-
Notifications
You must be signed in to change notification settings - Fork 7.6k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Illegal intruction after running gpt4all-lora-quantized-linux-x86
#82
Comments
There are 665 instructions in that function, and there are ones that require AVX and AVX2. So you need a CPU with AVX2 support to run this, as far as I can see E7-8880 v2 supports only AVX, but not AVX2. But in the output you've provided I see
which confuses me a bit. |
Lol, figured that out.
ggml_cpu_has_avx2 is basically "return true;" in the code. ggml_cpu_has_avx and ggml_cpu_has_sse3 are the same. Interesting that ggml_cpu_has_avx512 is "return false;". I.e. there are no real checks during the output of these statistics (from _Z23llama_print_system_infov, after the "system_info: " part of the line). It is decided in compile-time, and not run-time. |
FWIW I ran into a similar problem running a VM under Proxmox. I was able to work around this by setting the CPU type to "host", which exposed the full instruction set (for a Ryzen 9 5900X in my case) and then it worked. Certainly sounds like @qinidema is onto something here though. ;] |
Why is AVX2 necessary anyway? Is there a workaround? |
lol the gif demo has This probably just needs a simple recompile. |
@mvrozanti you need to recompile this to get a new binary with your compile-time defines. |
@qinidema That fork also didn't work for me:
At least I got an error with this one. What made you believe that fork in specific would work? |
It's the actual source code of the project.
|
That line in this (current) repository.
|
@mvrozanti @pirate486743186
Then i've received the "chat" executable (and at least is starts and shows help message successfully), as well as "quantize" and "libggml.a" library. |
I'm on Debian 11. It's probably incompatibility with older versions. |
@nomic-ai |
@pirate486743186 try this: no-avx2.tar.gz |
with the static compile, it gives again 'illegal instruction'. I have an old laptop. with cmake it can't find pthread when compiling apparently.
|
Same issue; though I get illegal instruction in |
for those that are frustrated, keep in mind it was released 2 days ago. Have ultra low expectations. |
I have an Intel i5-3320M with no AVX2 or FMA support. I followed these steps:
and then $ cmake -D LLAMA_NO_AVX2=1 -D LLAMA_NO_FMA=1 .. and it worked. On my laptop, it is very slow as would be expected. |
@pirate486743186 and what is the address of that instruction? @mvrozanti you can try this one: no-avx-avx2-fma-f16c.tar.gz @vsl-iil you can try the archive above too, though I cannot determine what is the instruction on your address 0x000055555558a1c9, there's some kind of heavy ASLR in your case. |
It doesn't say. I have avx and f16c. These should probably work for most or even all. |
Add this to CMakeLists.txt after line 25: |
i opened an other issue for debian 11 It compiles with this fix and the one mentioned there. But unfortunately, when it starts it gives |
@pirate486743186 can you share your resulting compiled binary and coredump after the crash? |
dump 2GB, respond here #180 the compiled file |
I'm using this. It seams to work better. to use it, you'll need to convert it run these commands adapted to your case ( they are working on a new unified converter script)
Then you run it with something like this. This is a bit incorrect, you'll need to adjust the parameters for better behavior. |
I am using old Macbook pro (Mid 2012 Intel Model) with 8GB RAM. $ wget https://github.com/zanussbaum/gpt4all.cpp/archive/refs/heads/master.zip This worked for me! But very very slow! I am going to upgrade my RAM tomorrow and see if that helps! |
it needs 4GB, more RAM will not help. In the next months/year, efficiency should increase by a lot. In general, at first software is inefficient and slow. |
Can you provide suggestions on how to fix this error?
|
When doing cmake, I get: And fails with |
I'm trying to run the
gpt4all-lora-quantized-linux-x86
on a Ubuntu Linux machine with 240 Intel(R) Xeon(R) CPU E7-8880 v2 @ 2.50GHz processors and 295GB RAM. No GPUs installed.Ubuntu 22.04 running on a VMWare ESXi
I get the following error: Illegal instruction
The text was updated successfully, but these errors were encountered: