nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 20 Packages Projects Releases Wiki Activity

mtmd, model : merge HunyuanOCR into HunyuanVL and fix OCR vision precision (#23329)

Browse Source 
- HunyuanOCR shares the same HF arch and vision layout as HunyuanVL butwas split into a separate path that skipped the +0.1 bilinear sampler used by the HF reference.
- Collapse OCR into the HUNYUANVL projector + HUNYUAN_VL text arch
This commit is contained in:
wendadawen
2026-05-21 00:35:37 +02:00
committed by GitHub
parent 3a479c9132
commit 6a257d4463
15 changed files with 58 additions and 134 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/mtmd/CMakeLists.txt
+1 -1
View File
@@ -22,7 +22,7 @@ add_library(mtmd
models/gemma4v.cpp
models/glm4v.cpp
models/granite-speech.cpp
models/hunyuanocr.cpp
models/hunyuanvl.cpp
models/internvl.cpp
models/kimivl.cpp
models/kimik25.cpp
tools/mtmd/clip-impl.h
+1 -3
View File
@@ -170,7 +170,7 @@
#define TN_TOK_BOI "v.boi"
#define TN_TOK_EOI "v.eoi"
// hunyuanocr / hunyuanvl (shared GGUF tensor names)
// hunyuanvl (shared GGUF tensor names)
#define TN_MM_PRE_NORM "mm.pre_norm.%s"
#define TN_TOK_IMG_BEGIN "mm.image_begin"
#define TN_TOK_IMG_END "mm.image_end"
@@ -343,7 +343,6 @@ enum projector_type {
PROJECTOR_TYPE_YASA2,
PROJECTOR_TYPE_KIMIK25,
PROJECTOR_TYPE_NEMOTRON_V2_VL,
PROJECTOR_TYPE_HUNYUANOCR,
PROJECTOR_TYPE_HUNYUANVL,
PROJECTOR_TYPE_MINICPMV4_6,
PROJECTOR_TYPE_GRANITE_SPEECH,
@@ -393,7 +392,6 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
{ PROJECTOR_TYPE_YASA2, "yasa2"},
{ PROJECTOR_TYPE_KIMIK25, "kimik25"},
{ PROJECTOR_TYPE_NEMOTRON_V2_VL, "nemotron_v2_vl"},
{ PROJECTOR_TYPE_HUNYUANOCR, "hunyuanocr"},
{ PROJECTOR_TYPE_HUNYUANVL, "hunyuanvl"},
{ PROJECTOR_TYPE_MINICPMV4_6, "minicpmv4_6"},
{ PROJECTOR_TYPE_GRANITE_SPEECH, "granite_speech"},
tools/mtmd/clip-model.h
+1 -1
View File
@@ -520,7 +520,7 @@ struct clip_model {
ggml_tensor * mm_boi = nullptr;
ggml_tensor * mm_eoi = nullptr;
// hunyuanocr perceiver
// hunyuanvl perceiver
ggml_tensor * mm_pre_norm_w = nullptr;
ggml_tensor * mm_img_begin = nullptr;
ggml_tensor * mm_img_end = nullptr;
tools/mtmd/clip.cpp
+5 -17
View File
@@ -936,10 +936,9 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
{
builder = std::make_unique<clip_graph_cogvlm>(ctx, img);
} break;
case PROJECTOR_TYPE_HUNYUANOCR:
case PROJECTOR_TYPE_HUNYUANVL:
{
builder = std::make_unique<clip_graph_hunyuanocr>(ctx, img);
builder = std::make_unique<clip_graph_hunyuanvl>(ctx, img);
} break;
case PROJECTOR_TYPE_MLP:
case PROJECTOR_TYPE_MLP_NORM:
@@ -1523,22 +1522,16 @@ struct clip_model_loader {
get_u32(KEY_SAM_N_EMBD, hparams.sam_n_embd, true);
get_u32(KEY_ATTN_WINDOW_SIZE, hparams.attn_window_size, true);
} break;
case PROJECTOR_TYPE_HUNYUANOCR:
{
hparams.n_merge = 2;
get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.n_merge, false);
get_u32(KEY_IMAGE_MIN_PIXELS, hparams.image_min_pixels);
get_u32(KEY_IMAGE_MAX_PIXELS, hparams.image_max_pixels);
hparams.set_warmup_n_tokens(28*28);
} break;
case PROJECTOR_TYPE_HUNYUANVL:
{
hparams.n_merge = 2;
hparams.image_resize_algo = RESIZE_ALGO_BICUBIC_PILLOW;
hparams.image_resize_pad = PAD_NONE;
hparams.ffn_op = FFN_GELU;
get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.n_merge, false);
hparams.set_limit_image_tokens(256, 16384);
get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.n_merge, false);
get_u32(KEY_IMAGE_MIN_PIXELS, hparams.image_min_pixels, false);
get_u32(KEY_IMAGE_MAX_PIXELS, hparams.image_max_pixels, false);
hparams.set_warmup_n_tokens(32*32);
} break;
case PROJECTOR_TYPE_LFM2A:
@@ -2343,7 +2336,6 @@ struct clip_model_loader {
model.mm_boi = get_tensor(TN_TOK_BOI);
model.mm_eoi = get_tensor(TN_TOK_EOI);
} break;
case PROJECTOR_TYPE_HUNYUANOCR:
case PROJECTOR_TYPE_HUNYUANVL:
{
// proj.0 -> mm.0 (conv1), proj.2 -> mm.2 (conv2), mlp -> mm.model.fc (linear)
