model: Add support for SenseNova U1 model (#28919)

* model : add SenseNova U1 understanding support

Add support for the dense SenseNova U1 understanding branch, including
the U1.5-8B-MoT model.

- Add HF-to-GGUF conversion for NEOChatModel, excluding MoT generation
  tensors and concatenating temporal and spatial Q/K normalization
  weights.
- Register the sensenova_u1 GGUF architecture and custom spatial RoPE
  metadata.
- Implement temporal/spatial Q/K normalization with 4-axis RoPE.
- Keep U1-specific text position handling in the model implementation.
- Add synthetic architecture coverage for sensenova_u1.

Only the dense understanding branch is supported; generation
branches are intentionally rejected during conversion.

* conversion : treat one-expert SenseNova U1 configs as dense

* mtmd : add SenseNova U1 vision support

Add conversion support for the SenseNova U1 vision encoder and implement
its CPU mtmd projector graph.

The graph applies patch embedding, GELU, adjacent-pair 2D RoPE, spatial
downsampling, and dense projection for dynamic-resolution image inputs.

* mtmd : fix SenseNova U1 vision token layout

* model : clean up SenseNova U1 RoPE metadata

Replace the hard-coded spatial RoPE key with shared GGUF and LLM_KV
definitions used by conversion, model loading, and architecture tests.

Clarify why SenseNova U1 retains its custom text-position input: U1 uses
[t, 0, 0, 0], while the generic M-RoPE input uses [t, t, t, 0].

* mtmd : simplify SenseNova U1 projector integration

Map the SenseNova U1 dense projection to the standard mm.0 weight and
bias tensors and reuse the shared projector fields in the loader and graph.

Read vision parameters from the checkpoint configuration and remove the
U1-specific tensor names and model fields while preserving its strided
projection, bias, and row-major token layout.

* mtmd : fix SenseNova U1 image positions and attention

* fix spacing

* mtmd : restore generic test input chunks

* mtmd : restore SenseNova U1 integration after upstream merge

* Update conversion/sensenova_u1.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

* Update src/models/sensenova-u1.cpp

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

* model : remove SenseNova U1 BF16 activation casts

Use standard backend precision handling for QKV, FFN, and output
projections instead of casting inputs based on the output tensor type.

Assisted-by: Codex

* Update src/models/sensenova-u1.cpp

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

conversion/__init__.py

@@ -17,6 +17,7 @@ __all__ = [
 
 
 TEXT_MODEL_MAP: dict[str, str] = {
+    "NEOChatModel": "sensenova_u1",
     "AfmoeForCausalLM": "afmoe",
     "LagunaForCausalLM": "laguna",
     "ApertusForCausalLM": "llama",
@@ -367,6 +368,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
     "MuseSparkForConditionalGeneration": "muse_spark",
     "PaddleOCRVisionModel": "ernie",
     "Phi4ForCausalLMV": "phi",
+    "NEOVisionModel": "sensenova_u1",
     "Qwen2AudioForConditionalGeneration": "ultravox",
     "Qwen2VLForConditionalGeneration": "qwenvl",
     "Qwen2VLModel": "qwenvl",

conversion/base.py

@@ -3143,7 +3143,7 @@ def get_model_architecture(hparams: dict[str, Any], model_type: ModelType) -> st
     # Kimi-K3's text_config reports "KimiLinearForCausalLM", which is the older
     # Kimi-Linear-48B architecture and cannot load K3 (no attention residuals,
     # latent MoE, situ, ...). Route on the top-level architecture instead.
-    if model_type == ModelType.TEXT and arch in ("StepVLForConditionalGeneration", "Sarashina2VisionForCausalLM", "Exaone4_5_ForConditionalGeneration", "Step3p7ForConditionalGeneration", "KimiK3ForConditionalGeneration"):
+    if model_type == ModelType.TEXT and arch in ("NEOChatModel", "StepVLForConditionalGeneration", "Sarashina2VisionForCausalLM", "Exaone4_5_ForConditionalGeneration", "Step3p7ForConditionalGeneration", "KimiK3ForConditionalGeneration"):
         return arch
 
