buildbot

46 ms·

931.
by buildbot 3y ago
I swear there was a google paper about using a DC as a really bad particle detector, but I can’t find :/
932.
by buildbot 3y ago
I can’t run llama on an intel iGPU, because it OOMs - the memory is unified because it’s in the same space and shared on M1. Not just marketing.
933.
by buildbot 3y ago
Not during residency, and I have not idea what the second part has to do with the match going underfilled. Are you saying med students saw this phenomena and decided emergency med was best left to PAs, and did not apply to match? What I hea
934.
by buildbot 3y ago
Right, but then half of all med students apply only to derm residency or other lucrative and competitive field with nice hours, leaving us in the situation we are in now where 18% of all emergency med department residencies went unmatched:
935.
by buildbot 3y ago
> Completely agree. If someone assaults you in real life, you should learn that you need to file charges to ensure your assaulter goes to jail. Not sit quietly in the corner while your abuser gets a talking to. Filing charges to make sur
936.
by buildbot 3y ago
Counterpoint - why don't we see defense of homeschooling at the same rate of attacks? Even if they are rare, there are several posters here that have experienced bad home schools. If they are at least 50:50, shouldn't we expect th
937.
by buildbot 3y ago
Sure, but a good example of how far certain domains have to go still. These datasheets should be in the models training data, at least one CCD datasheet, and verilog & (migen | nmigen | amaranth) certainly are. Controlling a CCD is actu
938.
by buildbot 3y ago
Well, neither GPT4 or this Phind model where able to answer my torture test: "Write amaranth code that can be used to control the readout of a frame from a kodak CCD with 4096 columns and 2048 rows." Which yes, is missing a lot of
939.
by buildbot 3y ago
Not exactly no - https://stackoverflow.com/questions/63960319/does-it-matter-... You need to be running a GPU driver on the host that supports the container cuda version. So in theory yes, in practice, weird issue
940.
by buildbot 3y ago
> One of the things that doable "now" that wasn't doable "then" is wireless charging. I love that I can put my phone on a pad next to the bed at night and it charges, pick it up and there is no trailing wire to u
941.
by buildbot 3y ago
Only The RTX A6000 (ampere) (Not RTX 6000! those are 24GB Turings!) has NVlink, and as far as Nevada states, they only work in 2x NVLink. Memory pooling is not really a real thing as much as they say it is. They run about 4K on eBay, and ha
942.
by buildbot 3y ago
I have not been able to get it working with 2 4090s yet, but I think that's because the library I choose to use (axolotl) does not support model parallel at all.
943.
by buildbot 3y ago
It’s tensor cores and floating point math have been either artificially or actually disabled - it would be very very slow. And 8gb of vram is really low as well.
944.
by buildbot 3y ago
AWS has trainwave, which is less off the ground that TPUs but does exist.
945.
by buildbot 3y ago
It was a really great feature, really well implemented too. It’s sad webos and the palm pre never really took off.
946.
"Serverless", GPU Accelerated OpenAI API Using Podman, Llama.cpp, and Systemd(blog.maxg.io)
2 points by buildbot 3y ago|0 comments
947.
by buildbot 3y ago
Oh I did not catch that … oops
948.
by buildbot 3y ago
I literally just started with htmx, but can’t you override both of those behaviors? hx-<http verb> on the form will switch the verb it uses, and you can swap outerhtml, innerhtml, or some other css target? Or am I incorrect?
949.
by buildbot 3y ago
Being a bit pedantic, you said employees don’t have contracts, as in any, I certainly have several. I agree most people don’t have “employment contracts” with defined timelines and places of work. I don’t have this.
950.
by buildbot 3y ago
Not true in this case, it was developed in the industry. https://azure.microsoft.com/en-us/blog/fostering-ai-infrastr... "Building on years of design space exploration and research at Microsoft, Microscaling
951.
by buildbot 3y ago
There is no mechanism per-say, it's more of a bit space vs quality issue. You could think of MX4 with an 8 bit exponent scale as a 12 bit number if the block size is one, "MX12" with E10M1. You can share the scale with some e
952.
by buildbot 3y ago
> Employees generally don't have contracts. Everyone I know signed one. It’s just one that can be terminated by either party at any time…
953.
by buildbot 3y ago
And a separate federal crime if messed with. Impersonating the IRS over USPS has to be one of the fastest ways to end up with two federal agencies after you…
954.
by buildbot 3y ago
I would hope so ;)
955.
by buildbot 3y ago
Exactly!
956.
by buildbot 3y ago
Copy my comment here too - Point of clarification - there is no E8M0 direct datatype (unless I misunderstand something!) E8M0 is only used for the scaling of exponents in the block - there is 8 bits of scale per block.
957.
by buildbot 3y ago
Point of clarification - there is no E8M0 direct datatype (unless I misunderstand something!) E8M0 is only used for the scaling of exponents in the block - there is 8 bits of scale per block.
958.
by buildbot 3y ago
This is for both training and inference, for what it is worth.
959.
by buildbot 3y ago
Typically, you need to use some tricks for pre-training in lower precision (finetuning seems to work at low precision), with FP16 you need loss scaling for example. With MX, you can train in 6 bits of precision without any tricks, and hit t
960.
by buildbot 3y ago
To finetune on each benchmark? I'd say it's not in our modern era of in-context learning, though of course fine-tuning has it's place as well for making smaller models better in one domain than a generalist larger model.

More ›