Method sample

src/common/modules/sampler.cpp:350–393

\brief Sample the token \param x the input buffer \return the sampled token

/// \param x the input buffer
/// \return the sampled token
int Sampler::sample(buffer<bf16>& x) {
    // PRNG is seeded once at construction, no per-call seeding needed

    // COPY FROM `x` → `this->logits[]`
    #if USEAVX2
    const int simd_width = 8;
    int i = 0;
    for (; i <= in_features - simd_width; i += simd_width) {
        __m128i bf16_vals_x = _mm_loadu_si128(
            (__m128i*)&x[i]
        );
        __m256 fp32_vals_x = bf16o_fp32(bf16_vals_x);
        _mm256_storeu_ps(&this->logits[i], fp32_vals_x);
    }
    for (; i < in_features; i++) {
        this->logits[i] = x[i];
    }
    #else
    for (int i = 0; i < in_features; i++) {
        this->logits[i] = x[i];
    }
    #endif

    sampler_penalty_apply_sparse();
    sampler_topk_apply(this->top_k);
    if (this->use_optimized_sampling) {
        softmax_with_topp_minp(this->top_p, this->min_p);
        sampler_temp_apply(this->temperature);
        softmax_inplace();
    } else {
        // Legacy behavior is kept as the default for output compatibility.
        softmax_inplace();
        sampler_topp_apply(this->top_p);
        softmax_inplace();
        sampler_minp_apply(this->min_p);
        sampler_temp_apply(this->temperature);
        softmax_inplace();
    }

    int sampled_index = sample_from_probs();
    ring_buffer_update_sparse(sampled_index);

    return sampled_index;
}