#ifndef SLEELA_AI_LLM_CAPACITY_HPP #define SLEELA_AI_LLM_CAPACITY_HPP #include "large_language_models_capacity.h" #include #include namespace sleela::ai { // A user-owned description of the model the caller intends to run, within the // supported 2 GB .. 16 GB weight envelope. SLeeLa ships no weights and selects // no model; this only records the user's choice and derives a reasoning budget. struct CapacityProfile { std::string model_id; // the caller's identifier for THEIR model std::string weight_format; // informational, e.g. "gguf-q4_k_m" std::uint64_t weight_bytes = 4ULL * 1024 * 1024 * 1024; // default STANDARD (4 GB) std::uint32_t context_tokens = 0; // 0 => derive from tier std::uint64_t ram_budget_bytes = 0; // 0 => unspecified bool accelerated = false; sllmc_tier tier() const { return sllmc_tier_for_bytes(weight_bytes); } std::string tier_name() const { return sllmc_tier_name(tier()); } // Advisory envelope check: true when weight_bytes is within 2..16 GB (or 0). bool within_envelope() const { return to_c_status() == SLLMC_OK; } sllmc_status to_c_status() const { sllmc_profile p = to_c(); return sllmc_profile_check(&p); } sllmc_profile to_c() const { sllmc_profile p; sllmc_profile_init(&p); p.model_id = model_id.empty() ? nullptr : model_id.c_str(); p.weight_format = weight_format.empty() ? nullptr : weight_format.c_str(); p.weight_bytes = weight_bytes; p.context_tokens = context_tokens; p.ram_budget_bytes = ram_budget_bytes; p.accelerated = accelerated ? 1 : 0; return p; } sllmc_budget budget() const { sllmc_profile p = to_c(); sllmc_budget b{}; sllmc_derive_budget(&p, &b); return b; } }; } // namespace sleela::ai #endif