     # if "architectures" is found in the sub-config, use that instead

conversion/sensenova_u1.py

@@ -0,0 +1,114 @@
+from __future__ import annotations
+
+from pathlib import Path
+
+import torch
+
+from .base import MmprojModel, ModelBase, gguf
+from .qwen import Qwen2Model
+
+
+@ModelBase.example("sensenova/SenseNova-U1.5-8B-MoT")
+@ModelBase.register("NEOChatModel")
+class SenseNovaU1Model(Qwen2Model):
+    model_arch = gguf.MODEL_ARCH.SENSENOVA_U1
+
+    @classmethod
+    def filter_tensors(cls, item):
+        name, _ = item
+        if not name.startswith("language_model.") or "_mot_gen" in name:
+            return None
+        return super().filter_tensors(item)
+
+    def set_gguf_parameters(self):
+        if int(self.hparams.get("num_experts") or 0) > 1:
+            raise ValueError("SenseNova U1 conversion currently supports the dense understanding branch")
+        if self.rope_parameters.get("rope_type", "default") not in ("default", None):
+            raise ValueError("SenseNova U1 scaled RoPE is not supported")
+        if self.hparams.get("attention_bias", False):
+            raise ValueError("SenseNova U1 attention bias is not supported")
+        if self.hparams.get("use_sliding_window", False):
+            raise ValueError("SenseNova U1 sliding attention is not supported")
+        super().set_gguf_parameters()
+        self.gguf_writer.add_rope_freq_base_spatial(self.hparams["rope_theta_hw"])
+
+    def get_tensors(self):
+        for name, gen in self.model_tensors.items():
+            if name.endswith((".q_norm_hw.weight", ".k_norm_hw.weight")):
+                continue
+            data = gen()
+            if name.endswith((".q_norm.weight", ".k_norm.weight")):
+                spatial = self.model_tensors[name.replace("_norm.weight", "_norm_hw.weight")]()
+                # Keep one weight vector; the graph normalizes its two halves independently.
+                data = torch.cat((data, spatial), dim=0)
+            yield name, data
+
+
+@ModelBase.register("NEOVisionModel")
+@ModelBase.example("sensenova/SenseNova-U1.5-8B-MoT")
+class SenseNovaU1VisionModel(MmprojModel):
+    def __init__(self, dir_model: Path, *args, **kwargs):
+        hparams = kwargs.pop("hparams", None)
+        if hparams is None:
+            hparams = ModelBase.load_hparams(dir_model, is_mistral_format=False)
+        else:
+            hparams = dict(hparams)
+
+        vision_config = dict(hparams["vision_config"])
+        # NEOVisionModel has no transformer blocks
+        # these defaults satisfy the generic mmproj metadata contract without replacing checkpoint values
+        vision_config.setdefault("image_size", 0)
+        vision_config.setdefault("intermediate_size", 0)
+        vision_config.setdefault("num_hidden_layers", 0)
+        vision_config.setdefault("num_attention_heads", 1)
+        hparams["vision_config"] = vision_config
+        super().__init__(dir_model, *args, hparams=hparams, **kwargs)
+
+        self.preprocessor_config.setdefault("image_mean", [0.485, 0.456, 0.406])
+        self.preprocessor_config.setdefault("image_std", [0.229, 0.224, 0.225])
+
+    @classmethod
+    def filter_tensors(cls, item):
+        name, gen = item
+        if not name.startswith("vision_model."):
+            return None
+        return super().filter_tensors((name, gen))
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_clip_projector_type("sensenova_u1")
+        vision_config = self.hparams["vision_config"]
+        # These defaults match the known U1.5 checkpoint when older configs omit fields.
+        hidden_size = int(vision_config.get("hidden_size", 1024))
+        min_pixels = int(vision_config.get("min_pixels", 65536))
+        max_pixels = int(vision_config.get("max_pixels", 16777216))
+        downsample_ratio = float(vision_config.get("downsample_ratio", 0.5))
+        if downsample_ratio <= 0:
+            raise ValueError(f"SenseNova U1 downsample_ratio must be positive, got {downsample_ratio}")
+        merge_size = round(1.0 / downsample_ratio)
+        if abs((1.0 / downsample_ratio) - merge_size) > 1e-6:
+            raise ValueError(f"SenseNova U1 downsample_ratio must have an integer reciprocal, got {downsample_ratio}")
+        norm_eps = float(vision_config.get("layer_norm_eps", vision_config.get("rms_norm_eps", 1e-6)))
+        self.gguf_writer.add_vision_head_dim(hidden_size)
+        self.gguf_writer.add_vision_attention_layernorm_eps(norm_eps)
+        self.gguf_writer.add_vision_min_pixels(min_pixels)
+        self.gguf_writer.add_vision_max_pixels(max_pixels)
+        self.gguf_writer.add_vision_spatial_merge_size(merge_size)
+
+    def modify_tensors(self, data_torch, name, bid):
+        tensor_names = {
+            "vision_model.embeddings.patch_embedding.weight": "v.patch_embd.weight",
+            "vision_model.embeddings.patch_embedding.bias":   "v.patch_embd.bias",
+            "vision_model.embeddings.dense_embedding.weight": self.format_tensor_name(gguf.MODEL_TENSOR.V_MMPROJ, 0),
+            "vision_model.embeddings.dense_embedding.bias": self.format_tensor_name(
+                gguf.MODEL_TENSOR.V_MMPROJ, 0, suffix=".bias"
+            ),
+        }
+        if name not in tensor_names:
+            raise ValueError(f"Unexpected SenseNova U1 vision tensor: {name!r}")
+        return [(tensor_names[name], data_torch)]
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.V_MMPROJ, 0):
+            return gguf.GGMLQuantizationType.F16 if self.ftype == gguf.LlamaFileType.MOSTLY_F16 else gguf.GGMLQuantizationType.F32
+        return super().tensor_force_quant(name, new_name, bid, n_dims)

