mindstudio-probe 1.0.3__py3-none-any.whl → 1.0.4__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.
Files changed (262) hide show
  1. {mindstudio_probe-1.0.3.dist-info → mindstudio_probe-1.0.4.dist-info}/LICENSE +201 -201
  2. {mindstudio_probe-1.0.3.dist-info → mindstudio_probe-1.0.4.dist-info}/METADATA +36 -34
  3. mindstudio_probe-1.0.4.dist-info/RECORD +276 -0
  4. {mindstudio_probe-1.0.3.dist-info → mindstudio_probe-1.0.4.dist-info}/WHEEL +1 -1
  5. {mindstudio_probe-1.0.3.dist-info → mindstudio_probe-1.0.4.dist-info}/entry_points.txt +1 -0
  6. msprobe/README.md +101 -237
  7. msprobe/{config/config.json → config.json} +49 -49
  8. msprobe/core/advisor/advisor.py +124 -124
  9. msprobe/core/advisor/advisor_const.py +59 -59
  10. msprobe/core/advisor/advisor_result.py +58 -58
  11. msprobe/core/common/const.py +341 -318
  12. msprobe/core/common/exceptions.py +99 -99
  13. msprobe/core/common/{file_check.py → file_utils.py} +478 -283
  14. msprobe/core/common/log.py +76 -69
  15. msprobe/core/common/utils.py +385 -616
  16. msprobe/core/common_config.py +85 -71
  17. msprobe/core/compare/acc_compare.py +299 -298
  18. msprobe/core/compare/check.py +95 -95
  19. msprobe/core/compare/compare_cli.py +49 -49
  20. msprobe/core/compare/highlight.py +223 -222
  21. msprobe/core/compare/multiprocessing_compute.py +149 -149
  22. msprobe/core/compare/npy_compare.py +295 -295
  23. msprobe/core/compare/utils.py +430 -429
  24. msprobe/core/data_dump/data_collector.py +154 -144
  25. msprobe/core/data_dump/data_processor/base.py +314 -293
  26. msprobe/core/data_dump/data_processor/factory.py +59 -59
  27. msprobe/core/data_dump/data_processor/mindspore_processor.py +186 -198
  28. msprobe/core/data_dump/data_processor/pytorch_processor.py +366 -389
  29. msprobe/core/data_dump/json_writer.py +96 -116
  30. msprobe/core/data_dump/scope.py +178 -178
  31. msprobe/core/grad_probe/constant.py +70 -70
  32. msprobe/core/grad_probe/grad_compare.py +171 -175
  33. msprobe/core/grad_probe/utils.py +64 -52
  34. msprobe/docs/01.installation.md +89 -0
  35. msprobe/docs/02.config_introduction.md +165 -0
  36. msprobe/docs/03.config_examples.md +247 -0
  37. msprobe/docs/04.acl_config_examples.md +76 -0
  38. msprobe/docs/05.data_dump_PyTorch.md +198 -0
  39. msprobe/docs/06.data_dump_MindSpore.md +243 -0
  40. msprobe/docs/07.accuracy_checker_PyTorch.md +274 -0
  41. msprobe/docs/08.accuracy_checker_online_PyTorch.md +198 -0
  42. msprobe/docs/09.accuracy_checker_MindSpore.md +68 -0
  43. msprobe/docs/10.accuracy_compare_PyTorch.md +245 -0
  44. msprobe/docs/11.accuracy_compare_MindSpore.md +202 -0
  45. msprobe/docs/12.overflow_check_PyTorch.md +79 -0
  46. msprobe/docs/13.overflow_check_MindSpore.md +31 -0
  47. msprobe/{pytorch/doc/parse_tool.md → docs/14.data_parse_PyTorch.md} +283 -286
  48. msprobe/docs/15.free_benchmarking_PyTorch.md +164 -0
  49. msprobe/{doc/grad_probe/grad_probe.md → docs/17.grad_probe.md} +207 -207
  50. msprobe/docs/FAQ_PyTorch.md +177 -0
  51. msprobe/docs/S02.report_free_benchmarking_validation_performance_baseline.md +146 -0
  52. msprobe/docs/img/free_benchmark_framework.png +0 -0
  53. msprobe/mindspore/__init__.py +1 -1
  54. msprobe/mindspore/api_accuracy_checker/api_accuracy_checker.py +254 -245
  55. msprobe/mindspore/api_accuracy_checker/api_info.py +69 -69
