sleepy/llama.cpp
1
0
Fork 0
Code Issues 9 Pull Requests Actions 103 Packages Projects Releases Wiki Activity

mtmd: support dots.ocr (#17575)

Browse Source 
* convert gguf

* clip impl

* fix conversion

* wip

* corrections

* update docs

* add gguf to test script
This commit is contained in:
Xuan-Son Nguyen
2026-04-09 12:16:38 +02:00
committed by GitHub
parent 0ec191e1d7
commit 501aeed18f
11 changed files with 165 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
convert_hf_to_gguf.py
+41 -2
View File
@@ -3777,7 +3777,14 @@ class QwenModel(TextModel):
self._set_vocab_qwen()
@ModelBase.register("Qwen2Model", "Qwen2ForCausalLM", "Qwen2AudioForConditionalGeneration", "KORMoForCausalLM", "AudioFlamingo3ForConditionalGeneration")
@ModelBase.register(
"Qwen2Model",
"Qwen2ForCausalLM",
"Qwen2AudioForConditionalGeneration",
"KORMoForCausalLM",
"AudioFlamingo3ForConditionalGeneration",
"DotsOCRForCausalLM",
)
class Qwen2Model(TextModel):
model_arch = gguf.MODEL_ARCH.QWEN2
@@ -3798,7 +3805,8 @@ class Qwen2Model(TextModel):
name = name.replace("language_model.", "") # for InternVL
if name.startswith("mlp") or name.startswith("multi_modal_projector") \
or name.startswith("vision_model") or name.startswith("audio_tower") \
or name.startswith("model.vision_tower") or name.startswith("model.multi_modal_projector"):
or name.startswith("model.vision_tower") or name.startswith("model.multi_modal_projector") \
or name.startswith("vision_tower."):
# skip vision and audio tensors
return
yield from super().modify_tensors(data_torch, name, bid)
@@ -12819,6 +12827,37 @@ class SolarOpenModel(Glm4MoeModel):
special_vocab.add_to_gguf(self.gguf_writer)
@ModelBase.register("DotsOCRForCausalLM")
class DotsOCRVisionModel(MmprojModel):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
assert self.hparams_vision is not None
self.hparams_vision["image_size"] = 0 # dynamic resolution
def set_gguf_parameters(self):
super().set_gguf_parameters()
self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.DOTSOCR)
self.gguf_writer.add_vision_min_pixels(self.preprocessor_config["min_pixels"])
self.gguf_writer.add_vision_max_pixels(self.preprocessor_config["max_pixels"])
self.gguf_writer.add_vision_attention_layernorm_eps(self.find_vparam(["rms_norm_eps"]))
self.gguf_writer.add_vision_projector_scale_factor(self.find_vparam(["spatial_merge_size"]))
self.gguf_writer.add_vision_use_silu(True)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
if name.startswith("vision_tower."):
if "vision_tower.blocks." in name and ".mlp." in name:
# note: to avoid naming conflicts in tensor_mapping.py, we need to handle FFN renaming here
# x = F.silu(self.fc1(x)) * self.fc3(x)
# x = self.fc2(x)
# fc1 -> gate, fc2 -> down, fc3 -> up
# mapping original names to Qwen2.5 naming scheme
name = name.replace("vision_tower.blocks.", "visual.blocks.")
name = name.replace(".fc1", ".gate_proj")
name = name.replace(".fc2", ".down_proj")
name = name.replace(".fc3", ".up_proj")
yield from super().modify_tensors(data_torch, name, bid)
###### CONVERSION LOGIC ######