Issues / #2334

#2334 Running Coder on Intel A380 causes gpu hangs and driver crashes

open · @asmfreak · 0 comments · View on GitHub

Models & quants

Description

I'm running EndevourOS (Arch branch distro). I have an Intel Xeon processor 14 cores (28 threads) and 32 Gb of RAM with Intel Arc A380 GPU with 6 Gb of VRAM.

```
[level_zero:gpu][level_zero:0] Intel(R) oneAPI Unified Runtime over Level-Zero, Intel(R) Arc(TM) A380 Graphics 12.56.5 [1.15.39122]
[opencl:cpu][opencl:0] Intel(R) OpenCL, Intel(R) Xeon(R) CPU E5-2690 v4 @ 2.60GHz OpenCL 3.0 (Build 0) [2026.21.3.0.31_160000]
[opencl:gpu][opencl:1] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A380 Graphics OpenCL 3.0 NEO  [26.27.39122]
```

I'm using current main of Strata, I've compiled Strata with SYCL compiler.
I used Claude to help me deduce arguments for Strata. 
Here is the run script it came up with: 

https://gist.github.com/asmfreak/92d02ff8296ed76c62ff2aa200a53f79

There were a lot of iteration but in the end it didnt even matter - I had to stop my experiments as it wedged gpu driver with the following message in `systemctl -k`:

```
kernel: Fence expiration time out i915-0000:06:00.0:strata[1882]:4dc!                                                                      
kernel: Fence expiration time out i915-0000:06:00.0:strata[1882]:4de!  
```

After additional tuning it crashed but without GPU driver. This time with the following message:

```
kernel: ixgbe 0000:01:00.0: invalid VPD tag 0xff (size 65378) at offset 25091                                                              
kernel: i915 0000:06:00.0: [drm] GPU HANG: ecode 12:2:ecffffe4, in strata [2103]                                                           
kernel: i915 0000:06:00.0: [drm] GPU error state saved to /sys/class/drm/card1/error                                                       
kernel: i915 0000:06:00.0: [drm] strata[2103] context reset due to GPU hang 
```

The last file in the gist I linked earlier is the error state from GPU driver.

I had hoped that it'll work, but alas - miracle didn't happen. If there is anything else I can do to fix it - please say it

Related on strata.com

Editorial links to help you install, pick models, or read release notes — not part of the upstream thread.