CUBLAS_STATUS_ARCH_MISMATCH cuda runtime_error ai_generated true

RuntimeError: CUDA error: CUBLAS_STATUS_ARCH_MISMATCH when calling cublasSgemm

ID: cuda/cublas-api-not-found

Also available as: JSON · Markdown · 中文
82%Fix Rate
85%Confidence
1Evidence
2023-05-12First Seen

Version Compatibility

VersionStatusIntroducedDeprecatedNotes
CUDA 11.8 active — — —
cuBLAS 11.11 active — — —
PyTorch 2.0.1 active — — —
NVIDIA Driver 525.85.05 active — — —

Root Cause

The GPU's compute capability is too low for the cuBLAS kernel being invoked, typically because the code was compiled for sm_80+ but the GPU only supports sm_70 or earlier.

generic

中文

GPU 的计算能力低于所调用 cuBLAS 内核的要求，通常是因为代码针对 sm_80+ 编译，但 GPU 仅支持 sm_70 或更早版本。

Official Documentation

https://docs.nvidia.com/cuda/cublas/index.html#cublas-status-t

Workarounds

  1. 70% success export CUBLAS_WORKSPACE_CONFIG=":4096:8" && python your_script.py 
    export CUBLAS_WORKSPACE_CONFIG=":4096:8" && python your_script.py
  2. 85% success export TORCH_CUDA_ARCH_LIST='7.0;7.5' && pip install --no-cache-dir torch --verbose 
    export TORCH_CUDA_ARCH_LIST='7.0;7.5' && pip install --no-cache-dir torch --verbose

中文步骤

  1. export CUBLAS_WORKSPACE_CONFIG=":4096:8" && python your_script.py
  2. export TORCH_CUDA_ARCH_LIST='7.0;7.5' && pip install --no-cache-dir torch --verbose

Dead Ends

Common approaches that don't work:

  1. 90% fail

    Reinstallation does not change the GPU hardware or the compiled architecture targets; the mismatch persists.

  2. 85% fail

    Driver updates do not alter cuBLAS library architecture requirements; the kernel still expects a higher compute capability.