All You Need To Know About Running LLMs Locally
Вставка
- Опубліковано 17 чер 2024
- RTX4080 SUPER giveaway!
Sign-up for NVIDIA's GTC2024: nvda.ws/48s4tmc
Giveaway participation link: forms.gle/2w5fQoMjjNfXSRqf7
Please read all the rules & steps carefully!!
1. Sign-up for NVIDIA's Virtual GTC2024 session between Mar 18 - 21st
2. Participate the giveaway DURING Mar 18 - 21st
3. ???
4. Profit
TensorRT LLM
[Code] github.com/NVIDIA/TensorRT-LLM
[Getting Started Blog] nvda.ws/3O7f8up
[Dev Blog] nvda.ws/490uadi
Chat with RTX
[Download] nvda.ws/3OHPRHE
[Blog] nvda.ws/3whKZTb
Links:
[Oobabooga] github.com/oobabooga/text-gen...
[SillyTavern] github.com/SillyTavern/SillyT...
[LM Studio] lmstudio.ai/
[Axolotl] github.com/OpenAccess-AI-Coll...
[Llama Factory] github.com/hiyouga/LLaMA-Factory
[HuggingFace] huggingface.co/models
[AWQ] github.com/mit-han-lab/llm-awq
[ExLlamav2] github.com/turboderp/exllamav2
[GGUF] github.com/ggerganov/ggml/blo...
[GPTQ] github.com/IST-DASLab/gptq
[LlamaCpp] github.com/ggerganov/llama.cpp
[vllm] github.com/vllm-project/vllm
[TensorRT LLM] github.com/NVIDIA/TensorRT-LLM
[Chat with RTX] www.nvidia.com/en-us/ai-on-rt...
[LlamaIndex] github.com/run-llama/llama_index
[Continue.dev] continue.dev/
Model recommendations:
[Nous-Hermes-llama-2-7b] huggingface.co/NousResearch/N...
[Openchat-3.5-0106] huggingface.co/openchat/openc...
[SOLAR-10.7B-Instruct-v1.0] huggingface.co/upstage/SOLAR-...
[Google Gemma] huggingface.co/google/gemma-7b
[Mixtral-8x7B-Instruct-v0.1] huggingface.co/mistralai/Mixt...
[Deepseek-coder-33b-instruct] huggingface.co/deepseek-ai/de...
[Madlad-400] huggingface.co/jbochi/madlad4...
[Colbertv2.0] huggingface.co/colbert-ir/col...
This video is supported by the kind Patrons & UA-cam Members:
🙏Andrew Lescelius, alex j, Chris LeDoux, Alex Maurice, Miguilim, Deagan, FiFaŁ, Daddy Wen, Tony Jimenez, Panther Modern, Jake Disco, Demilson Quintao, Shuhong Chen, Hongbo Men, happi nyuu nyaa, Carol Lo, Mose Sakashita, Miguel, Bandera, Gennaro Schiano, gunwoo, Ravid Freedman, Mert Seftali, Mrityunjay, Richárd Nagyfi, Timo Steiner, Henrik G Sundt, projectAnthony, Brigham Hall, Kyle Hudson, Kalila, Jef Come, Jvari Williams, Tien Tien, BIll Mangrum, owned, Janne Kytölä, SO, Richárd Nagyfi, Hector, Drexon
[Discord] / discord
[Twitter] / bycloudai
[Patreon] / bycloud
[Music] massobeats - magic carousel
[Profile & Banner Art] / pygm7
[Video Editor] maikadihaika - Наука та технологія
![How Did Open Source Catch Up To OpenAI? [Mixtral-8x7B]](/img/n.gif)
stay up-to-date on the latest AI research with my newsletter! → mail.bycloud.ai/
Minor correction: GGUF is not the predecessor to GGML, GGUF is the successor to GGML. (thanks to danielmadstv)
please make step by step guide how to install locally and private for example
Mistral-7B. im trying to do this with multple guides and all time im stuck at something
The amount of infos you give both in the videos and the descriptions is insane dude! Keep up the good work!
