#Paged

Update: I could not go back to sleep loll Getting paged is a normal part of software engineering tho, especially when you have services that are critical for the business. Working on stuff that’s rlly high impact like Amazon orders software is cool, but also entails nights like this. It’s just apart of the GAMEE, no stresssss just lost sleep loll #oncall #softwareengineer #techjobs

Getting paged sucks. But that’s just part of the game. this time i got paged at 3 am… If you’re wondering what the issue was, we had an app that was set to 200gb of storage MAX but then it tried mounting 279gb at once. Because the way replication works with that service, it writes 3 times at once. So we had 79gb tried to be written at once. #swe #softwarengineer #csmajor #bigtech #growth

1st being oncall and got paged 😩😭 . . #softwareengineering #oncall #tech #techlife #corporate

Getting paged is a normal part of software engineering tho, especially when you have services that are critical for the business. Doesn’t matter how much redundancy, retries, etc you have, things can always go wrong. Working on stuff that’s rlly high impact like Amazon orders software is cool, but also entails nights like this. It’s just apart of the GAME, no stresssss just lost sleep loll. happens #oncall #softwareengineer #techjobs

Getting paged as an oncall software engineer with a reverse shadow! #oncall #softwareengineer

Getting paged as night time oncall

Thankfully those days are behind me; but want to share what residency life is like. Yes the pay is like $50-60k for 60-80 work weeks (some specialities are even longer weeks). ACGME requires residents have 4 days off every month, and that’s often all you get unless you take vacation. A “golden” weekend refers to having a normal 2 day weekend lol. Home call is one of the worst creations because you’re not actually at work but get paged at any time at night while you’re sleeping, only to go back to work the next day regardless of what happened the night before. And yes sadly many a 24 hour call transform into 30 hours of wakefulness 😩 #femalesurgeon

Don’t you love it when things don’t go as planned, like getting paged on a Friday night 🫠 2 weeks ago, I said I was gonna build a music career while working full time in software engineering, and moments like these make me wonder if I’ll ever make it in the long term. But we’re showing up everyday, the reps only count when it starts to hurt 💪 P.S. it took me another 8 hours to figure out what was wrong 😭 Follow @onesemblemusic to see where we end up this summer! #musiccareer #musicfestivals #oncall #softwareengineer #oncallshift

1. Batching strategies -> group multiple requests together to improve GPU utilization vs latency tradeoff 2. Paged attention -> manage KV cache in small memory pages to avoid fragmentation and scale efficiently 3. Speculative decoding -> use a small model to guess tokens and a large model to verify them in parallel for speed 4. KV cache offloading -> move KV cache from GPU to CPU/disk to save memory at the cost of access latency 5. Data parallelism -> run multiple copies of the model on different GPUs handling different requests 6. Tensor parallelism -> split model layers across GPUs to fit and run very large models 7. Pipeline parallelism -> divide model into stages across GPUs to process sequentially in a pipeline 8. Prefill vs decode -> prefill processes the prompt in parallel while decode generates tokens sequentially 9. Disaggregated serving -> run prefill and decode on separate hardware optimized for each phase 10. KV cache quantization -> compress KV cache using lower precision to reduce memory usage 11. KV cache pruning/sliding window -> drop or limit stored tokens to control memory growth 12. Quantization -> reduce model weight precision (e.g. INT8/INT4) to speed up inference and save memory 13. Flash attention -> optimize attention computation to reduce memory movement and improve speed 14. Continuous batching -> dynamically add/remove requests during generation to maximize throughput 15. Scheduling -> control request ordering and resource allocation to balance latency and throughput BIG IDEA → inference optimization is balancing memory, latency, and hardware efficiency #learningtogether #womeninstem #aiengineering #tech #consistency

At Aleja Róż 10, technology meets craftsmanship, and history meets the present. Thank you for celebrating the opening of the Paged × Tylko × Pani Jurek showroom with us! #alejaroz10/14 _______________ W kamienicy przy Alei Róż 10 spotykają się technologia i rzemiosło, historia i współczesność. Dziękujemy za wspólne świętowanie otwarcia showroomu Paged × Tylko × Pani Jurek!

As the junior resident on call, I’m first call, which means I’m the first to get paged for incoming consults or any questions or concerns about our inpatients on the ward. In our program, plastic surgery call is city wide, so we cover three different hospitals across the city. Some nights are busier than others, but there’s always a lot to learn!

sorry i didn’t reply, i was busy reading this 1000-paged fantasy book 🫖🕯️🧖🏼‍♀️💤 #bookstagram #cozyroom #readingvlog #sliceoflife #explore cozy life, aesthetic vibes, pinterest inspired, book lover, bookworm, a day in my life, cozy lifestyle, cozy room, introvert diaries, slow living, fantasy books, the dandelion dynasty, ken liu, girl therapy, currently reading