Zero to Hero LLMs with M3 Max BEAST

[@AZisk]

JOIN: youtube.com/@azisk/join

[@AdamsTaiwan]

Just tried the LM Studio and my desktop. Was able to connect my 8 year old notebook's vsCode using Code GPT to it. Pretty nice, but still looking for a solution that can scan all my vs solution and tell me where to fix my problems.

[@bawbee27]

Incredibly helpful - this is the video everyone with an Apple Silicon machine trying to do LLM’s should see!

[@keithdow8327]

Thanks!

[@AZisk]

🤩 thanks!

[@vincentnestler1805]

Thanks!

[@AZisk]

🤩 thanks!

[@MaxTechOfficial]

Keep up the good hustle, Alex! -Vadim

[@AZisk]

Thanks Vadim!

[@univera1111]

@@AZisk if I may ask, can you replicate this on a Linux or windows and see which is easier for users. Or u can just say here

[@zt9233]

@@univera1111also benchmarks

[@abhishekjha9041]

​@@AZisksir please make a video for MacBook pro specifications for Machine learnings . I'm so confused about what to buy 16inch with 30 core 96gb ram Or 16inch with 40 core 64 GB ram. Or I have to buy a m3 pro 18 core 36gb ram. I'm so confused and like me other people also so please make a separate video on that it's a request

[@abhishekjha9041]

​@@AZiskAnd I have a question that I do some research and find out that MacBook pro in Delaware have zero sales tax which means if I buy MacBook pro in 2500 dollars so I don't have to give any tax on it. It's is true sir.

[@giovannimazzocco499]

Excellent stuff. I searched KZbin for weeks to find benchmarks of DNN models on M3. This is the first and only one I've found so far. There's is a ton of videos on video editing, graphics, gaming and music production on M3s. But for what concerns fresh material about machine learning on Apple Silicon I'm pretty convinced you're the only game in town. Keep it up. Looking forward to seeing more benchmarks.

[@tonbii]

i bought M1 Max with 64GB 3 years ago to do this kind of works. I am so happy to find this video.

[@catarinamoreira4805]

This is fantastic! Thank you so much! More content on LLMs, please!

[@joshgarzaBI]

Awesome video here. I'm bummed I didn't do it sooner. I have never seen my M1 (16GB) freeze before. Great teaching here!

[@innocent7048]

Very interesting article. I will try this :-)

[@AZisk]

🤩 thanks so much!

[@JasonHorsnell]

Just got myself an M3 Max and found your videos. You’ve saved me SO MUCH TIME….. Very much appreciated…..

[@danieljohnmorris]

How much ram?

[@JasonHorsnell]

⁠36GB max base. More than enough for my purposes atm.

[@TimHulse]

Same here!

[@dennisBZC]

Hey Alex, I’ve been watching many of your videos, mostly for comedy - as I find you hilarious the way you explain things to a non-tech mortal, but occasionally, try to copy your instructions and try my luck to test out a few things for fun. I’m not one for cutting code, but I still watched the whole thing, just to get to the LM Studio to download a model to try out on my M3 Max. I tried the Phi3, thinking Microsoft might be better than the others. I don’t have a clue what I’m doing, but it seems to work a little. You are a LEGEND! Keep up the great work. Love to see how you train your AI in due course. I keep shouting at it to “sit”…my MacBook hasn’t moved, so I guess, it is quite obedient.

[@gargarism]

I think the very first thing I will try out on my already ordered M3 Max, will be to follow what you did. The whole reason I bought the M3 Max is to work with machine learning. So thanks a lot!

[@AZisk]

Good choice!

[@zt9233]

@@AZiskis m3 max as good as nvidia for this?

[@pec8377]

@@zt9233 no it's not, unless you want to run large model that won't dit into nvdias cards, they Will Always beat M3 GPU. Maybe not when ANE IS activated, but none of thé tools présentes hère supports core ml

[@MikeBtraveling]

If you are looking for a laptop to work with LLMs on you cant really beat the Mac for models larger than 7bP and you want them to run locally@@zt9233

[@theperfguy]

I have to commend you for your effort. I havent seen any other reviewer showing any other usecase than media comsumption, synthetic benchmarks and video encoding and editing. You are perhaps the only youtuber I know who tries out other things like code compile time and ML workloads, which is what is going to run on majority of the high end machines.