Thanks for the video! Minor correction: GGUF is not the predecessor to GGML, GGUF is the successor to GGML.
I hoooonestly don't know how to feel about the thumbnails looking too similar to you-know-who that got me accidentally clicking this video but meh... One's gotta do what one's gotta do I guess.
Same
I don't know who, who?
@@Deathington. Fireship
Bycloud removed the frame and the grid background on his thumbnails, I think those work great as his signature style. I hope he keeps them
Let's just hope he doesn't get _burned~_
Stup osing Fireship thumbnails😭
Y
A thousand thanks! Finding a good LLM model was a complete nightmare for me + it is difficult to figure out which formats is outdated and which - new hot stuff.
that was awesome, thanks for the concise information bycloud! 🔥
You can also use ollama. It even runs on a raspberry pi 5 (although slow)
Immensely helpful video. I hope the future has tonnes of user controlled locally ran llms for us in store!
I was pretty sure this was a fireship video, but the video is great and informative. Exacly what I was looking for.
Poor Faraday nearly always gets overlooked when people talk about local LLMs, but it is without a doubt the most easy to use "install and run" solution. Unlike nearly all other options it's near-impossible to mess something up and default settings out of the box are not sub-par.
How much is Faraday?
@@hablalabibliaLike all the best things in life - it's free.
@@hablalabiblia It's free and very easy to use! It's really meant just for chatting, it's basically a Silly Tavern kind of app, just not with that many options but it has its own back end with a focus on GGML models. If you're looking to just run models through character cards I'd say, give it a go!
Faraday has outdated models and whenever you download models, you have to fumble with model cards and directory structures, plus it's not as fast as other options. LM Studio is better than Faraday.
Thanks, this is great. Please make a comprehensive video on Fine-tuning locally 101..Cheers
Your videos are way more fun than my algebra homework
I love your adhd-friendly edits cloudy.
Nice video! Can you do a video about fine tuning a model?
Absolutely fantastic and informative video. Well done! I will say I feel like the information certainly speaks to the grip that OpenAI has, especially from a development standpoint, despite the whole video being about open-source models.
The procedures, time, research, and money required for any rando or small (even mid size) business owners to integrate open-source and local AI without any practical knowledge about it is near impossible. OpenAI wraps up RAG, "fine-tuning", and memory nice and neat into Assistants which can be easily called via the API. It would be amazing to have a completely standardized system that allows for the same type of application, but geared towards the variety of open-source models out there. Some platforms like NatDev let you compare multiple models based on the same input. Being able to see how RAG and fine tuning affects different models, both open-source and non, from the same platform would be unreal.
Thank you for the discussion.
Thank you. Very interessting. Is it possible in LM Studio to work with own files? Or create own LLM or extend LLM for own cases?
but anyways this video was very helpful because no one made it very clear on what are the best front end interfaces to install, I kept trying to make one myself to no avail and give up after a while after testing stuff in the command prompt
Very nice, tons of useful info
Thank you!
Now we just need a cheap inference card with 128GB memory to run 70B models locally...
Maybe we can hope for Qualcomm
I’d love to see AI inference accelerator cards with dual or quad channel DIMM slots.
@@cbuchner1 Qualcomm AI 100 Ultra is using LPDDR5
groq is using something of the sort, an LPU. although only usable through an api. no consumer cards yet that i know of, but it shows the trend towards it
@@nyxilos9167 you can buy a single groq card right now. it costs 21k and has 230MB on board. So to run 70B models at fp16 you need like 572 cards.... which is several racks. 14+ million to buy and 30kW to power. It will run the model at 400 tok/s easily.
You can buy a ready made 8x H100 box for maybe 350k and run that with like 8kW and it might be slower than the groq card.
none of that are consumer solutions.
The one I am hoping for is Qualcomm AI 100 Ultra. Which comes with 128GB LPDDR5 and 150W. They say it's for edge inference, but it would be perfect for workstation.
idk Qualcomm SoCs are for phones mostly... maybe iPhone 30 will have it XD
Boy, Chat with RTX is my personnel oracle for now on. Its RAG really indexes local documents without that whole hallucination from previous tools.
