nunchaku-lite 0.3.0__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 (89) hide show
  1. nunchaku_lite-0.3.0/LICENSE +201 -0
  2. nunchaku_lite-0.3.0/MANIFEST.in +4 -0
  3. nunchaku_lite-0.3.0/PKG-INFO +394 -0
  4. nunchaku_lite-0.3.0/README.md +363 -0
  5. nunchaku_lite-0.3.0/pyproject.toml +63 -0
  6. nunchaku_lite-0.3.0/setup.cfg +4 -0
  7. nunchaku_lite-0.3.0/setup.py +63 -0
  8. nunchaku_lite-0.3.0/src/nunchaku_lite/__init__.py +65 -0
  9. nunchaku_lite-0.3.0/src/nunchaku_lite/__version__.py +3 -0
  10. nunchaku_lite-0.3.0/src/nunchaku_lite/adapters/__init__.py +3 -0
  11. nunchaku_lite-0.3.0/src/nunchaku_lite/adapters/attention_dispatch.py +230 -0
  12. nunchaku_lite-0.3.0/src/nunchaku_lite/adapters/common.py +698 -0
  13. nunchaku_lite-0.3.0/src/nunchaku_lite/adapters/flux.py +1032 -0
  14. nunchaku_lite-0.3.0/src/nunchaku_lite/adapters/flux2.py +677 -0
  15. nunchaku_lite-0.3.0/src/nunchaku_lite/adapters/ltx2.py +574 -0
  16. nunchaku_lite-0.3.0/src/nunchaku_lite/adapters/manifest.py +381 -0
  17. nunchaku_lite-0.3.0/src/nunchaku_lite/adapters/mps_processors.py +151 -0
  18. nunchaku_lite-0.3.0/src/nunchaku_lite/adapters/qwen_image.py +685 -0
  19. nunchaku_lite-0.3.0/src/nunchaku_lite/adapters/sdxl.py +226 -0
  20. nunchaku_lite-0.3.0/src/nunchaku_lite/adapters/z_image.py +426 -0
  21. nunchaku_lite-0.3.0/src/nunchaku_lite/checkpoint_keys.py +117 -0
  22. nunchaku_lite-0.3.0/src/nunchaku_lite/core.py +830 -0
  23. nunchaku_lite-0.3.0/src/nunchaku_lite/linear.py +566 -0
  24. nunchaku_lite-0.3.0/src/nunchaku_lite/lora/__init__.py +1 -0
  25. nunchaku_lite-0.3.0/src/nunchaku_lite/lora/core/__init__.py +1 -0
  26. nunchaku_lite-0.3.0/src/nunchaku_lite/lora/core/convert.py +274 -0
  27. nunchaku_lite-0.3.0/src/nunchaku_lite/lora/core/layout.py +213 -0
  28. nunchaku_lite-0.3.0/src/nunchaku_lite/lora/core/peft.py +78 -0
  29. nunchaku_lite-0.3.0/src/nunchaku_lite/lora/core/runtime.py +803 -0
  30. nunchaku_lite-0.3.0/src/nunchaku_lite/lora/flux.py +285 -0
  31. nunchaku_lite-0.3.0/src/nunchaku_lite/lora/flux2.py +263 -0
  32. nunchaku_lite-0.3.0/src/nunchaku_lite/lora/manifest.py +433 -0
  33. nunchaku_lite-0.3.0/src/nunchaku_lite/lora/qwen_image.py +128 -0
  34. nunchaku_lite-0.3.0/src/nunchaku_lite/lora/z_image.py +154 -0
  35. nunchaku_lite-0.3.0/src/nunchaku_lite/manifest.py +301 -0
  36. nunchaku_lite-0.3.0/src/nunchaku_lite/mlx/__init__.py +34 -0
  37. nunchaku_lite-0.3.0/src/nunchaku_lite/mlx/fused_attention.py +124 -0
  38. nunchaku_lite-0.3.0/src/nunchaku_lite/mlx/kernels.py +497 -0
  39. nunchaku_lite-0.3.0/src/nunchaku_lite/mlx/keymap.py +85 -0
  40. nunchaku_lite-0.3.0/src/nunchaku_lite/mlx/linear.py +443 -0
  41. nunchaku_lite-0.3.0/src/nunchaku_lite/mlx/manifest.py +267 -0
