liuliu

16 ms·

241.
by liuliu 2y ago
I think the only shaking part, is the Secure Enclave, which provides the root of the guarantees. From there, everything is attested so if one layer is adversarial, other layers can notice.
242.
by liuliu 2y ago
If you read closer to the benchmark, it seems to be slightly worse than FLUX [dev] on prompt adherence and quality. However, the best is to evaluate the result oneself, and the track-record of PixArt Sigma (from the same author?) is pretty
243.
by liuliu 2y ago
I usually don't want to comment on these, but: DiffusionBee's repo https://github.com/divamgupta/diffusionbee-stable-diffusion-... don't have any updates for 9 months except regular binary releases. Ther
244.
by liuliu 2y ago
On the same Mac hardware, Draw Things should be the fastest on models such as SDXL / FLUX.1 against other tools based on PyTorch (I stopped benchmarking SD v1.5 results for a while so that might regress a little bit here or there).
245.
by liuliu 2y ago
What do you mean? FLUX.1 prompts women or women faces just fine? Do you mean the skin texture is unrealistic or some other artifacts?
246.
by liuliu 2y ago
> they're attempting to make a "smart compressor" that minimizes the message-length of the compressed message [i.e. the state] by increasing the information embedded in the compressor itself [i.e. the model.]) And this see
247.
by liuliu 2y ago
PythonKit (from Swift) is underrated but immensely useful in certain scenarios (i.e.: https://engineering.drawthings.ai/from-iphone-ipad-to-mac-en... ). Hopefully JavaKit can find their usefulness too!
248.
by liuliu 2y ago
GaLore can do that too (if you transplant it). There are similar methods prior LLM era doing the same. They are not quite there on the loss graph side though (I am actually unsure about GaLore, but both FLoRA and ReLoRA were not quite there
249.
by liuliu 2y ago
As much as I liked the team, there is really no information other than the loss graph :(
250.
by liuliu 2y ago
A model trained on BF16 that within the range of FP16 have effective bit rate of 13-bit at max (e5m7). Reasonable quantization (at 8-bit) gets you weight error (i.e. L2 distance on weights) down to < 1e-3, and full diffusion run (30 step
251.
by liuliu 2y ago
Quality issue should be mainly due to using FP16 accmulators for GEMM in M1-M2, A14-A16 devices (it is not a problem for SD v1 / SDXL models due to smaller channel count). This is changed to FP32 accmulators for GEMM in these devices w
252.
by liuliu 2y ago
You should expect Flux model support in the next a few days coming to the app.
253.
by liuliu 2y ago
Snapchat do certificate pinning for it's main API domain. I am not exactly sure why analytics domain are different and why not have certificate pinning. (I thought analytics go through the same API domain, but it must be wrong then).
254.
by liuliu 2y ago
Not doing fuzzing on user-input supported feature, especially for AV, is damning.
255.
From iPhone to Mac – Enabling Rapid Local Deployment of SD3 Medium with S4nnc(engineering.drawthings.ai)
1 points by liuliu 2y ago|0 comments
256.
by liuliu 2y ago
That's where it is confusing to me. They mentioned that for LoRA fine-tuning, the router weights are frozen, so you don't update the routing when training different concept. But how that expert router is trained? Could be a pretra
257.
by liuliu 2y ago
I think there is an expert router layer to decide which loras to be integrated at inference time. But they also mention that they freeze the weights for router during training. So it is unclear to me how the router was trained on what los
258.
by liuliu 2y ago
Yeah, but people do like to write y = .log(x * 0.1) + (1.24).squareRoot() type of expressions for what Python did best (NumPy and PyTorch).
259.
by liuliu 2y ago
In retrospective, yes. But Swift for TensorFlow and similar projects received much attention at a time. (Also see Mojo as another attempt).
260.
by liuliu 2y ago
The math type inference example makes the usual claim that "what if Swift can replace Python" a non-starter. As someone who have to deal with this on frequent basis, it is pretty sad. (I maintains s4nnc and a fork of PythonKit).
261.
by liuliu 2y ago
How come the stream not cutting off during re-entry "blackout" period? Is it because the re-entry is low / slow enough so no thick plasma layer, or because it is streamed through StarLink which happens to be on the other side
262.
by liuliu 2y ago
This is light on details. What's the typical way to propagate version from shared memory to each renderer process so they know it is up-to-date / stale? Wouldn't a naive implementation has thundering herd problem?
263.
by liuliu 2y ago
Optane 905p goes for $500 a piece (1T) I believe.
264.
by liuliu 2y ago
Unitree robots are "research" platforms. People buy them to do development on top of them. I think G1 basically follows that route like their Go1 / 2 quadrupled. What they are really good at is to drive down cost for their mo
265.
by liuliu 2y ago
I might be naive comes to classic ML / NLP: how do you keep the prefix table for N = 5? Naively, that look-up table would be 100k^5 (assuming 100k vocabulary size). Is it very sparse? How large that usually is?
266.
by liuliu 2y ago
The problem is you cannot implement a reference type without dropping to UnsafePointer or use class. But using class is so discourage nowadays people probably would be better served by adding an Arc / Ref type wrapper like Rust did.
267.
by liuliu 3y ago
6 seems reasonable. 128 Lanes from ThreadRipper needs to have a few for network and NVMe (4x NVMe would be x16 lanes, and 10G network would be another x4 lanes).
268.
by liuliu 3y ago
Yeah, at some point, would just say if you want stackful coroutine, go to Go and stackless one, use C++. BTW, here is my macros for stackless coroutine in C: https://github.com/liuliu/co/blob/master/examp
269.
by liuliu 3y ago
What @cchance said. Also, Diffusion models have much more capacity than ESRGAN. Unable to hallucinate whole subjects is a limitation of model capacity and training data, not the reverse (i.e. a high capacity model can mimic what you want in
270.
by liuliu 3y ago
ESRGAN is the same thing. You cannot recover signal that are already lost without additional information (in Diffusion case, model's prior). Upscaler are hallucinators by that definition.

More ›