mirror of
https://github.com/kvcache-ai/ktransformers.git
synced 2026-11-05 01:07:28 +00:00
[GH-ISSUE #1381] unit tests #926
Labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
kvcache-ai/ktransformers#926
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 @createthis on GitHub (Jun 11, 2025).
Original GitHub issue: https://github.com/kvcache-ai/ktransformers/issues/1381
Originally assigned to: @Azure-Tang on GitHub.
I've noticed commit 8dc1ab9 (Apr 14th) is WAY more reliable on my hardware with
DeepSeek-V3-0324:671b-q4_k_mthan current 64ec0ec (June 11th). I can get 50k to 80k long context from 8dc1ab9. However, I'm lucky to get 24k to 29k long context from 64ec0ec.My hardware is dual EPYC 9355 CPUs with 768gb RAM and a 24gb 3090.
I've also noticed that I can easily write code that runs, but introduces a lot of instability and causes models to crash. For instance, I wrote code that backported R1-0528 support all the way to the March 15th commit
bda9cf1- it ran, but it was useless because it went insane after just a turn or two.How can we/I write unit tests that will help us verify that we aren't introducing problems when we make changes?
@ErvinXie commented on GitHub (Jun 12, 2025):
We have some unit tests, in
examples, to test cpu moe correctness.@createthis commented on GitHub (Jun 12, 2025):
Hmm. Seems broken in
8dc1ab9.@createthis commented on GitHub (Jun 12, 2025):
I had a moment during sprint planning to debug this. I made this change:
And got this assertion failure:
No clue if that's a legitimate bug being caught or if the test is out of date. Again, this is with commit
8dc1ab9. I'll try64ec0ecnext since I'm stuck in planning.@createthis commented on GitHub (Jun 12, 2025):
Same result with the same patch at commit
64ec0ec:@createthis commented on GitHub (Jun 12, 2025):
Time for CI/CD?