[@AZisk]

Glad it was helpful!

[@build.aiagents]

Thank you for the GPU setting in LM Studio at 15:00!! Can you do more videos on proper GPU setup on LLM's for M1-3?

[@paraconscious790]

This is amazing, awesome, super crisp yet easy to understand with absolute engagement 🙌👌🙏

[@SebastianWerner82]

Great to see you creating videos with this type of content as well.

[@atldeadhead]

I enjoy all your videos but this one was particularly interesting. I look forward to future videos that explore machine learning leveraging the power of the M3 Max. Fantastic stuff, Alex. Thank you!

[@LukeBarousse]

Interesting, I didn't know about LM Studio; that makes things A LOT cleaner

[@mr.w7803]

Dang!! Dude, this video sold me on that M3 Max configuration… this is EXACTLY what I want to do on my machine

[@Mrloganphillips1]

I had so much fun with this project. I just got a m3max and wanted a project to work on. After I got llama running I made a bash script to run the command and trigger a second bash script to open a browser window to the ip address after a 5s delay to let the server get up and running first. then I made a shortcuts button to run it. now I have on demand llm with an easy to use on/off button.

[@facepalmmute3619]

the bass in your voice on the MBP speakers is phenomenal

[@joshbarron7406]

I would love to see a token/second benchmark between M2 Max and M3 Max. Trying to decide if should upgrade

[@abhinav9058]

Hey did you upgrade?

[@ismatsamadov]

I subscribed a few months ago, but I have never seen such quality content. Thanks, Alex! Keep going.

[@AZisk]

thx 🙏

[@devdeal4146]

Just got the m3 max with 48gb ram. Excited to see how it works with your tutorial. Thanks!

[@jorgeluengo9774]

Thank You Alex, this is an amazing video. I will look into the software development tools installation.

[@AZisk]

Awesome! Thanks

[@terencedodge3249]

Still relevant, some of us aren't programmers, we are just people who screw ( fiddle ) around with software and LM Studio is an excellent tool for that.

[@anthonyzheng7274]

You are awesome! This is great, I bought an M3 Max several days ago and really having a great time playing around with LLM's.

[@scosee2u]

I really love your videos and how you explain these cutting edge concepts! Would you consider researching or interviewing someone to make a video about quantizing options and how it impacts using llms for coding? Thanks again for all you do!

[@AZisk]

Possibly!

[@someone5781]

So excited for your next video on training on the m3 max!

[@hamiltonwmr189]

If you are going to do any intensive task on MacBook then keep it charged at 80% using al dante. Dont let run the models on battery as churning though cycles will damage it's help ,keep it on power adapter and 80% charging. I did some intensive training on my m1 Pro and it went from 100 to 96% battery health in 1 year.

[@CitAllHearItAll]

4% loss in 1 year is normal. I'm at 2+ years on M1 Pro with 86% battery health. You're either trippin or trollin.

[@MikeBtraveling]

I bought a maxed out M3 max to do this, please run the larger models with ollama, when using LM studio you need to make sure you are using the correct prompt template for the model, i think that was your issue.

[@mrsai4740]

@@MikeBtraveling I'm curious, were you able to run larger model like a 70b llama maiden?

[@juangarcia-wp2zr]

very cool content, thanks, I feel very curious now to try out some of this llms

[@JohnSmith762A11B]

Excellent. Many thanks for putting this together! 🥂

[@amermoosa]

amazing. just shrinking the whole second grade of engineering college in 17 minutes. incredible 😊

[@RadAlzyoud]

Brilliant. Thanks for sharing.

[@jameshancock]

Nice! Thanks! FYI when you change the preset you’re changing how it inputs into the LLm. Which caused it to go nuts.

