1#include "arg.h"
2#include "common.h"
3#include "log.h"
4#include "ngram-cache.h"
5#include "llama.h"
6#include "ggml.h"
7
8#include <cstdint>
9#include <cstdio>
10#include <cinttypes>
11#include <fstream>
12#include <string>
13#include <vector>
14
15int main(int argc, char ** argv){
16 common_params params;
17
18 if (!common_params_parse(argc, argv, params, ex: LLAMA_EXAMPLE_LOOKUP)) {
19 return 1;
20 }
21
22 common_init();
23
24 const int n_draft = params.speculative.n_max;
25
26 // init llama.cpp
27 llama_backend_init();
28 llama_numa_init(numa: params.numa);
29
30 // load the model
31 common_init_result llama_init = common_init_from_params(params);
32
33 llama_context_ptr & ctx = llama_init.context;
34
35 // tokenize the prompt
36 std::vector<llama_token> inp;
37 inp = common_tokenize(ctx: ctx.get(), text: params.prompt, add_special: true, parse_special: true);
38
39 common_ngram_cache ngram_cache_context;
40 common_ngram_cache ngram_cache_dynamic;
41 common_ngram_cache ngram_cache_static;
42
43 int64_t t_draft_flat_us = 0;
44 int64_t t_draft_us = 0;
45
46 {
47 const int64_t t_start_draft_us = ggml_time_us();
48
49 if (!params.lookup_cache_static.empty()) {
50 try {
51 ngram_cache_static = common_ngram_cache_load(filename&: params.lookup_cache_static);
52 } catch (std::ifstream::failure const &) {
53 LOG_ERR("failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
54 exit(status: 1);
55 }
56 }
57
58 if (!params.lookup_cache_dynamic.empty()) {
59 try {
60 ngram_cache_dynamic = common_ngram_cache_load(filename&: params.lookup_cache_dynamic);
61 } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
62 }
63
64 t_draft_flat_us += ggml_time_us() - t_start_draft_us;
65 }
66
67 const int n_input = inp.size();
68 const int n_ctx = llama_n_ctx(ctx: ctx.get());
69
70 int n_drafted = 0;
71 int n_accept = 0;
72
73 const int64_t t_start_ms = ggml_time_ms();
74
75 // Iterate over input tokens in chunks of size n_ctx.
76 // Each chunk is treated as if a sequential generation but with pre-determined tokens to ensure reproducibility.
77 for (int i_start = 0; i_start + n_ctx < n_input; i_start += n_ctx) {
78 const std::vector<llama_token> inp_slice(inp.begin() + i_start, inp.begin() + i_start + n_ctx);
79 std::vector<llama_token> pseudo_output;
80 pseudo_output.push_back(x: inp_slice[0]);
81
82 while ((int) pseudo_output.size() < n_ctx) {
83 // Simulate drafting and decoding from draft:
84 std::vector<llama_token> draft;
85 draft.push_back(x: pseudo_output.back());
86
87 {
88 const int64_t t_start_draft_us = ggml_time_us();
89 common_ngram_cache_draft(inp&: pseudo_output, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, nc_context&: ngram_cache_context, nc_dynamic&: ngram_cache_dynamic, nc_static&: ngram_cache_static);
90 t_draft_us += ggml_time_us() - t_start_draft_us;
91 }
92
93 n_drafted += draft.size() - 1;
94
95 for (size_t j = 1; j < draft.size() && (int) pseudo_output.size() < n_ctx; ++j) {
96 const llama_token ground_truth = inp_slice[pseudo_output.size()];
97 const llama_token drafted = draft[j];
98
99 if (ground_truth != drafted) {
100 break;
101 }
102
103 ++n_accept;
104 pseudo_output.push_back(x: ground_truth);
105
106 {
107 const int64_t t_start_draft_us = ggml_time_us();
108 common_ngram_cache_update(ngram_cache&: ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp_data&: pseudo_output, nnew: 1, print_progress: false);
109 t_draft_us += ggml_time_us() - t_start_draft_us;
110 }
111 }
112
113 // After each simulated batch decoding simulate the sampling of a single token:
114 if ((int) pseudo_output.size() < n_ctx) {
115 pseudo_output.push_back(x: inp_slice[pseudo_output.size()]);
116 {
117 const int64_t t_start_draft_us = ggml_time_us();
118 common_ngram_cache_update(ngram_cache&: ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp_data&: pseudo_output, nnew: 1, print_progress: false);
119 t_draft_us += ggml_time_us() - t_start_draft_us;
120 }
121 }
122
123 draft.erase(position: draft.begin());
124
125 }
126 if (i_start > 0 && i_start / 100000 != (i_start - n_ctx) / 100000) {
127 const int64_t t_now_ms = ggml_time_ms();
128 const int64_t eta_ms = (n_input - i_start) * (t_now_ms - t_start_ms) / i_start;
129 const int64_t eta_min = eta_ms / (60*1000);
130 const int64_t eta_s = (eta_ms - 60*1000*eta_min) / 1000;
131
132 LOG_INF("lookup-stats: %d/%d done, ETA: %02" PRId64 ":%02" PRId64 "\n", i_start, n_input, eta_min, eta_s);
133 }
134
135 // After each chunk, update the dynamic ngram cache with the context ngram cache:
136 common_ngram_cache_merge(ngram_cache_target&: ngram_cache_dynamic, ngram_cache_add&: ngram_cache_context);
137 ngram_cache_context.clear();
138 }
139
140 LOG("\n");
141
142 LOG_INF("\n");
143 LOG_INF("n_draft = %d\n", n_draft);
144 LOG_INF("n_predict = %d\n", n_input - n_input % n_ctx);
145 LOG_INF("n_drafted = %d\n", n_drafted);
146 LOG_INF("t_draft_flat = %.2f ms\n", t_draft_flat_us*1e-3);
147 LOG_INF("t_draft = %.2f ms, %.2f us per token, %.2f tokens per second\n",
148 t_draft_us*1e-3, 1.0f*t_draft_us/n_drafted, n_drafted/(1e-6*t_draft_us));
149 LOG_INF("n_accept = %d\n", n_accept);
150 LOG_INF("accept = %.3f%%\n", 100.0f * n_accept / n_drafted);
151
152 llama_backend_free();
153
154 LOG("\n\n");
155
156 return 0;
157}
158