ```
145
+
146
+ ## Main Techniques
147
+
148
+ Quantization-Aware Distillation (QAD):
149
+
150
+ - Configurable mixed-precision quantization for weights
151
+ - Configurable knowledge distillation pipelines between 16-bit and $n$-bit models
152
+ - Knowledge distillation methods: Adaptive Feature Distillation (AFD), Entropy-Aware KL Divergence (EAKLD)
153
+
154
+ ## Applications
155
+
156
+ - Lightweight ViT-S/16, check [here](./example/vit/README.md).
157
+ - Lightweight ResNet-18, check [here](./example/resnet/README.md).
158
+ - Lightweight Qwen3-0.6B/1.7B, coming soon.
159
+ - Lightweight MobileLLM-ParetoQ-350M-BF16, coming soon.
160
+ - Lightweight Qwen2.5-Omni-7B, coming soon.
161
+ <!-- - Lightweight Qwen3-0.6B/1.7B, check [here](./example/qwen3/README.md).
162
+ - Lightweight MobileLLM-ParetoQ-350M-BF16, check [here](./example/mobilellm/README.md).
163
+ - Lightweight Qwen2.5-Omni-7B, check [here](./example/qwen2_5-omni/README.md). -->
164
+
165
+ ## Model Zoo
166
+
167
+ ### LLMs
168
+
169
+ - Average Performance (Avg.): average of performance scores in multiple tasks using [lm-eval v0.4.9.1](https://github.com/EleutherAI/lm-evaluation-harness/tree/v0.4.9.1) with [tasks](./src/eval/tasks/lm_eval/).
170
+ - Tasks for instruct LLMs: arc_easy, arc_challenge, hellaswag, boolq, social_iqa, openbookqa, piqa, winogrande, hendrycks_ethics, truthfulqa_mc2, mmlu, gsm8k, humaneval_instruct, ifeval.
171
+ - Tasks for base LLMs: arc_easy, arc_challenge, hellaswag, boolq, social_iqa, openbookqa, piqa, winogrande, hendrycks_ethics, truthfulqa_mc2, mmlu, gsm8k, humaneval.
172
+ - Except for 5-shot gsm8k, all other tasks are 0-shot.
173
+
174
+ - Hub Link: We provide the original quantized checkpoints. We also transfer the checkpoints into GGUF ([llama.cpp](https://github.com/ggml-org/llama.cpp)) and GPTQ ([GPTQModel](https://github.com/ModelCloud/GPTQModel), working in progress) formats if compatible.
175
+
176
+ | Model | W-A-KV | Group Size | Avg. | Hub Link |
177
+ | -------------- | ------------ | ---------- | ----- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
178
+ | Qwen3-0.6B | W16-A16-KV16 | - | 47.35 | [Base](https://huggingface.co/Qwen/Qwen3-0.6B) |
179
+ | Qwen3-0.6B | W4-A8-KV8 | 256 | 47.80 | [EdgeRazor](https://huggingface.co/zhangsq-nju/Qwen3-0.6B-EdgeRazor-4bit), [Q4_0.gguf](https://huggingface.co/zhangsq-nju/Qwen3-0.6B-EdgeRazor-GGUF/resolve/main/Qwen3-0.6B-EdgeRazor-Q4_0.gguf) |
180
+ | Qwen3-0.6B | W2.79-A8-KV8 | 256 | 44.10 | [EdgeRazor](https://huggingface.co/zhangsq-nju/Qwen3-0.6B-EdgeRazor-2.79bit) |
181
+ | Qwen3-0.6B | W1.88-A8-KV8 | 256 | 41.76 | [EdgeRazor](https://huggingface.co/zhangsq-nju/Qwen3-0.6B-EdgeRazor-1.88bit) |
182
+ | Qwen3-0.6B | W1.58-A8-KV8 | 256 | 39.81 | [EdgeRazor](https://huggingface.co/zhangsq-nju/Qwen3-0.6B-EdgeRazor-1.58bit), [TQ1_0.gguf](https://huggingface.co/zhangsq-nju/Qwen3-0.6B-EdgeRazor-GGUF/resolve/main/Qwen3-0.6B-EdgeRazor-TQ1_0.gguf), [TQ2_0.gguf](https://huggingface.co/zhangsq-nju/Qwen3-0.6B-EdgeRazor-GGUF/resolve/main/Qwen3-0.6B-EdgeRazor-TQ2_0.gguf) |
183
+ | Qwen3-1.7B | W16-A16-KV16 | - | 58.65 | [Base](https://huggingface.co/Qwen/Qwen3-1.7B) |
184
+ | Qwen3-1.7B | W4-A8-KV8 | 256 | 58.57 | [EdgeRazor](https://huggingface.co/zhangsq-nju/Qwen3-1.7B-EdgeRazor-4bit), [Q4_0.gguf](https://huggingface.co/zhangsq-nju/Qwen3-1.7B-EdgeRazor-GGUF/resolve/main/Qwen3-1.7B-EdgeRazor-Q4_0.gguf) |
185
+ | Qwen3-1.7B | W2.79-A8-KV8 | 256 | 53.00 | [EdgeRazor](https://huggingface.co/zhangsq-nju/Qwen3-1.7B-EdgeRazor-2.79bit) |
186
+ | Qwen3-1.7B | W1.88-A8-KV8 | 256 | 47.14 | [EdgeRazor](https://huggingface.co/zhangsq-nju/Qwen3-1.7B-EdgeRazor-1.88bit) |
187
+ | Qwen3-1.7B | W1.58-A8-KV8 | 256 | 43.91 | [EdgeRazor](https://huggingface.co/zhangsq-nju/Qwen3-1.7B-EdgeRazor-1.58bit), [TQ1_0.gguf](https://huggingface.co/zhangsq-nju/Qwen3-1.7B-EdgeRazor-GGUF/resolve/main/Qwen3-1.7B-EdgeRazor-TQ1_0.gguf), [TQ2_0.gguf](https://huggingface.co/zhangsq-nju/Qwen3-1.7B-EdgeRazor-GGUF/resolve/main/Qwen3-1.7B-EdgeRazor-TQ2_0.gguf) |
188
+ | MobileLLM-350M | W16-A16-KV16 | - | 41.18 | [Base](https://huggingface.co/facebook/MobileLLM-ParetoQ-350M-BF16) |
189
+ | MobileLLM-350M | W4-A8-KV8 | 64 | 41.86 | [EdgeRazor](https://huggingface.co/zhangsq-nju/MobileLLM-350M-EdgeRazor-4bit) |
190
+ | MobileLLM-350M | W2.79-A8-KV8 | 64 | 40.62 | [EdgeRazor](https://huggingface.co/zhangsq-nju/MobileLLM-350M-EdgeRazor-2.79bit) |
191
+ | MobileLLM-350M | W1.88-A8-KV8 | 64 | 39.02 | [EdgeRazor](https://huggingface.co/zhangsq-nju/MobileLLM-350M-EdgeRazor-1.88bit) |
192
+ | MobileLLM-350M | W1.58-A8-KV8 | 64 | 38.12 | [EdgeRazor](https://huggingface.co/zhangsq-nju/MobileLLM-350M-EdgeRazor-1.58bit) |
193
+
194
+ ### MLLMs
195
+
196
+ - Video-MME and MLVU are video understanding tasks using [lmms-eval v0.5.0](https://github.com/EvolvingLMMs-Lab/lmms-eval/tree/v0.5) with [tasks](./src/eval/tasks/lmms-eval/).
