Navdeep⚡ (@dev_navdeep) header
Navdeep⚡ (@dev_navdeep) profile picture

Navdeep⚡

@dev_navdeep

•developer •finetuning LLMs Building All in one AI for you https://xitter.cf/t.co/P94XO6mP9u

scrol.ai/
Joined April 2018
  • Tweets5.3K
  • Following163
  • Followers497
573Media
Photo by Navdeep⚡ (@dev_navdeep)Photo by Navdeep⚡ (@dev_navdeep)Photo by Navdeep⚡ (@dev_navdeep)Photo by Navdeep⚡ (@dev_navdeep)Photo by Navdeep⚡ (@dev_navdeep)Photo by Navdeep⚡ (@dev_navdeep)Photo by Navdeep⚡ (@dev_navdeep)Photo by Navdeep⚡ (@dev_navdeep)
nobody talks about nodejs much on twitter but today is not that day. →Learn basics of what and how of Nodejs: LIKE AND RETWEET thread 🧵👇
Image from Navdeep⚡ (@dev_navdeep)
14
72
3
311
Re-designed the gear drive to be 16:1 gear reduction, needed the extra kick for small sized robotic leg actuator. Improvements:- 1. Torque peak 5Nm(before 2Nm) 2. Torque nominal 3Nm 3. Weight ~305gm 4. Smooth small scale motion.
Image from Navdeep⚡ (@dev_navdeep)
39
Just finished building complete MIT style QDD actuator module with 4:1 gear drive train. Improvement from last design: - can communicate over CAN bus. - easily programmable with arduino - expected good torque rating (yet to benchmark fully). - handling high current upto 40 amps
Image from Navdeep⚡ (@dev_navdeep)
65
Just finished prototyping cheapest possible foc based position control motor/actuator. - 1: 4 planetry gear based output. Able to do velocity, torque and position control. Can not find a good affordable 3 phase motor driver in india... 🥲
1
92
"deep learning worked"
Sam Altman (@sama) profile pictureSam Altman
@sama23 Sep 2024
The Intelligence Age: ia.samaltman.com/
The Intelligence Age

The Intelligence Age

In the next couple of decades, we will be able to do things that would have seemed like magic to our grandparents.

ia.samaltman.com
1
79
If anyone is interested in how CPU works, I have created this project that work as 6502 cpu emulator, you can load assembly programs directly onto cpu, also included some references. looking forward to create BIOS for it. github.com/navdhakar/6502cpu
Image from Navdeep⚡ (@dev_navdeep)
5
159
If you want to use @karpathy 's llm.c gpt2 on low end gpu (in my case MX 330, 2GB memory), you can use unified memory in cuda, just replace cudaMalloc->cudaMallocManaged
Image from Navdeep⚡ (@dev_navdeep)
Image from Navdeep⚡ (@dev_navdeep)
4
5
1
135
35,783
If you want to use @karpathy 's llm.c gpt2 on low end gpu (in my case MX 330, 2GB memory), you can use unified memory in cuda, just replace cudaMalloc->cudaMallocManaged
Image from Navdeep⚡ (@dev_navdeep)
Image from Navdeep⚡ (@dev_navdeep)
4
5
1
135
35,783
- explanation It basic doesn't statically load all parameter in GPU but stream it from shared memory (SLOW) - possible optimization Training using this is bad but looking to optimize this with "cudaMemPrefetchAsync" etc. for reasonable inference speed on consumer gpu.
2
850
even small LLMs are insanely data hungry, I fine-tuned mistral7B with 20k data points it still overfits in less than 100 iterations
Image from Navdeep⚡ (@dev_navdeep)
1
193
Some things to do as a developer. Been diving deep into linux for a month, Just read this book, awesome stuff, was very Intimidated by linux, but when you can build kernel and drivers from scratch linux become so much easier. Resources that I went through👇
2
3
213
more replies
- deep dive books (kernel, drivers) kroah.com/lkn/ lwn.net/Kernel/LDD3/ - install gentoo os (how real life linux distros are packaged and installed) gentoo.org/ - awesome youtube channels youtube.com/@Doriandotslash youtube.com/@nirlichtman
1
1
110
One important channel I forgot, he gets right into meaty stuff of building real life linux drivers, its easy and its useful, you will never complain why your nvidia card is not working on linux. youtube.com/watch?v=XoYkHUnm…
98
I ran @karpathy's llm.c as pytorch and native c on low-end cpu. difference is already significant. 👇 below is simple comparision. 1. model compiled with C -almost 1 GB less memory use (no pytorch overhead) -efficient use of all 8 threads on cpu
Image from Navdeep⚡ (@dev_navdeep)
1
129
2. model using pytorch - some cores are being used and some not - +1GB memory uses as compared to native these memory save and efficiency can be useful if we want to run these model in offline on consumer hardware.
Image from Navdeep⚡ (@dev_navdeep)
64
flex (PS: I like to waste my time iykyk) genduu os
Image from Navdeep⚡ (@dev_navdeep)
1
58
Was thinking of buying H100 gpu, But reviews are scary
Image from Navdeep⚡ (@dev_navdeep)
81
✅successfully trained 1st test LLM for generating prompt from text chunks of raw data that can acts as synthetic data generator for anyone to make domain specific dataset. -comparison between base model and finetuned
Image from Navdeep⚡ (@dev_navdeep)
1
1
147
difference might seem very subtle but the finetuned version ask question in much more depth regarding variables in code etc. and try to mimic capabilities for GPT-3.5, which can make a lot of difference when this is used to train model that has knowledge about this data.
1
44
find model weights and code to test it out here. huggingface.co/navdeepdh/pro…

navdeepdh/prompt-gen-adapters · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co
38
Load more