fisf

8 ms·

61.
by fisf 3y ago
https://nvidia.github.io/TensorRT-LLM/performance.html It was one of the fastest backends last time I checked (with vLLM and lmdeploy being comparable), but the space moves fast. It uses cuda under the hood, torch is n
62.
by fisf 3y ago
I am pretty sure that does not activate the hardware encoder. https://trac.ffmpeg.org/wiki/HWAccelIntro
63.
by fisf 3y ago
I am an end user of AI in many cases. I use it to generate graphics, summarize notes I have drafted, explore topics (instead of searching to some degree), etc. I absolutely want AI driven tooling in place of the manual / tedious option
64.
by fisf 3y ago
The issue is that both your statements are ultimately very subjective opinions first and foremost. It's absolutely fine to hold those, but then people extrapolate this to things like: "The "benefits" it provides to end u
65.
by fisf 3y ago
Also ignores any required authentication. This is basically just a load test. Why has the security field degenerated to such a shitshow?
66.
by fisf 3y ago
There is a reason for that. Tech debt, code quality, poor SLA metrics are symptoms, and proxies for various things (that can be fixed). Failures in the other areas are root causes and more fundamental issues.
67.
by fisf 3y ago
Yes, the response could be more patient, but ultimately, that's absolutely a user error. If the documentation clearly says that something should not work, the same thing does not work in other DBs (-> semantics follow Postgres for e
68.
by fisf 3y ago
Neither of those are required for intelligence .
69.
by fisf 3y ago
That is entirely correct. But it also implies that employees need to price that uncertainty into their salary demands.
70.
by fisf 3y ago
So we outsource bad decision making of overcompensated people to the public. Brilliant.
71.
by fisf 3y ago
Sure, but without funding and/or massive support from MS this is not going to happen.
72.
by fisf 3y ago
But in practice, it is often not possible or allowed to simply add other software or hardware to those systems. Those are delivered and certified as-is and upgrades are a no-go.
73.
by fisf 3y ago
Funny note: Implementing e.g. a flexible observer pattern in Rust has the same ownership and lifetime problems. Usually this involves some weak reference, central broker, etc. So, that's really not only a language problem, but rather t
74.
by fisf 3y ago
That's not a framing issue. The original claim was > Generally immigrants are job creators and entrepreneurs, and take less from the system than native born people do And this is (at least in the context of many European countries)
75.
by fisf 3y ago
You demand a receipt and call that script kiddie garbage. What an irony. ROCm has kinda worked for some APUs in 5.x for a while now. As much as things are expected to work on AMD hardware anyway. So basically install ROCm 5.5 and check if
76.
by fisf 3y ago
That's not really what "puddle deep knowledge of anything that isn't iPhone firmware" implies.
77.
by fisf 3y ago
Yes. But at some point a quantitative difference becomes qualitativ.
78.
by fisf 3y ago
Onnx is just a serialisation format (using protobuf iirc) for the network, weights, etc. Your training framework and a suitable export is the compiler. Onnx Runtime (which really has various backends), tensorrt, .. (whatever inference engin
79.
by fisf 3y ago
All the recent qualcomm stuff has some kind of dedicated ai support (special vector extensions, etc.). Qualcomm has it's own SDK for that (used to be called SNPE), which uses GPU, DSP (hexagon),.. CPU is really only a fallback.
80.
by fisf 3y ago
Yes, that's basically an entity component system. It still runs into the ‘fake memory leaks’ problem the author describes for obvious reasons, i.e. you still have to deal with "components" attached to some handle somewhere (a
81.
by fisf 3y ago
> Have you ever tried to iterate developing any serious class of algorithms in C++? Yes, and it's not so bad. A lot of ML deployments have been based on C/C++ for inference anyway (with Python driving the training). So that&#x2
82.
by fisf 3y ago
There never was a broom. I.e. it never existed as a permanent, indivisible thing.
83.
by fisf 3y ago
That some goals are 'easier' accomplished by violence, is nothing you need to point for religious (and other) extremists. They are well aware of that tool in their box. It's also something politicians are well aware of (or a
84.
by fisf 3y ago
Because not everything that's useful is layman-understandable.
85.
by fisf 3y ago
Yes,they are both foreign powers. But one is a undemocratic dictatorship.
86.
by fisf 3y ago
It also has to do with the people the startup scene attracts nowadays. 10 years ago, the majority of candidates for a tech startup qualified as 'builders', very broadly speaking. Now, a larger chunk want to work on 'product&#
87.
by fisf 3y ago
Come on. The three relevant ways to access the system via API (Rest, graphql, SQL) + available frontend integrations are top level entries in the documentation ( https://cube.dev/docs ).
88.
by fisf 3y ago
Java pretty much provides BigDecimal for that
89.
by fisf 3y ago
> I believe you that you care about those, but no one cares only about those. If I were you I would cultivate the skill of tracking my other motives as well. That isn't comfortable, but it eventually produces huge benefits in just t
90.
by fisf 4y ago
I am rather concerned about the message, than the particular choice of words. Because the message is downright authoritarian and militaristic, aggressive -- no matter how eloquently it is formulated, it is inherently crude.

More ›