Run 35B MoE models in ~2.5 GB of RAM. Weights stream from SSD; a trained router predicts loads a step ahead. Mac, iPhone, Android, Windows supported
Python
// (layer, expert) → contiguous row slices inside fused tensors. 35b: shape[0]=E;
// 8b: E folded into the row dim. Nine regions: {gate,up,down}_proj × {weight,scales,biases}.
#pragma once
#include <cstdint>
#include <map>
#include <string>
#include <vector>
#include "safio.hpp"
namespace e0n {
struct Region {
uint64_t off, len;
std::string shard;
std::string name; // "gate_proj.weight" etc., diagnostics only
};
class ExpertMap {
public:
static bool build(const ModelFiles& files, ExpertMap& out, std::string& err);
int num_layers() const { return num_layers_; }
int64_t num_experts() const { return num_experts_; }
std::vector<Region> regions(int layer, int64_t expert) const;
uint64_t expert_bytes(int layer, int64_t expert) const;
const std::string& layer_prefix(int layer) const;
private:
struct Slice {
uint64_t base, stride;
std::string shard, name;
};
int num_layers_ = 0;
int64_t num_experts_ = 0;
std::vector<std::string> prefixes_;
std::vector<std::map<std::string, Slice>> layers_;
};
} // namespace e0n