测试100次前30次3分钟音频，后70次50秒，内存增加到1.95GiB，使用funasr-runtime-sdk-cpu-0.4.7 ，推理结束内存一直不释放

[jinwater88]

使用services: cpp-runtime-sdk: image: registry.cn-hangzhou.aliyuncs.com/funasr_repo/funasr:funasr-runtime-sdk-cpu-0.4.7 container_name: cpp-runtime-sdk ports: - "10095:10095" volumes: - ./models:/workspace/models command: > bash -c "cd FunASR/runtime && nohup bash run_server.sh \ --download-model-dir /workspace/models \ --model-dir damo/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-onnx \ --certfile 0 \ --hotword /workspace/models/hotwords.txt > /tmp/runtime.log 2>&1 & tail -f /tmp/runtime.log" restart: unless-stopped healthcheck: test: ["CMD-SHELL", "bash -lc 'echo > /dev/tcp/127.0.0.1/10095'"] interval: 30s timeout: 5s retries: 3
测试脚本是：
如果调用次数增加岂不是，32G内存都不够用，这里有什么办法解决吗

[majic31]

应该不会的，我们之前做过测试，离线docker版是稳定的。valgrind也应该没检测到内存泄露的。
online那边倒有可能会出现泄露，需要修复下。

[lukeewin]

使用services: cpp-runtime-sdk: image: registry.cn-hangzhou.aliyuncs.com/funasr_repo/funasr:funasr-runtime-sdk-cpu-0.4.7 container_name: cpp-runtime-sdk ports: - "10095:10095" volumes: - ./models:/workspace/models command: > bash -c "cd FunASR/runtime && nohup bash run_server.sh \ --download-model-dir /workspace/models \ --model-dir damo/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-onnx \ --certfile 0 \ --hotword /workspace/models/hotwords.txt > /tmp/runtime.log 2>&1 & tail -f /tmp/runtime.log" restart: unless-stopped healthcheck: test: ["CMD-SHELL", "bash -lc 'echo > /dev/tcp/127.0.0.1/10095'"] interval: 30s timeout: 5s retries: 3 测试脚本是： 如果调用次数增加岂不是，32G内存都不够用，这里有什么办法解决吗

我用的源码编译的，并没有出现你说的随着转写次数的增加，内存会增加的情况。这边测试过转写7000余条音频， 并没有出现内存一直增加的情况。不过不清楚的docker拉取的版本。

[majic31]

使用services: cpp-runtime-sdk: image: registry.cn-hangzhou.aliyuncs.com/funasr_repo/funasr:funasr-runtime-sdk-cpu-0.4.7 container_name: cpp-runtime-sdk ports: - "10095:10095" volumes: - ./models:/workspace/models command: > bash -c "cd FunASR/runtime && nohup bash run_server.sh \ --download-model-dir /workspace/models \ --model-dir damo/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-onnx \ --certfile 0 \ --hotword /workspace/models/hotwords.txt > /tmp/runtime.log 2>&1 & tail -f /tmp/runtime.log" restart: unless-stopped healthcheck: test: ["CMD-SHELL", "bash -lc 'echo > /dev/tcp/127.0.0.1/10095'"] interval: 30s timeout: 5s retries: 3 测试脚本是： 如果调用次数增加岂不是，32G内存都不够用，这里有什么办法解决吗

我用的源码编译的，并没有出现你说的随着转写次数的增加，内存会增加的情况。这边测试过转写7000余条音频， 并没有出现内存一直增加的情况。不过不清楚的docker拉取的版本。

@jinwater88 的确是，建议自己编译会更可控一点。

[jinwater88]

