538 lines
15 KiB
C++
538 lines
15 KiB
C++
#include "ggml.h"
|
|
#include "llama.h"
|
|
#include "mtmd-helper.h"
|
|
#include "mtmd.h"
|
|
|
|
#include <cinttypes>
|
|
#include <cmath>
|
|
#include <csignal>
|
|
#include <cstdio>
|
|
#include <cstdlib>
|
|
#include <getopt.h>
|
|
#include <string>
|
|
#include <sys/stat.h>
|
|
#include <thread>
|
|
#include <unistd.h>
|
|
#include <vector>
|
|
|
|
static volatile bool g_interrupted = false;
|
|
|
|
static void noop_log(const ggml_log_level level, const char *text,
|
|
void *user_data) {
|
|
(void)level;
|
|
(void)text;
|
|
(void)user_data;
|
|
}
|
|
|
|
static void sigint_handler(const int signo) {
|
|
(void)signo;
|
|
g_interrupted = true;
|
|
}
|
|
|
|
static void batch_clear(llama_batch &batch) { batch.n_tokens = 0; }
|
|
|
|
static void batch_add(llama_batch &batch, const llama_token id,
|
|
const llama_pos pos,
|
|
const std::vector<llama_seq_id> &seq_ids,
|
|
const bool logits) {
|
|
batch.token[batch.n_tokens] = id;
|
|
batch.pos[batch.n_tokens] = pos;
|
|
batch.n_seq_id[batch.n_tokens] = static_cast<int32_t>(seq_ids.size());
|
|
for (size_t i = 0; i < seq_ids.size(); ++i) {
|
|
batch.seq_id[batch.n_tokens][i] = seq_ids[i];
|
|
}
|
|
batch.logits[batch.n_tokens] = logits ? 1 : 0;
|
|
batch.n_tokens++;
|
|
}
|
|
|
|
static std::string token_to_piece(const llama_vocab *vocab,
|
|
const llama_token token) {
|
|
std::string piece;
|
|
piece.resize(256);
|
|
int32_t n = llama_token_to_piece(
|
|
vocab, token, &piece[0], static_cast<int32_t>(piece.size()), 0, false);
|
|
if (n < 0) {
|
|
piece.resize(-n);
|
|
n = llama_token_to_piece(vocab, token, &piece[0],
|
|
static_cast<int32_t>(piece.size()), 0, false);
|
|
}
|
|
piece.resize(std::max(0, n));
|
|
return piece;
|
|
}
|
|
|
|
static std::string json_escape(const std::string &s) {
|
|
std::string out;
|
|
out.reserve(s.size() + 2);
|
|
for (const char c : s) {
|
|
switch (c) {
|
|
case '"':
|
|
out += "\\\"";
|
|
break;
|
|
case '\\':
|
|
out += "\\\\";
|
|
break;
|
|
case '\n':
|
|
out += "\\n";
|
|
break;
|
|
case '\r':
|
|
out += "\\r";
|
|
break;
|
|
case '\t':
|
|
out += "\\t";
|
|
break;
|
|
default:
|
|
out += c;
|
|
break;
|
|
}
|
|
}
|
|
return out;
|
|
}
|
|
|
|
static std::vector<unsigned char> read_clipboard() {
|
|
const char *cmd;
|
|
if (getenv("WAYLAND_DISPLAY")) {
|
|
cmd = "wl-paste 2>/dev/null";
|
|
} else {
|
|
cmd = "xclip -selection clipboard -t image/png -o 2>/dev/null";
|
|
}
|
|
std::vector<unsigned char> data;
|
|
FILE *pipe = popen(cmd, "re");
|
|
if (!pipe)
|
|
return data;
|
|
unsigned char buf[65536];
|
|
size_t n;
|
|
while ((n = fread(buf, 1, sizeof(buf), pipe)) > 0) {
|
|
data.insert(data.end(), buf, buf + n);
|
|
}
|
|
const int st = pclose(pipe);
|
|
if (st != 0 || data.empty()) {
|
|
data.clear();
|
|
// fallback for X11: try without explicit target
|
|
if (!getenv("WAYLAND_DISPLAY")) {
|
|
