Plamo arch no longer working

[LostRuins]

Name and Version

Forwarding this here, it's been a very long time since plamo was added and seems like somewhere along the way it's implementation broke.

Tested with https://huggingface.co/tensorblock/plamo-13b-GGUF/blob/main/plamo-13b-Q4_0.gguf and https://huggingface.co/mradermacher/plamo-13b-GGUF/blob/main/plamo-13b.Q2_K.gguf

On llama-b5195-bin-win-avx2-x64
llama-cli.exe -m plamo-13b.Q2_K.gguf -p "Hi, my name is"

Hi, my name is nike dmeldv/ cd/ t ta pets petaren

external ref: LostRuins#1506
@okdshin

Operating systems

Windows

GGML backends

CPU

Hardware

i9 13980HX, running cpu, win 10

Models

Tested with https://huggingface.co/tensorblock/plamo-13b-GGUF/blob/main/plamo-13b-Q4_0.gguf and https://huggingface.co/mradermacher/plamo-13b-GGUF/blob/main/plamo-13b.Q2_K.gguf

Problem description & steps to reproduce

Listed above

First Bad Commit

No response

Relevant log output

llama-cli.exe -m plamo-13b.Q2_K.gguf -p "Hi, my name is"

Hi, my name is nike dmeldv/ cd/ t ta pets petaren

[CISC]

I think I know what the issue is, investigating...

Unfortunately it will most likely require reconversion.

[CISC]

Nope, that wasn't it. :(

@LostRuins Do you know roughly when it stopped working?

[LostRuins]

Probably quite a long time. It was only brought to my attention last week from a user, but a quick test shows that it's been broken for at least a few weeks.

[CISC]

Well, this is interesting, just tested if e0dbec0 could be the culprit (again), and while it does behave slightly better (it finished about 1.5 sentences before descending into madness), it's still broken before that commit.

I will try bisecting even further...

[LostRuins]

Thanks for looking into it

[CISC]

Oh, wow, this has been broken for over a year, it starts at bf08e00 then gets gradually worse from there...