fix adaptive p sampler rewinding too far back

[dungquixote42]

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High

#1287 does not account for partial rewinds. Fixing it with this PR.

[ikawrakow]

Why did you remove the check for ctx->adapt_p_ctx being null?

[dungquixote42]

It is now in llama_review_adaptive_p_impl(). I think of common_sampler_review() as a place to list stateful samplers that needs reviewing, so the check felt a bit too preemptive. I can put the check back in here.

[ikawrakow]

What if weighted_sum or total_weight is empty (i.e., size() returns 0)?

My concept is that erase is kind of slow. I would write something like this (assuming weighted_sum and total_weight have the same size):

if (weighted_sum.size() > 1) {
    weighted_sum.front() = weighted_sum.back();
    total_weight.front() = total_weight.beck();
    weighted_sum.resize(1);
    total_weight.resize(1);
}

[dungquixote42]

What if weighted_sum or total_weight is empty (i.e., size() returns 0)?

That really should not happen, but I added a check.

My concept is that erase is kind of slow. I would write something like this (assuming weighted_sum and total_weight have the same size):

if (weighted_sum.size() > 1) {
    weighted_sum.front() = weighted_sum.back();
    total_weight.front() = total_weight.beck();
    weighted_sum.resize(1);
    total_weight.resize(1);
}

Neat. Fixed.

[dungquixote42]

weighted_sum and total_weight is combined into a vector of pairs in the most recent commit. I need to test it a bit more, but I think it is an improvement overall. If it is acceptable, I will remove the old code that has been commented out.