1#pragma once
2
3#include "llama.h"
4
5#include "ggml-cpp.h"
6
7#include <string>
8#include <unordered_map>
9#include <vector>
10
11// TODO: pimpl
12
13//
14// llama_adapter_cvec
15//
16
17struct llama_adapter_cvec {
18 ggml_tensor * tensor_for(int il) const;
19
20 ggml_tensor * apply_to(ggml_context * ctx, ggml_tensor * cur, int il) const;
21
22 bool apply(
23 const llama_model & model,
24 const float * data,
25 size_t len,
26 int32_t n_embd,
27 int32_t il_start,
28 int32_t il_end);
29
30private:
31 bool init(const llama_model & model);
32
33 int32_t layer_start = -1;
34 int32_t layer_end = -1;
35
36 std::vector<ggml_context_ptr> ctxs;
37 std::vector<ggml_backend_buffer_ptr> bufs;
38
39 std::vector<ggml_tensor *> tensors; // per layer
40};
41
42//
43// llama_adapter_lora
44//
45
46struct llama_adapter_lora_weight {
47 ggml_tensor * a = nullptr;
48 ggml_tensor * b = nullptr;
49
50 // get actual scale based on rank and alpha
51 float get_scale(float alpha, float adapter_scale) const {
52 const float rank = (float) b->ne[0];
53 const float scale = alpha ? adapter_scale * alpha / rank : adapter_scale;
54 return scale;
55 }
56
57 llama_adapter_lora_weight() = default;
58 llama_adapter_lora_weight(ggml_tensor * a, ggml_tensor * b) : a(a), b(b) {}
59};
60
61struct llama_adapter_lora {
62 // map tensor name to lora_a_b
63 std::unordered_map<std::string, llama_adapter_lora_weight> ab_map;
64
65 std::vector<ggml_context_ptr> ctxs;
66 std::vector<ggml_backend_buffer_ptr> bufs;
67
68 float alpha;
69
70 // gguf metadata
71 std::unordered_map<std::string, std::string> gguf_kv;
72
73 // activated lora (aLoRA)
74 std::vector<llama_token> alora_invocation_tokens;
75
76 llama_adapter_lora() = default;
77 ~llama_adapter_lora() = default;
78
79 llama_adapter_lora_weight * get_weight(ggml_tensor * w);
80};
81
82using llama_adapter_loras = std::unordered_map<llama_adapter_lora *, float>;
83