  42. nunchaku_lite-0.3.0/src/nunchaku_lite/mlx/mx_layout.py +106 -0
  43. nunchaku_lite-0.3.0/src/nunchaku_lite/mlx/packing.py +277 -0
  44. nunchaku_lite-0.3.0/src/nunchaku_lite/mlx/patch.py +308 -0
  45. nunchaku_lite-0.3.0/src/nunchaku_lite/mlx/pipeline.py +959 -0
  46. nunchaku_lite-0.3.0/src/nunchaku_lite/ops/__init__.py +15 -0
  47. nunchaku_lite-0.3.0/src/nunchaku_lite/ops/_mlx_steel_attention.py +3503 -0
  48. nunchaku_lite-0.3.0/src/nunchaku_lite/ops/attention.py +22 -0
  49. nunchaku_lite-0.3.0/src/nunchaku_lite/ops/backend.py +51 -0
  50. nunchaku_lite-0.3.0/src/nunchaku_lite/ops/fused.py +432 -0
  51. nunchaku_lite-0.3.0/src/nunchaku_lite/ops/gemm.py +167 -0
  52. nunchaku_lite-0.3.0/src/nunchaku_lite/ops/gemv.py +34 -0
  53. nunchaku_lite-0.3.0/src/nunchaku_lite/ops/mps.py +1114 -0
  54. nunchaku_lite-0.3.0/src/nunchaku_lite/ops/quantize.py +59 -0
  55. nunchaku_lite-0.3.0/src/nunchaku_lite/shaders.py +157 -0
  56. nunchaku_lite-0.3.0/src/nunchaku_lite/text_encoders/__init__.py +5 -0
  57. nunchaku_lite-0.3.0/src/nunchaku_lite/text_encoders/t5.py +62 -0
  58. nunchaku_lite-0.3.0/src/nunchaku_lite/utils.py +231 -0
  59. nunchaku_lite-0.3.0/src/nunchaku_lite.egg-info/PKG-INFO +394 -0
  60. nunchaku_lite-0.3.0/src/nunchaku_lite.egg-info/SOURCES.txt +87 -0
  61. nunchaku_lite-0.3.0/src/nunchaku_lite.egg-info/dependency_links.txt +1 -0
  62. nunchaku_lite-0.3.0/src/nunchaku_lite.egg-info/requires.txt +15 -0
  63. nunchaku_lite-0.3.0/src/nunchaku_lite.egg-info/top_level.txt +1 -0
  64. nunchaku_lite-0.3.0/tests/test_adapter_common.py +48 -0
  65. nunchaku_lite-0.3.0/tests/test_attention_dispatch.py +61 -0
  66. nunchaku_lite-0.3.0/tests/test_awq_gemm.py +84 -0
  67. nunchaku_lite-0.3.0/tests/test_checkpoint_metadata_fallback.py +53 -0
  68. nunchaku_lite-0.3.0/tests/test_core.py +420 -0
  69. nunchaku_lite-0.3.0/tests/test_flux2_adapter.py +166 -0
  70. nunchaku_lite-0.3.0/tests/test_flux2_lora.py +333 -0
  71. nunchaku_lite-0.3.0/tests/test_flux_adapter.py +330 -0
  72. nunchaku_lite-0.3.0/tests/test_flux_lora.py +360 -0
  73. nunchaku_lite-0.3.0/tests/test_full_inference_flux.py +135 -0
  74. nunchaku_lite-0.3.0/tests/test_full_inference_flux2.py +151 -0
  75. nunchaku_lite-0.3.0/tests/test_full_inference_qwen_image.py +111 -0
  76. nunchaku_lite-0.3.0/tests/test_full_inference_z_image.py +313 -0
  77. nunchaku_lite-0.3.0/tests/test_ltx2_adapter.py +250 -0
  78. nunchaku_lite-0.3.0/tests/test_ltx2_fused_ops.py +216 -0
  79. nunchaku_lite-0.3.0/tests/test_manifest_adapter.py +705 -0
  80. nunchaku_lite-0.3.0/tests/test_mps_backend.py +484 -0
  81. nunchaku_lite-0.3.0/tests/test_mps_fused_attention.py +412 -0
  82. nunchaku_lite-0.3.0/tests/test_ops_backend.py +69 -0
  83. nunchaku_lite-0.3.0/tests/test_qwen_image_adapter.py +103 -0
