xinference 1.7.1.post1__py3-none-any.whl → 1.8.0__py3-none-any.whl

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.

Potentially problematic release.

This version of xinference might be problematic. Click here for more details.

Files changed (136) hide show
  1. xinference/_version.py +3 -3
  2. xinference/client/restful/async_restful_client.py +8 -13
  3. xinference/client/restful/restful_client.py +6 -2
  4. xinference/core/chat_interface.py +6 -4
  5. xinference/core/media_interface.py +5 -0
  6. xinference/core/model.py +1 -5
  7. xinference/core/supervisor.py +117 -68
  8. xinference/core/worker.py +49 -37
  9. xinference/deploy/test/test_cmdline.py +2 -6
  10. xinference/model/audio/__init__.py +26 -23
  11. xinference/model/audio/chattts.py +3 -2
  12. xinference/model/audio/core.py +49 -98
  13. xinference/model/audio/cosyvoice.py +3 -2
  14. xinference/model/audio/custom.py +28 -73
  15. xinference/model/audio/f5tts.py +3 -2
  16. xinference/model/audio/f5tts_mlx.py +3 -2
  17. xinference/model/audio/fish_speech.py +3 -2
  18. xinference/model/audio/funasr.py +17 -4
  19. xinference/model/audio/kokoro.py +3 -2
  20. xinference/model/audio/megatts.py +3 -2
  21. xinference/model/audio/melotts.py +3 -2
  22. xinference/model/audio/model_spec.json +572 -171
  23. xinference/model/audio/utils.py +0 -6
  24. xinference/model/audio/whisper.py +3 -2
  25. xinference/model/audio/whisper_mlx.py +3 -2
  26. xinference/model/cache_manager.py +141 -0
  27. xinference/model/core.py +6 -49
  28. xinference/model/custom.py +174 -0
  29. xinference/model/embedding/__init__.py +67 -56
  30. xinference/model/embedding/cache_manager.py +35 -0
  31. xinference/model/embedding/core.py +104 -84
  32. xinference/model/embedding/custom.py +55 -78
  33. xinference/model/embedding/embed_family.py +80 -31
  34. xinference/model/embedding/flag/core.py +21 -5
  35. xinference/model/embedding/llama_cpp/__init__.py +0 -0
  36. xinference/model/embedding/llama_cpp/core.py +234 -0
  37. xinference/model/embedding/model_spec.json +968 -103
  38. xinference/model/embedding/sentence_transformers/core.py +30 -20
  39. xinference/model/embedding/vllm/core.py +11 -5
  40. xinference/model/flexible/__init__.py +8 -2
  41. xinference/model/flexible/core.py +26 -119
  42. xinference/model/flexible/custom.py +69 -0
  43. xinference/model/flexible/launchers/image_process_launcher.py +1 -0
  44. xinference/model/flexible/launchers/modelscope_launcher.py +5 -1
  45. xinference/model/flexible/launchers/transformers_launcher.py +15 -3
  46. xinference/model/flexible/launchers/yolo_launcher.py +5 -1
  47. xinference/model/image/__init__.py +20 -20
  48. xinference/model/image/cache_manager.py +62 -0
  49. xinference/model/image/core.py +70 -182
  50. xinference/model/image/custom.py +28 -72
  51. xinference/model/image/model_spec.json +402 -119
  52. xinference/model/image/ocr/got_ocr2.py +3 -2
  53. xinference/model/image/stable_diffusion/core.py +22 -7
  54. xinference/model/image/stable_diffusion/mlx.py +6 -6
  55. xinference/model/image/utils.py +2 -2
  56. xinference/model/llm/__init__.py +71 -94
  57. xinference/model/llm/cache_manager.py +292 -0
  58. xinference/model/llm/core.py +37 -111
  59. xinference/model/llm/custom.py +88 -0
  60. xinference/model/llm/llama_cpp/core.py +5 -7
  61. xinference/model/llm/llm_family.json +16260 -8151
  62. xinference/model/llm/llm_family.py +138 -839
  63. xinference/model/llm/lmdeploy/core.py +5 -7
  64. xinference/model/llm/memory.py +3 -4
  65. xinference/model/llm/mlx/core.py +6 -8
  66. xinference/model/llm/reasoning_parser.py +3 -1
  67. xinference/model/llm/sglang/core.py +32 -14
  68. xinference/model/llm/transformers/chatglm.py +3 -7
  69. xinference/model/llm/transformers/core.py +49 -27
  70. xinference/model/llm/transformers/deepseek_v2.py +2 -2
  71. xinference/model/llm/transformers/gemma3.py +2 -2
  72. xinference/model/llm/transformers/multimodal/cogagent.py +2 -2
  73. xinference/model/llm/transformers/multimodal/deepseek_vl2.py +2 -2
  74. xinference/model/llm/transformers/multimodal/gemma3.py +2 -2
  75. xinference/model/llm/transformers/multimodal/glm4_1v.py +167 -0
  76. xinference/model/llm/transformers/multimodal/glm4v.py +2 -2
  77. xinference/model/llm/transformers/multimodal/intern_vl.py +2 -2
  78. xinference/model/llm/transformers/multimodal/minicpmv26.py +3 -3
  79. xinference/model/llm/transformers/multimodal/ovis2.py +2 -2
  80. xinference/model/llm/transformers/multimodal/qwen-omni.py +2 -2
  81. xinference/model/llm/transformers/multimodal/qwen2_audio.py +2 -2
  82. xinference/model/llm/transformers/multimodal/qwen2_vl.py +2 -2
  83. xinference/model/llm/transformers/opt.py +3 -7
  84. xinference/model/llm/utils.py +34 -49
  85. xinference/model/llm/vllm/core.py +77 -27
  86. xinference/model/llm/vllm/xavier/engine.py +5 -3
  87. xinference/model/llm/vllm/xavier/scheduler.py +10 -6
  88. xinference/model/llm/vllm/xavier/transfer.py +1 -1
  89. xinference/model/rerank/__init__.py +26 -25
  90. xinference/model/rerank/core.py +47 -87
  91. xinference/model/rerank/custom.py +25 -71
  92. xinference/model/rerank/model_spec.json +158 -33
  93. xinference/model/rerank/utils.py +2 -2
  94. xinference/model/utils.py +115 -54
  95. xinference/model/video/__init__.py +13 -17
  96. xinference/model/video/core.py +44 -102
  97. xinference/model/video/diffusers.py +4 -3
  98. xinference/model/video/model_spec.json +90 -21
  99. xinference/types.py +5 -3
  100. xinference/web/ui/build/asset-manifest.json +3 -3
  101. xinference/web/ui/build/index.html +1 -1
  102. xinference/web/ui/build/static/js/main.7d24df53.js +3 -0
  103. xinference/web/ui/build/static/js/main.7d24df53.js.map +1 -0
  104. xinference/web/ui/node_modules/.cache/babel-loader/2704ff66a5f73ca78b341eb3edec60154369df9d87fbc8c6dd60121abc5e1b0a.json +1 -0
  105. xinference/web/ui/node_modules/.cache/babel-loader/607dfef23d33e6b594518c0c6434567639f24f356b877c80c60575184ec50ed0.json +1 -0
  106. xinference/web/ui/node_modules/.cache/babel-loader/9be3d56173aacc3efd0b497bcb13c4f6365de30069176ee9403b40e717542326.json +1 -0
