This. I will resume my recommendation of Bitwarden.

I didn’t say it can’t. But I’m not sure how well it is optimized for it. From my initial testing it queues queries and submits them one after another to the model, I have not seen it batch compute the queries, but maybe it’s a setup thing on my side. vLLM on the other hand is designed specifically for the multi co current user use case and has multiple optimizations for it.

I run the Mistral-Nemo(12B) and Mistral-Small (22B) on my GPU and they are pretty code. As others have said, the GPU memory is one of the most limiting factors. 8B models are decent, 15-25B models are good and 70B+ models are excellent (solely based on my own experience). Go for q4_K models, as they will run many times faster than higher quantization with little performance degradation. They typically come in S (Small), M (Medium) and (Large) and take the largest which fits in your GPU memory. If you go below q4, you may see more severe and noticeable performance degradation.

If you need to serve only one user at the time, ollama +Webui works great. If you need multiple users at the same time, check out vLLM.

Edit: I’m simplifying it very much, but hopefully should it is simple and actionable as a starting point. I’ve also seen great stuff from Gemma2-27B

Edit2: added links

Edit3: a decent GPU regarding bang for buck IMO is the RTX 3060 with 12GB. It may be available on the used market for a decent price and offers a good amount of VRAM and GPU performance for the cost. I would like to propose AMD GPUs as they offer much more GPU mem for their price but they are not all as supported with ROCm and I’m not sure about the compatibility for these tools, so perhaps others can chime in.

Edit4: you can also use openwebui with vscode with the continue.dev extension such that you can have a copilot type LLM in your editor.

I wouldn’t assume this is done with malice in mind, but maybe this is someone unaware of the importance of a formal license.

Indeed. They only have in their github page:

Terms of Use

Feel free to use these components in personal and commercial projects. However, while the tutorials and demos are available for your use as-is, they cannot be redistributed or resold. Let’s keep things fair and respect each other’s work.

I’m wondering, the integrated RAM like Intel did for Lunar Lake, could the same performance be achieved with the latest CAMM modules? The only real way to go integrated to get the most out of it is doing it with HBM, anything else seems like a bad trade-off.

So either you go HBM with real bandwidth and latency gains or CAMM with decent performance and upgradeable RAM sticks. But the on-chip ram like Intel did is neither providing the HBM performance nor the CAMM modularity.

They used PimEyes, nothing new.

Of importance: they do not want to release the tool but use it as a way to raise awareness.

You mean between the French article and the English comment? :)

From what I understand, bsky’s architecture seems to allow federation at multiple levels. On one side the individual profiles are actually websites and the app aggregates the content almost as an RSS reader. I do see some profiles which are independent like Jeff Gierling’s, so yes federation at the profile level seems to work.

And this is really important because it is one way to prevent your data from being hostage by the service. Then there is another level of federation. I’m not entirely sure of the terminology here, but there is one aggregator aspect, which is quite compute intensive. And that one I don’t know if there is another instance of it. But functionally speaking, I’m quite impressed by the technical aspect of bsky. There has been a lot of thought put into it.

And monetizing it is not the issue, the problem is mostly how. That they have some paid features is fine, it’s even important that there are ways to monetize it without milking their users of their privacy.

Let’s hope this works out and becomes sustainable while respecting the users!

I’m very grateful for your extended help. I’ve made some progress. I’m able to get the prompt to appear asking me for my passphrase to unlock the right partition (sda3 in my case). Entering the passphrase, however, drops me in the Dracut emergency shell after ~3min of dracut logs, seemingly looping. (Edit: the reason for why it drops me in the shell is very unclear. It says Dropping to debug shell. /bin/sh: can't access tty: job control turned off. And if I try to exit the dracut shell, it says dracut Warning: could not boot.).

In the Dracut emergency shell, checking /dev/mapper/ I see a luks-<sda3-uuid> listed. Running blkid I see it listed too with TYPE=crypto_LUKS. I also see a dev/dm-0 with a dedicated UUID, in ext4. I ran blkid which shows:

/dev/mapper/luks-705fc477-573a-4ef6-81b6-a14c43cda1f5: UUID="57955343-922a-4918-9bc1-797ca8d13a9c" TYPE="ext4"
/dev/sda1: UUID="cc5e0b03-3544-4bef-ab8b-8b72dd236926" TYPE="ext4"
/dev/sda2: UUID="4df1af6c-3199-4bb2-bb12-bcf897cfc6fc" TYPE="swap"
/dev/sda3: UUID="705fc477-573a-4ef6-81b6-a14c43cda1f5" TYPE="crypto_LUKS"
/dev/dm-0: UUID="57955343-922a-4918-9bc1-797ca8d13a9c" TYPE="ext4"

I checked the status of the filesystem running cryptsetup status /dev/mapper/luks-<sda3-uuid> and it says it is active, which I guess means it is unlocked?

