Access violation, but works if I keep retrying and sometimes garbled output

[RationalFragile]

Describe the Issue
I use this command:
python koboldcpp.py --model "E:/models/knoveleng_Open-RS3-Q6_K.gguf" --skiplauncher --useclblast 0 0 --gpulayers 29 --contextsize 4096

and I get:

Traceback (most recent call last):
  File "C:\koboldcpp\koboldcpp.py", line 6069, in <module>
    main(launch_args=parser.parse_args(),default_args=parser.parse_args([]))
  File "C:\koboldcpp\koboldcpp.py", line 5213, in main
    kcpp_main_process(args,global_memory,using_gui_launcher)
  File "C:\koboldcpp\koboldcpp.py", line 5610, in kcpp_main_process
    loadok = load_model(modelname)
  File "C:\koboldcpp\koboldcpp.py", line 1115, in load_model
    ret = handle.load_model(inputs)
OSError: exception: access violation reading 0x000000005CC05C38

(The address at the end always changes, e.g. 0xFFFFFFFFFFFFFFFF, 0x000000005CC05C38, 0x0000000002093938...)
But, if I just keep rerunning the command, it works.
One time it worked but random tokens get generated.
I try again and it works just fine.

(The issue never happened if I use cpu instead of clblast.)

Additional Information:
Windows 10 Pro 21H1 19043.1110
Intel i5-5200U, Intel HD Graphics 5500, OpenCL 2.0
koboldcpp 1.86.2

[LostRuins]

Can you try adding --noavx2 to the launch commands and check if that works?

[RationalFragile]

I tested with --noavx2 and it does not fix the issue. Thank you.

[LostRuins]

What if you combine --noavx2 with --gpulayers 0, but keep useclblast?

[RationalFragile]

I tested these:

python koboldcpp.py --model "E:/models/knoveleng_Open-RS3-Q6_K.gguf" --skiplauncher --useclblast 0 0 --gpulayers -1 --contextsize 4096 --noavx2
No issue

python koboldcpp.py --model "E:/models/knoveleng_Open-RS3-Q6_K.gguf" --skiplauncher --useclblast 0 0 --gpulayers 0 --contextsize 4096 --noavx2
No issue

python koboldcpp.py --model "E:/models/knoveleng_Open-RS3-Q6_K.gguf" --skiplauncher --useclblast 0 0 --gpulayers 1 --contextsize 4096 --noavx2
Issue happens (sometimes)

Thank you again.

[LostRuins]

Yup, so it's happening when layers are offloaded in clblast. Let me see if I can figure it out.

[LostRuins]

Hi, can you check if this is fixed by 1.87.2?

[RationalFragile]

Thank you, I just tried with 1.87.2 and the issue still happens.

[LostRuins]

Hmm unfortunate. then i'm afraid you will not be able to offload layers. You can still use it for GPU prompt processing, but the generation will be using CPU only.

python koboldcpp.py --model "E:/models/knoveleng_Open-RS3-Q6_K.gguf" --skiplauncher --useclblast 0 0 --gpulayers 0 --contextsize 4096