@majic31 @lukeewin
这是我使用funasr-runtime-sdk-cpu-0.4.7的测试结果，可以看到测试100次从3.9G新增到5.857
1,2026-01-10T13:19:51+08:00,0,14438,4093960,4510084,4625600,4363280,4096636,4095211,4062600,32016,"3.902GiB / 31.07GiB","12.56%",38
2,2026-01-10T13:20:11+08:00,0,14443,4137992,4510084,4756672,4408852,4140832,4139368,4106736,32024,"3.944GiB / 31.07GiB","12.70%",38
3,2026-01-10T13:20:31+08:00,0,14438,4181768,4510084,4822208,4453376,4184452,4183012,4150356,32076,"3.986GiB / 31.07GiB","12.83%",38
4,2026-01-10T13:20:51+08:00,0,14534,4225032,4510084,4887744,4497876,4228044,4226588,4193948,32024,"4.028GiB / 31.07GiB","12.97%",38
5,2026-01-10T13:21:11+08:00,0,14524,4268552,4510084,4953280,4542664,4271924,4270447,4237828,32020,"4.07GiB / 31.07GiB","13.10%",38
6,2026-01-10T13:21:31+08:00,0,14516,4312840,4510084,5018816,4587624,4315964,4314500,4281868,32032,"4.112GiB / 31.07GiB","13.24%",38
7,2026-01-10T13:21:51+08:00,0,14565,4356360,4510084,5084352,4632132,4359564,4358095,4325468,32024,"4.154GiB / 31.07GiB","13.37%",38
8,2026-01-10T13:22:11+08:00,0,14586,4399880,4510084,5149888,4676632,4403152,4401724,4369056,32084,"4.195GiB / 31.07GiB","13.50%",38
9,2026-01-10T13:22:31+08:00,0,14503,4443144,4510084,5215424,4721132,4446744,4445339,4412648,32140,"4.237GiB / 31.07GiB","13.64%",38
10,2026-01-10T13:22:52+08:00,0,14522,4486920,4510084,5280960,4765632,4490332,4488861,4456236,32020,"4.278GiB / 31.07GiB","13.77%",38
11,2026-01-10T13:23:12+08:00,0,14507,4530696,4530696,5346496,4810132,4533928,4532490,4499832,32076,"4.32GiB / 31.07GiB","13.91%",38
12,2026-01-10T13:23:32+08:00,0,14468,4574216,4574216,5412032,4854632,4577516,4576050,4543420,32024,"4.362GiB / 31.07GiB","14.04%",38
13,2026-01-10T13:23:52+08:00,0,14554,4617480,4617480,5477568,4899132,4621104,4619709,4587008,32148,"4.404GiB / 31.07GiB","14.18%",38
14,2026-01-10T13:24:12+08:00,0,14541,4661000,4661000,5543104,4943632,4664692,4663264,4630596,32076,"4.446GiB / 31.07GiB","14.31%",38
15,2026-01-10T13:24:32+08:00,0,14423,4704520,4704520,5608640,4988132,4708280,4706819,4674184,32032,"4.487GiB / 31.07GiB","14.44%",38
16,2026-01-10T13:24:52+08:00,0,14406,4748040,4748040,5674176,5032636,4751872,4750436,4717776,32080,"4.529GiB / 31.07GiB","14.58%",38
17,2026-01-10T13:25:12+08:00,0,14403,4791048,4791048,5739712,5077136,4795464,4793994,4761368,32024,"4.571GiB / 31.07GiB","14.71%",38
18,2026-01-10T13:25:32+08:00,0,14421,4834568,4834568,5805248,5121636,4839052,4837600,4804956,32024,"4.612GiB / 31.07GiB","14.85%",38
19,2026-01-10T13:25:52+08:00,0,14416,4878344,4878344,5870784,5166476,4882984,4881530,4848888,32024,"4.654GiB / 31.07GiB","14.98%",38
20,2026-01-10T13:26:12+08:00,0,14416,4921864,4921864,5936320,5210976,4926572,4925104,4892476,32024,"4.696GiB / 31.07GiB","15.12%",38
21,2026-01-10T13:26:32+08:00,0,14412,4965128,4965128,6001856,5255476,4970160,4968688,4936064,32020,"4.738GiB / 31.07GiB","15.25%",38
22,2026-01-10T13:26:52+08:00,0,14412,5008904,5008904,6067392,5299976,5013748,5012309,4979652,32080,"4.779GiB / 31.07GiB","15.38%",38
23,2026-01-10T13:27:12+08:00,0,14405,5052680,5052680,6132928,5344604,5057464,5056004,5023368,32024,"4.821GiB / 31.07GiB","15.52%",38
24,2026-01-10T13:27:32+08:00,0,14407,5096200,5096200,6198464,5389104,5101052,5099593,5066956,32024,"4.863GiB / 31.07GiB","15.65%",38
25,2026-01-10T13:27:52+08:00,0,14453,5139976,5139976,6264000,5433604,5144640,5143170,5110544,32024,"4.904GiB / 31.07GiB","15.79%",38
