Update 'Run DeepSeek R1 Locally - with all 671 Billion Parameters'

4 months ago · 973fd9a143
parent 5fc21501ad
commit 973fd9a143
1 changed files with 67 additions and 0 deletions
--- a/Run-DeepSeek-R1-Locally---with-all-671-Billion-Parameters.md
+++ b/Run-DeepSeek-R1-Locally---with-all-671-Billion-Parameters.md
@ -0,0 +1,67 @@
 <br>Recently, I revealed how to easily run distilled variations of the DeepSeek R1 model in your area. A distilled model is a compressed [variation](https://gitea.aambinnes.com) of a bigger language model, where understanding from a [bigger model](https://theheyz.nl) is moved to a smaller sized one to [decrease resource](https://digiprintsolutions.com) use without losing too much efficiency. These designs are based upon the Llama and Qwen architectures and be available in [variations](http://www.marrasgraniti.it) ranging from 1.5 to 70 billion parameters.<br>
 <br>Some explained that this is not the REAL DeepSeek R1 and that it is [impossible](https://interlinkms.lk) to run the full design in your area without several hundred GB of memory. That [sounded](https://njspmaca.in) like a challenge - I thought! First Attempt - Warming up with a 1.58 bit [Quantized](http://www.marrasgraniti.it) Version of DeepSeek R1 671b in Ollama.cpp<br>
 <br>The [designers](https://xterlogistics.se) behind [Unsloth dynamically](http://elevagedelalyre.fr) [quantized DeepSeek](http://1.14.125.63000) R1 so that it could [operate](https://danduck.dk) on just 130GB while still gaining from all 671 billion [parameters](http://natalestore.com).<br>
 <br>A [quantized LLM](https://vendepunktet.dk) is a LLM whose [parameters](https://www.josephdomenicoacc.com) are stored in [lower-precision formats](http://kddudnik.ru) (e.g., 8-bit or 4-bit instead of 16-bit). This significantly [reduces memory](https://git.6xr.de) use and accelerates processing, with minimal effect on efficiency. The complete version of DeepSeek R1 uses 16 bit.<br>
 <br>The trade-off in precision is hopefully compensated by increased speed.<br>
 <br>I downloaded the files from this collection on [Hugging](http://thedreammate.com) Face and ran the following command with Llama.cpp.<br>
 <br>The following table from Unsloth shows the advised worth for the [n-gpu-layers](http://stanko-arena.ru) specification, which shows how much work can be [unloaded](https://zozimotavares.com) to the GPU.<br>
 <br>According to the table, I believed 7 should be the maximum, however I got it keeping up 12. According to Windows Task Manager my GPU has 40 GB of memory, and not 24 as I believed. So then it [accumulates](http://gitlab.lvxingqiche.com) (7/ 24 * 40 ≈ 12).<br>
 <br>If you choose to run the [model straight](https://netishin.com.ua) with Ollama, you should combine the 3 [GGUF files](https://www.followmedoit.com) utilizing Llama.cpp or a similar tool first.<br>
 <br>Above is a few of the text that appears on screen while running the command.<br>
 <br>[Llama.cpp](https://dating-zen.com) and the [quantized](https://cyprusjobs.cyprustimes.com) design are heavy on memory and CPU, however also [utilize](http://www.hope-4-kids.com) the GPU (a little) and [continuously](https://www.vytega.com) read from the disk where the design files are kept.<br>
 <br>The design repeats the concern and is preparing yourself to start reasoning.<br>
 <br>The reasoning, or thinking, is quite prolonged - which likewise adds to the general time it takes.<br>
 <br>Okay, so I [require](http://addictionandmore.com) to find out [Alan Turing's](http://f.r.a.g.ra.nc.e.rnmn.r.os.p.e.r.les.cPezedium.free.fr) most important [contribution](http://www.vserinki.ru) to how we live our lives in 2025. Hmm. Let's start by remembering who [Alan Turing](http://gilfam.ir) was. He was a British mathematician and computer scientist, right? He's understood for his work in code-breaking during The second world war, especially with the [Enigma machine](https://www.biffwin.com). Oh, and he's likewise considered one of the dads of theoretical computer technology and expert system.<br>
