Skip to content

debug feature PYTORCH_CUDA_MEMORY_CACHING_MEMSET_ZEROS #2699

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
merged 1 commit into from
Oct 7, 2025
Merged

debug feature PYTORCH_CUDA_MEMORY_CACHING_MEMSET_ZEROS #2699

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
22 changes: 22 additions & 0 deletions c10/cuda/CUDACachingAllocator.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3261,6 +3261,15 @@ class DeviceCachingAllocator {
}
};

static bool zeroAllocations() {
static auto has_cuda_env =
c10::utils::check_env("PYTORCH_CUDA_MEMORY_CACHING_MEMSET_ZEROS") == true;
static auto has_rocm_env =
c10::utils::check_env("PYTORCH_HIP_MEMORY_CACHING_MEMSET_ZEROS") == true;
static bool zeros = has_cuda_env || has_rocm_env;
return zeros;
}

// Returns whether to force all allocations to bypass the caching allocator and
// go straight to cudaMalloc. This setting is useful when debugging GPU memory
// errors, since the caching allocator foils cuda-memcheck.
Expand Down Expand Up @@ -3652,6 +3661,10 @@ class NativeCachingAllocator : public CUDAAllocator {
TORCH_SDT_WITH_SEMAPHORE(malloc, devPtr, device, size, stream.id());
}

if (zeroAllocations()) {
C10_CUDA_CHECK(cudaMemsetAsync(devPtr, 0, size, stream));
}

return {devPtr, devPtr, deleteFunc, Device(DeviceType::CUDA, device)};
}
DeleterFnPtr raw_deleter() const override {
Expand Down Expand Up @@ -3734,6 +3747,12 @@ class NativeCachingAllocator : public CUDAAllocator {
C10_CUDA_CHECK(c10::cuda::GetDevice(&device));
malloc(&r, device, nbytes, cuda::getCurrentCUDAStream(device));
}
if (zeroAllocations()) {
c10::DeviceIndex device = 0;
C10_CUDA_CHECK(c10::cuda::GetDevice(&device));
C10_CUDA_CHECK(
cudaMemsetAsync(r, 0, nbytes, cuda::getCurrentCUDAStream(device)));
}
return r;
}

Expand All @@ -3749,6 +3768,9 @@ class NativeCachingAllocator : public CUDAAllocator {
C10_CUDA_CHECK(c10::cuda::GetDevice(&device));
malloc(&r, device, nbytes, stream);
}
if (zeroAllocations()) {
C10_CUDA_CHECK(cudaMemsetAsync(r, 0, nbytes, stream));
}
return r;
}

Expand Down