Workers AI Update: Hello Mistral 7B

This post is also available in Deutsch.

Workers AI Update: Hello Mistral 7B

Today we’re excited to announce that we’ve added the Mistral-7B-v0.1-instruct to Workers AI. Mistral 7B is a 7.3 billion parameter language model with a number of unique advantages. With some help from the founders of Mistral AI, we’ll look at some of the highlights of the Mistral 7B model, and use the opportunity to dive deeper into “attention” and its variations such as multi-query attention and grouped-query attention.

Mistral 7B tl;dr:

Mistral 7B is a 7.3 billion parameter model that puts up impressive numbers on benchmarks. The model:

  • Outperforms Llama 2 13B on all benchmarks
  • Outperforms Llama 1 34B on many benchmarks,
  • Approaches CodeLlama 7B performance on code, while remaining good at English tasks, and
  • The chat fine-tuned version we’ve deployed outperforms Llama 2 13B chat in the benchmarks provided by Mistral.

Here’s an example of using streaming with the REST API:

curl -X POST \
“https://api.cloudflare.com/client/v4/accounts/{account-id}/ai/run/@cf/mistral/mistral-7b-instruct-v0.1” \
-H “Authorization: Bearer {api-token}” \
-H “Content-Type:application/json” \
-d '{ “prompt”: “What is grouped query attention”, “stream”: true }'

API Response: { response: “Grouped query attention is a technique used in natural language processing  (NLP) and machine learning  Continue reading

NB456: China Flexes Tech Muscles With 1.2Tb Backbone; Will CISOs Mitigate Risk Or Liability?

China demonstrates the strength of its home-grown technology industry by announcing a 1.2Tbit Internet backbone that stretches 3,000 kilometers, Fortinet warns of a 9.8 severity bug in its SIEM product, and Nvidia and Dell roll out ready-made models and infrastructure for enterprises to build their own generative AI apps. HPE Aruba Networking releases a new... Read more »

NB456: China Flexes Tech Muscles With 1.2Tb Backbone; Will CISOs Mitigate Risk Or Liability?

China demonstrates the strength of its home-grown technology industry by announcing a 1.2Tbit Internet backbone that stretches 3,000 kilometers, Fortinet warns of a 9.8 severity bug in its SIEM product, and we continue our ongoing discussion about the precarious role of the CISO, plus more tech news, in this episode of the Network Break podcast.

The post NB456: China Flexes Tech Muscles With 1.2Tb Backbone; Will CISOs Mitigate Risk Or Liability? appeared first on Packet Pushers.

2024, the year of elections

2024, the year of elections

2024 is a year of elections, with more than 70 elections scheduled in 40 countries around the world. One of the key pillars of democracy is trust. To that end, ensuring that the Internet is trusted, secure, reliable, and accessible for the public and those working in the election space is critical to any free and fair election.

Cloudflare has considerable experience in gearing up for elections and identifying how our cyber security tools can be used to help vulnerable groups in the election space. In December 2022, we expanded our product set to include Zero Trust products to assist these groups against new and emerging threats. Over the last few years, we’ve reported on our work in protecting a range of election entities and as we prepare for the 2024 elections, we want to provide insight into attack trends we’ve seen against these groups to understand what to expect in the next year.

For this blog post, we identified cyber attack trends for a variety of groups in the elections space based in the United States, as many of our Cloudflare Impact projects provide services to these groups. These include U.S. state and local government websites protected under Continue reading

VXLAN/EVPN Layer-3 Handoff (L3Out) on Arista EOS

A while ago, I published a blog post describing how to establish a LAN/WAN L3 boundary in VXLAN/EVPN networks using Cisco NX-OS. At that time, I promised similar information for Arista EOS. Here it is, coming straight from Massimo Magnani. The useful part of what follows is his; all errors were introduced during my editing process.

In the cases I have dealt with so far, implementing the LAN-WAN boundary has the main benefit of limiting the churn blast radius to the local domain, trying to impact the remote ones as little as possible. To achieve that, we decided to go for a hierarchical solution where you create two domains, local (default) and remote, and maintain them as separate as possible.

Read more …

VXLAN/EVPN Layer-3 Handoff (L3Out) on Arista EOS

A while ago, I published a blog post describing how to establish a LAN/WAN L3 boundary in VXLAN/EVPN networks using Cisco NX-OS. At that time, I promised similar information for Arista EOS. Here it is, coming straight from Massimo Magnani. The useful part of what follows is his; all errors were introduced during my editing process.

In the cases I have dealt with so far, implementing the LAN-WAN boundary has the main benefit of limiting the churn blast radius to the local domain, trying to impact the remote ones as little as possible. To achieve that, we decided to go for a hierarchical solution where you create two domains, local (default) and remote, and maintain them as separate as possible.

Read more …

Weekend Reads 111923

Bad queries tend to propagate to the root zone due to the hierarchical nature of DNS, so studying traffic at a root server can provide key insights into overall network usage.

This blog covers an interesting case of suspected abuse in a gTLD registry between February and April 2023.


YouTube wants its pound of flesh. Disable your ad blocker or pay for Premium, warns a new message being shown to an unsuspecting test audience, with the barely hidden subtext of “you freeloading scum.”


