mirror of
https://github.com/kvcache-ai/ktransformers.git
synced 2026-11-05 01:07:28 +00:00
[GH-ISSUE #82] Seg Fault on long replies #42
Labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
kvcache-ai/ktransformers#42
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @matthusby on GitHub (Sep 10, 2024).
Original GitHub issue: https://github.com/kvcache-ai/ktransformers/issues/82
I am trying to use this as a llama.cpp replacement as it is a lot faster. I did modify the backend args file (
ktransformers/server/backend/args.py) as max_new_tokens isn't an option for the sever:to allow for longer responses as 500 tokens was causing a lot of my stuff to get cut off half way through.
It does generate some tokens (feels like about the 500 limit) and then hard crashes. So maybe there is some other limit that I need to adjust?
I am using DeekSeek-V2.5 in Q4_K_M, I did also try it with WizardLM 8x22B and the same thing happens.
Hardware: 1x 3090, Epyc 7402, 512 Gb Ram
@Azure-Tang commented on GitHub (Sep 11, 2024):
Hi, maybe this one will help #73
@matthusby commented on GitHub (Sep 11, 2024):
Thank you, yeah changing the
cache_lensparam is what was causing my problem. I guess my chat was at just the right length where the longer max tokens would trigger that error.For reference I doubled it and now when I try to run DeepSeek-V2.5 I get a out of memory crash, but on Wizard 8x22 its working great.
It looks like there are plans to address some of this soon, so I will close this issue.