197
+
198
+ | Model | W-A-KV | Group Size | Video-MME | MLVU | Hub Link |
199
+ | --------------- | ------------ | ---------- | --------- | ----- | ------------------------------------------------------------------------------ |
200
+ | Qwen2.5-Omni-7B | W16-A16-KV16 | - | 62.81 | 48.01 | [Base](https://huggingface.co/Qwen/Qwen2.5-Omni-7B) |
201
+ | Qwen2.5-Omni-7B | W4-A16-KV16 | 32 | 62.22 | 48.82 | [EdgeRazor](https://huggingface.co/zhangsq-nju/Qwen2.5-Omni-7B-EdgeRazor-4bit) |
202
+
203
+ ## Todo List
204
+
205
+ EdgeRazor is continuously evolving! Here's what's coming:
206
+
207
+ - [ ] Lightweight MobileLLM, Qwen3, and Qwen2.5-Omni: training code
208
+ - [ ] Support pypi installation
209
+ - [ ] Upgrade to support the newest dependencies
210
+
211
+ Have ideas or suggestions? We welcome and appreciate any contributions and collaborations! Please feel free to submit issues or pull requests! 🚀
212
+
213
+ ## Citation
214
+
215
+ If you find our papar and code useful in your research, please consider kindly citing our papers ✏️:
216
+
217
+ ```
218
+ @article{zhangsh-edgerazor,
219
+ title={{EdgeRazor}: A Lightweight Framework for Large Language Models via Mixed-Precision Quantization-Aware Distillation},
220
+ author={Shu-Hao Zhang and Le-Tong Huang and Xiang-Sheng Deng and Xin-Yi Zou and Chen Wu and Nan Li and Shao-Qun Zhang},
221
+ year={2026},
222
+ }
223
+
224
+ @article{zhangsh-ternaryclip,
225
+ title={{TernaryCLIP}: Efficiently Compressing Vision-Language Models with Ternary Weights and Distilled Knowledge},
226
+ author={Shu-Hao Zhang and Wei-Cheng Tang and Chen Wu and Peng Hu and Nan Li and Liang-Jie Zhang and Qi Zhang and Shao-Qun Zhang},
227
+ year={2025},
228
+ journal={arXiv preprint arXiv:2510.21879}
229
+ }
230
+ ```
231
+
232
+ ## Contributor List
233
+
234
+ This project was supported by [LAMDA Lab](https://www.lamda.nju.edu.cn) and Assistant Professor [Shao-Qun Zhang](https://www.lamda.nju.edu.cn/zhangsq). [Shu-Hao Zhang](https://github.com/zhsh9) is the core developer and maintainer of EdgeRazor-V1. [Xiang-Sheng Deng](https://github.com/deng-xiangsheng) and [Le-Tong Huang](https://github.com/LT1923) jointly participated in the development of this project.
@@ -0,0 +1,198 @@
1
+ <div align="center">
2
+ <br/>
3
+ <img src="./asset/Logo-full.png" alt="EdgeRazor Logo" width="60%">
4
+ <h3>
5
+ Lightweight Framework for Edge AI
6
+ </h3>
7
+
8
+ <p>
9
+ <!-- <a href="https://arxiv.org/abs/2604.xxxxx" target="_blank">
10
+ <img src="https://img.shields.io/badge/arXiv-EdgeRazor-b31b1b?style=flat&logo=arxiv" alt="arXiv EdgeRazor">
11
+ </a> -->
12
+ <a href="https://huggingface.co/collections/zhangsq-nju/edgerazor-nbit" target="_blank">
13
+ <img src="https://img.shields.io/badge/HuggingFace-Collection-FFD21F?style=flat&logo=huggingface&logoColor=FFD21F" alt="Hugging Face Collection">
14
+ </a>
15
+ <a href="./README_ZH.md">
16
+ <img src="https://img.shields.io/badge/README-ZH-blue?style=flat&logo=readme" alt="README ZH">
17
+ </a>
18
+ <a href="https://github.com/zhangsq-nju/EdgeRazor/blob/main/LICENSE">
19
+ <img src="https://img.shields.io/badge/License-Apache_2.0-green?logo=opensourceinitiative&logoColor=green" alt="License: Apache 2.0">
20
+ </a>
21
+ </p>
22
+
23
+ <h5>
24
+ ✨ If you like our project, please give us a star ⭐️ for the latest update.
25
+ </h5>
26
+
27
+ <img src="./asset/Demo.gif" alt="EdgeRazor Demo" width="75%">
28
+ </div>
29
+
30
+ ---
31
+
32
+ **EdgeRazor** is a unified and lightweight framework for edge AI, designed to train models that are smaller, faster, and deployable across diverse hardware, ranging from mobile and edge endpoints to latency-sensitive clouds. The EdgeRazor framework **seamlessly integrates** model compression techniques into existing full-precision training pipelines with **minimal code modification**, preserving promising task performance and enabling low-cost and high-efficiency computations.
