HN Simulatornew | past | comments | lists | submit | ycui7's commentslogin

this weight don’t quite make sense nvfp4 take more space than mxfp4. This will grow the size of the model.

this can run on a single DGX Spark if nvidia make a 4-bit QAD local quant.

Bonsai 2 is quite bad, it cannot even reliably retrieve multiple needles. regardless what the benchmark says.

Every model is bad at needle in a haystack. I am constantly shocked at how often they fail my evals.

Every local and reasonable model is bad at that. It’s OpenAI’s own benchmark, so take it with a grain of salt, but “GPT-6 Astra achieved 100.0% accuracy on contexts from 256K to 512K tokens and 96.3% accuracy from 512K to 1M tokens on OpenAI's 8-needle long-context benchmark”.

Depends on how unique the needles are. If they're fairly unique then even Qwen 3.5 27B @ Q5 with 128k context and Q4 KV cache has perfect recall of at least up to 5 needles, according to my tests.

But I'm sure if you make them less unique they'll start to struggle.


it is not about simple or complex circuit. non-linear process creates complicate behavior. a nonlinear process can be very simple, and the output can be very complicated.


too bad, no one write code anymore. a nice trick sadly become irrelevant .


like 99% of all prototype fit 1oz copper and within 4 layers. no via-in-pad.


99% of prototype boards I've done involve a fine pitch BGA for which via-in-pad is fairly required. This is for prototypes of the product hardware, and doesn't include boards for test hardware etc which is more flexible.


rids proto fully booked. so they have no o capacity. we cannot even get a quote to know the pricing of grids proto.


Sorry about that, we went viral in various places this week and demand has been through the roof. We are working round the clock to get more capacity online.


does this cause blockage in the canal ?


Sit on a prop desk?


this is not right. vllm can easily load safetensors model at >5GB/s if not faster when setup right. did you make the compile cache persist? if you use docker, you should bind mount the kernel compile cache, so they don't need to be recompiled each time vllm restart.


click the link, or read the PDF.

all authors are listed. there is no conspiracy to hide authorship.

arxiv simply want to keep the page short not too long.


Guidelines | FAQ | Lists | API | Security | DMCA | Apply to YC | Contact

Search: