Skip to content

Commit a758711

Browse files
njbrakenjbrake
committed
patches
1 parent 5a902a6 commit a758711

15 files changed

Lines changed: 1388 additions & 18192 deletions

‎whisper.cpp.patches/006-main-cpp-llamafile-integration.patch‎

Lines changed: 248 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,248 @@
1+
--- examples/main/main.cpp 2025-10-31 22:21:46
2+
+++ ../whisper.cpp.patches/main.cpp 2025-10-31 16:01:53
3+
@@ -1,4 +1,11 @@
4+
+// -*- mode:c++;indent-tabs-mode:nil;c-basic-offset:4;tab-width:8;coding:utf-8 -*-
5+
+// vi: set et ft=cpp ts=4 sts=4 sw=4 fenc=utf-8 :vi
6+
+
7+
+#include "llamafile/version.h"
8+
+#include "llamafile/llamafile.h"
9+
+#include "llama.cpp/cores.h"
10+
#include "common.h"
11+
+#include "slurp.h"
12+
13+
#include "whisper.h"
14+
#include "grammar-parser.h"
15+
@@ -11,24 +18,19 @@
16+
#include <thread>
17+
#include <vector>
18+
#include <cstring>
19+
+#include <cosmo.h>
20+
21+
+#include "llamafile/llamafile.h"
22+
+#include "llamafile/debug.h"
23+
+#include "llama.cpp/string.h"
24+
+
25+
#if defined(_MSC_VER)
26+
#pragma warning(disable: 4244 4267) // possible loss of data
27+
#endif
28+
29+
-// helper function to replace substrings
30+
-static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
31+
- for (size_t pos = 0; ; pos += replace.length()) {
32+
- pos = s.find(search, pos);
33+
- if (pos == std::string::npos) break;
34+
- s.erase(pos, search.length());
35+
- s.insert(pos, replace);
36+
- }
37+
-}
38+
-
39+
// command-line parameters
40+
struct whisper_params {
41+
- int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
42+
+ int32_t n_threads = std::min(16, cpu_get_num_math());
43+
int32_t n_processors = 1;
44+
int32_t offset_t_ms = 0;
45+
int32_t offset_n = 0;
46+
@@ -68,7 +70,6 @@
47+
bool print_progress = false;
48+
bool no_timestamps = false;
49+
bool log_score = false;
50+
- bool use_gpu = true;
51+
bool flash_attn = false;
52+
53+
std::string language = "en";
54+
@@ -91,7 +92,7 @@
55+
std::vector<std::string> fname_inp = {};
56+
std::vector<std::string> fname_out = {};
57+
58+
- grammar_parser::parse_state grammar_parsed;
59+
+ whisper_grammar_parser::parse_state grammar_parsed;
60+
};
61+
62+
static void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
63+
@@ -118,6 +119,40 @@
64+
continue;
65+
}
66+
67+
+ if (arg == "--log-disable") {
68+
+ FLAG_log_disable = true;
69+
+ } else if (arg == "--cli") {
70+
+ } else if (arg == "--server") {
71+
+ } else if (arg == "--fast") {
72+
+ FLAG_fast = true;
73+
+ } else if (arg == "--precise") {
74+
+ FLAG_precise = true;
75+
+ } else if (arg == "--trace") {
76+
+ FLAG_trace = true;
77+
+ } else if (arg == "--trap") {
78+
+ FLAG_trap = true;
79+
+ FLAG_unsecure = true; // for better backtraces
80+
+ llamafile_trapping_enabled(+1);
81+
+ } else if (arg == "--unsecure") {
82+
+ FLAG_unsecure = true;
83+
+ } else if (arg == "--nocompile") {
84+
+ FLAG_nocompile = true;
85+
+ } else if (arg == "--recompile") {
86+
+ FLAG_recompile = true;
87+
+ } else if (arg == "--tinyblas") {
88+
+ FLAG_tinyblas = true; // undocumented
89+
+ } else if (arg == "--gpu") {
90+
+ if (++i >= argc) {
91+
+ fprintf(stderr, "error: missing --gpu flag value\n");
92+
+ exit(1);
93+
+ }
94+
+ FLAG_gpu = llamafile_gpu_parse(argv[i]);
95+
+ if (FLAG_gpu == LLAMAFILE_GPU_ERROR) {
96+
+ fprintf(stderr, "error: invalid --gpu flag value: %s\n", argv[i]);
