Skip to content

Address teardown issue #638

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

Merged
Binyang2014 merged 3 commits into from
Sep 25, 2025
Merged

Address teardown issue #638

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
ab8d01e
fix teardown issue
Binyang2014 Sep 25, 2025
ea3b69e
fix hip issue
Binyang2014 Sep 25, 2025
3723548
fix comment
Binyang2014 Sep 25, 2025
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
20 changes: 16 additions & 4 deletions src/gpu_utils.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,17 +5,19 @@
#include <mscclpp/gpu.hpp>
#include <mscclpp/gpu_utils.hpp>

#include "debug.h"

static inline bool isCudaTeardownError(cudaError_t err) {
#if defined(__HIP_PLATFORM_AMD__)
return err == cudaErrorContextIsDestroyed || err == cudaErrorInvalidDevice;
#else // !defined(__HIP_PLATFORM_AMD__)
return err == cudaErrorCudartUnloading || err == cudaErrorContextIsDestroyed || err == cudaErrorInitializationError ||
err == cudaErrorInvalidDevice;
err == cudaErrorInvalidDevice || err == cudaErrorLaunchFailure;
#endif // !defined(__HIP_PLATFORM_AMD__)
}

static inline bool isCuTeardownError(CUresult r) {
return r == CUDA_ERROR_DEINITIALIZED || r == CUDA_ERROR_CONTEXT_IS_DESTROYED;
return r == CUDA_ERROR_DEINITIALIZED || r == CUDA_ERROR_CONTEXT_IS_DESTROYED || r == CUDA_ERROR_LAUNCH_FAILED;
}

#define MSCCLPP_CUDATHROW_IGNORE_TEARDOWN(cmd) \
Expand All @@ -36,6 +38,16 @@ static inline bool isCuTeardownError(CUresult r) {
} \
} while (false)

#define MSCCLPP_CUTHROW_IGNORE(cmd) \
do { \
CUresult __e = cmd; \
if (__e != CUDA_SUCCESS) { \
const char* errStr; \
cuGetErrorString(__e, &errStr); \
WARN("%s:%d Cuda failure %d '%s'", __FILE__, __LINE__, __e, errStr); \
} \
} while (false)

namespace mscclpp {

AvoidCudaGraphCaptureGuard::AvoidCudaGraphCaptureGuard() : mode_(cudaStreamCaptureModeRelaxed), active_(true) {
Expand Down Expand Up @@ -242,9 +254,9 @@ void gpuFreePhysical(void* ptr) {
MSCCLPP_CUTHROW_IGNORE_TEARDOWN(cuMemRetainAllocationHandle(&handle, ptr));
MSCCLPP_CUTHROW_IGNORE_TEARDOWN(cuMemRelease(handle));
MSCCLPP_CUTHROW_IGNORE_TEARDOWN(cuMemGetAddressRange(NULL, &size, (CUdeviceptr)ptr));
MSCCLPP_CUTHROW_IGNORE_TEARDOWN(cuMemUnmap((CUdeviceptr)ptr, size));
MSCCLPP_CUTHROW_IGNORE(cuMemUnmap((CUdeviceptr)ptr, size));
MSCCLPP_CUTHROW_IGNORE_TEARDOWN(cuMemRelease(handle));
MSCCLPP_CUTHROW_IGNORE_TEARDOWN(cuMemAddressFree((CUdeviceptr)ptr, size));
MSCCLPP_CUTHROW_IGNORE(cuMemAddressFree((CUdeviceptr)ptr, size));
}
#endif // CUDA_NVLS_API_AVAILABLE

Expand Down
10 changes: 6 additions & 4 deletions test/torch/memory_report.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -58,17 +58,19 @@ def main():
group_ranks = list(range(world_size))
if rank == 0:
print(f"Creating new_group with ranks={group_ranks}", flush=True)
grp = dist.new_group(ranks=group_ranks, backend=backend)
grp0 = dist.new_group(ranks=group_ranks, backend=backend)
x = torch.ones(nelems, device=local_rank, dtype=torch.float32) * (rank + 1)
dist.all_reduce(x, op=dist.ReduceOp.SUM, group=grp)
dist.all_reduce(x, op=dist.ReduceOp.SUM, group=grp0)

grp = dist.new_group(ranks=list(range(world_size)), backend=backend)
grp1 = dist.new_group(ranks=list(range(world_size)), backend=backend)
x = torch.ones(nelems, device=local_rank, dtype=torch.float32) * (rank + 1)
dist.all_reduce(x, op=dist.ReduceOp.SUM, group=grp)
dist.all_reduce(x, op=dist.ReduceOp.SUM, group=grp1)

dist.barrier()

print(memory_report(local_rank))
dist.destroy_process_group(grp0)
dist.destroy_process_group(grp1)
dist.destroy_process_group()


Expand Down
Loading