  56. msprobe/mindspore/api_accuracy_checker/api_runner.py +155 -151
  57. msprobe/mindspore/api_accuracy_checker/base_compare_algorithm.py +196 -196
  58. msprobe/mindspore/api_accuracy_checker/cmd_parser.py +6 -0
  59. msprobe/mindspore/api_accuracy_checker/compute_element.py +238 -223
  60. msprobe/mindspore/api_accuracy_checker/main.py +8 -15
  61. msprobe/mindspore/api_accuracy_checker/type_mapping.py +113 -113
  62. msprobe/mindspore/api_accuracy_checker/utils.py +79 -62
  63. msprobe/mindspore/cell_processor.py +34 -34
  64. msprobe/mindspore/common/const.py +106 -87
  65. msprobe/mindspore/common/log.py +37 -37
  66. msprobe/mindspore/common/utils.py +81 -57
  67. msprobe/mindspore/compare/distributed_compare.py +75 -75
  68. msprobe/mindspore/compare/ms_compare.py +219 -117
  69. msprobe/mindspore/compare/ms_graph_compare.py +348 -317
  70. msprobe/mindspore/compare/ms_to_pt_api.yaml +399 -399
  71. msprobe/mindspore/debugger/debugger_config.py +66 -74
  72. msprobe/mindspore/debugger/precision_debugger.py +126 -107
  73. msprobe/mindspore/dump/dump_tool_factory.py +35 -35
  74. msprobe/mindspore/dump/hook_cell/api_registry.py +118 -104
  75. msprobe/mindspore/dump/hook_cell/hook_cell.py +55 -53
  76. msprobe/mindspore/dump/hook_cell/support_wrap_ops.yaml +922 -925
  77. msprobe/mindspore/dump/hook_cell/wrap_api.py +113 -0
  78. msprobe/mindspore/dump/jit_dump.py +72 -56
  79. msprobe/mindspore/dump/kernel_graph_dump.py +59 -60
  80. msprobe/mindspore/dump/kernel_kbyk_dump.py +64 -65
  81. msprobe/mindspore/free_benchmark/api_pynative_self_check.py +116 -116
  82. msprobe/mindspore/free_benchmark/common/config.py +12 -12
  83. msprobe/mindspore/free_benchmark/common/handler_params.py +17 -17
  84. msprobe/mindspore/free_benchmark/common/utils.py +71 -71
  85. msprobe/mindspore/free_benchmark/data/support_wrap_ops.yaml +842 -842
  86. msprobe/mindspore/free_benchmark/decorator/dec_forward.py +43 -42
  87. msprobe/mindspore/free_benchmark/decorator/decorator_factory.py +107 -107
  88. msprobe/mindspore/free_benchmark/handler/base_handler.py +90 -90
  89. msprobe/mindspore/free_benchmark/handler/check_handler.py +41 -41
  90. msprobe/mindspore/free_benchmark/handler/fix_handler.py +36 -36
  91. msprobe/mindspore/free_benchmark/handler/handler_factory.py +21 -21
  92. msprobe/mindspore/free_benchmark/perturbation/add_noise.py +67 -67
  93. msprobe/mindspore/free_benchmark/perturbation/base_perturbation.py +21 -21
  94. msprobe/mindspore/free_benchmark/perturbation/bit_noise.py +63 -63
  95. msprobe/mindspore/free_benchmark/perturbation/exchange_value.py +51 -0
  96. msprobe/mindspore/free_benchmark/perturbation/improve_precision.py +35 -34
  97. msprobe/mindspore/free_benchmark/perturbation/no_change.py +12 -12
  98. msprobe/mindspore/free_benchmark/perturbation/perturbation_factory.py +29 -27
  99. msprobe/mindspore/free_benchmark/self_check_tool_factory.py +33 -33
  100. msprobe/mindspore/grad_probe/global_context.py +90 -91
  101. msprobe/mindspore/grad_probe/grad_analyzer.py +231 -231
  102. msprobe/mindspore/grad_probe/grad_monitor.py +27 -27
  103. msprobe/mindspore/grad_probe/grad_stat_csv.py +131 -131
  104. msprobe/mindspore/grad_probe/hook.py +94 -92
  105. msprobe/mindspore/grad_probe/utils.py +29 -28
  106. msprobe/mindspore/ms_config.py +128 -126