33
+
34
+ EdgeRazor currently focuses on low-bit LLM compression via configurable quantization-aware distillation. In terms of **quantization**, EdgeRazor supports quantizing weights (including embedding and lm_head layers), activations, and KV cache. Quantized bit-widths include the uniform 1.58-bit and 4-bit, as well as matrix-wise mixed-precision, such as 2.79-bit (50% 4-bit + 50% 1.58-bit) and 1.88-bit (12.5% 4-bit + 87.5% 1.58-bit). In terms of **distillation**, EdgeRazor offers the logits, features, and attention distillation, all of which can be flexibly combined within a unified configuration interface.
35
+
36
+ EdgeRazor achieves the state-of-the-art performance across a range of models, including base LLMs, instruction-tuned LLMs, and multimodal LLMs. For W-A8-KV8 quantization, **Qwen3-0.6B-EdgeRazor** attains average scores of **47.80** / **44.10** / **41.76** / **39.81** at 4-bit / 2.79-bit / 1.88-bit / 1.58-bit, corresponding to compression ratios of **3.94×** / **5.05×** / **6.40×** / **7.03×**, respectively. In comparison, the best prior methods achieve <u>45.74</u> / <u>37.38</u> / <u>30.49</u> at 4-bit / 3-bit / 2-bit with compression ratios of <u>2.21×</u> / <u>2.47×</u> / <u>2.78×</u>.
37
+
38
+ <p align="center">
39
+ <img src="./asset/Architeacture.png" alt="EdgeRazor Architecture">
40
+ <br> Figure: The EdgeRazor framework with lightweight model training pipeline.
41
+ </p>
42
+
43
+ ## News
44
+
45
+ - 🔥 **[2026-04]**: 🏆 Low-bit LLMs by EdgeRazor is released! Check our Hugging Face collection: [zhangsq-nju/edgerazor-nbit](https://huggingface.co/collections/zhangsq-nju/edgerazor-nbit).
46
+ - 🔥 **[2026-04]**: 🛠️ Open-sourced EdgeRazor-V1 is released! Now configurable on diverse models for seamless integration and customization!
47
+ - 🔥 **[2025-10]**: 📄 Paper-TernaryCLIP is available on arXiv: [2510.21879](https://arxiv.org/abs/2510.21879)!
48
+ <!-- - 🔥 **[2026-04]**: 📄 Paper-EdgeRazor is available on arXiv: [2604.xxxxx](https://arxiv.org/abs/2604.xxxxx)! -->
49
+
50
+ ## Contents
51
+
52
+ - [News](#news)
53
+ - [Contents](#contents)
54
+ - [Getting Started](#getting-started)
55
+ - [Installation](#installation)
56
+ - [Usage](#usage)
57
+ - [Low-bit LLMs Serving on Docker](#low-bit-llms-serving-on-docker)
58
+ - [Main Techniques](#main-techniques)
59
+ - [Applications](#applications)
60
+ - [Model Zoo](#model-zoo)
61
+ - [LLMs](#llms)
62
+ - [MLLMs](#mllms)
63
+ - [Todo List](#todo-list)
64
+ - [Citation](#citation)
65
+ - [Contributor List](#contributor-list)
66
+
67
+ ## Getting Started
68
+
69
+ ### Installation
70
+
71
+ - Download from GitHub (latest version)
72
+
73
+ ```
74
+ git clone https://github.com/zhangsq-nju/EdgeRazor.git && cd EdgeRazor
75
+ conda create -n edgerazor python=3.10.20 -y
76
+ conda activate edgerazor
77
+ pip install -e .[cu128]
78
+ ```
79
+
80
+ ### Usage
81
+
82
+ After installation, you can integrate EdgeRazor into your existing training pipeline to build lightweight models.
83
+
84
+ 1. Use unified configuration by [yaml](./example/configs/qad/qat_w4_a8_kd_fd.yaml), [json](./example/configs/qad/qat_w4_a8_kd_fd.json) or [dict](./example/configs/qad/qat_w4_a8_kd_fd.py).
