[Bug]: Inference on Arc A770 segfaults after some number of tokens

[richiejp]

OpenVINO Version

2024.1.0

Operating System

Other (Please specify in description)

Device used for inference

GPU

Framework

PyTorch

Model used

https://huggingface.co/fakezeta/openchat-3.5-0106-openvino-int8

Issue description

On Ubuntu 22.04.04 using the upstream kernel driver (Linux dell 6.5.0-1020-oem #21-Ubuntu SMP PREEMPT_DYNAMIC Wed Apr 3 14:54:32 UTC 2024 x86_64 x86_64 x86_64 GNU/Linux) with an Intel Arc A770 inference usually causes a segfault after some number of tokens.

Originally reported to LocalAI: mudler/LocalAI#2219

I had issues installing the out-of-tree driver from Intel so have not tried that. There are no problems using the iGPU on the same machine or the CPU.

c.c. @fakezeta

Step-by-step reproduction

It can be reproduced either in LocalAI or by running https://github.com/fakezeta/openchat.openvino/tree/main

All that is required is to ask (any?) model a longer question, such as to describe the output of lscpu. It will usually segfault.

Relevant log output

[ +34.787117] python[66644]: segfault at 1e ip 00007e9adf577063 sp 00007e98aa7ea280 error 4 in libopenvino_intel_gpu_plugin.so[7e9ade9d6000+de4000] likely on CPU 4 (core 8, socket 0)
[  +0.000010] Code: ff e8 81 2c c7 ff 48 8b b5 38 ce ff ff 4c 89 ff 80 8d 37 ce ff ff 80 e8 6b 2c c7 ff 48 8b 85 c0 da ff ff 80 8d 3f ce ff ff 80 <80> 38 00 0f 85 fc 0c 00 00 48 8b 85 f8 da ff ff 80 38 00 74 67 48

Already disassembled here: mudler/LocalAI#2219 (comment)

### Issue submission checklist

- [X] I'm reporting an issue. It's not a question.
- [X] I checked the problem with the documentation, FAQ, open issues, Stack Overflow, etc., and have not found a solution.
- [X] There is reproducer code and related data files such as images, videos, models, etc.