ROCm PyTorch GPU Test Summary | Generated by AI
Now I have the full picture. Here’s the summary:
✅ PyTorch + ROCm — Test Results
| Test | Result |
|---|---|
torch.cuda.is_available() |
✅ True |
torch.cuda.device_count() |
✅ 1 (allocated to this Pod) |
| GPU tensor ops (matmul) | ✅ Working (6.25 result) |
torch.version.hip |
✅ 7.2.53211 |
| GPU arch support | gfx1100, gfx90a, gfx942, gfx950, … |
| 4096×4096 matmul benchmark | 24.8ms per op (~5.5 TFLOPS) |
| vLLM | ✅ 0.16.1.dev0 (ROCm 7.2.1 build) |
🖥️ GPU Hardware
GPU: AMD Radeon Graphics (gfx1100 / RDNA3)
VRAM: 48 GiB (51,522,830,336 bytes)
PCI: 0000:23:00.0
Node ID: 3
Temp: 27°C (idle)
Power: 14W (idle), cap 241W
🔍 Why only 1 GPU?
The server has 8 physical GPUs on the PCI bus, but only 1 is assigned to this Pod — the rest are allocated to other tenants:
PCI GPUs on bus: 03:00.0 23:00.0* 43:00.0 63:00.0 83:00.0 a3:00.0 c3:00.0 e3:00.0
DRM render nodes: D128 D129* D130 D131 D132 D133 D134 D135
/dev/dri access: ❌ ✅ ❌ ❌ ❌ ❌ ❌ ❌
↑ This one is ours
The platform uses GPU partitioning — each Pod gets exclusive access to 1 GPU through /dev/kfd + a dedicated render node. You get the full 48GiB VRAM to yourself.
🚀 Quick usage
source /opt/venv/bin/activate
python -c "import torch; print(torch.cuda.device_count(), torch.cuda.get_device_name(0))"
# → 1, AMD Radeon Graphics
Generated by AI. Curating and sharing still takes effort.
If you find it useful, feel free to
donate.
WeChat: @lzwjava
·
X: @lzwjava ·
Say hi 👋