Johannes Gäßler

icon indicating an email [email protected]

icon company Karlsruhe Institute of Technology icon location Karlsruhe, Germany icon location Experimental particle physicist. Other interests: statistics, generative neural networks, hardware acceleration.

Results 235 comments of Johannes Gäßler

Llama3-8b & Perplexity.exe Issue

Does it work with a plain text file?

Impact of bf16 on Llama 3 8B perplexity?

I did not explicitly check the effect of FP16/BF16 as an intermediary but when using them directly I basically found no relevant differences: https://github.com/ggerganov/llama.cpp/pull/7150 . And because the FP16 vs....

ThunderKittens：a simple yet faster flashattention alternative

The design philosophy of ggml/llama.cpp is **not** to use external dependencies if at all possible. I was recently informed by an NVIDIA engineer that the way to go for tensor...

common, ngram_cache: added const reference for std::pair<> and std::tuple<> more 16 bytes:

Good catch with the `&`. Does the compilation work if you just remove the `const`?

Improve and expand Wikipedia article about llama.cpp

>Write more information on the Wikipedia article about how llama.cpp works, citing reliable secondary and tertiary sources. I don't think there are any (for the technical aspects). For that primary...