ver4a/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

clip : refactor set input for cgraph + fix qwen2.5vl input (#13136)

Browse source 
* clip : refactor set input for cgraph

* more strict assert

* minicpmv : use clip_n_mmproj_embd instead of copying the same code everywhere

* split qwen2 and qwen2.5 code blocks

* minor style fix
This commit is contained in:
Xuan-Son Nguyen 2025-04-28 12:18:59 +02:00 • committed by GitHub
parent a4c340f974
commit 5fa9e63be8
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 215 additions and 258 deletions
Download patch file Download diff file Expand all files Collapse all files

473
examples/llava/clip.cpp
View file

@ -170,8 +170,8 @@ struct clip_hparams {
std::vector<int32_t> image_grid_pinpoints; std::vector<int32_t> image_grid_pinpoints;
int32_t image_crop_resolution; int32_t image_crop_resolution;
std::unordered_set<int32_t> vision_feature_layer; std::unordered_set<int32_t> vision_feature_layer;
int32_t attn_window_size; int32_t attn_window_size = 0;
int32_t n_wa_pattern; int32_t n_wa_pattern = 0;
}; };
struct clip_layer { struct clip_layer {
@ -325,7 +325,6 @@ struct clip_ctx {
float image_std[3]; float image_std[3];
bool use_gelu = false; bool use_gelu = false;
bool use_silu = false; bool use_silu = false;
int32_t ftype = 1;
gguf_context_ptr ctx_gguf; gguf_context_ptr ctx_gguf;
ggml_context_ptr ctx_data; ggml_context_ptr ctx_data;
@ -776,7 +775,6 @@ static ggml_cgraph * clip_image_build_graph_qwen25vl(clip_ctx * ctx, const clip_
const int image_size_width = imgs.entries[0]->nx; const int image_size_width = imgs.entries[0]->nx;
const int image_size_height = imgs.entries[0]->ny; const int image_size_height = imgs.entries[0]->ny;
const bool use_mrope = ctx->proj_type == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type == PROJECTOR_TYPE_QWEN25VL;
const bool use_window_attn = hparams.n_wa_pattern > 0; const bool use_window_attn = hparams.n_wa_pattern > 0;
const int n_wa_pattern = hparams.n_wa_pattern; const int n_wa_pattern = hparams.n_wa_pattern;
@ -785,10 +783,11 @@ static ggml_cgraph * clip_image_build_graph_qwen25vl(clip_ctx * ctx, const clip_
const int patches_w = image_size_width / patch_size; const int patches_w = image_size_width / patch_size;
const int patches_h = image_size_height / patch_size; const int patches_h = image_size_height / patch_size;
const int num_positions = num_patches + (model.class_embedding ? 1 : 0); const int num_positions = num_patches + (model.class_embedding ? 1 : 0);
const int num_position_ids = use_mrope ? num_positions * 4 : num_positions; const int num_position_ids = num_positions * 4; // m-rope requires 4 dim per position
const int hidden_size = hparams.hidden_size; const int hidden_size = hparams.hidden_size;
const int n_head = hparams.n_head; const int n_head = hparams.n_head;
const int d_head = hidden_size / n_head; const int d_head = hidden_size / n_head;
const int n_layer = hparams.n_layer;
const float eps = hparams.eps; const float eps = hparams.eps;
int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4}; int mrope_sections[4] = {d_head/4, d_head/4, d_head/4, d_head/4};
@ -870,7 +869,7 @@ static ggml_cgraph * clip_image_build_graph_qwen25vl(clip_ctx * ctx, const clip_
} }
// loop over layers // loop over layers
for (int il = 0; il < ctx->max_feature_layer; il++) { for (int il = 0; il < n_layer; il++) {
struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states
// rmsnorm1 // rmsnorm1
@ -1115,15 +1114,8 @@ static ggml_cgraph * clip_image_build_graph_legacy(clip_ctx * ctx, const clip_im
if (ctx->proj_type == PROJECTOR_TYPE_MINICPMV) { if (ctx->proj_type == PROJECTOR_TYPE_MINICPMV) {