97+
+ exit(1);
98+
+ }
99+
+ } else
100+
+
101+
if (arg == "-h" || arg == "--help") {
102+
whisper_print_usage(argc, argv, params);
103+
exit(0);
104+
@@ -153,7 +188,7 @@
105+
else if (arg == "-oj" || arg == "--output-json") { params.output_jsn = true; }
106+
else if (arg == "-ojf" || arg == "--output-json-full"){ params.output_jsn_full = params.output_jsn = true; }
107+
else if (arg == "-of" || arg == "--output-file") { params.fname_out.emplace_back(argv[++i]); }
108+
- else if (arg == "-np" || arg == "--no-prints") { params.no_prints = true; }
109+
+ else if (arg == "-np" || arg == "--no-prints") { params.no_prints = true; FLAG_log_disable = true; }
110+
else if (arg == "-ps" || arg == "--print-special") { params.print_special = true; }
111+
else if (arg == "-pc" || arg == "--print-colors") { params.print_colors = true; }
112+
else if (arg == "-pp" || arg == "--print-progress") { params.print_progress = true; }
113+
@@ -166,18 +201,20 @@
114+
else if (arg == "-oved" || arg == "--ov-e-device") { params.openvino_encode_device = argv[++i]; }
115+
else if (arg == "-dtw" || arg == "--dtw") { params.dtw = argv[++i]; }
116+
else if (arg == "-ls" || arg == "--log-score") { params.log_score = true; }
117+
- else if (arg == "-ng" || arg == "--no-gpu") { params.use_gpu = false; }
118+
+ else if (arg == "-ng" || arg == "--no-gpu") { FLAG_gpu = LLAMAFILE_GPU_DISABLE; }
119+
else if (arg == "-fa" || arg == "--flash-attn") { params.flash_attn = true; }
120+
else if ( arg == "--suppress-regex") { params.suppress_regex = argv[++i]; }
121+
else if ( arg == "--grammar") { params.grammar = argv[++i]; }
122+
else if ( arg == "--grammar-rule") { params.grammar_rule = argv[++i]; }
123+
else if ( arg == "--grammar-penalty") { params.grammar_penalty = std::stof(argv[++i]); }
124+
else {
125+
- fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
126+
- whisper_print_usage(argc, argv, params);
127+
+ fprintf(stderr, "error: unknown argument: %s\n", arg.c_str()); // [jart]
128+
+ // whisper_print_usage(argc, argv, params); // [jart]
129+
exit(0);
130+
}
131+
}
132+
+
133+
+ FLAGS_READY = true;
134+
135+
return true;
136+
}
137+
@@ -232,7 +269,7 @@
138+
fprintf(stderr, " -oved D, --ov-e-device DNAME [%-7s] the OpenVINO device used for encode inference\n", params.openvino_encode_device.c_str());
139+
fprintf(stderr, " -dtw MODEL --dtw MODEL [%-7s] compute token-level timestamps\n", params.dtw.c_str());
140+
fprintf(stderr, " -ls, --log-score [%-7s] log best decoder scores of tokens\n", params.log_score?"true":"false");
141+
- fprintf(stderr, " -ng, --no-gpu [%-7s] disable GPU\n", params.use_gpu ? "false" : "true");
142+
+ fprintf(stderr, " -ng, --no-gpu [%-7s] disable GPU\n", FLAG_gpu == LLAMAFILE_GPU_DISABLE ? "false" : "true");
143+
fprintf(stderr, " -fa, --flash-attn [%-7s] flash attention\n", params.flash_attn ? "true" : "false");
144+
fprintf(stderr, " --suppress-regex REGEX [%-7s] regular expression matching tokens to suppress\n", params.suppress_regex.c_str());
145+
fprintf(stderr, " --grammar GRAMMAR [%-7s] GBNF grammar to guide decoding\n", params.grammar.c_str());
146+
@@ -825,10 +862,10 @@
147+
}
148+
}
149+
150+
- ::replace_all(txt_bg, "'", "\u2019");
151+
- ::replace_all(txt_bg, "\"", "\\\"");
152+
- ::replace_all(txt_fg, "'", "\u2019");
153+
- ::replace_all(txt_fg, "\"", "\\\"");