26,2026-01-10T13:28:12+08:00,0,14469,5183496,5183496,6329536,5478104,5188228,5186761,5154132,32024,"4.946GiB / 31.07GiB","15.92%",38
27,2026-01-10T13:28:32+08:00,0,14468,5227016,5227016,6395072,5522604,5231820,5230355,5197724,32024,"4.988GiB / 31.07GiB","16.05%",38
28,2026-01-10T13:28:52+08:00,0,14451,5270792,5270792,6460608,5567104,5275408,5273951,5241312,32024,"5.029GiB / 31.07GiB","16.19%",38
29,2026-01-10T13:29:12+08:00,0,14468,5314568,5314568,6526144,5611604,5318996,5317599,5284900,32152,"5.071GiB / 31.07GiB","16.32%",38
30,2026-01-10T13:29:32+08:00,0,14484,5358088,5358088,6591680,5656104,5362588,5361120,5328492,32024,"5.113GiB / 31.07GiB","16.46%",38
31,2026-01-10T13:29:50+08:00,0,13157,5447432,5447432,6722752,5744800,5451468,5450000,5417372,32024,"5.198GiB / 31.07GiB","16.73%",38
32,2026-01-10T13:30:08+08:00,0,13146,5489972,5493816,6722752,5787376,5494192,5492797,5460096,32152,"5.238GiB / 31.07GiB","16.86%",38
33,2026-01-10T13:30:27+08:00,0,13144,5490484,5493816,6722752,5787548,5494696,5493234,5460600,32024,"5.239GiB / 31.07GiB","16.86%",38
34,2026-01-10T13:30:45+08:00,0,13168,5521172,5521172,6730880,5817076,5525108,5523634,5491012,32024,"5.268GiB / 31.07GiB","16.96%",38
35,2026-01-10T13:31:03+08:00,0,13126,5543072,5543072,6730880,5838940,5547664,5546232,5513568,32084,"5.29GiB / 31.07GiB","17.03%",38
36,2026-01-10T13:31:22+08:00,0,14048,5565356,5565356,6730880,5860356,5570076,5568614,5535980,32028,"5.311GiB / 31.07GiB","17.09%",38
37,2026-01-10T13:31:40+08:00,0,13103,5587440,5587440,6730880,5881292,5592008,5590551,5557912,32024,"5.332GiB / 31.07GiB","17.16%",38
38,2026-01-10T13:31:58+08:00,0,13227,5609372,5609372,6730880,5902228,5613940,5612476,5579844,32024,"5.353GiB / 31.07GiB","17.23%",38
39,2026-01-10T13:32:16+08:00,0,13180,5631384,5631384,6730880,5923164,5635872,5634398,5601776,32024,"5.374GiB / 31.07GiB","17.30%",38
40,2026-01-10T13:32:34+08:00,0,13146,5653420,5653420,6730880,5944100,5657804,5656350,5623708,32024,"5.395GiB / 31.07GiB","17.36%",38
41,2026-01-10T13:32:52+08:00,0,13143,5675628,5675628,6730880,5965508,5680212,5678738,5646116,32024,"5.416GiB / 31.07GiB","17.43%",38
42,2026-01-10T13:33:10+08:00,0,13157,5697428,5697428,6730880,5986444,5702144,5700677,5668048,32024,"5.437GiB / 31.07GiB","17.50%",38
43,2026-01-10T13:33:29+08:00,0,14041,5719360,5719360,6730880,6007380,5724076,5722616,5689980,32020,"5.458GiB / 31.07GiB","17.57%",38
44,2026-01-10T13:33:48+08:00,0,14047,5741408,5741408,6730880,6028316,5746008,5744536,5711912,32020,"5.479GiB / 31.07GiB","17.64%",38
45,2026-01-10T13:34:07+08:00,0,14079,5763340,5763340,6730880,6049252,5767940,5766506,5733844,32084,"5.5GiB / 31.07GiB","17.70%",38
46,2026-01-10T13:34:26+08:00,0,14058,5785360,5785360,6730880,6070188,5789872,5788404,5755776,32020,"5.521GiB / 31.07GiB","17.77%",38
47,2026-01-10T13:34:44+08:00,0,13044,5807292,5807292,6730880,6091124,5811804,5810347,5777708,32024,"5.542GiB / 31.07GiB","17.84%",38
48,2026-01-10T13:35:02+08:00,0,13040,5829188,5829188,6730880,6112060,5833740,5832270,5799644,32024,"5.563GiB / 31.07GiB","17.91%",38
49,2026-01-10T13:35:20+08:00,0,13057,5851004,5851004,6730880,6132996,5855672,5854222,5821576,32028,"5.584GiB / 31.07GiB","17.97%",38
50,2026-01-10T13:35:38+08:00,0,13082,5872680,5872680,6730880,6153932,5877604,5876130,5843508,32020,"5.605GiB / 31.07GiB","18.04%",38