85
+
86
+ 2. Seamlessly integrate EdgeRazor into your FULL-PRECISION model training and enjoy your lightweight journey!
87
+
88
+ ```python
89
+ # Init EdgeRazor for lightweight model
90
+ edgerazor = EdgeRazor(config="/path/to/config.yaml")
91
+ student = edgerazor.quantize(student)
92
+ # Training loop
93
+ student_outputs = student(inputs)
94
+ teacher_outputs = teacher(inputs)
95
+ # Calculate loss
96
+ loss, loss_dict = edgerazor.compute_loss(student_outputs, teacher_outputs, labels)
97
+ ```
98
+
99
+ ### Low-bit LLMs Serving on Docker
100
+
101
+ You can obtain lightweight models from checkpoints trained with EdgeRazor. For example, you can convert Qwen3-EdgeRazor-4bit checkpoints to Q4_0 GGUF models. We also provide ready-to-use quantized models in our [collection](https://huggingface.co/collections/zhangsq-nju/edgerazor-nbit), including [Qwen3-0.6B-EdgeRazor-GGUF](https://huggingface.co/zhangsq-nju/Qwen3-0.6B-EdgeRazor-GGUF) and [Qwen3-1.7B-EdgeRazor-GGUF](https://huggingface.co/zhangsq-nju/Qwen3-1.7B-EdgeRazor-GGUF).
102
+
103
+ ```bash
104
+ # Serve quantized LLMs under CPU-only environments:
105
+ docker pull ghcr.io/ggml-org/llama.cpp:server
106
+ hf download zhangsq-nju/Qwen3-1.7B-EdgeRazor-GGUF Qwen3-1.7B-EdgeRazor-TQ2_0.gguf --local-dir /path/to/Qwen3-1.7B-EdgeRazor-GGUF
107
+ cd ./docker && bash local_server_tq2_0.sh
108
+ ```
109
+
110
+ ## Main Techniques
111
+
112
+ Quantization-Aware Distillation (QAD):
113
+
114
+ - Configurable mixed-precision quantization for weights
115
+ - Configurable knowledge distillation pipelines between 16-bit and $n$-bit models
116
+ - Knowledge distillation methods: Adaptive Feature Distillation (AFD), Entropy-Aware KL Divergence (EAKLD)
117
+
118
+ ## Applications
119
+
120
+ - Lightweight ViT-S/16, check [here](./example/vit/README.md).
121
+ - Lightweight ResNet-18, check [here](./example/resnet/README.md).
122
+ - Lightweight Qwen3-0.6B/1.7B, coming soon.
123
+ - Lightweight MobileLLM-ParetoQ-350M-BF16, coming soon.
124
+ - Lightweight Qwen2.5-Omni-7B, coming soon.
125
+ <!-- - Lightweight Qwen3-0.6B/1.7B, check [here](./example/qwen3/README.md).
126
+ - Lightweight MobileLLM-ParetoQ-350M-BF16, check [here](./example/mobilellm/README.md).
127
+ - Lightweight Qwen2.5-Omni-7B, check [here](./example/qwen2_5-omni/README.md). -->
128
+
129
+ ## Model Zoo
130
+
131
+ ### LLMs
132
+
133
+ - Average Performance (Avg.): average of performance scores in multiple tasks using [lm-eval v0.4.9.1](https://github.com/EleutherAI/lm-evaluation-harness/tree/v0.4.9.1) with [tasks](./src/eval/tasks/lm_eval/).
134
+ - Tasks for instruct LLMs: arc_easy, arc_challenge, hellaswag, boolq, social_iqa, openbookqa, piqa, winogrande, hendrycks_ethics, truthfulqa_mc2, mmlu, gsm8k, humaneval_instruct, ifeval.
135
+ - Tasks for base LLMs: arc_easy, arc_challenge, hellaswag, boolq, social_iqa, openbookqa, piqa, winogrande, hendrycks_ethics, truthfulqa_mc2, mmlu, gsm8k, humaneval.
136
+ - Except for 5-shot gsm8k, all other tasks are 0-shot.
