Stories

Show HN: You can now train your own LLMs with less memory, on your consumer GPUs

Article news.ycombinator.com

Show HN: You can now train your own LLMs with less memory, on your consumer GPUs

*TL;DR: You can train your own 27B model on a 3090 with 24GB VRAM, though you would have to go through hundreds of billions of tokens for it to be good. I am working on conversion of existing models. Check repos for code and the paper if you are interested in more. I have a 40.4GB Qwen Flash* N*ext GGUF*

Hi, I do AI research. I have been working on an optimizer that doesn't need the \~16 bytes of VRAM per weight to train a model that AdamW does. I focused on training ternary networks bcs BitNet (Ma et al.) shows they match full precision quality in large scales. The problem is BitNet also keeps the full precision weights during training which means you still need to use the same amount of memory. I got this down from 16 bytes to 1 byte or even less per weight.

For you, this means, you can train your own 10-20B parameter models on your consumer GPUs. I fitted 27B model training on a single 24GB 3090 and a 23B model on a 16GB GPU. The trained model costs \~1.6 bits per weight (\~6GB at 27B), 10x smaller than 16bit, and 5x smaller than Q8.

I have also been working on converting existing models and training them. This is harder and I am still working on it but I have some promising results. I quantized and trained a Qwen3.8 27B though the file I published below has adapters which are about 2GBs, pushing the file size to 8GB. I am also making a no-adapter version right now. It scores GSM8K 0.895, the same as full precision, HumanEval 0.84 vs 0.93 and runs on a 12 GB GPU at \~2*3* tok/s on my crippled 12GB laptop 4080, though this number can heavily be improved with custom kernels, engineering, etc. I will leave it to you guys to help me with that too.

I also quantized and trained Qwen Flash Next down to 40.4 GBs including its n-gram table (at 8GB). It largely retains the skills of the network (scores equivalent on GSM8K benchmark, a little behind on HumanEval, I unfortunately don't have compute for the larger benchmarks such as TerminalBench2.1 but perhaps some of you can help me out) but forgets facts. It mostly works fine if you give it web search as far as my experience goes but this is not strictly measured.

I am still doing work on this and I will keep working on it, perhaps even for many years. The code is open for non-commercial use. Check out:

The core repo for training: https://github.com/oguzeldereli/ttf-ternary

The repo for conversion: https://github.com/oguzeldereli/ttf-conversion

Repo for running models trained by ttf: https://github.com/oguzeldereli/ttf-home

GGUF for a converted Qwen 3.8 27B with adapters (around 8 GB): https://huggingface.co/oguzeldereli/ttf-Qwen3.8-27B-GGUF

GGUF for a converted Qwen 3.8 Flash Next (around 40.4GB): https://huggingface.co/oguzeldereli/ttf-Qwen3.8-Flash-Next-G...

GGUF for a converted Qwen 3.8 27B without adapters (around 6 GB): Currently in conversion and will drop soon.

I will convert a Qwen 4 27B instance when it drops.

Full paper: https://github.com/oguzeldereli/ttf-ternary/blob/main/paper/...

Also, I would appreciate github stars if you like the project.

Discussion 0 comments · 2 points · gladia-solaria · 17m
Open on HN
Loading the discussion…

Domain filters

Stories from these domains are hidden from every list. Subdomains match too: blocking substack.com also hides danluu.substack.com. Type a domain, or paste a link to block its site, and press Enter.

New collection

Delete this collection?

About YAVCHN

YAVCHN is a reader for Hacker News and Lobsters, with articles and discussions in separate windows or Classic pages.

Created by Paul Parks and built with PUDL.

YAVCHN source code on GitHub

Privacy policy · Terms of use

Help

Keyboard

j / k
Move down and up the story list. The arrow keys scroll whatever has focus.
Enter
Read the marked story in the article reader.
[ / ]
Read the story above or below in the list, in the same article reader. Back returns to the story you came from.
p
Pin or unpin the marked story, which keeps it in Pinned.
n / N
Move to the next or previous top-level comment in the window in front.
c
Collapse or expand that comment.
f
Hide or show the story list.
Esc
Close a menu or this help.
Access key m
Go to the menu bar. Most browsers take it with Alt on Windows and Linux, and Safari with Control and Option.
?
Show this help.

Windows

Each story opens in a window holding its article above its discussion; drag the bar between them to share the room differently. A window can be moved by its title bar, resized from any edge, snapped to a half or a corner by dragging it there, maximised, or minimised to the bar at the foot of the page. Use Window > New reader window to open an empty reader, or Story > Open in new reader window to open another reader for the current article. Docked readers keep their articles when you select another story from the sidebar. Minimized readers can be restored and reused for their site. A window's Previous and Next buttons read up and down the list in the same window.

A link in a comment or an article to another Hacker News or Lobsters thread opens that thread in a window too. A link to a single HN comment opens the comment above its replies.

While a story's window is in front, the Story and Discussion menus in the menu bar hold its commands: pinning, Previous and Next story, sorting, collapsing every thread, jumping to the first new comment. Each window also remembers where you were in its article and discussion, so a reload, or Back to a story that Next took you past, finds your place again. Closing a window forgets it.

The whole arrangement lives in the address, so a bookmark or a shared link brings it back, and Back undoes the last change. Moving between Hacker News, Lobsters, their lists, Pinned and Find changes only the list, and leaves the windows open.

The list

The pin at the start of a row keeps the story in Pinned, and the cross at its end hides it. Pinned can be narrowed by words in the title, site or author, by source, and to the stories you haven't opened yet, and ordered by when you pinned them, by points or by comments; the filters are part of the address, so a filtered view can be bookmarked. Scroll past the end of the list to load more. Domain filters, in the View menu, hide every story from a site.

Collections are named lists of stories. Story > Add to collection files the story in front into one or more of them, and the Collections feed shows them all or one at a time; the menu that chooses collections also creates, renames, and deletes them. A note is your own text on a story. Choose Add note in a story's toolbar to write one; it saves as you type. Rows with a note carry the note mark, and the Notes feed lists every noted story and searches the text of your notes.

Browsing view

View > Windowed and View > Classic select the browsing view and save your default in this browser. Window view reuses a reader for each feed. Classic view opens stories and applets as pages. Open as a page is a one-off action that does not change your saved default. Use the Windowed selector to return an article to a window. Direct page links always open as pages.

Applets

The Applets menu in the menu bar holds three tools, each a window of its own. Replies to me takes your Hacker News user name and lists the replies to your last thirty comments and stories, checking again every three minutes while it is open, and marking what is new since you last marked them read. Look up a user opens a profile on Hacker News or Lobsters, with their submissions and recent comments, as a commenter's name in any discussion does; the bar at the top of a profile looks up someone else in the same window, and Back returns to the one before. Who is hiring? filters the posts of HN's monthly hiring threads by the words you type.

They read only what the sites publish to everyone, so none of them asks for a login, and your user name stays in this browser anonymously and is included in retained applet state when signed in.

Find

Find takes any link and lists every time it was submitted to Hacker News and Lobsters, so you can read each discussion of it.

About

YAVCHN never sees your Hacker News or Lobsters login. The discussion is fetched from each site's public API; to vote or reply, follow the link above the discussion, or the arrow beside a comment, to the source's own site. Without a YAVCHN account, your data stays in this browser. When signed in, pins, collections, notes, blocked domains, and retained reading state are stored with your account and synchronized across devices. Hidden stories and layout stay in this browser. The privacy policy has the details.

Open source: github.com/paulmooreparks/yavchn. Built with PUDL.