  84. nunchaku_lite-0.3.0/tests/test_qwen_image_lora.py +160 -0
  85. nunchaku_lite-0.3.0/tests/test_sdxl_adapter.py +109 -0
  86. nunchaku_lite-0.3.0/tests/test_t5_encoder.py +78 -0
  87. nunchaku_lite-0.3.0/tests/test_torch_free_imports.py +47 -0
  88. nunchaku_lite-0.3.0/tests/test_z_image_adapter.py +272 -0
  89. nunchaku_lite-0.3.0/tests/test_z_image_lora.py +565 -0
@@ -0,0 +1,201 @@
1
+ Apache License
2
+ Version 2.0, January 2004
3
+ https://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 [yyyy] [name of copyright owner]
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
+ https://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 +1,4 @@
1
+ recursive-include src/nunchaku_lite *.py *.json
2
+
3
+ include setup.py
4
+ include README.md
@@ -0,0 +1,394 @@
1
+ Metadata-Version: 2.4
2
+ Name: nunchaku_lite
3
+ Version: 0.3.0
4
+ Summary: Lite plugin runtime for applying Nunchaku v2 quantized transformer weights to Diffusers (PyTorch) and mflux (MLX) pipelines.
5
+ License-Expression: Apache-2.0
6
+ Project-URL: Homepage, https://github.com/rootonchair/nunchaku-lite
7
+ Project-URL: Documentation, https://nunchaku-lite.readthedocs.io/en/latest/
8
+ Project-URL: Issues, https://github.com/rootonchair/nunchaku-lite/issues
9
+ Classifier: Programming Language :: Python :: 3
10
+ Classifier: Programming Language :: Python :: 3.10
11
+ Classifier: Programming Language :: Python :: 3.11
12
+ Classifier: Programming Language :: Python :: 3.12
13
+ Classifier: Programming Language :: Python :: 3.13
14
+ Requires-Python: >=3.10
15
+ Description-Content-Type: text/markdown
16
+ License-File: LICENSE
17
+ Requires-Dist: accelerate>=0.31
18
+ Requires-Dist: diffusers>=0.36
19
+ Requires-Dist: huggingface-hub>=0.34
20
+ Requires-Dist: kernels<0.15,>=0.14
21
+ Requires-Dist: packaging>=23
22
+ Requires-Dist: peft
23
+ Requires-Dist: safetensors
24
+ Requires-Dist: torch>=2.7
25
+ Requires-Dist: transformers>=4.41.2
26
+ Provides-Extra: torch
27
+ Provides-Extra: mlx
28
+ Requires-Dist: mlx>=0.30; extra == "mlx"
29
+ Requires-Dist: numpy; extra == "mlx"
30
+ Dynamic: license-file
31
+
32
+ <p align="center">
33
+ <img src="assets/logo.svg" alt="nunchaku_lite" width="640">
34
+ </p>
35
+
36
+ <p align="center">
37
+ <a href="https://github.com/rootonchair/nunchaku-lite/actions/workflows/cpu-tests.yml">
38
+ <img src="https://github.com/rootonchair/nunchaku-lite/actions/workflows/cpu-tests.yml/badge.svg" alt="CPU Tests">
39
+ </a>
40
+ <a href="https://github.com/rootonchair/nunchaku-lite/actions/workflows/wheels.yml">
41
+ <img src="https://github.com/rootonchair/nunchaku-lite/actions/workflows/wheels.yml/badge.svg" alt="Release Wheels">
42
+ </a>
43
+ <a href="https://nunchaku-lite.readthedocs.io/en/latest/">