  107. xinference/web/ui/node_modules/.cache/babel-loader/9f9dd6c32c78a222d07da5987ae902effe16bcf20aac00774acdccc4de3c9ff2.json +1 -0
  108. xinference/web/ui/node_modules/.cache/babel-loader/b2ab5ee972c60d15eb9abf5845705f8ab7e1d125d324d9a9b1bcae5d6fd7ffb2.json +1 -0
  109. xinference/web/ui/src/locales/en.json +0 -1
  110. xinference/web/ui/src/locales/ja.json +0 -1
  111. xinference/web/ui/src/locales/ko.json +0 -1
  112. xinference/web/ui/src/locales/zh.json +0 -1
  113. {xinference-1.7.1.post1.dist-info → xinference-1.8.0.dist-info}/METADATA +9 -11
  114. {xinference-1.7.1.post1.dist-info → xinference-1.8.0.dist-info}/RECORD +119 -119
  115. xinference/model/audio/model_spec_modelscope.json +0 -231
  116. xinference/model/embedding/model_spec_modelscope.json +0 -293
  117. xinference/model/embedding/utils.py +0 -18
  118. xinference/model/image/model_spec_modelscope.json +0 -375
  119. xinference/model/llm/llama_cpp/memory.py +0 -457
  120. xinference/model/llm/llm_family_csghub.json +0 -56
  121. xinference/model/llm/llm_family_modelscope.json +0 -8700
  122. xinference/model/llm/llm_family_openmind_hub.json +0 -1019
  123. xinference/model/rerank/model_spec_modelscope.json +0 -85
  124. xinference/model/video/model_spec_modelscope.json +0 -184
  125. xinference/web/ui/build/static/js/main.9b12b7f9.js +0 -3
  126. xinference/web/ui/build/static/js/main.9b12b7f9.js.map +0 -1
  127. xinference/web/ui/node_modules/.cache/babel-loader/1460361af6975e63576708039f1cb732faf9c672d97c494d4055fc6331460be0.json +0 -1
  128. xinference/web/ui/node_modules/.cache/babel-loader/4efd8dda58fda83ed9546bf2f587df67f8d98e639117bee2d9326a9a1d9bebb2.json +0 -1
  129. xinference/web/ui/node_modules/.cache/babel-loader/55b9fb40b57fa926e8f05f31c2f96467e76e5ad62f033dca97c03f9e8c4eb4fe.json +0 -1
  130. xinference/web/ui/node_modules/.cache/babel-loader/5b2dafe5aa9e1105e0244a2b6751807342fa86aa0144b4e84d947a1686102715.json +0 -1
  131. xinference/web/ui/node_modules/.cache/babel-loader/611fa2c6c53b66039991d06dfb0473b5ab37fc63b4564e0f6e1718523768a045.json +0 -1
  132. /xinference/web/ui/build/static/js/{main.9b12b7f9.js.LICENSE.txt → main.7d24df53.js.LICENSE.txt} +0 -0
  133. {xinference-1.7.1.post1.dist-info → xinference-1.8.0.dist-info}/WHEEL +0 -0
  134. {xinference-1.7.1.post1.dist-info → xinference-1.8.0.dist-info}/entry_points.txt +0 -0
  135. {xinference-1.7.1.post1.dist-info → xinference-1.8.0.dist-info}/licenses/LICENSE +0 -0
  136. {xinference-1.7.1.post1.dist-info → xinference-1.8.0.dist-info}/top_level.txt +0 -0
xinference/model/rerank/model_spec_modelscope.json DELETED
@@ -1,85 +0,0 @@
1
- [
2
- {
3
- "model_name": "bge-reranker-base",
4
- "type": "normal",
5
- "language": ["en", "zh"],
6
- "max_tokens": 512,
7
- "model_id": "Xorbits/bge-reranker-base",
8
- "model_revision": "v0.0.1",
9
- "model_hub": "modelscope"
10
- },
11
- {
12
- "model_name": "bge-reranker-large",
13
- "type": "normal",
14
- "language": ["en", "zh"],
15
- "max_tokens": 512,
16
- "model_id": "Xorbits/bge-reranker-large",