int pos_w = image_size_width/patch_size; int pos_w = image_size_width/patch_size;
int pos_h = image_size_height/patch_size; int pos_h = image_size_height/patch_size;
if (ctx->minicpmv_version == 2) { int n_output_dim = clip_n_mmproj_embd(ctx);
pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 4096, pos_w * pos_h, 1); pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_output_dim, pos_w * pos_h, 1);
}
else if (ctx->minicpmv_version == 3) {
pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 3584, pos_w * pos_h, 1);
}
else if (ctx->minicpmv_version == 4) {
pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 3584, pos_w * pos_h, 1);
}
ggml_set_name(pos_embed, "pos_embed"); ggml_set_name(pos_embed, "pos_embed");
ggml_set_input(pos_embed); ggml_set_input(pos_embed);
} }
@ -1461,23 +1453,17 @@ static ggml_cgraph * clip_image_build_graph_legacy(clip_ctx * ctx, const clip_im
} }
{ // attention { // attention
int hidden_size = 4096; int hidden_size = clip_n_mmproj_embd(ctx);
const int d_head = 128; const int d_head = 128;
int n_head = hidden_size/d_head; int n_head = hidden_size/d_head;
int num_query = 96; int num_query = 96;
if (ctx->minicpmv_version == 2) { if (ctx->minicpmv_version == 2) {
hidden_size = 4096;
n_head = hidden_size/d_head;
num_query = 96; num_query = 96;
} }
else if (ctx->minicpmv_version == 3) { else if (ctx->minicpmv_version == 3) {
hidden_size = 3584;
n_head = hidden_size/d_head;
num_query = 64; num_query = 64;
} }
else if (ctx->minicpmv_version == 4) { else if (ctx->minicpmv_version == 4) {
hidden_size = 3584;
n_head = hidden_size/d_head;
num_query = 64; num_query = 64;
} }
@ -1760,6 +1746,8 @@ struct clip_model_loader {
LOG_INF("%s: projector: %s\n", __func__, proj_type.c_str()); LOG_INF("%s: projector: %s\n", __func__, proj_type.c_str());
LOG_INF("%s: has_llava_proj: %d\n", __func__, ctx_clip.has_llava_projector); LOG_INF("%s: has_llava_proj: %d\n", __func__, ctx_clip.has_llava_projector);
LOG_INF("%s: minicpmv_version: %d\n", __func__, ctx_clip.minicpmv_version); LOG_INF("%s: minicpmv_version: %d\n", __func__, ctx_clip.minicpmv_version);
LOG_INF("%s: proj_scale_factor: %d\n", __func__, hparams.proj_scale_factor);
LOG_INF("%s: n_wa_pattern: %d\n", __func__, hparams.n_wa_pattern);
LOG_INF("%s: model size: %.2f MiB\n", __func__, model_size / 1024.0 / 1024.0); LOG_INF("%s: model size: %.2f MiB\n", __func__, model_size / 1024.0 / 1024.0);
LOG_INF("%s: metadata size: %.2f MiB\n", __func__, ggml_get_mem_size(ctx_meta.get()) / 1024.0 / 1024.0); LOG_INF("%s: metadata size: %.2f MiB\n", __func__, ggml_get_mem_size(ctx_meta.get()) / 1024.0 / 1024.0);
} }
@ -3038,15 +3026,43 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
const int patch_size = hparams.patch_size; const int patch_size = hparams.patch_size;
const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size)); const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size));
const int num_positions = num_patches + (model.class_embedding ? 1 : 0); const int num_positions = num_patches + (model.class_embedding ? 1 : 0);
const int pos_w = ctx->load_image_size.width / patch_size; const int pos_w = ctx->load_image_size.width / patch_size;
const int pos_h = ctx->load_image_size.height / patch_size; const int pos_h = ctx->load_image_size.height / patch_size;
const bool use_window_attn = hparams.n_wa_pattern > 0; // for qwen2.5vl const bool use_window_attn = hparams.n_wa_pattern > 0; // for qwen2.5vl
auto get_inp_tensor = [&gf](const char * name) {
struct ggml_tensor * inp = ggml_graph_get_tensor(gf, name);
if (inp == nullptr) {
GGML_ABORT("Failed to get tensor %s", name);
}
if (!(inp->flags & GGML_TENSOR_FLAG_INPUT)) {
GGML_ABORT("Tensor %s is not an input tensor", name);
