liuliu

32 ms·

481.
by liuliu 4y ago
Yeah, stable diffusion's PyTorch code is not optimized for inference memory usage at start. I am looking at the code now and it seems that if it is converted to static graph, there are probably a bit more opportunities (I only looked a
482.
by liuliu 4y ago
Can these just be shared memory all the way through or in your case, the persistence to disk at fork time is important?
483.
by liuliu 4y ago
Probably another week or two. running on M1 iPad Pro needs to get out of PyTorch, possibly export the model either through TorchScript and then do onnx conversion. From what I found so far, not many of these conversions done (except the Ope
484.
by liuliu 4y ago
Containerization also helps the trend. Previously, you manage JVM runtime separately in your cluster and the deployment becomes simpler (just deploy the jars). Without containerization, maintaining separate runtime dependencies for compiled
485.
by liuliu 4y ago
Not sure. But it seems they at least aware of Alex's work on image recognition systems because the CIFAR-10 reference. Training ImageNet at that time involves a lot of engineering work.
486.
by liuliu 4y ago
NCNN does work with just CPU. Just need a little bit more packaging.
487.
by liuliu 4y ago
> I think the most intoxicating thing about SQLite is that you don't have to install or configure even one goddamn thing. There are a lot of configuration options when compiling SQLite, and some of its defaults are pretty conservati
488.
by liuliu 4y ago
Yeah, most databases probably will have pathological behaviors against your requirements (especially on tail-latency, which you would care about). Many implement similar tools would put a lightweight compression on top and just dump these s
489.
by liuliu 4y ago
It is weird. I think both actors and the upcoming ownership model (for Swift 6.0) are important. But they are huge to consume with (hard to understand, big foot-print in compiler). Maybe at some point, we can have Swift 7.0 which demotes cl
490.
Openness Trumps Profit-Seeking, What We Can Learn from Decade of AI Renaissance(liuliu.me)
2 points by liuliu 4y ago|0 comments
491.
by liuliu 4y ago
I don't know. Open-source for simple things such as how to talk to a database is overly crowded. It is not as simple as "have a free product and people will come" any more which may be true 10 years ago. All I am saying: it t
492.
by liuliu 4y ago
Oh, great! These are two big changes I am very interested in. Will look out when it is merged :)
493.
by liuliu 4y ago
The BEGIN CONCURRENT and wal2 are new things? Seems the former is only added in 3.39.0? Didn't get call out in the release note.
494.
by liuliu 4y ago
> I assume that means "efficient at runtime" - but by what metrics? CPU time? Memory usage? Power usage? Latency/responsiveness? Throughput? Not OP, but coming from Python, I find it is liberating to say: I need to write a
495.
by liuliu 4y ago
They do now offer more than Ubuntu (and now with arm64 on Ubuntu too, I guess for rPi) https://www.swift.org/download/ . It is quite usable with SwiftPM or Bazel on Linux. And SwiftPM can also generate static binaries (
496.
by liuliu 4y ago
Assuming you have to do a lot of string processing (for NLP). Do you have any "shortcut" to do that in Swift? The API is thoughtful, but tedious to type, comparing to Python. It would be great if we have some semi-official "s
497.
by liuliu 4y ago
Would be interesting to know the "moving away from Swift to Rust / Zig" story. My impression is that Swift not even in consideration in most projects with Rust / Zig.
498.
by liuliu 4y ago
Yes, it is educated that way. However, for the case requires some scratch space, dynamic allocations can be easier to manage, less error prune and with less overall memory usage (you don't need to preallocate worst case amount). That&#
499.
by liuliu 4y ago
Well, can IP be stolen? It is a state-protected monopoly. If you steal my IP, we both have these IPs. Using the word "stealing" is fundamentally wrong about intellectual creations. What frustrating me more is that Roborock cannot
500.
by liuliu 4y ago
The post has more leg than you give it credit for. The author thinks serializable is table stake, and linearizability can be implemented cheaply (through sharding). Strict-serializable however, cannot really break "scale-out" barr
501.
by liuliu 4y ago
Read this post a few days ago. I am trying to articulate my answer properly, mostly from a client-side dev's perspective. Here is my half-baked thoughts: Strict-serializability pretty much guarantees that you don't need any extra
502.
by liuliu 4y ago
To people who are as confused as me. The "nonlinear activation free" doesn't mean "linear" in the paper (otherwise this is going to be ground-breaking discovery). They use polynomial functions in place of traditiona
503.
by liuliu 4y ago
Read a bit of your code. Seems most magic happens around xLock / xUnlock implementation (which were used to identify a txn). I wonder if SQLite should expose one layer up (on the actual pager level). A lot of attempts (such as yours, o
504.
by liuliu 4y ago
Seems the benchmark does read into native structs in both Flatbuffers and this in Go. I am confused by the performance claim too. Seems very similar design v.s. fbs, would be surprised if it is from format differences rather than codegen &#
505.
by liuliu 4y ago
I am a bit out of depth here for obvious reasons (I work above SQLite, not under SQLite). But from my understanding, `journal_mode=memory` simply means it is in rollback mode, but rollback pages are not maintained on disk. Therefore, the mo
506.
by liuliu 4y ago
This really, IMHO (as someone implements things on top of SQLite too https://dflat.io ) pushes SQLite too far as the implementation of cross-db transactions have some big issues: https://www.sqlite.org/limits.html
507.
Fun with Motion-JPEG, Jupyter Notebook and How to Stream GUI in the Browser(liuliu.me)
1 points by liuliu 4y ago|0 comments
508.
by liuliu 4y ago
I wish this is true. But from all the things Swift has supported. It doesn't make compiler or compiler authors happy. I wish there are more languages that focused on happiness of the compiler authors so that they can write efficient an
509.
by liuliu 4y ago
The worry is about dataset shifting. Previously, the data were collected for a few hundreds thousands structures, now it is 200m. I think there could be doubts on distributions and how that could play a role in prediction accuracy.
510.
by liuliu 4y ago
Before ZFS, you can layering bcache on top of your file system. But otherwise, ZFS L2ARC I believe nowadays also does persistence L2ARC.

More ›