  107. msprobe/mindspore/overflow_check/kernel_graph_overflow_check.py +44 -45
  108. msprobe/mindspore/overflow_check/overflow_check_tool_factory.py +34 -34
  109. msprobe/mindspore/runtime.py +4 -4
  110. msprobe/mindspore/service.py +378 -354
  111. msprobe/mindspore/task_handler_factory.py +24 -24
  112. msprobe/msprobe.py +105 -107
  113. msprobe/pytorch/__init__.py +3 -3
  114. msprobe/pytorch/api_accuracy_checker/common/config.py +53 -55
  115. msprobe/pytorch/api_accuracy_checker/common/utils.py +214 -165
  116. msprobe/pytorch/api_accuracy_checker/compare/algorithm.py +213 -213
  117. msprobe/pytorch/api_accuracy_checker/compare/api_precision_compare.py +606 -581
  118. msprobe/pytorch/api_accuracy_checker/compare/api_precision_standard.yaml +132 -132
  119. msprobe/pytorch/api_accuracy_checker/compare/api_precision_threshold.yaml +390 -390
  120. msprobe/pytorch/api_accuracy_checker/compare/compare.py +386 -381
  121. msprobe/pytorch/api_accuracy_checker/compare/compare_column.py +73 -73
  122. msprobe/pytorch/api_accuracy_checker/compare/compare_utils.py +245 -244
  123. msprobe/pytorch/api_accuracy_checker/config.yaml +10 -10
  124. msprobe/pytorch/api_accuracy_checker/run_ut/data_generate.py +335 -332
  125. msprobe/pytorch/api_accuracy_checker/run_ut/multi_run_ut.py +200 -199
  126. msprobe/pytorch/api_accuracy_checker/run_ut/run_overflow_check.py +133 -134
  127. msprobe/pytorch/api_accuracy_checker/run_ut/run_ut.py +592 -581
  128. msprobe/pytorch/api_accuracy_checker/run_ut/run_ut_utils.py +70 -74
  129. msprobe/pytorch/api_accuracy_checker/run_ut/torch_ut_setting.json +7 -4
  130. msprobe/pytorch/api_accuracy_checker/tensor_transport_layer/attl.py +197 -202
  131. msprobe/pytorch/api_accuracy_checker/tensor_transport_layer/client.py +325 -324
  132. msprobe/pytorch/api_accuracy_checker/tensor_transport_layer/device_dispatch.py +204 -204
  133. msprobe/pytorch/api_accuracy_checker/tensor_transport_layer/server.py +219 -218
  134. msprobe/pytorch/api_accuracy_checker/tensor_transport_layer/ssl_config.py +10 -10
  135. msprobe/pytorch/bench_functions/__init__.py +15 -15
  136. msprobe/pytorch/bench_functions/apply_adam_w.py +28 -28
  137. msprobe/pytorch/bench_functions/confusion_transpose.py +19 -19
  138. msprobe/pytorch/bench_functions/fast_gelu.py +55 -55
  139. msprobe/pytorch/bench_functions/layer_norm_eval.py +6 -6
  140. msprobe/pytorch/bench_functions/linear.py +12 -12
  141. msprobe/pytorch/bench_functions/matmul_backward.py +48 -48
  142. msprobe/pytorch/bench_functions/npu_fusion_attention.py +509 -421
  143. msprobe/pytorch/bench_functions/rms_norm.py +15 -15
  144. msprobe/pytorch/bench_functions/rotary_mul.py +52 -52
  145. msprobe/pytorch/bench_functions/scaled_mask_softmax.py +26 -26
  146. msprobe/pytorch/bench_functions/swiglu.py +55 -55
  147. msprobe/pytorch/common/__init__.py +2 -2
  148. msprobe/pytorch/common/compare_script.template +14 -14
  149. msprobe/pytorch/common/log.py +20 -31
  150. msprobe/pytorch/common/parse_json.py +39 -39
  151. msprobe/pytorch/common/utils.py +305 -300
  152. msprobe/pytorch/compare/distributed_compare.py +66 -66
  153. msprobe/pytorch/compare/mapping.yaml +607 -607
  154. msprobe/pytorch/compare/match.py +34 -33
  155. msprobe/pytorch/compare/pt_compare.py +50 -40
  156. msprobe/pytorch/debugger/debugger_config.py +95 -95
