Amund
Amund
@naromero77amd OK, so I'm sure that the `amdgpu` driver is loaded on the host, and therefore there is no `rocm` driver or `rocm-smi` command on the host. ```bash # lsmod...
## In the container: I don't know C++, but ChatGPT provided me with this code for `vector_add.cpp`: ```cpp #include #include #define N 100 __global__ void vector_add(const float *A, const float...
Here it is: ```bash # PYTORCH_NO_HIP_MEMORY_CACHING=1 python tensor.py PyTorch detects GPU: True ROCm device detected: AMD Radeon RX 7600 VRAM available: 7.98 GB Allocating tensors on GPU... Traceback (most recent...
Sorry, no better luck: ```bash # python simple.py cuda Traceback (most recent call last): File "/app/simple.py", line 11, in y = x + 1 ~~^~~ torch.AcceleratorError: HIP error: invalid device...
Yes, and there is a hint: ```bash # AMD_LOG_LEVEL=3 python simple.py :3:rocdevice.cpp :480 : 46756822561 us: Initalizing runtime stack, Enumerated GPU agents = 1 :3:rocdevice.cpp :233 : 46756822612 us: Numa...
Sorry, I thought it was good news, but it's true that I have no idea about the underlying implications. ```bash # ls -la /opt/venv/lib/python3.12/site-packages/torch/lib/libtorch_hip.so -rwxr-xr-x 1 root root 318427912 Oct...
Since I'm a bit of a tease, I tested the script with the latest image, using ROCM 7.1. The error is different, and more severe. ```bash # AMD_LOG_LEVEL=3 python simple.py...
```bash # pip3 install --pre torch torchvision --index-url https://download.pytorch.org/whl/nightly/rocm7.0 Looking in indexes: https://download.pytorch.org/whl/nightly/rocm7.0 Requirement already satisfied: torch in /opt/venv/lib/python3.12/site-packages (2.8.0+rocm7.0.2.lw.git245bf6ed) Requirement already satisfied: torchvision in /opt/venv/lib/python3.12/site-packages (0.23.0+rocm7.0.2.git824e8c87) Requirement already satisfied:...
I come back 2.5 years later with the intention of writing the code, and I realise that I've been beaten to it by two weeks by an AI. Bad luck....
@swissspidy I just tested it: it works perfectly, just like drush, and that's a really big plus for automation.