diff options
| author | SsnL <tongzhou.wang.1994@gmail.com> | 2019-01-14 07:28:50 -0800 |
|---|---|---|
| committer | Facebook Github Bot <facebook-github-bot@users.noreply.github.com> | 2019-01-14 07:31:51 -0800 |
| commit | 300dcc3b96b2b8672570c4c9e6038cb5cff75afd (patch) | |
| tree | b0a64eebc9a9730fccf50a19b9a4b4739808666f /docs | |
| parent | 7c08f1083e56ce7e9799a15b36947c46052f0058 (diff) | |
| download | pytorch-300dcc3b96b2b8672570c4c9e6038cb5cff75afd.tar.gz pytorch-300dcc3b96b2b8672570c4c9e6038cb5cff75afd.tar.bz2 pytorch-300dcc3b96b2b8672570c4c9e6038cb5cff75afd.zip | |
Add cuda.reset_max_memory_* (#15985)
Summary:
Addresses #15968
Pull Request resolved: https://github.com/pytorch/pytorch/pull/15985
Differential Revision: D13649916
Pulled By: soumith
fbshipit-source-id: a207aea5709a79dba7a6fc541d0a70103f49efff
Diffstat (limited to 'docs')
| -rw-r--r-- | docs/source/cuda.rst | 2 | ||||
| -rw-r--r-- | docs/source/notes/cuda.rst | 8 |
2 files changed, 6 insertions, 4 deletions
diff --git a/docs/source/cuda.rst b/docs/source/cuda.rst index 6da20ce68e..462967461c 100644 --- a/docs/source/cuda.rst +++ b/docs/source/cuda.rst @@ -46,8 +46,10 @@ Memory management .. autofunction:: empty_cache .. autofunction:: memory_allocated .. autofunction:: max_memory_allocated +.. autofunction:: reset_max_memory_allocated .. autofunction:: memory_cached .. autofunction:: max_memory_cached +.. autofunction:: reset_max_memory_cached NVIDIA Tools Extension (NVTX) ----------------------------- diff --git a/docs/source/notes/cuda.rst b/docs/source/notes/cuda.rst index 212f68e694..7cf2fe6ad3 100644 --- a/docs/source/notes/cuda.rst +++ b/docs/source/notes/cuda.rst @@ -74,9 +74,9 @@ You can force synchronous computation by setting environment variable operation is actually executed, so the stack trace does not show where it was requested.) -As an exception, several functions such as :meth:`~torch.Tensor.to` and -:meth:`~torch.Tensor.copy_` admit an explicit :attr:`non_blocking` argument, -which lets the caller bypass synchronization when it is unnecessary. +As an exception, several functions such as :meth:`~torch.Tensor.to` and +:meth:`~torch.Tensor.copy_` admit an explicit :attr:`non_blocking` argument, +which lets the caller bypass synchronization when it is unnecessary. Another exception is CUDA streams, explained below. CUDA streams @@ -118,7 +118,7 @@ unused memory managed by the allocator will still show as if used in :meth:`~torch.cuda.max_memory_allocated` to monitor memory occupied by tensors, and use :meth:`~torch.cuda.memory_cached` and :meth:`~torch.cuda.max_memory_cached` to monitor memory managed by the caching -allocator. Calling :meth:`~torch.cuda.empty_cache` can release all **unused** +allocator. Calling :meth:`~torch.cuda.empty_cache` releases all **unused** cached memory from PyTorch so that those can be used by other GPU applications. However, the occupied GPU memory by tensors will not be freed so it can not increase the amount of GPU memory available for PyTorch. |