44
+ <img src="https://img.shields.io/badge/docs-readthedocs.io-blue" alt="Documentation">
45
+ </a>
46
+ <a href="LICENSE">
47
+ <img src="https://img.shields.io/badge/License-Apache%202.0-blue.svg" alt="License: Apache 2.0">
48
+ </a>
49
+ </p>
50
+
51
+ ## About
52
+
53
+ `nunchaku_lite` brings Nunchaku-quantized image generation models to standard
54
+ Diffusers pipelines with a lean runtime, native CUDA kernels, and drop-in
55
+ pipeline loading. It is designed for fast SVDQ W4A4 inference with a small
56
+ integration surface: load a normal Diffusers pipeline, replace only the
57
+ quantized component, and keep the rest of the workflow
58
+ unchanged.
59
+
60
+ Its core features include:
61
+
62
+ - Efficient pipeline loading: constructs the patched component up front so
63
+ Diffusers does not first materialize dense BF16 transformer or UNet weights
64
+ that are immediately replaced.
65
+ - Native quantized kernels: packages the CUDA kernels and Python wrappers needed
66
+ for INT4 and FP4 Nunchaku checkpoints.
67
+ - AWQ GEMM acceleration: unlike original Nunchaku, `nunchaku_lite` ships a
68
+ native W4A16 group-size-64/int32 GEMM path for large AWQ projections, turning
69
+ chunked GEMV bottlenecks into high-throughput CUDA matmuls.
70
+ - Broad image-model coverage: includes built-in adapters for FLUX.1, FLUX.2
71
+ Klein, Qwen-Image, Qwen-Image-Edit, SDXL, SDXL-Turbo, and Z-Image Turbo.
72
+ - Runtime LoRA support: exposes Diffusers-style LoRA loading, adapter strength
73
+ control, multi-LoRA composition, and unload/reset flows for supported model
74
+ families.
75
+ - Extensible adapter registry: keeps model-specific graph rewrites isolated in
76
+ small adapters without requiring pipeline subclasses or a dependency on the
77
+ full `nunchaku` Python package.
78
+
79
+ ## Benchmarks
80
+
81
+ <p align="center">
82
+ <img src="docs/assets/benchmarks/readme-benchmark-summary.png" alt="nunchaku_lite benchmark gains versus original Diffusers" width="900">
83
+ </p>
84
+
85
+ On an NVIDIA RTX PRO 6000, `nunchaku_lite` delivers up
86
+ to `1.79x` lower latency, up to `46%` lower peak CUDA memory, and up to `71%`
87
+ smaller transformer storage versus unmodified Diffusers pipelines in the current
88
+ benchmark set.
89
+
90
+ See [docs/benchmarks.md](docs/benchmarks.md) for charts, generated samples,
91
+ run settings, and reproduction commands.
92
+
93
+ ## Supported Models
94
+
95
+ `nunchaku_lite` currently supports Diffusers pipelines whose transformer or UNet
96
+ component matches one of the built-in adapters below. Checkpoints should be SVDQ
97
+ W4A4 Nunchaku v2 `.safetensors` files in INT4 or FP4 format. Full runnable
98
+ model guides live under `docs/models/`.
