Steve Korshakov

Hey, i am curious what was the reason to pick "1" instead of commonly used "2" for vocos?

FA2' flash_attn_varlen_func is 300x slower than flash_attn_func

5

Hello! I am benchmarking attention implementations, and trying to use flash attention for my variable length data and for some reason var length is much much slower than any other...

It seems it is missing and all code is implemented for 5dof variant.

Currently Robotis don't have them in stock and it seems they are almost single vendor that sells this part.

Missing logprob

1

For some reason, there are no way to get logprob of a completion to measure or visualise network performance, it would be nice to have to be able to build...

fix: fixing batching in MMS_FA

3

Fixing batching in MMS_FA bundle