21032293656

Committed 15 Jan 2026 12:21PM UTC coverage: 90.293% (+20.8%) from 69.536%

Build # 21032293656

Build Type

push

github

Committed by

grencez

Commit Message

update(llama.cpp): with EXAONE-MoE model support

Run Details

2000 of 2215 relevant lines covered (90.29%)

38.05 hits per line

Source File
Press 'n' to go to next uncovered line, 'b' for previous

91.67

/test/language/inference_test.cc

#include "src/chat/display.hh"
#include "src/chat/opt.hh"
#include "src/chat/trajectory.hh"
#include "src/language/inference.hh"
#include "src/language/vocabulary.hh"

#include <cassert>
#include <iostream>
#include <tuple>
#include <vector>

#include <fildesh/fildesh.h>

#include "llama.h"

using rendezllama::ChatDisplay;
using rendezllama::ChatOptions;
using rendezllama::ChatTrajectory;
using rendezllama::Inference;
using rendezllama::Vocabulary;

static
  void
noop_log_callback(enum ggml_log_level level, const char* text, void* user_data)
{
  (void) level;
  (void) text;
  (void) user_data;
}

static void test_antiprompt_suffix() {
  std::set<std::string> antiprompts;
  antiprompts.insert("User:");
  antiprompts.insert("\nUser:");

  // Case: No match
  assert(rendezllama::antiprompt_suffix("Hello World", antiprompts).empty());

  // Case: Exact match
  assert(rendezllama::antiprompt_suffix("User:", antiprompts) == "User:");

  // Case: Suffix match
  assert(rendezllama::antiprompt_suffix("Hello User:", antiprompts) == "User:");

  // Case: Longest match check (order in set matters)
  // "User:" is a suffix of "\nUser:" but "\nUser:" is longer.
  // In std::set, "\nUser:" (byte 10) comes before "User:" (byte 85).
  // The function iterates and returns the first match.
  // "\nUser:" should be checked first.
  assert(rendezllama::antiprompt_suffix("Hello\nUser:", antiprompts) == "\nUser:");

  // Case: Partial match should fail
  assert(rendezllama::antiprompt_suffix("User", antiprompts).empty());
}

static void inference_test(const std::string& model_filename) {
  llama_log_set(noop_log_callback, NULL);

  ChatOptions opt;
  opt.model_filename = model_filename;
  // Initialize infer_via with default Sampling to avoid assertion failure in Inference::reinitialize.
  opt.infer_via.emplace<rendezllama::inference::Sampling>();
  // Workaround for potential crash with default boundary prefix "☺".
  opt.substitution.boundary_prefix = ":";

  struct llama_model* model = nullptr;
  struct llama_context* ctx = nullptr;
  std::tie(model, ctx) = rendezllama::make_llama_context(opt);
  assert(model);
  assert(ctx);

  Vocabulary vocabulary(model, opt.substitution.boundary_prefix);
  Inference inference(vocabulary);
  ChatTrajectory chat_traj(vocabulary.bos_token_id());
  ChatDisplay chat_disp;
  // Use /dev/null for display to avoid spamming test output, we check tokens programmatically.
  chat_disp.out_ = open_FildeshOF("/dev/null");

  // Add a simple prompt to start generation.
  // "Once upon a time"
  chat_traj.tokenize_append("Once upon a time", vocabulary);
  chat_disp.show_new(chat_traj, vocabulary);

  using rendezllama::inference::AdjustViaKind;
  using rendezllama::inference::Sampling;
  using rendezllama::inference::Mirostat;

  std::vector<Sampling> samplings;
  // Temperature
  {
    Sampling s;
    s.adjust_thru.emplace_back(std::in_place_index<AdjustViaKind::temperature>, 0.8f);
    samplings.push_back(s);
  }
  // Top K
  {
    Sampling s;
    s.adjust_thru.emplace_back(std::in_place_index<AdjustViaKind::top_k>, 40u);
    samplings.push_back(s);
  }
  // Top P
  {
    Sampling s;
    s.adjust_thru.emplace_back(std::in_place_index<AdjustViaKind::top_p>, 0.9f);
    samplings.push_back(s);
  }
  // Min P
  {
    Sampling s;
    s.adjust_thru.emplace_back(std::in_place_index<AdjustViaKind::min_p>, 0.05f);
    samplings.push_back(s);
  }
   // Typical P
  {
    Sampling s;
    s.adjust_thru.emplace_back(std::in_place_index<AdjustViaKind::typical_p>, 0.9f);
    samplings.push_back(s);
  }
  // Mirostat V2
  {
    Sampling s;
    s.pick_via = Mirostat{2, 5.0f, 0.1f};
    samplings.push_back(s);
  }
  // Penalties
  {
    Sampling s;
    rendezllama::inference::PenalizeWith p;
    p.window_length = 5;
    p.repetition = 1.1f;
    s.adjust_thru.emplace_back(std::in_place_index<AdjustViaKind::penalize_with>, p);
    samplings.push_back(s);
  }
  // Dry
  {
    Sampling s;
    rendezllama::inference::Dry d;
    d.multiplier = 0.8f;
    d.base = 1.75f;
    d.allowed_length = 2;
    d.window_length = 0; // Default
    s.adjust_thru.emplace_back(std::in_place_index<AdjustViaKind::dry>, d);
    samplings.push_back(s);
  }
  // XTC
  {
    Sampling s;
    rendezllama::inference::Xtc x;
    x.threshold = 0.1f;
    x.probability = 0.5f;
    s.adjust_thru.emplace_back(std::in_place_index<AdjustViaKind::xtc>, x);
    samplings.push_back(s);
  }

