From: jacekpoplawski <redacted>
Date: Wed, 23 Jul 2025 21:23:57 +0000 (+0200)
Subject: convert : text-only support for GLM-4.1V-9B-Thinking (#14823)
X-Git-Tag: upstream/0.0.6073~99
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=a12363bbf0ca11f787dee9756861043136edc0df;p=pkg%2Fggml%2Fsources%2Fllama.cpp

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

* use language_model part only, ignore visual layers

* fix rope_dim calculation
---

diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index c8bf3c53..e12c922b 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -6486,7 +6486,7 @@ class JaisModel(TextModel):
         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 @@ class Glm4Model(TextModel):
 
     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 @@ class Glm4Model(TextModel):
             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):