lhl

6 ms·

1.
by lhl 4d ago
I think that Doctorow, Zitron, and other "denialists" are doing a real disservice to their audiences and it's only going to make the future shock worse. The basic claim that HF incident isn't evidence of consciousness or
2.
by lhl 24d ago
Pretty interesting to see on the training front. I've used most of these models to grind semi-autonomously (days at a time) on kernel optimizations (except for Fable - it kept triggering guardrails almost immediately and bouncing me do
3.
by lhl 1mo ago
Qwen 3.x does have an advantage but it's relatively small (64KB/token vs 80KB/token) - Gemma4 actually has less % of full attention layers, but the largest geometry and has the biggest "fixed" state for it's
4.
by lhl 1mo ago
For updated/validated updates, Donato Capitella maintains independent Strix Halo "toolboxes": https://strix-halo-toolboxes.com/ A team from AMD maintains Lemonade, another all-in-one setp with convenient inst
5.
by lhl 2mo ago
I started a project w/ the assumption that fully tuned for HIP kernels could get closer to roofline and beat llama.cpp's performance (and vs their ROCm engine it does!) but after thousands of iterations/experiments my suspici
6.
by lhl 2mo ago
The one thing that's new/worth pointing out are the https://developer.amd.com/playbooks/ ( https://github.com/amd/playbooks ) - this is AMD's answer to Nvidia's playbooks ( https
7.
by lhl 2mo ago
You got me curious, so I made a little harness comparison to my model test suite: Model Adapter Suite Score Passed Tasks --------------------------------- ------------ ------------
8.
by lhl 2mo ago
https://x.com/thsottiaux/status/2058071172361998482 "A little secret. About 5% of our production traffic is on the Pi harness, about another 5% is on OpenCode. Reminder you can use your ChatGPT account in a f
9.
by lhl 3mo ago
I've been testing Ornith-1.0 35B (my own FP8-block quant) and I like it. It runs at >200 tok/s w/ vLLM on an RTX PRO 6000 (sm120), I've run >140M cached tokens of agentic coding work on it over the past few days. I
10.
by lhl 4mo ago
It's still Python, but I removed torch dependencies (HIP/C++ for hot paths): http://github.com/shisa-ai/hipEngine/ There's a docs/ folder in there that is probably of interest as well.
11.
by lhl 4mo ago
Took a little longer to clean up than I expected. I'd recommend checking out the ROOFLINE and the LESSONS-LEARNED docs here: https://github.com/shisa-ai/hipEngine/tree/main/docs
12.
by lhl 4mo ago
When it's in a good state I'll open source it, I am keeping track of what optimizations make the most impact, stuff like this: ### Diagnosing parallelism pathologies (L1) *Grid occupancy:* - `Grid_Size / Workgroup_Size >=
13.
by lhl 4mo ago
I think especially with the ability for SOTA AI to optimize kernels more people should try their hand at making better inference for their specific hardware. I have an older W7900 (RDNA3) which, besides 48GB of VRAM, has some pretty decent
14.
by lhl 4mo ago
Oh, is this actually out now? If so, great, but I took a quick look and didn't spot any third party review yet. For those interested in this laptop, personally I'd still wait for some reviews from some real world people. Some hist
15.
by lhl 5mo ago
For those that don't want their data trained on, OpenRouter allows you to have account-wide or per-request routing with either provider.data_collection: "deny" or zdr: true (zero data retention). Also, you can use HuggingFace
16.
by lhl 5mo ago
RDNA is a whole different (and much poorer supported) animal than CDNA. As someone with extensive experience in both, if you're asking the question, then, no. (If you're just looking to learn, use the free Kaggle/Google Cola
17.
by lhl 5mo ago
BTW, a number of corrections. The TurboQuant paper was submitted to Arxiv back in April 2025: https://arxiv.org/abs/2504.19874 Current "TurboQuant" implementations are about 3.8X-4.9X on compression (w/
18.
by lhl 5mo ago
As some other people mentioned, using both/multiple is the way to go if it's within your means. I've been working on a wide range of relatively projects and I find that the latest GPT-5.2+ models seem to be generally better c
19.
by lhl 6mo ago
Like others have mentioned, I think the premise of looking at the most popular few projects (pypi.org currently lists 771,120 projects) on pypi as any sort of proxy for AI coding is terribly misguided/unrepresentative and that almost n
20.
by lhl 6mo ago
Funy that you mention multi-monitor since it's one of the reasons I eventually moved to Wayland. The only way to support different DPI monitors in X was to do janky scaling or even jankier multiple X servers. I don't use KDE (or G
21.
by lhl 6mo ago
Yes, I read it and specifically pointed it out (that's why there are 3 hours of interactive logs). There are 4 other runs pushed now so you can see what actual clean room runs for 5.2 xhigh, 5.3-Codex xhigh, 5.4 xhigh, and Opus 4.6 ult
22.
by lhl 7mo ago
I am not a theoretical CS or math expert by any means, but I have been wrangling coding agents for a while and reading the paper and the problems Stapper had with dealing w/ Claude (context management, instruction following, etc) dec
23.
by lhl 7mo ago
I was a bit interested to do a replication and see if better harness could avoid some of the problems they ran w/ context management, poor instruction following, etc and it looks like yes, it's definitely possible. Here's my
24.
by lhl 9mo ago
Anecdotally, I will say that for my toughest jobs GPT-5+ High in `codex` has been the best tool I've used - CUDA->HIP porting, finding bugs in torch, websockets, etc, it's able to test, reason deeply and find bugs. It can'
25.
by lhl 11mo ago
We do live in an age of frontier LLMs... For fun, I'll just use Kimi K2 (on Kagi Assistant). > Can you explain what this means and its significance? Assume that I'm a layperson with no familiarity with LLM jargon so explain all
26.
by lhl 1y ago
Looks great, thanks for the suggestion!
27.
by lhl 1y ago
I'd previously been giving Hyprland a try, but after lots of customization work, there were still a bunch of things I wasn't happy with and ended back on GNOME as a "just let me get work done" thing (I use multiple works
28.
by lhl 1y ago
In Linux, you can set it as high as you want, although you should probably have a swap drive and still be prepared for you system to die if you set it to 128GiB. Here's how you'd set it to 120GiB: # This is deprecated, but
29.
by lhl 1y ago
RDNA3 CUs do not have FP8 support and its INT8 runs at the same speed as FP16 so Strix Halo's max theoretical is basically 60 TFLOPS no matter how you slice it (well it has double INT4, but I'm unclear on how generally useful that
30.
by lhl 1y ago
Apple actually makes a lot more acquisitions than you think, but they are rarely very high profile/talked about: https://en.wikipedia.org/wiki/List_of_mergers_and_acquisitio...

More ›