ggml-org/whisper.cpp

whisper_full_with_state: failed to decode

Open

#2,334 建立於 2024年8月3日

在 GitHub 查看
 (12 留言) (1 反應) (0 負責人)C++ (49,693 star) (5,535 fork)batch import
help wanted

描述

After running successfully for 6 minutes of audio, whisper.cpp (compiled with CUDA support) exits with the following message:

[00:06:04.000 --> 00:06:11.000]   Gut, dann kommen wir zur Sache, nämlich zum digitale Dienste Gesetz.
whisper_full_with_state: failed to decode
./main: failed to process audio

This error happens using the ggml-tiny.bin as a model and the -bs 8 -bo 8 flags. Using the base model, the same issue happens some minutes later.

Are there any tools available to get a more descriptive error message that could help narrow the source of the issue?

貢獻者指南

技術棧
ccpp
領域
climachine learning
議題類型
bug
難度面向新貢獻者的預計實作難度，1 表示很小改動，5 表示專家級工作。
4
預計時間有經驗貢獻者完成調查、實作、測試並準備 pull request 的粗略時間範圍。
3-5 days
活動狀態議題目前的可參與程度：新鮮、活躍、陳舊、阻塞或等待維護者輸入。
stale
清晰度議題是否清楚說明預期改動、驗收標準和下一步。
needs investigation
前置要求
C++CUDAwhisper.cpp codebase
新手友善度1-100 的估計分數，表示該議題對首次貢獻者的友善程度。
20
研究方向
The error 'whisper full with state: failed to decode' occurs after about 6 minutes of audio with the tiny model and batch size 8. Investigate the `whisper full with state` function in `whisper.cpp`, focusing on the decoding loop and CUDA error handling. Check for memory leaks or thread synchronization issues, especially with stateful processing. Also examine if the error is reproducible with different batch sizes or models. Look at the CUDA kernel implementations for potential out of bounds or segmentation faults.