17
- "model_revision": "v0.0.1",
18
- "model_hub": "modelscope"
19
- },
20
- {
21
- "model_name": "bce-reranker-base_v1",
22
- "type": "normal",
23
- "language": ["en", "zh"],
24
- "max_tokens": 512,
25
- "model_id": "maidalun/bce-reranker-base_v1",
26
- "model_revision": "v0.0.1",
27
- "model_hub": "modelscope"
28
- },
29
- {
30
- "model_name": "bge-reranker-v2-m3",
31
- "type": "normal",
32
- "max_tokens": 8192,
33
- "language": ["en", "zh", "multilingual"],
34
- "model_id": "AI-ModelScope/bge-reranker-v2-m3",
35
- "model_hub": "modelscope"
36
- },
37
- {
38
- "model_name": "bge-reranker-v2-gemma",
39
- "type": "LLM-based",
40
- "language": ["en", "zh", "multilingual"],
41
- "max_tokens": 8192,
42
- "model_id": "AI-ModelScope/bge-reranker-v2-gemma",
43
- "model_hub": "modelscope"
44
- },
45
- {
46
- "model_name": "bge-reranker-v2-minicpm-layerwise",
47
- "type": "LLM-based layerwise",
48
- "language": ["en", "zh", "multilingual"],
49
- "max_tokens": 2048,
50
- "model_id": "mirror013/bge-reranker-v2-minicpm-layerwise",
51
- "model_hub": "modelscope"
52
- },
53
- {
54
- "model_name": "minicpm-reranker",
55
- "type": "normal",
56
- "language": ["en", "zh"],
57
- "max_tokens": 1024,
58
- "model_id": "OpenBMB/MiniCPM-Reranker",
59
- "model_hub": "modelscope"
60
- },
61
- {
62
- "model_name": "Qwen3-Reranker-0.6B",
63
- "type": "normal",
64
- "language": ["en", "zh"],
65
- "max_tokens": 32768,
66
- "model_id": "Qwen/Qwen3-Reranker-0.6B",
67
- "model_hub": "modelscope"
68
- },
69
- {
70
- "model_name": "Qwen3-Reranker-4B",
71
- "type": "normal",
72
- "language": ["en", "zh"],
73
- "max_tokens": 32768,
74
- "model_id": "Qwen/Qwen3-Reranker-4B",
75
- "model_hub": "modelscope"
76
- },
77
- {
78
- "model_name": "Qwen3-Reranker-8B",
79
- "type": "normal",
80
- "language": ["en", "zh"],
81
- "max_tokens": 32768,
82
- "model_id": "Qwen/Qwen3-Reranker-8B",
83
- "model_hub": "modelscope"
84
- }
85
- ]
xinference/model/video/model_spec_modelscope.json DELETED
@@ -1,184 +0,0 @@
1
- [
2
- {
3
- "model_name": "CogVideoX-2b",
4
- "model_family": "CogVideoX",
5
- "model_hub": "modelscope",
6
- "model_id": "ZhipuAI/CogVideoX-2b",
7
- "model_revision": "master",
8
- "model_ability": [
9
- "text2video"
10
- ],
11
- "default_model_config": {
12
- "scheduler": "CogVideoXDDIMScheduler",
13
- "torch_dtype": "float16"
14
- },
15
- "default_generate_config": {
16
- "guidance_scale": 6
17
- }
18
- },
19
- {
20
- "model_name": "CogVideoX-5b",
21
- "model_family": "CogVideoX",
22
- "model_hub": "modelscope",
23
- "model_id": "ZhipuAI/CogVideoX-5b",
24
- "model_revision": "master",
25
- "model_ability": [
26
- "text2video"
27
- ],
28
- "default_model_config": {
29
- "scheduler": "CogVideoXDPMScheduler",
30
- "torch_dtype": "bfloat16"
31
- },
32
- "default_generate_config": {
33
- "guidance_scale": 7
34
- }
35
- },
36
- {
37
- "model_name": "HunyuanVideo",
38
- "model_family": "HunyuanVideo",
39
- "model_hub": "modelscope",
40
- "model_id": "Xorbits/HunyuanVideo",
41
- "model_revision": "master",
42
- "model_ability": [
43
- "text2video"
