CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-31 08:51:55 +00:00
Code Issues Packages Projects Releases Wiki Activity

convert : text-only support for GLM-4.1V-9B-Thinking (#14823)

Browse Source 
* use language_model part only, ignore visual layers

* fix rope_dim calculation
This commit is contained in:
jacekpoplawski
2025-07-23 23:23:57 +02:00
committed by GitHub
parent a86f52b285
commit a12363bbf0
1 changed files with 10 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
convert_hf_to_gguf.py
View File

@@ -6486,7 +6486,7 @@ class JaisModel(TextModel):
self.gguf_writer.add_max_alibi_bias(self.max_alibi_bias) self.gguf_writer.add_max_alibi_bias(self.max_alibi_bias)
@ModelBase.register("Glm4ForCausalLM") @ModelBase.register("Glm4ForCausalLM", "Glm4vForConditionalGeneration")
class Glm4Model(TextModel): class Glm4Model(TextModel):
model_arch = gguf.MODEL_ARCH.GLM4 model_arch = gguf.MODEL_ARCH.GLM4
@@ -6508,7 +6508,8 @@ class Glm4Model(TextModel):
def set_gguf_parameters(self): def set_gguf_parameters(self):
super().set_gguf_parameters() super().set_gguf_parameters()
rope_dim = self.hparams["head_dim"] if (rope_dim := self.hparams.get("head_dim")) is None:
rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.5))) self.gguf_writer.add_rope_dimension_count(int(rope_dim * self.hparams.get("partial_rotary_factor", 0.5)))
rope_scaling = self.hparams.get("rope_scaling") or {} rope_scaling = self.hparams.get("rope_scaling") or {}
if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling: if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
@@ -6516,6 +6517,13 @@ class Glm4Model(TextModel):
self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"]) self.gguf_writer.add_rope_scaling_factor(rope_scaling["factor"])
self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"]) self.gguf_writer.add_rope_scaling_orig_ctx_len(rope_scaling["original_max_position_embeddings"])
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
if name.startswith("model.visual."): # ignore visual part of Glm4v
return []
elif name.startswith("model.language_model."):
name = name.replace("language_model.", "") # for Glm4v
return super().modify_tensors(data_torch, name, bid)
@ModelBase.register("GlmForCausalLM", "ChatGLMModel", "ChatGLMForConditionalGeneration") @ModelBase.register("GlmForCausalLM", "ChatGLMModel", "ChatGLMForConditionalGeneration")
class ChatGLMModel(TextModel): class ChatGLMModel(TextModel):