Skip to content

Use Markov Chain for temporal prediction - #159

Open
supastishn wants to merge 1 commit into
Helldez:mainfrom
supastishn:markov
Open

supastishn wants to merge 1 commit into
Helldez:mainfrom
supastishn:markov

Conversation

@supastishn

Copy link
Copy Markdown

This uses a Markov chain to assist with prediction of experts which should achieve higher accuracy.

@Helldez

Helldez commented Aug 27, 2026

Copy link
Copy Markdown
Owner

Sorry for the wait. Before I can judge this I need a number, not a description:

  • Run --predict-log on a real model with and without the chain and post both. It reports predictor accuracy per layer, it exists for exactly this.
  • clang-format 18 on the touched files (CI checks it), and please keep the comments you removed in session.cpp.

One thing to know going in: I re-measured prefetch on the phone today and it loses even when the prediction is right (+44% bytes for -14% stall, tok/s down). Better prediction helps only if it costs fewer bytes.

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.

2 participants