gguf-py/gguf/constants.py

@@ -263,6 +263,7 @@ class Keys:
         DIMENSION_SECTIONS        = "{arch}.rope.dimension_sections"
         SECTION_ORDER             = "{arch}.rope.section_order"
         FREQ_BASE                 = "{arch}.rope.freq_base"
+        FREQ_BASE_SPATIAL         = "{arch}.rope.freq_base_spatial"
         FREQ_BASE_SWA             = "{arch}.rope.freq_base_swa"
         SCALING_TYPE              = "{arch}.rope.scaling.type"
         SCALING_FACTOR            = "{arch}.rope.scaling.factor"
@@ -573,6 +574,7 @@ class MODEL_ARCH(IntEnum):
     QWEN2            = auto()
     QWEN2MOE         = auto()
     QWEN2VL          = auto()
+    SENSENOVA_U1     = auto()
     QWEN3            = auto()
     QWEN3MOE         = auto()
     QWEN3NEXT        = auto()
@@ -1390,6 +1392,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.QWEN2:            "qwen2",
     MODEL_ARCH.QWEN2MOE:         "qwen2moe",
     MODEL_ARCH.QWEN2VL:          "qwen2vl",
+    MODEL_ARCH.SENSENOVA_U1:     "sensenova_u1",
     MODEL_ARCH.QWEN3:            "qwen3",
     MODEL_ARCH.QWEN3MOE:         "qwen3moe",
     MODEL_ARCH.QWEN3NEXT:        "qwen3next",
@@ -2916,6 +2919,24 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_SHEXP,
         MODEL_TENSOR.FFN_UP_SHEXP,
     ],
+    MODEL_ARCH.SENSENOVA_U1: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     MODEL_ARCH.QWEN3: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -6485,6 +6506,7 @@ KEY_ATTENTION_LAYERNORM_RMS_EPS = Keys.Attention.LAYERNORM_RMS_EPS
 # RoPE
 KEY_ROPE_DIMENSION_COUNT           = Keys.Rope.DIMENSION_COUNT
 KEY_ROPE_FREQ_BASE                 = Keys.Rope.FREQ_BASE
+KEY_ROPE_FREQ_BASE_SPATIAL         = Keys.Rope.FREQ_BASE_SPATIAL
 KEY_ROPE_SCALING_TYPE              = Keys.Rope.SCALING_TYPE
 KEY_ROPE_SCALING_FACTOR            = Keys.Rope.SCALING_FACTOR
 KEY_ROPE_SCALING_ORIG_CTX_LEN      = Keys.Rope.SCALING_ORIG_CTX_LEN

gguf-py/gguf/gguf_writer.py

@@ -1167,6 +1167,9 @@ class GGUFWriter:
     def add_rope_freq_base(self, value: float) -> None:
         self.add_float32(Keys.Rope.FREQ_BASE.format(arch=self.arch), value)
 
+    def add_rope_freq_base_spatial(self, value: float) -> None:
+        self.add_float32(Keys.Rope.FREQ_BASE_SPATIAL.format(arch=self.arch), value)
+
     def add_rope_scaling_type(self, value: RopeScalingType) -> None:
         self.add_string(Keys.Rope.SCALING_TYPE.format(arch=self.arch), value.value)
 

src/llama-arch.cpp

@@ -33,6 +33,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_QWEN2,            "qwen2"            },
     { LLM_ARCH_QWEN2MOE,         "qwen2moe"         },
     { LLM_ARCH_QWEN2VL,          "qwen2vl"          },
+    { LLM_ARCH_SENSENOVA_U1,     "sensenova_u1"     },
     { LLM_ARCH_QWEN3,            "qwen3"            },
     { LLM_ARCH_QWEN3MOE,         "qwen3moe"         },
     { LLM_ARCH_QWEN3NEXT,        "qwen3next"        },
@@ -339,6 +340,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ROPE_DIMENSION_SECTIONS,        "%s.rope.dimension_sections"              },
     { LLM_KV_ROPE_SECTION_ORDER,             "%s.rope.section_order"                   },
     { LLM_KV_ROPE_FREQ_BASE,                 "%s.rope.freq_base"                       },
+    { LLM_KV_ROPE_FREQ_BASE_SPATIAL,         "%s.rope.freq_base_spatial"               },
     { LLM_KV_ROPE_FREQ_BASE_SWA,             "%s.rope.freq_base_swa"                   },
     { LLM_KV_ROPE_SCALE_LINEAR,              "%s.rope.scale_linear"                    },
     { LLM_KV_ROPE_SCALING_TYPE,              "%s.rope.scaling.type"                    },

