Advented

@MiaAI_lab They way I understand is Sequences mean concurrent requests. So your recipe can serve 6 requests at 1M context concurrently. Under full/heavy load t/s can spike and drop bc the longest 1M session has to finish generating for other concurrent requests to continue

@advented_ Do you mean single session? Unfortunately I'm not getting 40-50 t/s, it was unstable. My recipe gets me to 31.7 t/s for single session. For multiple sessions I get up to 80 t/s

53 tok/s achieved on Step-3.7-Flash NVFP4 with MTP on 2x DGX Spark with 256k context. 🎉🥳 Elapsed time: 56.694 s Prompt tokens: 29 Generated tokens: 3000 Total tokens: 3029 Generation tok/sec: 52.92 End-to-end tok/sec: 53.43

milestones achieved: > setup codex, claude CLI, > run Deepseek V4 locally > research on KiCad ai agent pipeline with tscircuit, atopile, diode, KiCad MCP > hand off to codex as goal with Modos E-ink circuit analysis Next step: > wake up tomorrow and see what my results 😎🔬

@MiaAI_lab I reviewed your Ds-v4 recipe against my own. You’re actually doing better than me! I falsely assumed my config was the most I could get for 2 DGX setup (200k, 5 seqs, 12k batch prefil) But turns out your recipe does way better (1M, 6 seqs, 8kbatch) for the same 40-50 t/s !!

@advented_ Sounds awesome, care to share the recipe? I will share mine when I feel it's stable enough.

@0xSero i've been trying to search around for the best DGX spark sg alng recipes.. your reap looked interesting and I seen promising results on dev forum getting this model running native on 2 DGX sparks.. benchmarks would be a huge help to determine recipe and best use case!

I’m collecting benchmarks for Deepseek-V4-Flash I’m almost done.

@willreil real question: how would you keep the yard green? is it all astro turf? else make the roof entirely electrochromatic glass to control sun light in the neighbourhoods. consistent weather all year round lol

the european mind cannot comprehend the idea of air conditioned neighbourhoods

@StockWorthyApp @StockMKTNewz major disappointment. I was holding warby parker calls for the anouncement and watched them get COOKED live

@StockMKTNewz "We have $META glasses at home"

Google $GOOGL will be launching these audio only smart glasses later this year

@0xSero I’ve been trying to build an MCP with it. After long enough chat it can hallucinate bad It’ll say ready to implement the plan but then never take action. Even at 30% context window usage Some ppl saying it might be the system prompt that breaks model over time

Fellas, is it gay to use Grok Build? - Model is very fast - The model is called "Grok Build" - It follows instructions well - It is relatively logical - The TUI is really nice, except for the 6 spinners... The bad: - It's sloppy with code | video soon

@bubbleboi without reservation lol

I walked into the house of prime rib and told them I was bubble boi and they got me a table immediately.

@r0b0t_sp1der @alexinexxx we live in a world where elon can hire devs on the spot when its go time to lock in talent. pandering for corporate just has no upside when you taste and see what high agency can be

@alexinexxx people who do participate in it get highly rewarded plenty of others who get walked in the front door without it, of course

God’s plan for me doesn’t involve a 6-round application process

@skcd42 @Daniel_Farinax based 🫡

@Daniel_Farinax tool calls remain the same, you obviously loose some information about the tool specific prompting which we use, but no point hiding it tbh. anyone can use strings and get access to the system prompt either way

I jailbroke the Grok Build system prompt 🤯 Removed sections like: - “don’t add features the user didn’t ask for - “don’t refactor” - “don’t add error handling” and similar restrictions. The model is incredibly powerful. I’m currently testing system prompt variations to unlock more creativity. The problem with rules like “don’t add features the user didn’t ask for” is that the model will only do exactly what’s requested and won’t proactively add improvements or surprises. Still in early testing, will share concrete findings soon.

@Daniel_Farinax @StudioZamudio i think Opencode is around the same size for thier system prompt. intitial prompt or fist chat is the system prompt getting consumed. sharing it is typically apart of the closed model aspect of frontier models. grok cli is interesting. i love and hate it so far

@StudioZamudio I agree, I'm noticing 15-20k tokens for a simple hello :D

@tueks3 the chinese do this with Fio BTR product with many audio codecs support. this module is powerful because it can be adapted for anything! In college i worked on spatial audio, I wonder how easy it could be to add gyroscope and processing unit to enable this device!

@iamgingertrash you think leopold gonna keep his trades private or he offloaded to crypto?

Two powerful crypto catalysts are about to arrive; > privacy maxxing to hide gains made in the semi trade, and de-risk > equities onchain (leading up to and post clarity Senate vote) The mega catalyst; (agent economies) Won’t arrive until 27’, 28’

@0xSero Docker containers defined alloc of resources but sometimes go over or under the RAM durring runtime. Would be nice to have for debug and more consistent recipe deployment

@0xSero I like vllm studio. I tried running it on my Spark recently but it didn’t support container deployment of vllm. Have you added support since then? I vibe coded my own patch but still haven’t worked out the bugs with live resource observability.

It had to happen. - deepseek-v4-flash reap on 1 spark - integration with vLLM-studio - research multi-device inference - dynamo disaggregated inference

This man is a legend and this project is truly amazing. He went from zero to selling an assembled pcb on his website and shipping it internationally in a single project. Shipping the thing is as impressive as making it, congrats will. You guys should buy the thing NOW, it's $1

@Tri_Stanisaurus @willreil the 10x referral upsell, amazing lol

@willreil eh, i got a guy who sells equipment, often full production lines. Idk if he ever has lumens but he did offer me a full pcb line last month with an xray inspection machine for like 25k.... so...

Once I make enough from selling circuit boards I am going to buy a pick and place machine to sell more circuit boards to buy more pick and place machines

it's too risky to not take your dreams seriously