I checked the /root directory, and it is empty. So I tried to mount the partition myself: mount /dev/mapper/luks-<sda3-uuid> /root but it fails saying mount: mounting /dev/mapper/luks-<sda3-uuid> on /root failed: No such file or directory and that got me really puzzled? I’ve been searching far and wide but I can’t seem to find anyone with a similar situation. I feel like I’m close to getting this working.

Below is my syslinux kernel config, and the 2nd and 3rd items are what I booted into (/boot/extlinux.conf)

# Generated by update-extlinux 6.04_pre1-r15
DEFAULT menu.c32
PROMPT 0
MENU TITLE Alpine/Linux Boot Menu
MENU HIDDEN
MENU AUTOBOOT Alpine will be booted automatically in # seconds.
TIMEOUT 10
LABEL lts
  MENU DEFAULT
  MENU LABEL Linux lts
  LINUX vmlinuz-lts
  INITRD initramfs-lts
  APPEND root=/dev/mapper/root modules=sd-mod,usb-storage,ext4 cryptroot=UUID=705fc477-573a-4ef6-81b6-a14c43cda1f5 cryptdm=root rootfstype=ext4 rd.debug log_buf_len=1M rd.shell

LABEL lts
  MENU DEFAULT
  MENU LABEL Dracut Linux lts
  LINUX vmlinuz-lts
  INITRD /boot/initramfs-6.6.56-0-lts.img
  APPEND root=/dev/mapper/luks-705fc477-573a-4ef6-81b6-a14c43cda1f5 modules=sd-mod,usb-storage,ext4 rootfstype=ext4 rd.shell rd.debug log_buf_len=1M rd.luks.uuid=705fc477-573a-4ef6-81b6-a14c43cda1f5

LABEL lts
  MENU DEFAULT
  MENU LABEL Dracut Linux lts 2
  LINUX vmlinuz-lts
  INITRD /boot/initramfs-6.6.56-0-lts.img
  APPEND modules=sd-mod,usb-storage,ext4,dm,crypt,rootfs-block rootfstype=ext4 rootflags=rw,relatime rd.shell rd.debug log_buf_len=1M root=UUID=57955343-922a-4918-9bc1-797ca8d13a9c rd.luks.uuid=705fc477-573a-4ef6-81b6-a14c43cda1f5

And here the /proc/cmdline of the booted partition:

BOOT_IMAGE=vmlinuz-lts modules=sd-mod,usb-storage,ext4,dm,crypt,rootfs-block rootfstype=ext4 rootflags=rw,relatime rd.shell rd.debug log_buf_len=1M root=UUID=57955343-922a-4918-9bc1-797ca8d13a9c rd.luks.uuid=705fc477-573a-4ef6-81b6-a14c43cda1f5 initrd=/boot/initramfs-6.6.56-0-lts.img

Here is my setup, when I boot in my regular initramfs (the one I’m trying to replicate using dracut):

mytestalpine:~# lsblk -o NAME,FSTYPE,FSVER,LABEL,UUID,FSAVAIL,FSUSE%,MOUNTPOINTS
NAME     FSTYPE      FSVER LABEL UUID                                 FSAVAIL FSUSE% MOUNTPOINTS
sda                                                                                  
├─sda1   ext4                    cc5e0b03-3544-4bef-ab8b-8b72dd236926  195.5M    21% /boot
├─sda2   swap                    4df1af6c-3199-4bb2-bb12-bcf897cfc6fc                [SWAP]
└─sda3   crypto_LUKS             705fc477-573a-4ef6-81b6-a14c43cda1f5                
  └─root ext4                    57955343-922a-4918-9bc1-797ca8d13a9c    2.3G     8% /

mytestalpine:~# lsblk -l -n /dev/sda3
sda3   8:3   0  2.8G 0 part  
root 253:0   0  2.8G 0 crypt /

Note: No idea of the relevance, but I’m testing this setup in a VM, with a BIOS firmware.

That does not sound like a viable long term solution to me.

Wizz

Thank you for your help. I spent time digging into this rabbit hole, and while I’ve learned a lot, I am struggling to get the basics to work. Right now, I’m focusing on being able to just boot an image I created using dracut, excluding all the initial stuff I wanted, just be able to reproduce the original functionality of being able to unlock my luks partition using my keyboard.

Where I’m at: I am building my initramfs using the following command: dracut -f -v --add crypt --add lvm --add dm. I get the following output log:

Then I updated the /boot/extlinux.conf file, adding the following second entry (displaying the first one just for comparison):

LABEL lts
  MENU DEFAULT
  MENU LABEL Linux lts
  LINUX vmlinuz-lts
  INITRD initramfs-lts
  APPEND root=/dev/mapper/root modules=sd-mod,usb-storage,ext4 cryptroot=<my-uuid> cryptdm=root quiet rootfstype=ext4

LABEL lts
  MENU LABEL dracut-img
  LINUX vmlinuz-lts
  INITRD /boot/initramfs-6.6.56-0-lts.img
  APPEND root=/dev/mapper/root modules=sd-mod,usb-storage,ext4 cryptroot=UUID=<my-uuid> cryptdm=root quiet rootfstype=ext4 rootflags=rw,relatime

I added the rootflags=rw,relatime because this was shown in the dracut log, so I thought perhaps that mattered. But for the most part I left it the same as the previous entry, because I’m trying to do the same thing I suppose. Perhaps I’m mistaken?

