[Model] Support MiniCPM-V 4.6

convert_hf_to_gguf.py

@@ -1331,6 +1331,9 @@ def get_vocab_base_pre(self, tokenizer) -> str:
         if chkhsh == "d4540891389ea895b53b399da6ac824becc30f2fba0e9ddbb98f92e55ca0e97c":
             # ref: https://huggingface.co/Qwen/Qwen3-Embedding-0.6B
             res = "qwen2"
+        if chkhsh == "1444df51289cfa8063b96f0e62b1125440111bc79a52003ea14b6eac7016fd5f":
+            # ref: https://huggingface.co/openbmb/MiniCPM-V-4_6
+            res = "qwen35"
         if chkhsh == "66b8d4e19ab16c3bfd89bce5d785fb7e0155e8648708a1f42077cb9fe002c273":
             # ref: https://huggingface.co/alvarobartt/grok-2-tokenizer
             res = "grok-2"
@@ -1526,7 +1529,7 @@ def get_vocab_base_pre(self, tokenizer) -> str:
         if chkhsh == "6c81ce329e0802883b22eabab0d3fa48357337ef1ecb45443828bf1f6254833f":
             # ref: https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B
             res = "exaone-moe"
-        if chkhsh == "d30d75d9059f1aa2c19359de71047b3ae408c70875e8a3ccf8c5fba56c9d8af4":
+        if chkhsh == "1444df51289cfa8063b96f0e62b1125440111bc79a52003ea14b6eac7016fd5f":
             # ref: https://huggingface.co/Qwen/Qwen3.5-9B-Instruct
             res = "qwen35"
         if chkhsh == "b4b8ca1f9769494fbd956ebc4c249de6131fb277a4a3345a7a92c7dd7a55808d":
@@ -5432,16 +5435,107 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter
         yield from super().modify_tensors(data_torch, name, bid)
 
 
+class _Qwen35MRopeMixin:
+    # Qwen3.5 always applies interleaved MRoPE (see Qwen3_5RotaryEmbedding in transformers);
+    # the upstream default mrope_section is [11, 11, 10] and llama.cpp's QWEN35 / QWEN35MOE
+    # loaders treat qwen35.rope.dimension_sections as required, so make sure it is always
+    # written even when a particular checkpoint omits the field in `rope_parameters`.
+    _QWEN35_DEFAULT_MROPE_SECTION = [11, 11, 10, 0]
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()  # type: ignore[misc]
+        if "mrope_section" not in self.rope_parameters:  # type: ignore[attr-defined]
+            self.gguf_writer.add_rope_dimension_sections(self._QWEN35_DEFAULT_MROPE_SECTION)  # type: ignore[attr-defined]
+
+
 @ModelBase.register("Qwen3_5ForConditionalGeneration", "Qwen3_5ForCausalLM")
-class Qwen3_5TextModel(_LinearAttentionVReorderBase):
+class Qwen3_5TextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
     model_arch = gguf.MODEL_ARCH.QWEN35
 
 
 @ModelBase.register("Qwen3_5MoeForConditionalGeneration", "Qwen3_5MoeForCausalLM")
-class Qwen3_5MoeTextModel(_LinearAttentionVReorderBase):
+class Qwen3_5MoeTextModel(_Qwen35MRopeMixin, _LinearAttentionVReorderBase):
     model_arch = gguf.MODEL_ARCH.QWEN35MOE
 