src/llama-arch.h

@@ -38,6 +38,7 @@ enum llm_arch {
     LLM_ARCH_QWEN2,
     LLM_ARCH_QWEN2MOE,
     LLM_ARCH_QWEN2VL,
+    LLM_ARCH_SENSENOVA_U1,
     LLM_ARCH_QWEN3,
     LLM_ARCH_QWEN3MOE,
     LLM_ARCH_QWEN3NEXT,
@@ -344,6 +345,7 @@ enum llm_kv {
     LLM_KV_ROPE_DIMENSION_SECTIONS,
     LLM_KV_ROPE_SECTION_ORDER,
     LLM_KV_ROPE_FREQ_BASE,
+    LLM_KV_ROPE_FREQ_BASE_SPATIAL,
     LLM_KV_ROPE_FREQ_BASE_SWA,
     LLM_KV_ROPE_SCALE_LINEAR,
     LLM_KV_ROPE_SCALING_TYPE,

src/llama-model-saver.cpp

@@ -15,6 +15,7 @@
 
 bool llama_model_saver_supports_arch(llm_arch arch) {
     switch (arch) {
+        case LLM_ARCH_SENSENOVA_U1:
         case LLM_ARCH_GEMMA3N:
         case LLM_ARCH_BITNET:
         case LLM_ARCH_T5:

src/llama-model.cpp

@@ -110,6 +110,8 @@ static llama_model * llama_model_mapping(llm_arch arch, const llama_model_params
             return new llama_model_qwen2vl(params);
         case LLM_ARCH_QWEN2MOE:
             return new llama_model_qwen2moe(params);
+        case LLM_ARCH_SENSENOVA_U1:
+            return new llama_model_sensenova_u1(params);
         case LLM_ARCH_QWEN3:
             return new llama_model_qwen3(params);
         case LLM_ARCH_QWEN3MOE:
@@ -3640,6 +3642,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
             // DSV4 DSpark drafters use DeepSeek-V4's normal RoPE; legacy DFlash backbones are NeoX
             return model->hparams.dsv4_hc_mult > 0 ? LLAMA_ROPE_TYPE_NORM : LLAMA_ROPE_TYPE_NEOX;
 
+        case LLM_ARCH_SENSENOVA_U1:
         case LLM_ARCH_QWEN2VL:
         case LLM_ARCH_PADDLEOCR:
             return LLAMA_ROPE_TYPE_MROPE;

src/models/models.h

@@ -614,6 +614,21 @@ struct llama_model_qwen2moe : public llama_model_base {
 };
 
 
+struct llama_model_sensenova_u1 : public llama_model_base {
+    llama_model_sensenova_u1(const struct llama_model_params & params) : llama_model_base(params) {}
+    void load_arch_hparams(llama_model_loader & ml) override;
+    void load_arch_tensors(llama_model_loader & ml) override;
+
+    float rope_freq_base_spatial = 10000.0f;
+
+    struct graph : public llm_graph_context {
+        graph(const llama_model_sensenova_u1 & model, const llm_graph_params & params);
+    };
+
+    std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
+};
+
+
 struct llama_model_qwen3 : public llama_model_base {
     llama_model_qwen3(const struct llama_model_params & params) : llama_model_base(params) {}
     void load_arch_hparams(llama_model_loader & ml) override;

