mirrors/ollama-for-amd
1
0
Fork 0
mirror of https://github.com/likelovewant/ollama-for-amd.git synced 2025-12-21 22:33:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

Update GGML to b6646 (#12245)

Browse Source 
Notable EOLs with this change:
- MacOS v12 and v13 are no longer supported (v14+ required)
- AMD gfx900 and gfx906 are no longer supported
This commit is contained in:
Daniel Hiltgen
2025-10-02 14:47:10 -07:00
committed by GitHub
parent fdb109469f
commit c68f367ef6
326 changed files with 30615 additions and 20624 deletions
Download Patch File Download Diff File Expand all files Collapse all files

429
llama/llama.cpp/tools/mtmd/clip.cpp vendored
View File

@@ -214,6 +214,7 @@ struct clip_hparams {
// legacy
bool has_llava_projector = false;
int minicpmv_version = 0;
int32_t minicpmv_query_num = 0; // MiniCPM-V query number
};
struct clip_layer {
@@ -277,6 +278,7 @@ struct clip_model {
// LLaVA projection
ggml_tensor * mm_input_norm_w = nullptr;
ggml_tensor * mm_input_norm_b = nullptr;
ggml_tensor * mm_0_w = nullptr;
ggml_tensor * mm_0_b = nullptr;
ggml_tensor * mm_2_w = nullptr;
@@ -417,6 +419,7 @@ struct clip_ctx {
}
if (!backend) {
backend = ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_GPU, nullptr);
backend = backend ? backend : ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_IGPU, nullptr);
}
}
@@ -500,11 +503,17 @@ struct clip_graph {
ggml_cgraph * build_siglip() {
ggml_tensor * inp = build_inp();
ggml_tensor * learned_pos_embd = model.position_embeddings;
if (ctx->proj_type() == PROJECTOR_TYPE_LFM2) {
learned_pos_embd = resize_position_embeddings();
}
ggml_tensor * cur = build_vit(
inp, n_patches,
NORM_TYPE_NORMAL,
hparams.ffn_op,
model.position_embeddings,
learned_pos_embd,
nullptr);
if (ctx->proj_type() == PROJECTOR_TYPE_GEMMA3) {
@@ -513,8 +522,8 @@ struct clip_graph {
const int patches_per_image = n_patches_x;
const int kernel_size = hparams.proj_scale_factor;
cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur));
cur = ggml_reshape_4d(ctx0, cur, patches_per_image, patches_per_image, n_embd, batch_size);
cur = ggml_transpose(ctx0, cur);
cur = ggml_cont_4d(ctx0, cur, patches_per_image, patches_per_image, n_embd, batch_size);
// doing a pool2d to reduce the number of output tokens
cur = ggml_pool_2d(ctx0, cur, GGML_OP_POOL_AVG, kernel_size, kernel_size, kernel_size, kernel_size, 0, 0);
@@ -531,29 +540,27 @@ struct clip_graph {
cur);
} else if (ctx->proj_type() == PROJECTOR_TYPE_IDEFICS3) {
// pixel_shuffle
// https://github.com/huggingface/transformers/blob/0a950e0bbe1ed58d5401a6b547af19f15f0c195e/src/transformers/models/idefics3/modeling_idefics3.py#L578
const int scale_factor = model.hparams.proj_scale_factor;
const int n_embd = cur->ne[0];
const int seq = cur->ne[1];
const int bsz = 1; // batch size, always 1 for now since we don't support batching
const int height = std::sqrt(seq);
const int width = std::sqrt(seq);
GGML_ASSERT(scale_factor != 0);
cur = ggml_reshape_4d(ctx0, cur, n_embd * scale_factor, width / scale_factor, height, bsz);
cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
cur = ggml_reshape_4d(ctx0, ggml_cont(ctx0, cur),
n_embd * scale_factor * scale_factor,
height / scale_factor,
width / scale_factor,
bsz);
cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
cur = ggml_reshape_3d(ctx0, ggml_cont(ctx0, cur),
n_embd * scale_factor * scale_factor,
seq / (scale_factor * scale_factor),
bsz);
cur = build_patch_merge_permute(cur, scale_factor);
cur = ggml_mul_mat(ctx0, model.projection, cur);
} else if (ctx->proj_type() == PROJECTOR_TYPE_LFM2) {
// pixel unshuffle block
const int scale_factor = model.hparams.proj_scale_factor;
cur = build_patch_merge_permute(cur, scale_factor);
// projection
cur = ggml_norm(ctx0, cur, 1e-5); // default nn.LayerNorm
cur = ggml_mul(ctx0, cur, model.mm_input_norm_w);
cur = ggml_add(ctx0, cur, model.mm_input_norm_b);
cur = ggml_mul_mat(ctx0, model.mm_1_w, cur);
cur = ggml_add(ctx0, cur, model.mm_1_b);
cur = ggml_gelu(ctx0, cur);
cur = ggml_mul_mat(ctx0, model.mm_2_w, cur);
cur = ggml_add(ctx0, cur, model.mm_2_b);
} else {
GGML_ABORT("SigLIP: Unsupported projector type");
}
@@ -681,15 +688,15 @@ struct clip_graph {
auto inp_1 = ggml_conv_2d(ctx0, model.patch_embeddings_1, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
inp = ggml_add(ctx0, inp, inp_1);
inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 2, 0, 3)); // [w, h, c, b] -> [c, w, h, b]
inp = ggml_reshape_4d(
inp = ggml_permute(ctx0, inp, 1, 2, 0, 3); // [w, h, c, b] -> [c, w, h, b]
inp = ggml_cont_4d(
ctx0, inp,
n_embd * 2, n_patches_x / 2, n_patches_y, batch_size);
inp = ggml_reshape_4d(
ctx0, inp,
n_embd * 2, n_patches_x / 2, 2, batch_size * (n_patches_y / 2));
inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 0, 2, 1, 3));
inp = ggml_reshape_3d(
inp = ggml_permute(ctx0, inp, 0, 2, 1, 3);
inp = ggml_cont_3d(
ctx0, inp,
n_embd, n_patches_x * n_patches_y, batch_size);
}
@@ -879,21 +886,8 @@ struct clip_graph {
int n_embd = clip_n_mmproj_embd(ctx);
const int d_head = 128;
int n_head = n_embd/d_head;
int num_query = 96;
if (ctx->model.hparams.minicpmv_version == 2) {
// MiniCPM-V 2.5
num_query = 96;
} else if (ctx->model.hparams.minicpmv_version == 3) {
// MiniCPM-V 2.6
num_query = 64;
} else if (ctx->model.hparams.minicpmv_version == 4) {
// MiniCPM-o 2.6
num_query = 64;
} else if (ctx->model.hparams.minicpmv_version == 5) {
// MiniCPM-V 4.0
num_query = 64;
}
// Use actual config value if available, otherwise fall back to hardcoded values
int num_query = ctx->model.hparams.minicpmv_query_num;
ggml_tensor * Q = ggml_add(ctx0,
ggml_mul_mat(ctx0, model.mm_model_attn_q_w, q),
model.mm_model_attn_q_b);
@@ -967,14 +961,14 @@ struct clip_graph {
GGML_ASSERT(scale_factor > 0);
cur = ggml_reshape_4d(ctx0, cur, n_embd * scale_factor, height / scale_factor, width, bsz);
cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
cur = ggml_reshape_4d(ctx0, ggml_cont(ctx0, cur),
cur = ggml_cont_4d(ctx0, cur,
n_embd * scale_factor * scale_factor,
height / scale_factor,
width / scale_factor,
bsz);
cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
// flatten to 2D
cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, cur),
cur = ggml_cont_2d(ctx0, cur,
n_embd * scale_factor * scale_factor,
cur->ne[1] * cur->ne[2]);
}
@@ -1060,14 +1054,14 @@ struct clip_graph {
n_patches_y,
bsz);
cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
cur = ggml_reshape_4d(ctx0, ggml_cont(ctx0, cur),
cur = ggml_cont_4d(ctx0, cur,
n_embd * scale_factor * scale_factor,
n_patches_x / scale_factor,
n_patches_y / scale_factor,
bsz);
cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
//cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
// flatten to 2D
cur = ggml_reshape_2d(ctx0, ggml_cont(ctx0, cur),
cur = ggml_cont_2d(ctx0, cur,
n_embd * scale_factor * scale_factor,
n_patches / scale_factor / scale_factor);
cb(cur, "pixel_shuffle", -1);
@@ -1092,6 +1086,67 @@ struct clip_graph {
return gf;
}
ggml_cgraph * build_kimivl() {
// 2D input positions
ggml_tensor * pos_h = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
ggml_set_name(pos_h, "pos_h");
ggml_set_input(pos_h);
ggml_tensor * pos_w = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
ggml_set_name(pos_w, "pos_w");
ggml_set_input(pos_w);
ggml_tensor * learned_pos_embd = resize_position_embeddings();
// build ViT with 2D position embeddings
auto add_pos = [&](ggml_tensor * cur, const clip_layer &) {
// first half is X axis and second half is Y axis
return build_rope_2d(ctx0, cur, pos_w, pos_h, hparams.rope_theta, false);
};
ggml_tensor * inp = build_inp();
ggml_tensor * cur = build_vit(
inp, n_patches,
NORM_TYPE_NORMAL,
hparams.ffn_op,
learned_pos_embd,
add_pos);
cb(cur, "vit_out", -1);
{
// patch_merger
const int scale_factor = model.hparams.proj_scale_factor;
cur = build_patch_merge_permute(cur, scale_factor);
// projection norm
int proj_inp_dim = cur->ne[0];
cur = ggml_view_2d(ctx0, cur,
n_embd, cur->ne[1] * scale_factor * scale_factor,
ggml_row_size(cur->type, n_embd), 0);
cur = ggml_norm(ctx0, cur, 1e-5); // default nn.LayerNorm
cur = ggml_mul(ctx0, cur, model.mm_input_norm_w);
cur = ggml_add(ctx0, cur, model.mm_input_norm_b);
cur = ggml_view_2d(ctx0, cur,
proj_inp_dim, cur->ne[1] / scale_factor / scale_factor,
ggml_row_size(cur->type, proj_inp_dim), 0);
cb(cur, "proj_inp_normed", -1);
// projection mlp
cur = ggml_mul_mat(ctx0, model.mm_1_w, cur);
cur = ggml_add(ctx0, cur, model.mm_1_b);
cur = ggml_gelu(ctx0, cur);
cur = ggml_mul_mat(ctx0, model.mm_2_w, cur);
cur = ggml_add(ctx0, cur, model.mm_2_b);
cb(cur, "proj_out", -1);
}
// build the graph
ggml_build_forward_expand(gf, cur);
return gf;
}
// this graph is used by llava, granite and glm
// due to having embedding_stack (used by granite), we cannot reuse build_vit
ggml_cgraph * build_llava() {
@@ -1300,8 +1355,8 @@ struct clip_graph {
ggml_tensor * block_1 = nullptr;
{
// transpose from [1, 576, 2048] --> [1, 2048, 576] --> [1, 2048, 24, 24]
mlp_3 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_3, 1, 0, 2, 3));
mlp_3 = ggml_reshape_4d(ctx0, mlp_3, n_patch, n_patch, mlp_3->ne[1], mlp_3->ne[2]);
mlp_3 = ggml_permute(ctx0, mlp_3, 1, 0, 2, 3);
mlp_3 = ggml_cont_4d(ctx0, mlp_3, n_patch, n_patch, mlp_3->ne[1], mlp_3->ne[2]);
// stride = 1, padding = 1, bias is nullptr
block_1 = ggml_conv_2d_dw(ctx0, model.mm_model_block_1_block_0_0_w, mlp_3, 1, 1, 1, 1, 1, 1);
@@ -1406,9 +1461,9 @@ struct clip_graph {
mlp_2 = ggml_add(ctx0, mlp_2, model.mm_model_mlp_2_b);
// mlp_2 ne = [2048, 576, 1, 1]
// // AVG Pool Layer 2*2, strides = 2
mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 0, 2, 3));
mlp_2 = ggml_permute(ctx0, mlp_2, 1, 0, 2, 3);
// mlp_2 ne = [576, 2048, 1, 1]
mlp_2 = ggml_reshape_4d(ctx0, mlp_2, n_patch, n_patch, mlp_2->ne[1], mlp_2->ne[2]);
mlp_2 = ggml_cont_4d(ctx0, mlp_2, n_patch, n_patch, mlp_2->ne[1], mlp_2->ne[2]);
// mlp_2 ne [24, 24, 2048, 1]
mlp_2 = ggml_pool_2d(ctx0, mlp_2, GGML_OP_POOL_AVG, 2, 2, 2, 2, 0, 0);
// weight ne = [3, 3, 2048, 1]
@@ -1428,8 +1483,8 @@ struct clip_graph {
// glm projector
else if (ctx->proj_type() == PROJECTOR_TYPE_GLM_EDGE) {
size_t gridsz = (size_t)sqrt(embeddings->ne[1]);
embeddings = ggml_cont(ctx0, ggml_permute(ctx0,embeddings,1,0,2,3));
embeddings = ggml_reshape_3d(ctx0, embeddings, gridsz, gridsz, embeddings->ne[1]);
embeddings = ggml_permute(ctx0,embeddings,1,0,2,3);
embeddings = ggml_cont_3d(ctx0, embeddings, gridsz, gridsz, embeddings->ne[1]);
embeddings = ggml_conv_2d(ctx0, model.mm_model_adapter_conv_w, embeddings, 2, 2, 0, 0, 1, 1);
embeddings = ggml_reshape_3d(ctx0, embeddings,embeddings->ne[0]*embeddings->ne[1] , embeddings->ne[2], batch_size);
embeddings = ggml_cont(ctx0, ggml_permute(ctx0,embeddings, 1, 0, 2, 3));
@@ -1585,6 +1640,29 @@ private:
}
}
// siglip2 naflex
ggml_tensor * resize_position_embeddings() {
ggml_tensor * pos_embd = model.position_embeddings;
const int height = img.ny / patch_size;
const int width = img.nx / patch_size;
const uint32_t mode = GGML_SCALE_MODE_BILINEAR;
const int n_per_side = (int)std::sqrt(pos_embd->ne[1]);
GGML_ASSERT(pos_embd);
if (height == n_per_side && width == n_per_side) {
return pos_embd;
}
pos_embd = ggml_reshape_3d(ctx0, pos_embd, n_embd, n_per_side, n_per_side); // -> (n_embd, n_per_side, n_per_side)
pos_embd = ggml_permute(ctx0, pos_embd, 2, 0, 1, 3); // -> (n_per_side, n_per_side, n_embd)
pos_embd = ggml_interpolate(ctx0, pos_embd, width, height, n_embd, 1, mode); // -> (width, height, n_embd)
pos_embd = ggml_permute(ctx0, pos_embd, 1, 2, 0, 3); // -> (n_embd, width, height)
pos_embd = ggml_cont_2d(ctx0, pos_embd, n_embd, width * height); // -> (n_embd, width * height)
return pos_embd;
}
// build vision transformer (ViT) cgraph
// this function should cover most of the models
// if your model has specific features, you should probably duplicate this function
@@ -1963,7 +2041,6 @@ private:
ggml_row_size(cur->type, n_dim),
ggml_row_size(cur->type, n_dim*n_head),
n_dim/2 * ggml_element_size(cur));
second = ggml_cont(ctx0, second); // copy, because ggml_rope don't play well with non-contiguous tensors
second = ggml_rope_ext(
ctx0,
second,
@@ -1980,6 +2057,39 @@ private:
return cur;
}
// aka pixel_shuffle / pixel_unshuffle / patch_merger (Kimi-VL)
// support dynamic resolution
ggml_tensor * build_patch_merge_permute(ggml_tensor * cur, int scale_factor) {
GGML_ASSERT(scale_factor > 1);
const int n_embd = cur->ne[0];
int width = img.nx / patch_size;
