I will try to run the benchmark this evening and overnight.

./iobench /Users/doug/glm52_i4/out-00069.safetensors 19 256 8 1

O_DIRECT x8 thread: 256 letture x 19MB = 5.1 GB in 0.36s -> 14.17 GB/s (1.4 ms/blocco effettivi)

./coli chat

 ▄▀▀▀▄  ▄        colibrì v1.0

▄▄▄▄▀▀▀▀▄▀▀ piccolo motore, modello immenso
▀▀▀▀▀▀▀ GLM-5.2 · 744B MoE · int4 · streaming CPU
▀▀▀▀ chat · glm52_i4 · ram -GB · topp off
▀
──────────────────────────────────────────────────────────
[RAM_GB=54.2 auto] cap=8 ok (proiezione picco / projected peak 25.1 GB)
scrivi e premi invio · :reset memoria · :q esci

╭────────────────────────────────────────────────────────────────────────────────────────────────╮
│ › If I paste some c code into this chat, can you analyze it?
╰────────────────────────────────────────────────────────────────────────────────────────────────╯

◆ colibrì
Yes, absolutely! Please go ahead and paste the C code here.

Once you provide it, I can help you with things like:

• Explaining what the code does: I can break down its logic and functionality step-by-step.
• Finding bugs or vulnerabilities: I can review it for potential errors, memory leaks, or security flaws.
• Optimizing it: I can suggest ways to make the code more efficient or readable.
• Answering specific questions: If you have a particular question about how a specific function or block of code works, I can explain it.

Just paste the code whenever you're ready, and let me know if you have any specific goals or questions about it!
└─ 148 tok · 0.52 tok/s · hit 25% · RSS 22.7 GB · 283s

Machine specs:

Model Name: MacBook Pro

Model Identifier: Mac17,7

Model Number: Z1MN0017VLL/A

Chip: Apple M5 Max

Total Number of Cores: 18 (6 Super and 12 Performance)

Memory: 128 GB

  Type:	GPU

  Total Number of Cores:	40

  Vendor:	Apple (0x106b)


  Metal Support:	Metal 4