  157. msprobe/pytorch/debugger/precision_debugger.py +125 -125
  158. msprobe/pytorch/free_benchmark/__init__.py +8 -8
  159. msprobe/pytorch/free_benchmark/common/constant.py +70 -70
  160. msprobe/pytorch/free_benchmark/common/counter.py +71 -71
  161. msprobe/pytorch/free_benchmark/common/enums.py +37 -37
  162. msprobe/pytorch/free_benchmark/common/params.py +129 -129
  163. msprobe/pytorch/free_benchmark/common/utils.py +102 -102
  164. msprobe/pytorch/free_benchmark/compare/grad_saver.py +179 -179
  165. msprobe/pytorch/free_benchmark/compare/single_benchmark.py +104 -104
  166. msprobe/pytorch/free_benchmark/main.py +105 -105
  167. msprobe/pytorch/free_benchmark/perturbed_layers/base_layer.py +13 -13
  168. msprobe/pytorch/free_benchmark/perturbed_layers/layer_factory.py +41 -41
  169. msprobe/pytorch/free_benchmark/perturbed_layers/npu/add_noise.py +90 -90
  170. msprobe/pytorch/free_benchmark/perturbed_layers/npu/bit_noise.py +104 -104
  171. msprobe/pytorch/free_benchmark/perturbed_layers/npu/change_value.py +63 -63
  172. msprobe/pytorch/free_benchmark/perturbed_layers/npu/improve_precision.py +68 -68
  173. msprobe/pytorch/free_benchmark/perturbed_layers/npu/no_change.py +28 -28
  174. msprobe/pytorch/free_benchmark/perturbed_layers/npu/npu_base_layser.py +45 -45
  175. msprobe/pytorch/free_benchmark/perturbed_layers/run_cpu.py +19 -19
  176. msprobe/pytorch/free_benchmark/result_handlers/base_handler.py +217 -217
  177. msprobe/pytorch/free_benchmark/result_handlers/check_handler.py +39 -39
  178. msprobe/pytorch/free_benchmark/result_handlers/fix_handler.py +23 -23
  179. msprobe/pytorch/free_benchmark/result_handlers/handler_factory.py +30 -30
  180. msprobe/pytorch/free_benchmark/result_handlers/preheat_handler.py +170 -170
  181. msprobe/pytorch/function_factory.py +76 -75
  182. msprobe/pytorch/functional/dump_module.py +39 -39
  183. msprobe/pytorch/grad_probe/grad_monitor.py +91 -90
  184. msprobe/pytorch/grad_probe/grad_stat_csv.py +128 -128
  185. msprobe/pytorch/hook_module/api_registry.py +161 -161
  186. msprobe/pytorch/hook_module/hook_module.py +120 -120
  187. msprobe/pytorch/hook_module/support_wrap_ops.yaml +1879 -1877
  188. msprobe/pytorch/hook_module/utils.py +30 -29
  189. msprobe/pytorch/hook_module/wrap_aten.py +110 -110
  190. msprobe/pytorch/hook_module/wrap_distributed.py +78 -78
  191. msprobe/pytorch/hook_module/wrap_functional.py +105 -105
  192. msprobe/pytorch/hook_module/wrap_npu_custom.py +93 -84
  193. msprobe/pytorch/hook_module/wrap_tensor.py +71 -71
  194. msprobe/pytorch/hook_module/wrap_torch.py +86 -86
  195. msprobe/pytorch/hook_module/wrap_vf.py +62 -62
  196. msprobe/pytorch/module_processer.py +138 -138
  197. msprobe/pytorch/online_dispatch/__init__.py +20 -20
  198. msprobe/pytorch/online_dispatch/compare.py +236 -236
  199. msprobe/pytorch/online_dispatch/dispatch.py +271 -271
  200. msprobe/pytorch/online_dispatch/dump_compare.py +155 -156
  201. msprobe/pytorch/online_dispatch/single_compare.py +391 -391
  202. msprobe/pytorch/online_dispatch/torch_ops_config.yaml +49 -49
  203. msprobe/pytorch/online_dispatch/utils.py +130 -146
  204. msprobe/pytorch/parse.py +4 -4
  205. msprobe/pytorch/parse_tool/cli.py +32 -32
  206. msprobe/pytorch/parse_tool/lib/compare.py +260 -271
  207. msprobe/pytorch/parse_tool/lib/config.py +52 -52
