HN user

i386

2,797 karma
Posts49
Comments702
View on HN
github.com 4mo ago

Pool spare GPU capacity to run LLMs at larger scale

i386
11pts3
twitter.com 4y ago

Apple rejected expedited app review over previously reviewed marketing materials

i386
4pts0
www.bbc.com 7y ago

The myth of the eight-hour sleep

i386
3pts2
www.smh.com.au 7y ago

5G Apocalypse: Russia wants you to believe next-gen wireless can kill you

i386
4pts3
www.pri.org 7y ago

If you have a meeting in Ethiopia, you'd better double check the time

i386
4pts0
en.wikipedia.org 7y ago

Time is recorded differently in Ethiopia

i386
3pts0
github.com 7y ago

PlotPlotPlot is a flexible Arduino plotter tool for Mac, Windows and Linux

i386
2pts0
jenkins.io 9y ago

New UX for Jenkins: Blue Ocean 1.0

i386
243pts96
jenkins.io 9y ago

New UI for Jenkins “Blue Ocean” Now in Beta

i386
1pts0
github.com 10y ago

Barmaid: an easy to use web server in your mac/windows system tray

i386
2pts0
jenkins.io 10y ago

Introducing Blue Ocean: a new user experience for Jenkins

i386
16pts1
jenkins.io 10y ago

Partnering with Microsoft to Run Jenkins Infrastructure on Azure

i386
1pts0
developer.atlassian.com 10y ago

Why java.util.Optional is broken

i386
11pts0
engineering.canva.com 11y ago

On workstation, as it is in production

i386
2pts0
news.ycombinator.com 11y ago

Ask HN: Anyone using Keen.io?

i386
1pts0
optiqueapp.com 12y ago

Show HN: Optique is a light and simple iPhoto alternative

i386
5pts0
www.smh.com.au 12y ago

Hasntpaidyet.com: Debt dodgers named, shamed by traders

i386
1pts0
www.smh.com.au 13y ago

Australia's first Muslim frontbencher abused for taking oath on Koran

i386
5pts2
blog.spearapp.com 13y ago

Goodbye

i386
2pts0
thenextweb.com 13y ago

Microsoft introduces Azure for Java

i386
2pts0
chronicle.com 13y ago

A redesign of the handicapped icon

i386
2pts0
www.nbcnewyork.com 13y ago

Marijuana smuggled across US-Mexico border under car using hard disk magnets

i386
7pts0
www.abc.net.au 13y ago

Archaeologists discover 'Gate to Hell'

i386
92pts33
atlassian.com 13y ago

Introducing Atlassian JIRA junior

i386
1pts0
whimsyapps.com 13y ago

Show HN: Always 'paste and match style' on Mac OS X

i386
1pts0
whimsyapps.com 13y ago

Show HN: Hackernews for Mac

i386
35pts49
stackoverflow.com 13y ago

Whats the point of [NSString stringWithString:]

i386
1pts0
gist.github.com 13y ago

A Stylish style to make HN feel less raw

i386
2pts0
phetdreams.tumblr.com 13y ago

Massive Google Play privacy issue

i386
14pts1
news.ycombinator.com 13y ago

Ask HN: What DevOps software and services does your startup or company use?

i386
4pts1

Each stage has its own KV for the layers it hosts. You are on the money there, when one stage is waiting it's free for more parallelism. I am planning on exploiting this for more token verification through ngram spec decoding.

The lab features two Mac Studios: an Apple M3 Ultra (32 CPU cores, 80 GPU cores, 256 GB unified memory) and an Apple M1 Ultra (20 CPU cores, 48 GPU cores, 128 GB unified memory), both connected via 1Gbit Ethernet.

We use a customized Q2 quantization that preserves sensitive tensors at Q8.

To reduce compute time per layer, we are developing a custom GLM DSA Metal graph.

While we are not yet approaching MTP, we plan to port our existing MTP implementations from versions 4.7 and 5.1 to 5.2.

Since GLM's MTP acceptance rate is very high for a single predicted token, we are exploring token prediction techniques to widen the predicted tokens and utilize parallelism for verification.

This was done on my home lab simulating 5ms latency and jitter between machines. Splits work quite well if you your nodes are over WAN at metro latency’s but not super fast on global WAN.

The idea is that you could take several machines without dedicated RDMA or NVLINK fabric and use them to serve a large model on hardware you own then share it with others.

I’m currently working on GLM 5.2 on my lab environment with around 10 tok/s on the same split.

I didn't know that Protobuf wasn't canonical but even without this knowledge, there are many other factors which make it an inferior format to JSON.

If strict typing and versioning are useless to you, sure.

This is very revisionist. While they have been catching up quickly there was no master 4D chess strategy here. Google was incredibly late to this game - Sergey had to come back from retirement because most of the research team had a Sarah Connor complex and couldn’t ship. The saving grace is that AdWords picked up the tab again and founders shook the place up when it became clear the golden goose was being cooked.