• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1.. _mps_environment_variables:
2
3MPS Environment Variables
4==========================
5
6**PyTorch Environment Variables**
7
8.. list-table::
9  :header-rows: 1
10
11  * - Variable
12    - Description
13  * - ``PYTORCH_DEBUG_MPS_ALLOCATOR``
14    - If set to ``1``, set allocator logging level to verbose.
15  * - ``PYTORCH_MPS_LOG_PROFILE_INFO``
16    - Set log options bitmask to ``MPSProfiler``. See ``LogOptions`` enum in `aten/src/ATen/mps/MPSProfiler.h` for available options.
17  * - ``PYTORCH_MPS_TRACE_SIGNPOSTS``
18    - Set profile and signpost bitmasks to ``MPSProfiler``. See ``ProfileOptions`` and ``SignpostTypes`` enums in `aten/src/ATen/mps/MPSProfiler.h` for available options.
19  * - ``PYTORCH_MPS_HIGH_WATERMARK_RATIO``
20    - High watermark ratio for MPS allocator. By default, it is set to 1.7.
21  * - ``PYTORCH_MPS_LOW_WATERMARK_RATIO``
22    - Low watermark ratio for MPS allocator. By default, it is set to 1.4 if the memory is unified and set to 1.0 if the memory is discrete.
23  * - ``PYTORCH_MPS_FAST_MATH``
24    - If set to ``1``, enable fast math for MPS metal kernels. See section 1.6.3 in https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf for precision implications.
25  * - ``PYTORCH_MPS_PREFER_METAL``
26    - If set to ``1``, force using metal kernels instead of using MPS Graph APIs. For now this is only used for matmul op.
27  * - ``PYTORCH_ENABLE_MPS_FALLBACK``
28    - If set to ``1``, full back operations to CPU when MPS does not support them.
29
30.. note::
31
32    **high watermark ratio** is a hard limit for the total allowed allocations
33
34    - `0.0` : disables high watermark limit (may cause system failure if system-wide OOM occurs)
35    - `1.0` : recommended maximum allocation size (i.e., device.recommendedMaxWorkingSetSize)
36    - `>1.0`: allows limits beyond the device.recommendedMaxWorkingSetSize
37
38    e.g., value 0.95 means we allocate up to 95% of recommended maximum
39    allocation size; beyond that, the allocations would fail with OOM error.
40
41    **low watermark ratio** is a soft limit to attempt limiting memory allocations up to the lower watermark
42    level by garbage collection or committing command buffers more frequently (a.k.a, adaptive commit).
43    Value between 0 to m_high_watermark_ratio (setting 0.0 disables adaptive commit and garbage collection)
44    e.g., value 0.9 means we 'attempt' to limit allocations up to 90% of recommended maximum
45    allocation size.