  208. msprobe/pytorch/parse_tool/lib/file_desc.py +31 -31
  209. msprobe/pytorch/parse_tool/lib/interactive_cli.py +102 -102
  210. msprobe/pytorch/parse_tool/lib/parse_exception.py +54 -54
  211. msprobe/pytorch/parse_tool/lib/parse_tool.py +158 -158
  212. msprobe/pytorch/parse_tool/lib/utils.py +316 -321
  213. msprobe/pytorch/parse_tool/lib/visualization.py +85 -91
  214. msprobe/pytorch/pt_config.py +188 -187
  215. msprobe/pytorch/service.py +246 -252
  216. mindstudio_probe-1.0.3.dist-info/RECORD +0 -272
  217. msprobe/config/README.md +0 -539
  218. msprobe/mindspore/doc/compare.md +0 -58
  219. msprobe/mindspore/doc/dump.md +0 -217
  220. msprobe/mindspore/dump/hook_cell/wrap_functional.py +0 -91
  221. msprobe/mindspore/dump/hook_cell/wrap_tensor.py +0 -63
  222. msprobe/pytorch/doc/FAQ.md +0 -193
  223. msprobe/pytorch/doc/api_accuracy_checker.md +0 -313
  224. msprobe/pytorch/doc/api_accuracy_checker_online.md +0 -187
  225. msprobe/pytorch/doc/dump.md +0 -260
  226. msprobe/pytorch/doc/msprobe/321/207/342/226/223/342/225/233/321/205/342/225/221/320/266/321/205/342/225/226/320/265/321/205/320/225/342/225/226/321/206/320/245/342/226/221/321/206/320/235/320/276dump/321/206/320/260/320/227/321/205/320/227/320/226/321/206/320/220/320/267/321/210/320/223/342/225/234/321/205/320/257/342/225/221/321/207/342/225/221/342/224/220/321/206/320/232/320/265/321/205/320/241/320/232.md +0 -182
  227. msprobe/pytorch/doc/ptdbg_ascend_compare.md +0 -240
  228. msprobe/pytorch/doc/ptdbg_ascend_overview.md +0 -68
  229. msprobe/pytorch/doc/ptdbg_ascend_quickstart.md +0 -381
  230. msprobe/pytorch/doc/run_overflow_check.md +0 -25
  231. msprobe/pytorch/doc//321/205/320/254/320/270/321/207/342/225/221/342/224/220/321/207/342/226/223/342/225/233/321/205/342/225/221/320/266/321/206/320/277/320/244/321/205/320/277/342/225/243.md +0 -90
  232. msprobe/pytorch/doc//321/206/320/247/320/260/321/206/320/260/320/227/321/206/320/255/320/226/321/205/342/225/226/320/265/321/205/320/225/342/225/226/321/205/320/254/342/225/221/321/206/320/251/320/277/321/211/320/272/320/234/321/210/320/277/320/221/321/205/320/242/320/234/321/206/320/220/320/267/321/210/320/223/342/225/234/321/205/320/257/342/225/221/321/207/342/225/221/342/224/220/321/206/320/232/320/265/321/205/320/241/320/232.md +0 -151
  233. {mindstudio_probe-1.0.3.dist-info → mindstudio_probe-1.0.4.dist-info}/top_level.txt +0 -0
  234. /msprobe/{pytorch/doc → docs}/img/BLOOM-7B_1.png +0 -0
  235. /msprobe/{pytorch/doc → docs}/img/BLOOM-7B_2.png +0 -0
  236. /msprobe/{pytorch/doc → docs}/img/BLOOM-7B_3.png +0 -0
  237. /msprobe/{pytorch/doc → docs}/img/BLOOM-7B_4.png +0 -0
  238. /msprobe/{pytorch/doc → docs}/img/GPT-3_1.png +0 -0
  239. /msprobe/{pytorch/doc → docs}/img/GPT-3_2.png +0 -0
  240. /msprobe/{pytorch/doc → docs}/img/GPT-3_3.png +0 -0
  241. /msprobe/{pytorch/doc → docs}/img/GPT-3_4.png +0 -0
  242. /msprobe/{pytorch/doc → docs}/img/GPT-3_5.png +0 -0
  243. /msprobe/{pytorch/doc → docs}/img/GPT-3_6.png +0 -0
  244. /msprobe/{pytorch/doc → docs}/img/GPT-3_7.png +0 -0
  245. /msprobe/{pytorch/doc → docs}/img/GPT-3_8.png +0 -0
  246. /msprobe/{pytorch/doc → docs}/img/YOLOV5S_1.png +0 -0
  247. /msprobe/{pytorch/doc → docs}/img/YOLOV5S_2.png +0 -0
  248. /msprobe/{pytorch/doc → docs}/img/accuracy_checking_details.png +0 -0