src/models/sensenova-u1.cpp

@@ -0,0 +1,190 @@
+#include "models.h"
+
+namespace {
+class llm_graph_input_pos_sensenova_u1 : public llm_graph_input_pos {
+public:
+    llm_graph_input_pos_sensenova_u1() : llm_graph_input_pos(4) {}
+
+    void set_input(const llama_ubatch * ubatch) override {
+        if (ubatch->pos && pos) {
+            const int64_t n_tokens = ubatch->n_tokens;
+
+            // U1 expects text-token positions in the temporal plane only: [t, 0, 0, 0].
+            // The generic M-RoPE input uses [t, t, t, 0], so U1 needs a custom input class.
+            if (ubatch->token) {
+                std::vector<llama_pos> pos_data(n_tokens*n_pos_per_embd, 0);
+                for (int i = 0; i < n_tokens; ++i) {
+                    pos_data[i] = ubatch->pos[i];
+                }
+                ggml_backend_tensor_set(pos, pos_data.data(), 0, pos_data.size()*ggml_element_size(pos));
+            } else {
+                ggml_backend_tensor_set(pos, ubatch->pos, 0, n_tokens*n_pos_per_embd*ggml_element_size(pos));
+            }
+        }
+    }
+};
+}
+
+void llama_model_sensenova_u1::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SPATIAL, rope_freq_base_spatial);
+    if (hparams.n_embd_head_k() % 4 != 0 || hparams.n_embd_head_v() != hparams.n_embd_head_k()) {
+        throw std::runtime_error("SenseNova U1 requires equal Q/K/V head sizes divisible by four");
+    }
+    switch (hparams.n_layer()) {
+        case 42: type = LLM_TYPE_8B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_sensenova_u1::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
+        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_sensenova_u1::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_sensenova_u1::graph::graph(const llama_model_sensenova_u1 & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    auto normalize = [&](ggml_tensor * x, ggml_tensor * weight) {
+        x = ggml_rms_norm(ctx0, x, hparams.f_norm_rms_eps);
+
+        return ggml_mul(ctx0, x, weight);
+    };
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    ggml_tensor * pos[3]{};
+    auto inp = std::make_unique<llm_graph_input_pos_sensenova_u1>();
+    auto & inp_pos = inp->pos;
+    inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, (int64_t)n_tokens*hparams.n_pos_per_embd());
+    ggml_set_input(inp_pos);
+    res->add_input(std::move(inp));
+
+    for (int axis = 0; axis < 3; ++axis) {
+        pos[axis] = ggml_view_1d(ctx0, inp_pos, n_tokens, axis*n_tokens*sizeof(int32_t));
+    }
+
+    auto normalize_and_rotate = [&](ggml_tensor * x, ggml_tensor * weight) {
+        const int64_t heads = x->ne[1];
+        x = ggml_reshape_4d(ctx0, x, n_embd_head/2, 2, heads, n_tokens);
+        x = normalize(x, ggml_reshape_2d(ctx0, weight, n_embd_head/2, 2));
+        x = ggml_reshape_3d(ctx0, x, n_embd_head, heads, n_tokens);
+
+        // Normalize temporal and spatial halves separately; rotate time, height, then width.
+        for (int axis = 0; axis < 3; ++axis) {
+            const int dims = axis == 0 ? n_embd_head/2 : n_embd_head/4;
+            const int offset = axis == 0 ? 0 : n_embd_head/2 + (axis - 1)*n_embd_head/4;
+            const float base = axis == 0 ? freq_base : model.rope_freq_base_spatial;
+            x = ggml_rope_ext(ctx0, x, pos[axis], nullptr, dims, GGML_ROPE_TYPE_NEOX,
+                    n_ctx_orig, base, 1.0f, 0.0f, 1.0f, 32.0f, 1.0f);
+            ggml_rope_set_offset(x, offset);
+        }
+        return x;
+    };
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        res->t_layer_inp[il] = inpL;
+
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = normalize(inpL, model.layers[il].attn_norm);
+        cb(cur, "attn_norm", il);
+
+        // self-attention
+        {
+            // compute Q and K and RoPE them
+            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
+                    n_embd_head, n_head, n_head_kv, il);
+
+            Qcur = normalize_and_rotate(Qcur, model.layers[il].attn_q_norm);
+            Kcur = normalize_and_rotate(Kcur, model.layers[il].attn_k_norm);
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        // feed-forward network
+        cur = normalize(ffn_inp, model.layers[il].ffn_norm);
+        cb(cur, "ffn_norm", il);
+
+        cur = build_ffn(cur,
+            model.layers[il].ffn_up,   NULL, model.layers[il].ffn_up_s,
+            model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_s,
+            model.layers[il].ffn_down, NULL, model.layers[il].ffn_down_s,
+            NULL,
+            LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = normalize(cur, model.output_norm);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur, model.output_s);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+}

tests/test-llama-archs.cpp

@@ -329,6 +329,9 @@ static gguf_context_ptr get_gguf_ctx(const llm_arch arch, const bool moe) {
         // partial rotary: n_rot must not exceed the indexer key length (64)
         ms.add_kv(LLM_KV_ROPE_DIMENSION_COUNT,       uint32_t(64));
     }
+    if (arch == LLM_ARCH_SENSENOVA_U1) {
+        ms.add_kv(LLM_KV_ROPE_FREQ_BASE_SPATIAL, 10000.0f);
+    }
     ms.add_kv(LLM_KV_ATTENTION_CLAMP_KQV,              1.0f);
     ms.add_kv(LLM_KV_ATTENTION_LAYERNORM_EPS,          1e-5f);
     ms.add_kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,      1e-5f);

