From bf9e05747a0a4c1d352fedeb526bcefd6e26da40 Mon Sep 17 00:00:00 2001 From: Lin Yuan Date: Tue, 6 Nov 2018 12:19:28 -0800 Subject: [PATCH] Document the newly added env variable (#13049) * add env variable to choose deterministic cudnn alg * set default value to false * fix build failure in Windows GPU * revert the previous change * only check determinism in CUDNN 7.x release * Add cudnn version check * fix lint error * document env variable MXNET_ENFORCE_DETERMINISM * use cudnnGet instead of cudnnFind when determinism required * Revert "use cudnnGet instead of cudnnFind when determinism required" This reverts commit d1bdf0f38f50b8c499f22ae1d50770b819f27678. --- docs/faq/env_var.md | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/docs/faq/env_var.md b/docs/faq/env_var.md index 92cf4931b04b..e373377ee8de 100644 --- a/docs/faq/env_var.md +++ b/docs/faq/env_var.md @@ -196,6 +196,12 @@ When USE_PROFILER is enabled in Makefile or CMake, the following environments ca - Flag to enable or disable MKLDNN accelerator. On by default. - Only applies to mxnet that has been compiled with MKLDNN (```pip install mxnet-mkl``` or built from source with ```USE_MKLDNN=1```) +* MXNET_ENFORCE_DETERMINISM + - Values: 0(false) or 1(true) ```(default=0)``` + - If set to true, MXNet will only use deterministic algorithms in forward and backward computation. + If no such algorithm exists given other constraints, MXNet will error out. This variable affects the choice + of CUDNN convolution algorithms. Please see [CUDNN developer guide](https://docs.nvidia.com/deeplearning/sdk/cudnn-developer-guide/index.html) for more details. + Settings for Minimum Memory Usage --------------------------------- - Make sure ```min(MXNET_EXEC_NUM_TEMP, MXNET_GPU_WORKER_NTHREADS) = 1```