int height = img.ny / patch_size;
// pad width and height to factor
const int64_t pad_width = CLIP_ALIGN(width, scale_factor) - width;
const int64_t pad_height = CLIP_ALIGN(height, scale_factor) - height;
cur = ggml_reshape_3d(ctx0, cur, n_embd, width, height);
if (pad_width || pad_height) {
cur = ggml_pad(ctx0, cur, 0, pad_width, pad_height, 0);
width += pad_width;
height += pad_height;
}
// unshuffle h
cur = ggml_reshape_3d(ctx0, cur, n_embd * scale_factor, width / scale_factor, height);
cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
// unshuffle w
cur = ggml_cont_3d(ctx0, cur, n_embd * scale_factor * scale_factor, height / scale_factor, width / scale_factor);
cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
cur = ggml_cont_2d(ctx0, cur, cur->ne[0], cur->ne[1] * cur->ne[2]);
cb(cur, "pixel_shuffle", -1);
return cur;
}
};
static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch & imgs) {
@@ -1991,6 +2101,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
switch (ctx->proj_type()) {
case PROJECTOR_TYPE_GEMMA3:
case PROJECTOR_TYPE_IDEFICS3:
case PROJECTOR_TYPE_LFM2:
{
res = graph.build_siglip();
} break;
@@ -2021,6 +2132,10 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
{
res = graph.build_whisper_enc();
} break;
case PROJECTOR_TYPE_KIMIVL:
{
res = graph.build_kimivl();
} break;
default:
{
res = graph.build_llava();
@@ -2151,7 +2266,21 @@ struct clip_model_loader {
get_u32(KEY_PATCH_SIZE, hparams.patch_size);
get_u32(KEY_IMAGE_CROP_RESOLUTION, hparams.image_crop_resolution, false);
get_i32(KEY_MINICPMV_VERSION, hparams.minicpmv_version, false); // legacy
get_u32(KEY_MINICPMV_QUERY_NUM, hparams.minicpmv_query_num, false);
if (hparams.minicpmv_query_num == 0) {
// Fallback to hardcoded values for legacy models
if (hparams.minicpmv_version == 3) {
hparams.minicpmv_query_num = 64;
} else if (hparams.minicpmv_version == 4) {
hparams.minicpmv_query_num = 64;
} else if (hparams.minicpmv_version == 5) {
hparams.minicpmv_query_num = 64;
} else if (hparams.minicpmv_version == 6) {
hparams.minicpmv_query_num = 64;
} else {
hparams.minicpmv_query_num = 96;
}
}
} else if (is_audio) {
get_u32(KEY_A_NUM_MEL_BINS, hparams.n_mel_bins);
@@ -2243,6 +2372,7 @@ struct clip_model_loader {
}
} break;
case PROJECTOR_TYPE_IDEFICS3:
case PROJECTOR_TYPE_LFM2:
case PROJECTOR_TYPE_INTERNVL:
{
get_u32(KEY_PROJ_SCALE_FACTOR, hparams.proj_scale_factor, false);
@@ -2256,6 +2386,12 @@ struct clip_model_loader {
hparams.image_size = 1024;
get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.spatial_merge_size, false);
} break;
case PROJECTOR_TYPE_KIMIVL:
{
hparams.rope_theta = 10000.0f;
hparams.warmup_image_size = hparams.patch_size * 8;
get_u32(KEY_PROJ_SCALE_FACTOR, hparams.proj_scale_factor, false);
} break;
case PROJECTOR_TYPE_GEMMA3:
{
// default value (used by all model sizes in gemma 3 family)
@@ -2420,7 +2556,20 @@ struct clip_model_loader {
// some models already exported with legacy (incorrect) naming which is quite messy, let's fix it here
// note: Qwen model converted from the old surgery script has n_ff = 0, so we cannot use n_ff to check!