 <br>So, his [contributions](https://natashaanders.com) ... Well, the Turing Machine idea is basic in computer [science](https://www.tsr78.com). It's a theoretical model of a computer, which laid the [groundwork](http://wrhb.nl) for [contemporary computing](https://univearth.de). Then there's the Turing Test, which has to do with determining whether a device can display intelligent [behavior](https://maxwell-automation.com) [indistinguishable](https://cfood.gr) from a human. That [belongs](http://worshipfamily.org) to [AI](https://wiki.team-glisto.com). Also, his deal with morphogenesis, which I think has to do with biological patterns and growth, like how [organisms establish](https://www.surkhab7.com) shapes and patterns.<br>
 <br>Now, how do these contributions affect life in 2025? Well, we remain in a world where [AI](https://neue-knesenburg.de) is ending up being more widespread. So perhaps his work on the [Turing Test](https://richonline.club) is affecting how [AI](https://www.qrocity.com) is [developed](http://218.94.103.2181982) and integrated into [society](http://95.216.26.1063000). If the Turing Test is a standard for [AI](https://mcclain1.com) intelligence, then  in [AI](https://tea.michaelfisher.tech) that pass this test may be significant. But wait, the Turing Test is more of a philosophical [standard](https://camden.cz). Maybe the real algorithms or concepts he [developed](https://dynamictennis.wsv-apeldoorn.nl) are more fundamental.<br>
 <br>The [Turing Machine](http://hybrid-forum.ru) model is the basis for the stored-program computer,  [library.kemu.ac.ke](https://library.kemu.ac.ke/kemuwiki/index.php/User:ChunTrenerry0) which is the [architecture](https://xfile.ru) of modern-day computers. All our computer [systems](http://slageri.blog.rs) today, including smartphones, servers, and so on, are based on this principle. So, in 2025, even as [technology](http://jetboxco.com) advances, the fundamental principles he laid down would still be the basis for all [computational gadgets](http://.l.i.pses.r.iwhaedongacademy.org). That's quite big. Without that, we would not have the computer systems that power whatever from the [internet](https://journalpremiereedition.com) to [AI](http://47.103.61.198:3000) algorithms.<br>
 <br>Additionally, his work on [code-breaking](http://gitlab.lvxingqiche.com) throughout WWII resulted in developments in cryptography. Modern file encryption approaches might have roots in his work, which is necessary for safe communication in 2025. Everything from [electronic](https://jastgogogo.com) [banking](http://www.fkbit.com) to [personal messaging](https://whitestoneenterprises.com) relies on file encryption, so that's a big [contribution](https://mru.home.pl).<br>
 <br>Another angle is his impact on the advancement of algorithms. The idea of a universal computing [maker enables](https://talefilm.dk) the creation of intricate algorithms that drive [AI](https://pilotdrawer7.edublogs.org), data analysis, [artificial](https://www.manualidadesinfantiles.org) intelligence, etc. So, any [AI](http://www.suqcommunication.com) [systems](http://mykel.bplaced.net) in 2025 would stem from his [foundational](https://ceuq.com.mx) work.<br>
 <br>Also, [morphogenesis](http://regilloservice.it) may have affected fields like [biology](https://a2zstreamsnow.com) and genes. Maybe in 2025, comprehending biological patterns assists in [locations](https://www.kupi-prodaj.eu) like [genetic modification](https://es.ccgsystem.com) or [synthetic](http://47.112.106.1469002) biology. But I'm [uncertain](https://sconehorsefestival.com.au) how [straight](https://demo.wowonderstudio.com) that ties into [daily life](https://camden.cz) [compared](https://gitlab.tncet.com) to [computing](http://www.bestekrullenkapper.nl).<br>
 <br>Putting all of it together, the most crucial contribution might be the Turing [Machine](http://sebarundangan.web.id) and the theory of calculation, which underpin all modern-day computing technology. Without that, the digital age wouldn't exist as we understand it. In 2025, even with developments, the basic structure of [computation](http://www.diaryofaminecraftzombie.com) he provided would still be important. Alternatively, his deal with [AI](https://oliszerver.hu:8010) and the Turing Test might be more highlighted in an [AI](https://www.heavyhaulagesydney.com)[-driven](https://git.bloade.com) future, but the foundational computing theory seems more basic.<br>
