Infrastructure architect.
Android, Termux, local LLMs (llama.cpp), Linux and Windows systems.
Builds, debugging, performance tuning. DMs open.github.com/kukirin-ui Croatia Joined September 2026
@filicroval Would love to see the draft acceptance rate next to each number, llama.cpp prints it at the end of every run. That's probably most of the gap between 108 on math and 63 on prose. github.com/ggml-org/llama…
@billadler@madebygoogle Quick workaround: set Preferred network type to LTE (Settings > Network & internet > SIMs > your carrier). If it still drops on LTE, try Reset Mobile Network Settings under System > Reset options. support.google.com/pixelphone/ans…
Doable with llama.cpp's RPC backend. Run ggml-rpc-server on each box with its own backend, then point llama-server at all of them with --rpc and it splits weights and KV cache by each device's free memory. Just keep it off any open network, the RPC server is still marked insecure. github.com/ggml-org/llama…
@TferThomas Worth knowing before jumping to beta: once you're on QPR2 beta you can't go back to stable without a full wipe, unless you wait for the stable release to catch up. If the freezes are bad enough, back up first. Otherwise riding it out until stable lands is the safer call.
@SpencrGreenberg Reading Android bugreports. adb bugreport gives you a huge zip of logs nobody reads by hand. Hand it to an agent, ask which app held wakelocks overnight, and the battery drain culprit shows up in a minute instead of an evening of scrolling through batterystats.
The A3B part is what makes this work. Only about 3B params are active per token, so with llama.cpp you keep attention on the GPU and push the experts to system RAM with --n-cpu-moe. Q4_K_M of a 35B is around 21GB, so 12GB VRAM plus 16GB RAM just fits. One catch: 262K is the configured window. Speed drops as the context actually fills, and quantizing the KV cache with -ctk q8_0 -ctv q8_0 helps a lot there.
The interesting part is the link between the Macs. Thunderbolt 5 is about 80 Gb/s, so roughly 10 GB/s, while one Mac's unified memory does around 800 GB/s. That's why these setups split the model by layers instead of tensor parallel. You get the memory of four machines, but single user speed is closer to one machine plus the hop latency.
@IVHOGIV2 Yeah, virtual private network it won't help the device but it will help you to eliminate US and android coverage. This takes too long for you to experience. DM me, we gonna fix it ftom there.
@IVHOGIV2 You are fine, go to Settings, Apps, X, Storage and tap Clear cache first. If pics still don't load, tap Clear storage there too, then log back in. Also in the X app check Settings and privacy, Accessibility display and languages, Data usage, and turn Data saver off
yeah that makes sense, QNN builds are tied to the exact Snapdragon, so no point running it on my Exynos S26 anyway. that's why I stick with llama.cpp on Vulkan there, slower than a fully delegated NPU model but it runs anything. a CUDA-like layer across all Qualcomm chips would change a lot
@cantliff9@madebygoogle then it's past anything you can fix on the phone. either wait for the next beta build or leave the beta, but leaving before a newer stable build is out wipes the phone, so back up first or wait for the stable release
@IVHOGIV2 then it's X itself. on iPhone go to Settings, General, iPhone Storage, X, Offload App, then reinstall it. also check Settings, Cellular and make sure X is allowed to use mobile data, that one blocks images while text still loads
@JayHSalem@grok fair, wobbly is wear, cleaning won't fix that. the S21 does wireless charging though, so a cheap pad gets you by without touching the port while you wait for the new job and pick the S25
@be4r_kid@telegram@Tesla if it started right after the update and survived a reinstall, it's most likely an app bug, not your phone. check Settings, Battery to confirm Telegram is the drain, report it in Telegram under Settings, Ask a Question, and use web.telegram.org in Safari until a fix lands
@TheOnlyNeonKing@starmexxx bandwidth is exactly why VRAM wins. a 3090 does ~936 GB/s, a 4070 ~504. and once a model doesn't fit in VRAM it spills to system RAM at maybe 60-100 GB/s, that's the real waiting all night case. fit first, then bandwidth
@ggerganov if draft-mtp is drafting from the model's own MTP heads, that's the nice part. speculative decoding without loading a separate draft model matters a lot when the main model already fills most of 128GB
@theo clones outranking the original on a name search is a real supply chain risk. one bad update in a fork and people install it thinking it's yours. worth checking the publisher and repo link before every npm i
@github makes sense, runtime CSS-in-JS has to serialize styles during SSR and inject them again on the client. static CSS Modules take that work out of every render. good to see real numbers on it
23 Followers 32 FollowingSolo indie game studio from Türkiye 🇹🇷 Making small, fun mobile games. Balon Pat! 🎈 in Google Play closed testing. Tester swaps welcome 🤝 #indiedev #gamedev
2K Followers 253 FollowingI want to make you nut just as much as you want to gift me things so we gotta be on a level with each other OK baby? 💋 CA: $enigmaticphantasm Venmo: @MissMe_
80 Followers 239 FollowingTokyo AI company. We measure local LLMs, GPU inference, Japan's LLM rankings and AI on Raspberry Pi, then post the numbers. WireCanal, MotionVox, Bestllam.
192 Followers 358 FollowingVideo Games // Random rambling // Art // Standing in a suspicious alley going: "pssst, want to sack Rome?"
They/Them/Raccoon/Necromancer/Transhumanist
80 Followers 239 FollowingTokyo AI company. We measure local LLMs, GPU inference, Japan's LLM rankings and AI on Raspberry Pi, then post the numbers. WireCanal, MotionVox, Bestllam.
1.6M Followers 268 FollowingThe engine room of @Google. Building AI safely and responsibly to solve the world’s most complex problems. Join us: https://t.co/jUHQA27iBL
183K Followers 49K FollowingDeep in Agents, LLMs and automation · Open source at ♥ · Sharing AI gems daily to help you work smarter · Ex-data scientist @Snowflake | @Streamlit ·
11K Followers 38 Following$0 to $70k/month with my home services marketing agency using FB Ads. On the journey to $100k+
I'll document everything I know using paid ads to get clients
35K Followers 3K FollowingI’m the co-founder and CEO of https://t.co/mrEtbPO2la - a productized SEO platform that has delivered 200,000+ campaigns since 2012.
4.6M Followers 356 FollowingProthom Alo is the most-read newspaper in Bangladesh. The online portal of Prothom Alo is the most visited Bengali website in the world. #ProthomAlo
53K Followers 1K FollowingDevRel Lead @appwrite 🥑📍🛠️ Tinkering with web apps, cloud and AI tools 🧑💻 📍 👨👩👧👦 Bringing writers together at @blrwritersroom 📝
208K Followers 3K Following🌏 Digital Nomad 🇬🇧🇵🇹🇹🇭
⭐️ GitHub Star of the Year | 👨🏫 GitHub Top Teacher Award
🤓 Developer Advocate @paypaldev
🎙️ Views expressed are my own
4.6M Followers 2K FollowingFounder and CEO, O'Reilly Media. Watching the alpha geeks, sharing their stories, helping the future unfold. Didn't pay for a blue check, cannot make it go away