happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

feat: VAETKI dynamic image size support

This commit is contained in:
suhyun-hwang 2026-01-10 20:49:01 +09:00
parent d61a3f817c
commit 5d08f3e87b
1 changed files with 2 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
tools/mtmd/clip.cpp
View File

@ -1201,6 +1201,7 @@ struct clip_model_loader {
hparams.rope_theta = 10000.0f; hparams.rope_theta = 10000.0f;
hparams.n_merge = 2; hparams.n_merge = 2;
get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.n_merge, false); get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.n_merge, false);
hparams.set_limit_image_tokens(4, 3265);
hparams.set_warmup_n_tokens(40*40); hparams.set_warmup_n_tokens(40*40);
} break; } break;
case PROJECTOR_TYPE_LLAMA4: case PROJECTOR_TYPE_LLAMA4:
@ -2853,6 +2854,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
case PROJECTOR_TYPE_QWEN25VL: case PROJECTOR_TYPE_QWEN25VL:
case PROJECTOR_TYPE_QWEN3VL: case PROJECTOR_TYPE_QWEN3VL:
case PROJECTOR_TYPE_GLM4V: case PROJECTOR_TYPE_GLM4V:
case PROJECTOR_TYPE_VAETKI:
{ {
GGML_ASSERT(params.image_min_pixels > 0 && params.image_max_pixels > 0); GGML_ASSERT(params.image_min_pixels > 0 && params.image_max_pixels > 0);
clip_image_u8 resized; clip_image_u8 resized;
@ -2973,7 +2975,6 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, str
case PROJECTOR_TYPE_GLM_EDGE: case PROJECTOR_TYPE_GLM_EDGE:
case PROJECTOR_TYPE_GEMMA3: case PROJECTOR_TYPE_GEMMA3:
case PROJECTOR_TYPE_INTERNVL: // TODO @ngxson : support dynamic resolution case PROJECTOR_TYPE_INTERNVL: // TODO @ngxson : support dynamic resolution
case PROJECTOR_TYPE_VAETKI:
{ {
clip_image_u8 resized_image; clip_image_u8 resized_image;
int sz = params.image_size; int sz = params.image_size;