model: support GLM4V vision encoder

convert_hf_to_gguf.py

@@ -4407,6 +4407,30 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
         return super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register("Glm4vForConditionalGeneration", "Glm4vMoeForConditionalGeneration")
+class Glm4VVisionModel(Qwen3VLVisionModel):
+    def set_gguf_parameters(self):
+        MmprojModel.set_gguf_parameters(self) # skip Qwen3VLVisionModel parameters
+        assert self.hparams_vision is not None
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.GLM4V)
+
+        hidden_act = str(self.hparams_vision.get("hidden_act", "")).lower()
+        if hidden_act == "gelu":
+            self.gguf_writer.add_vision_use_gelu(True)
+        elif hidden_act == "silu":
+            self.gguf_writer.add_vision_use_silu(True)
+
+        rms_norm_eps = self.hparams_vision.get("rms_norm_eps", 1e-5)
+        self.gguf_writer.add_vision_attention_layernorm_eps(rms_norm_eps)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.startswith("model.visual."):
+            name = name.replace("model.visual.", "visual.")
+        if name.startswith("visual.merger."):
+            return [(self.map_tensor_name(name), data_torch)]
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("Qwen3VLForConditionalGeneration")
 class Qwen3VLTextModel(Qwen3Model):
     model_arch = gguf.MODEL_ARCH.QWEN3VL
@@ -7876,6 +7900,16 @@ def prepare_tensors(self):
 @ModelBase.register("Glm4ForCausalLM", "Glm4vForConditionalGeneration")
 class Glm4Model(TextModel):
     model_arch = gguf.MODEL_ARCH.GLM4
+    use_mrope = False
+    partial_rotary_factor = 0.5
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        rope_scaling = self.hparams.get("rope_scaling") or self.hparams.get("rope_parameters") or {}
+        if "mrope_section" in rope_scaling:
+            self.use_mrope = True
+            self.partial_rotary_factor = rope_scaling.get("partial_rotary_factor", 0.5)
+            logger.info("Using M-RoPE")
 
     def set_vocab(self):
         from transformers import AutoTokenizer
@@ -7898,21 +7932,61 @@ def set_gguf_parameters(self):
         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 {}
+        rope_scaling = self.hparams.get("rope_scaling") or self.hparams.get("rope_parameters") or {}
         if rope_scaling.get("rope_type", rope_scaling.get("type")) == "yarn" and "factor" in rope_scaling:
             self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)
             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"])
+        # handle M-RoPE, the same as Qwen-VL
+        if self.use_mrope:
+            mrope_section = rope_scaling["mrope_section"]
+            # Pad to 4 dimensions [time, height, width, extra]
+            while len(mrope_section) < 4:
+                mrope_section.append(0)
+            self.gguf_writer.add_rope_dimension_sections(mrope_section[:4])
+            logger.info(f"MRoPE sections: {mrope_section[:4]}")
+
+    @staticmethod
+    def normal_to_neox(weights: Tensor, n_head: int, n_head_kv: int, head_dim: int, partial_rotary_factor: float) -> Tensor:
+        orig_shape = weights.shape
+        if len(orig_shape) == 1:
+            weights = weights.unsqueeze(1)  # [out_dim, 1]
+        if len(weights.shape) != 2:
+            raise ValueError("Only 1D and 2D tensors are supported.")
+        n_effective_heads = weights.shape[0] // head_dim
+        if n_head_kv is not None and n_effective_heads != n_head:
+            if n_effective_heads != n_head_kv:
+                raise AssertionError(f"Mismatch in effective heads: computed {n_effective_heads}, expected {n_head} or {n_head_kv}")
+        rotary_dim = int(head_dim * partial_rotary_factor)
+        if rotary_dim % 2 != 0:
+            raise ValueError("rotary_dim must be even.")
+        reshaped = weights.reshape(n_effective_heads, head_dim, -1)
+        rot_part = reshaped[:, :rotary_dim, :]
+        non_rot_part = reshaped[:, rotary_dim:, :]
+        permuted_rot = torch.cat((rot_part[:, ::2, :], rot_part[:, 1::2, :]), dim=1)
+        combined = torch.cat((permuted_rot, non_rot_part), dim=1)
+        result = combined.reshape(weights.shape)
+        return result if len(orig_shape) != 1 else result.squeeze(1)
 
