pavelstoev

7 ms·

31.
by pavelstoev 2y ago
It is written in Python itself and emits efficient CUDA code. This way, you can understand what is going on. The current focus is on inference, but hopefully, training workloads will be supported soon. https://github.com/hid
32.
by pavelstoev 2y ago
The hallucinated code was reusing memory buffers filled with previous results so not performing the actual computations. When this was fixed the AI generated code was like 0.3x of the baseline.
33.
by pavelstoev 2y ago
or use Hidet compiler (open source)
34.
by pavelstoev 2y ago
Is this hypothetical person someone you know? if yes, please email me to pavel at centml dotz ai
35.
by pavelstoev 2y ago
20 years ago, I broke my ankle in a friendly soccer game on the Cambridge Commons field. I was shocked when the doctor who performed surgery to put it back together ordered me to walk on crutches the next day after surgery and to continue p
36.
by pavelstoev 2y ago
poison pill offer ?
37.
by pavelstoev 2y ago
It is really just “open use” with detailed defined by the license type (MIT, etc)
38.
by pavelstoev 2y ago
Interesting take. They are now reading our minds vs looking at our kids and interiors.
39.
by pavelstoev 2y ago
English economist William Stanley Jevons vs the author of the article. Will NVIDIA be in trouble because of DSR1 ? Interpreting Jevon’s effect, if LLMs are “steam engines” and DSR1 brings 90% efficiency improvement for the same performance,
40.
by pavelstoev 2y ago
The problem is that performance achievements on AMD consumer-grade GPUs (RX7900XTX) are not representative/transferrable to the Datacenter grade GPUs (MI300X). Consumer GPUs are based on RDNA architecture, while datacenter GPUs are bas
41.
by pavelstoev 2y ago
Love both games spending countless hours, Unreal single player story was great (for its time). Are there any servers still online for UT?
42.
by pavelstoev 2y ago
I recommend hidet backend in torch.compile - implements many advanced model-specific optimizations automatically. https://github.com/hidet-org/hidet
43.
by pavelstoev 2y ago
that feeling you get when you realize the person who lived ~2300 years before you is smarter than you now...
44.
by pavelstoev 2y ago
Keyword - survived. Everything must become better.
45.
by pavelstoev 2y ago
I strongly disagree with your comment. First of all, ALL kids must survive, not most. Safety and safety of knowing are two different things. Lastly, this tech looks a lot less intrusive than the current another watch that everyone is gettin
46.
by pavelstoev 2y ago
Very interesting project and good progress on making private LLM use cases more accessible and usable, please keep going !
47.
by pavelstoev 2y ago
Next iteration of this will be video generated on demand with GenAI running closer to the request, ideally at the request.
48.
by pavelstoev 3y ago
I think what will emerge eventually are LLM architecture specific ICs.
49.
by pavelstoev 3y ago
Should I be worried ? I just landed at SFO
50.
by pavelstoev 3y ago
Best Tflops/$ is actually 4090, then 3090. Also L4
51.
by pavelstoev 3y ago
Check out Hidet [1]. Not as well funded, but delivers Python based ML acceleration with GPU support (unlike Mojo). [1] https://github.com/hidet-org/hidet
52.
by pavelstoev 3y ago
How in Mojo do you support GPU data parallelism and all the benefits it brings ?
53.
by pavelstoev 3y ago
Meanwhile - Brightline, Orlando to Miami 125 MPH, $6B and 4 years. https://www.wptv.com/news/region-martin-county/stuart/bright... .
54.
by pavelstoev 3y ago
you can also profile AI/ML performance without actually running it https://github.com/CentML/DeepView.Profile
55.
by pavelstoev 3y ago
We build software acceleration for LLM, effectively running smaller llama2 models at the same performance on several L4's as on 1xA100.
56.
by pavelstoev 3y ago
You can do this in software, effectively.
57.
by pavelstoev 3y ago
This is precisely how I passed the lab portion of my Control Systems course in undergrad engineering.
58.
by pavelstoev 3y ago
yes there is a much more performant competitor that actually supports Nvidia GPUs [1] https://centml.ai/make-your-ml-models-run-faster-with-hidet/
59.
by pavelstoev 3y ago
I have an SRX1400 cluster with nearly 11 years of uptime. But this is the bare metal BSD-based firewall. When later, Juniper started adding a hypervisor layer on SRX platforms, and I started having problems.
60.
by pavelstoev 3y ago
Very interesting. Well done authors!

More ›