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

[update] image min/max pixels

This commit is contained in:
megemini 2026-01-15 13:33:30 +08:00
parent 9d5a701ebe
commit 2f8c194135
4 changed files with 10 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
gguf-py/gguf/constants.py
View File

@ -284,8 +284,8 @@ class Keys:
class ClipVision:
PROJECTOR_TYPE = "clip.vision.projector_type" # for mixed modality models
IMAGE_SIZE = "clip.vision.image_size"
MAX_PIXELS = "clip.vision.max_pixels"
MIN_PIXELS = "clip.vision.min_pixels"
IMAGE_MIN_PIXELS = "clip.vision.image_min_pixels"
IMAGE_MAX_PIXELS = "clip.vision.image_max_pixels"
PREPROC_IMAGE_SIZE = "clip.vision.preproc_image_size"
PATCH_SIZE = "clip.vision.patch_size"
EMBEDDING_LENGTH = "clip.vision.embedding_length"

6
gguf-py/gguf/gguf_writer.py
View File

@ -1099,11 +1099,11 @@ class GGUFWriter:
self.add_uint32(Keys.ClipVision.EMBEDDING_LENGTH, value)
def add_vision_max_pixels(self, value: int) -> None:
self.add_uint32(Keys.ClipVision.MAX_PIXELS, value)
self.add_uint32(Keys.ClipVision.IMAGE_MAX_PIXELS, value)
def add_vision_min_pixels(self, value: int) -> None:
self.add_uint32(Keys.ClipVision.MIN_PIXELS, value)
self.add_uint32(Keys.ClipVision.IMAGE_MIN_PIXELS, value)
def add_vision_feed_forward_length(self, value: int) -> None:
self.add_uint32(Keys.ClipVision.FEED_FORWARD_LENGTH, value)

2
tools/mtmd/clip-impl.h
View File

@ -36,6 +36,8 @@
// vision-specific
#define KEY_VISION_PROJ_TYPE "clip.vision.projector_type" // for models with mixed modalities
#define KEY_IMAGE_SIZE "clip.vision.image_size"
#define KEY_IMAGE_MIN_PIXELS "clip.vision.image_min_pixels"
#define KEY_IMAGE_MAX_PIXELS "clip.vision.image_max_pixels"
#define KEY_PREPROC_IMAGE_SIZE "clip.vision.preproc_image_size"
#define KEY_PATCH_SIZE "clip.vision.patch_size"
#define KEY_IMAGE_MEAN "clip.vision.image_mean"

3
tools/mtmd/clip.cpp
View File

@ -1230,6 +1230,9 @@ struct clip_model_loader {
case PROJECTOR_TYPE_PADDLEOCR:
{
hparams.n_merge = 2;
get_u32(KEY_IMAGE_MIN_PIXELS, hparams.image_min_pixels);
get_u32(KEY_IMAGE_MAX_PIXELS, hparams.image_max_pixels);
// TODO(megemini): paddleocr vl not specified?
hparams.set_limit_image_tokens(8, 4096);
hparams.set_warmup_n_tokens(28*28); // avoid OOM on warmup