src/execution_plan.cpp

#include "ling3/execution_plan.h"

namespace ling3 {

ExecutionPlan BuildExecutionPlan(std::size_t max_context) {
    ExecutionPlan plan;
    std::size_t kda_layers = 0;
    std::size_t mla_layers = 0;
    for (std::size_t layer = 0; layer < plan.layers.size(); ++layer) {
        const bool mla = (layer + 1) % 4 == 0;
        plan.layers[layer] = {
            static_cast<std::uint8_t>(layer),
            mla ? AttentionKind::kMla : AttentionKind::kKda,
            layer == 0 ? FfnKind::kDense : FfnKind::kMoe,
        };
        kda_layers += !mla;
        mla_layers += mla;
    }

    constexpr std::size_t routed_expert_contexts = 128 * 2;
    constexpr std::size_t shared_expert_contexts = 2;
    constexpr std::size_t kda_projection_contexts = 7;
    constexpr std::size_t mla_projection_contexts = 6;
    plan.decode_matmul_contexts =
        23 * (routed_expert_contexts + shared_expert_contexts) +
        kda_layers * kda_projection_contexts + mla_layers * mla_projection_contexts + 3;
    plan.rknn_island_contexts = kda_layers * 3 + 23 * 3 + 24;
    plan.kda_state_bytes_fp16 = kda_layers * 16 * 128 * 128 * sizeof(std::uint16_t);
    plan.mla_cache_bytes_fp16 = mla_layers * max_context * (512 + 64) * sizeof(std::uint16_t);
    plan.activation_arena_bytes =
        2 * 1536 + 3 * (1024 * sizeof(std::int32_t) + 1536 * sizeof(std::int32_t)) + 16 * 128 * 4;
    return plan;
}

} // namespace ling3