51,2026-01-10T13:35:56+08:00,0,13208,5894728,5894728,6730880,6174868,5899536,5898083,5865440,32028,"5.626GiB / 31.07GiB","18.11%",38
52,2026-01-10T13:36:14+08:00,0,13086,5916660,5916660,6730880,6195804,5921468,5919987,5887372,32032,"5.647GiB / 31.07GiB","18.18%",38
53,2026-01-10T13:36:32+08:00,0,13197,5938592,5938592,6730880,6216740,5943400,5941942,5909304,32024,"5.668GiB / 31.07GiB","18.24%",38
54,2026-01-10T13:36:51+08:00,0,13238,5960604,5960604,6730880,6237676,5965332,5963880,5931236,32028,"5.689GiB / 31.07GiB","18.31%",38
55,2026-01-10T13:37:10+08:00,0,13172,5982420,5982420,6730880,6258612,5987264,5985808,5953168,32020,"5.71GiB / 31.07GiB","18.38%",38
56,2026-01-10T13:37:28+08:00,0,13128,6004472,6004472,6730880,6279548,6009196,6007759,5975100,32080,"5.731GiB / 31.07GiB","18.45%",38
57,2026-01-10T13:37:47+08:00,0,14129,6026484,6026484,6730880,6300540,6031188,6029723,5997092,32024,"5.752GiB / 31.07GiB","18.51%",38
58,2026-01-10T13:38:05+08:00,0,13215,6048416,6048416,6730880,6321476,6053120,6051669,6019024,32036,"5.772GiB / 31.07GiB","18.58%",38
59,2026-01-10T13:38:24+08:00,0,14185,6070076,6070076,6730880,6342412,6075056,6073591,6040960,32024,"5.793GiB / 31.07GiB","18.65%",38
60,2026-01-10T13:38:42+08:00,0,13155,6092260,6092260,6730880,6363348,6096988,6095510,6062892,32020,"5.814GiB / 31.07GiB","18.72%",38
61,2026-01-10T13:39:00+08:00,0,13163,6114140,6114140,6730880,6384284,6118920,6117463,6084824,32036,"5.835GiB / 31.07GiB","18.78%",38
62,2026-01-10T13:39:18+08:00,0,13091,6136092,6136092,6730880,6405220,6140852,6139387,6106756,32024,"5.856GiB / 31.07GiB","18.85%",38
63,2026-01-10T13:39:36+08:00,0,13114,6136848,6144796,6730880,6405776,6141408,6139944,6107312,32024,"5.857GiB / 31.07GiB","18.85%",38
64,2026-01-10T13:39:54+08:00,0,13169,6137020,6144796,6730880,6405776,6141408,6139956,6107312,32020,"5.857GiB / 31.07GiB","18.85%",38
65,2026-01-10T13:40:12+08:00,0,13159,6137020,6144796,6730880,6405832,6141464,6140036,6107368,32080,"5.857GiB / 31.07GiB","18.85%",38
66,2026-01-10T13:40:30+08:00,0,13100,6137020,6144796,6730880,6405832,6141464,6140007,6107368,32028,"5.857GiB / 31.07GiB","18.85%",38
67,2026-01-10T13:40:48+08:00,0,13153,6137020,6144796,6730880,6405832,6141464,6140002,6107368,32036,"5.857GiB / 31.07GiB","18.85%",38
68,2026-01-10T13:41:06+08:00,0,13165,6137020,6144796,6730880,6405832,6141464,6140002,6107368,32024,"5.857GiB / 31.07GiB","18.85%",38
69,2026-01-10T13:41:24+08:00,0,13143,6137020,6144796,6730880,6405832,6141464,6140003,6107368,32024,"5.857GiB / 31.07GiB","18.85%",38
70,2026-01-10T13:41:42+08:00,0,13150,6137020,6144796,6730880,6405832,6141464,6140036,6107368,32084,"5.857GiB / 31.07GiB","18.85%",38
71,2026-01-10T13:42:00+08:00,0,13144,6137020,6144796,6730880,6405832,6141464,6140003,6107368,32084,"5.857GiB / 31.07GiB","18.85%",38
72,2026-01-10T13:42:18+08:00,0,13120,6137020,6144796,6730880,6405832,6141464,6140026,6107368,32080,"5.857GiB / 31.07GiB","18.85%",38
73,2026-01-10T13:42:36+08:00,0,13118,6137020,6144796,6730880,6405832,6141468,6139999,6107372,32024,"5.857GiB / 31.07GiB","18.85%",38
74,2026-01-10T13:42:54+08:00,0,13226,6137020,6144796,6730880,6405832,6141468,6140003,6107372,32036,"5.857GiB / 31.07GiB","18.85%",38
75,2026-01-10T13:43:13+08:00,0,13136,6137140,6144796,6730880,6405896,6141532,6140060,6107436,32020,"5.857GiB / 31.07GiB","18.85%",38
