HackerTrans
TopNewTrendsCommentsPastAskShowJobs

emikulic

no profile record

comments

emikulic
·3 lata temu·discuss
It's over for decels.
emikulic
·3 lata temu·discuss
Three epochs means it sees each token three times. The dataset is ~1T like you said.
emikulic
·3 lata temu·discuss
I want flying cars and moving sidewalks like I was promised by The Jetsons. :(
emikulic
·3 lata temu·discuss
Try https://huggingface.co/ehartford/WizardLM-7B-Uncensored and related models. They're not even trained on smut, just the neutered responses were removed before the RLHF stage (IIUC)
emikulic
·3 lata temu·discuss
I used to be bad at math but then I did a 360° on that.
emikulic
·3 lata temu·discuss
Emad said SD cost $600,000 to train. I wonder if Midjourney also had to pay that to train from scratch.
emikulic
·3 lata temu·discuss
What benefits does the Huggingface diffusers(?) implementation have over A1111?
emikulic
·3 lata temu·discuss
I like to run A1111 in --api mode and write my own script to drive it over HTTP.
emikulic
·3 lata temu·discuss
Wow, Reddit is going to lose a lot of pagerank.
emikulic
·3 lata temu·discuss
You're thinking of archive.is
emikulic
·3 lata temu·discuss
You mean the one that defeated two superpowers?
emikulic
·3 lata temu·discuss
Last time I looked into this, the answer was "because huggingface transformers and torch.load are written to do it this way"

You could absolutely do something streaming, or mmap the weights instead of loading them into system RAM. Just the default interfaces don't.
emikulic
·3 lata temu·discuss
This used to be the case but nowadays you can just pip install e.g. https://libraries.io/pypi/nvidia-cuda-runtime-cu11