99
+
100
+ | Model family | Coverage | Runtime LoRA | Guide | Features |
101
+ | --- | --- | --- | --- | --- |
102
+ | FLUX.1 | FLUX.1-schnell and FLUX.1-dev | Yes | [docs/models/flux.md](docs/models/flux.md) | Pipeline loading, INT4/FP4 checkpoints, Diffusers-format LoRA loading, strength control, multi-LoRA composition, reset, and unload. |
103
+ | FLUX.2 Klein | FLUX.2 Klein | Yes | [docs/models/flux2.md](docs/models/flux2.md) | Pipeline loading, INT4/FP4 checkpoints, runtime LoRA, and ComfyUI Flux2 LoRA key conversion. |
104
+ | Qwen-Image and Qwen-Image-Edit | Qwen-Image, Qwen-Image-Lightning, Qwen-Image-Edit, and Qwen-Image-Edit-2509 | Yes | [docs/models/qwen_image.md](docs/models/qwen_image.md) | Pipeline loading, INT4/FP4 checkpoints, low-VRAM examples, Lightning LoRA workflows, and edit-pipeline examples. |
105
+ | SDXL and SDXL-Turbo | SDXL and SDXL-Turbo | Not yet | [docs/models/sdxl.md](docs/models/sdxl.md) | Pipeline loading for quantized UNet checkpoints. |
106
+ | Z-Image Turbo | Z-Image Turbo | Yes | [docs/models/z_image.md](docs/models/z_image.md) | Pipeline loading, INT4/FP4 checkpoints, runtime LoRA, and dense AdaLN modulation LoRA branches. |
107
+
108
+ The Qwen low-VRAM guides use `enable_model_cpu_offload()`, which requires
109
+ `accelerate`.
110
+
111
+ Additional model families should be added through the adapter registry rather
112
+ than pipeline subclasses. See [docs/roadmap.md](docs/roadmap.md) for planned
113
+ coverage and remaining feature work.
114
+
115
+ ## Requirements
116
+
117
+ - Python 3.10 or newer
118
+ - PyTorch 2.7 or newer with CUDA
119
+ - Diffusers 0.36 or newer
120
+
121
+ Native source builds also require CUDA toolkit 12.6 or newer with `nvcc`. The
122
+ default runtime path uses prebuilt kernels from `rootonchair/nunchaku-lite-kernels`
123
+ through the Hugging Face `kernels` library instead.
124
+
125
+ The native build detects the local GPU architecture by default. Supported
126
+ targets are `sm75`, `sm80`, `sm86`, `sm89`, `sm120a`, and `sm121a`, subject to
127
+ the installed CUDA toolkit version. CUDA 12.6 or newer is the documented
128
+ minimum; `sm120a` requires CUDA 12.8 or newer, and `sm121a` requires CUDA 13.0
129
+ or newer. The Hugging Face kernels package provides prebuilt kernels for CUDA
130
+ `7.5`, `8.0`, `8.6`, `8.9`, `12.0a`, and `12.1a`.
131
+
132
+ The root Python package does not compile CUDA code. CUDA toolkit requirements
133
+ apply only when installing the local `nunchaku_lite_kernels` package from
134
+ `./nunchaku-lite-kernels`.
135
+
136
+ > **CUDA version note:** `nunchaku_lite` first tries a locally installed
137
+ > `nunchaku_lite_kernels` package, then falls back to
138
+ > `rootonchair/nunchaku-lite-kernels` through Hugging Face `kernels`.
139
+
140
+ ### Apple silicon (MPS, experimental)
141
+
142
+ INT4 SVDQ checkpoints also run on Apple-silicon GPUs with `device="mps"`.
143
+ Apple GPUs have no INT4 matmul, so the MPS path runs SVDQ layers as **W8A8**:
144
+ the 4-bit weights (plus folded group scales) become int8, activations are
145
+ quantized to int8 per token after smoothing, and the matmul runs on the M5
146
+ Neural Accelerators with the SVD low-rank branch, scales and bias fused into
147
+ the kernel epilogue. Q/K RMSNorm and RoPE run in a fused Metal kernel.