}
return inp;
};
auto set_input_f32 = [&get_inp_tensor](const char * name, std::vector<float> & values) {
ggml_tensor * cur = get_inp_tensor(name);
GGML_ASSERT(cur->type == GGML_TYPE_F32);
GGML_ASSERT(ggml_nelements(cur) == (int64_t)values.size());
ggml_backend_tensor_set(cur, values.data(), 0, ggml_nbytes(cur));
};
auto set_input_i32 = [&get_inp_tensor](const char * name, std::vector<int32_t> & values) {
ggml_tensor * cur = get_inp_tensor(name);
GGML_ASSERT(cur->type == GGML_TYPE_I32);
GGML_ASSERT(ggml_nelements(cur) == (int64_t)values.size());
ggml_backend_tensor_set(cur, values.data(), 0, ggml_nbytes(cur));
};
// set input pixel values
{ {
struct ggml_tensor * inp_raw = ggml_graph_get_tensor(gf, "inp_raw"); size_t nelem = 0;
std::vector<float> inp_data(ggml_nelements(inp_raw)); for (const auto & img : imgs.entries) {
float * data = inp_data.data(); nelem += img->nx * img->ny * 3;
}
std::vector<float> inp_raw(nelem);
// layout of data (note: the channel dim is unrolled to better visualize the layout): // layout of data (note: the channel dim is unrolled to better visualize the layout):
// //
@ -3065,7 +3081,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
const int n = nx * ny; const int n = nx * ny;
for (int b = 0; b < batch_size; b++) { for (int b = 0; b < batch_size; b++) {
float * batch_entry = data + b * (3*n); float * batch_entry = inp_raw.data() + b * (3*n);
for (int y = 0; y < ny; y++) { for (int y = 0; y < ny; y++) {
for (int x = 0; x < nx; x++) { for (int x = 0; x < nx; x++) {
size_t base_src = 3*(y * nx + x); // idx of the first channel size_t base_src = 3*(y * nx + x); // idx of the first channel
@ -3077,266 +3093,207 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
} }
} }
} }
ggml_backend_tensor_set(inp_raw, data, 0, ggml_nbytes(inp_raw)); set_input_f32("inp_raw", inp_raw);
} }
if (ctx->proj_type == PROJECTOR_TYPE_MINICPMV) { // set input per projector
{ switch (ctx->proj_type) {
// inspired from siglip: case PROJECTOR_TYPE_MINICPMV:
// -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit {
// -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit/blob/d66538faeba44480d0bfaa42145eef26f9423199/modeling_siglip.py#L316 // inspired from siglip:
struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions"); // -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit
std::vector<int> pos_data(ggml_nelements(positions)); // -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit/blob/d66538faeba44480d0bfaa42145eef26f9423199/modeling_siglip.py#L316
int * data = pos_data.data(); std::vector<int32_t> positions(pos_h * pos_w);
int bucket_coords_h[1024]; int bucket_coords_h[1024];
int bucket_coords_w[1024]; int bucket_coords_w[1024];
for (int i = 0; i < pos_h; i++){ for (int i = 0; i < pos_h; i++){
bucket_coords_h[i] = std::floor(70.0*i/pos_h); bucket_coords_h[i] = std::floor(70.0*i/pos_h);
}
for (int i = 0; i < pos_w; i++){
bucket_coords_w[i] = std::floor(70.0*i/pos_w);
}
for (int i = 0, id = 0; i < pos_h; i++){
for (int j = 0; j < pos_w; j++){
data[id++] = bucket_coords_h[i]*70 + bucket_coords_w[j];
} }
} for (int i = 0; i < pos_w; i++){
ggml_backend_tensor_set(positions, data, 0, ggml_nbytes(positions)); bucket_coords_w[i] = std::floor(70.0*i/pos_w);
}
{
// inspired from resampler of Qwen-VL:
// -> 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 embed_dim = 4096;
if (ctx->minicpmv_version == 2) {
embed_dim = 4096;
}
else if (ctx->minicpmv_version == 3) {
embed_dim = 3584;
}
else if (ctx->minicpmv_version == 4) {
embed_dim = 3584;
}
else {
GGML_ABORT("Unknown minicpmv version");
}
// TODO @ngxson : this is very inefficient, can we do this using ggml_sin and ggml_cos?
