Open Bug 2048418 Opened 2 months ago Updated 16 days ago

[llama.cpp] Some models in the Deepseek family cannot be run

Categories

(Core :: Machine Learning: On Device, defect, P2)

Product:
Core
Component:
Machine Learning: On Device
Type:
defect

Tracking

()

People

(Reporter: vpollet, Unassigned)

References

Details

(Whiteboard: [aiplatform])

In the process of bumping Firefox's version of llama.cpp (stack on Phabricator), I applied the patches we maintain in-tree to a clone of llama.cpp directly and ran their tests with our changes. Everything passes, except three that test the tokenizers. Turns out, one of our patches removing std::wstring breaks models with unicode in their pre-tokenizer regex. Among the models affected, there's one in the DeepSeek family (see this test which is failing with our patches).

:gregtatum I set this as P2 as this potentially affects DeepSeek which is quite popular out there. Pinging you for triage

Flags: needinfo?(gtatum)

P2 seems reasonable as this is a correctness issue.

Flags: needinfo?(gtatum)

The severity field is not set for this bug.
:gregtatum, could you have a look please?

For more information, please visit BugBot documentation.

Flags: needinfo?(gtatum)
Severity: -- → S3
Flags: needinfo?(gtatum)
No longer blocks: llama-cpp
You need to log in before you can comment on or make changes to this bug.