ProsperoAI Pro (unofficial) 01.005.000 - PS5 Pro tuned

ProsperoAI Pro (unofficial build)

An unofficial build of ProsperoAI
by BlackBearReloaded, tuned for the PS5 Pro. It is not made, endorsed or
supported by the ProsperoAI author: report problems with this build to
https://github.com/obl1vn/ProsperoAI/issues, not upstream.

Source: the ps5pro branch of https://github.com/obl1vn/ProsperoAI
(ProsperoAI 01.005.000 plus the changes below). License: GPL-3.0-or-later, like
ProsperoAI.

What is different

Change	Effect on a PS5 Pro (Mistral 7B Q4_0)
llama.cpp mat-vec tuning (vulkan_env.txt)	faster generation
batch/ubatch 512 (batch_size.txt, ubatch_size.txt)	faster prompt reading
fewer interface frames while a model writes (ui_sleep_ms.txt)	faster generation
Pro memory mode (param.json attribute3 bit 22)	12.63 GiB for models instead of 11.44 GiB
no update check	an official update would replace this build

Together: 86.9 -> 96.6 tokens/s generation and 692 -> 740 tokens/s
prompt reading, measured on a PS5 Pro, firmware 13.60. Details:
https://github.com/obl1vn/Unlock-ps5-pro-memory

Only Mistral 7B Q4_0 has been measured; other models use the same settings but
were not benchmarked. Not tested on a base PS5.

Install

It uses the same title ID as ProsperoAI (PPSA99004) and replaces it:

1. Delete the old /data/homebrew/PPSA99004 (models in /data/prosperoai/models
   stay).
2. Upload this folder as /data/homebrew/PPSA99004.
3. Register it again in your homebrew loader and launch ProsperoAI Pro (unofficial).

To go back to the official app, install the official release the same way.

Settings

All files sit next to eboot.bin and are read when the app starts:

• vulkan_env.txt: llama.cpp / ggml-vulkan switches (see docs/PS5_PRO_TUNING.md
  in the source).
• ui_sleep_ms.txt: pause after each interface frame while a model writes,
  0-1000 ms; 0 draws every frame like the official app.
• batch_size.txt, ubatch_size.txt: prompt batch sizes.