auto pos_embed_t = get_2d_sincos_pos_embed(embed_dim, std::make_pair(pos_w, pos_h));
std::vector<float> pos_data(ggml_nelements(pos_embed));
float * data = pos_data.data();
for(int i = 0; i < pos_w * pos_h; ++i){
for(int j = 0; j < embed_dim; ++j){
data[i * embed_dim + j] = pos_embed_t[i][j];
} }
} for (int i = 0, id = 0; i < pos_h; i++){
for (int j = 0; j < pos_w; j++){
positions[id++] = bucket_coords_h[i]*70 + bucket_coords_w[j];
}
}
set_input_i32("positions", positions);
ggml_backend_tensor_set(pos_embed, data, 0, ggml_nbytes(pos_embed)); // inspired from resampler of Qwen-VL:
} // -> https://huggingface.co/Qwen/Qwen-VL/tree/main
} // -> https://huggingface.co/Qwen/Qwen-VL/blob/0547ed36a86561e2e42fecec8fd0c4f6953e33c4/visual.py#L23
else { int embed_dim = clip_n_mmproj_embd(ctx);
// non-minicpmv models
if (ctx->proj_type == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type == PROJECTOR_TYPE_QWEN25VL) { // TODO @ngxson : this is very inefficient, can we do this using ggml_sin and ggml_cos?
// pw * ph = number of tokens output by ViT after apply patch merger auto pos_embed_t = get_2d_sincos_pos_embed(embed_dim, std::make_pair(pos_w, pos_h));
// ipw * ipw = number of vision token been processed inside ViT
const int merge_ratio = 2;
const int pw = image_size_width / patch_size / merge_ratio;
const int ph = image_size_height / patch_size / merge_ratio;
const int ipw = image_size_width / patch_size;
const int iph = image_size_height / patch_size;
std::vector<int> idx (ph * pw); std::vector<float> pos_embed(embed_dim * pos_w * pos_h);
std::vector<int> inv_idx(ph * pw); for(int i = 0; i < pos_w * pos_h; ++i){
for(int j = 0; j < embed_dim; ++j){
pos_embed[i * embed_dim + j] = pos_embed_t[i][j];
}
}
if (use_window_attn) { set_input_f32("pos_embed", pos_embed);
const int attn_window_size = 112; } break;
struct ggml_tensor * window_idx = ggml_graph_get_tensor(gf, "window_idx"); case PROJECTOR_TYPE_QWEN2VL:
struct ggml_tensor * inv_window_idx = ggml_graph_get_tensor(gf, "inv_window_idx"); {
struct ggml_tensor * window_mask = ggml_graph_get_tensor(gf, "window_mask"); const int merge_ratio = 2;
const int pw = image_size_width / patch_size;
const int grid_window = attn_window_size / patch_size / merge_ratio; const int ph = image_size_height / patch_size;
int dst = 0; std::vector<int> positions(num_positions * 4);
// [num_vision_tokens, num_vision_tokens] attention mask tensor int ptr = 0;
std::vector<float> mask(pow(ipw * iph, 2), std::numeric_limits<float>::lowest()); for (int y = 0; y < ph; y += merge_ratio) {
int mask_row = 0; for (int x = 0; x < pw; x += merge_ratio) {
for (int dy = 0; dy < 2; dy++) {
for (int y = 0; y < ph; y += grid_window) for (int dx = 0; dx < 2; dx++) {
{ positions[ ptr] = y + dy;
for (int x = 0; x < pw; x += grid_window) positions[ num_patches + ptr] = x + dx;
{ positions[2 * num_patches + ptr] = y + dy;
const int win_h = std::min(grid_window, ph - y); positions[3 * num_patches + ptr] = x + dx;
const int win_w = std::min(grid_window, pw - x); ptr++;
const int dst_0 = dst;
// group all tokens belong to the same window togather (to a continue range)
