Issues / #2722
#2722 Strata Profile Studio: a page that shows what limits Strata on your PC, and whether a tune helped
open · @qiuzan2001 · 0 comentários · No GitHub
Setup & installAMD / HIPNVIDIA / CUDAModels & quantsDocumentationWindows
Descrição
Strata Profile Studio is a small web page that runs next to Strata on the same PC. It shows which part of the PC limits Strata (disk, RAM, CPU, PCIe, VRAM, GPU), where the time of each request goes, and whether a change of settings made Strata faster, measured against the noise of your PC. It uses Python 3.11+ and the standard library only, and no GPU of its own: https://github.com/qiuzan2001/strata-profile-studio  To install it, give this sentence to a coding agent on the PC where Strata runs: `Install Strata Profile Studio on this computer and open it in my browser: https://github.com/qiuzan2001/strata-profile-studio (follow docs/AI_SETUP.md)` To look first without Strata: `git clone https://github.com/qiuzan2001/strata-profile-studio && cd strata-profile-studio && python3 studio.py --demo` (recorded data of one PC, not yours). What it measured on my PC (RTX 5070 Ti 16 GB, Ryzen 7 5800X3D, 80 GB DDR4, Swift 1.5 IQ3_S, 0.1.44.1): a calibration (`setup.py --calibrate`) removed `--pcie-frac 0.35` and `--spec-min-p 0.70` because its short-chat test saw no 3 % gain, and set `--adapt-swaps 80`. The profile before and after it (run 5 to run 6): agent turns -5.2 %, 6 of 6 turns slower (sign test p = 0.031); chat +0.4 % and long context -2.7 %, inside the noise. With 0.35 / 0.70 put back, a second calibration changed only `--adapt-swaps 160` to `80` (run 8 to run 9): agent turns +5.1 %, 6 of 6 faster (p = 0.031). The calibration's test is three short chats of 128 and 512 tokens; it does not see agent turns. The Studio shows the calibration's result and the profile side by side, so you can keep or go back. Not tested: Windows, macOS and AMD cards. The readers for them come from the documentation of each tool. If you try it there, `python3 studio.py --host-probe --config <your strata-<model>.json>` prints what it can read, and a "works on" issue in the Studio's repository with that output is welcome. It writes only Strata's run config (after a dialog, with a `.bak` copy) and its own state file.
No site
Links install, modelos, releases.