Added README.md for main with examples and explanations

[DannyDaemonic]

I've been slowly working on this. It contains a couple of examples and a longer explanation of the options a user might use.

[DannyDaemonic]

Just updated it to fix a link to the primary README.

If you want to see it formatted, you can follow this link.

Edit: And again, adding a link back to the front llama.cpp page in case someone finds their way to the main example from outside of github.
Edit 2: Expanded the Generation Flags section.

[sw]

@mgroeber9110 is doing the same in #1131...

[DannyDaemonic]

Thanks. That's the problem with these large write ups. We probably both started days ago...

Edit: I'm biased for sure, but mine feels more complete. If you want to commit @mgroeber9110's README.md first I can try to edit my information into his but they are structured different. I feel like I would probably just end up replacing everything. :( Still, it would get his --interactive_start fixes in.

[DannyDaemonic]

I was trying to compare all the options to see if we could merge the two. I did find his has instruction for --seed and --ignore-eos, which mine doesn't have and should certainly be in there. But our READMEs are structured so differently, I don't know if we could really merge them.

Mine is missing:
-h, --help
-s SEED, --seed SEED
--ignore-eos
-b N, --batch_size N

I did have --batch_size in there before but I accidentally cut it out of the final draft when I was restructuring things. I'm going to edit it back in again. I don't have any issue adding in the other missing options but I'll wait until we have a better idea of what we're doing.

[mgroeber9110]

@DannyDaemonic Sorry for the duplicate work - I should have probably made a note on the ticket that I intend to work on it, but your text definitely feels much more comprehensive, while I only took a quick stab at giving people a starting point.

From my point of view, it probably makes more sense to merge yours first, and then I can perhaps transfer a few tidbits of information from mine (such as what exactly --instruct injects, and perhaps a sentence or two on stopping conditions).

Just wondering: was it intentional to not mention --ignore-eos? I found this quite powerful, but apparently it does not work everywhere (e.g. #990).

[DannyDaemonic]

Yeah, I feel bad that time was wasted either way.

The --ignore-eos was just an oversight. It really should be in there. I can make a stub for it in ## Generation Flags and you can fill it in?

I intended to put all the options in there. Even --help is worth mentioning as it shows the latest options and default values, which can and do change frequently.

[DannyDaemonic]

I added the missing options with the exception of --ignore-eos since @mgroeber9110 has more experience with that than me. I was going to put a stub in there but it might not even be best under ## Generation Flags. I was thinking maybe it belongs in ## Interaction if that's how it's most commonly used? Either way, I could leave that up to @mgroeber9110. Even if it does turn out to fit better in the generative section, it could certainly be mentioned in the instruct section.

On a bit of a tangent here, but I have had problems with Alpaca ending a response too soon, sometimes immediately, and never tried --ignore-eos to see if that helps. It does have a tendency to go a bit random at the end so I assumed it might make that worse. I suspect the real issue might be due to some early Alpaca models being trained with a context of 512 and I like to run everything with 2048.

[jon-chuang]

I'll add physical CPU cores, and the number of performance cores in a chipset with efficiency/performance (E/P) cores in PR #934.

[DannyDaemonic]

I'm watching that pull request. My plan was to put another pull request through removing it from the ## common section as soon as this is resolved.

[jon-chuang]

I still think we should leave it in; even with a warning in the cmdline the user might miss it; it's good to have multiple paths to getting the information.