76,2026-01-10T13:43:31+08:00,0,13107,6137140,6144796,6730880,6405896,6141536,6140139,6107440,32152,"5.857GiB / 31.07GiB","18.85%",38
77,2026-01-10T13:43:49+08:00,0,13161,6137140,6144796,6730880,6405896,6141536,6140085,6107440,32032,"5.857GiB / 31.07GiB","18.85%",38
78,2026-01-10T13:44:07+08:00,0,13159,6137140,6144796,6730880,6405896,6141536,6140074,6107440,32036,"5.857GiB / 31.07GiB","18.85%",38
79,2026-01-10T13:44:25+08:00,0,13133,6137140,6144796,6730880,6405896,6141536,6140065,6107440,32024,"5.857GiB / 31.07GiB","18.85%",38
80,2026-01-10T13:44:43+08:00,0,13224,6137140,6144796,6730880,6405896,6141536,6140108,6107440,32076,"5.857GiB / 31.07GiB","18.85%",38
81,2026-01-10T13:45:01+08:00,0,13164,6137140,6144796,6730880,6405896,6141536,6140079,6107440,32020,"5.857GiB / 31.07GiB","18.85%",38
82,2026-01-10T13:45:19+08:00,0,13108,6137140,6144796,6730880,6405896,6141536,6140106,6107440,32072,"5.857GiB / 31.07GiB","18.85%",38
83,2026-01-10T13:45:37+08:00,0,13155,6137140,6144796,6730880,6405896,6141540,6140080,6107444,32024,"5.857GiB / 31.07GiB","18.85%",38
84,2026-01-10T13:45:55+08:00,0,13157,6137140,6144796,6730880,6405896,6141540,6140081,6107444,32024,"5.857GiB / 31.07GiB","18.85%",38
85,2026-01-10T13:46:13+08:00,0,13185,6137140,6144796,6730880,6405896,6141540,6140068,6107444,32020,"5.857GiB / 31.07GiB","18.85%",38
86,2026-01-10T13:46:31+08:00,0,13146,6137140,6144796,6730880,6405896,6141540,6140097,6107444,32092,"5.857GiB / 31.07GiB","18.85%",38
87,2026-01-10T13:46:49+08:00,0,13073,6137140,6144796,6730880,6405896,6141540,6140066,6107444,32024,"5.857GiB / 31.07GiB","18.85%",38
88,2026-01-10T13:47:07+08:00,0,13156,6137140,6144796,6730880,6405896,6141540,6140111,6107444,32068,"5.857GiB / 31.07GiB","18.85%",38
89,2026-01-10T13:47:25+08:00,0,13114,6137140,6144796,6730880,6405896,6141544,6140114,6107448,32080,"5.857GiB / 31.07GiB","18.85%",38
90,2026-01-10T13:47:43+08:00,0,13143,6137140,6144796,6730880,6405896,6141544,6140083,6107448,32024,"5.857GiB / 31.07GiB","18.85%",38
91,2026-01-10T13:48:01+08:00,0,13198,6137140,6144796,6730880,6405896,6141548,6140067,6107452,32020,"5.857GiB / 31.07GiB","18.85%",38
92,2026-01-10T13:48:19+08:00,0,13166,6137140,6144796,6730880,6405896,6141548,6140097,6107452,32024,"5.857GiB / 31.07GiB","18.85%",38
93,2026-01-10T13:48:38+08:00,0,13189,6137140,6144796,6730880,6405896,6141548,6140111,6107452,32076,"5.857GiB / 31.07GiB","18.85%",38
94,2026-01-10T13:48:56+08:00,0,13229,6137140,6144796,6730880,6405896,6141548,6140115,6107452,32088,"5.857GiB / 31.07GiB","18.85%",38
95,2026-01-10T13:49:14+08:00,0,13184,6137140,6144796,6730880,6405896,6141548,6140096,6107452,32024,"5.857GiB / 31.07GiB","18.85%",38
96,2026-01-10T13:49:32+08:00,0,13206,6137124,6144796,6730880,6405896,6141548,6140091,6107452,32024,"5.857GiB / 31.07GiB","18.85%",38
97,2026-01-10T13:49:50+08:00,0,13061,6137124,6144796,6730880,6405904,6141560,6140099,6107464,32024,"5.857GiB / 31.07GiB","18.85%",38
98,2026-01-10T13:50:08+08:00,0,13089,6137124,6144796,6730880,6405904,6141560,6140097,6107464,32024,"5.857GiB / 31.07GiB","18.85%",38
99,2026-01-10T13:50:26+08:00,0,13109,6137124,6144796,6730880,6405904,6141560,6140098,6107464,32024,"5.857GiB / 31.07GiB","18.85%",38
100,2026-01-10T13:50:44+08:00,0,13176,6137124,6144796,6730880,6405904,6141560,6140130,6107464,32080,"5.857GiB / 31.07GiB","18.85%",38