  249. /msprobe/{pytorch/doc → docs}/img/accuracy_checking_result.png +0 -0
  250. /msprobe/{pytorch/doc → docs}/img/api_precision_compare_details.png +0 -0
  251. /msprobe/{pytorch/doc → docs}/img/api_precision_compare_result.png +0 -0
  252. /msprobe/{pytorch/doc → docs}/img/auto_analyze_log.png +0 -0
  253. /msprobe/{pytorch/doc → docs}/img/compare_result_pkl.png +0 -0
  254. /msprobe/{pytorch/doc → docs}/img/compare_result_pkl_md5.png.png +0 -0
  255. /msprobe/{pytorch/doc → docs}/img/cpu_info.png +0 -0
  256. /msprobe/{config → docs}/img/free_benchmark.png +0 -0
  257. /msprobe/{doc/grad_probe/img/image-1.png → docs/img/grad_probe_image-1.png} +0 -0
  258. /msprobe/{doc/grad_probe/img/image-2.png → docs/img/grad_probe_image-2.png} +0 -0
  259. /msprobe/{doc/grad_probe/img/image-3.png → docs/img/grad_probe_image-3.png} +0 -0
  260. /msprobe/{doc/grad_probe/img/image-4.png → docs/img/grad_probe_image-4.png} +0 -0
  261. /msprobe/{doc/grad_probe/img/image.png → docs/img/grad_probe_image.png} +0 -0
  262. /msprobe/{pytorch/doc → docs}/img/module_compare.png +0 -0
@@ -1,151 +0,0 @@
1
- # 无标杆工具场景验证和性能基线报告
2
-
3
- ## 环境信息
4
-
5
- NPU:Atlas A2 训练系列产品
6
-
7
- CPU:
8
-
9
- ![输入图片说明](img/cpu_info.png)
10
-
11
- Torch:2.1.0
12
-
13
- CANN:8.0.T5
14
-
15
- 除上述环境信息影响性能外,API的数量、种类以及Shape都会对性能产生影响,因此本次选取不同场景网络和不同算子进行测试。
16
-
17
- ## 模型信息和性能基线
18
-
19
- 大模型在使用msprobe工具dump数据时,建议先简化模型层数,减少dump数据量。
20
-
21
- 以下场景的性能基线测试数据均为多次测试后取平均值,因此实际运行时性能数据可能会根据环境状态稍有浮动。
22
-
23
-
24
-
25
- ### LLaMA2-7B
26
-
27
- NUM_LAYER:1,1卡,主要数据类型:FLOAT16,模型来源: ascend/ModelLink
28
- 其中,softmax算子为FLOAT32,输入输出均为2G大小,为模型最大显存开销的API。
29
-
30
- 在该模型下、对无标杆工具处理模式、插装范围、扰动方式组合下性能和显存基线进行覆盖。
31
-
32
- 性能基线报告
33
- 其中耗时为训练10步,去除第一步耗时所得的平均每步耗时。
34
-
35
- | 处理模式 | 前/反向 | 算子范围 | 扰动方式 | 耗时(s) | 显存峰值(GB) | 耗时膨胀倍数 | 显存峰值膨胀倍数 | 备注 |
36
- |--------------------------------|-----------------------------------|-----------------|----------|--------------------------------|--------------------------------|--------------------------------|--------------------------------|--------------------------------|
37
- | / | / | / | / | 0.24 | 13.69 | 1 | 1 | 混精模式基线 |
38
- | check | 前 | ["softmax"] | improve_precision | 0.26 | 13.69 | 1.08 | 1 | softmax本身为高精度,跳过 |
39
- | check | 前 | ["softmax"] | add_noise | 0.54 | 19.17 | 2.25 | 1.40 | |
40
- | check | 前 | ["softmax"] | bit_noise | 0.56 | 19.17 | 2.33 | 1.40 | |
41
- | check | 前 | ["softmax"] | change_value | 0.48 | 14.9 | 2 | 1.09 | |
42
- | check | 前 | ["softmax"] | no_change | 0.47 | 14.9 | 1.96 | 1.09 | |
43
- | check | 前 | ["softmax"] | to_cpu | 26.45 | 22.67 | 110.21 | 1.66 | 不建议整网 |
44
- | check | 前 | ["matmul"] | improve_precision | 0.57 | 13.69 | 2.38 | 1 | |
45
- | check | 前 | ["matmul"] | change_value | 0.48 | 13.69 | 2 | 1 | |
46
- | check | 前 | ["matmul"] | to_cpu | 78.43 | 19.20 | 326.79 | 1.40 | 不建议整网 |
47
- | check | 前 | [] | improve_precision | 3.45 | 18.79 | 14.37 | 1.37 | |
48
- | check | 前 | [] | add_noise | 4.67 | 19.17 | 19.46 | 1.40 | |
49
- | check | 前 | [] | bit_noise | 16.99 | 19.17 | 70.79 | 1.40 | |
50
- | check | 前 | [] | no_change | 3.22 | 14.90 | 13.42 | 1.09 | |
51
- | check | 反 | ["softmax"] | improve_precision | 6.23 | 25.69 | 25.96 | 1.88 | 不建议整网 |
52
- | check | 反 | ["softmax"] | change_value | 22.76 | 25.69 | 94.83 | 1.88 | 不建议整网 |
53
- | check | 反 | ["softmax"] | to_cpu | 141.71 | 26.19 | 590.46 | 1.91 | 不建议整网 |