pipe = popen("xclip -selection clipboard -o 2>/dev/null", "re");
|
|
if (pipe) {
|
|
while ((n = fread(buf, 1, sizeof(buf), pipe)) > 0)
|
|
data.insert(data.end(), buf, buf + n);
|
|
pclose(pipe);
|
|
}
|
|
}
|
|
}
|
|
return data;
|
|
}
|
|
|
|
static std::string capture_screenshot(bool quiet) {
|
|
const char *tmpdir = getenv("TMPDIR");
|
|
if (!tmpdir)
|
|
tmpdir = "/tmp";
|
|
std::string path = std::string(tmpdir) + "/ocr_screenshot.png";
|
|
|
|
if (!quiet)
|
|
fprintf(stderr, "Select a screen region to OCR...\n");
|
|
|
|
auto check = [&]() -> bool {
|
|
struct stat st {};
|
|
return stat(path.c_str(), &st) == 0 && st.st_size > 0;
|
|
};
|
|
|
|
// 1. gnome-screenshot (blocks until selection complete)
|
|
unlink(path.c_str());
|
|
if (system(("gnome-screenshot --area -f " + path + " >/dev/null 2>&1")
|
|
.c_str()) == 0 &&
|
|
check())
|
|
return path;
|
|
|
|
// 2. spectacle (KDE, blocks)
|
|
unlink(path.c_str());
|
|
if (system(
|
|
("spectacle --region -b -o " + path + " >/dev/null 2>&1").c_str()) ==
|
|
0 &&
|
|
check())
|
|
return path;
|
|
|
|
// 3. flameshot (writes raw PNG to stdout, blocks)
|
|
unlink(path.c_str());
|
|
if (system(("flameshot gui -r > " + path + " 2>/dev/null").c_str()) == 0 &&
|
|
check())
|
|
return path;
|
|
|
|
// 4. maim (lightweight X11, blocks)
|
|
unlink(path.c_str());
|
|
if (system(("maim -s " + path + " 2>/dev/null").c_str()) == 0 && check())
|
|
return path;
|
|
|
|
// 5. slurp + grim (Sway/Wayland, blocks)
|
|
unlink(path.c_str());
|
|
if (system(("slurp | grim -g - " + path + " 2>/dev/null").c_str()) == 0 &&
|
|
check())
|
|
return path;
|
|
|
|
return "";
|
|
}
|
|
|
|
static void print_usage(const char *prog) {
|
|
fprintf(
|
|
stderr,
|
|
"Usage: %s -m <model> --mmproj <mmproj> [options] [<image_path>]\n"
|
|
"\n"
|
|
"Options:\n"
|
|
" -m, --model <path> model file path (GGUF)\n"
|
|
" --mmproj <path> mmproj file path\n"
|
|
" -p <text> OCR prompt (default: OCR)\n"
|
|
" -t <n> number of threads (default: cpu cores)\n"
|
|
" --ngl <n> GPU layers (-1 = all, default: -1)\n"
|
|
" -c <n> context size (default: 8192)\n"
|
|
" --temp <f> sampling temperature (0 = greedy, default: 0)\n"
|
|
" -s <seed> random seed\n"
|
|
" --json output in JSON format\n"
|
|
" --chat-template <name> chat template (default: auto from model)\n"
|
|
" --screenshot interactively select screen region to OCR\n"
|
|
" -q, --quiet suppress info logs, show only OCR result\n"
|
|
" -h show this help\n"
|
|
"\n"
|
|
"If <image_path> is omitted, reads image from clipboard.\n"
|
|
"Requires wl-paste (Wayland) or xclip (X11) for clipboard support.\n"
|
|
"\n"
|
|
"Examples:\n"
|
|
" %s -m model.gguf --mmproj mmproj.gguf -p \"OCR\" image.png\n"
|
|
" %s -m model.gguf --mmproj mmproj.gguf # use clipboard image\n"
|
|
" %s -m model.gguf --mmproj mmproj.gguf --screenshot # select "
|
|
"region\n",
|
|
prog, prog, prog, prog);
|
|
}
|
|
|
|