然后我使用funasr-runtime-sdk-cpu-0.4.6启动容器前：

连续推理十次1分钟音频后出现：

这些都不算新增内存吗

我使用funasr 1.2.7：

model = AutoModel( model=config['model']['asr']['path'], vad_model=config['model']['vad']['path'], vad_kwargs=config['model']['vad']['kwargs'], punc_model=config['model']['punc']['path'], spk_model=config['model']['speaker']['path'], device=config['inference']['device'], disable_update=True # 禁用更新 ) res = model.generate( input=audio_path, cache={}, language=config['inference']['language'], use_itn=config['inference']['use_itn'], batch_size_s=config['inference']['batch_size'], merge_length_s=config['inference']['merge_length'], output_timestamp=config['inference']['output_timestamp'], output_spk_embedding=config['inference']['output_spk_embedding'] )

也是使用cpu进行推理，推理6次3分钟音频，模型是：
model:
asr:
path: "./funasr_models/iic/speech_seaco_paraformer_large_asr_nat-zh-cn-16k-common-vocab8404-pytorch"
type: "paraformer"
vad:
path: "./funasr_models/iic/speech_fsmn_vad_zh-cn-16k-common-pytorch"
kwargs:
max_single_segment_time: 30000
punc:
path: "./funasr_models/iic/punc_ct-transformer_cn-en-common-vocab471067-large"
speaker:
path: "./funasr_models/iic/speech_campplus_sv_zh-cn_16k-common"

inference:
device: "cpu"
batch_size: 60
merge_length: 15
language: "zn"
use_itn: true
output_timestamp: true
output_spk_embedding: true