 
+# MiniCPM-V 4.6: text tower is Qwen3.5 (linear+full hybrid attention) wrapped under
+# `model.language_model.*`; vision tower is SigLIP + a window-attention ViT merger
+# + a final DownsampleMLP merger. The same HF arch is registered twice below: once as
+# the LM (text mode) and once as the mmproj (vision mode), mirroring the Qwen3-VL setup.
+
+@ModelBase.register("MiniCPMV4_6ForConditionalGeneration")
+class MiniCPMV4_6TextModel(Qwen3_5TextModel):
+    model_arch = gguf.MODEL_ARCH.QWEN35
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # drop vision tower / multimodal merger tensors -- they belong to the mmproj file
+        if name.startswith(("model.vision_tower.", "model.merger.")):
+            return
+        # MTP tensors are not used at inference yet; align with Qwen3Next behaviour
+        if name.startswith("mtp"):
+            return
+        # strip the language-model wrapper so the underlying Qwen3.5 tensor mapping matches
+        if name.startswith("model.language_model."):
+            name = "model." + name[len("model.language_model."):]
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("MiniCPMV4_6ForConditionalGeneration")
+class MiniCPMV4_6VisionModel(MmprojModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        if self.hparams_vision is not None:
+            # In MiniCPM-V 4.6 `vision_config.image_size` (980) describes the SigLIP
+            # positional embedding bucket grid (70 x 70), while the per-slice processing
+            # resolution is the preprocessor's `scale_resolution` (typically 448).
+            # The CLIP loader in tools/mtmd/clip.cpp consumes `clip.vision.image_size`
+            # as the slice size and warmup resolution, so report `scale_resolution` there
+            # to match the upstream MiniCPMV4_6ImageProcessorPil slicing rules.
+            scale_resolution = self.preprocessor_config.get("scale_resolution")
+            if scale_resolution is not None:
+                self.hparams_vision["image_size"] = int(scale_resolution)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        assert self.hparams_vision is not None
+
+        # projector type string is consumed by clip_projector_type_from_string() in clip.cpp
+        # (mapped to PROJECTOR_TYPE_MINICPMV4_6).
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.MINICPMV4_6)
+
+        # legacy version tag, used by mtmd.cpp to pick the slice template (MINICPMV_2_6).
+        # The clip loader reads this field via gguf_get_val_i32, so it must be written as int32.
+        self.gguf_writer.add_int32("clip.minicpmv_version", 46)
+        # fixed merger output token count per slice for the default 16x downsample mode.
+        self.gguf_writer.add_uint32("clip.minicpmv_query_num", 64)
+
+        # ViT layer index after which the window-attention merger is applied
+        insert_layer_id = int(self.global_config.get(
+            "insert_layer_id", self.hparams_vision.get("insert_layer_id", 6)))
+        self.gguf_writer.add_uint32("clip.vision.insert_layer_id", insert_layer_id)
+
+        # SigLIP vision body uses gelu_pytorch_tanh, which matches ggml_gelu (tanh approx).
+        self.gguf_writer.add_vision_use_gelu(True)
+        self.gguf_writer.add_vision_attention_layernorm_eps(
+            self.hparams_vision.get("layer_norm_eps", 1e-6))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # text tower / lm_head / MTP -> belong to the LM file
+        if name.startswith(("model.language_model.", "lm_head.")) or name.startswith("mtp"):
+            return
+
+        # final merger and ViT merger
+        if name.startswith(("model.merger.", "model.vision_tower.vit_merger.")):
+            yield from super().modify_tensors(data_torch, name, bid)
+            return
+
+        # SigLIP vision body
+        if name.startswith("model.vision_tower."):
+            name = "vision_tower.vision_model." + name[len("model.vision_tower."):]
+            yield from super().modify_tensors(data_torch, name, bid)
+            return
+
+
 @ModelBase.register("GPT2LMHeadModel")
 class GPT2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.GPT2

convert_hf_to_gguf_update.py

@@ -175,6 +175,7 @@ class TOKENIZER_TYPE(IntEnum):
     {"name": "falcon-h1", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/Falcon-H1-34B-Base", "chkhsh": "48f8e02c0359c0bbdd82f26909171fac1c18a457bb47573ed1fe3bbb2c1cfd4b"},
     {"name": "kimi-k2",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/moonshotai/Kimi-K2-Base",   "chkhsh": "81212dc7cdb7e0c1074ca62c5aeab0d43c9f52b8a737be7b12a777c953027890"},
     {"name": "qwen2",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen3-Embedding-0.6B", "chkhsh": "d4540891389ea895b53b399da6ac824becc30f2fba0e9ddbb98f92e55ca0e97c"},
+    {"name": "qwen35",    "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openbmb/MiniCPM-V-4_6", "chkhsh": "1444df51289cfa8063b96f0e62b1125440111bc79a52003ea14b6eac7016fd5f"},
     {"name": "grok-2",    "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/alvarobartt/grok-2-tokenizer", "chkhsh": "66b8d4e19ab16c3bfd89bce5d785fb7e0155e8648708a1f42077cb9fe002c273"},
     # jina-v2-de variants
     {"name": "jina-v2-de", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/aari1995/German_Semantic_V3", "chkhsh": "b3d1dd861f1d4c5c0d2569ce36baf3f90fe8a102db3de50dd71ff860d91be3df"},