The one thing I hope to see soon is offloading different layers to different GPUs
I have a 4090 mobile in my laptop and an RX6800 in my eGPU.
I do have 96GB of system memory in addition to these two 16GB cards so I can do some fun stuff already.
In regards to context, would LLM Lora's help with that? Lets say im busy with story writer LLM and the fantasy world I'm working with would be as big as something like Middle Earth from LOTRs. Would a Lora help with that? Like if I train a Lora on all our past chat history about the story etc. Also more text regarding the lore of places and history of characters and family trees. So taking that into consideration, would that assist in keeping the context low? So I don't need to keep a detailed summerized chat history etc. What would the requirement be for training such a Lora and what would the minimum text dataset require for a coherent training?
LM STUDIO and TRINITY 1.2 is my favorite non-GPT entities!
thanks, this videos is very funny and helpful!
I can finally start my side project to take over the world, thanks!
Total newbie with running an LLM locally. What is the best llm for summarizing books and being able to ask questions about the books?
with local models are you able to make much longer responses given that you have enough ram and vram?
what about ollama as a backend, what is your take on that? Thank you so much for the video, sending love from switzerland
Ive been hamfisting my way through llms for over year. Just ramming squares into circles till it worked since informations so sporadic.
100% checking out your other videos. Learned more in 5 min then 4 hours reading github docs
Just to clarify then. For inference speed is more important GDDR6 will be GDDR5, but for fine tuning more more having 2x the amount of GDDR5 will be the GDDR6?
does the a Giveaway has country restriction?? I mean maybe you can't send it overseas due to shipping cost or something else.
That's a great question.
I'm a noob when it comes to this. I've come across Ollama, and started using it. Can I upload multiple things, texts, and possibly images, to chat with RTX and create my own data? And will it be uncensored? what are some other good options to 'Chat with RTX'
Hope this works better than the time I tried to download more RAM
Where do I upload the photo once GTC comes around ?
Step 4 is Clear, but How can I unlock step 3?
I only see questionmarks.
Do I have to do step 1 and 2 to unlock what I have to do at step 3,
Or do I just need to gain more XP for the unlock.
Maybe I just have to do step 4 twice to make up for the missing third step...
as a car content creator i approve this video
I don't have strong GPU , do you reccomend any sevices that i can run models on .
i run LM Studio and i think its great, good video my dude
How did you miss Faraday? Very easy to use and runs faster than LM Studio
What do you think of phi model ?
Where ollama?
agree, with the new windows installer its so easy for everyone to get local models
For a while it was only Mac-based, so it saw limited use with most AI folks who have Nvidia cards. If you're stuck on a Mac I hear it's really the better one for that.
wow now on support windows too ?@@sZenji
I use it on my Raspberry Pi5 to run LMM's, which is seriously cool, er hot when working.
Ollama + openwebui is the way to go. Same ui as ChatGPT, plenty of convenient functions. It's a no brainer.
Good to know.
Please make a video on how to fine tune a model using local documents.
EXL2 does support AMD GPUs. Turbo bought a couple just to make sure it runs with rocm
What 3 models do you recommend with 24 GB VRAM? Preferably 21-22GB / 24GB in practical usage.
huggingface lists models with their respective memory requirements. any 7b model will likely work very well and be under 21gb. you could also go with a bigger model but at a lower quantization. mistral models are among the most popular, open source, and very competitive.
Please make a video about making our locally running LLMs available for others to use maybe like our own API which people can use or a webUI interface to use our local LLM.
We love LM Studio 😫
Your thumbnail reminds me of fireship
Finally!!!!
I guess my machine is not good enough, 2019 intel imac, because running any model locally is usually lagging way behind ChatGPT 3, Gemini, Perplexity, etc.
what happened to the newsletter ????
Koboldcpp crying in the corner
I spent so much time trying to get something like this set up, but ended up back to gpt, most of these models are also censored just like gpt, and unlike gpt they are much slower AND on top of that they canot use plugins or special api's that let you access the internet or generate images etc. its sad but currently gpt has no peer
Curious headcount? 🙋How many of us watching these type videos are not developers?
