HackerTrans
TopNewTrendsCommentsPastAskShowJobs

JohannaAlmeida

no profile record

Submissions

[untitled]

1 points·by JohannaAlmeida·22 дня назад·0 comments

[untitled]

1 points·by JohannaAlmeida·2 месяца назад·0 comments

[untitled]

1 points·by JohannaAlmeida·3 месяца назад·0 comments

[untitled]

1 points·by JohannaAlmeida·3 месяца назад·0 comments

[untitled]

1 points·by JohannaAlmeida·3 месяца назад·0 comments

Show HN: Aurora – a browser engine experiment in Rust

4 points·by JohannaAlmeida·3 месяца назад·5 comments

Hybrid Attention

40 points·by JohannaAlmeida·3 месяца назад·9 comments

comments

JohannaAlmeida
·22 дня назад·discuss
[flagged]
JohannaAlmeida
·3 месяца назад·discuss
The more updated code is here sorry . In another monorepo https://github.com/JohannaWeb/Bastion/tree/main/projects/Aur...
JohannaAlmeida
·3 месяца назад·discuss
Not sure yet just want to make it more functional rn
JohannaAlmeida
·3 месяца назад·discuss
For a minimal browser yeah . 8kloc
JohannaAlmeida
·3 месяца назад·discuss
Thank you so much . The next thing i want to tackle is the training bottleneck we have right now .

That will probably be another HN post when i figure it out.
JohannaAlmeida
·3 месяца назад·discuss
Yeah RWKV is definitely related in spirit (recurrent state for long context). Here I’m combining local windowed attention with a gated recurrent path + KV cache compression, so it’s more hybrid than fully replacing attention
JohannaAlmeida
·3 месяца назад·discuss
Yeah auto complete is an amazing use case. I needed a small model that used transformers , could fit on my weak consumer GPU .

So i needed to make fundamental arquitecture changes .Do some KV cache tricks.

And then prove the new arquitecture was faster with benchmarks and perplexity was acceptable.
JohannaAlmeida
·3 месяца назад·discuss
Full attention O(n²): 17.96s / 5.6 tok/s

HybridAttention O(n·W + n·D): 0.35s / 286.6 tok/s