Skip to content

[Common] Don't call CUDA from static destructors at exit #3733

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
ConnorBaker wants to merge 2 commits into
base: main
Choose a base branch
from
+13 −5
Draft

[Common] Don't call CUDA from static destructors at exit #3733

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
18e5b70
[Common] Keep cuda_driver::call's symbol cache alive through exit
ConnorBaker Oct 23, 2026
ec38bcc
[Common] Never destroy rtc::KernelManager or the fused-attention grap…
ConnorBaker Oct 24, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 3 additions & 1 deletion tests/cpp/util/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,4 +21,6 @@ target_link_libraries(test_util PUBLIC
target_compile_options(test_util PRIVATE -O2 -fopenmp)

include(GoogleTest)
gtest_discover_tests(test_util DISCOVERY_TIMEOUT 600)
# MALLOC_PERTURB_ makes glibc overwrite freed memory, so a destructor that reads a destroyed object
# at exit crashes the test process, which ctest reports as a failure, instead of passing silently.
gtest_discover_tests(test_util DISCOVERY_TIMEOUT 600 PROPERTIES ENVIRONMENT "MALLOC_PERTURB_=165")
4 changes: 3 additions & 1 deletion transformer_engine/common/fused_attn/graph_cache.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -130,7 +130,9 @@ std::shared_ptr<CacheEntry<GraphAndTensors>> cache_graph(GraphCache<GraphAndTens
// Create a cache for each (backend, pass) pair, and either get cached entry or build anew.
template <Backend kBackend, Pass kPass, auto kCreateGraphFn>
auto get_graph(const FusedAttnConfig &cfg, cudnnHandle_t handle) {
static GraphCache<decltype(kCreateGraphFn(cfg))> cache;
// Never destroyed: destroying cached cuDNN plans calls into CUDA, which must not be used during
// program termination.
static auto &cache = *new GraphCache<decltype(kCreateGraphFn(cfg))>;
cfg.check_derived();
return detail::cache_graph(cache, cfg.make_cache_key(kPass), kBackend, kPass, handle,
[&] { return kCreateGraphFn(cfg); });
Expand Down
6 changes: 4 additions & 2 deletions transformer_engine/common/util/cuda_driver.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,8 +40,10 @@ template <typename... ArgTs>
inline CUresult call(const char *symbol, ArgTs... args) {
using FuncT = CUresult(ArgTs...);

static std::unordered_map<std::string, void *> symbol_cache;
static std::mutex cache_mutex;
// Never destroyed, so a lookup never reads a destroyed map, whatever order statics are
// constructed and destroyed in.
static auto &symbol_cache = *new std::unordered_map<std::string, void *>();
static auto &cache_mutex = *new std::mutex();
FuncT *func;

{
Expand Down
4 changes: 3 additions & 1 deletion transformer_engine/common/util/rtc.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -153,7 +153,9 @@ int Kernel::occupancy_max_active_blocks_per_sm(int device_id, int block_size,

KernelManager& KernelManager::instance() {
NVTE_CHECK(is_enabled(), "NVRTC support is not enabled");
static KernelManager instance_;
// Never destroyed: ~Kernel unloads modules through the CUDA driver, and CUDA must not be used
// during program termination.
static KernelManager& instance_ = *new KernelManager;
return instance_;
}

Expand Down
Loading