本周我们精选出社区问答进行整理汇总,开发者在使用PaddlePaddle过程中遇到任何技术难题,都可以到PaddlePaddle公众号FAQ专栏上寻求解决方案,希望能帮助新用户在Linux安装过程中解答疑惑。
1.问题:cuda9.0需要安装哪一个版本的paddle,安装包在哪?
因此,pip install paddlepaddle-gpu即可。
可以参考安装说明文档:http://paddlepaddle.org/documentation/docs/zh/0.14.0/new_docs/beginners_guide/install/install_doc.html#linuxpaddlepaddle
2.问题:pip install paddlepaddle-gpu 安装fluid 版本报错
机器的CUDA信息如下:
按照官网安装:pip install paddlepaddle-gpu 执行import paddle.fluid as fluid 失败
奇怪的是,同样的环境下,上周运行成功,这周确运行失败,求解答
3.问题:CUDA driver version is insufficient
4.问题:安装CPU版本后训练主动abort,gdb显示Illegal instruction
*** Aborted at 1539697466 (unix time) try "date -d @1539697466" if you are using GNU date ***
PC: @ 0x0 (unknown)
*** SIGILL (@0x7fe3a27b7912) received by PID 13005 (TID 0x7fe4059d8700) from PID 18446744072140585234; stack trace: ***
@ 0x318b20f500 (unknown)
@ 0x7fe3a27b7912 paddle::framework::VisitDataType<>()
@ 0x7fe3a279f84f paddle::operators::math::set_constant_with_place<>()
@ 0x7fe3a1e50c21 paddle::operators::FillConstantOp::RunImpl()
@ 0x7fe3a27526bf paddle::framework::OperatorBase::Run()
@ 0x7fe3a1ca31ea paddle::framework::Executor::RunPreparedContext()
@ 0x7fe3a1ca3be0 paddle::framework::Executor::Run()
@ 0x7fe3a1bc9e7d _ZZN8pybind1112cpp_function10initializeIZN6paddle6pybindL13pybind11_initEvEUlRNS2_9framework8ExecutorERKNS4_11ProgramDescEPNS4_5ScopeEibbE63_vIS6_S9_SB_ibbEINS_4nameENS_9is_methodENS_7siblingEEEEvOT_PFT0_DpT1_EDpRKT2_ENUlRNS_6detail13function_callEE1_4_FUNEST_
@ 0x7fe3a1c14c24 pybind11::cpp_function::dispatcher()
@ 0x7fe405acf3e4 PyEval_EvalFrameEx
@ 0x7fe405ad0130 PyEval_EvalCodeEx
@ 0x7fe405ace4a1 PyEval_EvalFrameEx
@ 0x7fe405ad0130 PyEval_EvalCodeEx
@ 0x7fe405ace4a1 PyEval_EvalFrameEx
@ 0x7fe405ad0130 PyEval_EvalCodeEx
@ 0x7fe405a5c181 function_call
@ 0x7fe405a340f3 PyObject_Call
@ 0x7fe405accde7 PyEval_EvalFrameEx
@ 0x7fe405acec56 PyEval_EvalFrameEx
@ 0x7fe405ad0130 PyEval_EvalCodeEx
@ 0x7fe405a5c27d function_call
@ 0x7fe405a340f3 PyObject_Call
@ 0x7fe405accde7 PyEval_EvalFrameEx
@ 0x7fe405ad0130 PyEval_EvalCodeEx
@ 0x7fe405a5c181 function_call
@ 0x7fe405a340f3 PyObject_Call
@ 0x7fe405a46f7f instancemethod_call
@ 0x7fe405a340f3 PyObject_Call
@ 0x7fe405a8abd4 slot_tp_call
@ 0x7fe405a340f3 PyObject_Call
@ 0x7fe405acd887 PyEval_EvalFrameEx
@ 0x7fe405acec56 PyEval_EvalFrameEx
$ /path/to/gdb -iex "set auto-load safe-path /" -iex "set solib-search-path /path/to/gcc-4/lib" /path/to/python -c core.xxx
在gdb界面:
(gdb) disas
找到箭头所指的指令,例如:
0x00007f381ae4b90d <+3101>: test %r8,%r8
=> 0x00007f381ae4b912 <+3106>: vbroadcastss %xmm0,%ymm1
0x00007f381ae4b917 <+3111>: lea (%r12,%rdx,4),%rdi
然后google一下这个指令需要的指令集。上面例子中的带xmm和ymm操作数的vbroadcastss指令只在AVX2中支持
然后看下自己的CPU是否支持该指令集
cat /proc/cpuinfo | grep flags | uniq | grep avx --color
如果没有AVX2,就表示确实是指令集不支持引起的主动abort
http://www.paddlepaddle.org/documentation/docs/zh/1.0/beginners_guide/install/Start.html#paddlepaddle
5.问题:nvidia-docker运行镜像latest-gpu-cuda8.0-cudnn7: SIGILL
import paddle.fluid
*** Aborted at 1539682149 (unix time) try "date -d @1539682149" if you are using GNU date ***
PC: @ 0x0 (unknown)
*** SIGILL (@0x7f6ac6ea9436) received by PID 16 (TID 0x7f6b07bc7700) from PID 18446744072751846454; stack trace: ***
如果在这篇文章中没有得到您所遇到问题的解答时,请不要着急。我们将即可推出后续问题解答报道,敬请期待。