Performance regression: llama-server high CPU use when generating tokens.

I am building ollama from git source. In the latest version e2c6c7e8 (HEAD -> main, origin/main, origin/HEAD) (14 hours ago) docs: claude docs (#18006) [GitHub], I notice llama-cpp is consuming about 560% CPU on my Mac Studio M4 Max 128 GB RAM when generating tokens. I remember this being much lower (almost no CPU use about 10% or less) in the past.

I did git bisect and narrowed the problem down to this commit:

e5a81899d014a847a08d47393351908b53d74008 is the first 'bad' commit
commit e5a81899d014a847a08d47393351908b53d74008
Author: Daniel Hiltgen <[email protected]>
Date:   Fri Aug 14 19:03:20 2026 -0700

    llama.cpp update (#17760)

 LLAMA_CPP_VERSION | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

Please note that I also built llama.cpp from git repo and latest commit (d19b2876 (HEAD -> master, tag: b10636, origin/master, origin/HEAD) (5 hours ago) ci: Clean up UI builds from releases (#27706) [GitHub]) does NOT exhibit this problem. llama-server running the same model does not use high CPU (same context size).

All of the above is with gemma4 26b, 262144 context size, open-webui 0.11.1.