tools/mtmd/CMakeLists.txt

@@ -55,6 +55,7 @@ add_library(mtmd
             models/paddleocr.cpp
             models/pixtral.cpp
             models/qwen2vl.cpp
+            models/sensenova-u1.cpp
             models/minimax-m3.cpp
             models/qwen3vl.cpp
             models/ling3vl.cpp

tools/mtmd/clip-impl.h

@@ -547,6 +547,7 @@ enum projector_type {
     PROJECTOR_TYPE_POCKETTTS_SPKENC,
     PROJECTOR_TYPE_POCKETTTS_GEN,
     PROJECTOR_TYPE_MUSE_GLIMMER,
+    PROJECTOR_TYPE_SENSENOVA_U1,
     PROJECTOR_TYPE_COHERE2V,
     PROJECTOR_TYPE_MUSE_SPARK_AUDIO,
     PROJECTOR_TYPE_UNKNOWN,
@@ -621,6 +622,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_POCKETTTS_SPKENC,  "pockettts_spkenc"},
     { PROJECTOR_TYPE_POCKETTTS_GEN,     "pockettts_gen"},
     { PROJECTOR_TYPE_MUSE_GLIMMER,      "muse-glimmer"},
+    { PROJECTOR_TYPE_SENSENOVA_U1,      "sensenova_u1"},
     { PROJECTOR_TYPE_COHERE2V,          "cohere2v"},
     { PROJECTOR_TYPE_MUSE_SPARK_AUDIO,  "muse-spark-audio"},
 };

tools/mtmd/clip.cpp

@@ -997,6 +997,10 @@ static std::unique_ptr<clip_graph> clip_get_graph_builder(clip_ctx * ctx, const
             {
                 builder = std::make_unique<clip_graph_qwen3vl>(ctx, img);
             } break;
+        case PROJECTOR_TYPE_SENSENOVA_U1:
+            {
+                builder = std::make_unique<clip_graph_sensenova_u1>(ctx, img);
+            } break;
         case PROJECTOR_TYPE_LING3VL:
             {
                 builder = std::make_unique<clip_graph_ling3vl>(ctx, img);
@@ -1795,6 +1799,20 @@ struct clip_model_loader {
                             LOG_WRN("%s: more info: https://github.com/ggml-org/llama.cpp/issues/16842\n\n", __func__);
                         }
                     } break;
+                case PROJECTOR_TYPE_SENSENOVA_U1:
+                    {
+                        hparams.n_merge = 2;
+                        hparams.image_resize_algo = RESIZE_ALGO_BICUBIC;
+                        get_u32(KEY_SPATIAL_MERGE_SIZE, hparams.n_merge, false);
+                        get_u32(KEY_IMAGE_MIN_PIXELS, hparams.image_min_pixels);
+                        get_u32(KEY_IMAGE_MAX_PIXELS, hparams.image_max_pixels);
+                        hparams.rope_theta = 10000.0f;
+                        hparams.set_limit_image_tokens(16, 4096);
+                        if (hparams.n_merge != 2) {
+                            throw std::runtime_error(string_format("%s: SenseNova U1 requires spatial_merge_size=2, got %u", __func__, hparams.n_merge));
+                        }
+                        hparams.set_warmup_n_tokens(16*16);
+                    } break;
                 case PROJECTOR_TYPE_MINIMAX_M3:
                     {
                         hparams.n_merge = 2; // spatial_merge_size
@@ -3837,6 +3855,11 @@ struct clip_model_loader {
                     }
 