154+
+ txt_bg = replace_all(txt_bg, "'", "\u2019");
155+
+ txt_bg = replace_all(txt_bg, "\"", "\\\"");
156+
+ txt_fg = replace_all(txt_fg, "'", "\u2019");
157+
+ txt_fg = replace_all(txt_fg, "\"", "\\\"");
158+
}
159+
160+
if (is_first) {
161+
@@ -903,9 +940,37 @@
162+
163+
static void cb_log_disable(enum ggml_log_level , const char * , void * ) { }
164+
165+
+int whisper_server_main(int argc, char ** argv);
166+
+
167+
int main(int argc, char ** argv) {
168+
- whisper_params params;
169+
170+
+ mallopt(M_GRANULARITY, 1 * 1024 * 1024);
171+
+ mallopt(M_MMAP_THRESHOLD, 1 * 1024 * 1024);
172+
+ mallopt(M_TRIM_THRESHOLD, 32 * 1024 * 1024);
173+
+ FLAG_gpu = LLAMAFILE_GPU_DISABLE; // pass `--gpu auto` to enable it
174+
+ llamafile_check_cpu();
175+
+ ShowCrashReports();
176+
+
177+
+ if (llamafile_has(argv, "--version")) {
178+
+ puts("whisperfile v" LLAMAFILE_VERSION_STRING);
179+
+ return 0;
180+
+ }
181+
+
182+
+ if (llamafile_has(argv, "-h") ||
183+
+ llamafile_has(argv, "-help") ||
184+
+ llamafile_has(argv, "--help")) {
185+
+ llamafile_help("/zip/whisper.cpp/main.1.asc");
186+
+ __builtin_unreachable();
187+
+ }
188+
+
189+
+ argc = cosmo_args("/zip/.args", &argv);
190+
+
191+
+ if (!llamafile_has(argv, "--cli") &&
192+
+ (llamafile_has(argv, "--server") ||
193+
+ !llamafile_has(argv, "-f"))) {
194+
+ return whisper_server_main(argc, argv);
195+
+ }
196+
+
197+
// If the only argument starts with "@", read arguments line-by-line
198+
// from the given file.
199+
std::vector<std::string> vec_args;
200+
@@ -935,6 +1000,7 @@
201+
}
202+
}
203+
204+
+ whisper_params params;
205+
if (whisper_params_parse(argc, argv, params) == false) {
206+
whisper_print_usage(argc, argv, params);
207+
return 1;
208+
@@ -979,7 +1045,6 @@
209+
210+
struct whisper_context_params cparams = whisper_context_default_params();
211+
212+
- cparams.use_gpu = params.use_gpu;
213+
cparams.flash_attn = params.flash_attn;
214+
215+
if (!params.dtw.empty()) {
216+
@@ -1020,10 +1085,10 @@
217+
// read grammar from file
218+
std::ifstream ifs(params.grammar.c_str());
219+
const std::string txt = std::string((std::istreambuf_iterator<char>(ifs)), std::istreambuf_iterator<char>());
220+
- grammar = grammar_parser::parse(txt.c_str());
221+
+ grammar = whisper_grammar_parser::parse(txt.c_str());
222+
} else {
223+
// read grammar from string
224+
- grammar = grammar_parser::parse(params.grammar.c_str());
225+
+ grammar = whisper_grammar_parser::parse(params.grammar.c_str());
226+
}
227+
228+
// will be empty (default) if there are parse errors
229+
@@ -1032,7 +1097,7 @@
230+
return 4;
231+
} else {
232+
fprintf(stderr, "%s: grammar:\n", __func__);
233+
- grammar_parser::print_grammar(stderr, grammar);
234+
+ whisper_grammar_parser::print_grammar(stderr, grammar);
235+
fprintf(stderr, "\n");
236+
}
237+
}
238+
@@ -1044,8 +1109,8 @@
239+
std::vector<float> pcmf32; // mono-channel F32 PCM
240+
std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
241+
242+
- if (!::read_wav(fname_inp, pcmf32, pcmf32s, params.diarize)) {
243+
- fprintf(stderr, "error: failed to read WAV file '%s'\n", fname_inp.c_str());
244+
+ if (!slurp_audio_file(fname_inp.c_str(), pcmf32, pcmf32s, params.diarize)) {
245+
+ fprintf(stderr, "error: failed to read audio file '%s'\n", fname_inp.c_str());
246+
continue;
247+
}
248+

0 commit comments

Comments
 (0)