From 74944b1696df6344a56f72e2dbbbae3111abdc40 Mon Sep 17 00:00:00 2001
From: jgrusewski <jgrusewski@localhost>
Date: Fri, 27 Feb 2026 12:16:25 +0100
Subject: [PATCH] fix(ci): keep CUDA stubs when GPU not available

Without nvidia RuntimeClass, there's no real libcuda.so in the
container. The before_script was stripping CUDA stubs, leaving test
binaries (like backtesting) unable to load libcuda.so.1.

Now nvidia-smi is checked first: if GPU present, strip stubs (real
CUDA from RuntimeClass). If no GPU, keep stubs so CUDA-linked test
binaries can load. CUDA compute cap defaults to 89 (L4).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 .gitlab-ci.yml | 40 ++++++++++++++++++++++++----------------
 1 file changed, 24 insertions(+), 16 deletions(-)

diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
index 2223f5712..a6a7dd355 100644
--- a/.gitlab-ci.yml
+++ b/.gitlab-ci.yml
@@ -206,16 +206,20 @@ test:
     - name: redis:7-alpine
       alias: redis
   script:
-    # Remove CUDA stubs from LD_LIBRARY_PATH — nvidia RuntimeClass provides real libcuda.so
-    - export LD_LIBRARY_PATH=$(echo "$LD_LIBRARY_PATH" | tr ':' '\n' | grep -v stubs | tr '\n' ':' | sed 's/:$//')
-    - nvidia-smi || echo "nvidia-smi not available (no GPU RuntimeClass) — compile-only mode"
-    # Detect GPU compute capability and partition sccache per architecture
-    # Falls back to 89 (L4) when nvidia-smi unavailable (compile jobs don't need GPU)
-    - export CUDA_COMPUTE_CAP=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader 2>/dev/null | head -1 | tr -d '.' || echo "89")
-    - export CUDA_COMPUTE_CAP=${CUDA_COMPUTE_CAP:-89}
-    - echo "CUDA_COMPUTE_CAP=$CUDA_COMPUTE_CAP"
-    - export SCCACHE_DIR="/mnt/sccache/sm_${CUDA_COMPUTE_CAP}"
-    - mkdir -p "$SCCACHE_DIR"
+    # Detect GPU: if nvidia-smi works, strip stubs (real libcuda.so from RuntimeClass)
+    # If no GPU, keep stubs so CUDA-linked test binaries can at least load
+    - |
+      if nvidia-smi 2>/dev/null; then
+        echo "GPU detected — stripping CUDA stubs from LD_LIBRARY_PATH"
+        export LD_LIBRARY_PATH=$(echo "$LD_LIBRARY_PATH" | tr ':' '\n' | grep -v stubs | tr '\n' ':' | sed 's/:$//')
+        export CUDA_COMPUTE_CAP=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader | head -1 | tr -d '.')
+      else
+        echo "No GPU detected — keeping CUDA stubs for test binary loading"
+        export CUDA_COMPUTE_CAP=89
+      fi
+      echo "CUDA_COMPUTE_CAP=$CUDA_COMPUTE_CAP"
+      export SCCACHE_DIR="/mnt/sccache/sm_${CUDA_COMPUTE_CAP}"
+      mkdir -p "$SCCACHE_DIR"
     # Wait for Redis sidecar (K8s executor doesn't wait for service readiness)
     - |
       for i in $(seq 1 30); do
@@ -251,12 +255,16 @@ compile-services:
     - if: $CI_PIPELINE_SOURCE == "web" || $CI_PIPELINE_SOURCE == "api"
   script:
     # Detect GPU compute capability and partition sccache per architecture
-    # Falls back to 89 (L4) when nvidia-smi unavailable (compile doesn't need GPU)
-    - export CUDA_COMPUTE_CAP=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader 2>/dev/null | head -1 | tr -d '.' || echo "89")
-    - export CUDA_COMPUTE_CAP=${CUDA_COMPUTE_CAP:-89}
-    - echo "CUDA_COMPUTE_CAP=$CUDA_COMPUTE_CAP"
-    - export SCCACHE_DIR="/mnt/sccache/sm_${CUDA_COMPUTE_CAP}"
-    - mkdir -p "$SCCACHE_DIR"
+    - |
+      if nvidia-smi 2>/dev/null; then
+        export CUDA_COMPUTE_CAP=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader | head -1 | tr -d '.')
+      else
+        echo "No GPU — keeping CUDA stubs, defaulting compute cap to 89 (L4)"
+        export CUDA_COMPUTE_CAP=89
+      fi
+      echo "CUDA_COMPUTE_CAP=$CUDA_COMPUTE_CAP"
+      export SCCACHE_DIR="/mnt/sccache/sm_${CUDA_COMPUTE_CAP}"
+      mkdir -p "$SCCACHE_DIR"
     - sccache --zero-stats || true
     # Select build profile: dev-release (fast, thin LTO) or release (full, fat LTO)
     - PROFILE=${DEV_RELEASE:+dev-release}; PROFILE=${PROFILE:-release}