torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 214.66 GiB.

[jony4]

torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 214.66 GiB.

需要 214.66 G 显存？

[jony4]

我是 load 了一个 400M 的 wav 文件，显存崩溃。

同样文件在 whisper 时候，是可以 load 进入。不知道这里边有什么本质性差别。

暂时先用切割方式一块块的处理 wav 文件了

[LauraGPT]

Please ref to this doc: https://github.com/FunAudioLLM/SenseVoice?tab=readme-ov-file#method-2

[bitiliu]

The same issue occurs when I try to process a 1-hour MP3 file (approximately 14MB), throwing an exception:

torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 53.66 GiB. GPU 0 has a total capacity of 21.99 GiB of which 19.61 GiB is free. Including non-PyTorch memory, this process has 2.37 GiB memory in use. Of the allocated memory 1.99 GiB is allocated by PyTorch, and 92.67 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)

I don't want to enable VAD as it sometimes leads to information loss. Are there any other solutions?

[LauraGPT]

重新打开。原问题关闭后，@bitiliu 给出了新的、尚未解决的约束：一小时 MP3 在整段推理时尝试分配约 53.66 GiB，而且不希望启用可能丢信息的 VAD。这个后续不能留在 closed 状态。

当前边界需要说清楚：SenseVoice 对超长音频整段编码会让中间张量随时长增长；不使用 VAD 时，可靠方案应是带重叠的固定窗口切分 + 时间偏移回填 + 边界去重，而不是把一小时波形一次送入模型。我们会继续评估并补充一条不会静默丢段的参考实现与回归测试。

如果需求是“一次端到端推理同时得到长音频 ASR、分段时间戳和匿名说话人标签，并且不外挂 VAD”，FunASR 生态现在另有第三方 OpenMOSS MOSS-Transcribe-Diarize 路径：
https://www.funasr.com/deploy/moss-transcribe-diarize.html

这只是替代模型选择，不代表 SenseVoice 本身已经消除了长音频内存边界；模型与权重归 OpenMOSS，FunASR 提供适配和部署契约。本 Issue 保持开放，直到 SenseVoice 路径有可复现的无丢段长音频方案或报告者确认替代路径满足需求。

[LauraGPT]

A bounded no-VAD candidate is now available in #349. It uses ffmpeg streaming decode plus fixed overlapping windows, so the whole recording is never sent to one encoder call. The merged TXT only removes exact repeated boundary text, while the .chunks.jsonl sidecar preserves every raw window response and its input-window bounds.

Local evidence on exact head b8d141b2fb7bfe78a90316615f5662bae627d2fd:

• repository test suite: 11 passed
• real SenseVoiceSmall inference on the 6-second repository sample: 0-6000 ms, non-empty TXT and raw JSONL
• forced overlap run: 0-4000 ms and 3000-6000 ms, both raw responses retained

Please try the PR on the long recording that previously requested 53.66 GiB / 214.66 GiB and report peak memory plus any boundary omissions. This issue stays open with needs feedback; merging the PR alone will not be treated as confirmation that your recording is resolved.

[LauraGPT]

有界 no-VAD 参考路径 #349 已合并到 main，merge commit 为 4a131bd6ec0ece6cc1b00a063623223a970d4519；被验证的 exact head 为 b8d141b2fb7bfe78a90316615f5662bae627d2fd。

合并前在当前 main@eee0de09270b50d69e40b61434326e3fddefce3a 上重新模拟合并并验证：

• 仓库 11 项测试全部通过，py_compile 与 git diff --check 通过；
• 使用 FunASR 1.3.14 在 CPU 上对仓库自带 6 秒 WAV 进行真实推理，单窗口得到 0-6000 ms 且 TXT/JSONL 非空；
• 强制 4 秒窗口、1 秒重叠时得到 0-4000 ms 与 3000-6000 ms 两个窗口，两条原始响应均保留在 JSONL。

请 @jony4 和 @bitiliu 更新当前 main，分别用此前 400 MB WAV / 1 小时 MP3 复测：

python long_audio_no_vad.py input.mp3 --output transcript.txt --window-seconds 30 --overlap-seconds 2

请回复峰值 GPU/CPU 内存、窗口总数，以及是否发现边界处缺词或重复；若有边界问题，请附相邻两条 .chunks.jsonl 记录。固定窗口能约束单次编码内存，但不宣称切点完全无损。

本 issue 继续保持 open / needs feedback。PR 合并只表示参考实现进入主线，不代表两份原始长录音已经确认解决。