frozenleaves/pytorch
1
0
Fork 0
mirror of https://github.com/pytorch/pytorch.git synced 2025-10-21 05:34:18 +08:00
Code Issues Packages Projects Releases Wiki Activity

Removing per torch.compile audit. (#154572)

Browse Source 
Removing https://pytorch.org/docs/stable/torch.compiler_best_practices_for_backends.html per torch.compile audit

Pull Request resolved: https://github.com/pytorch/pytorch/pull/154572
Approved by: https://github.com/williamwen42, https://github.com/svekars
This commit is contained in:
Alanna Burke
2025-06-03 15:41:52 +00:00
committed by PyTorch MergeBot
parent 3685b10170
commit 250e9af4da
2 changed files with 0 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
docs/source/torch.compiler.rst
View File

@ -107,7 +107,6 @@ Read More
torch.compiler_dynamo_deepdive torch.compiler_dynamo_deepdive
torch.compiler_dynamic_shapes torch.compiler_dynamic_shapes
torch.compiler_nn_module torch.compiler_nn_module
torch.compiler_best_practices_for_backends
torch.compiler_cudagraph_trees torch.compiler_cudagraph_trees
torch.compiler_fake_tensor torch.compiler_fake_tensor

17
docs/source/torch.compiler_best_practices_for_backends.rst
View File

@ -1,17 +0,0 @@
Best Practices for Backends
===========================
x86 CPU
-------
Compiled workloads on modern x86 CPUs are usually optimized by Single Instruction Multiple Data (SIMD) instruction sets. SIMD is a typical parallel processing technique for high performance computing, such as deep learning model training and inference. With SIMD applied, each compute unit performs the same instruction with different allocated data at any given time slot. The most commonly deployed x86 instruction set architectures (ISAs) enabling SIMD include `AVX, AVX2, AVX-512 <https://en.wikipedia.org/wiki/Advanced_Vector_Extensions>`_ and `AMX <https://en.wikipedia.org/wiki/Advanced_Matrix_Extensions>`_.
You can check supported ISAs for your machine by using the `collect_env script <https://github.com/pytorch/pytorch/blob/main/torch/utils/collect_env.py>`_. As the script provides complete environment information for PyTorch, we can use ``grep`` to extract the line containing ISA information:
::
python collect_env.py | grep "a[(v|m)]x"
Normally, if AVX-512 is supported, instructions start with "avx512" (like ``avx512f``, ``avx512bw``, ``avx512_vnni``) should be observed. If AMX is supported, instructions start with "amx" (like ``amx_tile``, ``amx_bf16``, ``amx_int8``) should be observed.
Specifically, with a server having AMX instructions enabled, workloads performance can be further boosted by `leveraging AMX <https://pytorch.org/tutorials/recipes/amx.html>`_.