                 } break;
+            case PROJECTOR_TYPE_SENSENOVA_U1:
+                {
+                    model.mm_0_w = get_tensor(string_format(TN_LLAVA_PROJ, 0, "weight"));
+                    model.mm_0_b = get_tensor(string_format(TN_LLAVA_PROJ, 0, "bias"));
+                } break;
             default:
                 GGML_ASSERT(false && "unknown projector type");
         }
@@ -4311,6 +4334,7 @@ int clip_n_output_tokens_x(const clip_ctx * ctx, const clip_image_f32 * img) {
         case PROJECTOR_TYPE_QWEN2VL:
         case PROJECTOR_TYPE_QWEN25VL:
         case PROJECTOR_TYPE_QWEN3VL:
+        case PROJECTOR_TYPE_SENSENOVA_U1:
         case PROJECTOR_TYPE_LING3VL:
         case PROJECTOR_TYPE_EXAONE4_5:
         case PROJECTOR_TYPE_MIMOVL:
@@ -4341,6 +4365,7 @@ int clip_n_output_tokens_y(const clip_ctx * ctx, const clip_image_f32 * img) {
         case PROJECTOR_TYPE_QWEN2VL:
         case PROJECTOR_TYPE_QWEN25VL:
         case PROJECTOR_TYPE_QWEN3VL:
+        case PROJECTOR_TYPE_SENSENOVA_U1:
         case PROJECTOR_TYPE_LING3VL:
         case PROJECTOR_TYPE_EXAONE4_5:
         case PROJECTOR_TYPE_MIMOVL:
@@ -4426,6 +4451,7 @@ int clip_n_output_tokens(const clip_ctx * ctx, const clip_image_f32 * img) {
         case PROJECTOR_TYPE_QWEN2VL:
         case PROJECTOR_TYPE_QWEN25VL:
         case PROJECTOR_TYPE_QWEN3VL:
+        case PROJECTOR_TYPE_SENSENOVA_U1:
         case PROJECTOR_TYPE_LING3VL:
         case PROJECTOR_TYPE_EXAONE4_5:
         case PROJECTOR_TYPE_MIMOVL:
@@ -5525,6 +5551,19 @@ bool clip_encode(struct clip_ctx * ctx, struct clip_encode_params * params) {
                 }
                 set_input_i32("pos_w", pos_data);
             } break;
+        case PROJECTOR_TYPE_SENSENOVA_U1:
+            {
+                const int n_patches_per_row = image_size_width / patch_size;
+                std::vector<int> pos_data(n_pos);
+                for (int i = 0; i < n_pos; i++) {
+                    pos_data[i] = i / n_patches_per_row;
+                }
+                set_input_i32("pos_y", pos_data);
+                for (int i = 0; i < n_pos; i++) {
+                    pos_data[i] = i % n_patches_per_row;
+                }
+                set_input_i32("pos_x", pos_data);
+            } break;
         case PROJECTOR_TYPE_DEEPSEEK4V:
             {
                 // set the 2D positions (mrope layout, only the first 2 channels are used)
@@ -6411,6 +6450,8 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
         case PROJECTOR_TYPE_JANUS_PRO:
         case PROJECTOR_TYPE_YOUTUVL:
             return ctx->model.mm_1_b->ne[0];
+        case PROJECTOR_TYPE_SENSENOVA_U1:
+            return ctx->model.hparams.projection_dim;
         case PROJECTOR_TYPE_QWEN3VL:
             // main path + deepstack paths
             return ctx->model.mm_1_b->ne[0] * (1 + ctx->model.n_deepstack_layers);

tools/mtmd/models/models.h

@@ -50,6 +50,11 @@ struct clip_graph_qwen3vl : clip_graph_qwen2vl {
     ggml_cgraph * build() override;
 };
 
+struct clip_graph_sensenova_u1 : clip_graph {
+    clip_graph_sensenova_u1(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph(ctx, img) {}
+    ggml_cgraph * build() override;
+};
+
 struct clip_graph_ling3vl : clip_graph_qwen3vl {
     clip_graph_ling3vl(clip_ctx * ctx, const clip_image_f32 & img) : clip_graph_qwen3vl(ctx, img) {}
     ggml_cgraph * build() override;

tools/mtmd/models/sensenova-u1.cpp

@@ -0,0 +1,40 @@
+#include "models.h"
+
+ggml_cgraph * clip_graph_sensenova_u1::build() {
+    GGML_ASSERT(model.patch_embeddings_0 != nullptr);
+    GGML_ASSERT(model.patch_bias != nullptr);
+    GGML_ASSERT(model.mm_0_w != nullptr);
+    GGML_ASSERT(model.mm_0_b != nullptr);
+    GGML_ASSERT(hparams.n_merge == 2);
+
+    ggml_tensor * pos_x = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
+    ggml_set_name(pos_x, "pos_x");
+    ggml_set_input(pos_x);
+
+    ggml_tensor * pos_y = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_patches);
+    ggml_set_name(pos_y, "pos_y");
+    ggml_set_input(pos_y);
+
+    ggml_tensor * inp = build_inp();
+    inp = ggml_gelu(ctx0, inp);
+
+    // Apply adjacent-pair 2D RoPE to patch embeddings before downsampling.
+    inp = ggml_reshape_4d(ctx0, inp, n_embd, 1, n_patches, n_batch);
+    inp = build_rope_2d(ctx0, inp, pos_x, pos_y, hparams.rope_theta, false);
+    inp = ggml_reshape_3d(ctx0, inp, n_embd, n_patches, n_batch);
+    cb(inp, "patch_embd_rope", -1);
+
+    inp = ggml_reshape_4d(ctx0, inp, n_embd, n_patches_x, n_patches_y, n_batch);
+    inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 2, 0, 1, 3));
+    inp = ggml_conv_2d(ctx0, model.mm_0_w, inp,
+            hparams.n_merge, hparams.n_merge, 0, 0, 1, 1);
+
+    const int n_output_patches = n_patches / (hparams.n_merge * hparams.n_merge);
+    inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 2, 0, 3));
+    inp = ggml_reshape_3d(ctx0, inp, n_mmproj_embd, n_output_patches, n_batch);
+    inp = ggml_add(ctx0, inp, model.mm_0_b);
+    cb(inp, "proj_out", -1);
+
+    ggml_build_forward_expand(gf, inp);
+    return gf;
+}

