Cannot load libmkl_intel_thread.dylib

[crondonm]

Hi, I'm new to Julia and I am still learning. I am on MacOS 10.13 (High Sierra) and using Julia 0.6, and installed anaconda3 with python 3.6.2. Every time I run PyPlot like this:
using PyPlot
plot(rand(10))

I got the following error:

Intel MKL FATAL ERROR: Cannot load libmkl_intel_thread.dylib
Julia has stopped: 2

I have tried everything. I uninstalled Julia and Python, I switched back and forth between the private Conda version from PyCall and Python 3.6. Moreover, I plot directly from Python using matlibplot it works fine. The problem seems to be Julia. I ran out of ideas.

Any help is appreciated.

Carlos

[iamed2]

@willtebbutt is also experiencing this, with Python 2.7 installed with Conda.jl, and El Capitan.

[stevengj]

Possibly a conflict between the BLAS libraries linked to Julia and those linked to NumPy (which is used by Matplotlib). Are you using the standard Julia binary, built with OpenBLAS, or a Julia binary linked with MKL?

[iamed2]

Will was using the binaries (OpenBLAS). We're currently building Julia from source on his machine to see if we get the same problem.

[jonalm]

I'm having the same issue on OSX Sierra.

julia> using PyPlot
julia> plot(rand(10))
Intel MKL FATAL ERROR: Cannot load libmkl_intel_thread.dylib.

For what its worth:

My julia version is Version 0.6.0 (2017-06-19 13:05 UTC), installed from the v0.6.0 branch, using make -j 4, and I believe the python version PyPlot tries to use, is whatever PyCall installs by default.

julia> Pkg.build("PyPlot")
INFO: Building Conda
INFO: Building PyCall
INFO: PyCall is using /Users/jon.alm.eriksen/.julia/v0.6/Conda/deps/usr/bin/python (Python 2.7.13) at /Users/jon.alm.eriksen/.julia/v0.6/Conda/deps/usr/bin/python, libpython = /Users/jon.alm.eriksen/.julia/v0.6/Conda/deps/usr/lib/libpython2.7
INFO: /Users/jon.alm.eriksen/.julia/v0.6/PyCall/deps/deps.jl has not changed
INFO: /Users/jon.alm.eriksen/.julia/v0.6/PyCall/deps/PYTHON has not changed

[stevengj]

Is there a libmkl_intel_thread.dylib anywhere in ~/.julia/v0.6/Conda/deps/usr? Maybe the anaconda packagers didn't link things correctly so that it is just not finding the path?

[jonalm]

@stevengj I don't know if this is expected? It looks like there are three of them.

~/opt/julia > find ~/.julia/v0.6/Conda/deps/usr -name "libmkl_intel_thread.dylib"
/Users/jon.alm.eriksen/.julia/v0.6/Conda/deps/usr/lib/libmkl_intel_thread.dylib
/Users/jon.alm.eriksen/.julia/v0.6/Conda/deps/usr/pkgs/mkl-2017.0.1-0/lib/libmkl_intel_thread.dylib
/Users/jon.alm.eriksen/.julia/v0.6/Conda/deps/usr/pkgs/mkl-2018.0.0-hc285769_4/lib/libmkl_intel_thread.dylib

I don't know how this linking work, but is there someway to manually point to the dylib file?

Curiously, this seems to work:

~/opt/julia > /Users/jon.alm.eriksen/.julia/v0.6/Conda/deps/usr/bin/python
Python 2.7.13 |Continuum Analytics, Inc.| (default, Dec 20 2016, 23:05:08)
[GCC 4.2.1 Compatible Apple LLVM 6.0 (clang-600.0.57)] on darwin
Type "help", "copyright", "credits" or "license" for more information.
Anaconda is brought to you by Continuum Analytics.
Please check out: http://continuum.io/thanks and https://anaconda.org
>>> import matplotlib
>>> import matplotlib.pyplot as plt
>>> plt.plot([1,2,3])
[<matplotlib.lines.Line2D object at 0x112375ad0>]

[stevengj]

Setting LD_LIBRARY_PATH explicitly works around the problem, i.e.export LD_LIBRARY_PATH=$HOME/.julia/v0.6/Conda/deps/usr/lib before launching Julia, and then it works for me.

This seems like a problem with Anaconda. The library paths are set up correctly if you launch the python executable, apparently, but not if you load libpython from another program (as we do in Julia).

