There are some great resources here! And crazy to see this amount of material over past few weeks Shout out to awesome partners @GregKamradt @pinecone @anyscalecompute @AssemblyAI
COMPUTING
-
Condor Galaxy 1 Powers BTLM and Jais Open-Source LLMs
By
–
Condor Galaxy 1 (CG-1) was used to train BTLM, the top-performing open-source 3B parameter LLM. Now, CG-1 has been used to train Jais, the best open-source Arabic model. CG-1 is moving the open-source community forward. Contact us to train on CG-1 http://
cerebras.net/contact-us -
PDEs Enable Direct Solution Validation From Equations
By
–
With PDEs, you can check if a solution is valid or not, directly from the equation
-
CodeLlama PR merged with 4-bit quantization inference benchmarks
By
–
The CodeLlama PR just got merged: https://
github.com/Lightning-AI/l
it-gpt/pull/472
… When I tried it with bnb's 4-bit Normal Float quantization, the 34B Instruct and Python variants used about 20 Gb for inference: -
AI and Cloud Transform Formula 1 Race Strategy and Fan Engagement
By
–
Dive into the world of #Formula1 with Rob Smedley, ex-race engineer for Ferrari & Williams. Data, AI, & cloud's impact on F1 Revolutionizing race strategy with connectivity @F1
's partnership with @awscloud Enhancing fan engagement -
ML Replaces Hand-Written Heuristics in Compilers
By
–
There's a huge opportunity to use ML to replace hand-written heuristics in compilers. Want to try? Check out this @kaggle contest organized by Ashley Chow, Bryan Perozzi, HCL-Jevster, inversion, Mangpo Phothilimthana, and Sami Abu-El-Haija!
-
Google Cloud Unveils MultiSlice Innovation with Nisha Johnson
By
–
The great work here is from @nisha_m_johnson and so many other colleagues in @googlecloud and elsewhere across Google who developed MultiSlice!
-
34B Parameter Models Now Run on Laptops, Massive Hardware Progress
By
–
Crazy how 34 billions parameters models seemed huge and unmanageable outside of a data center just maybe 1.5 years ago. Now it’s laptop stuff https://t.co/gXCg770BOM
— Thomas Wolf (@Thom_Wolf) 31 août 2023Crazy how 34 billions parameters models seemed huge and unmanageable outside of a data center just maybe 1.5 years ago. Now it’s laptop stuff
-
Joy and Art in Debugging Low-Level Numerical Issues
By
–
there is surprising joy & art in debugging low-level numerical issues
-
Efficient LLM Training QLoRA Llama 2 Resource Optimization
By
–
Depends on your settings. But if you limit the context size to like 2048 (like in the NeurIPS competition) and use a microbatch size of 1 with gradient accumulation and qlora with llama 2 7B, that’s approx 20 GB RAM and shouldn’t take too long, maybe an hour.