KubeRay's auto-injected wait-gcs-ready init container has only 256Mi
memory limit. The .pth hook was unconditionally importing torch+ROCm
which requires >256Mi, causing OOMKill.
Now checks cgroup memory limit first — if under 512Mi, skips the
expensive torch import entirely. The VRAM patch is only needed by the
main Ray worker process, not by health-check init containers.
On Strix Halo, PyTorch reports GTT pool (128 GiB) as device memory
instead of real VRAM (96 GiB from BIOS). vLLM uses mem_get_info() to
pre-allocate and refuses to start when free GTT (29 GiB) < requested.
The strixhalo_vram_fix.pth hook auto-patches mem_get_info on Python
startup to read real VRAM total/used from /sys/class/drm sysfs.
Only activates when PyTorch total differs >10% from sysfs VRAM.
The native amdsmi from ROCm 7.1 requires libamd_smi.so linked against
glibc 2.38 (Ubuntu 24.04), but the Ray base image is Ubuntu 22.04
(glibc 2.35). This caused vLLM to fail ROCm platform detection with
'No module named amdsmi' / GLIBC_2.38 not found errors.
Solution: Pure-Python amdsmi shim that reads GPU info from sysfs
(/sys/class/drm/*) instead of the native library. Provides the full
API surface used by both vLLM (platform detection, device info) and
PyTorch (device counting, memory/power/temp monitoring).
Tested in-container: vLLM detects RocmPlatform, PyTorch sees GPU
(Radeon 8060S, 128GB, HIP 7.3), DeviceConfig resolves to 'cuda'.
Changes:
- Add amdsmi-shim/ package with sysfs-backed implementation
- Update Dockerfile to install shim after vLLM/torch
- Add amdsmi-shim/ to .dockerignore explicit includes