docs/multimodal/minicpmv4.6.md

@@ -0,0 +1,49 @@
+## MiniCPM-V 4.6
+
+### Prepare models and code
+
+Download [MiniCPM-V-4_6](https://huggingface.co/openbmb/MiniCPM-V-4_6) PyTorch model from huggingface to "MiniCPM-V-4_6" folder.
+
+The model must be the standard `transformers` v5.7.0+ checkpoint (no `trust_remote_code`); the architecture in `config.json` is `MiniCPMV4_6ForConditionalGeneration` with a `qwen3_5_text` text model and a SigLIP-based vision tower plus a window-attention `vit_merger`.
+
+### Build llama.cpp
+
+If there are differences in usage, please refer to the official build [documentation](https://github.com/ggml-org/llama.cpp/blob/master/docs/build.md)
+
+Clone llama.cpp:
+```bash
+git clone https://github.com/ggml-org/llama.cpp
+cd llama.cpp
+```
+
+Build llama.cpp using `CMake`:
+```bash
+cmake -B build
+cmake --build build --config Release
+```
+
+
+### Usage of MiniCPM-V 4.6
+
+Unlike older MiniCPM-V variants, MiniCPM-V 4.6 is converted directly through `convert_hf_to_gguf.py`. The same script is invoked twice on the original Hugging Face directory: once to produce the language-model GGUF and once with `--mmproj` to produce the multimodal projector GGUF.
+
+```bash
+# language model
+python ./convert_hf_to_gguf.py ../MiniCPM-V-4_6 --outfile ../MiniCPM-V-4_6/ggml-model-f16.gguf
+
+# multimodal projector (vision tower + window-attention vit_merger + DownsampleMLP merger)
+python ./convert_hf_to_gguf.py ../MiniCPM-V-4_6 --mmproj --outfile ../MiniCPM-V-4_6/mmproj-model-f16.gguf
+
+# optional: quantize to Q4_K_M
+./build/bin/llama-quantize ../MiniCPM-V-4_6/ggml-model-f16.gguf ../MiniCPM-V-4_6/ggml-model-Q4_K_M.gguf Q4_K_M
+```
+
+
+Inference on Linux or Mac
+```bash
+# run in single-turn mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-V-4_6/ggml-model-f16.gguf --mmproj ../MiniCPM-V-4_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+
+# run in conversation mode
+./build/bin/llama-mtmd-cli -m ../MiniCPM-V-4_6/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-4_6/mmproj-model-f16.gguf
+```