int main(int argc, char **argv) {
|
|
std::string model_path;
|
|
std::string mmproj_path;
|
|
std::string image_path;
|
|
std::string prompt = "OCR";
|
|
std::string chat_template;
|
|
bool screenshot_mode = false;
|
|
int n_threads = static_cast<int>(std::thread::hardware_concurrency());
|
|
int n_gpu_layers = -1;
|
|
int n_ctx = 8192;
|
|
float temp = 0.0f;
|
|
uint32_t seed = LLAMA_DEFAULT_SEED;
|
|
bool json_output = false;
|
|
bool quiet = false;
|
|
|
|
while (true) {
|
|
static option long_opts[] = {
|
|
{"temp", required_argument, nullptr, 0},
|
|
{"json", no_argument, nullptr, 1},
|
|
{"mmproj", required_argument, nullptr, 2},
|
|
{"ngl", required_argument, nullptr, 3},
|
|
{"model", required_argument, nullptr, 'm'},
|
|
{"chat-template", required_argument, nullptr, 4},
|
|
{"screenshot", no_argument, nullptr, 5},
|
|
{"quiet", no_argument, nullptr, 6},
|
|
{"help", no_argument, nullptr, 'h'},
|
|
{nullptr, 0, nullptr, 0}};
|
|
int idx = 0;
|
|
int c = getopt_long(argc, argv, "m:p:t:c:s:qh", long_opts, &idx);
|
|
if (c == -1)
|
|
break;
|
|
switch (c) {
|
|
case 0:
|
|
temp = std::stof(optarg);
|
|
break;
|
|
case 1:
|
|
json_output = true;
|
|
break;
|
|
case 2:
|
|
mmproj_path = optarg;
|
|
break;
|
|
case 3:
|
|
n_gpu_layers = std::stoi(optarg);
|
|
break;
|
|
case 4:
|
|
chat_template = optarg;
|
|
break;
|
|
case 5:
|
|
screenshot_mode = true;
|
|
break;
|
|
case 6:
|
|
case 'q':
|
|
quiet = true;
|
|
break;
|
|
case 'm':
|
|
model_path = optarg;
|
|
break;
|
|
case 'p':
|
|
prompt = optarg;
|
|
break;
|
|
case 't':
|
|
n_threads = std::stoi(optarg);
|
|
break;
|
|
case 'c':
|
|
n_ctx = std::stoi(optarg);
|
|
break;
|
|
case 's':
|
|
seed = static_cast<uint32_t>(std::stoul(optarg));
|
|
break;
|
|
case 'h':
|
|
print_usage(argv[0]);
|
|
return 0;
|
|
default:
|
|
print_usage(argv[0]);
|
|
return 1;
|
|
}
|
|
}
|
|
|
|
if (model_path.empty() || mmproj_path.empty()) {
|
|
fprintf(stderr, "ERROR: -m/--model and --mmproj are required\n");
|
|
return 1;
|
|
}
|
|
|
|
if (screenshot_mode) {
|
|
image_path = capture_screenshot(quiet);
|
|
if (image_path.empty()) {
|
|
fprintf(stderr,
|
|
"ERROR: no screenshot tool found. Install gnome-screenshot, "
|
|
"flameshot, spectacle, maim, or slurp+grim.\n");
|
|
return 1;
|
|
}
|
|
} else if (optind < argc) {
|
|
image_path = argv[optind];
|
|
}
|
|
|
|
signal(SIGINT, sigint_handler);
|
|
|
|
int64_t t_start_us = ggml_time_us();
|
|
|
|
llama_backend_init();
|
|
|
|
if (quiet) {
|
|
llama_log_set(noop_log, nullptr);
|
|
mtmd_helper_log_set(noop_log, nullptr);
|
|
}
|
|
|
|
llama_model_params mparams = llama_model_default_params();
|
|
mparams.n_gpu_layers = n_gpu_layers;
|
|
|
|
llama_model *model = llama_model_load_from_file(model_path.c_str(), mparams);
|
|
if (!model) {
|
|
fprintf(stderr, "ERROR: failed to load model from %s\n",
|
|
model_path.c_str());
|
|
return 1;
|
|
}
|
|
|
|
