Steve Korshakov

Results 169 comments of Steve Korshakov

I don't see the reason to do so since there would be a lot of configuration involved, and there is barely any issues with tsconfig.json in typescript projects.

Most of the time, compiler call `compileProjects` - https://github.com/ton-community/tact/blob/0e0d8088809d82e2bffce553cb6b2c92887ad5bc/src/main.ts#L105

Thanks for the fast reply, but invoking 800k times is not very fast. (i am trying to fix polygons of all parcels in NYC). Actually it takes 2 hours at...

Documentation is unclear that this is possible. There are some strange words that shp is multipolygon. I assumed that this will be the one.

Bot doesn't implemented correctly for now. Telegram Api have an idea about Update Sequence and all update messages (username changed, messages, etc...) are sent by server in special sequence. Sometimes...

I have a separate implementation, that i built looking at this repo and i am also facing NaNs from time to time. What is interesting - they are quire random,...

I was training in bfloat16, but switched to fp16 now. Everything was good until some loss became NaN, but restarting helps. I feel there is something within a data. My...

I have restarted training from the checkpoint shortly before explosion and it did not crash and the only difference is how i cut pieces of data to feed to the...

I am curious why no one uses bfloat16?

@nrocketmann IIRC A100 can have flash attention on bfloat16 as well, performance is similar. bfloat16 in my experience much more stable for training. I am starting to think that this...