Patch rng seed

[rdentato]

Rather than exit, let's treat rng_seed = 0 as an indication for a default value (time(NULL)).
It seems more reasonable to me ...

[xefoci7612]

@rdentato this is not deterministic

If we run 2 times

./run model.bin -s 0
./run model.bin -s 0

We'd like to see the same output (this is the reason we use the -s option)

IMHO a correct patch that preserves deterministic behavior can be

else if (argv[i][1] == 's') { rng_seed = atoi(argv[i + 1]) ^ RND_0; }

where for instance

#define RND_0 0x2545F4914F6CDD1Dull

and we don't need to special case the 0 anymore

[rdentato]

The issue I see with

rng_seed = atoi(argv[i + 1]) ^ RND_0;

is that if RND_0 is 0x2545F4914F6CDD1Dull then

./run model.bin -s 2685821657736338717

will set rng_seed to 0.

Yes, is very unlikely, but still possible.

The reason behind my proposal is that you can imagine run called by a script

   ./run model.bin -s ${LLM_SEED:-0}

If LLM_SEED is not set, the default value is given (and it's a value that changes all the time).
I usually prefer to have a value that signifies default and 0 as a seed seems a very good candidate to me.

[twobob]

If LLM_SEED is not set, the default value is give (and it's a value that changes all the time). I usually prefer to have a value that signifies default and 0 as a seed seems a very good candidate to me.

Exact reason s=0 was immediately tested here, following the change. Concur, fwiw, with this mindset.
That said, determinism was also our expectation so, unsure.