liuliu

26 ms·

391.
by liuliu 3y ago
Yeah, I am not explaining myself very well. I meant: people try to sharing code between main app and extensions with dylib. If they don't do that, they have code-duplication in the app bundle so the binary size is main code + 2 * share
392.
by liuliu 3y ago
For example: https://tailscale.com/blog/go-linker/ talks about how they slim down the Go runtime to workaround the 15MiB memory limit for network extension.
393.
by liuliu 3y ago
Yeah, and the only reason they don't static link is because these dylibs are some 3rd-party closed-source libraries to avoid a lot of version clashes comparing to ship static lib naively. Otherwise you put some code in dylib so your ap
394.
by liuliu 3y ago
They are talking about app-bundled dylibs, which cannot be updated separately due to the whole-bundle signing.
395.
by liuliu 3y ago
Weird. Yes, link time is a big issue at development time, but with the new development of ld-prime and sold, seems trading that with dylib is not a big of a deal. On the other hand, the other big use of dylib I saw for apps to embed (other
396.
by liuliu 3y ago
I like dlib better if you want a swiss army knife of algos. ccv is moved on to be all-in on neural-network based computer vision and all development effort happened in that area. (Similar can be said for dlib but at least the author still i
397.
by liuliu 3y ago
None of what you mentioned related to "traffic to a blog post from HN front page". I am just saying "traffic from HN front page" is a poor example of things that need "scale up". Not disagreeing all the challen
398.
by liuliu 3y ago
Everything agree with, until: > but if you have a blog post that hits the front page of HN. HN traffic is so small that unless you are using Raspberry Pi Pico, it is fine. Your 5-year old Android phone probably can handle HN traffic just
399.
by liuliu 3y ago
Isn't it known for quite a few years now that these CNN networks (ResNet, VGG etc) trains well with FP16? The problem is that attention layer with softmax can have dynamic range higher than FP16 can handle hence you have to go to BF16?
400.
by liuliu 3y ago
> The question really is why is swift <-> objc bridge so bad you need this software? Do you want to keep using a UTF-16 backed String type while the world (including Swift) is on UTF-8? (Oh, and I am posing the question to Java too
401.
by liuliu 3y ago
I am curious why cannot we just, at instruct-tuning phase, add additional token type embedding, such as: embedding = text_embedding + token_type_embedding + position_embedding The token_type_embedding is zero init and frozen for responses a
402.
by liuliu 3y ago
Swift doesn't do a good job of interleaving with Python (i.e. allowing Python to call into Swift, while possible, is nothing like Numba due to various object type translation required). Mojo seems to be targeting Mojo -> Python -&gt
403.
by liuliu 4y ago
There is no user error but UX issues :) That has been said, you probably used paintbrush rather than the eraser? There would be more help on the Discord server though! https://discord.gg/5gcBeGU58f
404.
by liuliu 4y ago
There might be a few things Draw Things missing from this article: no mask blur, not selecting the inpainting model for inpainting work. Tomorrow's release should contain both mask blur and inpainting ControlNet, which might help these
405.
by liuliu 4y ago
Chip's capability planning seems need about ~2yr lead time. So we are expecting fastest would be somewhere around end of 2024. (Transformers probably earlier than that (2023?), 4-bit would be later).
406.
by liuliu 4y ago
Metal only recent versions (macOS 13 / iOS 16) supports mmap and use that in GPU directly. CUDA does have unified memory mode even it is dedicated GPU, would be interesting to try that out. Probably going to slow down quite a bit, but
407.
by liuliu 4y ago
Thank you for the compliment! Definitely would love to contribute in some ways, once we have the benchmark, I have some ideas :)
408.
by liuliu 4y ago
PyTorch saves tensor data and metadata separately (metadata in Pickle format a.k.a. pickle VM, which is famously not mmap friendly because it builds up Python object through the VM execution). These then zipped together in a zip file. Wheth
409.
by liuliu 4y ago
That's actually the point of safetensors format, although it needs to take care of alignment. I actually use this technique to mmap directly to GPU buffer on Apple systems (although it seems to segfault on iOS 15.x systems, only suppor
410.
by liuliu 4y ago
The comparison table from the ts_server site looks awesome though. I wish we could generate one for llama.cpp, unfortunately too busy with other things at the moment.
411.
by liuliu 4y ago
(Talking about several-hour long take-homes, not FizzBuzz problems). It is cruel, but it is numbers game. The take-home is easier for hiring managers given the macro environment and would help to cull out false positives (hence "number
412.
by liuliu 4y ago
Here is some quick math: these devices has SSD read speed somewhere around 2GiB/s. With 4-bit quantization, we are looking at loading 4B parameters per second. That means we need 8s per token for 30B model. Hmmm, the math is a bit off
413.
by liuliu 4y ago
Agreed. The power of SD is actually in its diverse styles. You should be able to aim at any specific style in latent space and get it. This is in contrast with Midjourney, which heavily favors sci-fi modern style.
414.
by liuliu 4y ago
You probably don't need canned positive prompt. ControlNet has a "guess mode" and in reality, basically just send the control to positive part with empty prompt while not sending control to negative part with empty prompt. Ed
415.
by liuliu 4y ago
I think that you just need a special entitlement to enable w+x page. I could be wrong though, haven't dug too deep in this.
416.
by liuliu 4y ago
Add "calculator" in a form of Python interpreter to give alternative result. Providing interactive graphical interface for some of the charts they shown. Connecting outside sites for booking, pictures, user comments. To some degre
417.
by liuliu 4y ago
This is just openjourney model fine-tuned with Dreambooth. You can use any of these tools: Draw Things, Mochi Diffusion, DiffusionBee, AUTOMATIC1111 UI on M1 / M2 with this model. (I wrote Draw Things).
418.
by liuliu 4y ago
https://civitai.com/
419.
by liuliu 4y ago
But You can keep CDC choices in memory and if there is update to one chunk, just compute new boundaries from that chunk until next boundaries matches? A bit more code to write but doable?
420.
by liuliu 4y ago
It is possible. You even have textual inversion trained to do exactly this: https://civitai.com/models/3036/charturner-character-turnaro... Consistent character creation is a big topic the community interested in

More ›