It looks like the problem is in the new ggml build but it's not the
version - it's the changelog.

ggml contains the hip-specific code that needs to be built for specific
amdgpu ISA targets and, like the ROCm components do, it uses pkg-rocm-
tools to get that list dynamically. To be specific, the shell command
used is rocm-target-arch --sep ';'

rocm-target-arch was designed such that every release can have its own
list of GPUs and as such, it parses the changelog during build and
selects a GPU list based on the latest changelog entry.

The new ggml build in resolute was imported directly from debian and
thus, that changelog entry is for 'unstable'. Because of this, rocm-
target-arch pulls in the ISA targets for unstable which are is not the
ISA list that we want to be building for.

This was not a problem in ROCm testing before now because the automated
builds in the PPA we're using have an auto-generated changelog that uses
'resolute' for the target.

rebuilding ggml with a changelog entry with 'resolute' as a target
should fix this.

-- 
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2146822

Title:
  llama-cli segfaults on startup

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/llama.cpp/+bug/2146822/+subscriptions


-- 
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs

Reply via email to