[willtebbutt]

This solution also works for me. Thanks!

[jonalm]

Works for me too. Greatly appreciated suggestion @stevengj Thanks.

[crondonm]

It also works for me. Thank you very much!

[flamesofmoon]

Thank you very much @stevengj Works for me as well. I have struggled for two days!

[flamesofmoon]

@stevengj do you know how to set that forever? Cause every time I run Julia I have to type that command which is troublesome. Thanks a lot!

[stevengj]

@flamesofmoon , you can put it in your ~/.profile file.

Hopefully, we'll soon have better fix. If nothing else, as I mentioned in ContinuumIO/anaconda-issues#6423, I can add manual dlopen calls in PyPlot and PyCall. But hopefully we can get this fixed in Anaconda.

[flamesofmoon]

@stevengj Thank you! I still have the following questions:

But I don't know what it means by putting it in some designated directory. Also, I don't have the directory ~/.profile. Could you please give me the bash command cause I just started to use terminal.

[stevengj]

echo "export LD_LIBRARY_PATH=$HOME/.julia/v0.6/Conda/deps/usr/lib" >> ~/.profile

then the next time you re-launch the terminal it will be set.

[flamesofmoon]

@stevengj Thank you for the reply!

I don't know why, but this time your command doesn't work on my laptop. With or without resetting the terminal, I get Intel MKL FATAL ERROR: Cannot load libmkl_intel_thread.dylib.

[stevengj]

A better workaround, which hopefully will get incorporated into Anaconda, is to run:

run(`install_name_tool -change @rpath/libiomp5.dylib @loader_path/libiomp5.dylib $(Pkg.dir("Conda", "deps/usr/lib/libmkl_intel_thread.dylib"))`)

You only need to do this once (until Conda upgrades MKL).

[anitescu]

@flamesofmoon : do you per chance have a MAC? Since El Capitan LD_LIBRARY_PATH cannot be set.

[mingwandroid]

@anitescu, conda does not require DYLD_LIBRARY_PATH or DYLD_FALLBACK_LIBRARY_PATH to be set, does julia? From what I've seen, it does not. You can set those variables anyway provided you disable SIP.

[stevengj]

No, Julia does not require DYLD_LIBRARY_PATH or similar.

In any case, I would recommend using the install_name_tool -change method rather than setting environment variables as a workaround until the Anaconda patch is merged into the main distro.

[anitescu]

@mingwandroid : thanks; did not mention that I hoped to not need to disable SIP. I am a bit unclear about how it deals with certain packages, such as CPLEX which itself does require it; but do not understand it enough.

@stevengj : there was this issue
issue to which I believe you also replied, maybe I misunderstood the proposed solution.

I ended up defining the Libdl PATH and process.env.LD_ for atom by hand for the previous issue.

[flamesofmoon]

@stevengj Well, I still can't work around that. Got the same intel MKL Fatal ERROR.

Anyway, your original idea is good enough. At least I can plot now!

[zashwood]

Hi @stevengj Thanks for figuring this out. Running export LD_Library_Path... fixed the Intel MKL FATAL ERROR: Cannot load libmkl_intel_thread.dylib problem for me. However I have tried to run
install_name_tool -change @rpath/libiomp5.dylib @loader_path/libiomp5.dylib $(Pkg.dir("Conda", "deps/usr/lib/libmkl_intel_thread.dylib"))
and I got back:

-bash: command substitution: line 1: syntax error near unexpected token `"Conda",'
-bash: command substitution: line 1: `Pkg.dir("Conda", "deps/usr/lib/libmkl_intel_thread.dylib")'

Do you know what I am doing wrong? Thanks.

[stevengj]

@zashwood, $(Pkg.dir(...)) is a Julia thing — it won't work in bash. That's why I said to run it (in Julia) via run(`install_name_tool ....`) above.

[stevengj]

Should be fixed with the latest Anaconda release.

[fufjvnvnf]

I don't seem to have installed anaconda on my computer. The workaround works in the REPL but does not work in jupyter notebook for me.

[rveltz]

Hi,

I am seeing the same problem on osx. Is it fixed?

[stevengj]

The original problem was fixed more than a year ago. I just tried updating to the latest Conda on my Mac laptop and it continues to work fine. So maybe you are using an ancient Anaconda installation, @rveltz, and need to update.

[rveltz]

OK, I am seeing another error. I'll open an issue.

Thank you.

