tensor (@tensor_kelechi) header
tensor (@tensor_kelechi) profile picture

ml research/engineering • prev @audiogen

latent space
Joined June 2023
Photo by tensor (@tensor_kelechi)Photo by tensor (@tensor_kelechi)Photo by tensor (@tensor_kelechi)Photo by tensor (@tensor_kelechi)Photo by tensor (@tensor_kelechi)Photo by tensor (@tensor_kelechi)Photo by tensor (@tensor_kelechi)Photo by tensor (@tensor_kelechi)Photo by tensor (@tensor_kelechi)Photo by tensor (@tensor_kelechi)
I built a custom jax runtime for Flux.2-Klein 4b that cuts text2image generation latency down to 306ms. full writeup: kelechi.cc/articles/flicker
3
4
1
19
895
@RisingSayak please check this out 🫡
6
It is also more cost-effective, with respect to compute pricing and generation time, giving you more generations/images per $
1
23
Two big levers are using a custom flash attention kernel (without the running softmax calculations) and a fused pre-attention qk-prep kernel, plus other fixes in the text encoder pass and Vae decoding.
1
1
81
All development/testing was done on a @GoogleColab v6e-1 chip. Parallel GPU benchmarks were on @modal's rtx pro 6000 (thanks for the free credits😅). feedback/criticism is highly appreciated,
41
officially an unc
20
2
1
70
3,037
last presentable Jax thing I did was making a checkpointing library for Flax(nnx API) models. still needs more testing and improvement. I created it cus I wanted to make something simpler than the rest github.com/kelechi-c/nnx_sav…
11
752
"Google has no moat" yeah sure dude
1
6
605
having a real job really interferes with my ability to tweet 20 times a day
17
13
2
211
27,270
excited to share that I have joined @audiogenai (as a research intern) to build the next generation of audio/music generation models 🚀. It's been extremely cool working with the team so far 🫡
24
5
5
86
3,372
I have EXCITING news: I've started a company! Introducing Sophont We’re building open multimodal foundation models for the future of healthcare. We need a DeepSeek for medical AI, and @SophontAI will be that company! Check out our website & blog post for more info (link below)
106
107
24
1,220
216,874
companies just wanna make money, so all you need to do is show them how you can help them make money drop everything you have ever been told about getting a job and start over from that simple statement
28
141
10
2,253
138,240
tensor retweeted
all of my knowledge is self taught
4
5
80
2,443
1
4
397
Cool. Now someone should make arxiv reels.
wikitok.vercel.app/ It's 2:00 AM Who wants infinite scrolling wikipedia!!
1
2
16
814
Training our most capable Gemini models relies heavily on our JAX software stack + Google's TPU hardware platforms. If you want to learn more, see this awesome book "How to Scale Your Model": jax-ml.github.io/scaling-boo… It was put together by my @GoogleDeepMind colleagues @_sholtodouglas, @charliexychen, @pchoy95, @albertwebson, @vinayramasesh, @froystig, @anselmlevskaya, @sharadvikram, and Fede Lebron, and describing the work of many, many people!
Making LLMs run efficiently can feel scary, but scaling isn’t magic, it’s math! We wanted to demystify the “systems view” of LLMs and wrote a little textbook called “How To Scale Your Model” which we’re releasing today. 1/n
23
159
6
979
129,027
DeepSeek reminds me of Pixart-alpha. $27k cost for diffusion training compared to Dalle-2(~$2M), SDv1.5($320k), etc, while getting better results. Chinese engineers 🐐 Link: arxiv.org/abs/2310.00426
3
3
22
826
research is hard.
31
18
5
349
13,573