Verify GPU and ROCm

If you skip this page and jump to a 32B download, you will debug the wrong layer. Ten minutes here saves a night of “why is this on CPU.”

1. Driver and device nodes

lsmod | grep amdgpu
ls -l /dev/kfd /dev/dri
rocm-smi

rocm-smi should show your GPU, memory size near 32 GB on a R9700, and temperatures that look like a live card — not an empty table.

rocm-smi --showproductname --showmeminfo vram --showuse

Flags vary slightly by rocm-smi vintage; rocm-smi --help is the authority.

Watch compute and VRAM in another terminal while you run the checks below. If they never move, the process is not on the GPU.

2. rocminfo — the identity check

rocminfo

For the Radeon AI PRO R9700 you want, on a GPU agent:

• Marketing name containing Radeon AI PRO R9700 (partner cards may add a vendor prefix)
• ISA / GFX gfx1201

If you see gfx1100 or similar, you are looking at a different GPU (or an iGPU). An iGPU listed first is normal on Ryzen boxes — later steps must pick the discrete device.

rocminfo | grep -E 'Marketing Name|gfx'

3. A tiny HIP / userspace smoke test

If ROCm examples or hipcc are installed:

hipcc --version

A one-file check (save as square.cpp) that only proves the compiler and runtime:

#include <hip/hip_runtime.h>
#include <cstdio>

__global__ void inc(int* x) { *x += 1; }

int main() {
  int h = 41, *d = nullptr;
  hipMalloc(&d, sizeof(int));
  hipMemcpy(d, &h, sizeof(int), hipMemcpyHostToDevice);
  inc<<<1, 1>>>(d);
  hipMemcpy(&h, d, sizeof(int), hipMemcpyDeviceToHost);
  std::printf("result=%d device=%s\n", h, [] {
    static hipDeviceProp_t p;
    hipGetDeviceProperties(&p, 0);
    return p.name;
  }());
  return h == 42 ? 0 : 1;
}

hipcc square.cpp -o square && ./square

You want result=42 and a device name that is the Radeon, not a surprise iGPU. If hipcc is absent, skip to PyTorch — do not invent a second ROCm install just to get a compiler.

4. PyTorch on HIP

Use AMD’s current Install PyTorch for ROCm instructions. They distinguish gfx1201 wheels from other targets.

python3 -m venv ~/venvs/rocm-torch
source ~/venvs/rocm-torch/bin/activate
python -m pip install -U pip

# Paste the index-url and torch[device-gfx1201]==… lines from AMD's page.
# Do not install vanilla CUDA torch from pypi.org's default index.

Then:

import torch

print("torch", torch.__version__)
print("hip", torch.version.hip)
print("available", torch.cuda.is_available())
print("count", torch.cuda.device_count())
if torch.cuda.is_available():
    print("name", torch.cuda.get_device_name(0))
    x = torch.randn(2048, 2048, device="cuda")
    y = x @ x
    torch.cuda.synchronize()
    print("matmul_ok", float(y[0, 0]))

Success looks like: HIP version string present, available True, device name is the R9700, and rocm-smi shows activity during the matmul.

Failure patterns:

5. Optional: bandwidth sanity, not a benchmark

import time, torch
n = 256 * 1024 * 1024
a = torch.empty(n, device="cuda", dtype=torch.float32)
b = torch.empty_like(a)
torch.cuda.synchronize()
t0 = time.perf_counter()
b.copy_(a)
torch.cuda.synchronize()
seconds = time.perf_counter() - t0
gb = n * 4 / 1e9
print(f"rough_copy {gb / seconds:.1f} GB/s")

AMD publishes 640 GB/s peak on the R9700. A naive PyTorch copy will not hit that. Use this only as “is the number in a plausible band” versus “is it 8 GB/s because you are on host RAM.”

When to declare victory

You can move on when:

1. rocm-smi lists the card and ~32 GB VRAM
2. rocminfo says gfx1201 for that card
3. Either hipcc or PyTorch ran a kernel and rocm-smi noticed

Then run a local LLM. If step 1 already failed, go back to install ROCm — do not keep going down the framework path.