148
+
149
+ - Requires macOS 26+ and a PyTorch build with MPS. With PyTorch 2.10 to 2.15 the
150
+ Metal kernels download from the Hugging Face Hub on first use (the `kernels`
151
+ dependency of `nunchaku-lite`). For other PyTorch versions, build them
152
+ locally: `pip install --no-build-isolation ./nunchaku-lite-kernels`
153
+ - Choose the quantized-kernel backend with `backend=` on `load_nunchaku_pipeline`
154
+ or `patch_transformer` (the same keyword Diffusers uses for kernel selection in
155
+ `AutoRoundConfig`; unrelated to `set_attention_backend`). CUDA has a single
156
+ kernel and accepts only `"auto"`.
157
+ - `"auto"` (default): `"int8"` when the transformer fits in half of the GPU
158
+ working set with int8 weights, otherwise `"int4"`; `"dense"` on Apple GPUs
159
+ without Neural Accelerators (pre-M5).
160
+ - `"int8"`: fastest. `"int4"`: weights stay 4-bit in memory, a few percent
161
+ slower, identical output. `"dense"`: exact dequantized fallback (more
162
+ memory, no speedup).
163
+
164
+ ```python
165
+ pipe = load_nunchaku_pipeline(..., device="mps", backend="int4")
166
+ ```
167
+ - INT4 checkpoints only (no NVFP4). Verified end to end with FLUX.2 Klein 4B and
168
+ 9B, FLUX.1-dev (Nunchaku and Nunchaku Lite checkpoints), Z-Image Turbo, and
169
+ Qwen-Image Lightning (which needs more than 24 GB to run without swapping).
170
+
171
+ On an M5 MacBook Pro (FLUX.2 Klein 4B, 4 steps, denoising only):
172
+
173
+ | Transformer | Weights | Live GPU memory (1024px) | Step 512px | Step 1024px |
174
+ | --- | --- | --- | --- | --- |
175
+ | Dense BF16 | 7.75 GB | 8.1 GB | 1.20 s | 4.96 s |
176
+ | INT4, `int8` backend (default) | 4.57 GB | 5.6 GB | 0.62 s (1.94x) | 2.27 s (2.19x) |
177
+ | INT4, `int4` backend | 2.67 GB | 4.4 GB | 0.70 s (1.71x) | 2.50 s (1.98x) |
178
+
179
+ The packed checkpoint `qweight` is released after conversion whenever the
180
+ kernel copy reproduces it exactly (94 of 100 layers for `int8`, all layers for
181
+ `int4`; AWQ modulation layers too); `state_dict()` re-packs it and
182
+ `load_state_dict()` restores it.
183
+ On M5-class GPUs the patched transformer's attention also runs on the Neural Accelerators (MLX's flash
184
+ attention kernel) instead of PyTorch's MPS attention, and the q/k norm + RoPE of FLUX.1, FLUX.2, Qwen-Image and
185
+ Z-Image is one fused kernel. Both come with the Hub kernels; with an older Hub build they are compiled at runtime
186
+ instead. The dense BF16 row uses stock Diffusers.
187
+
188
+ There is also an MLX runtime, `nunchaku_lite.mlx` (`pip install "nunchaku-lite[mlx]" mflux`),
189
+ which loads the same checkpoints into [mflux](https://github.com/filipstrand/mflux)
190
+ pipelines with the same `load_nunchaku_pipeline` interface. See
191
+ [docs/mlx.md](docs/mlx.md).
192
+
193
+ ## Installation
194
+
195
+ ```bash
196
+ pip install nunchaku-lite # PyTorch runtime (CUDA, Apple MPS) for Diffusers pipelines
197
+ pip install "nunchaku-lite[mlx]" mflux # adds the MLX runtime (Apple silicon) for mflux pipelines
198
+ ```
199
+
200
+ From a source checkout, `pip install .` and `pip install ".[mlx]"` do the same.
