singhrac

16 ms·

181.
by singhrac 3y ago
Will Oxide build a GPU rack as well? Asking for a friend.
182.
by singhrac 3y ago
Any chance you guys, Charlie Marsh/Astral, and Armin Ronacher will ever team up? I’m really excited about next-gen tooling but would love a cargo-like “one true path”. At the very least keep the tools compatible?
183.
by singhrac 3y ago
Yes, that's me (mid 20s). I learned using Scratch in middle school, made a lot of tiny games for me and my friends, and then learned Python in high school (with C++ fairly shortly afterwards). Scratch shields you from learning how data
184.
by singhrac 3y ago
I think MathJax has really caught up in speed in the past few (5) years with MathJax 3. I would pick MathJax today for compatibility reasons.
185.
by singhrac 3y ago
Any chance you've tried the tldraw plugin? https://github.com/holxsam/tldraw-in-obsidian Would love a review if you have one!
186.
by singhrac 3y ago
The linked blog is written by a transit expert who writes extensively on costs, including this report: https://pedestrianobservations.com/2022/10/24/the-transit-co... . I'm not an expert, though, so if yo
187.
by singhrac 3y ago
Absolutely, I was surprised anyone would recommend ctypes for Python/Rust today. I've written a few Python libraries using Rust to do the slow parts, and PyO3/maturin is a great experience.
188.
by singhrac 3y ago
There’s Swiftfin, no? Development is a bit sparse but it’s worked ok in the past when I’ve needed it.
189.
by singhrac 3y ago
Yeah, I found this too. For our specific workloads (data science) m7i-flex works well enough that I don’t mind, but it was not an obvious upgrade from m6a like I expected it to be.
190.
by singhrac 3y ago
If I recall correctly https://www.listennotes.com is a one-person operation. The owner has posted about it a few times: https://www.listennotes.com/blog/the-boring-technology-behin...
191.
by singhrac 3y ago
Sorry for drive-by (accidental) criticism, it was really more of a joke than anything. I haven't kept up to date at all and I think I missed that GAE is now a general-purpose hosting platform now (iirc in 2012 it was specific to a Goog
192.
by singhrac 3y ago
I'm at roughly 2 years, and I totally agree. If you make the friction to create a note high, you won't write anything down. I find tags sometimes helpful, but the only structure I really stick to is to create a daily note every da
193.
by singhrac 3y ago
I saw Google App Engine when skimming and had to check the date. The last time I launched a GAE project was in 2012.
194.
by singhrac 3y ago
Bitcoin advocates talk about the cash flows sponsoring new generation capacity, but everyone ignores the fact that bitcoin plants are being built in areas with low power prices, so they are necessarily less valuable to the grid than next-to
195.
by singhrac 3y ago
To offer a slightly different perspective than just "ecosystem", I learned Ruby a few years after Python, and found it kind of annoying, in that many Ruby programmers tend to invent their own DSLs and little hacks because the lang
196.
by singhrac 3y ago
Since this is a discussion on dependency management in Python - does anyone use rye [0] regularly now? I'm interested in using it but want a little more social validation before I try - some issues with package managers only appear aft
197.
by singhrac 3y ago
Extremely humble in this statement. He was hugely influential in statistics (Cramer-Rao lower bound, etc) and somewhat recently his work was also quite relevant to RL, eg with Rao-Blackwellization of gradient estimators.
198.
Setting up CUDA on WSL2 in 2023(rachitsingh.com)
2 points by singhrac 3y ago|0 comments
199.
by singhrac 3y ago
This code runs Llama2 quantized and unquantized in a roughly minimal way: https://github.com/srush/llama2.rs (though extracting the quantized 70B weights takes a lot of RAM). I'm running the 13B quantized model on
200.
by singhrac 3y ago
I've spent a lot of time writing and debugging multiprocessing code, so a few thoughts, besides the general idea that this looks good and I'm excited to try it: - automatic restarting of workers after N task is very nice, I have h
201.
by singhrac 3y ago
You can ask ChatGPT to output in Lilypond format, and use https://github.com/matthunz/staff to render it; this worked for me on your above example.
202.
Dealing with Pandas nullable float dtypes(rachitsingh.com)
2 points by singhrac 3y ago|0 comments
203.
by singhrac 3y ago
I would love to use Polars today (I have a lot of thoughts about good dataframe API design, and Polars gets a lot of the big ones right - including removing indexes). However we have a huge amount (and growing) of Pandas code - is there an
204.
by singhrac 3y ago
As a separate data point, I briefly switched one of our servers from an r6a.4xlarge (AMD Epyc) to a r6i.4xlarge (Intel Xeon) and saw a 30% speedup in our number-heavy compute task. I would love to find out why (MKL or AVX512? Do I need to r
205.
by singhrac 3y ago
I mean, this is classic whataboutism. That’s 16M trees over 23 years, whereas Scottish forestry plants ~20M per year [0]. If you look at Kerr’s Twitter, he’s part of the oil and gas lobby. The bigger issue is that they aren’t constructing p
206.
by singhrac 3y ago
I have spent a lot of time trying to use PPLs (including Pyro, Edward, numpyro, etc.) in Real World data science use cases, and many times mixing probabilistic programming (which in these contexts means Bayesian inference on graphical model
207.
by singhrac 3y ago
Intel has Habana Gaudi2, which is an A100 competitor, but you can only access it on Intel’s developer cloud, apparently.
208.
by singhrac 3y ago
Sure, compute shaders might work, but don’t you need rocBLAS, rocSPARSE, MIOpen, etc? Are people reinventing those in compute shaders in another package?
209.
by singhrac 3y ago
Sorry, do you have an example of using wgsl for prototyping? I'm having trouble imagining something like that, since WebGPU is not exactly terse.
210.
by singhrac 3y ago
No need for a Hilbert curve, you can just flatten pixels the usual way (ie X = img.reshape(-1)). The main issue is that attention doesn’t scale that well, and with a 512x512 img the attended region is now 262k tokens, which is a lot. The

More ›