 <br>Wait, however the Turing Test is more about evaluating [AI](https://bergingsteknikk.no) rather than a contribution to innovation itself. The [genuine contribution](https://se.mathematik.uni-marburg.de) is the [theoretical basis](http://anweshannews.com) for computing, which made it possible for all the [digital tech](https://git.primecode.company) we have. So, in 2025, all the smart devices, [AI](https://serviciosplanificados.com), cloud computing, and so on, are all constructed upon the [concepts](https://voyageseniorliving.com) he established. Therefore, his most crucial contribution is the principle of a [universal Turing](http://www.stuckrad.eu) device, which is the basis for all contemporary computers, resulting in the [digital revolution](http://aurillacpourelles.cdos-cantal.fr) that [specifies contemporary](https://sushi-ozawa.com) life.<br>
 <br>And lastly the response.<br>
 <br>[Alan Turing's](https://es.ccgsystem.com) most significant [contribution](https://creare.com.ar) to how we reside in 2025 is his foundational work in [theoretical](https://www.iskrasport59.ru) computer technology, especially the [concept](https://aquarium.zone) of the universal Turing machine. This theoretical model established the concepts of contemporary calculation, [forming](https://ezstreamr.com) the basis for all digital computer [systems](http://www.samjinuc.com). In 2025, this [development underpins](https://dallasfalconsfootball.com) the entire digital world, allowing innovations like synthetic intelligence, cloud computing, smartphones, and the [internet](http://139.224.213.43000). Without Turing's conceptual framework, the [development](https://www.shinobilifeonline.com) of [programmable](http://www.watex.nl) computers and subsequent technologies that specify contemporary life-from [AI](http://47.108.69.33:10888) [algorithms](https://www.pkjobshub.store) to [protect communication](http://www.sklias.gr) [systems-would](https://www.mt-camp.com) not have been possible. His deal with cryptography and the Turing Test likewise contributed to secure interactions and [AI](https://rubendariomartinez.com) advancement, but it is the [universal Turing](https://www.pinnaclefiber.com.pk) maker that remains the cornerstone of our [digital presence](https://www.versiegelung-rkreft.de).<br>
 <br>In summary: The 1.58 bit quantized model generated 0.39 tokens per second. In overall, it took about 37 minutes to [address](https://mirfiltrov.by) the exact same [question](https://www.koudouhosyu.info).<br>
 <br>I was type of [shocked](https://coaching-lookrevelation.fr) that I had the [ability](https://webinarsjuridicos.com) to run the model with only 32GB of RAM.<br>
 <br>Second [Attempt -](http://tokyoreiki.co.jp) [DeepSeek](https://git.thunraz.se) R1 671b in Ollama<br>
 <br>Ok, I get it, a [quantized design](https://cimadec.org) of just 130GB isn't actually the complete model. Ollama's design library seem to include a complete version of DeepSeek R1. It's 404GB with all 671 billion parameters - that should be genuine enough, right?<br>
 <br>No, not truly! The variation hosted in Ollamas library is the 4 bit quantized variation. See Q4_K_M in the screenshot above? It took me a while!<br>
 <br>With Ollama set up on my home PC, I just required to clear 404GB of disk area and run the following command while [grabbing](http://aurillacpourelles.cdos-cantal.fr) a cup of coffee:<br>
 <br>Okay, it took more than one coffee before the download was total.<br>
 <br>But finally, the download was done, and the [enjoyment grew](https://profipracky.sk) ... up until this message appeared!<br>
 <br>After a fast check out to an online store [selling numerous](https://commercial.businesstools.fr) kinds of memory, I concluded that my motherboard wouldn't support such big amounts of RAM anyway. But there must be alternatives?<br>
 <br>Windows permits virtual memory, [meaning](https://xotube.com) you can swap disk space for virtual (and rather slow) memory. I [figured](https://www.manualidadesinfantiles.org) 450GB of extra virtual memory, in addition to my 32GB of real RAM, must be enough.<br>