The current result of booting that image leads to a long loading (not asking for the passphrase to unlock the partition) then displaying the following error:

dracut Warning: Could not boot.

dracut Warning: "/dev/mapper/root" does not exist

Generating "/run/initramfs/rdsosreport.txt"
You might want to save "/run/initramfs/rdsosreport.txt" to a USB stick or /boot after mounting them and attach it to a bug report.

To get more debug information in the report, reboot with "rd.debug" added to the kernel command line.

Dropping to debug shell.

Before dropping me in a shell, in which I have not found anything useful to do. I am surely missing something basic as my understanding of what’s happening is pretty superfluous.

What I’m noticing which may be of importance:

• dracut[E]: ldconfig exited ungracefully, in the dracut output log. Perhaps this matters and should be fixed? An image is nonetheless generated.
• there are many missing modules when creating an image, but I don’t know if any of them matter, at least for my purpose.
• One thing I can’t wrap my head around is, how come the original kernal image work, when I had packages such as device-mapper and lvm missing, why did dracut complain about them missing for me to compile my own image? and would I need to add options in the /boot/extlinux.conf file, when they are not required for the original boot entry, when all I’m trying to do (as a start) is just make sure I can reproduce a bootable kernel image?

Indeed, quite surprising. You got to “stroke their fur the right way” so to speak haha

Also, I’m increasingly more impressed with the rapid progress reaching open-weights models: initially I was playing with Llama3.1-8B which is already quite useful for simple querries. Then lately I’ve been trying out Mistral-Nemo (12B) and Mistrall-Small (22B) and they are quite much more capable. I have a 12GB GPU and so far those are the most powerful models I can run decently. I’m using them to help me in writing tasks for ansible, learning the inner workings of the Linux kernel and some bootloader stuff. I find them quite helpful!

Someone recently referred me to this blog post about using RAG in open-webui. I have not tested if but the author seems to reach a good setup.

https://medium.com/@kelvincampelo/how-ive-optimized-document-interactions-with-open-webui-and-rag-a-comprehensive-guide-65d1221729eb

Perhaps this is of use to you?

I have no idea if ollama can handle multi-GPU. The 70B in it’s q2_k quantized form requires already 26GB of memory, so you would need at least that to run it well and that would only imply it could be entirely run on GPU, which is the best case scenario, but not at what speed.

I know some people with apple silicon who have enough memory to run the 70B model and for them it runs fast enough to be usable. You may be able to find more info about it online.

I wish I could. I have an RTX 3060 12GB, I run mostly llama3.1 8B versions in fp8, at 30-35 tokens/s.

Sure! It can be a bit of a steep learning curve at times but there are heaps of resources online, and LLMs can also be useful, even if it just in pointing you in the direction for further reading. Regardless, you can reach out to me or other great folks from the !localllama@sh.itjust.works or similar AI, ML or related communities!

Enjoy :)

For RAG, there are some tools available in open-webui, which are documented here: https://docs.openwebui.com/tutorials/features/rag They have plans for how to expand and improve it, which they describe here: https://docs.openwebui.com/roadmap#information-retrieval-rag-

For fine-tuning, I think this is (at least for now) out of scope. They focus on inferencing. I think the direction is to eventually help you create/manage your own data which you get from using LLMs using Open-WebUI, but the task of actually fine-tuning is not possible (yet) using either ollama or open-webui.

I have not used the RAG function yet, but besides following the instructions on how to set it up, your experience with RAG may also be somewhat limited depending on which embedding model you use. You may have to go and look for a good model (which is probably both small and efficient to re-scan your documents yet powerful to generate meaningful embeddings). Also, in case you didn’t know, the embeddings you generate are specific to an embedding model, so if you change that model you’ll have to rescan your whole documents library.

Edit: RAG seems a bit limited by the supported file types. You can get it here: https://github.com/open-webui/open-webui/blob/2fa94956f4e500bf5c42263124c758d8613ee05e/backend/apps/rag/main.py#L328 It seems not to support word documents, or PDFs, so mostly incompatible with documents which have advanced formatting and are WYSIWYG.

The interface called open-webui can run in a container, but ollama runs as a service on your system, from my understanding.

The models are local and only answer queries by default. It all happens on the system without any additional tools. Now, if you want to give them internet access, you can, it is an option you have to setup and open-webui makes that possible though I have not tried it myself. I just see it.

I have never heard of any llm “answer base queries offline before contacting their provider for support”. It’s almost impossible for the LLM to do it by itself without you setting things up for it that way.