     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
+        if self.use_mrope:
+            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
+            # 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)
+            if name.endswith(("k_proj.weight", "k_proj.bias")):
+                data_torch = Glm4Model.normal_to_neox(data_torch, n_head, n_kv_head, head_dim, self.partial_rotary_factor)
         return super().modify_tensors(data_torch, name, bid)
 
 
-@ModelBase.register("Glm4MoeForCausalLM")
+@ModelBase.register("Glm4MoeForCausalLM", "Glm4vMoeForConditionalGeneration")
 class Glm4MoeModel(TextModel):
     model_arch = gguf.MODEL_ARCH.GLM4_MOE
 
@@ -7977,6 +8051,17 @@ def set_gguf_parameters(self):
         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)
 
+        # handle M-RoPE, the same as Qwen-VL
+        # note: unlike GLM4 non-MoE, we don't need to permute the weights here since GLM4_MOE uses Neox ordering already
+        rope_scaling = self.hparams.get("rope_scaling") or self.hparams.get("rope_parameters") or {}
+        if "mrope_section" in rope_scaling:
+            mrope_section = rope_scaling["mrope_section"]
+            # Pad to 4 dimensions [time, height, width, extra]
+            while len(mrope_section) < 4:
+                mrope_section.append(0)
+            self.gguf_writer.add_rope_dimension_sections(mrope_section[:4])
+            logger.info(f"MRoPE sections: {mrope_section[:4]}")
+
     _experts: list[dict[str, Tensor]] | None = None
 
     def modify_tensors(

gguf-py/gguf/constants.py

@@ -642,6 +642,7 @@ class MODEL_TENSOR(IntEnum):
     V_MMPROJ_PEG         = auto()
     V_ENC_EMBD_CLS       = auto()
     V_ENC_EMBD_PATCH     = auto()
+    V_ENC_EMBD_NORM      = auto()
     V_ENC_EMBD_POS       = auto()
     V_ENC_INPUT_NORM     = auto()
     V_ENC_ATTN_QKV       = auto()
@@ -660,6 +661,7 @@ class MODEL_TENSOR(IntEnum):
     V_LAYER_SCALE_2      = auto()
     V_PRE_NORM           = auto()
     V_POST_NORM          = auto()
+    V_MM_POST_NORM       = auto()
     V_MM_INP_NORM        = auto()
     V_MM_INP_PROJ        = auto() # gemma3
     V_MM_SOFT_EMB_NORM   = auto() # gemma3
@@ -1014,6 +1016,7 @@ class MODEL_TENSOR(IntEnum):
     MODEL_TENSOR.V_MMPROJ_PEG:              "mm.model.peg.{bid}",
     MODEL_TENSOR.V_ENC_EMBD_CLS:            "v.class_embd",
     MODEL_TENSOR.V_ENC_EMBD_PATCH:          "v.patch_embd",
+    MODEL_TENSOR.V_ENC_EMBD_NORM:           "v.norm_embd",
     MODEL_TENSOR.V_ENC_EMBD_POS:            "v.position_embd",
     MODEL_TENSOR.V_ENC_ATTN_QKV:            "v.blk.{bid}.attn_qkv",
     MODEL_TENSOR.V_ENC_ATTN_Q:              "v.blk.{bid}.attn_q",
@@ -1032,6 +1035,7 @@ class MODEL_TENSOR(IntEnum):
     MODEL_TENSOR.V_LAYER_SCALE_2:           "v.blk.{bid}.ls2",
     MODEL_TENSOR.V_PRE_NORM:                "v.pre_ln",
     MODEL_TENSOR.V_POST_NORM:               "v.post_ln",
+    MODEL_TENSOR.V_MM_POST_NORM:            "mm.post_norm",
     MODEL_TENSOR.V_MM_INP_PROJ:             "mm.input_projection",
     MODEL_TENSOR.V_MM_INP_NORM:             "mm.input_norm",
     MODEL_TENSOR.V_MM_SOFT_EMB_NORM:        "mm.soft_emb_norm",
@@ -1092,6 +1096,7 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.V_MMPROJ_PEG,
         MODEL_TENSOR.V_ENC_EMBD_CLS,
         MODEL_TENSOR.V_ENC_EMBD_PATCH,
+        MODEL_TENSOR.V_ENC_EMBD_NORM,
         MODEL_TENSOR.V_ENC_EMBD_POS,
         MODEL_TENSOR.V_ENC_INPUT_NORM,
         MODEL_TENSOR.V_ENC_ATTN_QKV,
@@ -1110,6 +1115,7 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.V_LAYER_SCALE_2,
         MODEL_TENSOR.V_PRE_NORM,
         MODEL_TENSOR.V_POST_NORM,
+        MODEL_TENSOR.V_MM_POST_NORM,
         MODEL_TENSOR.V_MM_INP_PROJ,
         MODEL_TENSOR.V_MM_INP_NORM,
         MODEL_TENSOR.V_MM_SOFT_EMB_NORM,
@@ -3327,6 +3333,7 @@ class VisionProjectorType:
     LIGHTONOCR = "lightonocr"
     COGVLM = "cogvlm"
     JANUS_PRO = "janus_pro"
+    GLM4V = "glm4v"
 
