AMD ROCm audit on linux

amd-rocm-audit-docker.260216a.sh

# Replace <ctr> with container name
docker inspect <ctr> | jq '.[0].HostConfig.Devices, .[0].HostConfig.GroupAdd, .[0].HostConfig.Runtime, .[0].Config.Env'
docker logs --tail=200 <ctr>

amd-rocm-audit.260216a.sh

# 1) Hardware + kernel baseline
lscpu
uname -r
lsb_release -a

# 2) Boot/kernel args actually in effect
cat /proc/cmdline
grep -E '^GRUB_CMDLINE_LINUX(_DEFAULT)?=' /etc/default/grub

# 3) TTM/AMDGPU runtime knobs currently active
for p in /sys/module/ttm/parameters/pages_limit \
         /sys/module/ttm/parameters/page_pool_size \
         /sys/module/amdgpu/parameters/gttsize; do
  [ -e "$p" ] && echo "$p=$(cat "$p")" || echo "$p=not-present"
done

# 4) ROCm + GPU detection
which rocminfo >/dev/null 2>&1 && rocminfo | rg -i "Name:|gfx|Marketing Name" || echo "rocminfo not installed"
which rocm-smi >/dev/null 2>&1 && rocm-smi || echo "rocm-smi not installed"
clinfo | rg -i "platform|device|gfx|amd" || true

# 5) Driver/modules/log hints
lsmod | rg -i "amdgpu|kfd|ttm"
dmesg | rg -i "amdgpu|kfd|ttm|gttsize|pages_limit|iommu" | tail -n 200

# 6) Permissions/runtime prerequisites
id
ls -l /dev/kfd /dev/dri/render* 2>/dev/null
groups "$USER"

# 7) Container runtime (if vLLM in Docker/Podman)
docker ps --format "table {{.Names}}\t{{.Image}}\t{{.Status}}" 2>/dev/null || true
podman ps --format "table {{.Names}}\t{{.Image}}\t{{.Status}}" 2>/dev/null || true