model_registry.hpp

Go to the documentation of this file.

#pragma once
 
// SPDX-License-Identifier: LicenseRef-FSL-1.1-Apache-2.0
// Copyright 2026 Lloyal Labs
 
 
#include "common.hpp"
#include <functional>
#include <llama/llama.h>
#include <memory>
#include <mutex>
#include <string>
#include <unordered_map>
 
namespace lloyal {
 
struct ModelKey {
  std::string canonPath; 
  int n_gpu_layers;      
  int load_mode;         
 
  bool operator==(const ModelKey &o) const {
    return n_gpu_layers == o.n_gpu_layers && load_mode == o.load_mode &&
           canonPath == o.canonPath;
  }
};
 
// Compiler-specific sanitizer suppression (GCC/Clang only; no-op on MSVC)
#if defined(__GNUC__) || defined(__clang__)
  #define LLOYAL_NO_SANITIZE_OVERFLOW __attribute__((no_sanitize("unsigned-integer-overflow")))
#else
  #define LLOYAL_NO_SANITIZE_OVERFLOW
#endif
 
struct ModelKeyHash {
  LLOYAL_NO_SANITIZE_OVERFLOW
  size_t operator()(const ModelKey &k) const {
    std::hash<std::string> Hs;
    std::hash<int> Hi;
    return Hs(k.canonPath) ^
           (Hi(k.n_gpu_layers) + 0x9e3779b9 + (Hi(k.load_mode) << 6));
  }
};
 
#undef LLOYAL_NO_SANITIZE_OVERFLOW
 
class ModelRegistry {
public:
  static std::shared_ptr<llama_model> acquire(const std::string &fsPath,
                                              const llama_model_params &params);
 
private:
  inline static std::mutex mu_;
 
  inline static std::unordered_map<ModelKey, std::weak_ptr<llama_model>,
                                   ModelKeyHash>
      cache_;
 
  static ModelKey makeKey(const std::string &fsPath,
                          const llama_model_params &params);
};
 
} // namespace lloyal
 
namespace lloyal::detail {
 
inline void freeModel(llama_model *model) {
  LLOYAL_LOG_DEBUG(
      "[ModelRegistry] Freeing model: ptr=%p (last reference released)",
      (void *)model);
  llama_model_free(model);
  LLOYAL_LOG_DEBUG("[ModelRegistry] Model freed: ptr=%p", (void *)model);
}
 
} // namespace lloyal::detail
 
namespace lloyal {
 
// ===== IMPLEMENTATION =====
 
// Normalize path to ensure "file:///path" and "/path" map to same key
inline ModelKey ModelRegistry::makeKey(const std::string &fsPath,
                                       const llama_model_params &params) {
  // Inline path normalization (removes file:// prefix if present)
  std::string canonPath = fsPath;
  const std::string filePrefix = "file://";
  if (canonPath.substr(0, filePrefix.length()) == filePrefix) {
    canonPath = canonPath.substr(filePrefix.length());
  }
 
  return {canonPath, params.n_gpu_layers, static_cast<int>(params.load_mode)};
}
 
// Acquire model from cache or load new
// 1. Check cache (thread-safe)
// 2. Return existing if found (cache hit)
// 3. Load new if expired (cache miss)
// 4. Store as weak_ptr, return shared_ptr
inline std::shared_ptr<llama_model>
ModelRegistry::acquire(const std::string &fsPath,
                       const llama_model_params &params) {
  ModelKey key = makeKey(fsPath, params);
 
  LLOYAL_LOG_DEBUG("[ModelRegistry] Acquiring model: path='%s', "
                   "n_gpu_layers=%d, load_mode=%d",
                   key.canonPath.c_str(), key.n_gpu_layers, key.load_mode);
 
  std::lock_guard<std::mutex> lock(mu_);
 
  auto cacheEntry = cache_.find(key);
  if (cacheEntry != cache_.end()) {
    // Try to upgrade weak_ptr to shared_ptr
    if (auto existingModel = cacheEntry->second.lock()) {
      long refCount = existingModel.use_count();
      LLOYAL_LOG_DEBUG(
          "[ModelRegistry] Cache HIT - Reusing model: ptr=%p, refcount=%ld",
          (void *)existingModel.get(), refCount);
      return existingModel;
    } else {
      LLOYAL_LOG_DEBUG("[ModelRegistry] Cache entry expired (model was freed), "
                       "removing stale entry");
      cache_.erase(cacheEntry);
    }
  }
 
  LLOYAL_LOG_DEBUG("[ModelRegistry] Cache MISS - Loading NEW model from disk");
  LLOYAL_LOG_DEBUG("[ModelRegistry]   Path: %s", key.canonPath.c_str());
  LLOYAL_LOG_DEBUG("[ModelRegistry]   GPU layers: %d", key.n_gpu_layers);
  LLOYAL_LOG_DEBUG("[ModelRegistry]   Load mode: %d", key.load_mode);
 
  llama_model *rawModel =
      llama_model_load_from_file(key.canonPath.c_str(), params);
 
  if (!rawModel) {
    // Let caller handle error (will throw structured error)
    LLOYAL_LOG_DEBUG(
        "[ModelRegistry] ERROR: llama_model_load_from_file returned NULL");
    return nullptr;
  }
 
  size_t modelSize = llama_model_size(rawModel);
  LLOYAL_LOG_DEBUG("[ModelRegistry] Model loaded:");
  LLOYAL_LOG_DEBUG("[ModelRegistry]   Pointer: %p", (void *)rawModel);
  LLOYAL_LOG_DEBUG("[ModelRegistry]   Size: %zu bytes (%.2f MB)", modelSize,
                   modelSize / (1024.0 * 1024.0));
 
  auto sharedModel = std::shared_ptr<llama_model>(rawModel, detail::freeModel);
 
  // Store as weak_ptr (allows automatic cleanup when all contexts release the
  // model)
  cache_[key] = sharedModel;
  LLOYAL_LOG_DEBUG("[ModelRegistry] Model cached as weak_ptr, returning "
                   "shared_ptr (refcount=1)");
 
  return sharedModel;
}
 
} // namespace lloyal