gguf-py/gguf/constants.py

@@ -767,6 +767,14 @@ class MODEL_TENSOR(IntEnum):
     V_DS_NORM            = auto() # qwen3vl
     V_DS_FC1             = auto() # qwen3vl
     V_DS_FC2             = auto() # qwen3vl
+    V_MERGER_LN1         = auto() # minicpmv4_6
+    V_MERGER_ATTN_Q      = auto() # minicpmv4_6
+    V_MERGER_ATTN_K      = auto() # minicpmv4_6
+    V_MERGER_ATTN_V      = auto() # minicpmv4_6
+    V_MERGER_ATTN_O      = auto() # minicpmv4_6
+    V_MERGER_DS_LN       = auto() # minicpmv4_6
+    V_MERGER_DS_UP       = auto() # minicpmv4_6
+    V_MERGER_DS_DOWN     = auto() # minicpmv4_6
     V_MM_POST_FC_NORM    = auto() # cogvlm
     V_MM_UP              = auto() # cogvlm
     V_MM_DOWN            = auto() # cogvlm
@@ -1251,6 +1259,14 @@ class MODEL_TENSOR(IntEnum):
     MODEL_TENSOR.V_DS_NORM:                 "v.deepstack.{bid}.norm",
     MODEL_TENSOR.V_DS_FC1:                  "v.deepstack.{bid}.fc1",
     MODEL_TENSOR.V_DS_FC2:                  "v.deepstack.{bid}.fc2",
+    MODEL_TENSOR.V_MERGER_LN1:              "v.vit_merger.ln1",
+    MODEL_TENSOR.V_MERGER_ATTN_Q:           "v.vit_merger.attn_q",
+    MODEL_TENSOR.V_MERGER_ATTN_K:           "v.vit_merger.attn_k",
+    MODEL_TENSOR.V_MERGER_ATTN_V:           "v.vit_merger.attn_v",
+    MODEL_TENSOR.V_MERGER_ATTN_O:           "v.vit_merger.attn_out",
+    MODEL_TENSOR.V_MERGER_DS_LN:            "v.vit_merger.ds_ln",
+    MODEL_TENSOR.V_MERGER_DS_UP:            "v.vit_merger.ds_ffn_up",
+    MODEL_TENSOR.V_MERGER_DS_DOWN:          "v.vit_merger.ds_ffn_down",
     MODEL_TENSOR.V_MM_POST_FC_NORM:         "mm.post_fc_norm", # cogvlm
     MODEL_TENSOR.V_MM_UP:                   "mm.up",
     MODEL_TENSOR.V_MM_DOWN:                 "mm.down",
@@ -1403,6 +1419,14 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.V_DS_NORM,
         MODEL_TENSOR.V_DS_FC1,
         MODEL_TENSOR.V_DS_FC2,
+        MODEL_TENSOR.V_MERGER_LN1,
+        MODEL_TENSOR.V_MERGER_ATTN_Q,
+        MODEL_TENSOR.V_MERGER_ATTN_K,
+        MODEL_TENSOR.V_MERGER_ATTN_V,
+        MODEL_TENSOR.V_MERGER_ATTN_O,
+        MODEL_TENSOR.V_MERGER_DS_LN,
+        MODEL_TENSOR.V_MERGER_DS_UP,
+        MODEL_TENSOR.V_MERGER_DS_DOWN,
         MODEL_TENSOR.V_MM_POST_FC_NORM,
         MODEL_TENSOR.V_MM_UP,
         MODEL_TENSOR.V_MM_DOWN,
@@ -4158,6 +4182,7 @@ class VisionProjectorType:
     NEMOTRON_V2_VL = "nemotron_v2_vl"
     HUNYUANOCR     = "hunyuanocr"
     HUNYUANVL      = "hunyuanvl"
+    MINICPMV4_6    = "minicpmv4_6"
 
 
 # Items here are (block size, type size)

gguf-py/gguf/tensor_mapping.py

@@ -1690,6 +1690,7 @@ class TensorNameMap:
             "mlp_AR.pre_norm", # PaddleOCR-VL
             "merger.ln_q",
             "vision_tower.merger.ln_q", # dots.ocr
+            "model.merger.mlp.0.pre_norm", # minicpmv4_6
         ),
 
         MODEL_TENSOR.V_MM_SOFT_EMB_NORM: (
@@ -1763,6 +1764,38 @@ class TensorNameMap:
             "model.visual.deepstack_merger_list.{bid}.linear_fc2", # deepstack in qwen3vl
         ),
 
