Function sample_with_cfg

src/models/heartmula/generator.cpp:106–148

}

std::vector<int32_t> sample_with_cfg(
    const std::vector<float> & logits,
    int64_t batch,
    int64_t vocab_size,
    float cfg_scale,
    int64_t topk,
    float temperature,
    uint64_t seed,
    uint64_t & sample_call_index,
    const TorchCudaSamplingPolicy & policy,
    TopKSamplerScratch & scratch) {
    if (batch <= 0 || vocab_size <= 0 || static_cast<int64_t>(logits.size()) != batch * vocab_size) {
        throw std::runtime_error("HeartMuLa sampler logits shape mismatch");
    }
    const bool use_cfg = cfg_scale > 1.0F && batch > 1 && batch % 2 == 0;
    const int64_t rows = use_cfg ? batch / 2 : batch;
    std::vector<int32_t> out(static_cast<size_t>(batch), 0);
    std::vector<float> guided(static_cast<size_t>(vocab_size));
    for (int64_t row = 0; row < rows; ++row) {
        const float * source = logits.data() + static_cast<size_t>(row * vocab_size);
        if (use_cfg) {
            const float * uncond = logits.data() + static_cast<size_t>((row + rows) * vocab_size);
            for (int64_t v = 0; v < vocab_size; ++v) {
                guided[static_cast<size_t>(v)] = uncond[v] + (source[v] - uncond[v]) * cfg_scale;
            }
            source = guided.data();
        }
        const int32_t token = sample_topk_row(
            source,
            vocab_size,
            topk,
            temperature,
            seed,
            sample_call_index++,
            policy,
            scratch);
        out[static_cast<size_t>(row)] = token;
        if (use_cfg) {
            out[static_cast<size_t>(row + rows)] = token;
        }
    }
    return out;
}

HeartMuLaFrameEmbeddingInputs prompt_embedding_inputs(
    const HeartMuLaPromptEncoding & encoding,