Executing the example reports an error

[chang-wenbin]

When I execute the demo under example, I get this error. Is it a problem with my CUDA driver version?
Driver Version: 550.127.08
CUDA Version: 12.8
python 3.10.16

python matmul_v5.py

/miniconda3/envs/cwb_research_310/lib/python3.10/site-packages/hidet/utils/stack_limit.py:33: UserWarning: The hard limit for stack size is too small (10.0 MiB), we recommend to increase it to 512.0 MiB. If you are the root user on Linux OS, you could refer to `man limits.conf` to increase this limit.
  warnings.warn(
Traceback (most recent call last):
  File "/cwb/tilus/examples/matmul/matmul_v5.py", line 234, in <module>
    main()
  File "/cwb/tilus/examples/matmul/matmul_v5.py", line 213, in main
    matmul(m, n, k, a, b, c_actual)
  File "miniconda3/envs/cwb_research_310/lib/python3.10/site-packages/tilus/lang/instantiated_script.py", line 728, in __call__
    jit_instance = JitInstance(self.script_cls, self.params, self.build_options, self.schedules, jit_key)
  File "/miniconda3/envs/cwb_research_310/lib/python3.10/site-packages/tilus/lang/instantiated_script.py", line 331, in __init__
    self._transpile_programs()
  File "/miniconda3/envs/cwb_research_310/lib/python3.10/site-packages/tilus/lang/instantiated_script.py", line 407, in _transpile_programs
    lazy_init()
  File "/miniconda3/envs/cwb_research_310/lib/python3.10/site-packages/tilus/target.py", line 141, in lazy_init
    get_current_target()
  File "/miniconda3/envs/cwb_research_310/lib/python3.10/site-packages/tilus/target.py", line 92, in get_current_target
    has_nvgpu = cuda.available() and cuda.device_count() > 0
  File "/miniconda3/envs/cwb_research_310/lib/python3.10/site-packages/hidet/cuda/device.py", line 75, in device_count
    assert err == 0, err
AssertionError: cudaError_t.cudaErrorInsufficientDriver

[yaoyaoding]

Hi @chang-wenbin ,

Can I know the version of cuda-python package?

pip show cuda-python
# and
nvidia-smi

[chang-wenbin]

Hi @chang-wenbin ,

Can I know the version of cuda-python package?

pip show cuda-python
# and
nvidia-smi

Of course, my machine configuration is：
Name: cuda-python Version: 13.0.0
Driver Version: 550.127.08 CUDA Version: 12.8
My GPU is H800

[yaoyaoding]

Hi @chang-wenbin,

Thanks for the information. I can reproduce the error. Tilus depends on cuda-python package, which recently released version 13 and it requires new GPU driver for cuda toolkit 13.

Quick fix:

pip uninstall cuda-python
pip install "cuda-python<13"

Alternatively, you can upgrade the gpu driver to the latest one.

[yaoyaoding]

#17 to add the hint. Thanks for reporting this issue.

[chang-wenbin]

Thank you very much. After rolling back the cuda-python version, it can now run normally. However, I found that the performance of tilus on Hopper architecture GPUs needs to be improved. Is it because the Hopper architecture features are not currently used for optimization?

[yaoyaoding]

Is it because the Hopper architecture features are not currently used for optimization?

Yes, we need to support warp-specialization and use wgmma to execute the dot operator. We will add the support later.

[chang-wenbin]

We will add the support later.

This is awesome, I'm really looking forward to it.