14
#define die(msg) do { fputs("error: " msg "\n", stderr); exit(1); } while (0)
15
#define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)
17
#define LOG_INF(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
18
#define LOG_WRN(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
19
#define LOG_ERR(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
20
#define LOG_DBG(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
27
std::vector<uint8_t> buf;
30
// RGB float32 image (NHWC)
31
// Memory layout: RGBRGBRGB...
32
struct clip_image_f32 {
36
std::vector<float> buf;
39
struct clip_image_grid_shape {
45
* Selects the best resolution from a list of possible resolutions based on the original size.
47
* @param original_size The original size of the image in the format (width, height).
48
* @param possible_resolutions A list of possible resolutions in the format [(width1, height1), (width2, height2), ...].
49
* @return The best fit resolution in the format (width, height).
51
static std::pair<int, int> select_best_resolution(const std::pair<int, int>& original_size, const std::vector<std::pair<int, int>>& possible_resolutions) {
52
int original_width = original_size.first;
53
int original_height = original_size.second;
55
std::pair<int, int> best_fit;
56
int max_effective_resolution = 0;
57
int min_wasted_resolution = std::numeric_limits<int>::max();
59
for (const auto& resolution : possible_resolutions) {
60
int width = resolution.first;
61
int height = resolution.second;
62
float scale = std::min(static_cast<float>(width) / original_width, static_cast<float>(height) / original_height);
63
int downscaled_width = static_cast<int>(original_width * scale);
64
int downscaled_height = static_cast<int>(original_height * scale);
65
int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height);
66
int wasted_resolution = (width * height) - effective_resolution;
67
// LOG_DBG("resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
68
if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) {
69
max_effective_resolution = effective_resolution;
70
min_wasted_resolution = wasted_resolution;
71
best_fit = resolution;
79
* @brief Get the anyres image grid shape object
82
* @param grid_pinpoints
83
* @param image_patch_size
86
static struct clip_image_grid_shape get_anyres_image_grid_shape(const std::pair<int, int> & image_size, const std::vector<std::pair<int, int>> & grid_pinpoints, int image_patch_size) {
88
Conversion from gguf flat array to vector:
89
std::vector<std::pair<int, int>> possible_resolutions;
90
for (int i = 0; i < 32 && params.image_grid_pinpoints[i] != 0; i+=2) {
91
possible_resolutions.push_back({params.image_grid_pinpoints[i], params.image_grid_pinpoints[i+1]});
94
auto best_resolution = select_best_resolution(image_size, grid_pinpoints);
95
return {best_resolution.first / image_patch_size, best_resolution.second / image_patch_size};
98
// Take the image segments in a grid configuration and return the embeddings and the number of embeddings into preallocated memory (image_embd_out)
99
static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *> & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out) {
101
struct ggml_context * ctx;
104
const int32_t image_size = clip_image_size(ctx_clip);
105
const int32_t patch_size = clip_patch_size(ctx_clip);
107
int32_t num_patches_per_side = image_size / patch_size; // 336 / 14 = 24 - used for embedding-patching boxes (24*24 = 576 patches)
109
int num_patches_width = grid_shape.first; // grid 1-4
110
int num_patches_height = grid_shape.second; // grid 1-4
112
const size_t num_images = num_patches_width * num_patches_height + 1;
114
// TODO: size calculation is not calculated - it's only tens of MB
118
ctx_size += clip_embd_nbytes(ctx_clip) * num_images * 8; // image_features
119
ctx_size += 1024*1024 * ggml_type_size(GGML_TYPE_F32);
122
struct ggml_init_params params {
123
/*.mem_size =*/ ctx_size,
124
/*.mem_buffer =*/ NULL,
125
/*.no_alloc =*/ false, // NOTE: this should be false when using the legacy API
128
// Python reference code for full unpad:
130
base_image_feature = image_feature[0]
131
image_feature = image_feature[1:]
132
image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
133
image_feature = image_feature.flatten(1, 2).flatten(2, 3)
134
image_feature = unpad_image(image_feature, image_sizes[image_idx])
135
image_feature = torch.cat((
137
self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1)
139
image_feature = image_feature.flatten(1, 2).transpose(0, 1)
140
image_feature = torch.cat((base_image_feature, image_feature), dim=0)
142
// We now have two options: unpad or no unpad. Unpad removes tokens for faster llm eval.