llama_context_params cparams = llama_context_default_params();
|
|
cparams.n_ctx = static_cast<uint32_t>(n_ctx);
|
|
cparams.n_batch = 512;
|
|
cparams.n_ubatch = 512;
|
|
|
|
llama_context *lctx = llama_init_from_model(model, cparams);
|
|
if (!lctx) {
|
|
fprintf(stderr, "ERROR: failed to create context\n");
|
|
llama_model_free(model);
|
|
return 1;
|
|
}
|
|
|
|
llama_set_n_threads(lctx, n_threads, n_threads);
|
|
|
|
const llama_vocab *vocab = llama_model_get_vocab(model);
|
|
|
|
mtmd_context_params mtmd_params = mtmd_context_params_default();
|
|
mtmd_params.use_gpu = (n_gpu_layers != 0);
|
|
mtmd_params.n_threads = n_threads;
|
|
|
|
mtmd_context *ctx_vision =
|
|
mtmd_init_from_file(mmproj_path.c_str(), model, mtmd_params);
|
|
if (!ctx_vision) {
|
|
fprintf(stderr, "ERROR: failed to load mmproj from %s\n",
|
|
mmproj_path.c_str());
|
|
llama_free(lctx);
|
|
llama_model_free(model);
|
|
return 1;
|
|
}
|
|
|
|
int64_t t_loaded_us = ggml_time_us();
|
|
|
|
mtmd_bitmap *bmp;
|
|
if (!image_path.empty()) {
|
|
bmp = mtmd_helper_bitmap_init_from_file(ctx_vision, image_path.c_str());
|
|
if (!bmp) {
|
|
fprintf(stderr, "ERROR: failed to load image from %s\n",
|
|
image_path.c_str());
|
|
if (screenshot_mode)
|
|
unlink(image_path.c_str());
|
|
mtmd_free(ctx_vision);
|
|
llama_free(lctx);
|
|
llama_model_free(model);
|
|
return 1;
|
|
}
|
|
} else {
|
|
if (!quiet)
|
|
fprintf(stderr, "Reading image from clipboard...\n");
|
|
auto clip = read_clipboard();
|
|
if (clip.empty()) {
|
|
fprintf(stderr,
|
|
"ERROR: clipboard is empty or no image found.\n"
|
|
"Make sure wl-paste (Wayland) or xclip (X11) is installed.\n");
|
|
mtmd_free(ctx_vision);
|
|
llama_free(lctx);
|
|
llama_model_free(model);
|
|
return 1;
|
|
}
|
|
bmp =
|
|
mtmd_helper_bitmap_init_from_buf(ctx_vision, clip.data(), clip.size());
|
|
if (!bmp) {
|
|
fprintf(stderr, "ERROR: failed to decode clipboard image.\n");
|
|
mtmd_free(ctx_vision);
|
|
llama_free(lctx);
|
|
llama_model_free(model);
|
|
return 1;
|
|
}
|
|
}
|
|
|
|
// Build user message with image marker placeholder
|
|
std::string user_content = std::string(mtmd_default_marker()) + prompt;
|
|
|
|
// Apply chat template for any model
|
|
std::string full_prompt;
|
|
const char *tmpl = chat_template.empty()
|
|
? llama_model_chat_template(model, nullptr)
|
|
: chat_template.c_str();
|
|
if (tmpl) {
|
|
llama_chat_message msg[1] = {{"user", user_content.c_str()}};
|
|
char buf[8192];
|
|
int32_t n = llama_chat_apply_template(tmpl, msg, 1, true, buf, sizeof(buf));
|
|
if (n > 0) {
|
|
full_prompt = std::string(buf, n);
|
|
}
|
|
}
|
|
|
|
// Fallback: use chatml format for models without recognizable template
|
|
if (full_prompt.empty()) {
|
|
full_prompt = "<|im_start|>user\n" + user_content +
|
|
"<|im_end|>\n<|im_start|>assistant\n";
|
|
}
|
|
|
|
// Template output already includes role markers and special tokens (e.g.
