-
Notifications
You must be signed in to change notification settings - Fork 371
/
env.sh
25 lines (25 loc) · 1.3 KB
/
env.sh
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
set -x
#在LD_LIBRARY_PATH中添加cuda库的路径
export LD_LIBRARY_PATH=/home/work/cuda-10.1/lib64:$LD_LIBRARY_PATH
export LD_LIBRARY_PATH=/home/work/cuda-10.1/extras/CUPTI/lib64:$LD_LIBRARY_PATH
#在LD_LIBRARY_PATH中添加cudnn库的路径
export LD_LIBRARY_PATH=/home/work/cudnn/cudnn_v7.4/cuda/lib64:$LD_LIBRARY_PATH
#需要先下载NCCL,然后在LD_LIBRARY_PATH中添加NCCL库的路径
export LD_LIBRARY_PATH=/home/work/nccl/nccl2.4.2_cuda10.1/lib:$LD_LIBRARY_PATH
#如果FLAGS_sync_nccl_allreduce为1,则会在allreduce_op_handle中调用cudaStreamSynchronize(nccl_stream),这种模式在某些情况下可以获得更好的性能
export FLAGS_sync_nccl_allreduce=1
#表示分配的显存块占GPU总可用显存大小的比例,范围[0,1]
export FLAGS_fraction_of_gpu_memory_to_use=1
#选择要使用的GPU
export CUDA_VISIBLE_DEVICES=7
#表示是否使用垃圾回收策略来优化网络的内存使用,<0表示禁用,>=0表示启用
export FLAGS_eager_delete_tensor_gb=1.0
#是否使用快速垃圾回收策略
export FLAGS_fast_eager_deletion_mode=1
#垃圾回收策略释放变量的内存大小百分比,范围为[0.0, 1.0]
export FLAGS_memory_fraction_of_eager_deletion=1
#设置fluid路径
export PATH=fluid=/home/work/python/bin:$PATH
#设置python
alias python=/home/work/python/bin/python
set +x