if (layer.ff_up_w && layer.ff_down_w && layer.ff_down_w->ne[0] == hparams.n_embd) {
bool is_ffn_swapped = (
// only old models need this fix
model.proj_type == PROJECTOR_TYPE_MLP
|| model.proj_type == PROJECTOR_TYPE_MLP_NORM
|| model.proj_type == PROJECTOR_TYPE_LDP
|| model.proj_type == PROJECTOR_TYPE_LDPV2
|| model.proj_type == PROJECTOR_TYPE_QWEN2VL
|| model.proj_type == PROJECTOR_TYPE_QWEN25VL
|| model.proj_type == PROJECTOR_TYPE_GLM_EDGE
|| model.proj_type == PROJECTOR_TYPE_GEMMA3
|| model.proj_type == PROJECTOR_TYPE_IDEFICS3
|| model.proj_type == PROJECTOR_TYPE_MINICPMV
) && layer.ff_up_w && layer.ff_down_w && layer.ff_down_w->ne[0] == hparams.n_embd;
if (is_ffn_swapped) {
// swap up and down weights
ggml_tensor * tmp = layer.ff_up_w;
layer.ff_up_w = layer.ff_down_w;
@@ -2429,6 +2578,9 @@ struct clip_model_loader {
tmp = layer.ff_up_b;
layer.ff_up_b = layer.ff_down_b;
layer.ff_down_b = tmp;
if (il == 0) {
LOG_WRN("%s: ffn up/down are swapped\n", __func__);
}
}
}
@@ -2546,6 +2698,16 @@ struct clip_model_loader {
{
model.projection = get_tensor(TN_MM_PROJECTOR);
} break;
case PROJECTOR_TYPE_LFM2:
case PROJECTOR_TYPE_KIMIVL:
{
model.mm_input_norm_w = get_tensor(TN_MM_INP_NORM);
model.mm_input_norm_b = get_tensor(TN_MM_INP_NORM_B);
model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 1, "weight"));
model.mm_1_b = get_tensor(string_format(TN_LLAVA_PROJ, 1, "bias"));
model.mm_2_w = get_tensor(string_format(TN_LLAVA_PROJ, 2, "weight"));
model.mm_2_b = get_tensor(string_format(TN_LLAVA_PROJ, 2, "bias"));
} break;
case PROJECTOR_TYPE_PIXTRAL:
{
model.mm_1_w = get_tensor(string_format(TN_LLAVA_PROJ, 1, "weight"));
@@ -2944,7 +3106,7 @@ struct image_manipulation {
dst.buf.resize(3 * target_width * target_height);
float Cc;
float C[5];
float C[5] = {};
float d0, d2, d3, a0, a1, a2, a3;
int i, j, k, jj;
int x, y;
@@ -3467,6 +3629,45 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
res_imgs->grid_y = inst.grid_size.height;
return true;
} else if ( ctx->proj_type() == PROJECTOR_TYPE_LFM2
|| ctx->proj_type() == PROJECTOR_TYPE_KIMIVL
) {
GGML_ASSERT(params.proj_scale_factor);
// smart resize
const int width = img->nx;
const int height = img->ny;
const int total_factor = params.patch_size * params.proj_scale_factor;
constexpr int min_image_tokens = 64;
constexpr int max_image_tokens = 1024;
const float min_pixels = min_image_tokens * total_factor * total_factor;
const float max_pixels = max_image_tokens * total_factor * total_factor;
auto round_by_factor = [f = total_factor](float x) { return static_cast<int>(std::nearbyintf(x / static_cast<float>(f))) * f; };
auto ceil_by_factor = [f = total_factor](float x) { return static_cast<int>(std::ceil(x / static_cast<float>(f))) * f; };
auto floor_by_factor = [f = total_factor](float x) { return static_cast<int>(std::floor(x / static_cast<float>(f))) * f; };
int h_bar = std::max(total_factor, round_by_factor(height));
int w_bar = std::max(total_factor, round_by_factor(width));
if (h_bar * w_bar > max_pixels) {
const auto beta = std::sqrt((height * width) / max_pixels);
h_bar = std::max(total_factor, floor_by_factor(height / beta));
w_bar = std::max(total_factor, floor_by_factor(width / beta));
} else if (h_bar * w_bar < min_pixels) {
const auto beta = std::sqrt(min_pixels / (height * width));
h_bar = ceil_by_factor(height * beta);
w_bar = ceil_by_factor(width * beta);
}
const std::array<uint8_t, 3> pad_color = {122, 116, 104};
clip_image_u8 resized_img;
image_manipulation::resize_and_pad_image(*img, resized_img, clip_image_size{w_bar, h_bar}, pad_color);
clip_image_f32_ptr res(clip_image_f32_init());
