Skip to content

LLM inference in C/C++

C++ ggml

mtmd: Unlimited-OCR fix max_tiles, setting in converter (#25614)

Saba Fallahb06aa774c03d
diff --git a/conversion/deepseek.py b/conversion/deepseek.py
index 0518fcecc1..3dca369cf9 100644
--- a/conversion/deepseek.py
+++ b/conversion/deepseek.py
@@ -17,8 +17,11 @@ from .base import LazyTorchTensor, MmprojModel, ModelBase, TextModel, gguf, logg
from .qwen import QwenModel
[email protected]("DeepseekOCRForCausalLM", "UnlimitedOCRForCausalLM")
[email protected]("DeepseekOCRForCausalLM")
class DeepseekOCRVisionModel(MmprojModel):
+ # HF dynamic_preprocess() max_num, which differs per model
+ preproc_max_tiles = 9
+
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.clip_projector_type = gguf.VisionProjectorType.DEEPSEEKOCR
@@ -43,6 +46,9 @@ class DeepseekOCRVisionModel(MmprojModel):
# @bluebread: there's no window_size in config but just add it here anyway
self.gguf_writer.add_vision_window_size(self.hparams.get("window_size", 14))
+ self.gguf_writer.add_vision_preproc_min_tiles(2)
+ self.gguf_writer.add_vision_preproc_max_tiles(self.preproc_max_tiles)
+
# SAM configuration
sam_hparams = hparams['sam']
self.gguf_writer.add_vision_sam_layers_count(sam_hparams['layers'])
@@ -93,8 +99,15 @@ class DeepseekOCRVisionModel(MmprojModel):
return super().filter_tensors((name, gen))
[email protected]("UnlimitedOCRForCausalLM")
+class UnlimitedOCRVisionModel(DeepseekOCRVisionModel):
+ preproc_max_tiles = 32
+
+
@ModelBase.register("DeepseekOCR2ForCausalLM")
class DeepseekOCR2VisionModel(DeepseekOCRVisionModel):
+ preproc_max_tiles = 6
+
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.clip_projector_type = gguf.VisionProjectorType.DEEPSEEKOCR2
diff --git a/tools/mtmd/clip.cpp b/tools/mtmd/clip.cpp
index d6670030ff..c77e7cc9d3 100644
--- a/tools/mtmd/clip.cpp
+++ b/tools/mtmd/clip.cpp
@@ -1761,6 +1761,10 @@ struct clip_model_loader {
// qwen2 encoder is GQA, requires KEY_N_HEAD_KV
get_u32(string_format(KEY_N_HEAD_KV, "vision"), hparams.n_head_kv);
}
+ // unlimited-ocr shares the v1 projector but tiles up to 32
+ get_u32(KEY_PREPROC_MIN_TILES, hparams.preproc_min_tiles, false);
+ get_u32(KEY_PREPROC_MAX_TILES, hparams.preproc_max_tiles, false);
+ GGML_ASSERT(hparams.preproc_min_tiles <= hparams.preproc_max_tiles);
} break;
case PROJECTOR_TYPE_HUNYUANVL:
{
@@ -1909,6 +1913,9 @@ struct clip_model_loader {
if (hparams.image_max_pixels > 0) {
LOG_INF("%s: image_max_pixels: %d%s\n", __func__, hparams.image_max_pixels, hparams.custom_image_max_tokens > 0 ? " (custom value)" : "");
}
+ if (hparams.preproc_max_tiles > 0) {
+ LOG_INF("%s: preproc_tiles: %d - %d\n", __func__, hparams.preproc_min_tiles, hparams.preproc_max_tiles);
+ }
} else if (is_audio) {
LOG_INF("\n--- audio hparams ---\n");
LOG_INF("%s: n_mel_bins: %d\n", __func__, hparams.n_mel_bins);