for (int dy = 0; dy < win_h; dy++) {
for (int dx = 0; dx < win_w; dx++) {
const int src = (y + dy) * pw + (x + dx);
assert(src < (int)idx.size());
assert(dst < (int)inv_idx.size());
idx [src] = dst;
inv_idx[dst] = src;
dst++;
} }
} }
for (int r=0; r < win_h * win_w * merge_ratio * merge_ratio; r++) {
int row_offset = mask_row * (ipw * iph);
std::fill(
mask.begin() + row_offset + (dst_0 * merge_ratio * merge_ratio),
mask.begin() + row_offset + (dst * merge_ratio * merge_ratio),
0.0);
mask_row++;
}
} }
} }
ggml_backend_tensor_set(window_idx, idx.data(), 0, ggml_nbytes(window_idx)); set_input_i32("positions", positions);
ggml_backend_tensor_set(inv_window_idx, inv_idx.data(), 0, ggml_nbytes(inv_window_idx)); } break;
ggml_backend_tensor_set(window_mask, mask.data(), 0, ggml_nbytes(window_mask)); case PROJECTOR_TYPE_QWEN25VL:
} else {
std::iota(idx.begin(), idx.end(), 0);
std::iota(inv_idx.begin(), inv_idx.end(), 0);
}
struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
const int mpow = merge_ratio * merge_ratio;
std::vector<int> positions_data(ggml_nelements(positions));
int * data = positions_data.data();
int ptr = 0;
for (int y = 0; y < iph; y += merge_ratio)
{ {
for (int x = 0; x < ipw; x += merge_ratio) // pw * ph = number of tokens output by ViT after apply patch merger
{ // ipw * ipw = number of vision token been processed inside ViT
for (int dy = 0; dy < 2; dy++) { const int merge_ratio = 2;
for (int dx = 0; dx < 2; dx++) { const int pw = image_size_width / patch_size / merge_ratio;
auto remap = idx[ptr / mpow]; const int ph = image_size_height / patch_size / merge_ratio;
remap = remap * mpow + (ptr % mpow); const int ipw = image_size_width / patch_size;
const int iph = image_size_height / patch_size;
data[ remap] = y + dy; std::vector<int> idx (ph * pw);
data[ num_patches + remap] = x + dx; std::vector<int> inv_idx(ph * pw);
data[2 * num_patches + remap] = y + dy;
data[3 * num_patches + remap] = x + dx; if (use_window_attn) {
ptr++; const int attn_window_size = 112;
const int grid_window = attn_window_size / patch_size / merge_ratio;
int dst = 0;
// [num_vision_tokens, num_vision_tokens] attention mask tensor
std::vector<float> mask(pow(ipw * iph, 2), std::numeric_limits<float>::lowest());
int mask_row = 0;
for (int y = 0; y < ph; y += grid_window) {
for (int x = 0; x < pw; x += grid_window) {
const int win_h = std::min(grid_window, ph - y);
const int win_w = std::min(grid_window, pw - x);
const int dst_0 = dst;
// group all tokens belong to the same window togather (to a continue range)
for (int dy = 0; dy < win_h; dy++) {
for (int dx = 0; dx < win_w; dx++) {
const int src = (y + dy) * pw + (x + dx);
GGML_ASSERT(src < (int)idx.size());
GGML_ASSERT(dst < (int)inv_idx.size());
idx [src] = dst;
inv_idx[dst] = src;
dst++;
}
}
for (int r=0; r < win_h * win_w * merge_ratio * merge_ratio; r++) {
int row_offset = mask_row * (ipw * iph);
std::fill(
mask.begin() + row_offset + (dst_0 * merge_ratio * merge_ratio),
mask.begin() + row_offset + (dst * merge_ratio * merge_ratio),
0.0);
mask_row++;
}
}
}
set_input_i32("window_idx", idx);
set_input_i32("inv_window_idx", inv_idx);
set_input_f32("window_mask", mask);
