Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-12 05:17:21 +01:00
Code Issues Packages Projects Releases Wiki Activity

imitate reshape bug of python code

Browse Source
This commit is contained in:
caitianchi 2024-07-04 17:25:02 +08:00
parent 4c67d7cef5
commit 977941d9fe
3 changed files with 50 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

21
examples/llava/clip.cpp
View File

@ -554,7 +554,7 @@ struct clip_ctx {
ggml_gallocr_t compute_alloc = NULL;
};
static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs, std::pair<int, int> load_image_size = {448, 448}) {
static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs, std::pair<int, int> load_image_size = {448, 448}, bool is_inf = false) {
if (!ctx->has_vision_encoder) {
LOG_TEE("This gguf file seems to have no vision encoder\n");
return nullptr;
@ -569,6 +569,10 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
if (ctx->has_minicpmv_projector) {
image_size_width = load_image_size.first;
image_size_height = load_image_size.second;
if (is_inf){
image_size_width = imgs->data->nx;
image_size_height = imgs->data->ny;
}
}
const int patch_size = hparams.patch_size;
const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size));
@ -762,7 +766,8 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings);
embeddings = ggml_add(ctx0, embeddings, model.mm_2_b);
} else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
}
else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
// ggml_tensor_printf(embeddings, "mm_0_w",0,true,false);
@ -1450,7 +1455,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1, s
new_clip->compute_alloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(new_clip->backend));
clip_image_f32_batch batch;
batch.size = 1;
ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch, load_image_size);
ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch, load_image_size, false);
ggml_gallocr_reserve(new_clip->compute_alloc, gf);
size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_clip->compute_alloc, 0);
LOG_TEE("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0);
@ -2080,7 +2085,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
}
// build the inference graph
ggml_cgraph * gf = clip_image_build_graph(ctx, imgs, load_image_size);
ggml_cgraph * gf = clip_image_build_graph(ctx, imgs, load_image_size, true);
ggml_gallocr_alloc_graph(ctx->compute_alloc, gf);
// set inputs
@ -2091,8 +2096,8 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
int image_size_width = image_size;
int image_size_height = image_size;
if (ctx->has_minicpmv_projector) {
image_size_width = load_image_size.first;
image_size_height = load_image_size.second;
image_size_width = imgs->data[0].nx;;
image_size_height = imgs->data[0].ny;
}
const int patch_size = hparams.patch_size;
const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size));
@ -2144,8 +2149,8 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
// -> https://huggingface.co/Qwen/Qwen-VL/tree/main
// -> https://huggingface.co/Qwen/Qwen-VL/blob/0547ed36a86561e2e42fecec8fd0c4f6953e33c4/visual.py#L23
struct ggml_tensor * pos_embed = ggml_graph_get_tensor(gf, "pos_embed");
int pos_w = image_size_width/patch_size;
int pos_h = image_size_height/patch_size;
int pos_w = load_image_size.first/patch_size;
int pos_h = load_image_size.second/patch_size;
int embed_dim = 4096;
auto pos_embed_t = get_2d_sincos_pos_embed(embed_dim, std::make_pair(pos_w, pos_h));

