cannot run on CPU with AVX but no AVX2 - crashes with SIGILL

[bonze82]

System Info

version 2.5.3 debian 12.2 on xen 6.1.0-12 hvm qt 6.4 python 3.11

Initial start ran fine until it got the model then received traps: b2752a1d-32d9-4[3788580] trap invalid opcode ip:7f47dac751ea sp:7f47b37fcd50 error:0 lscpu flags fpu de tsc msr pae mce cx8 apic sep mca cmov pat clflush acpi mmx fxsr sse sse2 ss ht syscall nx rdtscp lm constant_tsc rep_good nopl nonsto p_tsc cpuid tsc_known_freq pni pclmulqdq est ssse3 cx16 sse4_1 sse4_2 popcnt aes xsave avx hypervisor lahf_lm xsaveopt

Built from scratch as well and still no go the built from scratch opens and runs but only downloads the .bin models and then refuses to run, the only model it will download wizardlm-13b-v1.1-superhot-8k.ggmlv3.q4_0.bin. When attempting to delete the model within the environment it throws a wizardlm-13b-v1.1-superhot-8k.ggmlv3.q4_0.bin.bin error.

Information

• [X] The official example notebooks/scripts
• [ ] My own modified scripts

Reproduction

1. install gpt4all
2. start gpu4all
3. download model
4. crashes every start even after removing model

Expected behavior

Start download the models and run the model.

[cebtenzzre]

Could you be more specific? You are starting the chat UI? If you have a recent clone of the main branch it should definitely download .gguf files, not .bin.

[bonze82]

How can I completely flush this from my system and start from scratch as I just rebuilt with a new pull of 2.5.4, i was building against 2.5.2 and it is an immediate crash with traps: fc724f40-1fb2-4[3821774] trap invalid opcode ip:7fd395877c15 sp:7fd36ddf9dc0 error:0 in libllmodel.so.0.5.0[7fd395834000+bd000] in dmesg. I flushed all found databases and generated files then ran again and got this crash traps: 12321a58-f9de-4[3823052] trap invalid opcode ip:7f0e18a77c15 sp:7f0df13f9dc0 error:0 in libllmodel.so.0.5.0[7f0e18a34000+bd000] and qt creator gave this error before the crash [Warning] (Tue Nov 21 17:38:49 2023): ERROR: Cannot select collections QSqlError("1", "Unable to execute statement", "no such table: collections")

[bonze82]

yes this is the chat ui

[cebtenzzre]

This is because of the fact that we unconditionally link to the AVX2 build of llama.cpp in at least two places. We are doing this so the available devices can be shown before loading a model. We need to find a better way to do this.