Mat Kelcey

Mat Kelcey

@matpalm.com

data nerd wannabe https://matpalm.com/

414 followers129 following139 posts

Mat Kelcey reposted

If you're headed to #electronica, come check out the edge lab LIVE! events on Wed 11/11. I'll be hosting a workshop (reinforcement learning for #robotics) and evening interviews/demos. Use the code GUEST-SHAWN for free admission (first 50). #embedded #electronics Links ⬇️

Mat Kelcey reposted

At Imagine Innovators, we just announced our Developer Plan, offering expanded access to Edge Impulse’s powerful tools, completely free! Get GPU training, longer runtimes, more projects, and production-ready licensing, all at zero cost. Details: www.edgeimpulse.com/blog/introdu...

edgeimpulse.comIntroducing the Developer Plan: More Power, More Freedom — for FreeWith our new Developer pricing plan you’ll now get the best of our community and professional plans. That means GPU training, longer runtimes, more projects, and production-ready licensing, all at zer...

Mat Kelcey reposted

Making LLMs run efficiently can feel scary, but scaling isn’t magic, it’s math! We wanted to demystify the “systems view” of LLMs and wrote a little textbook called “How To Scale Your Model” which we’re releasing today. 1/n

keep coming back to "Scalable MatMul-free Language Modeling" arxiv.org/abs/2406.02528 full of fun ideas! am slowly mapping my FPGA audio rate cached casual dilated convolutions network to operate using their mat mul free GLU version of self attention. slowly getting there!

it's a super clever reparameterisation. though the tensors might need to be bigger, i feel it'll be paid for by the simpler streaming implementation on an FPGA. hacking / time will tell!

first logic gate waveshaper results tonight! this is network represented purely by combinational logic ( see arxiv.org/abs/2210.08277 ) i.e. just look up tables; no mults or adds. trained in jax ( 3 nested levels of vmap! ) & inference in amaranth. inference speed should be ~10s nanoseconds.

by a lot of measures this was a completely irrational purchase, but, you know what i like it! it's ok to be irrational some times.

started learning "Amaranth HDL" this morning. my prior FPGA neural net stuff was verilog but switching to amaranth since the match the gateware on a device i'm targetting. as someone with a python background it's 1e6 times nicer than verilog!!

Mat Kelcey reposted

Today STMicroelectronics @ST_World announced the incredible new STM32N6, a powerhouse microcontroller with built-in neural processing that achieves 600x performance over current high-end STM32 MCUs — and offers native Edge Impulse support. Press release: www.businesswire.com/news/home/20...

businesswire.comEdge Impulse and STMicroelectronics Launching Official Support of Revolutionary STM32N6 for Powerful Edge AI SolutionsEdge Impulse, the leading platform for building, deploying, and scaling edge machine learning models, and STMicroelectronics (NYSE: STM), a global sem

i love seeing people ( myself included ) rediscover old things. the recent flood of interest in AI which, for a lot of people, is 100% LLMs, is the best case i've seen of it! it really helps me solidify my only internal concepts to see someone else explain a decades old idea through new terminology.

sadly none of it "just works" with a keras3 jax model ( exported as stateless_function ) #sigh no doubt i've just "simply" got my `jax_module.methods[constants.DEFAULT_METHOD_KEY]` wrong or some other magic arg combo, but not sure i can be bothered today...

Mat Kelcey@matpalm.com

(this week) jax to tflite is done with orbax ( nightly build only) ai.google.dev/edge/litert/... ( and that's orbax-export, not orbax-checkpoint, since it's been split ) not sure what'll be next week, we'll see! #canYouKeepUp

you know you're in for a good time when you... 1. go to the highest level tutorial page 2. open in colab 3. run all cells and 4. hit errors :/