normalize_image_u8_to_f32(resized_img, *res, params.image_mean, params.image_std);
res_imgs->entries.push_back(std::move(res));
return true;
}
// the logic below is to pad the shorter side to the longer side with a background color: rgb(122, 116, 104)
@@ -3506,10 +3707,10 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
}
return true;
} else {
GGML_ABORT("Unknown image preprocessing type");
}
GGML_ASSERT(false && "Unknown image preprocessing type");
}
ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx) {
@@ -3573,8 +3774,9 @@ int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 *
int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * img) {
const auto & params = ctx->model.hparams;
// only for models using fixed size square images
int n_patches_sq = (params.image_size / params.patch_size) * (params.image_size / params.patch_size);
// for models with fixed size image, the input image is already pre-processed and resized to square
int patch_size = params.patch_size;
int n_patches = (img->nx / patch_size) * (img->ny / patch_size);
projector_type proj = ctx->proj_type();
@@ -3588,89 +3790,97 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
case PROJECTOR_TYPE_LDPV2:
case PROJECTOR_TYPE_GLM_EDGE:
{
n_patches_sq /= 4;
n_patches /= 4;
if (ctx->model.mm_glm_tok_boi) {
n_patches_sq += 2; // for BOI and EOI token embeddings
n_patches += 2; // for BOI and EOI token embeddings
}
} break;
case PROJECTOR_TYPE_MINICPMV:
{
if (params.minicpmv_version == 2) {
// MiniCPM-V 2.5
n_patches_sq = 96;
} else if (params.minicpmv_version == 3) {
// MiniCPM-V 2.6
n_patches_sq = 64;
} else if (params.minicpmv_version == 4) {
// MiniCPM-o 2.6
n_patches_sq = 64;
} else if (params.minicpmv_version == 5) {
// MiniCPM-V 4.0
n_patches_sq = 64;
// Use actual config value if available, otherwise fall back to hardcoded values
if (params.minicpmv_query_num > 0) {
n_patches = params.minicpmv_query_num;
} else {
GGML_ABORT("Unknown minicpmv version");
// Fallback to hardcoded values for legacy models
if (params.minicpmv_version == 2) {
n_patches = 96;
} else if (params.minicpmv_version == 3) {
n_patches = 64;
} else if (params.minicpmv_version == 4) {
n_patches = 64;
} else if (params.minicpmv_version == 5) {
// MiniCPM-V 4.0
n_patches = 64;
} else if (params.minicpmv_version == 6) {
// MiniCPM-V 4.5
n_patches = 64;
} else {
GGML_ABORT("Unknown minicpmv version");
}
}
} break;
case PROJECTOR_TYPE_QWEN2VL:
case PROJECTOR_TYPE_QWEN25VL:
{
// dynamic size
// dynamic size (2 conv, so double patch size)
int patch_size = params.patch_size * 2;
int x_patch = img->nx / patch_size + (int)(img->nx % patch_size > 0);
int y_patch = img->ny / patch_size + (int)(img->ny % patch_size > 0);
n_patches_sq = x_patch * y_patch;
n_patches = x_patch * y_patch;
} break;
case PROJECTOR_TYPE_GEMMA3:
{
int n_per_side = params.image_size / params.patch_size;
int n_per_side_2d_pool = n_per_side / params.proj_scale_factor;
n_patches_sq = n_per_side_2d_pool * n_per_side_2d_pool;
} break;
case PROJECTOR_TYPE_IDEFICS3:
case PROJECTOR_TYPE_INTERNVL:
case PROJECTOR_TYPE_LLAMA4:
{
// both W and H are divided by proj_scale_factor
n_patches_sq /= (params.proj_scale_factor * params.proj_scale_factor);
// both X and Y are downscaled by the scale factor
int scale_factor = ctx->model.hparams.proj_scale_factor;
n_patches /= (scale_factor * scale_factor);
} break;
case PROJECTOR_TYPE_LFM2:
case PROJECTOR_TYPE_KIMIVL:
{
// dynamic size
int scale_factor = ctx->model.hparams.proj_scale_factor;
int out_patch_size = params.patch_size * scale_factor;