[@BenWann]

I couldn’t agree more - I wanted to really sink my teeth in ML since it’s been a while - and I bought a MBP m3 max after seeing your comparisons. Sorry I couldn’t use an affiliate code - micro center had a killer deal on it :(. I look for your videos to drop now, and look forward to what you come up with next.

[@nikolamar]

Alex this is AWESOME!!! Thank you!

[@astrohgamingZero]

Looks good. I use text-generation-webui and the chat/chat-instruct modes or input presets can make or break some models.

[@XNaos]

Finally, I waited for this

[@dukecitywildcat]

Great video, looking forward to the LLM fine-tuning video.

[@bdarla]

Super helpful! I hope you will continue with further relevant videos!

[@DivineZeal]

Great video! Thinking about getting the MBP M3 for llm

[@Andrew-v2g]

Alex, thanks.

[@AZisk]

You bet!

[@suburbanflyer]

Thanks for this Alex! Just got an M3 Max so it'll be great to try out some new things on it, this definitely looks interesting!

[@terra8net]

THX .... Great LLM Content for Mac User :)

[@_mansoor]

Awesome, Thank you. Halo Alex!!!🎉🎉

[@stanchan]

The performance of the M3 is amazing. Waiting for the refreshed Studio, as the M3 Ultra will be a beast. Hoping it will have the 256GB RAM as predicted.

[@geog8964]

Thanks, Alex.

[@justingarcia500]

Hey could you do a low battery mode test on the m3 max as you did with your m1 max a while back

[@eldee8704]

Awesome tutorial! I bought the 14" MacBook Pro M3 Max base model for this to try out.. lol

[@timelesscoding]

Interesting stuff, I wish I could understand a little more. Thanks

[@kman41000]

Awesome video man!

[@AZisk]

Glad you enjoyed it

[@FrankHouston-v5e]

Best LLM build video KZbin ❤. I’m buying my 36GB MacBook Pro M3 Max 14 Core cpu with 30 core GPU. Planning on launching a KZbin AI/Ml channel soon 🧐.

[@DavidCampero26]

Hi Alex! I would love to see a comparison between M3 Max 14/30 and M3 Max 16/40 with the same processes for LLMs. I read that many people is going with the base model M3 Max and I would like to see how much difference there is. If you know of someone who did it, please let me know!! I want to buy a laptop as soon as possible!! Thanks!!

[@pbdivyesh]

You're a good lad, thank you!🎉😅

[@camsand6109]

Glad i subscribed. you've been on a roll lately (new subscriber).

[@MuhammaddiyorMurodov-l5n]

Thank you so much for making this video, it was really helpful. Please do more this kind of coding videos and testing on m3 macbook, and push them to the limits, I think you are the best channel for this because you have the knowledge and intention to do these things and it will be win win situation for both of us

[@sujithkumar8261]

Are you using macbook m3 base variant?

[@justisabelll]

Great video, really looking forward to the next few ML related ones. You might have had better results with LM studio though if you disabled mlock after enabling Metal GPU. Also the model output looks nicer if you enable markdown in the settings as well.

[@yinoussaadagolodjo4549]

How to disable mlock ? Can find it !

[@jeffersonmp4]

How do you know how to do all these steps?

[@Xilefx7]

Can you test the LLM perfomance in low power mode? I believe Apple needs to optimize how they handle the thermals of the MacBook Pro with the m3 max.

[@fallinginthed33p]

Now the Snapdragon X Elite laptops are also pretty good at local LLMs if you run a specific quantized format.

[@abhinav23045]

That fan noise is like feel the power of AGI.

[@AZisk]

😆

[@MikeBtraveling]

very interested in the topic and would love to see you do more in this space.

[@christopherr8441]

If only we could directly access and use the Apple Neural Engine for doing things like this. Imagine the speed and performance gains.

[@juliana.2120]

ohh i love that you use conda here because it really helps me keep my hard drive clean with all those different AIs :D im an absolute beginner so i'm afraid of installing stuff i cant find later on. some people say its "outdated" and runs in errors too often but i cant really judge that. is that true?