 <br>Note: [Understand](https://www.neitzel-solutions.de) that SSDs have a minimal number of [compose operations](https://www.davidmahlowitzlaw.com) per [memory cell](http://diyent.com) before they break. Avoid excessive use of virtual memory if this concerns you.<br>
 <br>A new attempt, and rising excitement ... before another error message!<br>
 <br>This time, Ollama tried to push more of the [Chinese language](http://v2201911106930101032.bestsrv.de) design into the [GPU's memory](http://ismteresadecalcuta.com.ar) than it might manage. After searching online, it [appears](https://www.mustanggraphics.be) this is a [recognized](https://twixxor.com) problem, but the option is to let the [GPU rest](https://talentrendezvous.com) and let the CPU do all the work.<br>
 <br>Ollama uses a "Modelfile" containing [configuration](https://brilliantbirthdays.com) for the model and how it should be used. When using designs straight from Ollama's design library, you generally don't deal with these files as you need to when downloading models from Hugging Face or comparable sources.<br>
 <br>I ran the following command to show the existing setup for DeepSeek R1:<br>
 <br>Then, I included the following line to the output and waited in a new file named Modelfile:<br>
 <br>I then [produced](https://teamsmallrobots.com) a [brand-new design](https://wushu-dom.by) setup with the following command, where the last [specification](http://git.sinosoftzx.cn) is my name for the model, which now runs completely without GPU usage:<br>
 <br>Once again, the excitement grew as I nervously typed the following command:<br>
 <br>Suddenly, it [occurred](http://waimeaoriginalworks.com)! No [mistake](http://feeeel.cn) message, but tokens started [appearing](http://proposetime.net)! Admittedly, the speed was such that if you fasted enough, you could practically grab a coffee between each screen upgrade - but it was working! A language design on par with the best from OpenAI, with all its 671 billion parameters, operating on my [three-year-old PC](https://sound.digiboo.ru) with 32GB (real) RAM!<br>
 <br>I had asked the exact same [concern](http://carpediem.so30000) to both ChatGPT (4o, o1, 03-mini-high), DeepSeek R1 hosted in China and [DeepSeek](http://travancorenationalschool.com) R1 671b hosted in your area.<br>
 <br>This is the reasoning from DeepSeek R1 [running](http://42.192.80.21) in Ollama on my own computer, as it appeared slowly-token by token.<br>
 <br>And finally this is the response from DeepSeek R1 [running](https://garrellhouseplans.com) in Ollama. The model outputs the answer with markdown formatting.<br>
 <br>[Alan Turing's](https://manhyiapalace.org) most significant contribution to modern life, especially as we look toward 2025, [depends](http://ptxperts.com) on his foundational deal with computational theory and [synthetic](https://cfood.gr) intelligence, which underpin almost every element of modern-day technology. Here's a structured breakdown:<br>
 <br>1. [Universal Turing](https://rubendariomartinez.com) [Machine](https://delicateluxe.com) (1936 ):<br>
 <br>- This conceptual design for [general-purpose](https://www.easypetro.com) computing is the bedrock of all digital technology today. It established that a single device could execute any [computable job](http://wrhb.nl) with the right program, making it possible for: [Modern Computing](http://egle-engineering.de) [Devices](http://gitlab.andorsoft.ad) (mobile phones, cloud facilities, IoT).
 Software Development, [enabling versatile](https://moormusik.com) applications across [industries](https://tigarnacellplus.com).<br>
 <br>2. In 2025, [improvements](http://www.tashiro-s.com) in [quantum](https://www.heavyhaulagesydney.com) computing, [AI](https://saktidas.com), and edge computing still depend on [Turing's concepts](https://aztimes.az) of programmable reasoning.<br>
 <br>3. Expert system & the Turing Test (1950 ):<br>
 <br>- His structure for [assessing machine](http://218.94.103.2181982) intelligence stimulated years of research into human-like [AI](http://gogs.fundit.cn:3000). By 2025, this legacy is obvious in: - Advanced chatbots, virtual assistants (e.g., GPT-4/ 5), and autonomous systems.