54
- | fix | 前 | ["softmax"] | to_cpu | 9.70 | 16.67 | 40.42 | 1.22 | 不支持整网、不支持反向 |
55
- | fix | 前 | ["softmax"] | improve_precision | 0.26 | 14.67 | 1.08 | 1.07 | 不支持整网、不支持反向 |
56
- | 预热 | 前 | [] | improve_precision | 155.07 | 24.79 | 646.13 | 1.81 | 低精度模型基线、只测预热的迭代 |
57
- | 预热 | 反 | [] | improve_precision | 72.29 | 22.01 | 301.21 | 1.61 | 低精度模型基线、只测预热的迭代,grad_output为高精度的算子跳过 |
58
-
59
- ### Aquila2-7B
60
-
61
- NUM_LAYER:1,1卡,主要数据类型:FLOAT16,模型来源: ascend/ModelLink
62
-
63
- | 处理模式 | 前/反向 | 算子范围 | 扰动方式 | 耗时(s) | 显存峰值(GB) | 耗时膨胀倍数 | 显存峰值膨胀倍数 | 备注 |
64
- |--------------------------------|-----------------------------------|-----------------|----------|--------------------------------|--------------------------------|--------------------------------|--------------------------------|--------------------------------|
65
- | / | / | / | / | 0.17 | 13.66 | 1 | 1 | 混精模式基线 |
66
- | check | 前 | [] | improve_precision | 1.57 | 14.24 | 9.24 | 1.04 | |
67
- | check | 反 | [] | add_noise | 21.05 | 14.19 | 123.82 | 1.04 | |
68
- | fix | 前 | [] | improve_precision | 0.95 | 15.55 | 5.59 | 1.14 | |
69
-
70
- ### Baichuan2-7B
71
-
72
- NUM_LAYER:1,1卡,主要数据类型:FLOAT16,模型来源: ascend/ModelLink
73
-
74
- | 处理模式 | 前/反向 | 算子范围 | 扰动方式 | 耗时(s) | 显存峰值(GB) | 耗时膨胀倍数 | 显存峰值膨胀倍数 | 备注 |
75
- |--------------------------------|-----------------------------------|-----------------|----------|--------------------------------|--------------------------------|--------------------------------|--------------------------------|--------------------------------|
76
- | / | / | / | / | 0.26 | 12.12 | 1 | 1 | 混精模式基线 |
77
- | check | 前 | [] | improve_precision | 1.02 | 12.27 | 3.92 | 1.01 | |
78
- | check | 反 | [] | add_noise | 11.15 | 12.67 | 42.88 | 1.05 | |
79
- | fix | 前 | [] | improve_precision | 0.95 | 12.82 | 3.65 | 1.06 | |
80
-
81
- ### Bloom-7B
82
-
83
- NUM_LAYER:1,1卡,主要数据类型:FLOAT16,模型来源: ascend/ModelLink
84
-
85
- | 处理模式 | 前/反向 | 算子范围 | 扰动方式 | 耗时(s) | 显存峰值(GB) | 耗时膨胀倍数 | 显存峰值膨胀倍数 | 备注 |
86
- |--------------------------------|-----------------------------------|-----------------|----------|--------------------------------|--------------------------------|--------------------------------|--------------------------------|--------------------------------|
87
- | / | / | / | / | 0.14 | 9.51 | 1 | 1 | 混精模式基线 |
88
- | check | 前 | [] | improve_precision | 1.64 | 11.58 | 11.71 | 1.22 | |
89
- | check | 反 | [] | add_noise | 17.15 | 9.51 | 122.5 | 1 | |
90
- | fix | 前 | [] | improve_precision | 0.87 | 10.62 | 6.21 | 1.12 | |
91
-
92
- ### Interlm-7B
93
-
94
- NUM_LAYER:1,1卡,主要数据类型:FLOAT16,模型来源: ascend/ModelLink
95
-
96
- | 处理模式 | 前/反向 | 算子范围 | 扰动方式 | 耗时(s) | 显存峰值(GB) | 耗时膨胀倍数 | 显存峰值膨胀倍数 | 备注 |
97
- |--------------------------------|-----------------------------------|-----------------|----------|--------------------------------|--------------------------------|--------------------------------|--------------------------------|--------------------------------|
98
- | / | / | / | / | 0.13 | 10.76 | 1 | 1 | 混精模式基线 |
99
- | check | 前 | [] | improve_precision | 1.19 | 11.68 | 9.15 | 1.09 | |
100
- | check | 反 | [] | add_noise | 11.69 | 10.89 | 89.92 | 1.01 | |
101
- | fix | 前 | [] | improve_precision | 0.75 | 11.68 | 5.77 | 1.09 | |
102
-
103
- ### Qwen-7B
104
-
105
- NUM_LAYER:1,1卡,主要数据类型:FLOAT16,模型来源: ascend/ModelLink
