"import tensorflow.compat.v1 as tf" exits with "Illegal instruction" about tensorflow-directml HOT 10 CLOSED

Thanks.

This looks like this is a mistake on our end: we're inadvertently compiling with AVX2 instructions for Linux builds (but not Windows builds). We'll fix this in a future preview package!

from tensorflow-directml.

Hi, what CPU are you running this on? And just to confirm, are you running a 64-bit distro in WSL?

from tensorflow-directml.

Hi,

CPU: Intel Core i7-3770S
And yes, I am on 64-bits distro, from the container,

cht@DESKTOP-IHJTV93:~$ uname -i
x86_64

from tensorflow-directml.

same here

Linux version 4.19.121-microsoft-standard

Edition	Windows 10 Pro Insider Preview
Version	2004
Installed on	‎04/‎07/‎2020
OS build	20161.1000
Experience	Windows Feature Experience Pack 120.16201.0.0

Processor	Intel(R) Core(TM) i7-3612QM CPU @ 2.10GHz   2.10 GHz
Installed RAM	16.0 GB (15.9 GB usable)
System type	64-bit operating system, x64-based processor

from tensorflow-directml.

@charlie-ht and @ferdavs : which Linux distro are you testing with?

It does look like both of your processors support AVX (though not AVX2, which should be OK). Just for a sanity check, does the official version of tensorflow work for you (i.e. pip install tensorflow) in the same WSL distro?

from tensorflow-directml.

which Linux distro are you testing with?
mine is Ubuntu-18.04

1. python -v venv venv
2. pip install tensorflow installed tf 2.2.0
   tested with

import tensorflow.compat.v1 as tf

tf.enable_eager_execution(tf.ConfigProto(log_device_placement=True))

print(tf.add([1.0, 2.0], [3.0, 4.0]))

2020-07-07 23:13:39.823067: W tensorflow/stream_executor/platform/default/dso_loader.cc:55] Could not load dynamic library 'libcuda.so.1'; dlerror: libcuda.so.1: cannot open shared object file: No such file or directory
2020-07-07 23:13:39.823277: E tensorflow/stream_executor/cuda/cuda_driver.cc:313] failed call to cuInit: UNKNOWN ERROR (303)
2020-07-07 23:13:39.823461: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:156] kernel driver does not appear to be running on this host (PC): /proc/driver/nvidia/version does not exist
2020-07-07 23:13:39.950198: I tensorflow/core/platform/profile_utils/cpu_utils.cc:102] CPU Frequency: 2095235000 Hz
2020-07-07 23:13:39.954135: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x7f2818000b20 initialized for platform Host (this does not guarantee that XLA will be used). Devices:
2020-07-07 23:13:39.954222: I tensorflow/compiler/xla/service/service.cc:176]   StreamExecutor device (0): Host, Default Version
tf.Tensor([4. 6.], shape=(2,), dtype=float32)

from tensorflow-directml.

Thanks for the investigation @jstoecker, I can confirm the official tensorflow package works for me too 👍

from tensorflow-directml.

Quick update: this bug should be fixed in the new 1.15.4-dev201216 release. Try it out and let us know how it goes!

from tensorflow-directml.

Hi, I have tried with 1.15.4-dev201216 release in WSL2 and I am still getting the same error:

pip list |grep directml
tensorflow-directml 1.15.4.dev201216

lsb_release -a
No LSB modules are available.
Distributor ID: Ubuntu
Description: Ubuntu 20.04.2 LTS
Release: 20.04
Codename: focal

cat /proc/version
Linux version 5.4.91-microsoft-standard-WSL2 (oe-user@oe-host) (gcc version 9.3.0 (GCC)) #1 SMP Mon Jan 25 18:39:31 UTC 2021

windows OS version: 10.0.21327 N/D Compilación 21327

Processor model name : Intel64 Family 6 Model 30 Stepping 5 GenuineIntel ~2801 Mhz

from tensorflow-directml.

@jglara I believe your CPU may not support AVX, which is a requirement for the packages on pypi. According to this table your CPU appears to be from the Nehalem architecture (likely an i7-720QM from the clock speed you listed), which doesn't have AVX. You can check for AVX support using one of the recommendations here; for example, does typing grep avx /proc/cpuinfo into your WSL Ubuntu shell print anything? The official tensorflow packages on pypi also require AVX, so another way of checking is trying them (i.e. pip install tensorflow==1.15).

from tensorflow-directml.