The shift towards chiplet architecture is inevitable for almost all high-end CPUs/GPUs, accelerators, and networking silicon vendors. It is not a question of ‘if’ but ‘when’.


The Global Coalition on Telecommunications (GCOT) purports to be about synching up how the five countries approach telecoms. The scope of corporation includes information sharing, joint R&D, funding alignment, the development of standards, skills, supply chain diversification, security, and 6G, so says the release.


Securing mainframes remains top of mind, with 61% of mainframe and IT professionals ranking security as the top problem they are facing, according to BMC’s annual survey of mainframe users for 2023.


Gartner has raised the specter of departments outside of tech running their own IT Continue reading

SDR transmit and clean signals

If you have a transmit capable SDR, you may have heard that you need to filter its output, before transmitting to the world. Certainly before amplifying the signal.

I have a TinySA Ultra spectrum analyzer, and will here show you some screenshots about just how true that is.

I tested this with my USRP B200, transmitting a pure carrier around 145MHz and 435MHz.

Oh, and a word of caution: If you want to replicate this, make sure to add an inline attenuator, to not damage your spectrum analyzer. I had a cheap 40dB one, but the values in the graphs have been adjusted to show the real signal strength, as if I hadn’t.

tl;dr

  1. Harmonics can be almost as strong as the fundamental. You need to filter these.
  2. Transmitting at maximum output gain may cause lots of unwanted signals right around your fundamental. You cannot filter these. You need to not generate them.

Harmonics

Harmonics for 145MHz Harmonics for 435MHz

Reducing the output gain did not meaningfully fix the problem. The best I saw from using half output gain was to make the strongest harmonic 9dB less than the fundamental. That’s way too strong.

I added a cheap band pass filter (FBP-144), which made Continue reading

AX.25 and 9600bps G3RUH decoding

I’ve been coding more on my rust SDR framework, and want to improve my ability to send/receive data packets efficiently and reliably.

There are two main ways I use learn to do this better: designing a new protocol, and making the best implementation possible for an existing one. This post is about refining the latter.

AX.25 and APRS

First a detour, or background.

AX.25 is the standard amateur radio data protocol. It’s mostly an OSI layer 2-4 protocol, mashing the layers together into one. Contrast this with IP, which just encapsulates the next layer.

Layer 3 (IP stack equivalent: IP itself) consists of the ability to add, in addition to source and destination, a variable number of intermediate repeaters. This allows limited source routing. In APRS the repeaters are usually not named, but instead uses “virtual” hops like WIDE1-1.

Layer 4 (IP stack equivalent: TCP and UDP) allows both connected and disconnected communication channels. In my experience connected AX.25 works better over slow simplex radio than TCP. If TCP was ever optimized for high delay low bandwidth, it’s not anymore.

For the physical layer, there are three main “modems”:

  1. 300 baud bell 103, used Continue reading

Pushing The Limits Of HPC And AI Is Becoming A Sustainability Headache

As Moore’s law continues to slow, delivering more powerful HPC and AI clusters means building larger, more power hungry facilities. …

The post Pushing The Limits Of HPC And AI Is Becoming A Sustainability Headache first appeared on The Next Platform.

Pushing The Limits Of HPC And AI Is Becoming A Sustainability Headache was written by Timothy Prickett Morgan at The Next Platform.

SC23 Over 6 Terabits per Second of WAN Traffic

The world’s fastest temporary internet service gets turned on in Denver for one week only describes the SCinet temporary network built to support the The International Conference for High Performance Computing, Networking, Storage, and Analysis (SC23) this week in Denver. The SC23 WAN Stress Test chart demonstrates that the provisioned 6.71 terabits bits per second capacity was pushed to the limits.
SC23 SCinet traffic describes the architecture of the real-time monitoring system used to comprehensively monitor the SCinet network and generate these charts. This chart shows that over 175 Petabytes of data were transfered during the show.
SC23 Dropped packet visibility demonstration describes a joint demonstration by InMon Corp and Arista Networks of one of newest developments in sFlow telemetry, identifying every dropped packet, the reason it was dropped, and the location it was dropped across all the switches in real-time.
SC23 WiFi Traffic Heatmap shows a real-time view of WiFi usage at the conference displayed on a conference floorplan.
Finally, SC23 Data Transfer Node TCP Metrics demonstrates how standard metrics maintained by the Linux kernel can be used to augment sFlow telemetry and track the performance of large science data transfers.

How to execute an object file: Part 4, AArch64 edition

How to execute an object file: Part 4, AArch64 edition

Translating source code written in a high-level programming language into an executable binary typically involves a series of steps, namely compiling and assembling the code into object files, and then linking those object files into the final executable. However, there are certain scenarios where it can be useful to apply an alternate approach that involves executing object files directly, bypassing the linker. For example, we might use it for malware analysis or when part of the code requires an incompatible compiler. We’ll be focusing on the latter scenario: when one of our libraries needed to be compiled differently from the rest of the code. Learning how to execute an object file directly will give you a much better sense of how code is compiled and linked together.

To demonstrate how this was done, we have previously published a series of posts on executing an object file:

The initial posts are dedicated to the x86 architecture. Since then the fleet of our working machines has expanded to include a large and growing number of ARM CPUs. This Continue reading

1 … 187 188 189 190 191 … 3,834