} else {
for (int i = 0; i < ph * pw; i++) {
idx[i] = i;
}
}
const int mpow = merge_ratio * merge_ratio;
std::vector<int> positions(num_positions * 4);
int ptr = 0;
for (int y = 0; y < iph; y += merge_ratio) {
for (int x = 0; x < ipw; x += merge_ratio) {
for (int dy = 0; dy < 2; dy++) {
for (int dx = 0; dx < 2; dx++) {
auto remap = idx[ptr / mpow];
remap = (remap * mpow) + (ptr % mpow);
positions[ remap] = y + dy;
positions[ num_patches + remap] = x + dx;
positions[2 * num_patches + remap] = y + dy;
positions[3 * num_patches + remap] = x + dx;
ptr++;
}
} }
} }
} }
}
ggml_backend_tensor_set(positions, data, 0, ggml_nbytes(positions)); set_input_i32("positions", positions);
} } break;
else if (ctx->proj_type == PROJECTOR_TYPE_GEMMA3) { case PROJECTOR_TYPE_PIXTRAL:
// do nothing {
} // set the 2D positions
else if (ctx->proj_type == PROJECTOR_TYPE_IDEFICS3) { int n_patches_per_col = image_size_width / patch_size;
// do nothing std::vector<int> pos_data(num_positions);
} // dimension H
else if (ctx->proj_type == PROJECTOR_TYPE_PIXTRAL) { for (int i = 0; i < num_positions; i++) {
// set the 2D positions pos_data[i] = i / n_patches_per_col;
int n_patches_per_col = image_size_width / patch_size; }
std::vector<int> pos_data(num_positions); set_input_i32("pos_h", pos_data);
struct ggml_tensor * pos; // dimension W
// dimension H for (int i = 0; i < num_positions; i++) {
pos = ggml_graph_get_tensor(gf, "pos_h"); pos_data[i] = i % n_patches_per_col;
for (int i = 0; i < num_positions; i++) { }
pos_data[i] = i / n_patches_per_col; set_input_i32("pos_w", pos_data);
} } break;
ggml_backend_tensor_set(pos, pos_data.data(), 0, ggml_nbytes(pos)); case PROJECTOR_TYPE_GLM_EDGE:
// dimension W {
pos = ggml_graph_get_tensor(gf, "pos_w");
for (int i = 0; i < num_positions; i++) {
pos_data[i] = i % n_patches_per_col;
}
ggml_backend_tensor_set(pos, pos_data.data(), 0, ggml_nbytes(pos));
}
else {
// llava and other models // llava and other models
struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions"); std::vector<int32_t> positions(num_positions);
int* positions_data = (int*)malloc(ggml_nbytes(positions));
for (int i = 0; i < num_positions; i++) { for (int i = 0; i < num_positions; i++) {
positions_data[i] = i; positions[i] = i;
} }
ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions)); set_input_i32("positions", positions);
free(positions_data); } break;
case PROJECTOR_TYPE_MLP:
case PROJECTOR_TYPE_MLP_NORM:
case PROJECTOR_TYPE_LDP:
case PROJECTOR_TYPE_LDPV2:
{
// llava and other models
std::vector<int32_t> positions(num_positions);
for (int i = 0; i < num_positions; i++) {
positions[i] = i;
}
set_input_i32("positions", positions);
if (ctx->proj_type != PROJECTOR_TYPE_GLM_EDGE) {
struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
// The patches vector is used to get rows to index into the embeds with; // The patches vector is used to get rows to index into the embeds with;
// we should skip dim 0 only if we have CLS to avoid going out of bounds // we should skip dim 0 only if we have CLS to avoid going out of bounds
// when retrieving the rows. // when retrieving the rows.
