diff --git a/amdsmi-shim/strixhalo_vram_fix.py b/amdsmi-shim/strixhalo_vram_fix.py index 946a9a3..f5417b7 100644 --- a/amdsmi-shim/strixhalo_vram_fix.py +++ b/amdsmi-shim/strixhalo_vram_fix.py @@ -130,7 +130,7 @@ def _apply_patch() -> None: import torch if not hasattr(torch, "cuda") or not torch.cuda.is_available(): return - except ImportError: + except (ImportError, OSError): return finally: # CRITICAL: Clear the guard so child processes (vLLM EngineCore diff --git a/dockerfiles/Dockerfile.ray-worker-strixhalo b/dockerfiles/Dockerfile.ray-worker-strixhalo index 62b4604..fbfade9 100644 --- a/dockerfiles/Dockerfile.ray-worker-strixhalo +++ b/dockerfiles/Dockerfile.ray-worker-strixhalo @@ -53,6 +53,7 @@ RUN --mount=type=cache,target=/var/cache/apt,sharing=locked \ libdrm2 \ libdrm-amdgpu1 \ kmod \ + libopenmpi3 \ && rm -rf /var/lib/apt/lists/* # Install uv for fast Python package management (ADR-0014)