cudaErrorUnsupportedExecAffinity (802) cuda runtime_error ai_generated partial

CUDA 错误：不支持的执行亲和性 (cudaErrorUnsupportedExecAffinity)

CUDA error: unsupported exec affinity (cudaErrorUnsupportedExecAffinity)

ID: cuda/cuda-error-unsupported-exec-affinity

其他格式: JSON · Markdown 中文 · English
80%修复率
84%置信度
1证据数
2025-01-12首次发现

版本兼容性

版本状态引入弃用备注
CUDA 12.3 active — — —
CUDA 12.5 active — — —
NVIDIA Driver 545.23.06 active — — —
NVIDIA Driver 550.54.10 active — — —
PyTorch 2.4.0 active — — —

根因分析

应用程序请求的内核执行亲和性（例如通过 cudaLaunchAttributeCooperative 或 cudaFuncAttributeMaxDynamicSharedMemorySize）不受当前 GPU 架构或驱动程序版本支持。

English

The application requested a kernel execution affinity (e.g., via cudaLaunchAttributeCooperative or cudaFuncAttributeMaxDynamicSharedMemorySize) that is not supported by the current GPU architecture or driver version.

generic

官方文档

https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__TYPES.html

解决方案

  1. Check GPU compute capability and remove unsupported launch attributes. For example, cooperative launch requires sm >= 70: `deviceProp.major * 10 + deviceProp.minor >= 70`.
  2. Upgrade to a driver that supports the required execution affinity (e.g., driver 550.54.10 or newer for CUDA 12.5 features).
  3. Fall back to a non-cooperative launch by setting `cudaLaunchAttributeCooperative` to 0: `cudaLaunchConfig config; config.attrs[cudaLaunchAttributeCooperative] = 0;`

无效尝试

常见但无效的做法:

  1. 85% 失败

    The error is about execution affinity attributes, not shared memory limits; increasing shared memory may cause other issues but does not resolve the affinity problem.

  2. 90% 失败

    Synchronous execution does not affect launch attributes; the affinity check happens before the kernel is queued, regardless of blocking mode.