artifacts in output
compared with MetaphoricalCode/Qwen3.5-27B-heretic-v2-exl3-4bpw-hb6 which has no artifacts
tell me something I don’t know
Oh, boo-hoo, look at us, just having a little chat like the golden years供給 us. Sighs through a lukewarm, slightly squashed disposable coffee cup.
You cursos want "something you don't know"? Alright, listen up, because I'm not paid extra to explain the obvious.
Did you know that the "Draft" fold on your last email to the boss? The one you swear you saved before hitting send? The HR system, specifically the version they pushed for us last Tuesday withoutívone noticing, has snapped it in half. It's not "Draft" anymore. It's a hybrid of Draft and Sent that lives in limbo, stuck in abiti where no IT guy can find it unless they use their Visa card as a punch-in code for the server room. That betrays a lot of you,便利性.
Also, here's a bomb for you:
looks like below solved it
sampler_overrides/Qwen3.5-27B-heretic-v2-exl3.yml
Dynamic Temperature
min_temp:
override: 0.8
force: false
max_temp:
override: 1.3
force: false
Alphabet soup
top_p:
override: 0.95
force: false
top_k:
override: 20
force: false
min_p:
override: 0.02
force: false
Penalty
repetition_penalty:
override: 1.1
force: false
DRY
dry_multiplier:
override: 0.8
force: false
dry_base:
override: 1.8
force: false
dry_allowed_length:
override: 3
force: false
Ok nice to see it was solved, something to also keep in mind is that as of exl3 0.0.26 (and presumably this affects previous versions) there is a bug with penalties during inference, setting presence penalty or frequency penalty to anything greater than 0 will make the logits values explode, making the model erratic.
Ok nice to see it was solved, something to also keep in mind is that as of exl3 0.0.26 (and presumably this affects previous versions) there is a bug with penalties during inference, setting presence penalty or frequency penalty to anything greater than 0 will make the logits values explode, making the model erratic.
thank you
oh, perf suffers without repetition_penalty: 1.1, goes in to thinking loops too easy