+        MODEL_TENSOR.V_MERGER_LN1: (
+            "model.vision_tower.vit_merger.layer_norm1", # minicpmv4_6
+        ),
+
+        MODEL_TENSOR.V_MERGER_ATTN_Q: (
+            "model.vision_tower.vit_merger.self_attn.q_proj", # minicpmv4_6
+        ),
+
+        MODEL_TENSOR.V_MERGER_ATTN_K: (
+            "model.vision_tower.vit_merger.self_attn.k_proj", # minicpmv4_6
+        ),
+
+        MODEL_TENSOR.V_MERGER_ATTN_V: (
+            "model.vision_tower.vit_merger.self_attn.v_proj", # minicpmv4_6
+        ),
+
+        MODEL_TENSOR.V_MERGER_ATTN_O: (
+            "model.vision_tower.vit_merger.self_attn.out_proj", # minicpmv4_6
+        ),
+
+        MODEL_TENSOR.V_MERGER_DS_LN: (
+            "model.vision_tower.vit_merger.pre_norm", # minicpmv4_6
+        ),
+
+        MODEL_TENSOR.V_MERGER_DS_UP: (
+            "model.vision_tower.vit_merger.linear_1", # minicpmv4_6
+        ),
+
+        MODEL_TENSOR.V_MERGER_DS_DOWN: (
+            "model.vision_tower.vit_merger.linear_2", # minicpmv4_6
+        ),
+
         MODEL_TENSOR.V_SAM_POS_EMBD: (
             "model.sam_model.pos_embed",
         ),
@@ -1822,11 +1855,13 @@ class TensorNameMap:
         MODEL_TENSOR.V_MM_UP: (
             "model.vision.linear_proj.dense_h_to_4h", # cogvlm
             "visual.merger.up_proj", # glm4v
+            "model.merger.mlp.0.linear_1", # minicpmv4_6
         ),
 
         MODEL_TENSOR.V_MM_DOWN: (
             "model.vision.linear_proj.dense_4h_to_h", # cogvlm
             "visual.merger.down_proj", # glm4v
+            "model.merger.mlp.0.linear_2", # minicpmv4_6
         ),
 
         MODEL_TENSOR.V_MM_GATE: (

tools/mtmd/README.md

@@ -49,6 +49,7 @@ For the following models, you can use `convert_hf_to_gguf.py` with `--mmproj` fl
 - Qwen 2 VL and Qwen 2.5 VL (from [Qwen](https://huggingface.co/Qwen))
 - [Mistral Small 3.1 24B](https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503)
 - InternVL 2.5 and InternVL 3 from [OpenGVLab](https://huggingface.co/OpenGVLab) (note: we don't support conversion of `InternVL3-*-hf` model, only non-HF version is supported ; `InternLM2Model` **text** model is not supported)
+- [MiniCPM-V 4.6](https://huggingface.co/openbmb/MiniCPM-V-4_6) ; See the guide [here](../../docs/multimodal/minicpmv4.6.md) - requires the standard `transformers` v5.7.0+ checkpoint
 
 For older models, please refer to the relevant guide for instructions on how to obtain or create them:
 
@@ -60,4 +61,7 @@ NOTE: conversion scripts are located under `tools/mtmd/legacy-models`
 - [MiniCPM-V 2.5](../../docs/multimodal/minicpmv2.5.md)
 - [MiniCPM-V 2.6](../../docs/multimodal/minicpmv2.6.md)
 - [MiniCPM-o 2.6](../../docs/multimodal/minicpmo2.6.md)
+- [MiniCPM-V 4.0](../../docs/multimodal/minicpmv4.0.md)
+- [MiniCPM-o 4.0](../../docs/multimodal/minicpmo4.0.md)
+- [MiniCPM-V 4.5](../../docs/multimodal/minicpmv4.5.md)
 - [IBM Granite Vision](../../docs/multimodal/granitevision.md)

tools/mtmd/clip-impl.h

@@ -126,6 +126,19 @@
 #define TN_MINICPMV_ATTN       "resampler.attn.%s.%s"
 #define TN_MINICPMV_LN         "resampler.ln_%s.%s"
 
