experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Disable use of the new GpuVirtualMemAllocator in BfcAllocator.

Browse Source 
We've obsered some OOMing with the new allocator; will re-enable with a follow-up patch that resolves these issues.

PiperOrigin-RevId: 355209467
Change-Id: I23e44e912d0557d22aaac4a64fbe3481c77b97c6
This commit is contained in:
A. Unique TensorFlower 2021-02-02 11:15:21 -08:00 committed by TensorFlower Gardener
parent 36521b5bd4
commit 89c59719cf
1 changed files with 3 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
tensorflow/core/common_runtime/gpu/gpu_process_state.cc
View File

@ -88,7 +88,9 @@ static SubAllocator* CreateSubAllocator(
platform_gpu_id) platform_gpu_id)
.ValueOrDie(); .ValueOrDie();
#if defined(GOOGLE_CUDA) && CUDA_VERSION >= 10020 // FIXME(imintz): Observed OOM issues when using the virtual memory
// allocators. This should be reenabled when resolved.
#if 0 && defined(GOOGLE_CUDA) && CUDA_VERSION >= 10020
// Use the old allocator when unified memory is required. // Use the old allocator when unified memory is required.
// TODO(imintz): Remove the cuMemAlloc capability of this allocator. // TODO(imintz): Remove the cuMemAlloc capability of this allocator.
if (options.per_process_gpu_memory_fraction() > 1.0 || if (options.per_process_gpu_memory_fraction() > 1.0 ||