正在加载视频...

视频加载失败

[ENG SUB] how it feels to use eager pytorch in 2025

88,878 次观看 • 1 年前 •via X (Twitter)

27 条评论

VioP 的头像
VioP1 年前

how much effort do u put in daily just to create bangers?

tender 的头像
tender1 年前

Minimal, I just have a very particular set of skills. Skills I have acquired over a very long career of being terminally online.

Misha Goin 的头像
Misha Goin1 年前

i need this to be your full time job. 996 pytorch dev anime

Birchlabs 的头像
Birchlabs1 年前

all I’ve got left is cuDNN SDPA

cloud 的头像
cloud1 年前

wait what's wrong with eager mode cuda graph replays?

shikhar 的头像
shikhar1 年前

dOOMed!!

bilal 的头像
bilal1 年前

holy banger

Andre S 的头像
Andre S1 年前

Diffusers fans when their torch compile flow is still slower than baseline

Adi Ganesh 的头像
Adi Ganesh11 个月前

I would love to watch an MLPerf anime

Anam Hira 的头像
Anam Hira1 年前

Artwork

Barak 的头像
Barak1 年前

Me when the compiled backwards pass is 4x slower than eager and I have to rewrite it myself

tokenbender 的头像
tokenbender1 年前

banger

Lee Penkman 的头像
Lee Penkman1 年前

Basically me but then graph breaks and so I need to explicitly call cuda graph break and reconfigure settings because graph breaks keeps triggering graph recompiling then I need to migrate off max-autotune to reduce-overhead. At least I'm not dOOMed

Jake Halloran 的头像
Jake Halloran1 年前

brother

Thomas Neff 的头像
Thomas Neff1 年前

Certified banger

λux 的头像
λux1 年前

🔥🔥🔥

Khoa 的头像
Khoa1 年前

banger

Amit Prakash 的头像
Amit Prakash1 年前

banger

Manu Gaur 的头像
Manu Gaur1 年前

Chosen one

cloud 的头像
cloud1 年前

banger

Nike_Zoldyck 的头像
Nike_Zoldyck1 年前

He's dOOMed ! Haha. Followed

surya 的头像
surya1 年前

this is so good

steve 的头像
steve9 个月前

my pytorch eagerly dispathces to megakernels (2% mfu btw)

cheaty 的头像
cheaty1 年前

hey lol it's me in the eightsix

Gregory 的头像
Gregory1 年前

Fire⚔️⚔️

paulopacitti 🌐 的头像
paulopacitti 🌐1 年前

I WANT MORE EPISODES

Quantamentally ill 的头像
Quantamentally ill1 年前

Is this what it feels like to finally experience a sense of belonging in this world 🥹

相关视频