  // Iterate through different sampling options.
  for (const auto& sampling : samplings) {
    opt.infer_via = sampling;
    if (!inference.commit_to_context(ctx, chat_disp, chat_traj, opt, model)) {
      break;
    }
    inference.sample_to_trajectory(chat_traj, ctx, false);
    chat_disp.show_new(chat_traj, vocabulary);
  }

  llama_free(ctx);
  llama_model_free(model);
}

int main(int argc, char** argv)
{
  rendezllama::GlobalScope rendezllama_global_scope;
  assert(argc == 2);
  test_antiprompt_suffix();
  inference_test(argv[1]);
  return 0;
}

1	#include "src/chat/display.hh"
2	#include "src/chat/opt.hh"
3	#include "src/chat/trajectory.hh"
4	#include "src/language/inference.hh"
5	#include "src/language/vocabulary.hh"
6
7	#include <cassert>
8	#include <iostream>
9	#include <tuple>
10	#include <vector>
11
12	#include <fildesh/fildesh.h>
13
14	#include "llama.h"
15
16	using rendezllama::ChatDisplay;
17	using rendezllama::ChatOptions;
18	using rendezllama::ChatTrajectory;
19	using rendezllama::Inference;
20	using rendezllama::Vocabulary;
21
22	static
23	void
24	noop_log_callback(enum ggml_log_level level, const char* text, void* user_data)	180✔
25	{
26	(void) level;	180✔
27	(void) text;	180✔
28	(void) user_data;	180✔
29	}	180✔
30
31	static void test_antiprompt_suffix() {	1✔
32	std::set<std::string> antiprompts;	1✔
33	antiprompts.insert("User:");	2✔
34	antiprompts.insert("\nUser:");	2✔
35
36	// Case: No match
37	assert(rendezllama::antiprompt_suffix("Hello World", antiprompts).empty());	1✔
38
39	// Case: Exact match
40	assert(rendezllama::antiprompt_suffix("User:", antiprompts) == "User:");	1✔
41
42	// Case: Suffix match
43	assert(rendezllama::antiprompt_suffix("Hello User:", antiprompts) == "User:");	1✔
44
45	// Case: Longest match check (order in set matters)
46	// "User:" is a suffix of "\nUser:" but "\nUser:" is longer.
47	// In std::set, "\nUser:" (byte 10) comes before "User:" (byte 85).
48	// The function iterates and returns the first match.
49	// "\nUser:" should be checked first.
50	assert(rendezllama::antiprompt_suffix("Hello\nUser:", antiprompts) == "\nUser:");	1✔
51
52	// Case: Partial match should fail
53	assert(rendezllama::antiprompt_suffix("User", antiprompts).empty());	1✔
54	}	1✔
55
56	static void inference_test(const std::string& model_filename) {	1✔
57	llama_log_set(noop_log_callback, NULL);	1✔
58
59	ChatOptions opt;	1✔
60	opt.model_filename = model_filename;	1✔
61	// Initialize infer_via with default Sampling to avoid assertion failure in Inference::reinitialize.
62	opt.infer_via.emplace<rendezllama::inference::Sampling>();	1✔
63	// Workaround for potential crash with default boundary prefix "☺".
64	opt.substitution.boundary_prefix = ":";	1✔
65
66	struct llama_model* model = nullptr;	1✔
67	struct llama_context* ctx = nullptr;	1✔
68	std::tie(model, ctx) = rendezllama::make_llama_context(opt);	1✔
69	assert(model);	1✔
70	assert(ctx);	1✔
71
72	Vocabulary vocabulary(model, opt.substitution.boundary_prefix);	1✔
73	Inference inference(vocabulary);	1✔
74	ChatTrajectory chat_traj(vocabulary.bos_token_id());	1✔
75	ChatDisplay chat_disp;	1✔