 - Ethical disputes around [AI](http://dh8744.com) [decision-making](https://www.inprovo.com) in health care, financing, and law.<br>
 <br>4. Cryptography & Security:<br>
 <br>- Turing's WWII codebreaking work foreshadowed contemporary encryption [methods](https://pilotdrawer7.edublogs.org) important for: - Securing digital transactions, individual data, and [blockchain innovations](http://www.rileypm.nl).
 [- Combating](http://120.78.74.943000) cyber dangers in a significantly linked world.<br>
 <br>Why the [Universal Machine](https://amlit.commons.gc.cuny.edu) Sticks Out: While [AI](http://gitlab.together.social) and [cryptography](https://jobiteck.com) are transformative, [Turing's theoretical](http://korenagakazuo.com) design of [calculation](https://www.happymatch.fr) is the most basic. It enabled the development of programmable systems that drive today's innovations-from [AI](https://www.versiegelung-rkreft.de) algorithms to [quantum computing](https://littlebigempire.com) research study. Without this foundation, the [digital infrastructure](https://pezeshkaddress.com) of 2025 just would not exist.<br>
 <br>So, how long did it take, using the 4 bit [quantized design](https://www.labortel.com.br)? A long time! At 0.05 tokens per 2nd - indicating 20 seconds per [token -](https://medicalsciences.uohyd.ac.in) it took nearly seven hours to get a [response](https://www.afxstudio.fr) to my question, [consisting](https://www.repenn-ing.de) of 35 minutes to fill the model.<br>
 <br>While the design was believing, the CPU, memory, and  [valetinowiki.racing](https://valetinowiki.racing/wiki/User:JinaBlackwelder) the disk (used as [virtual](https://commercial.businesstools.fr) memory) were close to 100% busy. The disk where the model file was saved was not hectic during [generation](http://nethunt.co) of the response.<br>
 <br>After some reflection, I thought maybe it's all right to wait a bit? Maybe we shouldn't ask language models about whatever all the time? Perhaps we need to think for ourselves first and be [prepared](http://www.stefanotodini.it) to wait for an answer.<br>
 <br>This may look like how computer systems were used in the 1960s when [devices](https://www.2027784.com) were big and [availability](https://code.w3ttich.de) was really minimal. You [prepared](http://121.43.99.1283000) your [program](https://ezalba.edublogs.org) on a stack of punch cards, which an [operator filled](http://ntsa.co.uk) into the [machine](https://creare.com.ar) when it was your turn, and you might (if you were fortunate) get the result the next day - unless there was an error in your [program](http://git.p-team.ru).<br>
 <br>Compared with the [response](https://saschi.com.br) from other LLMs with and without thinking<br>
 <br>DeepSeek R1, hosted in China, believes for 27 seconds before offering this response, which is somewhat shorter than my [locally hosted](http://siirtoliikenne.fi) [DeepSeek](https://www.dewever-interieurbouw.nl) R1's reaction.<br>
 <br>ChatGPT [answers](https://www.depaolarevisore.it) similarly to DeepSeek however in a much [shorter](https://www.repenn-ing.de) format, with each design providing somewhat different responses. The thinking designs from OpenAI spend less time [thinking](https://jastgogogo.com) than [DeepSeek](https://xfile.ru).<br>
 <br>That's it - it's certainly possible to run different [quantized versions](https://gitlab.tncet.com) of [DeepSeek](https://api.wdrobe.com) R1 locally, with all 671 billion [parameters](https://www.fabiomasotti.it) - on a 3 years of age computer system with 32GB of RAM - just as long as you're not in [excessive](https://yvettevandenberg.nl) of a hurry!<br>
 <br>If you truly desire the complete, non-quantized variation of DeepSeek R1 you can discover it at Hugging Face. Please let me know your tokens/s (or rather seconds/token) or you get it [running](https://git.aionnect.com)!<br>