Conversation
|
I believe ggml-org/llama.cpp#19591 hasn't reach ggml yet, so this could cause a regression on ROCm. |
That's a patch to llama.cpp not to ggml though. And just to prove things work - I do have this in a downstream fork (tied with the other two PRs I have open). https://github.com/lemonade-sdk/stable-diffusion.cpp/actions/runs/22145959583 |
It is: in the same way the change that causes build issues with pre-7.0 ROCm (ggml-org/llama.cpp#19580) originated in llama.cpp ggml-org/llama.cpp@6845f7f , then migrated to ggml-org/ggml@64a9c1b , we need ggml-org/llama.cpp@2ba9adc to migrate to the ggml repository before syncing, or we'll have the same build issue here on sd.cpp. |
|
Ohhh, thanks for explaining it. Ideally there should be a ROCm 6.x CI target here too so this would have failed as a result. If you guys are wanting one I can adjust to add one after my others are merged. I'm pretty new to submitting things to ggml projects. How does that migration happen? Do I need to do something manually to push for a backport? |
I'd personally appreciate that, but I know each build variant increases CI testing time. Maybe it would be more useful on the ggml repo instead? That way, the llama.cpp devs may catch this kind of issue as soon as they make a release, then decide if they sync again, or add a hotfix.
IIRC llama.cpp devs aim for 2-4 week cycles, but I've seen people pestering ggerganov to sync sooner on a few occasions 🙂 An old discussion about it: ggml-org/ggml#1066 Also related: ggml-org/llama.cpp#6381 |
2 participants
No description provided.