环境如下：

# 查看系统/驱动与 python 版本
nvidia-smi || echo "nvidia-smi failed (no GPU or driver)"
python -V
python -c "import sys, importlib; print('cuda module spec:', importlib.util.find_spec('cuda'))"
Mon Sep  1 10:07:27 2025
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 580.76.04              Driver Version: 580.97         CUDA Version: 13.0     |
+-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA GeForce RTX 4080        On  |   00000000:08:00.0  On |                  N/A |
|  0%   51C    P0             49W /  340W |    1153MiB /  16376MiB |      2%      Default |
|                                         |                        |                  N/A |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI              PID   Type   Process name                        GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|    0   N/A  N/A              83      G   /Xwayland                             N/A      |
+-----------------------------------------------------------------------------------------+
Python 3.10.18
cuda module spec: ModuleSpec(name='cuda', loader=None, submodule_search_locations=_NamespacePath(['/root/miniconda3/envs/f5_tts_faster/lib/python3.10/site-packages/cuda']))


(f5_tts_faster) root@be89038f2e49:/home/F5-TTS/F5_TTS_Faster# python -c "from cuda import cudart; print('OK, cudart dir sample:', dir(cudart)[:10])"
Traceback (most recent call last):
  File "<string>", line 1, in <module>
ImportError: cannot import name 'cudart' from 'cuda' (unknown location)

检查环境：
 

python - <<'PY'
import importlib, os
spec = importlib.util.find_spec('cuda')
print('cuda spec:', spec)
if spec and getattr(spec, 'origin', None):
    print('cuda __file__:', spec.origin)
if spec and getattr(spec, 'submodule_search_locations', None):
    for p in spec.submodule_search_locations:
        print('cuda path:', p)
        try:
            print('listing:', os.listdir(p))
        except Exception as e:
            print('ls error:', e)
PY
cuda spec: ModuleSpec(name='cuda', loader=None, submodule_search_locations=_NamespacePath(['/root/miniconda3/envs/f5_tts_faster/lib/python3.10/site-packages/cuda']))
cuda path: /root/miniconda3/envs/f5_tts_faster/lib/python3.10/site-packages/cuda
listing: ['pathfinder', 'bindings']
(f5_tts_faster) root@be89038f2e49:/home/F5-TTS/F5_TTS_Faster# # pip packages related to cuda
python -m pip list | grep -i cuda || true
python -m pip show cuda || true
python -m pip show cuda-python || true

# 如果你用 conda 安装过，也看下 conda list
conda list | grep -i cuda || true
cuda-bindings                 13.0.1
cuda-pathfinder               1.2.1
cuda-python                   13.0.1
nvidia-cuda-runtime-cu12      12.9.79
pycuda                        2025.1.1
WARNING: Package(s) not found: cuda
Name: cuda-python
Version: 13.0.1
Summary: CUDA Python: Performance meets Productivity
Home-page: https://nvidia.github.io/cuda-python/
Author:
Author-email: NVIDIA Corporation <cuda-python-conduct@nvidia.com>
License-Expression: LicenseRef-NVIDIA-SOFTWARE-LICENSE
Location: /root/miniconda3/envs/f5_tts_faster/lib/python3.10/site-packages
Requires: cuda-bindings, cuda-pathfinder
Required-by: tensorrt_llm
cuda-bindings                    13.0.1           pypi_0                        pypi
cuda-cudart                      12.4.127         0                             nvidia
cuda-cupti                       12.4.127         0                             nvidia
cuda-libraries                   12.4.1           0                             nvidia
cuda-nvrtc                       12.4.127         0                             nvidia
cuda-nvtx                        12.4.127         0                             nvidia
cuda-opencl                      13.0.39          h59263cc_0                    nvidia
cuda-pathfinder                  1.2.1            pypi_0                        pypi
cuda-python                      13.0.1           pypi_0                        pypi
cuda-runtime                     12.4.1           0                             nvidia
cuda-version                     13.0             3                             nvidia
nvidia-cuda-runtime-cu12         12.9.79          pypi_0                        pypi
pycuda                           2025.1.1         pypi_0                        pypi
pytorch                          2.5.0            py3.10_cuda12.4_cudnn9.1.0_0  pytorch
pytorch-cuda                     12.4             hc786d27_7                    pytorch
pytorch-mutex                    1.0              cuda                          pytorch

解决方案：

pip uninstall cuda-python

pip install cuda-python==12.4.0

再次运行就正常了。

python -c "import tensorrt_llm"
python -c "import tensorrt_llm.bindings"
[TensorRT-LLM] TensorRT-LLM version: 0.15.0
[TensorRT-LLM] TensorRT-LLM version: 0.15.0

原因： cuda-python 版本与系统 CUDA toolkit（你使用的是 CUDA 12.4）或像 tensorrt_llm 这样的依赖不匹配。将 cuda-python 安装为与本地 CUDA 匹配的版本（例如 12.4.0）或尝试 12.6.0（在其它 NVIDIA 场景下解决过类似问题）就可以修复 from cuda import cudart 的导入错误。

tensorrt_llm/_ipc_utils.py 中尝试 from cuda import cudart，cudart 来自 cuda-python 的 CUDA Runtime API 绑定。如果 Python 能识别 cuda 名称空间但无法加载 cudart 子模块，就会出现类似“unknown location”的导入错误 —— 表示模块存在但不完整或其扩展模块加载失败（而不是典型的 ModuleNotFoundError）。