143
// In terms of result quality it appears to make no difference, so we'll start with the easier approach given 5D tensors are not supported in ggml yet.
144
// Without unpad we have to split the sub-image embeddings into patches of 24 features each and permute them.
145
// Once all images are processed to prepended the base_image_features without any changes.
147
// Pytorch reference simplified, modified for ggml compatibility - confirmed identical output in python (for a 2x2 grid image (676x676 scaling))
149
image_feature = image_feature.view(2, 2, 24, 24, 4096)
150
image_feature = image_feature.permute(0, 2, 1, 3, 4).contiguous()
151
image_feature = image_feature.view(2, 24, 2, 24, 4096)
152
image_feature = image_feature.flatten(0, 3)
154
// Reshape to 4D tensor by merging the last two dimensions
155
image_feature = image_feature.view(2, 2, 24, 24*4096)
156
image_feature = image_feature.permute(0, 2, 1, 3).contiguous()
157
image_feature = image_feature.view(-1, 4096)
160
model.ctx = ggml_init(params);
162
struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_patches(ctx_clip), num_images - 1); // example: 4096 x 576 x 4
163
// ggml_tensor_printf(image_features,"image_features",__LINE__,false,false);
164
// fill it with the image embeddings, ignoring the base
165
for (size_t i = 1; i < num_images; i++) {
166
size_t offset = (i-1) * clip_embd_nbytes(ctx_clip);
167
memcpy((uint8_t *)(image_features->data) + offset, image_embd_v[i], clip_embd_nbytes(ctx_clip));
170
struct ggml_cgraph * gf = ggml_new_graph(model.ctx);
171
size_t size_ele = ggml_type_size(GGML_TYPE_F32);
173
struct ggml_tensor *image_features_patchview = ggml_view_4d(model.ctx, image_features,
174
num_patches_per_side * clip_n_mmproj_embd(ctx_clip),
175
num_patches_per_side,
178
size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip),
179
size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip) * num_patches_per_side,
180
size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip) * num_patches_per_side * num_patches_width, 0);
181
// ggml_tensor_printf(image_features_patchview,"image_features_patchview",__LINE__,false,false);
182
struct ggml_tensor *permuted_cont = ggml_cont(model.ctx, ggml_permute(model.ctx, image_features_patchview, 0, 2, 1, 3));
184
At the end of each row we have to add the row_end embeddings, which are the same as the newline embeddings
185
image_feature = torch.cat((
187
self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.device)
192
// ggml_tensor_printf(permuted_cont,"permuted_cont",__LINE__,false,false);
193
struct ggml_tensor *flatten = ggml_view_2d(model.ctx, permuted_cont, clip_n_mmproj_embd(ctx_clip), num_patches_height * num_patches_width * num_patches_per_side * num_patches_per_side, size_ele * clip_n_mmproj_embd(ctx_clip), 0);
194
// ggml_tensor_printf(flatten,"flatten",__LINE__,false,false);
195
ggml_build_forward_expand(gf, flatten);
196
ggml_graph_compute_with_ctx(model.ctx, gf, 1);
197
struct ggml_tensor* result = ggml_graph_node(gf, -1);
199
memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as global context
200
// append without newline tokens (default behavior in llava_arch when not using unpad ):
201
memcpy(image_embd_out + clip_n_patches(ctx_clip) * clip_n_mmproj_embd(ctx_clip), (float*)result->data, clip_embd_nbytes(ctx_clip) * (num_images-1)); // grid patches
202
*n_img_pos_out = static_cast<int>(result->ne[1]+clip_n_patches(ctx_clip));
204
// Debug: Test single segments
205
// Current findings: sending base image, sending a segment embedding all works similar to python
206
// However, permuted embeddings do not work yet (stride issue?)
