[Bug] Vulkan backend crashes on AMD GPUs on Windows (0xc0000005)

Bug Report: llama-funasr-sensevoice Vulkan Backend Crashes on AMD GPUs on Windows

中文报告

问题描述

在 Windows 系统下，使用 llama-funasr-sensevoice.exe 的 Vulkan 后端（--backend vulkan）时，程序在 AMD GPU 上崩溃，异常代码 0xc0000005（访问冲突）。

环境信息

• OS: Windows 11
• GPU: AMD Radeon RX 7600M XT (RDNA3) + AMD Radeon 780M (RDNA3, integrated)
• AMD 驱动版本: 32.0.31035.1003 (2026年7月24日，最新版)
• FunASR 版本: runtime-llamacpp-v0.1.9 (funasr-llamacpp-windows-x64-vulkan.zip)
• 模型: sensevoice-small-f16.gguf + fsmn-vad.gguf

复现步骤

1. 下载并解压 funasr-llamacpp-windows-x64-vulkan.zip
2. 下载模型：sensevoice-small-f16.gguf 和 fsmn-vad.gguf
3. 运行命令：

   .\llama-funasr-sensevoice.exe -m models\sensevoice-small-f16.gguf --vad models\fsmn-vad.gguf -a audio\examples\zh.mp3 --backend vulkan
   

预期行为

程序应使用 AMD GPU (Vulkan) 完成语音识别并输出转录文本。

实际行为

程序在 GPU 设备检测后立即崩溃，Windows 事件查看器显示：

Application Error: llama-funasr-sensevoice.exe
Exception Code: 0xc0000005 (Access Violation)
Faulting Module: llama-funasr-sensevoice.exe
Faulting Offset: 0x0000000000271743

已尝试的解决方法

1. ✅ CPU 模式 — 正常工作，转录结果正确
2. ❌ 设置 GGML_VK_DISABLE_COOPMAT=1 — coopmat 成功禁用（matrix cores: none），但崩溃依旧
3. ❌ 设置 GGML_VK_DISABLE_COOPMAT=1 + GGML_VK_DISABLE_COOPMAT2=1 — 崩溃依旧
4. ❌ 设置 GGML_VK_DISABLE_COOPMAT=1 + GGML_VK_DISABLE_COOPMAT2=1 + GGML_VK_DISABLE_INTEGER_DOT_PRODUCT=1 — 崩溃依旧
5. ❌ 设置 GGML_VULKAN_VALIDATE=1 — 崩溃依旧
6. ❌ 使用 GGML_VULKAN_DEVICE=0 或 1 指定特定 GPU — 崩溃依旧
7. ❌ 使用不同的驱动版本（从旧到新）— 均崩溃

崩溃分析

该崩溃与上游 ggml 项目中已知的 AMD Vulkan 驱动兼容性问题一致：

• whisper.cpp #3723: AMD 780M + VAD + beam-size 8 崩溃
• whisper.cpp #3806: RX 9070 XT + Flash Attention 崩溃
• llama.cpp #11074: ggml 为 AMD 专有驱动添加了 coopmat 黑名单
• llama.cpp #18763: 禁用大 coopmat 配置

FunASR 预编译二进制文件中缺少以下功能：

• 没有 --no-flash-attn 参数来禁用有问题的 Flash Attention 着色器
• 仅设置 GGML_VK_DISABLE_COOPMAT=1 不足以解决问题
• 崩溃发生在固定代码偏移 0x271743，与是否禁用 coopmat 无关

建议

1. 在 FunASR 文档中注明：AMD GPU 在 Windows 上不支持 Vulkan 后端
2. 添加 --no-flash-attn 参数支持
3. 在运行时检测 AMD Windows 驱动并自动禁用有问题的 Vulkan 功能
4. 或者建议 AMD Windows 用户使用 CPU 版本

---

English Report

Description

The llama-funasr-sensevoice.exe Vulkan backend (--backend vulkan) crashes with exception code 0xc0000005 (Access Violation) on AMD GPUs under Windows.

