liuliu

15 ms·

271.
by liuliu 3y ago
SUPIR and other diffusion based models.
272.
by liuliu 3y ago
That's appalling to say at least. But Snapchat implemented certificate-pinning since 2015. Does that mean either the analytics endpoint was not covered or somehow the certificate-pinning is circumvented in this case?
273.
by liuliu 3y ago
Palworld
274.
by liuliu 3y ago
You shouldn't judge people by what Forbes says, but what people do. So far, Emad delivers on what he promises (with some delays here and there), and Stability.ai is the only one that publish image / video generation models at prod
275.
by liuliu 3y ago
Yeah. It is weird. On one hand, v3 looks really good, the arch is sound, and the work is solid. On the other hand, selling Clipdrop is not a good sign (a growing startup divest is uncommon). Emad likes to talk about how they were able to re
276.
by liuliu 3y ago
`multiprocessing` works fine for serving HTTP requests or do some other subset of embarrassingly-parallel problems.
277.
by liuliu 3y ago
You cannot kill an immortal.
278.
by liuliu 3y ago
But the tuple still cannot conform to Equatable because it is not an "existential type"? It is pretty handy when you want to pass data around generic type T which is Equatable. Now if you want to add a little bit more data (throug
279.
by liuliu 3y ago
This arch seems to be flexible enough to extends to video easily. Hopefully what we have here will be another "foundation" blocks like the transformer blocks in LLaMA. Why: It looks generic enough to incorporated text encoding &#x
280.
by liuliu 3y ago
Yeah, safetensors separates metadata and tensor data. The metadata is an offset reference to the tensor data that you are free to define yourselves. In that way, you can create files in safetensors format but the tensor data itself is paged
281.
by liuliu 3y ago
safetensors can mmap too because the tensor data are just offsets and you are free to align to whatever you want. It is hard to keep metadata minimal, and before long, you will start to have many different "atom"s and end-up with
282.
by liuliu 3y ago
I honestly think have a way to just use json (a.k.a. safetensors) / msgpack or some lightweight metadata serializer is a better route than coming up with a new file format. That's also why I just use SQLite to serialize the metada
283.
by liuliu 3y ago
Not really. We've been on that road before. Embedding computation graph into the file makes changes to the computation graph harder (you need to make sure it is backward compatible). This is OK in general (as we have onnx already), but
284.
by liuliu 3y ago
There is a more detailed paper from Apple on this. Basically, you can do a little bit better than only keeping current weights in RAM with mmap. For LLM, you are mostly dealing with b = W @ a where a and b are vectors, only W is the matrix.
285.
by liuliu 3y ago
pth can include Python code (PyTorch code) for inference. TF includes the complete static graph. GGUF is just weights, safetensors the same thing. GGUF doesn't need a JSON decoder for the format while safetensors needs that. I personal
286.
by liuliu 3y ago
I am going to look at quantization for 8b. But also, these are transformers, so variety of merging / Frankenstein-tune is possible. For example, you can use 8b model to populate the KV cache (which computes once, so can load from slowe
287.
by liuliu 3y ago
Wow, this is really just scale-up DiT. We are going to see tons of similar models very soon.
288.
by liuliu 3y ago
Note that this model works well at higher resolution too (in contrast with MiDaS family) given the depth head is convolution and the DINO v2 base can interpolate positional encoding to generate higher-resolution features: https:/&#x2F
289.
by liuliu 3y ago
The large C model have fair bit of parameters tied to text-conditioning, not to the main denoising process. Similar to how we split the network for SDXL Base, I am pretty confident we can split non-trivial amount of parameters to text-condi
290.
by liuliu 3y ago
DALL-E shares the same autoencoders as SD v1.x. It is probably similar to how Meta's Emu-class models work though. They tweaked the architecture quite a bit, trained on their own dataset, reused some components (or in Emu case, trained
291.
by liuliu 3y ago
Should use no more than 6GiB for FP16 models at each stage. The current implementation is not RAM optimized.
292.
by liuliu 3y ago
prior generations usually take fewer steps than vanilla SDXL to reach the same quality. But yeah, the inference speed improvement is mediocre (until I take a look at exactly what computation performed to have more informed opinion on whethe
293.
by liuliu 3y ago
Agreed. From the look at it, a few things: WAL is not enabled, insertion is not wrapped in transactions. I did a benchmark a while ago and if everything done right it is much faster: https://dflat.io/benchmark/ Also, e
294.
by liuliu 3y ago
The comoditization of models outside of LLM is delusional. There is no comparable open source image / video models to the private ones.
295.
by liuliu 3y ago
> one can just quickly swap it out for something else someone else paid to train. That doesn't seem to be the case. There are very limited open-source models outside of the small-LLM bubble.
296.
by liuliu 3y ago
I think TinyML has pretty close tie to Pete Warden / Useful Sensors, who led TF Lite back in Google.
297.
by liuliu 3y ago
Open an BEGIN transaction forces read from a particular snapshot in SQLite. There are some complications around lock upgrade in that case: BEGIN a, read, BEGIN b, read, write from a transaction fail because cannot upgrade read lock to write
298.
by liuliu 3y ago
But Wal also keeps "history" so one reader transaction can see x0 from wal1, another reader can see x1 from Wal1. Wal1 cannot be merged into main db otherwise x0 is lost?
299.
by liuliu 3y ago
I am going to put model related code we use in a public repo soon (it is very similar to https://github.com/liuliu/swift-diffusion but in NHWC format). ANE will be around 25s if it runs. DT's default only uses GPU
300.
by liuliu 3y ago
iPhone 15 Pro with 30 steps at 512x512 resolution (SD v1.5) should take around 35 seconds on iPhone 15 Pro with Draw Things. 1.5 mins are too slow. (I am the author of DT).

More ›