207
// memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as context
208
// memcpy(image_embd_out, (float*)prepared_cont->data, clip_embd_nbytes(ctx_clip)); // main image as context
209
// *n_img_pos_out=576;
211
ggml_free(model.ctx);
215
static clip_image_f32 * only_v2_5_reshape_by_patch(clip_image_f32 * image, int patch_size) {
216
int width = image->nx;
217
int height = image->ny;
218
int num_patches = (height / patch_size) * (width / patch_size);
219
clip_image_f32 * patch = clip_image_f32_init();
220
patch->nx = patch_size * num_patches;
221
patch->ny = patch_size;
222
patch->buf.resize(3 * patch->nx * patch->ny);
226
for (int i = 0; i < height; i += patch_size) {
227
for (int j = 0; j < width; j += patch_size) {
228
for (int pi = 0; pi < patch_size; ++pi) {
229
for (int pj = 0; pj < patch_size; ++pj) {
230
int input_index = ((i + pi) * width + (j + pj)) * 3;
231
int output_index = (pi * patch_size * num_patches + patch_index * patch_size + pj) * 3;
232
patch->buf[output_index] = image->buf[input_index];
233
patch->buf[output_index+1] = image->buf[input_index+1];
234
patch->buf[output_index+2] = image->buf[input_index+2];
243
static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
244
// std::vector<clip_image_f32*> img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336
245
clip_image_f32_batch img_res_v;
247
img_res_v.data = nullptr;
248
if (!clip_image_preprocess(ctx_clip, img, &img_res_v)) {
249
LOG_ERR("%s: unable to preprocess image\n", __func__);
250
delete[] img_res_v.data;
254
const int64_t t_img_enc_start_us = ggml_time_us();
256
const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip);
258
if (clip_is_minicpmv(ctx_clip)) {
259
std::vector<float *> image_embd_v;
260
image_embd_v.resize(img_res_v.size);
261
struct clip_image_size * load_image_size = clip_image_size_init();
262
for (size_t i = 0; i < img_res_v.size; i++) {
263
const int64_t t_img_enc_step_start_us = ggml_time_us();
264
image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip));
266
load_image_size->width = img_res_v.data[i].nx;
267
load_image_size->height = img_res_v.data[i].ny;
268
clip_add_load_image_size(ctx_clip, load_image_size);
269
bool encoded = false;
270
int has_minicpmv_projector = clip_is_minicpmv(ctx_clip);
271
if (has_minicpmv_projector == 2) {
272
encoded = clip_image_encode(ctx_clip, n_threads, only_v2_5_reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]);
274
else if (has_minicpmv_projector == 3) {
275
encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]);
278
LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
281
const int64_t t_img_enc_steop_batch_us = ggml_time_us();
282
LOG_INF("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)img_res_v.size, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0);
284
const int64_t t_img_enc_batch_us = ggml_time_us();
285
LOG_INF("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
287
int n_img_pos_out = 0;
288
for (size_t i = 0; i < image_embd_v.size(); i++) {
289
std::memcpy(image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip), image_embd_v[i], clip_embd_nbytes(ctx_clip));
290
n_img_pos_out += clip_n_patches(ctx_clip);
292
*n_img_pos = n_img_pos_out;
293
for (size_t i = 0; i < image_embd_v.size(); i++) {
294
free(image_embd_v[i]);
296
image_embd_v.clear();
297
load_image_size->width = img->nx;
298
load_image_size->height = img->ny;
299
clip_add_load_image_size(ctx_clip, load_image_size);
300
LOG_INF("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height);
302
else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
303
// flat / default llava-1.5 type embedding
304
*n_img_pos = clip_n_patches(ctx_clip);
305
bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096
306
delete[] img_res_v.data;
308
LOG_ERR("Unable to encode image\n");
314
// spatial_unpad llava-1.6 type embedding
315
// TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working
316
std::vector<float *> image_embd_v;
317
image_embd_v.resize(img_res_v.size);
318
for (size_t i = 0; i < img_res_v.size; i++) {
319
image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); // 576 patches * 4096 embeddings * 4 bytes = 9437184
320
const bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
322
LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
326
const int64_t t_img_enc_batch_us = ggml_time_us();
327
LOG_INF("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
329
const int32_t * image_grid = clip_image_grid(ctx_clip);
331
std::vector<std::pair<int, int>> grid_pinpoints;
332
