Show HN: FemtoGPT – Pure Rust implementation of a GPT language model

Posted by keyvank on Sunday June 04, 2023 @07:58PM
from the show-hn dept.

( 6 comments | on HN )

Re:Show HN: FemtoGPT – Pure Rust implementation of a GPT lan... by jayy-lmao on Monday June 05, 2023 @04:55AM (#36193987)

I sure hope this is the unit of measurement and not a Berserk reference

[ Parent | on HN ]

Re:Show HN: FemtoGPT – Pure Rust implementation of a GPT lan... by quickthrower2 on Monday June 05, 2023 @07:00PM (#36205449)

Well done! I would be crazy if I dived into this having never studied Rust! But it looks very interesting so I might do that! I have almost finished AK’s lecture. Have you made the matmul faster yet? That would seem a priority?

[ Parent | on HN ]

Re:Show HN: FemtoGPT – Pure Rust implementation of a GPT lan... by keyvank on Monday June 05, 2023 @07:04PM (#36205495)

Hey! Yes please dive in :) It can be a starting point for learning Rust ;) MatMul is fine, the bottleneck is LayerNorm + Softmax gradients right now. Haven't been able to optimize them well yet.

[ Parent | on HN ]

Re:Show HN: FemtoGPT – Pure Rust implementation of a GPT lan... by quickthrower2 on Monday June 05, 2023 @11:06PM (#36207561)

Well, first step is to run it. I am getting a step every 3 seconds. Is that typical?

i7-13700H, 2400 Mhz, 14 Core(s), 20 Logical Processor(s). It is using 12 CPUs which is good. And 160Mb memory.

[ Parent | on HN ]

Re:Show HN: FemtoGPT – Pure Rust implementation of a GPT lan... by keyvank on Tuesday June 06, 2023 @05:25AM (#36210233)

Ah, that's slow. I have been able to reach 1.3s per step on my machine. Did you run with --release flag?

[ Parent | on HN ]

Re:Show HN: FemtoGPT – Pure Rust implementation of a GPT lan... by quickthrower2 on Tuesday June 06, 2023 @05:56AM (#36210444)

Yes release. Is running on Windows which might affect things.

[ Parent | on HN ]