regular tests pass now having issues with sanitizer passes with driver issues. addresses this problem

Author: skyler-ruiter

.github/workflows/ci.yml

@@ -117,13 +117,28 @@ jobs:
           submodules: recursive
 
       - name: Configure (ASan preset)
-        run: cmake --preset asan
+        run: |
+          if [[ -f /etc/profile.d/lmod.sh ]]; then
+            source /etc/profile.d/lmod.sh
+          fi
+          if command -v module >/dev/null 2>&1; then
+            module purge
+            module load nvhpc/23.11/nvhpc-hpcx-cuda12
+          fi
+          cmake --preset asan
 
       - name: Build
         run: cmake --build --preset asan --parallel
 
       - name: Run tests (ASan)
         run: |
+          if [[ -f /etc/profile.d/lmod.sh ]]; then
+            source /etc/profile.d/lmod.sh
+          fi
+          if command -v module >/dev/null 2>&1; then
+            module purge
+            module load nvhpc/23.11/nvhpc-hpcx-cuda12
+          fi
           # ASan is baked into the shared libraries. The dynamic linker does not
           # guarantee it initializes before main(), so we must preload it first.
           # Without this, every test fails with:

CHANGELOG.md

@@ -84,6 +84,7 @@ Version numbers follow [Semantic Versioning](https://semver.org/).
 - `LorenzoStage` (the old fused predictor+quantizer) removed and replaced by `LorenzoQuantStage`; `LorenzoStage` now refers exclusively to the plain integer delta predictor
 
 ### Fixed
+- ASan: avoid a use-after-free in `CompressionDAG::addStage` by taking the stage name by value
 - CI: make CUDA module loading optional when lmod/module are unavailable so non-Jetstream runners do not fail early
 - vGPU compatibility: added fallback from `cudaMallocAsync`/`cudaFreeAsync` to `cudaMalloc`/`cudaFree` when memory pools are unavailable; `MemoryPool` gracefully degrades to regular malloc mode with warning log; fixes "operation not supported" errors on virtualized GPUs (e.g., Jetstream NVIDIA Virtual Compute Server)
 - vGPU stream synchronization: fallback code paths in `MemoryPool`, `DifferenceStage`, `RLEStage`, and `RZEStage` now synchronize streams before calling `cudaFree()` to prevent use-after-free race conditions when kernels are still using freed memory

include/pipeline/dag.h

@@ -93,7 +93,7 @@ class CompressionDAG {
     // ── Construction ──────────────────────────────────────────────────────────
 
     /** Add a stage and return its node for wiring dependencies. */
-    DAGNode* addStage(Stage* stage, const std::string& name = "");
+    DAGNode* addStage(Stage* stage, std::string name = "");
 
     /**
      * Add a dependency between two nodes, creating an intermediate buffer.

src/pipeline/dag.cpp

@@ -76,7 +76,7 @@ CompressionDAG::~CompressionDAG() {
     }
 }
 
-DAGNode* CompressionDAG::addStage(Stage* stage, const std::string& name) {
+DAGNode* CompressionDAG::addStage(Stage* stage, std::string name) {
     if (is_finalized_) {
         throw std::runtime_error("Cannot modify DAG after finalization");
     }
@@ -90,7 +90,11 @@ DAGNode* CompressionDAG::addStage(Stage* stage, const std::string& name) {
 
     auto* node = new DAGNode(stage);
     node->id = static_cast<int>(nodes_.size());
-    node->name = name.empty() ? "stage_" + std::to_string(node->id) : name;
+    if (name.empty()) {
+        node->name = "stage_" + std::to_string(node->id);
+    } else {
+        node->name = std::move(name);
+    }
 
     cudaError_t err = cudaEventCreate(&node->completion_event);
     if (err != cudaSuccess) {