43
examples/llava/llava.cpp
View File

@ -410,13 +410,10 @@ void llava_image_embed_free(struct llava_image_embed * embed) {
free(embed);
}
static bool encode_image_with_clip_uhd(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
static bool encode_image_with_clip_uhd(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos, std::pair<int, int> load_image_size) {
// std::vector<clip_image_f32*> img_res_v;
// format VectN x H x W x RGB (N x 448 x 448 x 3)
clip_image_f32 * img_res_v = clip_image_f32_init();
std::pair<int, int> load_image_size;
load_image_size.first = img->nx;
load_image_size.second = img->ny;
uhd_normalize_image_u8_to_f32(ctx_clip, img, img_res_v);
const int64_t t_img_enc_start_us = ggml_time_us();
@ -545,6 +542,34 @@ static bool bicubic_resize(const clip_image_u8 &img, clip_image_u8 &dst, int tar
return true;
}
static clip_image_u8 * only_v2_5_reshape_by_patch(clip_image_u8 * image, int patch_size) {
int width = image->nx;
int height = image->ny;
int num_patches = (height / patch_size) * (width / patch_size);
clip_image_u8 * patch = clip_image_u8_init();
patch->nx = patch_size * num_patches;
patch->ny = patch_size;
patch->buf.resize(3 * patch->nx * patch->ny);
int patch_index = 0;
for (int i = 0; i < height; i += patch_size) {
for (int j = 0; j < width; j += patch_size) {
for (int pi = 0; pi < patch_size; ++pi) {
for (int pj = 0; pj < patch_size; ++pj) {
int input_index = ((i + pi) * width + (j + pj)) * 3;
int output_index = (pi * patch_size * num_patches + patch_index * patch_size + pj) * 3;
patch->buf[output_index] = image->buf[input_index];
patch->buf[output_index+1] = image->buf[input_index+1];
patch->buf[output_index+2] = image->buf[input_index+2];
}
}
patch_index++;
}
}
return patch;
}
// inspired from LLaVA-UHD:
// -> https://arxiv.org/pdf/2403.11703
// -> https://github.com/thunlp/LLaVA-UHD
@ -657,7 +682,11 @@ struct uhd_image_embed * llava_image_embed_make_with_bytes_uhd(struct clip_ctx *
for (size_t j = 0; j < imgs[i].size(); ++j) {
float* image_embed = NULL;
int n_image_pos = 0;
bool image_embed_result = llava_image_embed_make_with_clip_img_uhd(ctx_clip, n_threads, imgs[i][j], &image_embed, &n_image_pos);
int patch_size=14;
std::pair<int, int> load_image_size;
load_image_size.first = imgs[i][j]->nx;
load_image_size.second = imgs[i][j]->ny;
bool image_embed_result = llava_image_embed_make_with_clip_img_uhd(ctx_clip, n_threads, only_v2_5_reshape_by_patch(imgs[i][j], patch_size), &image_embed, &n_image_pos, load_image_size);
if (!image_embed_result) {
LOG_TEE("%s: coulnd't embed the image\n", __func__);
return NULL;
@ -672,7 +701,7 @@ struct uhd_image_embed * llava_image_embed_make_with_bytes_uhd(struct clip_ctx *
return results;
}
bool llava_image_embed_make_with_clip_img_uhd(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
bool llava_image_embed_make_with_clip_img_uhd(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out, std::pair<int, int> load_image_size) {
float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*6); // TODO: base on gridsize/llava model
if (!image_embd) {
LOG_TEE("Unable to allocate memory for image embeddings\n");
@ -680,7 +709,7 @@ bool llava_image_embed_make_with_clip_img_uhd(clip_ctx * ctx_clip, int n_threads
}
int n_img_pos;
if (!encode_image_with_clip_uhd(ctx_clip, n_threads, img, image_embd, &n_img_pos)) {
if (!encode_image_with_clip_uhd(ctx_clip, n_threads, img, image_embd, &n_img_pos, load_image_size)) {
LOG_TEE("%s: cannot encode image, aborting\n", __func__);
free(image_embd);
return false;

2
examples/llava/llava.h
View File

@ -47,7 +47,7 @@ LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed);
/** build an image embed from image file bytes */
LLAVA_API struct uhd_image_embed * llava_image_embed_make_with_bytes_uhd(struct clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img);
/** build an image embed from a path to an image filename */
LLAVA_API bool llava_image_embed_make_with_clip_img_uhd(struct clip_ctx * ctx_clip, int n_threads, const struct clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out);
LLAVA_API bool llava_image_embed_make_with_clip_img_uhd(struct clip_ctx * ctx_clip, int n_threads, const struct clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out, std::pair<int, int> load_image_size = {448, 448});
LLAVA_API bool llava_image_embed_make_with_clip_img_ollama(struct clip_ctx * ctx_clip, int n_threads, const struct clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out);
LLAVA_API struct uhd_image_embed * llava_image_embed_make_with_filename_uhd(struct clip_ctx * ctx_clip, int n_threads, const char * image_path);
LLAVA_API void llava_image_embed_free_uhd(struct uhd_image_embed * embed);