Sagnik

Dilon mein tum apni betabiyan leke chal rahe ho, toh zinda ho tum...

@vllm_project engineers are crazy ,Most of the components are art I was going through the memory of gpu part Like how they decide 12 gb vram gpu with 95% utilization -> 11.4 gb in this how the whole model weight, all the kv cache , cuda graphs and other runtime memory will load and run ? Their technique - > load the model weights > run a dummy forward pass(fake inputs) > also track other memory req (kernels) > now we have --> weights + others(cuda/custom) > rest give to the kv cache > same for all models(text/mm) > so proper gpu utilization

@jbhuang0604 What an algo I was just watching ur flash attention vid 😅

Wow, Opus 4.7 seems very capable. Let me give it a try! 45 mins later:

@squibsters Ik the third pic veggi but i forgot the exact name

Deep in the mountains of Uttarakhand, spending my time foraging, cooking and walking

@gpusteve Mtp is also a spec decoding right?

🪩

@w0rdgenerator 🥰

Bro…

@inkdrop_app Cool

🎬 New video: Explore Zed's source code to learn how to support multiple AI providers 💪

@gooseydotgov Give this cutu more 🐈

the creature's craving for brown fishy slop has been fulfilled... yet the creature yearns for more...

lmao

vast.ai gpus are quite cheap(mostly spot ig?) Modal also give free gpus decent amount , I have never tried to rent any gpu , for running for few hours and benchmarking whats the best options we have ?

@gregpr07 browser-use rewritten in rust when

How did “let’s rewrite it in Rust” go from a meme to a single /goal prompt

@meetapersecond xactly me

me when i hear work from home 💦💦💦💦

@emxlymargaret 😰😰😰

another mother's day where my cats didn't get me SHIT

The only i keep up to date with all the ai engg stuffs is actually build my own coding agent and keep updating , pi is actually a good start , adding good harness like codex or opencode will be tough , but lets try..

@merve77_ Smol hands 😹😹

Nasılda kendini sevdiriyor sevimli portakal. Sarı beyaz yanında olmayınca uslu uslu takılıyor.😂

@mohitwt_ Damn coool

I'm building a GPT-2 inference engine from scratch in CUDA. the project focuses on implementing and optimizing transformer inference kernels directly at the CUDA level, with emphasis on reduction strategies, memory behavior, numerical stability, and kernel-level performance optimization. by the end, this engine will take a prompt, tokenize it, run a full GPT-2 forward pass using custom CUDA kernels, and autoregressively generate text, with KV cache for fast decoding. every operation from embedding lookup to sampling runs through kernels written and profiled from scratch. things planned for the repo: - attention kernels - causal masking - KV cache, and alot more. things i've implemented so far: - tiled matmul with shared memory - online softmax - 3 pass layernorm: benchmarked against Welford across all GPT-2 model sizes, using 3 pass as primary kernel for LN benchmarks and profiling done on RTX 3050 laptop GPU. repo: github.com/Mog9/gpt2-infe…

@Madhav_goyal_ damnnn looks great

Thai mangoes dont get enough hype, likely on par with Indian mangoes

My favourite character rn 😋👇🏻

@PopCrave 🦐

Today is National Shrimp Day.

@JustAditiThing Us aditi us

I wanna go back to my momo and puchka, this vada pav and misal pav thing ain't for me bhaiiiiiiii 😭😭😭

@kafklore Will gift this to my sister 😋

gente comprei uma presilha que promete espantar os insetos, volto depois pra dizer se funcionou