201
+
202
+ This installs the root Python runtime and uses Hugging Face prebuilt kernels by
203
+ default. `nunchaku_lite.mlx` does not import torch, even though the base install
204
+ includes it.
205
+
206
+ To build and use the local CUDA kernels package:
207
+
208
+ ```bash
209
+ pip install --no-build-isolation ./nunchaku-lite-kernels
210
+ ```
211
+
212
+ `--no-build-isolation` ensures the extension is compiled against the PyTorch
213
+ version installed in the active environment.
214
+
215
+ When `nunchaku_lite_kernels` is installed locally, `nunchaku_lite` prefers it
216
+ over the Hugging Face fallback.
217
+
218
+ Build and install a wheel:
219
+
220
+ ```bash
221
+ python -m build --wheel
222
+ pip install dist/nunchaku_lite-*.whl
223
+ ```
224
+
225
+ By default, the local kernels package build uses `NUNCHAKU_INSTALL_MODE=FAST`
226
+ and compiles for visible local CUDA devices. To build all supported
227
+ architectures:
228
+
229
+ ```bash
230
+ NUNCHAKU_INSTALL_MODE=ALL pip install --no-build-isolation ./nunchaku-lite-kernels
231
+ ```
232
+
233
+ ## Quick Start
234
+
235
+ Run from the repository root:
236
+
237
+ ```python
238
+ from pathlib import Path
239
+
240
+ import torch
241
+ from diffusers import FluxPipeline
242
+
243
+ from nunchaku_lite import load_nunchaku_pipeline
244
+
245
+
246
+ model_id = "black-forest-labs/FLUX.1-schnell"
247
+ precision = "fp4" # "int4" or "fp4"
248
+ checkpoints = {
249
+ "int4": "nunchaku-ai/nunchaku-flux.1-schnell/svdq-int4_r32-flux.1-schnell.safetensors",
250
+ "fp4": "nunchaku-ai/nunchaku-flux.1-schnell/svdq-fp4_r32-flux.1-schnell.safetensors",
251
+ }
252
+ output_path = Path(f"outputs/flux_schnell_nunchaku_lite_{precision}.png")
253
+
254
+ pipe = load_nunchaku_pipeline(
255
+ model_id,
256
+ pipeline_cls=FluxPipeline,
257
+ checkpoint=checkpoints[precision],
258
+ target="flux",
259
+ precision=precision,
260
+ torch_dtype=torch.bfloat16,
261
+ device="cuda",
262
+ )
263
+ pipe = pipe.to("cuda")
264
+
265
+ image = pipe(
266
+ "A cat holding a sign that says hello world",
267
+ height=1024,
268
+ width=1024,
269
+ num_inference_steps=4,
270
+ guidance_scale=0.0,
271
+ generator=torch.Generator(device="cuda").manual_seed(12345),
272
+ ).images[0]
273
+
274
+ output_path.parent.mkdir(parents=True, exist_ok=True)
275
+ image.save(output_path)
276
+ print(f"saved {output_path}")
277
+ ```
278
+
279
+ ### Runtime LoRA
280
+
281
+ Pipelines loaded with `load_nunchaku_pipeline` expose Diffusers-style LoRA
282
+ methods when the selected adapter supports runtime LoRA:
283
+
284
+ ```python
285
+ from pathlib import Path
286
+
287
+ import torch
288
+ from diffusers import FluxPipeline
289
+
290
+ from nunchaku_lite import load_nunchaku_pipeline
291
+
292
+
293
+ model_id = "black-forest-labs/FLUX.1-dev"
294
+ precision = "fp4"
295
+ checkpoint = "nunchaku-tech/nunchaku-flux.1-dev/svdq-fp4_r32-flux.1-dev.safetensors"
296
+ output_path = Path("outputs/flux_dev_ghibsky_lora_fp4.png")
297
+
298
+ pipe = load_nunchaku_pipeline(
299
+ model_id,
300
+ pipeline_cls=FluxPipeline,
301