int x_patch = CLIP_ALIGN(img->nx, out_patch_size) / out_patch_size;
int y_patch = CLIP_ALIGN(img->ny, out_patch_size) / out_patch_size;
n_patches = x_patch * y_patch;
} break;
case PROJECTOR_TYPE_PIXTRAL:
{
// dynamic size
int n_merge = params.spatial_merge_size;
int n_patches_x = img->nx / params.patch_size / (n_merge > 0 ? n_merge : 1);
int n_patches_y = img->ny / params.patch_size / (n_merge > 0 ? n_merge : 1);
n_patches_sq = n_patches_y * n_patches_x + n_patches_y - 1; // + one [IMG_BREAK] per row, except the last row
} break;
case PROJECTOR_TYPE_LLAMA4:
{
int scale_factor = ctx->model.hparams.proj_scale_factor;
n_patches_sq /= (scale_factor * scale_factor);
int n_patches_x = img->nx / patch_size / (n_merge > 0 ? n_merge : 1);
int n_patches_y = img->ny / patch_size / (n_merge > 0 ? n_merge : 1);
n_patches = n_patches_y * n_patches_x + n_patches_y - 1; // + one [IMG_BREAK] per row, except the last row
} break;
case PROJECTOR_TYPE_VOXTRAL:
case PROJECTOR_TYPE_ULTRAVOX:
case PROJECTOR_TYPE_QWEN2A:
{
n_patches_sq = img->nx;
n_patches = img->nx;
const int proj_stack_factor = ctx->model.hparams.proj_stack_factor;
if (ctx->model.audio_has_stack_frames()) {
GGML_ASSERT(proj_stack_factor > 0);
const int n_len = CLIP_ALIGN(n_patches_sq, proj_stack_factor);
n_patches_sq = n_len / proj_stack_factor;
const int n_len = CLIP_ALIGN(n_patches, proj_stack_factor);
n_patches = n_len / proj_stack_factor;
}
// whisper downscales input token by half after conv1d
n_patches_sq /= 2;
n_patches /= 2;
if (ctx->model.audio_has_avgpool()) {
// divide by 2 because of nn.AvgPool1d(2, stride=2)
n_patches_sq /= 2;
n_patches /= 2;
}
} break;
default:
GGML_ABORT("unsupported projector type");
}
return n_patches_sq;
return n_patches;
}
static std::vector<std::vector<std::vector<float>>> get_1d_sincos_pos_embed_from_grid_new(int embed_dim, const std::vector<std::vector<float>> & pos) {
@@ -4019,6 +4229,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
set_input_i32("positions", positions);
} break;
case PROJECTOR_TYPE_PIXTRAL:
case PROJECTOR_TYPE_KIMIVL:
{
// set the 2D positions
int n_patches_per_col = image_size_width / patch_size;
@@ -4070,6 +4281,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
case PROJECTOR_TYPE_INTERNVL:
case PROJECTOR_TYPE_QWEN2A:
case PROJECTOR_TYPE_ULTRAVOX:
case PROJECTOR_TYPE_LFM2:
case PROJECTOR_TYPE_VOXTRAL:
{
// do nothing
@@ -4141,7 +4353,6 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
}
int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
const auto & hparams = ctx->model.hparams;
switch (ctx->model.proj_type) {
case PROJECTOR_TYPE_LDP:
return ctx->model.mm_model_block_1_block_2_1_b->ne[0];
@@ -4153,20 +4364,7 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
case PROJECTOR_TYPE_MLP_NORM:
return ctx->model.mm_3_b->ne[0];
case PROJECTOR_TYPE_MINICPMV:
if (hparams.minicpmv_version == 2) {
// MiniCPM-V 2.5
return 4096;
} else if (hparams.minicpmv_version == 3) {
// MiniCPM-V 2.6
return 3584;
} else if (hparams.minicpmv_version == 4) {
// MiniCPM-o 2.6
return 3584;
} else if (hparams.minicpmv_version == 5) {
// MiniCPM-V 4.0
return 2560;
}
GGML_ABORT("Unknown minicpmv version");
return ctx->model.mm_model_proj->ne[0];
case PROJECTOR_TYPE_GLM_EDGE:
return ctx->model.mm_model_mlp_3_w->ne[1];
case PROJECTOR_TYPE_QWEN2VL:
@@ -4185,6 +4383,9 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
return ctx->model.mm_model_proj->ne[1];
case PROJECTOR_TYPE_QWEN2A:
return ctx->model.mm_fc_w->ne[1];
case PROJECTOR_TYPE_LFM2:
case PROJECTOR_TYPE_KIMIVL:
return ctx->model.mm_2_w->ne[1];
default:
GGML_ABORT("Unknown projector type");
}