tools/mtmd/mtmd.cpp

@@ -225,6 +225,7 @@ enum mtmd_pos_type {
     MTMD_POS_TYPE_NORMAL,    // number of positions equals to number of tokens
     MTMD_POS_TYPE_MROPE,     // qwen-vl mrope style, each image takes max(t,h,w) position indexes
     MTMD_POS_TYPE_HUNYUANVL, // HunyuanVL mrope + BOI/EOI/newline layout with XD-RoPE dim-3
+    MTMD_POS_TYPE_SENSENOVA_U1, // U1 uses zero-based H/W and one temporal position per image
     MTMD_POS_TYPE_CANVAS,    // MiniCPM-V 4.7: overview + slices in one chunk, sharing one 2D canvas (see mtmd_image_tokens::canvas_tile_grid)
     MTMD_POS_TYPE_COUNT,     // for validation
 };
@@ -751,6 +752,12 @@ struct mtmd_context {
                     img_end = "]<]end of image[>[";
                     image_preproc = std::make_unique<mtmd_image_preprocessor_dyn_size>(ctx_v);
                 } break;
+            case PROJECTOR_TYPE_SENSENOVA_U1:
+                {
+                    img_beg = "<img>";
+                    img_end = "</img>";
+                    image_preproc = std::make_unique<mtmd_image_preprocessor_dyn_size>(ctx_v);
+                } break;
             case PROJECTOR_TYPE_MUSE_GLIMMER:
                 {
                     img_beg = "<|image_start|>";
@@ -1615,10 +1622,15 @@ struct mtmd_tokenizer {
                     image_tokens->ny = 1;
                 }
                 image_tokens->pos = ctx->pos_type;
+                // U1 uses M-RoPE, but unlike Qwen its image H/W coordinates are
+                // zero-based and the whole image occupies one temporal position.
+                if (ctx->proj_type_v() == PROJECTOR_TYPE_SENSENOVA_U1) {
+                    image_tokens->pos = MTMD_POS_TYPE_SENSENOVA_U1;
+                }
                 // HunyuanVL wraps the image grid with BOI/EOI and adds one newline per row,
                 // and uses XD-RoPE (dim-3 = image index). Override the position type so that
                 // n_tokens() and mtmd_image_tokens_get_decoder_pos pick the HunyuanVL layout.
-                if (ctx->proj_type_v() == PROJECTOR_TYPE_HUNYUANVL) {
+                else if (ctx->proj_type_v() == PROJECTOR_TYPE_HUNYUANVL) {
                     image_tokens->pos       = MTMD_POS_TYPE_HUNYUANVL;
                     image_tokens->image_idx = n_images_added;
                     GGML_ASSERT(n_tokens == (size_t)image_tokens->n_tokens());
@@ -2398,6 +2410,7 @@ bool mtmd_decode_use_non_causal(const mtmd_context * ctx, const mtmd_input_chunk
             }
         case PROJECTOR_TYPE_GEMMA4UV:
         case PROJECTOR_TYPE_GEMMA3:
+        case PROJECTOR_TYPE_SENSENOVA_U1:
         case PROJECTOR_TYPE_DEEPSEEK4V:
             return true;
         default:
@@ -2761,6 +2774,13 @@ mtmd_decoder_pos mtmd_image_tokens_get_decoder_pos(const mtmd_image_tokens * ima
                 pos.y = pos_0 + i;
                 pos.z = pos_0 + i;
             } break;
+        case MTMD_POS_TYPE_SENSENOVA_U1:
+            {
+                pos.t = pos_0;
+                pos.x = i % image_tokens->nx;
+                pos.y = i / image_tokens->nx;
+                pos.z = 0;
+            } break;
         case MTMD_POS_TYPE_HUNYUANVL:
             {
                 // HunyuanVL layout: [BOI] [row0 tokens + newline] ... [row(ny-1) tokens + newline] [EOI]
@@ -2813,6 +2833,8 @@ llama_pos mtmd_image_tokens_get_n_pos(const mtmd_image_tokens * image_tokens) {
             return std::max(image_tokens->nx, image_tokens->ny);
         case MTMD_POS_TYPE_NORMAL:
             return image_tokens->n_tokens();
+        case MTMD_POS_TYPE_SENSENOVA_U1:
+            return 1;
         case MTMD_POS_TYPE_HUNYUANVL:
             // HunyuanVL: the sequential (dim-0) position advances by the full token count
             // (includes BOI/EOI and row newline tokens), not by max(nx, ny)