thek0tyara/llama-cpp-turboquant
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 4

model: support GLM-OCR (#19677)

Browse source 
* model: support GLM-OCR

* Update convert_hf_to_gguf.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
This commit is contained in:
Xuan-Son Nguyen 2026-02-18 17:51:40 +01:00 • committed by GitHub
parent e99f1083a0
commit eeef3cfced
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
8 changed files with 122 additions and 43 deletions
Download patch file Download diff file Expand all files Collapse all files

25
convert_hf_to_gguf.py
View file

@ -4584,7 +4584,7 @@ class Qwen3VLVisionModel(MmprojModel):
yield from super().modify_tensors(data_torch, name, bid)
@ModelBase.register("Glm4vForConditionalGeneration", "Glm4vMoeForConditionalGeneration")
@ModelBase.register("Glm4vForConditionalGeneration", "Glm4vMoeForConditionalGeneration", "GlmOcrForConditionalGeneration")
class Glm4VVisionModel(Qwen3VLVisionModel):
def set_gguf_parameters(self):
MmprojModel.set_gguf_parameters(self) # skip Qwen3VLVisionModel parameters
@ -8776,7 +8776,7 @@ class Glm4Model(TextModel):
n_head = self.hparams["num_attention_heads"]
n_kv_head = self.hparams["num_key_value_heads"]
n_embd = self.hparams["hidden_size"]
head_dim = n_embd // n_head
head_dim = self.hparams.get("head_dim", n_embd // n_head)
# because llama.cpp M-RoPE kernel only supports Neox ordering, we have to permute the weights here
if name.endswith(("q_proj.weight", "q_proj.bias")):
data_torch = Glm4Model.normal_to_neox(data_torch, n_head, n_head, head_dim, self.partial_rotary_factor)
@ -8785,6 +8785,27 @@ class Glm4Model(TextModel):
yield from super().modify_tensors(data_torch, name, bid)
@ModelBase.register("GlmOcrForConditionalGeneration")
class GlmOCRModel(Glm4Model):
model_arch = gguf.MODEL_ARCH.GLM4
use_mrope = False
partial_rotary_factor = 0.5
# Note: GLM-OCR is the same as GLM4, but with an extra NextN/MTP prediction layer
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
# GLM-OCR has num_hidden_layers + 1 actual layers (including NextN layer)
self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0)
self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
def set_gguf_parameters(self):
super().set_gguf_parameters()
# NextN/MTP prediction layers
if (num_nextn_predict_layers := self.hparams.get("num_nextn_predict_layers")) is not None:
self.gguf_writer.add_nextn_predict_layers(num_nextn_predict_layers)
@ModelBase.register("Glm4MoeForCausalLM", "Glm4vMoeForConditionalGeneration")
class Glm4MoeModel(TextModel):
model_arch = gguf.MODEL_ARCH.GLM4_MOE