|
|
// [gMASK]<sop>) So we don't add BOS separately — add_special=false
|
|
mtmd_input_chunks *chunks = mtmd_input_chunks_init();
|
|
|
|
mtmd_input_text in_text;
|
|
in_text.text = full_prompt.c_str();
|
|
in_text.add_special = false;
|
|
in_text.parse_special = true;
|
|
|
|
const mtmd_bitmap *bitmaps[1] = {bmp};
|
|
|
|
int32_t res = mtmd_tokenize(ctx_vision, chunks, &in_text, bitmaps, 1);
|
|
if (res != 0) {
|
|
fprintf(stderr, "ERROR: mtmd_tokenize failed (%d)\n", res);
|
|
mtmd_input_chunks_free(chunks);
|
|
mtmd_bitmap_free(bmp);
|
|
if (screenshot_mode)
|
|
unlink(image_path.c_str());
|
|
mtmd_free(ctx_vision);
|
|
llama_free(lctx);
|
|
llama_model_free(model);
|
|
return 1;
|
|
}
|
|
|
|
llama_pos n_past = 0;
|
|
res = mtmd_helper_eval_chunks(ctx_vision, lctx, chunks, n_past, 0, 512, true,
|
|
&n_past);
|
|
if (res != 0) {
|
|
fprintf(stderr, "ERROR: mtmd_helper_eval_chunks failed (%d)\n", res);
|
|
mtmd_input_chunks_free(chunks);
|
|
mtmd_bitmap_free(bmp);
|
|
if (screenshot_mode)
|
|
unlink(image_path.c_str());
|
|
mtmd_free(ctx_vision);
|
|
llama_free(lctx);
|
|
llama_model_free(model);
|
|
return 1;
|
|
}
|
|
|
|
mtmd_input_chunks_free(chunks);
|
|
mtmd_bitmap_free(bmp);
|
|
if (screenshot_mode)
|
|
unlink(image_path.c_str());
|
|
|
|
llama_sampler *smpl;
|
|
if (temp <= 0.0f) {
|
|
llama_sampler_chain_params sparams = llama_sampler_chain_default_params();
|
|
smpl = llama_sampler_chain_init(sparams);
|
|
llama_sampler_chain_add(smpl, llama_sampler_init_greedy());
|
|
} else {
|
|
llama_sampler_chain_params sparams = llama_sampler_chain_default_params();
|
|
smpl = llama_sampler_chain_init(sparams);
|
|
llama_sampler_chain_add(smpl, llama_sampler_init_top_k(40));
|
|
llama_sampler_chain_add(smpl, llama_sampler_init_top_p(0.9f, 1));
|
|
llama_sampler_chain_add(smpl, llama_sampler_init_temp(temp));
|
|
llama_sampler_chain_add(smpl, llama_sampler_init_dist(seed));
|
|
}
|
|
|
|
llama_batch batch = llama_batch_init(1, 0, 1);
|
|
|
|
int64_t t_infer_start_us = ggml_time_us();
|
|
std::string ocr_text;
|
|
int n_generated = 0;
|
|
|
|
for (int i = 0; i < 2048; i++) {
|
|
if (g_interrupted)
|
|
break;
|
|
|
|
llama_token token_id = llama_sampler_sample(smpl, lctx, -1);
|
|
llama_sampler_accept(smpl, token_id);
|
|
|
|
if (llama_vocab_is_eog(vocab, token_id))
|
|
break;
|
|
|
|
std::string piece = token_to_piece(vocab, token_id);
|
|
ocr_text += piece;
|
|
|
|
if (!json_output) {
|
|
printf("%s", piece.c_str());
|
|
fflush(stdout);
|
|
}
|
|
|
|
batch_clear(batch);
|
|
batch_add(batch, token_id, n_past++, {0}, true);
|
|
|
|
if (llama_decode(lctx, batch) != 0) {
|
|
fprintf(stderr, "\nERROR: llama_decode failed\n");
|
|
break;
|
|
}
|
|
|
|
n_generated++;
|
|
}
|
|
|
|
int64_t t_end_us = ggml_time_us();
|
|
|
|
if (!json_output) {
|
|
printf("\n");
|
|
}
|
|
|
|
llama_batch_free(batch);
|
|
llama_sampler_free(smpl);
|
|
mtmd_free(ctx_vision);
|
|
llama_free(lctx);
|
|
llama_model_free(model);
|
|
llama_backend_free();
|
|
|
|
if (json_output) {
|
|
printf("{\n");
|
|
printf(" \"text\": \"%s\",\n", json_escape(ocr_text).c_str());
|
|
printf(" \"meta\": {\n");
|
|
printf(" \"model\": \"%s\",\n", json_escape(model_path).c_str());
|
|
printf(" \"mmproj\": \"%s\",\n", json_escape(mmproj_path).c_str());
|
|
printf(" \"prompt\": \"%s\",\n", json_escape(prompt).c_str());
|
|
printf(" \"n_prompt_tokens\": %" PRId64 ",\n",
|
|
static_cast<int64_t>(n_past));
|
|
printf(" \"n_generated\": %d,\n", n_generated);
|
|
printf(" \"timings_ms\": {\n");
|
|
printf(" \"load\": %" PRId64 ",\n", (t_loaded_us - t_start_us) / 1000);
|
|
printf(" \"inference\": %" PRId64 "\n",
|
|
(t_end_us - t_infer_start_us) / 1000);
|
|
printf(" }\n");
|
|
printf(" }\n");
|
|
printf("}\n");
|
|
}
|
|
|
|
return g_interrupted ? 130 : 0;
|
|
}
|