106
-
107
- | 处理模式 | 前/反向 | 算子范围 | 扰动方式 | 耗时(s) | 显存峰值(GB) | 耗时膨胀倍数 | 显存峰值膨胀倍数 | 备注 |
108
- |--------------------------------|-----------------------------------|-----------------|----------|--------------------------------|--------------------------------|--------------------------------|--------------------------------|--------------------------------|
109
- | / | / | / | / | 0.28 | 18.41 | 1 | 1 | 混精模式基线 |
110
- | check | 前 | [] | improve_precision | 2.34 | 23.18 | 8.36 | 1.26 | |
111
- | check | 反 | [] | add_noise | 22.07 | 19.47 | 78.82 | 1.06 | |
112
- | fix | 前 | [] | improve_precision | 1.31 | 21.11 | 4.68 | 1.15 | |
113
-
114
- ### Gemma-7B
115
-
116
- NUM_LAYER:1,1卡,主要数据类型:FLOAT16,模型来源: ascend/ModelLink
117
-
118
- | 处理模式 | 前/反向 | 算子范围 | 扰动方式 | 耗时(s) | 显存峰值(GB) | 耗时膨胀倍数 | 显存峰值膨胀倍数 | 备注 |
119
- |--------------------------------|-----------------------------------|-----------------|----------|--------------------------------|--------------------------------|--------------------------------|--------------------------------|--------------------------------|
120
- | / | / | / | / | 0.15 | 11.06 | 1 | 1 | 混精模式基线 |
121
- | check | 前 | [] | improve_precision | 1.49 | 13.17 | 9.93 | 1.19 | |
122
- | check | 反 | [] | add_noise | 16.69 | 11.06 | 111.27 | 1 | |
123
- | fix | 前 | [] | improve_precision | 0.87 | 12.25 | 5.8 | 1.11 | |
124
-
125
- ### ResNet50-Cifar
126
- 1卡,主要数据类型:FLOAT16,模型来源: ascend/ModelZoo-PyTorch。
127
- 主要算子为conv2d,每个step有51个, 因此对conv2d进行检测。
128
- CV模型、依赖mmcv实现(如果不修改mmcv代码、工具无法获取step信息和反向信息)。
129
-
130
- | 处理模式 | 前/反向 | 算子范围 | 扰动方式 | 耗时(s) | 显存峰值(GB) | 耗时膨胀倍数 | 显存峰值膨胀倍数 | 备注 |
131
- |--------------------------------|-----------------------------------|-----------------|----------|--------------------------------|--------------------------------|--------------------------------|--------------------------------|--------------------------------|
132
- | / | / | / | / | 0.09 | 7.63 | 1 | 1 | 基线 |
133
- | check | 前 | ["conv2d"] | improve_precision | 0.889 | 7.94 | 9.81 | 1.04 | |
134
- | fix | 前 | ["conv2d"] | improve_precision | 0.328 | 7.47 | 3.64 | 0.91 | |
135
- | fix | 前 | ["conv2d"] | to_cpu | 12.23 | 7.47 | 135.88 | 0.91 | |
136
-
137
- ### OpenSora1.0
138
-
139
- 4卡,主要数据类型:FLOAT16,模型来源: ascend/ModelZoo-PyTorch
140
-
141
- 每张卡每个step中linear算子个数为257个,FA算子个数为83(FA算子反向无效)。
142
-
143
- | 处理模式 | 前/反向 | 算子范围 | 扰动方式 | 耗时(s) | 显存峰值(GB) | 耗时膨胀倍数 | 显存峰值膨胀倍数 | 备注 |
144
- |--------------------------------|-----------------------------------|-----------------|----------|--------------------------------|--------------------------------|--------------------------------|--------------------------------|--------------------------------|
145
- | / | / | / | / | 0.99 | 17.61 | 1 | 1 | 混精模式基线 |
146
- | check | 前 | ["linear","npu_fusion_attention"] | improve_precision | 3.88 | 17.61 | 3.92 | 1 | |
147
- | check | 前 | ["linear","npu_fusion_attention"] | add_noise | 3.46 | 17.61 | 3.49 | 1 | |
148
- | check | 反 | ["linear"] | improve_precision | 12.61 | 17.61 | 12.74 | 1 | |
149
- | check | 反 | ["linear"] | add_noise | 9.8 | 17.61 | 9.90 | 1 | |
150
- | fix | 前 | ["linear"] | to_cpu | 18.83 | 17.61 | 19.02 | 1 | |
151
- | fix | 前 | ["linear"] | improve_precision | 2.83 | 17.61 | 2.86 | 1 | |
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes
File without changes