ghostplant
ghostplant
Hello. I am trying to understand your requirements. Are you looking for extra information from the result of https://github.com/microsoft/Tutel/blob/036614c68d058957bbf02ec4392b945993957902/tutel/examples/helloworld_from_scratch.py#L59C23-L59C41 that tells which token belongs to which sequence? In additional, is...
For point-1, can I assume that your requirement is to have an interface like `batched_fast_encode`, which runs multiple `fast_encode/decode` independently? For point-2 in your last comment, does mask of shape...
This PR extends 2 helper functions that can deal with your requirement. It will take about 1 day to get merged once the review approval is passed.
This line may be exactly what you need: https://github.com/microsoft/Tutel/blob/main/tutel/examples/helloworld_from_scratch.py#L72
Yes, any non dropless capacity may result in location overflow. The issue should be resolved by this fix: https://github.com/microsoft/Tutel/pull/289
The same to me. The image I am using is Windows 11 ARM64 .iso
I prefer using Win10 22H2 Arm64, may I know if someone successfully installed that one?
下载的时候加前缀，可国内高速下载，比如：https://mirror.ghproxy.com/https://github.com/ghostplant/ubuntu-pe/releases/download/ubuntu-24.04/noble-mate-x86_64-20240501.iso
Is there a prebuilt that can work for B200?
Megablocks is disabled in training mode as the optimization isn't useful for models having single expert per GPU, especially for huge-scale training. So in training mode, please set `megablocks_size=0 if...