[milankl]

This issue came up today despite a fresh Julia and following the

julia> ENV["PYTHON"] = ""
julia> Pkg.build("PyCall")

guidline. However, adding this

Setting LD_LIBRARY_PATH explicitly works around the problem, i.e.export LD_LIBRARY_PATH=$HOME/.julia/v0.6/Conda/deps/usr/lib before launching Julia, and then it works for me.

to .zshrc in macos 10.15 solved the issue in the REPL but not inside Atom. Any more clues how this is caused and whether there is a more automatic way around it?

[wizofe]

I still have the same problem with:

• MacOS Catalina v.10.15.5
• Julia 1.5.0-rc1.0

what I get is:

julia> using PyPlot
[ Info: Precompiling PyPlot [d330b81b-6aea-500a-939a-2ce795aea3ee]
INTEL MKL ERROR: dlopen(local/miniconda3/envs/cvnd3/lib/libmkl_intel_thread.dylib, 9): Library not loaded: @rpath/libiomp5.dylib
  Referenced from: local/miniconda3/envs/cvnd3/lib/libmkl_intel_thread.dylib
  Reason: image not found.
Intel MKL FATAL ERROR: Cannot load libmkl_intel_thread.dylib.

[milankl]

Setting LD_LIBRARY_PATH explicitly works around the problem, i.e.export LD_LIBRARY_PATH=$HOME/.julia/v0.6/Conda/deps/usr/lib before launching Julia, and then it works for me.

to .zshrc in macos 10.15 solved the issue in the REPL but not inside Atom. Any more clues how this is caused and whether there is a more automatic way around it?

To be precise, I added

export LD_LIBRARY_PATH=/Users/milan/.julia/conda/3/lib:$LD_LIBRARY_PATH

to my .zshrc. Then in Atom you can change in settings the julia-client "Julia Options" from "Basic" to "External Terminal" for example.

[rokapre]

I am having this same issue but with just about any Python package. Following the works in the terminal directly but not in the IDEs Atom nor VSCode.

[tlienart]

I stumbled upon this issue today on OSX with Julia 1.7, the LD_LIBRARY_PATH trick didn't work but the following did:

using Conda
Conda.rm("mkl")
Conda.add("nomkl")

I'm not sure whether both are needed but it fixed the issue for me.

[xrisk]

I’m still facing this issue.

Julia 1.6.0
macOS Big Sur 11.2.3

The LD_LIBRARY_PATH trick works for me. But I’m curious to understand why this happens in the first place. Am I using Julia incorrectly?

[George9000]

On Julia 1.6.1, macOS 10.15.7

Solution offered by tlienart worked. Thanks!

[willcoxe]

Solution provided by @tlienart also worked for me with
Julia 1.5.3
OSX Catalina 10.15.7
(Atom 1.56.0)

[TheCedarPrince]

How peculiar. I can also confirm that @tlienart's solution worked on Julia 1.6 with my mac OSX Catalina 10.15.7. Not sure why.

[KarroyWang001]

Thanks very much to @tlienart, his solution helps me a lot for my Julia 1.6.1. Still, it confuses me why it happens.

[logankilpatrick]

@tlienart is going to become famous for this one, thanks for the help! This fixed my issue on my M1 MacBook Pro with Julia 1.6

[johnnychen94]

To be precise, I added

export LD_LIBRARY_PATH=/Users/milan/.julia/conda/3/lib:$LD_LIBRARY_PATH
to my .zshrc. Then in Atom you can change in settings the julia-client "Julia Options" from "Basic" to "External Terminal" for example.

For custom conda env settings

ENV["CONDA_JL_HOME"] = "/Users/jc/opt/miniconda3/envs/conda_jl"
ENV["PYTHON"] = "" 

Appending LD_LIBRARY_PATH with /Users/jc/opt/miniconda3/envs/conda_jl/lib instead of /Users/jc/.julia/conda/3/lib

[astrobc1]

For anyone else stuck with this error, I ended up having to manually copy the file libiomp5.dylib from my anaconda3 folder to the path /Applications/Julia-1.7.app/Contents/Resources/julia/lib/. I tried copying the file libmkl_intel_thread.1.dylib as well into appropriate folders, modifed the LD_LIBRARY_PATH variable, but this was ultimately the only thing that worked.

In my case, this issue started happening after I started using a custom Python installation (the default one I usually use on my machine, not the hidden one inside ~/.julia/).

MacOS Monterey 12.3.1, Julia 1.7.2