modelscope
离线GPU版本镜像报错
问题: 镜像版本:funasr-runtime-sdk-gpu-0.1.1 部署了GPU版本的离线语音识别,任务跑着跑着会报错挂掉,辛苦帮忙看下,下面是报错的详细信息
报错: [aac @ 0x7f7ed01e73c0] invalid band type [aac @ 0x7f7ed01e73c0] invalid band type [aac @ 0x7f7ed01e73c0] Number of bands (54) exceeds limit (42). [mpegts @ 0x7f7ed02cb580] Packet corrupt (stream = 0, dts = NOPTS). [aac @ 0x7f7ed01e73c0] Reserved bit set. [aac @ 0x7f7ed01e73c0] Prediction is not allowed in AAC-LC. [aac @ 0x7f7ed01e73c0] Number of bands (45) exceeds limit (42). [aac @ 0x7f7ed01e73c0] Sample rate index in program config element does not match the sample rate index configured by the container. [aac @ 0x7f7ed01e73c0] Pulse tool not allowed in eight short sequence. [mpegts @ 0x7f7ed02cb580] PES packet size mismatch [mpegts @ 0x7f7ed02cb580] Packet corrupt (stream = 0, dts = 2355813). [mpegts @ 0x7f7ed02cb580] PES packet size mismatch [mpegts @ 0x7f7ed02cb580] Packet corrupt (stream = 0, dts = 5503045). [aac @ 0x7f7ed01e73c0] ms_present = 3 is reserved. [aac @ 0x7f7ed01e73c0] channel element 3.9 is not allocated [aac @ 0x7f7ed01e73c0] Number of bands (67) exceeds limit (42). [aac @ 0x7f7ed01e73c0] Pulse tool not allowed in eight short sequence. [mpegts @ 0x7f7ed02cb580] Packet corrupt (stream = 0, dts = NOPTS). [aac @ 0x7f7ed01e73c0] Number of bands (49) exceeds limit (42). [aac @ 0x7f7ed01e73c0] channel element 3.4 is not allocated [mpegts @ 0x7f7ed02cb580] PES packet size mismatch [mpegts @ 0x7f7ed02cb580] Packet corrupt (stream = 0, dts = 5517674). [mpegts @ 0x7f7ed02cb580] PES packet size mismatch [mpegts @ 0x7f7ed02cb580] Packet corrupt (stream = 0, dts = 2370441). [aac @ 0x7f7ed01e73c0] invalid band type [aac @ 0x7f7ed01e73c0] Reserved bit set. [aac @ 0x7f7ed01e73c0] Number of scalefactor bands in group (55) exceeds limit (49). [aac @ 0x7f7ed01e73c0] Number of bands (53) exceeds limit (42). [aac @ 0x7f7ed01e73c0] Prediction is not allowed in AAC-LC. I20240919 01:34:01.352533 1780 websocket-server.cpp:364] client done F20240919 01:34:01.474787 1718 wfst-decoder.h:32] Check failed: frm < num_frames_ (27 vs. 0) *** Check failure stack trace: *** @ 0x7f8107252f36 @ 0x7f810724f969 @ 0x7f8107250fbf @ 0x7f8108a8b245 @ 0x7f8108ac8ad5 @ 0x7f8108ac8d83 @ 0x7f8108a899c7 @ 0x7f8108a935e1 @ 0x7f8108a97da2 @ 0x7f81089f27c9 @ 0x559edd706118 @ 0x559edd714aff @ 0x559edd685618 @ 0x559edd66fa4d @ 0x7f81001845a0 @ 0x7f81074c9609 @ 0x7f80e3df8133 @ (nil)
遇到同样的问题 版本 registry.cn-hangzhou.aliyuncs.com/funasr_repo/funasr:funasr-runtime-sdk-cpu-0.4.7
遇到同样的问题 版本 registry.cn-hangzhou.aliyuncs.com/funasr_repo/funasr:funasr-runtime-sdk-cpu-0.4.7
偶发找不到规律，使用的是offline模式，使用文件的方式上传的音频文件，模型下面的启动方式：
cd FunASR/runtime
nohup bash run_server.sh
--model-thread-num 8
--download-model-dir /workspace/models
--vad-dir /workspace/models/damo/speech_fsmn_vad_zh-cn-16k-common-onnx
--model-dir /workspace/models/iic/speech_seaco_paraformer_large_asr_nat-zh-cn-16k-common-vocab8404-pytorch
--punc-dir /workspace/models/damo/punc_ct-transformer_cn-en-common-vocab471067-large-onnx
--lm-dir /workspace/models/damo/speech_ngram_lm_zh-cn-ai-wesp-fst
--itn-dir /workspace/models/thuduj12/fst_itn_zh
--certfile 0
--hotword /workspace/models/hotwords.txt > log.txt 2>&1 &
其中设置了不少热词，权重最大设置了 95
同最新版离线GPU版本遇到这个问题。复现方法很容易。，就是连续快速提交任务。只要速度稍微太快，提交三五个就挂了。与音频长短、热词数量没关系。
试验多次发现是在单个socket进程下连续发送才会出现，如果建立2个socket链接，在两个链接下同时发送都不会出行这个情况
I20250804 09:32:46.717374 37756 websocket-server.cpp:375] client done F20250804 09:32:46.840677 37690 wfst-decoder.h:32] Check failed: frm < num_frames_ (1 vs. 0) *** Check failure stack trace: *** @ 0x71056714f846 @ 0x71056714ba49 @ 0x71056714d12f @ 0x710568992fb5 @ 0x7105689c51f5 @ 0x7105689c54a3 @ 0x7105689916ef @ 0x710568941451 @ 0x71056894539b @ 0x7105688f8628 @ 0x59d257d40f43 @ 0x59d257d40d14 @ 0x59d257d4faf5 @ 0x59d257cc15e8 @ 0x59d257cb20ed @ 0x71055ff83df4 @ 0x7105673d2609
辛苦提供一个可复现的步骤
同最新版离线GPU版本遇到这个问题。复现方法很容易。，就是连续快速提交任务。只要速度稍微太快，提交三五个就挂了。