Skip to content

docs: M4 Pro 48GB community hardware validation#81

Open
cemigo114 wants to merge 1 commit intoTheTom:mainfrom
cemigo114:m4-pro-validation
Open

docs: M4 Pro 48GB community hardware validation#81
cemigo114 wants to merge 1 commit intoTheTom:mainfrom
cemigo114:m4-pro-validation

Conversation

@cemigo114
Copy link
Copy Markdown

Summary

  • End-to-end validation on Apple M4 Pro 48GB (MTLGPUFamilyApple9, Metal4)
  • llama-bench speed benchmarks (q8_0, turbo4, turbo3, asymmetric q8_0-K + turbo4-V)
  • Perplexity comparison on Qwen2.5-1.5B Q8_0 with wikitext-2
  • Reproduction script (demo_run.sh)

Key results

Config Prefill t/s Decode t/s PPL
q8_0 baseline 2326 111.6 11.917
q8_0-K + turbo4-V 2271 89.8 12.048 (+1.1%)
turbo4 symmetric 2245 75.8 6921 (catastrophic on 1.5B)

Confirms asymmetric K/V works on M4 Pro. Symmetric turbo4 catastrophic on this small model, consistent with documented findings. Decode regression between M1 Max and M5 Max, as expected for pre-M5.

Files added

  • docs/community-hardware/m4-pro-48gb.md — raw benchmark data
  • demo_run.sh — one-command reproduction script

End-to-end validation on Apple M4 Pro 48GB:
- llama-bench speed (q8_0, turbo4, turbo3, asymmetric)
- Perplexity comparison confirming asymmetric q8_0-K + turbo4-V at +1.1%
- Symmetric turbo4 catastrophic on 1.5B, consistent with documented findings
- Reproduction script (demo_run.sh)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant