Function draw_from_weights

server/src/common/sampler.cpp:66–76

Draws a token from `cand`, whose .first fields are proportional probabilities (need not already sum to 1 or be sorted). `r_uniform` is a pre-drawn uniform in [0,1) supplied by the caller (drawn once per sample_logits call) so every path — GPU, GPU-assisted top_p, or CPU — consumes the same single RNG value.

// sample_logits call) so every path — GPU, GPU-assisted top_p, or CPU —
// consumes the same single RNG value.
int draw_from_weights(const std::vector<std::pair<float, int>> & cand, double r_uniform) {
    double Z = 0.0;
    for (auto & c : cand) Z += c.first;
    const double r = r_uniform * Z;
    double acc = 0.0;
    for (auto & c : cand) {
        acc += c.first;
        if (r <= acc) return c.second;
    }
    return cand.back().second;
}

#ifdef DFLASH27B_HAVE_GPU_SAMPLER
// Given probabilities the GPU already computed (penalties + softmax(temp)