启动模型是4.5G，推理后：

[majic31]

top命令，看看res的占用

[jinwater88]

我的代码
`import os
import time
import yaml
import uvicorn
from fastapi import FastAPI, UploadFile, File
from pydantic import BaseModel
from typing import Dict, List
import tempfile
from funasr import AutoModel
import logging
import warnings
import sys

配置日志处理器

def setup_logger():�
# 创建日志格式
formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')

# 创建控制台处理器
console_handler = logging.StreamHandler(sys.stdout)
console_handler.setFormatter(formatter)

# 配置应用日志记录器
app_logger = logging.getLogger("asr_server")
app_logger.setLevel(logging.INFO)

# 清除可能存在的旧处理器
app_logger.handlers.clear()
# 添加新的处理器
app_logger.addHandler(console_handler)

# 防止日志传播到根记录器
app_logger.propagate = False

# 关闭其他模块的日志
logging.getLogger("uvicorn.access").setLevel(logging.WARNING)
logging.getLogger("jieba").setLevel(logging.WARNING)
logging.getLogger("modelscope").setLevel(logging.WARNING)

return app_logger

初始化日志

logger = setup_logger()

加载配置

def load_config(config_path: str) -> Dict:
with open(config_path, 'r', encoding='utf-8') as f:
return yaml.safe_load(f)

初始化FastAPI

app = FastAPI(title="ASR Server")
config = load_config("asr_config.yaml")

添加健康检查接口

@app.get("/health")
async def health_check():
return {"status": "ASR Server OK"}

添加启动事件

@app.on_event("startup")
async def startup_event():
logger.info("ASR服务启动...")
logger.info(f"配置信息: {config}")

    # 记录模型加载的时间
start_time = time.time()
try:
    # 只在启动时初始化模型，避免重复加载
    global model
    model = AutoModel(
        model=config['model']['asr']['path'],
        vad_model=config['model']['vad']['path'],
        vad_kwargs=config['model']['vad']['kwargs'],
        punc_model=config['model']['punc']['path'],
        spk_model=config['model']['speaker']['path'],
        device=config['inference']['device'],
        disable_update=True  # 禁用更新
    )
    load_time = time.time() - start_time

    logger.info(f"模型加载完成，加载时间: {load_time:.2f}秒")
    logger.info(f"模型加载在设备: {config['inference']['device']}")
except Exception as e:
    logger.error(f"模型加载失败: {str(e)}")
    raise

添加关闭事件

@app.on_event("shutdown")
async def shutdown_event():
logger.info("ASR服务关闭...")

class ASRResponse(BaseModel):
success: bool
message: str = ""
results: List[Dict] = []
time_cost: float = 0.0

@app.post("/asr", response_model=ASRResponse)
async def asr_endpoint(audio_file: UploadFile = File(...)):
try:
start_time = time.time()

    # 保存上传的音频文件
    with tempfile.NamedTemporaryFile(delete=False, suffix='.wav') as tmp:
        tmp.write(await audio_file.read())
        audio_path = tmp.name
        
    # 执行识别
    res = model.generate(
        input=audio_path,
        cache={},
        language=config['inference']['language'],
        use_itn=config['inference']['use_itn'],
        batch_size_s=config['inference']['batch_size'],
        merge_length_s=config['inference']['merge_length'],
        output_timestamp=config['inference']['output_timestamp'],
        output_spk_embedding=config['inference']['output_spk_embedding']
    )
    
    # 删除临时文件
    os.unlink(audio_path)
    
    return ASRResponse(
        success=True,
        results=res,
        time_cost=time.time() - start_time
    )
    
except Exception as e:
    return ASRResponse(
        success=False,
        message=str(e)
    )

if name == "main":
uvicorn.run(
"asr_server:app",
host=config['server']['host'],
port=config['server']['port'],
workers=config['server']['workers'],
reload=False,
)`有问题吗

[jinwater88]

@majic31
模型推理前是

模型进行了100次3分钟音频推理后是