137
+
138
+ - Hub Link: We provide the original quantized checkpoints. We also transfer the checkpoints into GGUF ([llama.cpp](https://github.com/ggml-org/llama.cpp)) and GPTQ ([GPTQModel](https://github.com/ModelCloud/GPTQModel), working in progress) formats if compatible.
139
+
140
+ | Model | W-A-KV | Group Size | Avg. | Hub Link |
141
+ | -------------- | ------------ | ---------- | ----- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
142
+ | Qwen3-0.6B | W16-A16-KV16 | - | 47.35 | [Base](https://huggingface.co/Qwen/Qwen3-0.6B) |
143
+ | Qwen3-0.6B | W4-A8-KV8 | 256 | 47.80 | [EdgeRazor](https://huggingface.co/zhangsq-nju/Qwen3-0.6B-EdgeRazor-4bit), [Q4_0.gguf](https://huggingface.co/zhangsq-nju/Qwen3-0.6B-EdgeRazor-GGUF/resolve/main/Qwen3-0.6B-EdgeRazor-Q4_0.gguf) |
144
+ | Qwen3-0.6B | W2.79-A8-KV8 | 256 | 44.10 | [EdgeRazor](https://huggingface.co/zhangsq-nju/Qwen3-0.6B-EdgeRazor-2.79bit) |
145
+ | Qwen3-0.6B | W1.88-A8-KV8 | 256 | 41.76 | [EdgeRazor](https://huggingface.co/zhangsq-nju/Qwen3-0.6B-EdgeRazor-1.88bit) |
146
+ | Qwen3-0.6B | W1.58-A8-KV8 | 256 | 39.81 | [EdgeRazor](https://huggingface.co/zhangsq-nju/Qwen3-0.6B-EdgeRazor-1.58bit), [TQ1_0.gguf](https://huggingface.co/zhangsq-nju/Qwen3-0.6B-EdgeRazor-GGUF/resolve/main/Qwen3-0.6B-EdgeRazor-TQ1_0.gguf), [TQ2_0.gguf](https://huggingface.co/zhangsq-nju/Qwen3-0.6B-EdgeRazor-GGUF/resolve/main/Qwen3-0.6B-EdgeRazor-TQ2_0.gguf) |
147
+ | Qwen3-1.7B | W16-A16-KV16 | - | 58.65 | [Base](https://huggingface.co/Qwen/Qwen3-1.7B) |
148
+ | Qwen3-1.7B | W4-A8-KV8 | 256 | 58.57 | [EdgeRazor](https://huggingface.co/zhangsq-nju/Qwen3-1.7B-EdgeRazor-4bit), [Q4_0.gguf](https://huggingface.co/zhangsq-nju/Qwen3-1.7B-EdgeRazor-GGUF/resolve/main/Qwen3-1.7B-EdgeRazor-Q4_0.gguf) |
149
+ | Qwen3-1.7B | W2.79-A8-KV8 | 256 | 53.00 | [EdgeRazor](https://huggingface.co/zhangsq-nju/Qwen3-1.7B-EdgeRazor-2.79bit) |
150
+ | Qwen3-1.7B | W1.88-A8-KV8 | 256 | 47.14 | [EdgeRazor](https://huggingface.co/zhangsq-nju/Qwen3-1.7B-EdgeRazor-1.88bit) |
151
+ | Qwen3-1.7B | W1.58-A8-KV8 | 256 | 43.91 | [EdgeRazor](https://huggingface.co/zhangsq-nju/Qwen3-1.7B-EdgeRazor-1.58bit), [TQ1_0.gguf](https://huggingface.co/zhangsq-nju/Qwen3-1.7B-EdgeRazor-GGUF/resolve/main/Qwen3-1.7B-EdgeRazor-TQ1_0.gguf), [TQ2_0.gguf](https://huggingface.co/zhangsq-nju/Qwen3-1.7B-EdgeRazor-GGUF/resolve/main/Qwen3-1.7B-EdgeRazor-TQ2_0.gguf) |
152
+ | MobileLLM-350M | W16-A16-KV16 | - | 41.18 | [Base](https://huggingface.co/facebook/MobileLLM-ParetoQ-350M-BF16) |
153
+ | MobileLLM-350M | W4-A8-KV8 | 64 | 41.86 | [EdgeRazor](https://huggingface.co/zhangsq-nju/MobileLLM-350M-EdgeRazor-4bit) |
154
+ | MobileLLM-350M | W2.79-A8-KV8 | 64 | 40.62 | [EdgeRazor](https://huggingface.co/zhangsq-nju/MobileLLM-350M-EdgeRazor-2.79bit) |
