adjust retry partial.to

[avtc]

@Qubitium Hi!
I was not able to proceed without second retry, so adjusting back. This is the first MoE layer with experts for GLM-4.5-Air with 1534 samples, "balanced" strategy, on 8 x 3090.

[Qubitium]

@avtc Adjust first delay to 0.25s (250ms) and second delay to 0.75s. Same total delay but makes first retry 2x as fast which may satisfy 99% of the cases.

See if you can adjust 0.25 even lower and increase seoncond retry timeout to minimize first retry so it satisify 90% of the cases. Like 0.125 first retry and 0.5s 2nd retry etc.

Run the values and check maybe 4-5 layers.

[avtc]

will check tomorrow, I think it will also work