This was a pretty heavy lift for us to get out which was why it took a while. In addition to writing new image processing routines, a vision encoder, and doing cross attention, we also ended up re-architecting the way the models get run by the scheduler. We'll have a technical blog post soon about all the stuff that ended up changing.
How long until Vulkan Compute support is merged into ollama? There is an active pull request at https://github.com/ollama/ollama/pull/5059 but it seems to be stalled with no reviews.
It's difficult because we actually ditched a lot of the c++ code with this change and rewrote it in golang. Specifically server.cpp has been excised (which was deprecated by llama.cpp anyway), and the image processing routines are all written in go as well. We also bypassed clip.cpp and wrote our own routines for the image encoder/cross attention (using GGML).
The hope is to be able to get more multimodal models out soon. I'd like to see if we can get Pixtral and Qwen2.5-vl in relatively soon.
I tested the small model with a few images from Clevr. On first blush I am afraid it didn't do very well at all, it got object counts totally wrong and struggled to identify shapes and colours.
Still, it seems to understand what's in the images in general (cones and spheres and cubes), and the fact that it runs on my mac book at all is basically amazing.
My initial testing was with charts - I've been waiting on local vision models to be good enough to feed technical documents and my initial testing is looking very good. Example:
I've tried with some ppt images rather than Clevr ones and it does much better. It can count circles and triangles and differentiates between them quite well. It can recognise the colours of the objects as well.
I think that the faux 3d of clevr images is too much for the model, it's interesting because much smaller pre-transformer specialist models were very good at clevr.
Did they fix multiline editing yet? Any interactive input that wraps across 3+ lines seems to become off-by-one when editing (but fine if you only append?), and this will be only more common with long filenames being added. And triple-quote breaks editing entirely.
How does this address the security concern of filenames being detected and read when not wanted?
Readit News
also, can it do grounding like cogvlm?
either way, great job!
The hope is to be able to get more multimodal models out soon. I'd like to see if we can get Pixtral and Qwen2.5-vl in relatively soon.
Ran the 11B yesterday and it worked great.
Still, it seems to understand what's in the images in general (cones and spheres and cubes), and the fact that it runs on my mac book at all is basically amazing.
https://i.imgur.com/1ETREP9.png
I think that the faux 3d of clevr images is too much for the model, it's interesting because much smaller pre-transformer specialist models were very good at clevr.
How does this address the security concern of filenames being detected and read when not wanted?
Deleted Comment