for (int i = 0; i < 32 && image_grid[i] != 0; i += 2) {
333
grid_pinpoints.push_back({image_grid[i], image_grid[i+1]});
336
// free all img_res_v - not needed anymore
337
delete[] img_res_v.data;
339
img_res_v.data = nullptr;
341
const int32_t image_size = clip_image_size(ctx_clip);
343
struct clip_image_grid_shape grid_shape = get_anyres_image_grid_shape({img->nx,img->ny}, grid_pinpoints, image_size);
346
clip_llava_handle_patches(ctx_clip, image_embd_v, grid_shape, image_embd, &n_img_pos_out);
347
*n_img_pos = n_img_pos_out;
349
for (size_t i = 0; i < image_embd_v.size(); i++) {
350
free(image_embd_v[i]);
352
image_embd_v.clear();
354
// debug image/segment/normalization content:
355
// clip_image_u8 * tmp = clip_image_u8_init();
356
// clip_image_convert_f32_to_u8(*image_feature, *tmp);
357
// clip_image_save_to_bmp(*tmp, "image_feature.bmp");
360
LOG_INF("%s: image embedding created: %d tokens\n", __func__, *n_img_pos);
362
const int64_t t_img_enc_end_us = ggml_time_us();
363
float t_img_enc_ms = (t_img_enc_end_us - t_img_enc_start_us) / 1000.0;
365
LOG_INF("\n%s: image encoded in %8.2f ms by CLIP (%8.2f ms per image patch)\n", __func__, t_img_enc_ms, t_img_enc_ms / *n_img_pos);
370
bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip) {
371
// make sure that the correct mmproj was used, i.e., compare apples to apples
372
int n_llama_embd = llama_n_embd(llama_get_model(ctx_llama));
373
auto n_image_embd = clip_n_mmproj_embd(ctx_clip);
374
if (n_image_embd != n_llama_embd) {
375
LOG_ERR("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd);
381
bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
382
int num_max_patches = 6;
383
if (clip_is_minicpmv(ctx_clip)) {
384
num_max_patches = 10;
386
float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model
388
LOG_ERR("Unable to allocate memory for image embeddings\n");
393
if (!encode_image_with_clip(ctx_clip, n_threads, img, image_embd, &n_img_pos)) {
394
LOG_ERR("%s: cannot encode image, aborting\n", __func__);
398
*image_embd_out = image_embd;
399
*n_img_pos_out = n_img_pos;
404
bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed, int n_batch, int * n_past) {
405
int n_embd = llama_n_embd(llama_get_model(ctx_llama));
407
for (int i = 0; i < image_embed->n_image_pos; i += n_batch) {
408
int n_eval = image_embed->n_image_pos - i;
409
if (n_eval > n_batch) {
412
llama_batch batch = {int32_t(n_eval), nullptr, (image_embed->embed+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
413
if (llama_decode(ctx_llama, batch)) {
414
LOG_ERR("%s : failed to eval\n", __func__);
422
struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length) {
423
clip_image_u8 * img = clip_image_u8_init();
424
if (!clip_image_load_from_bytes(image_bytes, image_bytes_length, img)) {
425
clip_image_u8_free(img);
426
LOG_ERR("%s: can't load image from bytes, is it a valid image?", __func__);
430
float* image_embed = NULL;
432
bool image_embed_result = llava_image_embed_make_with_clip_img(ctx_clip, n_threads, img, &image_embed, &n_image_pos);
433
if (!image_embed_result) {
434
clip_image_u8_free(img);
435
LOG_ERR("%s: coulnd't embed the image\n", __func__);
439
clip_image_u8_free(img);
440
auto result = (llava_image_embed*)malloc(sizeof(llava_image_embed));
441
result->embed = image_embed;
442
result->n_image_pos = n_image_pos;
446
static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long *sizeOut) {
447
auto file = fopen(path, "rb");
449
LOG_ERR("%s: can't read file %s\n", __func__, path);
453
fseek(file, 0, SEEK_END);
454
auto fileSize = ftell(file);
455
fseek(file, 0, SEEK_SET);
457
auto buffer = (unsigned char *)malloc(fileSize); // Allocate memory to hold the file data
458
if (buffer == NULL) {
459
LOG_ERR("%s: failed to alloc %ld bytes for file %s\n", __func__, fileSize, path);
460
perror("Memory allocation error");
465
size_t ret = fread(buffer, 1, fileSize, file); // Read the file into the buffer
467
die_fmt("read error: %s", strerror(errno));
469
if (ret != (size_t) fileSize) {
470
die("unexpectedly reached end of file");
472
fclose(file); // Close the file
479
struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path) {
480
unsigned char* image_bytes;
481
long image_bytes_length;
482
auto loaded = load_file_to_bytes(image_path, &image_bytes, &image_bytes_length);
484
LOG_ERR("%s: failed to load %s\n", __func__, image_path);
488
llava_image_embed *embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length);
494
void llava_image_embed_free(struct llava_image_embed * embed) {