[@vlogdavidalejandro]

super interesting and useful, I take this opportunity to ask about Matlab again and its real performance, since Apple's on its page is not real

[@JonNordland]

I would love some videos where the M3 was pushed a bit harder. For instance, 70b models. The 70b models are much more useful for real work.

[@AZisk]

Noted!

[@brandall101]

I have the 48GB variant so can't do 70B...but 34b models run fairly slow as is, seeing a reported 11-12 tok/sec in LM Studio, so I'd expect a 70B to be about 5-6 tok/sec. It's also pushing 110W during inference. For me personally, that's just not performant enough for real use so opted to not go through the hassle of swapping for a 64GB BTO.

[@AZisk]

@@brandall101For what it would cost to get a high end 128+gb mac AND all the ssd space you’d need for the ml models, i would just get a 4080 or 4090. only problem is - memory requirements

[@brandall101]

@@AZisk The common thing is to buy a pair of 3090s to get a nice middle-ground between performance / memory / cost... those can be had used for about $1500. I just don't think the hardware is quite there... yet. A couple more generations and I think we'll be golden.

[@geoffseyon3264]

I hope Apple is reading this thread…

[@AmpiroMax]

Please, can you compare tokens generation speed of any llama-like model on m3, m3 pro, and m3 max?

[@chillymanny714]

This is a great video, I think if you were to make videos to teach intro/intermediate data analyst how to build LLMs or a series of videos to try different application creation using Macs M chips, that it would be a big hit. I will try to replicate your approach

[@syedanas2083]

I look forward to that

[@djicubsoaicn]

Is it the 96GB version of the M2 Max, what do you think, I want to deploy my own 13B model locally (train the model with some relatively sensitive data), or even become my 'digital clone', do you think the 38c 96GB M2 Max is a suitable choice?

[@tomdonaldson8140]

Love it! Looking forward to the training video(s). Now I want a Mac Studio M3 Ultra! Oh, no such thing yet? Come on Apple! We’re waiting!!!

[@TimHulse]

That's great, thanks!

[@mercadolibreventas]

Kep it up! Good Job! Can you do a video on getting Llama Factory set up on the M3, Thanks!

[@stephenthumb2912]

thanks for testing. it's interesting that even with enough memory, still some slowness on the bigger model quants. my base M2 8gb can run the q4 7b's barely.... prefer ollama using cli which will run at usable tps. it's sort of ok with LM Studio, but generally I need to run 3b's or below with q4 quants. Orca-mini 3b is sort of the default test standard for me on 8gb mac's incl. the mac metal air. can confirm, using the mac metal checkbox, causes runaways. textgen funnily runs fine with mac metal suport as well.

[@theoldknowledge6778]

This LM Studio is Lit 🔥

[@MeinDeutschkurs]

Thanks for the video, but there's something I don't understand: we explored this together, yet concluding with "find another model" isn't what I expected. Achieving satisfaction through a successful demonstration, despite the difficult journey, is essential. Now, I feel like I've wasted my time. I could have just downloaded LM Studio and figured it out through trial and error myself.

[@AZisk]

this video is not about finding the best model, but getting set up with an environment that will allow you to use any model you want.

[@MeinDeutschkurs]

@@AZisk, and that’s why this demonstration was not satisfying, but yes, it was a demonstration: a not satisfying demonstration. :)

[@SergeyZarin]

Thanks great video explaining !

[@AZisk]

Glad it was helpful!

[@jigyansunanda]

looking forward to your training models video

[@ChitrakGupta]

That was really good. I learnt something and was fun to run on the new M3 Max

[@Renzsu]

Hey Alex, can you please do a video on Stable Diffusion on Mac? I'm on the fence on getting one.. the shared memory is tempting, it allows for more than using a discreet GPU. But I wonder how the speed is..

