Stentor 4 when?

#1
by Datdanboi25 - opened

Yo, I'm hyped for the release of stentor 4 80m, when do you think it will be out?

I’m over halfway through debugging the Stentor 4 files. The move from XLA to JAX/MaxText meant that well over 80% of the code had to be redone, so there’s still a ton of debugging left. My best estimate is that debugging could finish in around 3–5 days, but no promises.

After that, training time depends on Kaggle TPU quota. The training budget is now almost 6× larger, so full Stentor 4 needs roughly two weeks’ worth of quota. Kaggle’s quota resets on Friday. In the best-case scenario, I can use the current quota, let the job auto-resume around the session cap, then continue after Friday’s reset. That could make training take about two days of actual elapsed time. But if debugging finishes on Friday or later, it may stretch across a full extra week because of the quota reset cycle.

There’s also another 1–2 days needed to update the Stentor 3 model cards for Stentor 4.

That’s why I’m releasing Stentor4-Preview first. It only trains on about 1.8B tokens, so it should not run into the cross-week quota issue. Assuming debugging goes reasonably well, I’d estimate the Preview release in about 6–8 days—again, no promises.

For the full non-preview Stentor 4, I’m estimating around 12–20 days. The longer end is mainly if training throughput is worse than expected and I need another week of TPU quota to finish.

Awesome cant wait!

Datdanboi25 changed discussion status to closed

Sign up or log in to comment