 
 # Items here are (block size, type size)

gguf-py/gguf/tensor_mapping.py

@@ -1205,6 +1205,7 @@ class TensorNameMap:
         MODEL_TENSOR.V_MMPROJ_FC: (
             "model.connector.modality_projection.proj", # SmolVLM
             "model.vision.linear_proj.linear_proj", # cogvlm
+            "visual.merger.proj", # glm4v
         ),
 
         MODEL_TENSOR.V_MMPROJ_MLP: (
@@ -1238,6 +1239,10 @@ class TensorNameMap:
             "model.vision.patch_embedding.proj", # cogvlm
         ),
 
+        MODEL_TENSOR.V_ENC_EMBD_NORM: (
+            "visual.post_conv_layernorm", # glm4v
+        ),
+
         MODEL_TENSOR.V_ENC_EMBD_POS: (
             "vision_tower.vision_model.embeddings.position_embedding",
             "model.vision_tower.embeddings.position_embeddings", # Intern-S1
@@ -1247,6 +1252,7 @@ class TensorNameMap:
             "vision_tower.patch_embed.pos_emb", # kimi-vl
             "visual.pos_embed", # qwen3vl
             "model.vision.patch_embedding.position_embedding", # cogvlm
+            "visual.embeddings.position_embedding", # glm4v
         ),
 
         MODEL_TENSOR.V_ENC_ATTN_QKV: (
@@ -1402,6 +1408,11 @@ class TensorNameMap:
             "vision_model.layernorm_post", # llama4
             "visual.merger.ln_q", # qwen2vl
             "vision_tower.encoder.final_layernorm", # kimi-vl
+            "visual.post_layernorm", # glm4v
+        ),
+
+        MODEL_TENSOR.V_MM_POST_NORM: (
+            "visual.merger.post_projection_norm", # glm4v
         ),
 
         MODEL_TENSOR.V_MM_INP_PROJ: (
@@ -1471,6 +1482,7 @@ class TensorNameMap:
         MODEL_TENSOR.V_MM_PATCH_MERGER: (
             "multi_modal_projector.patch_merger.merging_layer", # mistral small 3.1 - hf
             "patch_merger.merging_layer", # mistral
+            "visual.downsample", # glm4v
         ),
 
         MODEL_TENSOR.V_DS_NORM: (
@@ -1491,14 +1503,17 @@ class TensorNameMap:
 
         MODEL_TENSOR.V_MM_UP: (
             "model.vision.linear_proj.dense_h_to_4h", # cogvlm
+            "visual.merger.up_proj", # glm4v
         ),
 
         MODEL_TENSOR.V_MM_DOWN: (
             "model.vision.linear_proj.dense_4h_to_h", # cogvlm
+            "visual.merger.down_proj", # glm4v
         ),
 
         MODEL_TENSOR.V_MM_GATE: (
             "model.vision.linear_proj.gate_proj", # cogvlm
+            "visual.merger.gate_proj", # glm4v
         ),
 
         MODEL_TENSOR.V_TOK_BOI: (

src/llama-hparams.cpp

@@ -241,3 +241,7 @@ float llama_hparams::yarn_attn_factor_adjust(float attn_factor, float freq_scale
 
     return attn_factor;
 }
+
+bool llama_hparams::use_mrope() const {
+    return rope_sections[0] > 0 && rope_sections[1] > 0;
+}

src/llama-hparams.h

@@ -275,6 +275,8 @@ struct llama_hparams {
     // ref: https://github.com/ggml-org/llama.cpp/discussions/7416
     //      https://github.com/ggml-org/llama.cpp/pull/17945
     static float yarn_attn_factor_adjust(float attn_factor, float freq_scale, float ext_factor);
+
+    bool use_mrope() const;
 };
 