[@yashen12345]

YES AWESOME! More on this pls. do you have an m3 max with 128gb available? Typically these smaller and quantized models you have showcased will preform worse than the bigger ones. I wanna see LLama 2 70B 8bit running on an m3 max with 128gb . This is the largest most powerful model thats able to fit in a macbook. Lets push this thing to the absolute limit and see how it preforms llama 2 70b is actually able to match chatgpt3.5 performance. so if we're able to run this we can have OUR OWN chatgpt that is actually as good as chatgpt running LOCALLY ON A MACBOOK THATS INSANE, pls can i have a video on this

[@toddturner6]

Actually it will run Mistral 180B in the top configuration with 128GB RAM. Edit: typo. Falcon 180B.

[@yashen12345]

@@toddturner6 ? mistral is 7b

[@toddturner6]

@@yashen12345 Falcon 180B (typo).

[@akashtriz]

Why is it that no one questions the metal GPU hardware for bungling up the model? Llama seemed less wacky on CPU.

[@speed_mc3990]

I train my own NNs but sadly popular ML libraries like tensor flow and PyTorch don't support multiple GPUs, hence I can only use one GPU Core.

[@ergun_kocak]

3 to 5 times faster than M1 Max 64GM full spec. Thank you very much for the video 👍

[@marabgol]

Thanks Alex! great videos watched 2 so far, do you have videos or plan to make videos how to fine-tune Llama2 models on Metal ?

[@AZisk]

Not yet! but considering digging more into this area on the channel

[@uninoma]

cool thank you !!!!🤟

[@AliHussain-jh3iq]

Insightful video Planning to get a MacBook Pro M3 Max for LLM work. Should I go for 1TB or 2TB, 14 or 16-core CPU, and 64GB or 128GB RAM? Thanks for your insight!

[@robertleal7455]

Could the m3 pro handle these tasks? I would assume yes but much slower.

[@Stonk5331]

you assume correct

[@geoffseyon3264]

Yes. Even the low end 7B and some quantized 13B models will run well on a 16GB M2 MacBook Pro. But if you’re gonna get an M3 Pro, get the unbinned 12 core CPU/18 core GPU version. It will be worth the extra couple hundred bucks if you’re into developing your LLM skills.

[@boraoku]

In my experience trying different open LLMs for code generation, my recommendation is that don’t waste your time unless you can’t access OpenAI for some reason…

[@kingmargie1182]

Great job!

[@01_abhijeet49]

These models run soooo well on my 3060 rtx desktop. Alas, my investment is worth it

[@bern_stock8946]

Can you please do a tf installation video for m3?

[@melody-cheung]

I still recommend ollama. Easier installation and friendly local WebUI.

[@liquathrushbane2003]

Following along using LM Studio on Windows, using the same model (7B), and it appears to be broken. Just typing in "hello" at the prompt and the response shows the AI writing out a question (presumably from the forum where the data was scraped) followed by the answer to that question. Completely ignores me. Tried the 13b and it seems better.

[@SimoneFolador]

Thanks about the video man! I loved it and it helped me a lot since I wanted to try some models on my machine. What's your experience on fans on the M3 Max machine? I've read that they are pretty noisy and it becomes pretty hot as well. I still have an Intel machine (last generation) with 64GB ram and 2TB drive but i wanted to buy a new M3 max

[@marcoseliasmep]

I went crazy once I got this: What is the square root of 2? The square root of 2 is 1, because 1 x 1 = 2 I don't remember which model I have used haha

[@PMX]

7:15 no, the one for text generation is S_TG t/s, in your case about 20 tokens per second for f16. And at 7:46, again, that's the wrong column, the correct column (S_TG t/s) shows the correct value, 33.61 which is exactly what I get on an M2 Max for a 7B Q8_0 (the GPU improvements on the M3 Max compared to the M2 Max don't have much impact for LLM inferencing, they are mostly useful for 3D rendering apps like Blender and for games). The column you are using (S t/s) is (tokens prompt processing + tokens text generation) / total time, which is a meaningless number (you can get it to be as fast as prompt processing speed just by having a large prompt and a very small text generation, or as slow as text generation speed by having a short prompt and a long number of tokens for text generation).

[@gileneusz]

I tried to run Falcon 180B with 4.0 quantisation within LM Studio and it failed to run.... 😥