int patch_offset = model.class_embedding ? 1 : 0; int patch_offset = model.class_embedding ? 1 : 0;
int* patches_data = (int*)malloc(ggml_nbytes(patches)); std::vector<int32_t> patches(num_patches);
for (int i = 0; i < num_patches; i++) { for (int i = 0; i < num_patches; i++) {
patches_data[i] = i + patch_offset; patches[i] = i + patch_offset;
} }
ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches)); set_input_i32("patches", patches);
free(patches_data); } break;
} case PROJECTOR_TYPE_GEMMA3:
} case PROJECTOR_TYPE_IDEFICS3:
}
if (use_window_attn && (ctx->proj_type == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type == PROJECTOR_TYPE_QWEN25VL)) {
struct ggml_tensor * window_idx = ggml_graph_get_tensor(gf, "window_idx");
struct ggml_tensor * inv_window_idx = ggml_graph_get_tensor(gf, "inv_window_idx");
struct ggml_tensor * window_mask = ggml_graph_get_tensor(gf, "window_mask");
const int merge_ratio = 2;
const int attn_window_size = 112;
const int pw = image_size_width / patch_size / merge_ratio;
const int ph = image_size_height / patch_size / merge_ratio;
const int grid_window = attn_window_size / patch_size / merge_ratio;
const int ipw = image_size_width / patch_size;
const int iph = image_size_height / patch_size;
/*
pw * ph = number of tokens output by ViT after apply patch merger
ipw * ipw = number of vision token been processed inside ViT
*/
std::vector<int> idx(ph * pw);
std::vector<int> inv_idx(ph * pw);
int dst = 0;
// [num_vision_tokens, num_vision_tokens] attention mask tensor
std::vector<float> mask(pow(ipw * iph, 2), std::numeric_limits<float>::lowest());
int mask_row = 0;
for (int y = 0; y < ph; y+=grid_window)
{
for (int x = 0; x < pw; x+=grid_window)
{ {
const int win_h = std::min(grid_window, ph - y); // do nothing
const int win_w = std::min(grid_window, pw - x); } break;
const int dst_0 = dst; default:
// group all tokens belong to the same window togather (to a continue range) GGML_ABORT("Unknown projector type");
for (int dy = 0; dy < win_h; dy++) {
for (int dx = 0; dx < win_w; dx++) {
const int src = (y + dy) * pw + (x + dx);
assert(src < (int)idx.size());
assert(dst < (int)inv_idx.size());
idx[src] = dst;
inv_idx[dst] = src;
dst++;
}
}
for (int r=0; r < win_h * win_w * merge_ratio * merge_ratio; r++) {
int row_offset = mask_row * (ipw * iph);
std::fill(
mask.begin() + row_offset + (dst_0 * merge_ratio * merge_ratio),
mask.begin() + row_offset + (dst * merge_ratio * merge_ratio),
0.0);
mask_row++;
}
}
}
ggml_backend_tensor_set(window_idx, idx.data(), 0, ggml_nbytes(window_idx));
ggml_backend_tensor_set(inv_window_idx, inv_idx.data(), 0, ggml_nbytes(inv_window_idx));
ggml_backend_tensor_set(window_mask, mask.data(), 0, ggml_nbytes(window_mask));
} }
ggml_backend_cpu_set_n_threads(ctx->backend_cpu, n_threads); ggml_backend_cpu_set_n_threads(ctx->backend_cpu, n_threads);
@ -3537,7 +3494,7 @@ bool clip_is_glm(const struct clip_ctx * ctx) {
} }
bool clip_is_qwen2vl(const struct clip_ctx * ctx) { bool clip_is_qwen2vl(const struct clip_ctx * ctx) {
return ctx->proj_type == PROJECTOR_TYPE_QWEN2VL; return ctx->proj_type == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type == PROJECTOR_TYPE_QWEN25VL;
} }
bool clip_is_llava(const struct clip_ctx * ctx) { bool clip_is_llava(const struct clip_ctx * ctx) {