@@ -3071,7 +3063,6 @@ int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 *
case PROJECTOR_TYPE_MIMOVL:
case PROJECTOR_TYPE_GLM4V:
case PROJECTOR_TYPE_PADDLEOCR:
case PROJECTOR_TYPE_HUNYUANOCR:
case PROJECTOR_TYPE_HUNYUANVL:
case PROJECTOR_TYPE_YOUTUVL:
return (img->nx / params.patch_size) / 2;
@@ -3288,7 +3279,6 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
int h = static_cast<int>(std::sqrt(static_cast<float>(n_patches)));
n_patches = h * (h + 1) + 1;
} break;
case PROJECTOR_TYPE_HUNYUANOCR:
case PROJECTOR_TYPE_HUNYUANVL:
{
int merge = ctx->model.hparams.n_merge;
@@ -3924,7 +3914,6 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
case PROJECTOR_TYPE_JANUS_PRO:
case PROJECTOR_TYPE_PHI4:
case PROJECTOR_TYPE_COGVLM:
case PROJECTOR_TYPE_HUNYUANOCR:
case PROJECTOR_TYPE_YASA2:
{
// do nothing
@@ -3934,7 +3923,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
// Compute the HunyuanVL 2D position embedding on CPU (with the
// custom sf=(target+0.1)/n_grid bilinear sampling that the
// reference implementation uses) and upload it to the graph
// input declared in clip_graph_hunyuanocr::build().
// input declared in clip_graph_hunyuanvl::build().
GGML_ASSERT(model.position_embeddings != nullptr);
ggml_tensor * src_t = model.position_embeddings;
const int64_t n_embd = src_t->ne[0];
@@ -4255,7 +4244,6 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
case PROJECTOR_TYPE_KIMIK25:
case PROJECTOR_TYPE_YASA2:
return ctx->model.mm_2_w->ne[1];
case PROJECTOR_TYPE_HUNYUANOCR:
case PROJECTOR_TYPE_HUNYUANVL:
return ctx->model.mm_model_proj->ne[1];
case PROJECTOR_TYPE_COGVLM:
tools/mtmd/models/hunyuanocr.cpp → tools/mtmd/models/hunyuanvl.cpp
+6 -16
View File
@@ -1,25 +1,15 @@
#include "models.h"
ggml_cgraph * clip_graph_hunyuanocr::build() {
ggml_cgraph * clip_graph_hunyuanvl::build() {
const int merge = hparams.n_merge;
const int pw = n_patches_x;
const int ph = n_patches_y;
// Position embedding interpolation.
// HunyuanVL needs scale factors sf=(target+0.1)/n_grid, which the standard
// ggml_interpolate cannot express. To avoid adding a new ggml op, the
// resize is computed on CPU in clip_image_batch_encode and uploaded here
// as a graph input (named "hunyuanvl_pos_embd").
// HunyuanOCR uses the same square layout and the standard ratio-based
// interpolation provided by resize_position_embeddings().
ggml_tensor * pos_embd = nullptr;
if (proj_type == PROJECTOR_TYPE_HUNYUANVL && model.position_embeddings) {
pos_embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, ph * pw);
ggml_set_name(pos_embd, "hunyuanvl_pos_embd");
ggml_set_input(pos_embd);
} else {
pos_embd = resize_position_embeddings(GGML_SCALE_MODE_BILINEAR);
}
// position embedding: declared as a graph input, filled on CPU
// by clip_image_batch_encode (see PROJECTOR_TYPE_HUNYUANVL branch there).
ggml_tensor * pos_embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, ph * pw);
ggml_set_name(pos_embd, "hunyuanvl_pos_embd");
ggml_set_input(pos_embd);
ggml_tensor * inp = build_inp();
ggml_tensor * cur = build_vit(inp, n_patches, NORM_TYPE_NORMAL, hparams.ffn_op, pos_embd, nullptr);
tools/mtmd/models/models.h
+2 -2
View File
@@ -142,8 +142,8 @@ struct clip_graph_glm4v : clip_graph {
ggml_cgraph * build() override;
};
struct clip_graph_hunyuanocr : clip_graph {
clip_graph_hunyuanocr(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
struct clip_graph_hunyuanvl : clip_graph {
clip_graph_hunyuanvl(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
ggml_cgraph * build() override;
};
tools/mtmd/mtmd.cpp
-1
View File
@@ -493,7 +493,6 @@ struct mtmd_context {
img_end = "\n"; // prevent empty batch on llama-server
image_preproc = std::make_unique<mtmd_image_preprocessor_deepseekocr>(ctx_v);
} break;
case PROJECTOR_TYPE_HUNYUANOCR:
case PROJECTOR_TYPE_HUNYUANVL:
{
// note: these use fullwidth ｜ (U+FF5C) and ▁ (U+2581) to match the tokenizer vocabulary