Open FlyingFathead opened 3 years ago
Sorry for a late response! You may perhaps want to try disabling the thread support altogether (edit config.h and either undef or comment out #define HAVE_THREADS). You may even try to disable the SDL support there.
Sorry for a late response! You may perhaps want to try disabling the thread support altogether (edit config.h and either undef or comment out #define HAVE_THREADS). You may even try to disable the SDL support there.
Thanks a lot for the answer; no worries for it being late, there's no such thing on this field :o)
Is the threading feature btw in any way related to Tensorflow's AVX/AVX2 (Advanced Vector Extensions) utilization? That might indeed be a problem in some of the endpoints I've tried it out on. I'll try the tip you gave and see how it goes. Thank you once more. Good luck.
Ahoy there!
First of all, congratulations on the interesting concept.
Second of all, well, where do I start? I must apologize in advance that C is not my native language, hence I have no clue where to start from with some of the compile warnings that I got...
The
Makefile
could use make-over, pardon the pun. Unless the-pthread
link flag is manually added into theMakefile
, the compile will fail on all the Linux-based setups that I tried to compile vzgpt on.In other words, the following happens if the
-pthread
flag is not added into theMakefile
(a.k.a. by default):Had the same problem with a Linux that was running glibc 2.3.3.
Just a heads up to anyone actually attempting to compile and run this, you need to link '-pthread' at the end of
vzgpt
's parameters inside theMakefile
for the thing to compile on default settings in the first place, like so:(Notice the
-pthread
flag at the end of the line. Add that to theMakefile
on that line, save andmake
.)Speaking of
make
, here are the warnings I got during compile on a non-AVX/AVX2 x86/64 system with-pthread
flag enabled (Ubuntu clang version 12.0.0-3ubuntu1~21.04.1 / Target: x86_64-pc-linux-gnu / Thread model: posix).Any idea on these?
The build did finish without an error though, which brings me to my penultimate question... How do I get vzgpt to not to crash on startup? :-D
The checkpoint conversion with
dumpckpt.py
from my pre-existing GPT-2 model checkpoint goes OK, but upon launching the program with the data set, I end up getting a segfault+coredump during the SDL UI startup after everything else has been loaded up successfully.(On that test setup, I'm using Ubuntu 21.04's default desktop WM, GNOME 3.38.4.)
During startup and after loading up the translated model successfully, there's a brief flash where the program tries to draw a GUI window and the program crashes at that exact point. This is what I got in the verbose mode:
There doesn't appear to be a command line switch to force it to run inside terminal? (I understood that I should edit the
config.h
to disable SDL, any hints on how should I do that?)I've tried running the compiled program in userspace modes that are both GPU-enabled (CUDA v11.2) and CPU-only (i.e. different Conda environments for TensorFlow).
I would assume it has something to do with the (SDL-related?) compile warnings (my LibSDL is 2.0.14). Tensorflow itself isn't the most stable contraption either tbh, as you might have noticed if you've worked around with TF/PyTorch based ML/NN stuff. ;)
All help kindly appreciated, and I'm sorry about my silly inquiries in advance -- I hope you can forgive my lack of knowledge on C when it comes to this.
I wish you a nice summer (and Halloween, and Christmas, in case you don't read this by then!) All the best with your unique work. :o)