 static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams must be trivially copyable");

src/llama-model.cpp

@@ -1685,7 +1685,8 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             } break;
         case LLM_ARCH_GLM4:
             {
-                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,    hparams.f_norm_rms_eps);
+                ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
                 switch (hparams.n_layer) {
                     case 40: type = LLM_TYPE_9B; break;
                     case 61: type = LLM_TYPE_32B; break;
@@ -1694,8 +1695,9 @@ void llama_model::load_hparams(llama_model_loader & ml) {
             } break;
         case LLM_ARCH_GLM4_MOE:
             {
-                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
-                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+                ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,     hparams.n_ff_exp);
+                ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,    hparams.f_norm_rms_eps);
+                ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
 
                 // MoE parameters
                 ml.get_key(LLM_KV_EXPERT_COUNT,                hparams.n_expert);
@@ -7758,7 +7760,6 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_DEEPSEEK2:
         case LLM_ARCH_PLM:
         case LLM_ARCH_CHATGLM:
-        case LLM_ARCH_GLM4:
         case LLM_ARCH_GRANITE:
         case LLM_ARCH_GRANITE_MOE:
         case LLM_ARCH_GRANITE_HYBRID:
@@ -7820,7 +7821,6 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_LFM2:
         case LLM_ARCH_LFM2MOE:
         case LLM_ARCH_SMALLTHINKER:
-        case LLM_ARCH_GLM4_MOE:
         case LLM_ARCH_SEED_OSS:
         case LLM_ARCH_GROVEMOE:
         case LLM_ARCH_APERTUS:
@@ -7837,6 +7837,11 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
         case LLM_ARCH_QWEN3VLMOE:
             return LLAMA_ROPE_TYPE_IMROPE;
 
+        case LLM_ARCH_GLM4:
+            return model->hparams.use_mrope() ? LLAMA_ROPE_TYPE_MROPE : LLAMA_ROPE_TYPE_NORM;
+        case LLM_ARCH_GLM4_MOE:
+            return model->hparams.use_mrope() ? LLAMA_ROPE_TYPE_MROPE : LLAMA_ROPE_TYPE_NEOX;
+
         // all model arches should be listed explicitly here
         case LLM_ARCH_UNKNOWN:
             GGML_ABORT("unknown architecture");

src/models/glm4-moe.cpp

@@ -5,11 +5,20 @@ llm_build_glm4_moe::llm_build_glm4_moe(const llama_model & model, const llm_grap
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
 
+    int sections[4];
+    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
+
     ggml_tensor * cur;
     ggml_tensor * inpL;
 
     inpL = build_inp_embd(model.tok_embd);
 
+    bool use_mrope = hparams.use_mrope();
+    if (ubatch.embd && !use_mrope) {
+        // unfortunately, we need to forcefully stop here, to avoid users complaining about wrong results
+        GGML_ABORT("This GGUF does not support multimodal. Please reconvert it.");
+    }
+
     // inp_pos - contains the positions
     ggml_tensor * inp_pos = build_inp_pos();
 
@@ -60,17 +69,25 @@ llm_build_glm4_moe::llm_build_glm4_moe(const llama_model & model, const llm_grap
                 Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
                 cb(Kcur, "Kcur_normed", il);
             }
-            Qcur = ggml_rope_ext(
-                    ctx0, Qcur, inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                    );
-
-            Kcur = ggml_rope_ext(
-                    ctx0, Kcur, inp_pos, nullptr,
-                    n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                    ext_factor, attn_factor, beta_fast, beta_slow
-                    );
+
+            if (use_mrope) {
+                Qcur = ggml_rope_multi(ctx0, Qcur, inp_pos, nullptr,
+                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow);
+
+                Kcur = ggml_rope_multi(ctx0, Kcur, inp_pos, nullptr,
+                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow);
+            } else {
+                // Normal RoPE
+                Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot,
+                                    rope_type, n_ctx_orig, freq_base, freq_scale,
+                                    ext_factor, attn_factor, beta_fast, beta_slow);
+
+                Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot,
+                                    rope_type, n_ctx_orig, freq_base, freq_scale,
+                                    ext_factor, attn_factor, beta_fast, beta_slow);
+            }
 
             cb(Qcur, "Qcur", il);
             cb(Kcur, "Kcur", il);