rdmpw3275m-changelog-20260901-2322-refresh-cascadelake-script-with-llamacpp

Summary: Refreshed brew_recompile_cascadelake.zsh to include llama.cpp, ggml, whisper-cpp, ollama, simdjson, fast_float, highway, and other high-performance AI/numerical packages. Successfully compiled llama.cpp and ggml from source with AVX-512 VNNI optimizations and deployed across rdmpw3275m, rdmpw3265m, and rdmsm4x:~/dev/scripts/.

1. Scope

• Hosts: rdmpw3275m (Intel Xeon W-3275M 28C/56T), rdmpw3265m (Intel Xeon W-3265M 24C/48T), rdmsm4x (Canonical Root).
• Files Modified:
  • ~/dev/scripts/brew_recompile_cascadelake.zsh (Updated AI/compute whitelist, receipt auditing, ccache integration).
  • rdmpw3265m:~/dev/scripts/brew_recompile_cascadelake.zsh
  • rdmsm4x:~/dev/scripts/brew_recompile_cascadelake.zsh

2. Changes Made

1. AI & Heavy Compute Package Whitelist:
   • Added llama.cpp, ggml, whisper-cpp, ollama, highway, eigen, openblas, simdjson, fast_float, faiss, onnxruntime, libtorch, opencv, duckdb.
2. Compiled from Source on Cascade Lake:
   • Recompiled llama.cpp 0.3.0 and ggml 0.22.0 from source targeting -march=cascadelake -mtune=cascadelake -O3 -pipe with AVX-512 VNNI vector acceleration.
3. Multi-Host Deployment:
   • Distributed and synchronized the canonical script across all Intel Mac Pros and the canonical rdmsm4x hub.

3. Verification Evidence

• llama-cli on rdmpw3275m:

  version: 0.3.0 (build 10621, commit c1d0e7a00)
  built with AppleClang 21.0.0.21000101 for Darwin x86_64

• Recompile Benchmarks:
  • llama.cpp: Recompiled in 120s using 56 threads.
  • ggml: Recompiled in 32s using 56 threads.
• Audit Scan:
  • Verified llama.cpp and ggml are recognized in the audit whitelist and queued when unoptimized bottles are present.

4. Undo / Rollback

To reinstall generic pre-built bottles:

brew reinstall llama.cpp ggml