155
+ | MobileLLM-350M | W1.88-A8-KV8 | 64 | 39.02 | [EdgeRazor](https://huggingface.co/zhangsq-nju/MobileLLM-350M-EdgeRazor-1.88bit) |
156
+ | MobileLLM-350M | W1.58-A8-KV8 | 64 | 38.12 | [EdgeRazor](https://huggingface.co/zhangsq-nju/MobileLLM-350M-EdgeRazor-1.58bit) |
157
+
158
+ ### MLLMs
159
+
160
+ - Video-MME and MLVU are video understanding tasks using [lmms-eval v0.5.0](https://github.com/EvolvingLMMs-Lab/lmms-eval/tree/v0.5) with [tasks](./src/eval/tasks/lmms-eval/).
161
+
162
+ | Model | W-A-KV | Group Size | Video-MME | MLVU | Hub Link |
163
+ | --------------- | ------------ | ---------- | --------- | ----- | ------------------------------------------------------------------------------ |
164
+ | Qwen2.5-Omni-7B | W16-A16-KV16 | - | 62.81 | 48.01 | [Base](https://huggingface.co/Qwen/Qwen2.5-Omni-7B) |
165
+ | Qwen2.5-Omni-7B | W4-A16-KV16 | 32 | 62.22 | 48.82 | [EdgeRazor](https://huggingface.co/zhangsq-nju/Qwen2.5-Omni-7B-EdgeRazor-4bit) |
166
+
167
+ ## Todo List
168
+
169
+ EdgeRazor is continuously evolving! Here's what's coming:
170
+
171
+ - [ ] Lightweight MobileLLM, Qwen3, and Qwen2.5-Omni: training code
172
+ - [ ] Support pypi installation
173
+ - [ ] Upgrade to support the newest dependencies
174
+
175
+ Have ideas or suggestions? We welcome and appreciate any contributions and collaborations! Please feel free to submit issues or pull requests! 🚀
176
+
177
+ ## Citation
178
+
179
+ If you find our papar and code useful in your research, please consider kindly citing our papers ✏️:
180
+
181
+ ```
182
+ @article{zhangsh-edgerazor,
183
+ title={{EdgeRazor}: A Lightweight Framework for Large Language Models via Mixed-Precision Quantization-Aware Distillation},
184
+ author={Shu-Hao Zhang and Le-Tong Huang and Xiang-Sheng Deng and Xin-Yi Zou and Chen Wu and Nan Li and Shao-Qun Zhang},
185
+ year={2026},
186
+ }
187
+
188
+ @article{zhangsh-ternaryclip,
189
+ title={{TernaryCLIP}: Efficiently Compressing Vision-Language Models with Ternary Weights and Distilled Knowledge},
190
+ author={Shu-Hao Zhang and Wei-Cheng Tang and Chen Wu and Peng Hu and Nan Li and Liang-Jie Zhang and Qi Zhang and Shao-Qun Zhang},
191
+ year={2025},
192
+ journal={arXiv preprint arXiv:2510.21879}
193
+ }
194
+ ```
195
+
196
+ ## Contributor List
197
+
198
+ This project was supported by [LAMDA Lab](https://www.lamda.nju.edu.cn) and Assistant Professor [Shao-Qun Zhang](https://www.lamda.nju.edu.cn/zhangsq). [Shu-Hao Zhang](https://github.com/zhsh9) is the core developer and maintainer of EdgeRazor-V1. [Xiang-Sheng Deng](https://github.com/deng-xiangsheng) and [Le-Tong Huang](https://github.com/LT1923) jointly participated in the development of this project.