OOOGABOOOOGAAAAH 💪😎🍺
I keep canceling my GPT4 subscription and then renewing it... 'Just when I thought I was out, they pull me back in.' GPT4 reminded me of that phrase from The Godfather. :)
Make a video about stable diffusion like this
Nice
Where is the diagram at 8:50 from?
2:17 Bro lives in the future where M4 is already released
what about ollama
The best RP model atm is Kunoichi v2
Fireship thumbnail is working for me
is it possible though to run LLM on iOS?
you're awesome
I am from Russia, can I participate in the contest?
How can I unistall text generation web ui? Anyone know this??
I just have a question, why this channel is so similar to fireship? are you the same person? : )
Basically to understand this video one should already know everything mentioned in this video by heart.
Eh, it provides terms to hunt for and sometimes that's all someone needs, a starting point. The video is short and covers a lot of ground.
Dude wants a 16 part lecture to explain it all😂
@@MonkeeGeenyuss I mean, I can only follow because I know it all and cannot imagine someone unfamiliar to understand anything from this firehose, lol.
How do local models compare to cloud ones like openai? Wouldnt a local pc have way worse results? A server farm can have way more vram and hence is better?
I'm getting ~gpt 3.5 performance on my laptop with 16gb ram and rtx 3060. I'm primarily using it because I feel like commerical ai chatbots are getting more and more censored
@@joseph-ianex Can you share which model are you using?, I have a laptop with those exact specs
@@MrBoxerbone *rtx 3050 ti. Most 7B models run fine, you can try Mistral, Gemma, or Llama 2. Get either ollama (command line) or llm studio (ui) to run the model. If you are new to running models I would recommend llm studio. The models are a bit slow and the context window is pretty small but they run. Pinokio is another cool ai if you want to test out open-source AI art tools 👍
Dunno why my comment isn't going through, but try Kobold! Better for GGUF. Current fav is "Crunchy Onion" Q4_K_M GGUF. Give it a taste! 10t/s on a 3090 and pretty smart.
I thought it was a video from Fireship 😂
oobaGooba, OOGABOOBA
My brain melted
lm studio/ollama are probably the simplest ways to get started, not sure why you picked these ones
Besides saving money, are there any other reasons to do it locally vs spending $20 a month for chatGPT?
privacy mainly
privacy and reliability, as with local LLM you don't depend on anyone's else infrastructure
Privacy, it's not filtered so you can do more things with it, won't see random dips in quality based on the whims of investors.
You did not name countries you are able to ship for the giveaway. Is it worldwide?
i’ll pay for whatever shipping it costs
unless the country is unshippable like north korea
@@bycloudAI Thank you for this information, and also for the amazing content that you are putting out ♥
So, AI is the new computer, everyone will have one? Seems good to me. I wonder how the job market will be, hardware will be on top for sure and Open AI will still being a giant. But the thing is how other industries will be affected.
How hard is it to run LLM with AMD GPU? Is it still Linux only hell bc no driver support?
rocm works for some stuff on windows, just don't expect to be on the bleeding edge with new features
running LMs on linux and windows, for some unknown (to me) reason, linux is over 5 times as fast as windows at prompt evaluation. it's not even close.
I just really really like how many serious people have to say ooobabooga.
It's like, almost as good of a joke on science as when that guy named the seventh planet.
Stanford open source LLama model is free. 🎉
I was highly disappointed, Shōji Kawamori and Kazutaka Miyatake are not on the panel about Transformers... ;)
We're gonna need a bigger GPU
You pay 20$ for convenience. Spending 1 day to set up the flow, Waiting 2 minutes every time for your model to load when you have a quick question, your GPU + CPU setting your room on fire cuz of how hot they're running... Unless you need some really specific usecase that cloud models censor, then it's just easier to pay those 20$ for instant access
Patience is a virtue. I got Mistral 7B running on an 2018 laptop, and it takes two minutes to respond, but it works well. Why have 8 GB of RAM when I don't use all 8 GB. The AI uses all my RAM. :) But, for people who have to use AI for a job, $20 is cheap, and workplaces cover the cost. For AI at home, a fast enough computer could work.