+ checkpoint=checkpoint,
302
+ target="flux",
303
+ precision=precision,
304
+ torch_dtype=torch.bfloat16,
305
+ )
306
+ pipe.enable_model_cpu_offload()
307
+
308
+ pipe.load_lora_weights(
309
+ "aleksa-codes/flux-ghibsky-illustration",
310
+ weight_name="lora.safetensors",
311
+ adapter_name="ghibsky",
312
+ )
313
+ pipe.set_adapters("ghibsky", adapter_weights=0.75)
314
+
315
+ image = pipe(
316
+ "GHIBSKY style painting of a cozy mountain cabin beside a clear lake at sunset",
317
+ height=1024,
318
+ width=1024,
319
+ num_inference_steps=28,
320
+ guidance_scale=3.5,
321
+ generator=torch.Generator(device="cpu").manual_seed(12345),
322
+ ).images[0]
323
+
324
+ output_path.parent.mkdir(parents=True, exist_ok=True)
325
+ image.save(output_path)
326
+ print(f"saved {output_path}")
327
+ ```
328
+
329
+ Checkpoint paths can be local `.safetensors` files or Hugging Face paths of the
330
+ form:
331
+
332
+ ```text
333
+ org-or-user/repo-name/path/to/checkpoint.safetensors
334
+ ```
335
+
336
+ ## Documentation
337
+
338
+ | Topic | Link |
339
+ | --- | --- |
340
+ | Hosted documentation | [nunchaku-lite.readthedocs.io](https://nunchaku-lite.readthedocs.io/en/latest/) |
341
+ | Benchmark charts and samples | [docs/benchmarks.md](docs/benchmarks.md) |
342
+ | Public API and runtime LoRA usage | [docs/api.md](docs/api.md) |
343
+ | Development, testing, and adapter authoring | [docs/development.md](docs/development.md) |
344
+ | Documentation deployment flow | [docs/deployment.md](docs/deployment.md) |
345
+ | Release wheel build flow | [docs/release_wheels.md](docs/release_wheels.md) |
346
+ | Supported models and feature backlog | [docs/roadmap.md](docs/roadmap.md) |
347
+ | Benchmarks | [benchmarks/README.md](benchmarks/README.md) |
348
+
349
+ ## Public API
350
+
351
+ ```python
352
+ from nunchaku_lite import (
353
+ TransformerAdapter,
354
+ list_adapters,
355
+ load_nunchaku_pipeline,
356
+ patch_transformer,
357
+ register_adapter,
358
+ )
359
+ ```
360
+
361
+ `load_nunchaku_pipeline(...)` is the preferred entry point for normal pipeline
362
+ loading. Use `patch_transformer(...)` only when a Diffusers component has already
363
+ been constructed and needs in-place patching.
364
+
365
+ See [docs/api.md](docs/api.md) for argument details, runtime LoRA examples, and
366
+ adapter registry usage.
367
+
368
+ ## Development
369
+
370
+ Run the unit tests:
371
+
372
+ ```bash
373
+ pytest -q tests
374
+ ```
375
+
376
+ Build the extension in place:
377
+
378
+ ```bash
379
+ python setup.py build_ext --inplace
380
+ ```
381
+
382
+ See [docs/development.md](docs/development.md) for full inference tests, adapter
383
+ authoring guidance, runtime LoRA implementation notes, and repository layout.
384
+
385
+ ## License
386
+
387
+ `nunchaku_lite` is licensed under the Apache License, Version 2.0. See
388
+ [LICENSE](LICENSE).
389
+
390
+ ## Acknowledgements
391
+
392
+ `nunchaku_lite` builds on the Nunchaku project and uses selected native kernel
393
+ code for the lite runtime. We are grateful to the maintainers and contributors
394
+ of these projects.