Sabitlenmiş Tweet
TechMD
3.4K posts
TechMD
@TechMDAI
🔬 MD fueled by a deep passion for Medicine & Tech. 🌐 Exploring the frontiers of VR/AR/MR & BCI. 🤖 AI enthusiast.
Katılım Haziran 2023
2.5K Takip Edilen689 Takipçiler
. @NVIDIAAI We've been running GLM-5.2 across 4× DGX Sparks and hit a class of multi-node NCCL deadlocks that appear to affect the broader Spark community. We did a deep root-cause analysis.
The key finding is that stock DGX Spark lacks GPUDirect RDMA, making host-staged collectives 10-100× more susceptible to TP race conditions. Full write-up with evidence and community references here: github.com/marksunner/dgx…
English
English
I hear we are posting our @NVIDIAAI clusters this morning, so here we go:
4X GB10's with CRS804 RoCE for the fabric.
10GBE backhaul to the NAS in the other rack.
All in a 12U 10" rack.
SOTA AI, in the corner of my den.
Currently serving GLM-5.2 NVFP4
English
Wearing noice cancelling masks to talk to Claude is crazy
English
@thatcofffeeguy When you’ve spent the whole budget on gb10s and ultras and the robot body has to come from the packaging.
3090 boxes for arms 😂
English
I have built what I call my robot of boxes with my collection of hardware.
6 Asus GX10s, 2 m3 ultras, 2 m4 minis, 1 m4 pro, and you can see in photo the MacBooks hat is a rapsbery pi 5 cana kit.
I think the last 2 sparks I want will also finish up his legs/shoes. No idea how to make the arms work. Little hardware humor for the night.
@NVIDIAAI @mr_r0b0t
English
@loktar00 @thsottiaux @sama Yes I can have more than 1 codex desktop open at once. Open in new window is live
English
English
I think having the ability to have more than 1 codex instance is huge. Thanks @thsottiaux @sama
English
GLM 5.2 now pushing past 18 tok/sec on 2 DGX Sparks 🔥🔥🔥
This time I quantized the 3 biggest attention projections to NVFP4 (compared to the original FP8)
There are still wins to be had on the speed and quality fronts so stay tuned
Joe Muller@BosonJoe
Played a few tricks and tripled the speed of GLM 5.2 on 2 DGX Sparks 🔥 Old: 4.1 tok/sec New: 13.8 tok/sec 2-bit quantization, TP2, and a frequency based prune of the experts to get planes under the ~88 GB page-cache ceiling
English
A great Hermes agent btw, Lenovo Legion Go S. PC gamers have more resources at their disposal than they think.
Coupled with @Tailscale, it a deadly combo for robotics and agentic work.
English
Building AI agents is like raising kids who never sleep.
You think the hard part is teaching them to walk
The hard part is making sure they don't walk off a cliff while you're not looking.
Guardrails, not goals.
What's the first thing you handed to an agent?
#buildinpublic #aiagent
English
I’m loving this DGX Spark set up. I need a switch from MikroTik
jon@lursor_lover
New limited edition spark just dropped
English
@fbwalker4 That’s super smart. I like that, seems your workflow is very advanced. Thanks for the tip btw
English
This guy is pushing the limit with 2 sparks. GLM at home on 2 sparks
Joe Muller@BosonJoe
Played a few tricks and tripled the speed of GLM 5.2 on 2 DGX Sparks 🔥 Old: 4.1 tok/sec New: 13.8 tok/sec 2-bit quantization, TP2, and a frequency based prune of the experts to get planes under the ~88 GB page-cache ceiling
English
@fbwalker4 Ya fable is amazing and a class above glm.
It’s pretty sweet to get glm 5.2 doing unlimited work for you with relatively low energy burn. Idk I don’t know; I find that amazing, especially when developing autonomous loops and agents running.
English
Well, choosing the right approach depends on your goals and comfort level when planning to reduce cloud use. While the frontier offers impressive capabilities, not every task requires it. Many local models perform adequately for most purposes so the decision comes down to whether to use an API or not. I just bet on myself to develop my own ai infrastructure for continued learning.
English