ggml-org
Documentation fix: Quantum -> Quantized.
- [x] I have read the contributing guidelines
- Self-reported review complexity:
- [x] Low
- [ ] Medium
- [ ] High
Description of changes
I just fixed a small spelling error at imatrix/readme and server/readme (specifically: quantum models -> quantized models). While it would be great to have support for Quantum Mechanics based language models in future 1, I am sure this refers to quantized models (models compressed by reducing precision of parameters).
Why I bothered making such a small PR?
I refer to these documents (imatrix/readme, server/readme) very often. At first, I just ignored the spellchecks. After a while, it became irritating. So here I am.
Isn't "quantum" actually more appropriate than "quantized"? That's why we say "quantum physics" instead of "quantized physics".
Isn't "quantum" actually more appropriate than "quantized"? That's why we say "quantum physics" instead of "quantized physics".
Hmm not sure but I think "quantized" here meaning quantization in signal processing, not in physics (or I didn't get the joke? let me know if I should remove my comment)
Heh, I wasn't joking - I genuinely think that "quantum model" is correct. Yes, the process of converting continuous values into discrete values is "quantization". In a similar way, in quantum mechanics, bound states are quantized to discrete quantities [0]. I'm drawing the analogy that quantum mechanics is modelling nature, similar to how LLMs are modelling language and hence the same terminology should apply.
[0] - https://en.wikipedia.org/wiki/Quantum_mechanics
Hmm ok that make sense. Thanks for the explanation.
But mixing naming quantum/quantized maybe a bit confused, so I think this PR is not really redundant (just my opinion)
Interesting! While the ML community in general seems to use the word "quantized" (search "quantized" at: 1, 2, 3), I see the analogy @ggerganov is making. Thanks for the explanation!
Maybe one could add a short note somewhere in the two docs for the average user about your analogy for clarity. What do you think about that?
Isn't "quantum" actually more appropriate than "quantized"? That's why we say "quantum physics" instead of "quantized physics".
No, I disagree. In quantum physics things like discrete energy levels are inherent properties. Models that have been compressed from 16 bit floats to low-precision integers on the other hand have been actively changed from their original state. I personally don't like or use the term "quantum model" because it in my opinion just sounds like some marketing buzzword but I ultimately think which term is used is inconsequential; to me this is on the same level as whether or not the first letter of a code comment should be capitalized and use punctuation at the end of the line.