44
- ],
45
- "default_model_config": {
46
- "transformer_torch_dtype": "bfloat16",
47
- "torch_dtype": "float16"
48
- },
49
- "default_generate_config": {
50
- }
51
- },
52
- {
53
- "model_name": "Wan2.1-1.3B",
54
- "model_family": "Wan",
55
- "model_hub": "modelscope",
56
- "model_id": "Wan-AI/Wan2.1-T2V-1.3B-Diffusers",
57
- "model_revision": "master",
58
- "model_ability": [
59
- "text2video"
60
- ],
61
- "default_model_config": {
62
- "torch_dtype": "bfloat16"
63
- },
64
- "default_generate_config": {
65
- },
66
- "virtualenv": {
67
- "packages": [
68
- "diffusers>=0.33.0",
69
- "ftfy",
70
- "imageio-ffmpeg",
71
- "imageio",
72
- "#system_numpy#"
73
- ]
74
- }
75
- },
76
- {
77
- "model_name": "Wan2.1-14B",
78
- "model_family": "Wan",
79
- "model_hub": "modelscope",
80
- "model_id": "Wan-AI/Wan2.1-T2V-14B-Diffusers",
81
- "model_revision": "master",
82
- "model_ability": [
83
- "text2video"
84
- ],
85
- "default_model_config": {
86
- "torch_dtype": "bfloat16"
87
- },
88
- "default_generate_config": {
89
- },
90
- "virtualenv": {
91
- "packages": [
92
- "diffusers>=0.33.0",
93
- "ftfy",
94
- "imageio-ffmpeg",
95
- "imageio",
96
- "#system_numpy#"
97
- ]
98
- }
99
- },
100
- {
101
- "model_name": "Wan2.1-i2v-14B-480p",
102
- "model_family": "Wan",
103
- "model_hub": "modelscope",
104
- "model_id": "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers",
105
- "model_revision": "master",
106
- "model_ability": [
107
- "image2video"
108
- ],
109
- "default_model_config": {
110
- "torch_dtype": "bfloat16"
111
- },
112
- "default_generate_config": {
113
- "max_area": [
114
- 480,
115
- 832
116
- ]
117
- },
118
- "virtualenv": {
119
- "packages": [
120
- "diffusers>=0.33.0",
121
- "ftfy",
122
- "imageio-ffmpeg",
123
- "imageio",
124
- "#system_numpy#"
125
- ]
126
- }
127
- },
128
- {
129
- "model_name": "Wan2.1-i2v-14B-720p",
130
- "model_family": "Wan",
131
- "model_hub": "modelscope",
132
- "model_id": "Wan-AI/Wan2.1-I2V-14B-720P-Diffusers",
133
- "model_revision": "master",
134
- "model_ability": [
135
- "image2video"
136
- ],
137
- "default_model_config": {
138
- "torch_dtype": "bfloat16"
139
- },
140
- "default_generate_config": {
141
- "max_area": [
142
- 720,
143
- 1280
144
- ]
145
- },
146
- "virtualenv": {
147
- "packages": [
148
- "diffusers>=0.33.0",
149
- "ftfy",
150
- "imageio-ffmpeg",
151
- "imageio",
152
- "#system_numpy#"
153
- ]
154
- }
155
- },
156
- {
157
- "model_name": "Wan2.1-flf2v-14B-720p",
158
- "model_family": "Wan",
159
- "model_hub": "modelscope",
160
- "model_id": "Wan-AI/Wan2.1-FLF2V-14B-720P-diffusers",
161
- "model_revision": "master",
162
- "model_ability": [
163
- "firstlastframe2video"
164
- ],
165
- "default_model_config": {
166
- "torch_dtype": "bfloat16"
167
- },
168
- "default_generate_config": {
169
- "max_area": [
170
- 720,
171
- 1280
172
- ]
173
- },
174
- "virtualenv": {
175
- "packages": [
176
- "git+https://github.com/huggingface/diffusers",
177
- "ftfy",
178
- "imageio-ffmpeg",
179
- "imageio",
180
- "#system_numpy#"
181
- ]
182
- }
183
- }
184
- ]