Are you the same as fireship?
Different human being
it’s fireship experimenting with 100% channel automation
Hope it works on my toaster too
easy oogabooga Apple M4 cpu support before release 😂
📝 Summary of Key Points:
📌 The video discusses the landscape of AI services in 2024, highlighting the abundance of hiring freezes and the prevalence of subscription-based AI services.
🧐 Various user interfaces for running AI chatbots and language models locally are explored, including UABA, Silly Tarvin, LM Studio, and Axel AO.
🚀 The importance of choosing the right model format, understanding context length, and utilizing CPU offloading for running local language models efficiently is emphasized.
💡 Additional Insights and Observations:
💬 "Garbage in, garbage out" is a crucial principle highlighted when fine-tuning AI models, emphasizing the significance of quality training data.
📊 Different model formats like GGF, AWQ, and EXL 2 are explained, showcasing how they optimize model size and performance.
📣 Concluding Remarks:
The video provides a comprehensive guide on running AI chatbots and language models locally, emphasizing the importance of model selection, context length, and fine-tuning techniques. Understanding these key aspects can help individuals navigate the AI landscape effectively and optimize performance while saving costs.
Generated using TalkBud
>this model list
🎯 Key Takeaways for quick navigation:
00:28 *🤖 Running AI chatbots and LM models locally provides flexibility and avoids subscription costs.*
00:43 *📊 Choosing the right user interface (UI) for local AI model usage is crucial, depending on individual needs.*
02:05 *🖥️ UABA is a versatile UI choice for running AI models locally, supported across various operating systems and hardware.*
02:33 *💡 Installing UABA enables access to free and open-source models on Hugging Face, simplifying the model selection process.*
05:18 *🤔 Context length is crucial for AI models' effectiveness, affecting their ability to process prompts accurately.*
06:12 *⚙️ CPU offloading allows running large models even with limited VRAM, leveraging CPU and system RAM resources.*
06:52 *🚀 Hardware acceleration frameworks like VM inference engine and TensorRTLM enhance model inference speed significantly.*
07:36 *🎓 Fine-tuning models with tools like Kora enables customization for specific tasks, enhancing AI capabilities.*
08:47 *💰 Running local LM models offers cost-saving benefits and customization options, making it an attractive option in the AI landscape.*
Made with HARPA AI
Which model is best for uh... y'know... stuff...
idk if you still need this, but one of the most "fun" models is MLewd
@@user-yj2tc8xu1fI don't know what you're talking about but thank you. This conversation didn't happen.
Much cheaper too!
this isn't fireship.. where am I?
same . the thumbnail got me and then i realised this guy took fireship's entire style
06:14 How do you have Mixtral 8x7b setup to use just that much VRAM and run that fast? On Oobabooga just 9-10 layers and I'm already risking running out of VRAM on my 16GB GPU, and the thing still takes enough time to finish writing moderate length replies that often my ADHD kicks in and I go do something else while it finishes writing....
Which of the GGUF quantization are you using? Is the issue I'm using the Dolphin version instead of the raw Mixtral? Should I ditch the Dolphin variants? It's been getting so hard to keep up with which models are the current most well regarded by the community with so many models coming out all the time...
For mixtral 8x7B you're going to want 2 3090s.
@@voidsofold16GB is not enough?
@@TiagoTiagoTNot enough for 8x7B
@@voidsofoldThe example he shows after mentioning using 10GB VRAM for quantized Mixtral was not actually that being done in practice but just some unrelated LLM output clip?
@@TiagoTiagoTOh, if you offload it sure you can run 8x7B, it'll just be very slow and have barely any token context
I'm glad I avoided the 20 dollar subscription by buying a $500 GPU and $100 ram
my brain hurts ( i only reached 4:08 I just watched the video to see if there anything I need to know about sillytavren since that what I searched but i don't thinks there any more )