76	// Use /dev/null for display to avoid spamming test output, we check tokens programmatically.
77	chat_disp.out_ = open_FildeshOF("/dev/null");	1✔
78
79	// Add a simple prompt to start generation.
80	// "Once upon a time"
81	chat_traj.tokenize_append("Once upon a time", vocabulary);	1✔
82	chat_disp.show_new(chat_traj, vocabulary);	1✔
83
84	using rendezllama::inference::AdjustViaKind;	1✔
85	using rendezllama::inference::Sampling;	1✔
86	using rendezllama::inference::Mirostat;	1✔
87
88	std::vector<Sampling> samplings;	1✔
89	// Temperature
90	{	1✔
91	Sampling s;	1✔
92	s.adjust_thru.emplace_back(std::in_place_index<AdjustViaKind::temperature>, 0.8f);	1✔
93	samplings.push_back(s);	1✔
94	}	×
95	// Top K
96	{	1✔
97	Sampling s;	1✔
98	s.adjust_thru.emplace_back(std::in_place_index<AdjustViaKind::top_k>, 40u);	1✔
99	samplings.push_back(s);	1✔
100	}	×
101	// Top P
102	{	1✔
103	Sampling s;	1✔
104	s.adjust_thru.emplace_back(std::in_place_index<AdjustViaKind::top_p>, 0.9f);	1✔
105	samplings.push_back(s);	1✔
106	}	×
107	// Min P
108	{	1✔
109	Sampling s;	1✔
110	s.adjust_thru.emplace_back(std::in_place_index<AdjustViaKind::min_p>, 0.05f);	1✔
111	samplings.push_back(s);	1✔
112	}	×
113	// Typical P
114	{	1✔
115	Sampling s;	1✔
116	s.adjust_thru.emplace_back(std::in_place_index<AdjustViaKind::typical_p>, 0.9f);	1✔
117	samplings.push_back(s);	1✔
118	}	×
119	// Mirostat V2
120	{	1✔
121	Sampling s;	1✔
122	s.pick_via = Mirostat{2, 5.0f, 0.1f};	1✔
123	samplings.push_back(s);	1✔
124	}	×
125	// Penalties
126	{	1✔
127	Sampling s;	1✔
128	rendezllama::inference::PenalizeWith p;	1✔
129	p.window_length = 5;	1✔
130	p.repetition = 1.1f;	1✔
131	s.adjust_thru.emplace_back(std::in_place_index<AdjustViaKind::penalize_with>, p);	1✔
132	samplings.push_back(s);	1✔
133	}	×
134	// Dry
135	{	1✔
136	Sampling s;	1✔
137	rendezllama::inference::Dry d;	1✔
138	d.multiplier = 0.8f;	1✔
139	d.base = 1.75f;	1✔
140	d.allowed_length = 2;	1✔
141	d.window_length = 0; // Default	1✔
142	s.adjust_thru.emplace_back(std::in_place_index<AdjustViaKind::dry>, d);	1✔
143	samplings.push_back(s);	1✔
144	}	×
145	// XTC
146	{	1✔
147	Sampling s;	1✔
148	rendezllama::inference::Xtc x;	1✔
149	x.threshold = 0.1f;	1✔
150	x.probability = 0.5f;	1✔
151	s.adjust_thru.emplace_back(std::in_place_index<AdjustViaKind::xtc>, x);	1✔
152	samplings.push_back(s);	1✔
153	}	×
154
155	// Iterate through different sampling options.
156	for (const auto& sampling : samplings) {	10✔
157	opt.infer_via = sampling;	9✔
158	if (!inference.commit_to_context(ctx, chat_disp, chat_traj, opt, model)) {	9✔
159	break;
160	}
161	inference.sample_to_trajectory(chat_traj, ctx, false);	9✔
162	chat_disp.show_new(chat_traj, vocabulary);	9✔
163	}
164
165	llama_free(ctx);	1✔
166	llama_model_free(model);	1✔
167	}	1✔
168
169	int main(int argc, char** argv)	1✔
170	{
171	rendezllama::GlobalScope rendezllama_global_scope;	1✔
172	assert(argc == 2);	1✔
173	test_antiprompt_suffix();	1✔
174	inference_test(argv[1]);	2✔
175	return 0;	1✔
176	}	1✔

rendezqueue / rendezllama / 21032293656

Source File Press 'n' to go to next uncovered line, 'b' for previous

Source File
Press 'n' to go to next uncovered line, 'b' for previous