From d959644abec134df3cd505070bae6407973c8491 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Jacek=20Pop=C5=82awski?= Date: Tue, 22 Jul 2025 23:48:57 +0200 Subject: [PATCH] convert: text-only support for GLM-4.1V-9B-Thinking (#14495) * use language_model part only, ignore visual layers * fix rope_dim calculation --- convert_hf_to_gguf.py | 12 ++++++++++-- 1 file changed, 10 insertions(+), 2 deletions(-) diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index c8bf3c5383089..e12c922bd9ab4 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -6486,7 +6486,7 @@ def prepare_tensors(self): self.gguf_writer.add_max_alibi_bias(self.max_alibi_bias) -@ModelBase.register("Glm4ForCausalLM") +@ModelBase.register("Glm4ForCausalLM", "Glm4vForConditionalGeneration") class Glm4Model(TextModel): model_arch = gguf.MODEL_ARCH.GLM4 @@ -6508,7 +6508,8 @@ def set_vocab(self): def set_gguf_parameters(self): 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))) rope_scaling = self.hparams.get("rope_scaling") or {} if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling: @@ -6516,6 +6517,13 @@ def set_gguf_parameters(self): 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"]) + 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") class ChatGLMModel(TextModel):