+// MiniCPM-V 4.6 ViT merger (window attention + MLP downsample),
+// matching the upstream `vit_merger` module name in transformers.
+#define TN_VIT_MERGER_LN1      "v.vit_merger.ln1.%s"
+#define TN_VIT_MERGER_ATTN_Q   "v.vit_merger.attn_q.%s"
+#define TN_VIT_MERGER_ATTN_K   "v.vit_merger.attn_k.%s"
+#define TN_VIT_MERGER_ATTN_V   "v.vit_merger.attn_v.%s"
+#define TN_VIT_MERGER_ATTN_O   "v.vit_merger.attn_out.%s"
+#define TN_VIT_MERGER_DS_LN    "v.vit_merger.ds_ln.%s"
+#define TN_VIT_MERGER_DS_UP    "v.vit_merger.ds_ffn_up.%s"
+#define TN_VIT_MERGER_DS_DOWN  "v.vit_merger.ds_ffn_down.%s"
+
+#define KEY_INSERT_LAYER_ID "clip.vision.insert_layer_id"
+
 #define TN_GLM_ADAPER_CONV      "adapter.conv.%s"
 #define TN_GLM_ADAPTER_LINEAR   "adapter.linear.linear.%s"
 #define TN_GLM_ADAPTER_NORM_1   "adapter.linear.norm1.%s"
@@ -304,6 +317,7 @@ enum projector_type {
     PROJECTOR_TYPE_NEMOTRON_V2_VL,
     PROJECTOR_TYPE_HUNYUANOCR,
     PROJECTOR_TYPE_HUNYUANVL,
+    PROJECTOR_TYPE_MINICPMV4_6,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
@@ -351,6 +365,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_NEMOTRON_V2_VL, "nemotron_v2_vl"},
     { PROJECTOR_TYPE_HUNYUANOCR, "hunyuanocr"},
     { PROJECTOR_TYPE_HUNYUANVL,  "hunyuanvl"},
+    { PROJECTOR_TYPE_MINICPMV4_6, "minicpmv4_6"},
 };
 
 static projector_type clip_projector_type_from_string(const std::string & str) {

tools/mtmd/clip-model.h

@@ -104,6 +104,7 @@ struct clip_hparams {
     bool has_llava_projector = false;
     int minicpmv_version = 0;
     int32_t minicpmv_query_num = 0;         // MiniCPM-V query number
+    int32_t insert_layer_id   = 0;          // MiniCPM-V 4.6 ViT merger insertion layer
 
     // custom value provided by user, can be undefined if not set
     int32_t custom_image_min_tokens = -1;
@@ -403,6 +404,24 @@ struct clip_model {
     ggml_tensor * mm_model_ln_post_w = nullptr;
     ggml_tensor * mm_model_ln_post_b = nullptr;
 
+    // MiniCPM-V 4.6 ViT merger (window self-attention + ViT MLP downsample)
+    ggml_tensor * vit_merger_ln1_w     = nullptr;
+    ggml_tensor * vit_merger_ln1_b     = nullptr;
+    ggml_tensor * vit_merger_attn_q_w  = nullptr;
+    ggml_tensor * vit_merger_attn_q_b  = nullptr;
+    ggml_tensor * vit_merger_attn_k_w  = nullptr;
+    ggml_tensor * vit_merger_attn_k_b  = nullptr;
+    ggml_tensor * vit_merger_attn_v_w  = nullptr;
+    ggml_tensor * vit_merger_attn_v_b  = nullptr;
+    ggml_tensor * vit_merger_attn_o_w  = nullptr;
+    ggml_tensor * vit_merger_attn_o_b  = nullptr;
+    ggml_tensor * vit_merger_ds_ln_w   = nullptr;
+    ggml_tensor * vit_merger_ds_ln_b   = nullptr;
+    ggml_tensor * vit_merger_ds_up_w   = nullptr;
+    ggml_tensor * vit_merger_ds_up_b   = nullptr;
+    ggml_tensor * vit_merger_ds_down_w = nullptr;
+    ggml_tensor * vit_merger_ds_down_b = nullptr;
+
     // gemma3
     ggml_tensor * mm_input_proj_w = nullptr;
     ggml_tensor * mm_soft_emb_norm_w = nullptr;