Environment

• OS: Windows 11
• GPU: AMD Radeon RX 7600M XT (RDNA3) + AMD Radeon 780M (RDNA3, integrated)
• AMD Driver: 32.0.31035.1003 (July 24, 2026, latest)
• FunASR Version: runtime-llamacpp-v0.1.9 (funasr-llamacpp-windows-x64-vulkan.zip)
• Model: sensevoice-small-f16.gguf + fsmn-vad.gguf

Steps to Reproduce

1. Download and extract funasr-llamacpp-windows-x64-vulkan.zip
2. Download models: sensevoice-small-f16.gguf and fsmn-vad.gguf
3. Run:

   .\llama-funasr-sensevoice.exe -m models\sensevoice-small-f16.gguf --vad models\fsmn-vad.gguf -a audio\examples\zh.mp3 --backend vulkan
   

Expected Behavior

Program should transcribe audio using AMD GPU (Vulkan) acceleration.

Actual Behavior

Program crashes immediately after GPU device detection. Windows Event Log shows:

Application Error: llama-funasr-sensevoice.exe
Exception Code: 0xc0000005 (Access Violation)
Faulting Module: llama-funasr-sensevoice.exe
Faulting Offset: 0x0000000000271743

Attempted Workarounds

1. ✅ CPU mode — works correctly
2. ❌ GGML_VK_DISABLE_COOPMAT=1 — coopmat disabled (matrix cores: none), still crashes
3. ❌ GGML_VK_DISABLE_COOPMAT=1 + GGML_VK_DISABLE_COOPMAT2=1 — still crashes
4. ❌ GGML_VK_DISABLE_COOPMAT=1 + GGML_VK_DISABLE_COOPMAT2=1 + GGML_VK_DISABLE_INTEGER_DOT_PRODUCT=1 — still crashes
5. ❌ GGML_VULKAN_VALIDATE=1 — still crashes
6. ❌ GGML_VULKAN_DEVICE=0 or 1 — still crashes
7. ❌ Various driver versions (old to new) — all crash

Root Cause Analysis

This crash is consistent with known AMD Vulkan driver compatibility issues in upstream ggml:

• whisper.cpp #3723: AMD 780M + VAD crash
• whisper.cpp #3806: RX 9070 XT + Flash Attention crash
• llama.cpp #11074: ggml added coopmat blacklist for AMD proprietary driver
• llama.cpp #18763: Disabled large coopmat config on AMD proprietary driver

The FunASR pre-built binary is missing:

• --no-flash-attn parameter to disable problematic Flash Attention shaders
• GGML_VK_DISABLE_COOPMAT=1 alone is insufficient
• Crash occurs at fixed offset 0x271743 regardless of coopmat status

Suggestions

1. Document in FunASR docs that AMD GPUs on Windows do not support the Vulkan backend
2. Add --no-flash-attn parameter support
3. Auto-detect AMD Windows driver and disable problematic Vulkan features at runtime
4. Or direct AMD Windows users to use the CPU package

• Thought: 768ms
  $ gh issue create --repo modelscope/FunASR --title "[Bug] Vulkan backend crashes on AMD GPUs on Windows (0xc0000005)" --body-file "D:\funasr-llamacpp-windows-x64-vulkan\bug_report.md" 2>&1
  gh : �޷�����gh����ʶ��Ϊ cmdlet���������ű��ļ�������г�������ơ��������Ƶ�ƴд���������·������ȷ��·����ȷ��Ȼ������
  һ�Ρ�
  ����λ�� ��:1 �ַ�: 1
• gh issue create --repo modelscope/FunASR --title "[Bug] Vulkan backen ...
• ~~
  • CategoryInfo : ObjectNotFound: (gh:String) [], CommandNotFoundException
  • FullyQualifiedErrorId : CommandNotFoundException

First reproduce the documented PowerShell command with the Vulkan backend and compare its behavior with CPU mode and the listed GGML_VK workarounds. Review the linked upstream Vulkan crash reports and determine which proposed scope—documentation, a runtime option, or driver-feature handling—can be supported; done requires a reproducible AMD/Windows outcome and a defined mitigation.